CN116204825A - Production line equipment fault detection method based on data driving - Google Patents
Production line equipment fault detection method based on data driving Download PDFInfo
- Publication number
- CN116204825A CN116204825A CN202310122557.1A CN202310122557A CN116204825A CN 116204825 A CN116204825 A CN 116204825A CN 202310122557 A CN202310122557 A CN 202310122557A CN 116204825 A CN116204825 A CN 116204825A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- production line
- value
- sampling
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000004519 manufacturing process Methods 0.000 title claims abstract description 78
- 238000001514 detection method Methods 0.000 title claims abstract description 59
- 238000000034 method Methods 0.000 claims abstract description 116
- 238000005070 sampling Methods 0.000 claims abstract description 51
- 238000012549 training Methods 0.000 claims abstract description 42
- 238000012360 testing method Methods 0.000 claims abstract description 9
- 230000002159 abnormal effect Effects 0.000 claims abstract description 8
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 238000012795 verification Methods 0.000 claims abstract description 8
- 238000012544 monitoring process Methods 0.000 claims abstract description 7
- 238000007637 random forest analysis Methods 0.000 claims description 46
- 238000004422 calculation algorithm Methods 0.000 claims description 28
- 230000008569 process Effects 0.000 claims description 17
- 238000005457 optimization Methods 0.000 claims description 16
- 210000002569 neuron Anatomy 0.000 claims description 14
- 238000003066 decision tree Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 238000002474 experimental method Methods 0.000 claims description 9
- 230000001419 dependent effect Effects 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 7
- 238000000540 analysis of variance Methods 0.000 claims description 6
- 238000004458 analytical method Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000013145 classification model Methods 0.000 claims description 6
- 238000012706 support-vector machine Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 5
- 230000011218 segmentation Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 238000009434 installation Methods 0.000 claims description 4
- 238000009826 distribution Methods 0.000 claims description 3
- 238000005259 measurement Methods 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 2
- 230000008447 perception Effects 0.000 claims description 2
- 230000001953 sensory effect Effects 0.000 claims 1
- 238000011425 standardization method Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 description 8
- 230000007547 defect Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 238000003745 diagnosis Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 238000007405 data analysis Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 238000013499 data model Methods 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 238000009776 industrial production Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000013178 mathematical model Methods 0.000 description 2
- 230000006855 networking Effects 0.000 description 2
- 238000003908 quality control method Methods 0.000 description 2
- 230000010076 replication Effects 0.000 description 2
- 239000002699 waste material Substances 0.000 description 2
- 230000032683 aging Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000037396 body weight Effects 0.000 description 1
- 239000003245 coal Substances 0.000 description 1
- 238000005094 computer simulation Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000012372 quality testing Methods 0.000 description 1
- 239000002994 raw material Substances 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000010183 spectrum analysis Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Automation & Control Theory (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a production line equipment fault detection method based on data driving, which comprises the steps of firstly collecting operation data of each equipment on a production line in real time, wherein the operation data comprise monitoring data during normal operation and abnormal operation of the equipment; then preprocessing the data; feature selection is carried out on the training set; data sampling is carried out by adopting a method combining over-sampling and under-sampling aiming at a training set; inputting the sampling data into a fault detection model of production line equipment, and training the model; defining different performance indexes, evaluating the advantages and disadvantages of the fault detection model of the production line equipment by using the precision, and selecting the model with the best performance as a test model; finally, verifying the selected model by using the collected and tidied verification set; judging whether the model meets the requirement or not; if the model meets the requirement, adopting; and if the model does not meet the requirements, discarding. The invention can realize the detection and early warning of the running state of the electrical equipment on the production line by collecting the state data of the electrical equipment.
Description
Technical Field
The invention belongs to the technical field of production equipment fault detection, relates to a production line equipment fault detection method, and in particular relates to a production line equipment fault detection method based on data driving in the electrical industry.
Background
As people's lives are increasingly dependent on industrial products, so is the desire for high quality products. Therefore, it is necessary to provide a healthy production line with a low failure rate for equipment on the production line in the electrical industry. However, during the production of the pipeline operation, the failure of the electrical equipment is often unavoidable. Because once equipment is in trouble, a certain amount of economic loss is caused to the production line, and the waste of raw materials of the product caused by high failure rate also leads to the increase of defective rate of the product and energy consumption of the equipment.
Based on the above phenomena and reasons, an effective quality control strategy for production line management is in need of research. During production in a production line, different methods may be used to monitor for faults in electrical equipment. Sometimes, a manual detection method is adopted to detect equipment faults, but the method is generally poor in effect, high in price and time-consuming. However, if specialized quality testing equipment is used, extensive adjustments to the production line may be required, as well as high capital investment. Based on this, chun et al developed a solution that can perform a failure check for each equipment operation link and shipment only a failure-free checked product (reference 1). However, kang et al (reference 2) believe that the equipment may still randomly fail and be unpredictable due to unsatisfactory equipment inspection procedures, inspection quality control standard deviations, and continual changes in production environment. The result of the lack of the equipment failure detection technology may cause user dissatisfaction and economic disputes (reference 3). Thus, techniques for identifying equipment failures on a production line using low cost and efficient methods are challenging.
In order to overcome the above-described problem of performing equipment fault diagnosis by using a manual method, technical means and methods related to predictive analysis are increasingly applied to different fields and scenes (references 4 to 6). These predictive models are well predicted and have good effects against equipment failure in the industrial process (reference 7). Analysis of important variables in the model by Kang et al (reference 8) helps to find the root cause of equipment failure and to improve the quality of future products. Research shows that related algorithms in the machine learning field are likely to produce good effects on the industrial production line for equipment fault diagnosis, evaluation, product quality prediction and other problems. On an industrial line, most equipment types produce large amounts of industrial data. When equipment faults occur, certain equipment abnormal data are often generated in the industrial Internet and on industrial production lines. Therefore, the prediction model can be constructed by using the generated data better by adopting a machine learning correlation algorithm. By using the prediction model, the production line can be prevented from being additionally modified, and meanwhile, the additional labor cost investment is reduced.
The current algorithms for equipment anomaly detection are mainly of three types:
(1) A process model-based method: such methods compare the output of the measurement system with the output of the mathematical model. The residual of the comparison result is then used to adjust and refine the mathematical model. Many studies have applied different process model-based approaches, including parity equations (reference 9), state observers (reference 10) and parameter estimation (references 11-12).
(2) Knowledge-based methods: such methods are rule-based, relying primarily on expert knowledge. Such models are easy to interpret and have high operating efficiency. However, such methods are not flexible enough and are expensive to maintain. Angeli et al (reference 13) developed an on-line system for applying the above method to equipment fault diagnosis. The results of the work according to Milkovic et al (reference 14) indicate that methods based on expert knowledge are more suitable for well-defined processes.
(3) Data-driven based methods: such methods can be categorized into subclasses of methods such as signal analysis, spectral analysis, pattern analysis, and the like. Isermann et al (reference 15) have given some studies to identify equipment faults by analyzing normal and fault signals from sensors.
The manual detection is that after equipment on the assembly line breaks down, a worker checks the whole assembly line, and overhauls and replaces the broken equipment, so that the defects of time and labor waste and high overhauling cost are the biggest defects.
The automatic detection is to locate the specific equipment and the fault position (but not the specific components to be replaced) of the equipment fault, but the time of the hidden trouble of the equipment is not known, the equipment fault caused by the problem of the specific components is not known, the time and the fault position cannot be accurately located, and once the fault is automatically detected, the components are required to be replaced by manually locating the fault position.
Reference is made to:
[1]Chun,Y.H.(2016).Improved method of estimating the product quality after multiple inspections.International Journal of Production Research,54(19),5686–5696.
[2]Kang,S.,Kim,E.,Shim,J.,Chang,W.,&Cho,S.(2018).Product failure prediction with missing data.International Journal of Production Research,56(14),4849–4859.
[3]Kang,S.,Kim,E.,Shim,J.,Chang,W.,&Cho,S.(2018).Product failure prediction with missing data.International Journal of Production Research,56(14),4849–4859.
[4]K¨oksal,G.,Batmaz,˙I.,&Testik,M.C.(2011).A review of data mining applications for quality improvement in manufacturing industry.Expert systems with Applications,38(10),13448–13467.
[5]Choudhary,A.K.,Harding,J.A.,&Tiwari,M.K.(2009).Data mining in manufacturing:Areview based on the kind of knowledge.Journal of Intelligent Manufacturing,20(5),501.
[6]Kusiak,A.(2006).Data mining:Manufacturing and service applications.International Journal of Production Research,44(18–19),4175–4191.
[7]Lughofer,E.,Pollak,R.,Zavoianu,A.C.,Meyer-Heye,P.,Z¨orrer,H.,Eitzinger,C.,...&Radauer,T.(2017,June).Self-adaptive time-series based forecast models for predicting quality criteria in microfluidics chip production.In:2017 3rd IEEE International Conference on Cybernetics(CYBCONF)(pp.1-8).IEEE.
[8]Kang,S.,Kim,E.,Shim,J.,Chang,W.,&Cho,S.(2018).Product failure prediction with missing data.International Journal of Production Research,56(14),4849–4859.
[9]Frank,P.M.(1990).Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy:A survey and some new results.Auto-matica,26(3),459–474.
[10]Isermann,R.(2005).Model-based fault-detection and diagnosis–status and applications.Annual Reviews in control,29(1),71–85.
[11]Isermann,R.(2006).Fault-diagnosis systems:An introduction from fault detection to fault tolerance.Springer Science&Business Media.
[12]Venkatasubramanian,V.,Rengaswamy,R.,Yin,K.,&Kavuri,S.N.(2003).A review of process fault detection and diagnosis:Part I:Quantitative model-based methods.Computers&Chemical Engineering,27(3),293–311.
[13]Angeli,C.(2010).Diagnostic expert systems:From expert’s knowledge to real-time systems.Advanced Knowledge Based Systems:Model,Applications&Zesearch,1,50–73.
[14]Miljkovi′c,D.(2011).Fault detection methods:A literature survey.In 2011Proceedings of the34th international convention MIPRO(pp.750-755).IEEE.Natekin,A.,&Knoll,A.(2013).Gradient Boosting Machines,A Tutorial.Frontiers in neurorobotics,7,21.
[15]Isermann,R.(2006).Fault-diagnosis systems:An introduction from fault detection to fault tolerance.Springer Science&Business Media.
disclosure of Invention
Aiming at the problems of flexible detection modes and challenges of controlling detection cost in the fault detection of production line equipment in the electrical industry, and combining the defects of the existing detection technology (including manual detection, automatic detection and the like), the invention provides a fault detection method of the production line equipment based on data driving.
The technical scheme adopted by the method is as follows: a production line equipment fault detection method based on data driving comprises the following steps:
step 1: collecting operation data of each device on the production line in real time, wherein the operation data comprise monitoring data of the device in normal operation and abnormal operation;
Step 2: data preprocessing, including data segmentation, data normalization and data classification;
the data segmentation is to divide the data into a training set, a testing set and a verification set; the data classification is to divide the known result into two types of data with equipment failure and without equipment failure;
step 3: selecting characteristics for the training set, wherein the characteristics comprise voltage, current, power, frequency, electric energy and required quantity;
the method for selecting ANOVA f-score by using variance analysis features is used for selecting a plurality of related features, judging whether a feature is important for dependent variables or not by using scores, and selecting the percentage of the variable with the highest score of the feature f as a training feature;
step 4: data sampling is carried out by adopting a method combining over-sampling and under-sampling aiming at a training set;
step 5: inputting the sampling data into the fault detection model of the production line equipment, and training the model;
the production line equipment fault detection model is as follows:
wherein, 
phase voltage U, line voltage U l Average phase voltage->
Average line voltage->
Phase current I, average phase current->
Active power P, reactive power Q, apparent power S, power factor PF; actual frequency F of various power reference values r Sampling frequency F of various sensors s The method comprises the steps of carrying out a first treatment on the surface of the Active power EP, reactive power EQ, combined active total power W EP Combined reactive power total power W EQ Apparent total electric energy W VAh The method comprises the steps of carrying out a first treatment on the surface of the Active demand P xl Reactive power demand Q xl Apparent required amount S xl The method comprises the steps of carrying out a first treatment on the surface of the The "|" in the value set of S indicates or indicates that the value of S can be replaced by any value in the set; t=t, T represents the characteristic value sampling period corresponding to the value of S, therefore +.>
Representing the original value of a certain position i at initial installation, < >>
Representation->
The difference of (1) if->
σ l For a certain kindThe sensor to be monitored can set an early warning threshold value; l= { L num The position of the detecting sensor is marked by a string of character string code set L num Indicating num= { f n ～w n ～pl n ～ep n ～c n }，f n Indicating the factory number, w n Indicating the number of workshops, pl n Representing pipeline number, ep n Indicating the device number, c n Representing the component number;
the training process comprises the processes of input, model wild configuration, numerical calculation, data comparison, parameter optimization and output; firstly, inputting parameters:
model wildcard is then performed:
Step 6: defining different performance indexes, evaluating the advantages and disadvantages of the fault detection model of the production line equipment by using the precision, and selecting the model with the best performance as a test model;
step 7: verifying the selected model by using the collected and tidied verification set; judging whether the model meets the requirement or not; the classification accuracy is greater than or equal to a threshold value, and the model requirement is considered to be met;
if the model meets the requirement, adopting;
and if the model does not meet the requirements, discarding.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention can not only know the fault time of the equipment, but also know the specific components of the equipment, and can also predict when the equipment will fail, the fault probability of the components in the equipment and the aging degree of the equipment and the components;
2. the invention provides a production line equipment fault detection method based on data driving in the electrical industry, which can realize detection and early warning of the running state of electrical equipment on a production line by collecting state data in the electrical equipment;
3. the data model and the method proposed in the invention: the thought and the method comprising a data sampling method, a feature selection method, a data analysis and preprocessing can be transplanted to other production and manufacturing industry equipment to detect and early warn the running state of the production and manufacturing equipment;
4. The method and the method for detecting the equipment faults of the production line based on data driving are applicable to the technology and the method for detecting the equipment faults of other industrial manufacturing enterprises on the production line in the aspects of model configuration, model optimization and model verification.
Drawings
FIG. 1 is a flow chart of an embodiment of the invention;
FIG. 2 is a schematic diagram of training a fault detection model of production line equipment according to an embodiment of the present invention;
FIG. 3 is a graph of experimental results of random forest algorithm before and after sampling according to an embodiment of the present invention;
FIG. 4 is a graph of experimental results of GDBT algorithm before and after sampling according to an embodiment of the present invention;
FIG. 5 is a graph of the results of a random forest oversampling experiment before and after feature selection in an embodiment of the present invention;
FIG. 6 is a PR curve and ROC plot for various model combinations and modes of embodiment of the invention.
Detailed Description
In order to facilitate the understanding and practice of the invention, those of ordinary skill in the art will now make further details with reference to the drawings and examples, it being understood that the examples described herein are for the purpose of illustration and explanation only and are not intended to limit the invention thereto.
As industrial field device operational data sets in the electrical industry exist in different modes and are discovered with the aid of technology in the relevant fields of machine learning and the like. Thus, a large number of data points are required to be used as the study base and precondition for the study of the present invention during the data training phase.
Aiming at the problems of flexible detection modes and challenges of controlling detection cost in the fault detection of production line equipment in the electrical industry, and combining the defects of the existing detection technology (including manual detection, automatic detection and the like), the invention provides a fault detection method of the production line equipment based on data driving. The implementation process of the method mainly comprises the stages of data preprocessing, feature selection, data sampling, parameter optimization and the like. The data of the invention are all networking state data generated in the running process of equipment in the electrical industry, and the data are used as data sources of the method for detecting the equipment faults of the production line.
Referring to fig. 1, the method for detecting the fault of the production line equipment based on the data driving provided by the invention comprises the following steps:
step 1: collecting operation data of each device on the production line in real time, wherein the operation data comprise monitoring data of the device in normal operation and abnormal operation;
this phase is the data preparation phase of all the work of the invention. Various sensors are additionally arranged on various devices in a production line in the electrical industry to serve as a device for monitoring the running state of the devices. When the equipment starts to operate after being started, the sensors can realize an ad hoc industrial internet through a built-in networking module, and collect operation data of all electrical equipment on a pipeline in real time, wherein the data comprise monitoring data sets during normal operation and abnormal operation (equipment failure) of the equipment.
Production line equipment operating state data in the electrical industry is mostly derived from semiconductor sensors installed at the production or operating end of the equipment, which are mainly used to monitor the operating state of the electrical equipment and the manufacturing process of the production line. At the same time, the manufacturing process data generated by the sensor can also be fed back and used for fault detection of the electrical equipment. Through extensive research basis and results in the early stage, four challenging problems need to be overcome in order to utilize manufacturing process data generated by production manufacturing equipment on a pipeline for equipment failure detection. The method comprises the following steps:
(1) the dataset has the problem of too high a feature-sample ratio. Research has found that features account for 1/3 of the data set, which also means that less data information is used to train each feature;
(2) the problem of unbalance of the height of the data set is that the sample data volume of normal operation of the equipment in the data set is far greater than the sample data volume when the equipment fails;
(3) many features in the data set have noise information or uncorrelated signals, i.e., in the device state data, only a few features have absolute correlation, and the correlation between most features is weak. It can thus be stated that most features are noisy or uncorrelated for the dependent variable;
(4) The equipment operating state data in the electrical industry contains a large number of missing values and the value scales for measuring different features vary greatly. Some sensors are graduated in thousands, while some features are graduated in decimal.
Step 2: data preprocessing, including data segmentation, data normalization and data classification;
in this embodiment, the data division is performed by assuming that a data set composed of running state data of the production line equipment is I. Then the data set is divided into training set I by the holdout method train And test set I test (e.g., 80% of the full set for training and 20% for testing). Training set I before training the model train Is the data that has been preprocessed.
In this embodiment, data is normalized, and the missing values are calculated and filled by an average interpolation method. A minimum maximum metric normalization method is used for data normalization. In general, the minimum maximum metric may be used to normalize the feature values between 0,1, as shown in equation 1,
wherein F is i t Indicating the i-th characteristic value in the sampling time slice t among the electrical equipment operation state values,
representing the minimum value of the ith characteristic value in the sample time slice t,/i >
Representing the maximum value of the ith eigenvalue in the sampling time slice t; />
The normalized value obtained by normalizing the ith characteristic value in the sampling time slice t by the above formula is represented.
The purpose of the data classification in this embodiment is to provide a priori knowledge for the training of the model, in other words to divide the known results into two classes of data with and without equipment failure in advance. The purpose of such classification is to tell the model to be trained that the data with equipment failure is to be extracted to the data with equipment failure to extract the features, and the data without equipment failure is to be extracted to the data without equipment failure to extract the features.
In this embodiment, the data classification is performed by using algorithms such as a support vector classifier (Support Vector Classifier, SVC), a multi-layer perceptron (Multilayer Perceptron, MLP), a Random Forest (RF), a gradient lifting Tree (Gradient Boosted Trees, GDBT), and a Random undersampled lifting Tree (Random under-sampling boosting Tree, RUS-Boost Tree). Wherein,
the support vector classifier of the present embodiment is an implementation of a support vector machine for performing classification tasks. It has a low computational cost and a method with a good classification effect in high-dimensional data of a low sample size. As shown in equation 2, is a linear representation of the support vector machine.
f(x)＝w Tr ·x+b(2)
Where x represents the input variable, w represents the weight matrix, b represents the deviation, and Tr represents the transpose of the matrix. The formula is intended to allow for range
The error value is minimized. The present invention is intended to solve the optimization problem using the following equation 3,
wherein P represents a penalty factor,
and->
Respectively representing positive (+) penalty and negative (-) penalty of the ith characteristic value related to training data, w represents a weight matrix, and alpha represents a dynamically adjustable parameter, and the value of the dynamically adjustable parameter is between 0 and 1. Min represents that equation 3 needs to be satisfied +.>
Is>
(/>
To be a dynamically adjustable parameter with an upper limit of values), the minimum value of equation 3 is taken. X is x i And y i Input variables and output variables respectively representing the ith feature value of training data; n represents the total number of feature sets.
The multi-layer perceptron of the embodiment is also called a multi-layer perceptron neural network, and can be widely applied in recent years because the advanced deep learning algorithm can cope with a plurality of problems and can increase the system calculation force. However, neural network related algorithms generally have certain requirements on the data volume of the training data. The multi-layer sensor proposed by the present invention is given by the following formula 4,
wherein eta j A weight coefficient, a representing the neuron j of the multi-layer perception neural network in the hidden layer j Represents the activation function of the corresponding neuron j, n represents the number of neurons in the input layer, m represents the number of neurons in the hidden layer, w i Input layer variable x representing input to neural network using neuron i i (k) Weights, x i (k) Representing input layer variables, w, input by neuron i as neural network 0 Representing the deviation of the input layer, eta 0 Indicating the deviation of the output layer.
The random forest of the present embodiment is a tree-based method and is used for guidance during training. A random forest consists of a number of trees, each tree using only a subset of all features. Each tree generates one prediction, so the final prediction is the set of all predictions. Random forests consist of three main parameters, the size of the tree, the number of predicted variables and the depth of the tree, respectively. Because the random forest can process unbalanced data by adjusting the weight of each class, the unbalanced data is classified, the performance of the random forest is generally better than that of a single tree prediction algorithm with weights and balanced random forests, and the classification result is better than that of other algorithms. As shown in equation 5, a random forest algorithm to be constructed by the present invention is given.
As can be seen from equation 5, D (x) represents the combination of random forests as classification models, D l Representing a decision tree classification model corresponding to a single decision tree l, wherein H represents an indication function, and x represents an input variable; z represents an output variable, and small Z represents the scope of Z, which is a collective distribution of Z, and the meaning of Z is that the scope of all output Z takes on the scope of small Z; l represents the total set of decision trees or the total number of decision trees.
The gradient-lifted tree of this embodiment converts a series of weak learners into stronger learners. Each GDBT tree may improve the prediction results of previous tree algorithms. This makes the performance of the GDBT tree more flexible and increasingly popular in recent years. GDBT is mainly used to perform prognostic genetic tasks, which are often unbalanced classification problems, and has better classification performance than random forests and support vector machines. The GBDT algorithm aims to minimize the regularized objective function. As shown in formula 6, a GDBT tree is to be constructed according to the present invention.
Wherein, assuming a given dataset has n instances and d features, μ (z' i ,z i ) Is a given convex type loss function. Wherein, 
i.e. representing the parameter z' i ,z i The constructed parameter pairs are respectively composed of characteristic sets
And example set->
And (3) generating. />
It is a regularization term, each β j Is a variable associated with the decision tree. Wherein χ is j Representing the variable beta j Corresponding regularization parameters, V j Representing the variable beta j The weights of the corresponding leaf nodes, i and j, respectively represent different node numbers in the random forest.
The random undersampled lifting tree (RUSBT) of the present embodiment is a complex tree with lifting and undersampling functions, which saves time in data preprocessing. The RUSBT is used to balance the classes by performing random undersampling before lifting. During the boosting process, the RUSBT combines multiple weak learners into one strong learner. As shown in formula 7, the invention constructs a random undersampled lifting tree for calculating the pseudo-loss value lambda t . The meaning of T in formulas (7) - (11) is the number of iterative iterations of the algorithm, and T in formula 12 is the maximum number of iterations.
Wherein a is i Representing a point numbered i, b in feature space A i Representing a class label set B with a number i as a class label (class labels are labels that are manually given to a class when classifying data, for example, class labels may be represented by "equipment failure" as 1 and "equipment normal" as 0, where Boolean values 1 and 0 are class labels), each instance in the training dataset may be represented by a tuple (a i ,b i ) The representation is performed. θ t Representing Weak hypotheses (trained in the present invention using the Weak classification learning algorithm Weak-Learn). θ t (a i ,b i ) And theta t (a i B) represents a weak hypothesis θ t Is provided. For example a i (which may be a confidence rating number), E t (i) Representing the weight after t iterations of the ith instance. Equation 8 is used to calculate the weight update parameter β t 。
Wherein lambda is t The meaning of the expression is the same as in equation 7.
The meaning of each parameter in formula 9 is given in formula 7 and formula 8, and will not be described here.
Equation 10 is used to iterate the weight E after t+1 times for the ith instance t+1 (i) And (5) performing specification standardization treatment.
Equation 11 is used to calculate the final hypothesis theta (a),
wherein, Θ (a) returns a weight value after T iterations of weak hypothesis, and other parameters are mentioned in formulas (7) - (10), and are not repeated here.
Step 3: selecting characteristics of the training set, wherein the characteristics comprise voltage, current, power, frequency, electric energy and required quantity;
since equipment operating state data in the electrical industry has a high feature-to-sample ratio, the correlation of a large number of features to dependent variables is low. In this case, feature selection is required to reduce the number of uncorrelated features before the data is applied to model training. Thus, the present invention employs an analysis of variance feature selection ANOVA f-score method for selecting a large number of relevant features. The effect is to use the score to determine if a feature is important to the dependent variable. That is, a higher f-value rejects the zero hypothesis, which also means that the variance of the variable has an effect on the variance of the dependent variable. Thus, the percentage of the highest f-score variable may be selected as the training feature.
In the present invention, analysis of variance ANOVA is used to measure the correlation between one feature and all features. In view of this, the f statistic of the feature can be utilized to satisfy the characteristic of f distribution for saliency checking.
In the electrical industry, data-driven-based production line equipment fault detection methods require extraction of features for equipment fault diagnosis including, but not limited to, the following features:
1) Voltage: comprises a phase voltage U and a line voltage U l Average phase voltage
Average line voltage->
Wherein the phase voltage U can pass the phase voltage harmonic content H U Giving characteristic feedback. When the actual voltage value is stable, but the voltage value acquired by the sensor on the equipment has local maximum value or minimum value at certain moment, the equipment is displayed to be faulty from the angle of data analysis, so that the sensor with instantaneous extremum in the acquisition equipment needs to be replaced;
2) Current flow: comprising phase current I, average phase current
Wherein the phase currents may give characteristic feedback through the phase current harmonic content. And the voltage value
Similarly, if the current value on the electrical device is compared to the original value, a local extremum occurs at a time, then it is assumed that the corresponding sensor needs to be replaced at that time.
3) Power: the system comprises active power P, reactive power Q, apparent power S and power factor PF;
4) Frequency: actual frequency F including various power references r Sampling frequency F of various sensors s ；
5) Electric powerThe energy can be: comprises active power EP, reactive power EQ and combined active total power W EP Combined reactive power total power W EQ Apparent total electric energy W VAh ；
6) The required amount is as follows: including the active demand P xl Reactive power demand Q xl Apparent required amount S xl ；
1-6 are detection characteristic values generated in the application process of the electric equipment in the electric power industry; here, an original value is defined, which refers to a measurement initial value acquired by various sensors at the beginning of installation of the apparatus. In the electrical industry, data acquired by an acquisition device on brand new equipment is generally considered as an accurate value, and the measured value can be used as an original reference value of various characteristic indexes to be used as a reference standard. If a significant characteristic difference (such as sampling extremum, sampling missing, sampling unreadable, etc.) occurs between the detected sampling value and the original value, it can be basically determined that the corresponding electrical component is aged and should be located and replaced.
Besides the characteristic variables which can be defined and extracted according to the mode in the electric power industry are used as the basis for detecting the faults of the equipment, the characteristic variables can be defined and extracted by adopting the mode in other industries such as industries of coal, water conservancy and hydropower, chemical industry and the like, and the fault detection method of the production line equipment is implemented by adopting the fault detection model of the production line equipment in the step 5.
Step 4: data sampling is carried out by adopting a method combining over-sampling and under-sampling aiming at a training set;
as the training data acquired by the operation of equipment on the production line in the electrical industry is highly unbalanced, the data points in all the equipment operation state data are found to be only 7% of equipment abnormal data. In response to this phenomenon, the present invention employs a combination of over-sampling and under-sampling to improve the performance of the algorithm. Details of the method used to implement this step are detailed below:
the undersampling of this embodiment samples all minority class samples and randomly selects an equal number of majority class samples. It then combines the two sample subsets to form a new balanced data set. In the training set, normal samples belong to the majority class and failure samples belong to the minority class. In other words, the undersampling method is implemented on the principle that data is balanced by losing some data points and useful information.
Oversampling of this embodiment balances data by copying a few classes of samples, as opposed to undersampling. Although the oversampling method balances the data, increasing the amount of data; but has the disadvantage that the over-fitting easily occurs due to the mechanical replication of the data. Therefore, the invention intends to use a synthetic minority oversampling technique SMOTE for solving the over-fitting problem. This technique overcomes the over-fitting problem caused by mechanical replication of data points by synthesizing similar data points, i.e., finding its k-neighbors with Euclidean distance through a selected few points, and then creating one or more new points therein.
Step 5: inputting the sampling data into a fault detection model of production line equipment, and training the model;
the production line equipment fault detection model of the embodiment is as follows:
wherein, 
phase voltage U, line voltage U l Average phase voltage->
Average line voltage->
Phase current I, average phase current->
Active power P, reactive power Q, apparent power S, power factor PF; actual frequency F of various power reference values r Sampling frequency of various sensorsF s The method comprises the steps of carrying out a first treatment on the surface of the Active power EP, reactive power EQ, combined active total power W EP Combined reactive power total power W EQ Apparent total electric energy W VAh The method comprises the steps of carrying out a first treatment on the surface of the Active demand P xl Reactive power demand Q xl Apparent required amount S xl The method comprises the steps of carrying out a first treatment on the surface of the The "|" in the value set of S indicates or indicates that the value of S can be replaced by any value in the set; t=t, T represents the characteristic value sampling period corresponding to the value of S, therefore +.>
Representing the original value of a certain position i at initial installation, < >>
Representation->
The difference of (1) if->
σ l A settable early warning threshold value for a certain sensor to be monitored; l= { L num The position of the detecting sensor is marked by a string of character string code set L num Indicating num= { f n ～w n ～pl n ～ep n ～c n }，f n Indicating the factory number, w n Indicating the number of workshops, pl n Representing pipeline number, ep n Indicating the device number, c n Representing the component number; for example, a string 001 ～ 011 ～ 035 ～ 026 ～ 20141021 of codes can be used for identifying the occurrence position of abnormal components at the factory number 001, the workshop number 011, the pipeline number 035, the equipment number 026 and the components number 20141021 (initial use date), and the codes are made into two-dimensional codes, so that fault tracking and remote monitoring of production line equipment can be realized.
Please refer to fig. 2, the training process of the present embodiment includes input, model wildcard, numerical calculation, data comparison, parameter optimization and output processes;
firstly, inputting parameters:
model wildcard is then performed:
in FIG. 2, (-)>(less than or equal to) the comparison result is alternatively selected, sigma l* And representing the target set value of parameter optimization, wherein the target set value is determined by the clustering result of the numerical characteristics.
The output result is indicated to be selected from 1 or 0, if the result is 1, the early warning of the component at the t time l is indicated, and replacement is required to be reminded; if the result is 0, the component is normal and does not need to be replaced.
Training and applying a random forest algorithm with default parameters in the training process; optimizing super parameters in the model (including regularization parameters in the support vector classifier (for determining the strength of regularization) and kernel coefficients (for controlling the width of kernels), and determining the number of estimated quantities and maximum tree depth parameters using a parameter optimization method; optimizing hyper-parameters (comprising the maximum leaf node number and the maximum depth of a single number in a random forest) in the random forest by a Grid Search cross validation (Grid-Search CV) method; utilizing a plurality of super parameters provided by a support vector classifier SVC (Support Vector Classifier, SVC) and performing parameter tuning by using a Grid-Search CV method; parameter optimization is carried out in a Grid Search CV method in an MLP algorithm; the parameters mainly comprise node connectivity Connection, number of units of the neuron, data point Input dimension, and activation functions (including Relu, linear and the like) required in the modeling process.
Step 6: defining different performance indexes, evaluating the advantages and disadvantages of the fault detection model of the production line equipment by using the precision, and selecting the model with the best performance as a test model;
for the classification problem, the invention adopts the confusion matrix to define different performance indexes, and utilizes the precision to evaluate the performance indexes of the classification model. The invention is presented to evaluate the performance of the classification model listed above for use in electrical industry, production line equipment fault detection applications, as in equations (13) - (16), respectively.
Where Acc represents the accuracy of classification and Prec represents the accuracy of model prediction. TP and FP are true and false positives, respectively. The representative meaning is that the number of data pieces in which a real fault occurs in the equipment operation state data is a percentage of the number of predicted fault data pieces.
Where Rec represents the recall of the computational model. TN and FN are true negative and false negative respectively; equation 15 is used to measure the percentage of predicted faults to actual faults. In practical applications, increasing recall tends to decrease accuracy because a lower threshold is required for high recall. Thus, the present invention balances accuracy and recall by introducing F1.
The F1 score is selected as the classification performance index, and the best model should have the highest F1 score as determined by the calculation of equation 15.
Notably, the accuracy rate, prec, in equation 14 and the recall rate, rec, in equation 15 are typically used to evaluate the performance of unbalanced classification tasks.
Step 7: verifying the selected model by using the collected and tidied verification set; judging whether the model meets the requirements, and considering that the model meets the requirements if the classification accuracy is more than or equal to 95%;
if the model meets the requirement, adopting;
and if the model does not meet the requirements, discarding.
In this embodiment, the step 7 specifically includes the following steps:
step 7.1: the experimental results of the random forest algorithm before and after sampling (as shown in fig. 3) were verified. The method adopts the following steps: (1) oversampling (Over-Sampling) and Random Forest (RF), (2) undersampling (underworkings), feature Selection (FS) and Random Forest (RF), (3) using only Random Forest (RF), three sets of models and methods can be seen from the experimental results of experiments conducted with equivalent body weight samples:
the Accuracy (Accuracy) is as follows: (1) (2) (3);
precision (Precision) high-low ordering is: (1) (2) (3);
recall (Recall) high-low ordering is: (3) (2) (1);
the F1 Score (F1-Score) was ranked high and low: (1) (2) (3);
it is clear from this that the experimental results using the model and method of group (1) are generally better, but the model and method of groups (2) and (3) can overcome the deficiencies of group (1) in recall.
Step 7.2: the experimental results of the gradient-lifted tree algorithm before and after sampling were validated (as shown in fig. 4). The method adopts the following steps: (1) oversampling (Over-Sampling) and gradient lifting tree (GDBT), (2) undersampling (underwriter-Sampling), feature Selection (FS) and gradient lifting tree (GDBT), (3) using gradient lifting tree (GDBT) only, three sets of models and methods can be seen from the experimental results of experiments performed with homobody samples:
the Accuracy (Accuracy) is as follows: (1) (3) (2);
precision (Precision) high-low ordering is: (1) (2);
recall (Recall) high-low ordering is: (2) (1);
the F1 Score (F1-Score) was ranked high and low: (1) (2);
from this, it can be seen that the experimental effect is generally better with the model and method of group (1), but the deficiency of recall rate of group (1) can be overcome with the model and method of group (2), and the almost equivalent effect of the model and method of group (1) can still be achieved with the model of group (3).
Step 7.3: the experimental results of the gradient-lifted tree algorithm before and after sampling were validated (as shown in fig. 5). The method adopts the following steps: (1) undersampling (underworkings-Sampling), feature Selection (FS) and Random Forest (RF), (2) undersampling (underworkings-Sampling), random Forest (RF), two sets of models and methods can be seen from the experimental results of experiments performed with samples of the same volume:
The Accuracy (Accuracy) is as follows: (1) (2);
precision (Precision) high-low ordering is: (2) (1);
recall (Recall) high-low ordering is: (1) (2);
the F1 Score (F1-Score) was ranked high and low: (1) (2);
it is clear from this that the experimental results using the model and method of group (1) are generally better, but the model and method of group (2) can overcome the disadvantages of group (1) in terms of accuracy.
Step 7.4: the PR curves and ROC curves (as shown in fig. 6) were validated for various model combinations and modes. The method adopts the following steps: (1) random Forest (RF), oversampling (OS) and Feature Selection (FS); (2) gradient-lifted tree (GDBT), oversampling (OS), and Feature Selection (FS); (3) random Forest (RF), undersampling (US), and Feature Selection (FS); (4) gradient lifting tree (GDBT), undersampling (US), and Feature Selection (FS); (5) random undersampled lifting trees (RUS) and Feature Selection (FS).
Five sets of models and methods can be seen from the experimental results of experiments performed with equivalent body mass samples:
in terms of Recall rate (Recall), the Recall rates of (3), 4) and (5) can reach 100%, but under the same condition, the Accuracy (Accuracy) is ranked as follows: (4) (3) (5). The recall rate of (1) is inferior to (3) (4) (5), only about 78%; worst is (2), only about 68%;
By comparing the ratio of the Accuracy (Accuracy) and the Recall (Recall), the Accuracy is ranked as (2) (1) (4) (3) (5) along with the increase of the Recall,
from this, it can be seen that the experimental results using the (4) th model and method are generally better, and the (1) th model and method, although ultimately achieving better accuracy, have low recall. The defects of the recall rate of the medicine (1) can be overcome by adopting the model and the method of the group (4).
As can be seen by comparing the ROC curves of the models in fig. 6:
when the FPR (true positive ratio) of the model and the method of the group (1) is increased, the TPR (false positive ratio) is also increased, but both the ratios are not high;
the FPR (true positive ratio) of the model and method of group (2) was 0.28 before no TPR (false positive ratio) appeared, but after that TPR was increased to 0.55, but the FPR of group (2) was only up to 0.48, but TPR was as high as 0.8;
no TPR (false positive ratio) occurred before the FPR (true positive ratio) of group (3) model and method was 0.58, but after that the TPR was increased to 0.82, after which the TPR was as high as 0.92 although the FPR reached 1;
FPR (false positive ratio) of group (4) models and methods did not appear TPR (true positive ratio) before 0.15. After this time the FPR goes to 0.23 and its TPR goes up to 0.43. Thereafter, the TPR steadily rises to 1, but the FPR also eventually rises to 0.92;
Before FPR (false positive ratio) of group (5) model and method was 0.44, TPR (true positive ratio) did not appear, and TPR was raised to 0.18. But then at the 0.18 position the FPR also reaches 0.57. Thereafter, as the TPR increases, the FPR steadily increases. Finally, when TPR is 1, FPR also reaches 0.9.
The present invention contemplates techniques and methods related to equipment failure detection on a production line using data driven models in the electrical industry. The invention relates to a core algorithm and key technical means of various aspects such as data analysis, data sampling, data preprocessing, data feature selection, data classification, validity verification of a data model and the like.
It should be understood that the foregoing description of the preferred embodiments is not intended to limit the scope of the invention, but rather to limit the scope of the claims, and that those skilled in the art can make substitutions or modifications without departing from the scope of the invention as set forth in the appended claims.
Claims (10)
1. The production line equipment fault detection method based on data driving is characterized by comprising the following steps of:
Step 1: collecting operation data of each device on the production line in real time, wherein the operation data comprise monitoring data of the device in normal operation and abnormal operation;
step 2: data preprocessing, including data segmentation, data normalization and data classification;
the data segmentation is to divide the data into a training set, a testing set and a verification set; the data classification is to divide the known result into two types of data with equipment failure and without equipment failure;
step 3: selecting characteristics for the training set, wherein the characteristics comprise voltage, current, power, frequency, electric energy and required quantity;
the method for selecting ANOVA f-score by using variance analysis features is used for selecting a plurality of related features, judging whether a feature is important for dependent variables or not by using scores, and selecting the percentage of the variable with the highest score of the feature f as a training feature;
step 4: data sampling is carried out by adopting a method combining over-sampling and under-sampling aiming at a training set;
step 5: inputting the sampling data into the fault detection model of the production line equipment, and training the model;
the production line equipment fault detection model is as follows:
wherein, 
phase voltage U, line voltage U l Average phase voltage
Average line voltage- >
Phase current I, average phase current->
Active power P, reactive power Q, apparent power S, power factor PF; actual frequency F of various power reference values r Sampling frequency F of various sensors s The method comprises the steps of carrying out a first treatment on the surface of the Active power EP, reactive power EQ, combined active total power W EP Combined reactive power total power W EQ Apparent total electric energy W VAh The method comprises the steps of carrying out a first treatment on the surface of the Active demand P xl Reactive power demand Q xl Apparent required amount S xl The method comprises the steps of carrying out a first treatment on the surface of the The "|" in the value set of S indicates or indicates that the value of S can be replaced by any value in the set; t=t, T represents the characteristic value sampling period corresponding to the value of S, therefore +.>
Representing the original value of a certain position i at initial installation, < >>
Representation->
The difference of (1) if->
σ l A settable early warning threshold value for a certain sensor to be monitored; l= { L num The position of the detecting sensor is marked by a string of character string code set L num Indicating num= { f n ～w n ～pl n ～ep n ～c n }，f n Indicating the factory number, w n Indicating the number of workshops, pl n Representing pipeline number, ep n Indicating the device number, c n Representing the component number;
the training process comprises the processes of input, model wild configuration, numerical calculation, data comparison, parameter optimization and output; firstly, inputting parameters:
model wildcard is then performed:
step 6: defining different performance indexes, evaluating the advantages and disadvantages of the fault detection model of the production line equipment by using the precision, and selecting the model with the best performance as a test model;
step 7: verifying the selected model by using the collected and tidied verification set; judging whether the model meets the requirement or not; the classification accuracy is greater than or equal to a threshold value, and the model requirement is considered to be met;
if the model meets the requirement, adopting;
and if the model does not meet the requirements, discarding.
2. The data-driven production line equipment failure detection method according to claim 1, wherein: step 2, normalizing the data, and calculating and filling a missing value by adopting an average interpolation method; a minimum and maximum value measurement standardization method is adopted for data standardization;
wherein F is i t Indicating the i-th characteristic value in the sampling time slice t among the electrical equipment operation state values,
representing the minimum value of the ith characteristic value in the sample time slice t,/i >
Representing the maximum value of the ith eigenvalue in the sampling time slice t; />
The normalized value obtained by normalizing the ith characteristic value in the sampling time slice t by the above formula is represented.
3. The data-driven production line equipment failure detection method according to claim 1, wherein: step 2, classifying the data, and executing classification tasks by adopting a support vector machine;
the linear representation form of the support vector machine is as follows:
f(x)＝w Tr ·x+b (2)；
wherein x represents an input variable, w represents a weight matrix, b represents a deviation, and Tr represents a transpose of the matrix;
the support vector machine solves the optimization problem using the following formula,
wherein P represents a penalty factor,
and->
Respectively representing positive penalty and negative penalty of the ith characteristic value related to training data, wherein the I < w > I represents a weight matrix, and alpha represents a dynamic adjustable parameter, and the value of the dynamic adjustable parameter is between 0 and 1; min represents that equation 3 needs to be satisfied +.>
Is>
The minimum value in +_>
Is a dynamically adjustable parameter with a value upper limit; x is x i And y i Input variables and output variables respectively representing the ith feature value of training data; n represents the total number of feature sets.
4. The data-driven production line equipment failure detection method according to claim 1, wherein: the data classification in the step 2 adopts a multi-layer perception neural network N (k);
Wherein eta j Weight coefficient, a, representing neuron j of multi-layer perceptual neural network in hidden layer j Represents the activation function of the corresponding neuron j, n represents the number of neurons in the input layer, m represents the number of neurons in the hidden layer, w i Input layer variable x representing input to a multi-layer sensory neural network with neuron i i (k) Weights, x i (k) Representing input layer variables, w, input by using neuron i as a multi-layer perceptual neural network 0 Representing the deviation of the input layer, eta 0 Indicating the deviation of the output layer.
5. The data-driven production line equipment failure detection method according to claim 1, wherein: the data classification in the step 2 adopts a random forest algorithm;
wherein D (x) represents a combination of random forests as classification models, D l Representing a decision tree classification model corresponding to a single decision tree l, wherein H represents an indication function, and x represents an input variable; z represents an output variable, and small Z represents the scope of Z, which is a collective distribution of Z, and the meaning of Z is that the scope of all output Z takes on the scope of small Z; l represents the total set of decision trees or the total number of decision trees.
6. The data-driven production line equipment failure detection method according to claim 1, wherein: step 2, classifying the data, and converting a series of weak learners into stronger learners by adopting a gradient lifting tree GDBT;
One GDBT tree is:
wherein if a given dataset has n instances and d features, μ (z' i ,z i ) Is the loss function of a given convex shape,
i.e. representing the parameter z' i ,z i The constructed parameter pairs are each composed of a feature set +.>
And example set->
Generating; />
Is a regularization term, each beta j Is a variable, χ, related to the decision tree j Representing the variable beta j Corresponding regularization parameters, V j Representing the variable beta j The weights of the corresponding leaf nodes, i and j, respectively represent different node numbers in the random forest.
7. The data-driven production line equipment failure detection method according to claim 1, wherein: step 2, classifying the data, and combining a plurality of weak learners into a strong learner by adopting a random undersampling lifting tree RUSBT;
calculating a pseudo-loss value lambda by randomly undersampling the lifting tree t The method comprises the following steps:
wherein a is i Representing a point numbered i, b in feature space A i If a number i in the class label set B is a class label, then each instance in the training data set may be labeled with a tuple (a i ,b i ) Representing; θ t Representing weak hypothesis θ t (a i ,b i ) And theta t (a i B) represents a weak hypothesis θ t An output of (2); for example a i ，E t (i) The weight of the ith example after t times of iteration is represented, and t represents the number of repeated iterations; wherein, the class labels are labels which are manually marked for a certain class when data classification is carried out;
Calculating a weight update parameter beta t ：
Calculating the weight E after t+1 times of the ith example iteration t+1 (i)：
Weights E after t+1 iterations for the ith instance t+1 (i) Performing specification standardization treatment;
the final hypothesis theta (a) is calculated,
wherein Θ (a) returns a weight value after T iterations of weak hypothesis, T represents the maximum value of the iteration times.
8. The data-driven production line equipment failure detection method according to claim 1, wherein: in the training process in the step 5, training and applying a random forest algorithm with default parameters; super parameters in the optimization model comprise regularization parameters and nuclear coefficients in the support vector classifier SVC; determining the number of estimated values and the maximum tree depth parameter by using a parameter optimization method; optimizing hyper-parameters in the random forest by a grid search cross-validation method, wherein the hyper-parameters comprise the maximum leaf node number and the maximum depth of a single number in the random forest; performing parameter tuning by using a plurality of super parameters provided by a support vector classifier SVC and a Grid-Search CV method; parameter optimization is carried out in a Grid Search CV method in an MLP algorithm; the parameters mainly comprise node connectivity Connection, neuron unit Number of units, data point Input dimension and activation function.
9. The data-driven production line equipment failure detection method according to claim 1, wherein: defining different performance indexes in step 6 includes:
acc represents the classification accuracy, TP and FP are true positive and false positive respectively, and the representative meaning is that the number of data bars with true faults in the running state data of the equipment accounts for the percentage of the number of predicted fault data bars; TN and FN are true negative and false negative respectively; prec represents the accuracy of model prediction; rec represents the recall of the calculation model, and is used for measuring the percentage of the predicted faults to the actual faults;
balancing precision and recall through F1;
the F1 score is selected as the classification performance indicator, with the best model having the highest F1 score.
10. The data-driven production line equipment fault detection method according to any one of claims 1 to 9, wherein the specific implementation of step 7 comprises the following sub-steps:
step 7.1: verifying the experimental results of random forest algorithms before and after sampling; the method adopts the following steps: (1) oversampling and random forests, (2) undersampling, feature selection and random forests, (3) using only random forests; obtaining experimental results of experiments of three groups of models and methods by adopting an equivalent sample;
Step 7.2: verifying the experimental result of a gradient lifting tree algorithm before and after sampling; the method adopts the following steps: (1) an oversampling and gradient lifting tree, (2) an undersampling, feature selection and gradient lifting tree, (3) a gradient lifting tree is used only; obtaining experimental results of experiments of three groups of models and methods by adopting an equivalent sample;
step 7.3: verifying the experimental result of a gradient lifting tree algorithm before and after sampling; the method adopts the following steps: (1) undersampling, feature selection and random forests, (2) undersampling, random forests; obtaining experimental results of two groups of models and methods in experiments performed by adopting an equivalent sample;
step 7.4: verifying PR curves and ROC curves of various model combinations and modes; the method adopts the following steps: (1) random forest, oversampling and feature selection; (2) gradient lifting tree, over-sampling and feature selection; (3) random forest, undersampling and feature selection; (4) gradient lifting tree, undersampling and feature selection; (5) random undersampling lifting tree and feature selection; experimental results of five sets of models and methods in experiments using equivalent body mass samples were obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310122557.1A CN116204825A (en) | 2023-02-10 | 2023-02-10 | Production line equipment fault detection method based on data driving |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310122557.1A CN116204825A (en) | 2023-02-10 | 2023-02-10 | Production line equipment fault detection method based on data driving |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116204825A true CN116204825A (en) | 2023-06-02 |
Family
ID=86516818
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310122557.1A Pending CN116204825A (en) | 2023-02-10 | 2023-02-10 | Production line equipment fault detection method based on data driving |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116204825A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451142A (en) * | 2023-06-09 | 2023-07-18 | 山东云泷水务环境科技有限公司 | Water quality sensor fault detection method based on machine learning algorithm |
| CN118519411A (en) * | 2024-07-24 | 2024-08-20 | 陕西智引科技有限公司 | Intelligent real-time monitoring system for coal safety production |
-
2023
- 2023-02-10 CN CN202310122557.1A patent/CN116204825A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116451142A (en) * | 2023-06-09 | 2023-07-18 | 山东云泷水务环境科技有限公司 | Water quality sensor fault detection method based on machine learning algorithm |
| CN118519411A (en) * | 2024-07-24 | 2024-08-20 | 陕西智引科技有限公司 | Intelligent real-time monitoring system for coal safety production |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113255848B (en) | Water turbine cavitation sound signal identification method based on big data learning | |
| Zheng et al. | A rolling bearing fault diagnosis method based on multi-scale fuzzy entropy and variable predictive model-based class discrimination | |
| CN114813105A (en) | Gear box fault early warning method and system based on working condition similarity evaluation | |
| Liu et al. | A two-stage approach for predicting the remaining useful life of tools using bidirectional long short-term memory | |
| CN111598150B (en) | Transformer fault diagnosis method considering operation state grade | |
| CN116380445B (en) | Equipment state diagnosis method and related device based on vibration waveform | |
| CN115187832A (en) | Energy system fault diagnosis method based on deep learning and gram angular field image | |
| CN105700518A (en) | Fault diagnosis method during industrial process | |
| CN104596780B (en) | Diagnosis method for sensor faults of motor train unit braking system | |
| CN116204825A (en) | Production line equipment fault detection method based on data driving | |
| CN114091525A (en) | Rolling bearing degradation trend prediction method | |
| CN109298633A (en) | Chemical production process fault monitoring method based on adaptive piecemeal Non-negative Matrix Factorization | |
| CN110503133A (en) | A kind of centrifugal compressor failure prediction method based on deep learning | |
| Wen et al. | A new method for identifying the ball screw degradation level based on the multiple classifier system | |
| Wang et al. | Two-stage imbalanced learning-based quality prediction method for wheel hub assembly | |
| CN115496108A (en) | Fault monitoring method and system based on manifold learning and big data analysis | |
| CN110956331A (en) | Method, system and device for predicting operation state of digital factory | |
| Song et al. | A fault feature extraction method for rolling bearings based on refined composite multi-scale amplitude-aware permutation entropy | |
| CN118244110B (en) | Industrial robot motor fault early warning method and system | |
| CN111474476B (en) | Motor fault prediction method | |
| CN117312972A (en) | Method for identifying health state of scraper conveyor speed reducer | |
| KR20240015457A (en) | Method for creating adaptive isolation filter models for anomaly detection in semiconductor products, method for detecting anomalies in semiconductor products and anomaly detection recording medium and system for the same | |
| CN114660931A (en) | Method and system for diagnosing and identifying petrochemical process fault | |
| Brandao et al. | Fault Diagnosis of Rotary Machines Using Machine Learning | |
| CN118501692B (en) | Servo motor fault diagnosis method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |