Rail transit key equipment fault prediction method and system

Abstract

The invention discloses a rail transit key equipment fault prediction method and system, and the method comprises the steps: carrying out the data collection and marking, and preliminarily calibrating an initial data state credibility coefficient; initializing and defining a feature set representing the health state of the equipment, associating abnormal item points with features, preliminarily calibrating correlation coefficients of the abnormal item points and the features, and performing initial feature extraction; establishing a fault prediction model by adopting a machine learning algorithm; using the established fault prediction model to diagnose main line operation data, outputting an online prediction result, performing data mark correction based on the result, and updating an abnormal item point and feature correlation coefficient and a corresponding initial data state credibility coefficient; and feature updating is carried out based on the corrected mark data, so that the fault prediction model is further updated and trained. According to the method, automatic data marking and data marking updating can be realized, online automatic model updating is realized in combination with a prediction result and actual maintenance feedback, and the model performance is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault prediction, and in particular to a method and system for predicting faults of key rail transit equipment. Background Art

[0002] The normal operation of key rail transit equipment (hereinafter referred to as equipment) is crucial to ensure the safe and punctual operation of trains. Equipment failure will not only lead to train delays and cancellations, but may even cause safety accidents. Therefore, predicting and preventing failures in advance is of great significance for the safe, reliable and efficient operation of rail transit systems.

[0003] Equipment fault prediction technology can identify possible equipment failures. Traditional fault prediction methods based on expert experience have the problems of strong subjectivity and low accuracy; with the development of big data and artificial intelligence technology, data-driven fault prediction methods have gradually become a research hotspot. At present, data-driven equipment fault prediction technology has the following problems: on the one hand, the fault prediction method is relatively simple, the single signal monitoring method has large fluctuations, and the prediction deviation is obvious; on the other hand, fault prediction often lacks high-quality labeled data, and the labeled data is not further fed back according to the actual effect. The quality of the labeled data is difficult to guarantee, which affects the actual model effect. In addition, the prediction model has weak generalization ability. As time goes by and the scene switches, model updates require a lot of manpower and take a long time, and cannot cope with complex and changeable actual working conditions. Summary of the invention

[0004] Purpose of the invention: The first purpose of the present invention is to provide a method for predicting faults of key rail transit equipment, which can realize automatic labeling and updating of data by combining prediction results and actual maintenance feedback, and then realize automatic online updating of models, thereby improving the quality of label data and model prediction performance; The second purpose of the present invention is to provide a system for predicting faults of key rail transit equipment.

[0005] Technical solution: A method for predicting failures of key rail transit equipment described in the present invention comprises the following steps: (1) Data collection and labeling, preliminary calibration of the initial data state credibility coefficient; (2) Based on the experience of equipment diagnosis experts, the feature set that characterizes the health status of the equipment is initialized and defined, the abnormal items are associated with the features, the correlation coefficients between the abnormal items and the features are preliminarily calibrated, and the initial feature extraction is performed; (3) Use machine learning algorithms to establish a fault prediction model; (4) Use the established fault prediction model to diagnose the mainline operation data and output the online prediction results. Based on the results, correct the data labeling, update the correlation coefficient between abnormal items and features, and the corresponding initial data state credibility coefficient; update the features based on the corrected labeled data, and then further update the trained fault prediction model.

[0006] Preferably, the data labeling in step (1) includes: first defining multi-source data labels, wherein the data labels include equipment operating conditions, working modes, and data states; and then performing initial data labeling according to the data labels.

[0007] Preferably, the initial feature extraction in step (2) includes: setting an initial data state credibility coefficient threshold, screening data reaching the initial data state credibility coefficient threshold from the marked data, and preliminarily calculating its feature value.

[0008] Preferably, step (3) includes: using the feature set obtained in step (2) as the input of the fault prediction model to carry out model training; and performing model evaluation based on the external control factor index and the internal control factor index; if the model meets the external control factor index requirements, then the model training is terminated; otherwise, the model internal control factor index is calculated, and the training sample is updated and the model update training is performed.

[0009] Preferably, the external control factor indicators include model accuracy, macro-average or micro-average precision, recall rate, and F1 score; the internal control factor indicator is model credibility.

[0010] Preferably, the training sample update and model update training include: The threshold of the initial data state credibility coefficient is dynamically adjusted according to the model credibility; under the premise of ensuring the number of training samples, samples whose initial data state credibility coefficient does not reach the adjusted threshold are eliminated, and then the model update training is carried out.

[0011] Preferably, the data mark correction in step (4) includes: First, the reliability of the positive prediction is calculated and compared with the set threshold. If the reliability of the positive prediction reaches the set threshold, an early warning is pushed for known abnormal types, and the data is strongly marked. If the reliability of the positive prediction does not reach the set threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, an early warning is pushed for unknown abnormal types, and the data is weakly marked. Secondly, compare the warning push result with the actual maintenance feedback result: if the warning push result is consistent with the actual maintenance feedback result, the strengthening mark is confirmed. First, according to the feature importance matrix output by the model, the abnormal item point and feature correlation coefficient are updated; second, according to the confidence result output by the fault prediction model, the corresponding initial data state credibility coefficient is increased; if the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed and the corresponding initial data state credibility coefficient is lowered.

[0012] Preferably, the feature update in step (4) includes: According to the feature importance matrix output by the prediction model, the correlation coefficient between abnormal items and features is updated, and the threshold of the correlation coefficient between abnormal items and features is set. When the model update history records or the normal operation data accumulate to the set sample size threshold, the features that reach the threshold of the correlation coefficient between abnormal items and features are screened, the feature set is updated, and the corresponding feature values ​​are obtained.

[0013] A rail transit key equipment fault prediction system according to the present invention comprises: The data collection and labeling module is used to collect multi-source data, perform initial data labeling according to data labels, and preliminarily calibrate the initial data state credibility coefficient; The feature extraction module is used for initial feature extraction and feature update. The initial feature extraction includes initializing the feature set that characterizes the health status of the equipment based on the experience of equipment diagnosis experts, associating abnormal items with features, preliminarily calibrating the correlation coefficient between abnormal items and features, and performing initial feature extraction. The feature update includes updating the features based on the corrected labeled data. A fault prediction model building module is used to build a fault prediction model based on a machine learning algorithm; The online prediction module is used to diagnose the main line operation data using the established fault prediction model, output the online prediction results, perform data labeling corrections based on the results, update the abnormal item point and feature correlation coefficient and the corresponding initial data state credibility coefficient; update the features based on the corrected labeled data, and further update the trained fault prediction model.

[0014] Preferably, the initial feature extraction includes setting an initial data state credibility coefficient threshold, screening data reaching the initial data state credibility coefficient threshold from the marked data, and preliminarily calculating its feature value.

[0015] Preferably, the data label correction based on the positive line prediction result includes: First, the reliability of the positive prediction is calculated and compared with the set threshold. If the reliability of the positive prediction reaches the set threshold, an early warning is pushed for known abnormal types, and the data is strongly marked. If the reliability of the positive prediction does not reach the set threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, an early warning is pushed for unknown abnormal types, and the data is weakly marked. Secondly, compare the warning push result with the actual maintenance feedback result: if the warning push result is consistent with the actual maintenance feedback result, the strengthening mark is confirmed. First, according to the feature importance matrix output by the model, the abnormal item point and feature correlation coefficient are updated; second, according to the confidence result output by the fault prediction model, the corresponding initial data state credibility coefficient is increased; if the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed and the corresponding initial data state credibility coefficient is lowered.

[0016] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the above-mentioned method for predicting failures of key rail transit equipment is implemented.

[0017] The present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method for predicting failures of key rail transit equipment is implemented.

[0018] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) After online diagnosis, the data is reinforced or weakly marked according to the reliability of the positive line prediction and the type of label recognized by the model; combined with the positive line prediction results and the actual maintenance feedback results, the data is confirmed to be reinforced or weakly marked to correct the initial data label, realizing positive line data mining and automatic data labeling update. Automatic data labeling takes into account the output of the diagnosis process and the positive line feedback, making data labeling less dependent on manual work and more reliable.

[0019] (2) In the early stage of model establishment, the initial data is used to extract the full set of features, conduct model training, and establish a diagnostic model based on the machine learning algorithm. During online prediction, the feature importance matrix output by the machine learning algorithm diagnostic model is recorded, and the correlation coefficient between abnormal items and features is updated. When the positive line data accumulates to the set sample size threshold, the features are updated to obtain the features whose correlation coefficient between abnormal items and features reaches the threshold, and the diagnostic model is updated to reduce feature redundancy and accurately identify anomalies.

[0020] (3) In order to solve the problems of weak generalization ability of prediction models, poor adaptability to scenarios, high labor cost and long time consumption of modeling, an automatic model update mechanism based on model credibility evaluation is established to realize online automatic model update. Through model credibility calculation, samples are eliminated during model training, the model is optimized and model performance is improved.

[0021] (4) In view of the single fault prediction method, a multi-source data fusion diagnosis method is proposed. By fusing and processing multiple types of data, more accurate and comprehensive diagnosis results can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the equipment failure prediction method of the present invention; Figure 2 It is a data labeling flow chart of the present invention; Figure 3 Automatically update the training flow chart for the equipment failure prediction model of the present invention; Figure 4 This is a flow chart of the reliability calculation of the equipment failure prediction model in an embodiment of the present invention; Figure 5 A data tagging correction flow chart for the present invention; Figure 6 It is a schematic diagram of the structure of the equipment failure prediction system of the present invention. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0024] like Figure 1 As shown, the method for predicting the failure of key rail transit equipment described in the present invention includes data collection and marking, feature extraction, model establishment, and online prediction. The specific steps are as follows: (1) Data collection and labeling (1.1) Collect multi-source data and pre-process the data, including filling missing values ​​and correcting outliers, to obtain the corrected equipment monitoring data, i.e., the initial data.

[0025] (1.2) Initial data labeling, such as Figure 2 shown.

[0026] (1.2.1) Define multi-source data labels. The data labels include equipment conditions, working modes, and data status. Operating conditions generally include normal line and test; the working mode is marked according to the actual operating mode of the equipment; based on the experience of equipment diagnosis experts, the data status includes healthy and abnormal, where the abnormality includes different abnormal items, and the initial data and abnormal label association relationship are set according to the operating conditions and data status. Among them, the abnormal label association relationship is defined as [(data status label, data status credibility coefficient)], such as [(0, 0.5)]: 0 means that the data status is healthy, and 0.5 means that the data status label credibility coefficient is 0.5. The data status credibility coefficient, the initial value defaults to 0.5 for normal line conditions and 1 for test conditions.

[0027] (1.2.2) Label the initial data according to the data labels.

[0028] (2) Feature extraction stage, including: (2.1) Define dynamic feature sets. Based on the experience of equipment diagnosis experts, initialize and define the feature set that represents the health status of the equipment.

[0029] (2.2) Define the correlation between abnormal items and features. Based on the experience of equipment diagnosis experts, associate abnormal items with features and calibrate the correlation coefficient between abnormal items and features, with a value range of [0,1]. The sum of the correlation coefficients of multiple features corresponding to one abnormal item is 1.

[0030] (2.3) Calculate the characteristic value based on the initial data. Based on the experience of equipment diagnosis experts, set the initial data state credibility coefficient threshold, filter the data that reaches the initial data state credibility coefficient threshold from the data marked in step (1.2.2), preliminarily calculate the characteristic value, and record the characteristic set and initial data information.

[0031] (3) Model building stage, such as Figure 3 shown.

[0032] (3.1) Use machine learning algorithms to establish fault prediction models, such as support vector machine (SVM), random forest, neural network models, etc.

[0033] (3.2) Set the maximum number of rounds for model training.

[0034] (3.3) Use the feature set obtained in step (2) as the input of the fault prediction model, and the "(label, confidence)" list as the output of the fault prediction model to carry out model training.

[0035] (3.4) Model evaluation.

[0036] Select indicators such as model accuracy, macro-average or micro-average precision, recall rate or F1 score as external control factor indicators for model evaluation, and design model credibility indicators as internal control factor indicators for model training. If the model meets the requirements of external control factor indicators, the model training is terminated, and the abnormal item point and feature correlation coefficient are updated according to the feature importance matrix output by the model, and the model information is stored. Otherwise, calculate the model credibility, update the training samples, and retrain the model.

[0037] Among them, the model credibility index refers to: the credibility coefficient of the model's diagnosis of a certain label sample, which is positively correlated with the confidence of the actual label of the sample, and negatively correlated with the standard deviation of the candidate label confidence and the number of candidate labels.

[0038] The calculation of the model credibility index includes: filtering out candidate labels according to the "(label, confidence)" list output by the model, and calculating the standard deviation of the candidate label confidence. The smaller the standard deviation of the candidate label confidence, the smaller the distinction between the candidate labels in the prediction results, the lower the certainty of the model for the prediction type, and the lower the credibility of the model, and the need to update the model training samples and update the model training.

[0039] (3.5) Training sample update and model update training.

[0040] Dynamically adjust the threshold of the initial data state credibility coefficient according to the model credibility; under the premise of ensuring the number of training samples, remove the samples whose initial data state credibility coefficient does not reach the adjusted new threshold, and then perform model update training. Determine whether the new round reaches the maximum round of model training. If so, end the model training; otherwise, repeat steps (3.3)-(3.5) to enter a new round of model training.

[0041] (4) Online prediction: data labeling and correction based on the prediction results, including: (4.1) Mainline data diagnosis. The collected equipment mainline operation data is processed through steps (1) and (2); the extracted feature values ​​are input into the fault prediction model established in step (3) to obtain the initial diagnosis results.

[0042] (4.2) Evaluation of initial diagnosis results of the positive line. The calculation of the positive line prediction credibility is consistent with the calculation of model credibility during model training.

[0043] If the credibility of the positive line prediction reaches the set threshold, the warning will push the known abnormal type and strengthen the data marking. The initial data involved in the calculation will be marked as specific abnormal points; according to the confidence of the model calculation output, the initial data state credibility coefficient will be updated.

[0044] If the positive prediction credibility does not reach the threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, the warning pushes the unknown abnormal type and weakly marks the data. The initial data involved in the calculation is marked as abnormal.

[0045] (4.3) Compare the warning push results with the actual maintenance feedback results.

[0046] If the warning push result is consistent with the actual maintenance feedback result, the enhanced mark is confirmed. First, according to the feature importance matrix output by the model, the correlation coefficient between the abnormal item point and the feature is updated (the correlation coefficient between the abnormal item point and the feature is positively correlated with the feature importance); second, according to the "(label, confidence)" list output by the fault prediction model, the corresponding initial data state credibility coefficient is updated, that is, for the confirmed abnormal label, the corresponding initial data state credibility coefficient is adjusted upward, and the value is close to 1. If the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed, that is, for the misjudged abnormal label, the corresponding initial data state credibility coefficient is adjusted downward, and the value is [0,0.5), so as to realize the automatic correction of data labeling.

[0047] (5) Further update and train the model based on the statistics of the model prediction results.

[0048] According to the statistics of the online prediction results, when the performance of the model's mainline operation decreases or the quality of samples accumulated through mainline data mining or experiments increases and the quantity is sufficient, the feature update is performed, thereby triggering the model to further update and train, thereby forming a closed-loop processing of the prediction model and realizing the online automatic update of the model. The feature update includes: updating the abnormal item point and feature correlation coefficient according to the feature importance matrix output by the model, setting the abnormal item point and feature correlation coefficient threshold, when the model update history record or the mainline operation data accumulates to the set sample volume threshold, screening the features that reach the abnormal item point and feature correlation coefficient threshold, updating the feature set, and obtaining the corresponding feature value.

[0049] Based on the same inventive concept, a rail transit key equipment fault prediction system described in the present invention is as follows: Figure 6 As shown, including: The data collection and labeling module is used to collect multi-source data, perform initial data labeling according to data labels, and preliminarily calibrate the initial data state credibility coefficient; Feature extraction module, used for initial feature extraction and feature update; initial feature extraction includes initializing and defining a feature set that characterizes the health status of the equipment based on the experience of equipment diagnosis experts; associating abnormal items with features and calibrating the correlation coefficient between abnormal items and features; setting an initial data state credibility coefficient threshold, screening data that reaches the initial data state credibility coefficient threshold from the marked data, and preliminarily calculating its feature value; feature update includes updating features based on the corrected marked data; The fault prediction model establishment module is used to establish a fault prediction model using a machine learning algorithm; the acquired feature set is used as the input of the fault prediction model to carry out model training; and the model is evaluated based on the external control factor index and the internal control factor index; if the model meets the external control factor index requirements, the model training is terminated; otherwise, the model internal control factor index is calculated, and the training sample and prediction model update training are carried out; The online prediction module is used to perform online prediction based on the established fault prediction model, perform data labeling correction based on the positive line prediction results, update the correlation coefficient between abnormal items and features and the corresponding initial data state credibility coefficient; perform feature update based on the corrected labeled data, and further update the trained fault prediction model.

[0050] The data mark correction based on the positive line prediction result includes: First, the reliability of the positive prediction is calculated and compared with the set threshold. If the reliability of the positive prediction reaches the set threshold, an early warning is pushed for known abnormal types, and the data is strongly marked. If the reliability of the positive prediction does not reach the set threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, an early warning is pushed for unknown abnormal types, and the data is weakly marked. Secondly, compare the warning push result with the actual maintenance feedback result: if the warning push result is consistent with the actual maintenance feedback result, the strengthening mark is confirmed. First, according to the feature importance matrix output by the model, the abnormal item point and feature correlation coefficient are updated; second, according to the confidence result output by the fault prediction model, the corresponding initial data state credibility coefficient is increased; if the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed and the corresponding initial data state credibility coefficient is lowered.

[0051] Based on the same inventive concept, an electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the above-mentioned method for predicting faults of key rail transit equipment is implemented.

[0052] Based on the same inventive concept, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned method for predicting failures of key rail transit equipment is implemented.

[0053] Next, two specific embodiments are used to further illustrate the content of the present invention.

[0054] Example 1: Battery Failure Prediction A battery fault prediction method according to the present invention comprises the following steps: (1) Data collection and labeling (1.1) Collect multi-source data, including the voltage, current, and capacity of the battery pack, the voltage, internal resistance, and temperature of the single battery, and the ambient temperature and humidity, as shown in Table 1. Preprocess the data, including filling in missing values ​​and correcting outliers, to obtain the corrected equipment monitoring data, i.e., the initial data.

[0055] Table 1 Battery collection data

[0056] (1.2) Initial data labeling (1.2.1) Define multi-source data labels. The data labels include equipment operating conditions, working modes, and data status. Operating conditions include positive line and test; battery working modes include equalization charge, floating charge, discharge, and static; battery data status includes healthy and abnormal. Label 0 is defined as "healthy", 1 is defined as "abnormal", and 2, 3, 4... are defined as specific abnormal items, such as 2 for "battery pack overvoltage". Set the association relationship between initial data and abnormal labels according to the data status. The multi-source data labels are shown in Table 2, and the abnormal item labels are shown in Table 3.

[0057] Table 2 Multi-source data labels

[0058] Table 3 Abnormal item labels

[0059] (1.2.2) Label the initial data according to the data labels, as shown in Table 4 below.

[0060] Table 4 Initial data label marking

[0061] (2) Feature extraction (2.1) Based on the experience of equipment diagnosis experts, the feature set characterizing the health status of single battery cells is initialized and defined, including the feature parameter name and feature parameter unit, as shown in Table 5 below.

[0062] Table 5. Feature set of single battery health status

[0063] (2.2) Define the correlation between abnormal points and features. Associate abnormal points with features, calibrate the correlation coefficient between abnormal points and features, and take the value range [0,1]. The correlation coefficient between abnormal points and features is initially set by expert experience, and is subsequently corrected according to the fault prediction model update. The sum of the correlation coefficients of multiple features corresponding to one abnormal point is 1, as shown in Table 6 below.

[0064] Table 6 Correlation between abnormal items and features

[0065] (2.3) Calculate the characteristic value based on the initial data. Based on the experience of equipment diagnosis experts, set the initial data state credibility coefficient threshold in the abnormal label association relationship, select the data that reaches the threshold from the data marked in step (1.2.2), preliminarily calculate its characteristic value, and record the characteristic set and initial data information, as shown in Table 7 below.

[0066] Table 7 Characteristic values ​​of the selected samples

[0067] (3) Model building stage, such as Figure 3 shown.

[0068] (3.1) The random forest algorithm is used to establish a battery failure prediction model.

[0069] (3.2) Set the maximum number of rounds for model training.

[0070] (3.3) Use the feature set obtained in step (2) as the input of the fault prediction model, and the "(label, confidence)" list as the output of the fault prediction model to carry out model training.

[0071] (3.4) Model evaluation The model accuracy is selected as the external control factor of model evaluation, and the model credibility index is designed as the internal control factor of model training. If the model accuracy reaches the threshold, the model training ends, and the abnormal item point and feature correlation coefficient are updated according to the feature importance matrix output by the model, and the model information is stored; otherwise, the model credibility is calculated, the training samples are updated, and the training is repeated.

[0072] The format of the model information excerpt is as follows: model_info = { "modelParam":{ # Model parameters ...... "featureImportances":[ ], # Feature importance matrix "modelRelMicro":0, # Model micro credibility coefficient "modelRelMacro":0, # Model macro credibility coefficient ...... }, "processData":{ # process data } } The calculation of the model credibility index includes: screening candidate labels according to the "(label, confidence)" list output by the model, and calculating the standard deviation of the candidate label confidence. The smaller the standard deviation of the candidate label confidence, the smaller the distinction between the candidate labels in the prediction results, the lower the certainty of the model for the prediction type, and the lower the credibility of the model, and the need to update the model training samples.

[0073] The specific calculation process is as follows Figure 4 As shown: 1) Assume that the model output covers a sample set of m abnormal labels, S, and the label sequence is L(S) = [L1,L2,...L i ,...L m ], where i is a positive integer, 1≤i≤m, L i is the i-th sample. The label of the initial label is defined as a "pseudo label". In this embodiment, m=4, then L(S) = [0,2,3,4,5].

[0074] 2) Set the model credibility assessment threshold, including the confidence threshold P th , label data threshold N th , the credibility coefficient threshold of the model is M th In this embodiment, P th =0.5,N th =3,M th =0.6.

[0075] 3) Using the label L i The verification data is used as the input of the fault prediction model, and the output result is a list of "(label, confidence)", that is, LP(L i ) = [(L1,P L1 (L i )), (L2,P L2 (L i )), ..., (L i ,P Li (L i )), ...(L m ,P Lm (L i ))],P Lm (L i ) indicates that the label is L i The data is diagnosed as label L m confidence level.

[0076] In this embodiment, if L i =4, then LP(4) = [(0,0.3),(2,0.5),(3,0.3),(4,0.8),(5,0.6)].

[0077] 4) Initialize the model pair label L i Predicted model credibility P model (L i ) = P Li (L i ).

[0078] In this embodiment, the model confidence level P of the initialization model for label 4 is model (4) = P4(4) = 0.8.

[0079] 5) Candidate label screening. Determine label L i Whether the corresponding confidence reaches the confidence threshold P th ; If reached, it will be used as a candidate label for model credibility calculation, and the number of candidate labels N is counted over (L i ); otherwise, jump to 9).

[0080] In this embodiment, the confidence level of abnormal label 4 reaches the set confidence threshold 0.5, and the number of candidate labels N with a confidence level greater than 0.5 is counted and outputted. over (4) = 3, the corresponding result set is LPN(4) = [(2,0.5),(4,0.8),(5,0.6)], and the candidate labels are 2, 4, and 5.

[0081] 6) If N over (L i )≥N th , then the standard deviation P of the corresponding confidence of the candidate label is calculated std (i); otherwise, jump to 9).

[0082] In this embodiment, since N over (4)≥N th , calculate the standard deviation P of the confidence level corresponding to the label in LPN (4) std (4)=0.07.

[0083] P std (4) is small, indicating that the distinction between the prediction result labels is small and the model has low certainty about the prediction type. Therefore, it is judged that the model's credibility for predicting label 4 is low.

[0084] 7) Correct the model's confidence in the prediction of label Li

[0085] in, For the model pair label L i The micro-correction coefficient of the predicted model credibility is Positively correlated with N over (Li ), P std (L i ) is negatively correlated, and its value range is [0,0.1].

[0086] The corrected model credibility of the prediction model for label 4 in this embodiment

[0087] .

[0088] 8) If i+1≤m, then according to steps 4)-8), calculate the model for label L i+1 The model credibility of the prediction; otherwise, end the model credibility calculation for each prediction label in this round of model, and record the list Record the uncertainty list of each label data prediction result: “Positive label” confidence list P Li (S) = [P L1 (L1),P L2 (L2),...,P Li (L i ),...,P Lm (L m )] Confidence standard deviation list P std (S) = [P std (L1),P std (L2),...,P std (Li),...,P std (m)] The result is an uncertain number of lists N over (S) = [N over (L1),N over (L2),...,N over (L i ),...,N over (L m )].

[0089] In this embodiment, according to steps 4)-8), the model credibility of the model prediction for label L5 is calculated; the credibility calculation of each predicted label by the model in this round is ended, and the list is recorded.

[0090] Record the uncertainty list of each label data prediction result: “Positive label” confidence list P Li (S) = [0.8, 0.7, 0.7,0.8, 0.73] Confidence standard deviation list P std (S) = [0.02, 0.11, 0.01, 0.07, 0.01] The result is an uncertain number of lists N over (S) = [2,3,2,3,3] .

[0091] 9) Calculate the model credibility coefficient of the comprehensive prediction of the model for the sample set S Micro coefficient

[0092] Macro coefficient

[0093] P Li_mean (S)=0.746, P std_mean (S)=0.044, N over_mode (S)=3.

[0094] is the macro correction coefficient for the model's credibility predicted by the model for the sample set S, and the mean P of the "positive label" confidence list Li_mean (S) positive correlation, with the result of uncertain number list mode N over_mode (S), confidence standard deviation list mean P std_mean (S) is negatively correlated, and its value is generally distributed in [0,0.1].

[0095] In this embodiment, .

[0096] (3.5) Training sample update and model update training The threshold of the initial data state credibility coefficient is dynamically adjusted according to the model credibility; under the premise of ensuring the number of training samples, samples whose initial data state credibility coefficient does not reach the adjusted new threshold are eliminated to determine whether the new round reaches the maximum round of model training. If so, the model training is terminated; otherwise, steps (3.3)-(3.5) are repeated to enter a new round of model training.

[0097] (4) Use the established fault prediction model to diagnose the mainline operation data, output the online prediction results, and perform data labeling correction based on the results.

[0098] (4.1) Mainline data diagnosis. The collected equipment mainline operation data is processed through steps (1) and (2); the extracted feature values ​​are input into the fault prediction model established in step (3) to obtain the initial diagnosis results.

[0099] (4.2) Evaluation of initial diagnosis results of the positive line. Calculate the positive line prediction credibility. The calculation of the positive line prediction credibility does not distinguish between macro and micro coefficients. It is similar to the calculation method of the model credibility of the model prediction of the label Li during the model training period, but uses the confidence corresponding to the predicted output label instead of the corresponding confidence of the "positive label" during model training.

[0100] If the online prediction credibility reaches the threshold, the known abnormal type will be pushed out as an early warning, and the data will be marked. The initial data involved in the calculation will be marked as specific abnormal items; according to the confidence of the model calculation output, the initial data state credibility coefficient will be updated.

[0101] If the online prediction credibility does not reach the threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, an unknown abnormal type is pushed as an early warning, and the data is weakly marked. The initial data involved in the calculation is marked as abnormal.

[0102] (4.3) Compare the warning push results with the actual maintenance feedback results.

[0103] If the warning push result is consistent with the actual maintenance feedback result, the enhanced mark is confirmed. First, according to the feature importance matrix output by the model, the abnormal item point and feature correlation coefficient are updated according to the positive correlation between the abnormal item point and the feature importance; second, according to the "(label, confidence)" list output by the fault prediction model, the corresponding initial data state credibility coefficient is updated, that is, for the confirmed abnormal label, the initial data state credibility coefficient is adjusted upward, and the value is close to 1. If the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed, that is, for the misjudged abnormal label, the initial data state credibility coefficient is adjusted downward, and the value is [0,0.5).

[0104] Preferably, the data tagging correction process is as follows Figure 5 As shown: If the diagnosis result of a battery fault prediction is confirmed as abnormal point 4 (low battery pack SOC), the data mark correction method of each level is as follows: (a) Update of the correlation coefficient between abnormal items and features. According to the feature importance matrix output by the fault prediction model, the correlation coefficient between abnormal items and features is updated. The higher the importance, the higher the correlation coefficient between abnormal items and features is adjusted. The results before and after the adjustment are shown in Table 8.

[0105] For the abnormal label L i , its associated characteristic parameter serial number list F(L i )=[F1,F2,...F j ,...Fn], where j is a positive integer, 1≤j≤n. The corresponding abnormal item point and feature correlation coefficient vector before updating the mark is R Li_F = [RLi_F1 ,R Li_F2 ,...,R Li_Fj ,...R Li_Fn ] The model output feature importance matrix for this model training round is P' Li_F = [P' Li_F1 ,P' Li_F2 ,...,P' Li_Fj ,...P' Li_Fn ] Then, the updated abnormal item point and feature correlation coefficient vector is R' Li_F = [R Li_F1 ×P' Li_F1 , ... ,R Li_Fj ×P' Li_Fj , ..., R Li_Fn ×P' Li_Fn ] / (R Li_F ·P' Li_F ) In this embodiment, for the abnormal point Fault4 (battery pack SOC is low), its associated characteristic parameter serial number list is F(4)=[3, 4, 5, 6], The corresponding abnormal item point and feature correlation coefficient vector before updating the mark is R 4_F = [R 4_3 , R 4_4 , ..., R 4_5 , R 4_6 ] = [0.2, 0.3, 0.1, 0.4] The model output feature importance matrix for this model training round is P' 4_F = [P' 4_3 , P' 4_4 , ..., P' 4_5 , P' 4_6 ] = [0.2, 0.2, 0.1, 0.5], The updated abnormal item point and feature correlation coefficient vector is: R' 4_F = [R 4_3 ×P' 4_3 , R 4_4 ×P' 4_4 , R 4_5 ×P' 4_5 , R 4_6 ×P' 4_6 ] / (R4_F · P' 4_F ) =[0.04,0.06,0.01,0.20] / 0.31 = [0.129, 0.194, 0.032, 0.645] Table 8 Update table of abnormal item points and feature correlation coefficients

[0106] (b) Update the association between the initial data and the abnormal label. For the abnormal label L i ,According to the “(label, confidence)” list output by the fault prediction, the corresponding initial data and abnormal label association relationship are updated, as shown in Table 9.

[0107] In this embodiment, for this fault prediction, the initial data sequence number used for feature extraction is: "OrigData6,OrigData16,OrigData26,...,OrigData676" For the initial data with serial number OrigData6, the relationship between the initial data and the abnormal label before the data label correction is: DF(6) = [(0,0.5)] The corresponding fault prediction confidence output is: LP(6) = [(0,0.3),(2,0.8),(3,0.5),(4,0.3),(5,0.6)] Then the relationship between the corrected initial data and the abnormal label is: DF(6) = [(0,0.3),(2,0.8)] Table 9 Initial data and abnormal label association update table .

[0108] Example 2: Fault prediction of rail transit vehicle door system

[0109] A rail transit vehicle door system fault prediction method according to the present invention comprises the following steps: (1) Data collection and labeling (1.1) Collect multi-source data, including the door motor’s rotation angle, speed, and current, and the door controller’s IO signal data, as shown in Table 10. Preprocess the data, including missing value filling and outlier correction, to obtain the corrected equipment monitoring data, i.e., the initial data.

[0110] Table 10 Door collection data

[0111] (1.2) Initial data labeling (1.2.1) Define multi-source data labels. The data labels include equipment operating conditions, working modes, and data status. Operating conditions include normal line and test. Working modes include door opening and door closing. Door data status includes healthy and abnormal. Label 0 is defined as "healthy", label 1 is defined as "abnormal", and labels 2, 3, 4... are defined as specific abnormal items. Set the association relationship between initial data and abnormal labels according to the data status. Multi-source data labels are shown in Table 11, and abnormal item labels are shown in Table 12.

[0112] Table 11 Multi-source data labels

[0113] Table 12 Abnormal item labels

[0114] (1.2.2) Label the initial data according to the data labels, as shown in Table 13.

[0115] Table 13 Door initial data mark

[0116] (2) Feature extraction (2.1) Based on the experience of door diagnosis experts, the feature set that characterizes the health status of the door is initialized and defined, including the feature parameter name and feature parameter unit, as shown in Table 14.

[0117] Table 14 Feature set of door health status

[0118] (2.2) Define the correlation between abnormal points and features. Associate abnormal points with features, calibrate the correlation coefficient between abnormal points and features, and take the value range [0,1]. The correlation coefficient between abnormal points and features is initially set by expert experience, and is subsequently corrected according to the fault prediction model update. The sum of the correlation coefficients of multiple features corresponding to one abnormal point is 1, as shown in Table 15 below.

[0119] Table 15 Correlation between abnormal items and features

[0120] (2.3) Calculate the characteristic value based on the initial data. Based on the experience of equipment diagnosis experts, set the initial data status credibility coefficient threshold, filter the data that reaches the threshold from the data marked in step (1.2.2), preliminarily calculate the characteristic value, and record the characteristic set and initial data information, as shown in Table 16.

[0121] Table 16 Characteristic values ​​of the selected samples

[0122] (3) Model building stage (3.1) The SVM algorithm is used to establish a fault prediction model.

[0123] (3.2) Set the maximum number of rounds for model training.

[0124] (3.3) Use the feature set obtained in step (2) as the input of the fault prediction model, and the "(label, confidence)" list as the output of the fault prediction model to carry out model training.

[0125] (3.4) Model evaluation The model accuracy index is selected as the external control factor of model evaluation, and the model credibility index is designed as the internal control factor of model training. If the model accuracy reaches the threshold, the model training ends, and the abnormal item point and feature correlation coefficient are updated according to the feature importance matrix output by the model, and the model information is stored; otherwise, the model credibility is calculated, the training samples are updated, and retraining is performed.

[0126] The calculation of the model credibility index includes: screening out candidate labels according to the "(label, confidence)" list output by the model, and calculating the standard deviation of the candidate label confidence; the smaller the standard deviation of the candidate label confidence, the smaller the distinction between the candidate labels in the prediction results, the lower the certainty of the model for the prediction type, and the lower the credibility of the model, and the need to update the model training samples.

[0127] (3.5) Model update training The threshold of the initial data state credibility coefficient is dynamically adjusted according to the model credibility; under the premise of ensuring the number of training samples, samples whose initial data state credibility coefficient does not reach the adjusted new threshold are eliminated to determine whether the new round reaches the maximum round of model training. If so, the model training is terminated; otherwise, steps (3.3)-(3.5) are repeated to enter a new round of model training.

[0128] (4) Online prediction, data labeling and correction based on prediction results (4.1) Mainline data diagnosis. The collected equipment mainline operation data is processed through steps (1) and (2); the extracted feature values ​​are input into the fault prediction model established in step (3) to obtain the initial diagnosis results.

[0129] (4.2) Evaluation of initial diagnosis results of the positive line. Calculate the positive line prediction credibility. The calculation of the positive line prediction credibility does not distinguish between macro and micro coefficients. It is similar to the calculation method of the model prediction credibility of the label Li during the model training period, but the confidence corresponding to the predicted output label is used instead of the corresponding confidence of the "positive label" during model training.

[0130] If the online prediction credibility reaches the threshold, the known abnormal type will be pushed out as an early warning, and the data will be marked. The initial data involved in the calculation will be marked as specific abnormal items; according to the confidence of the model calculation output, the initial data state credibility coefficient will be updated.

[0131] If the online prediction credibility does not reach the threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, an unknown abnormal type is pushed as an early warning, and the data is weakly marked. The initial data involved in the calculation is marked as abnormal.

[0132] (4.3) Compare the warning push results with the actual maintenance feedback results.

[0133] If the warning push result is consistent with the actual maintenance feedback result, the enhanced mark is confirmed. First, according to the feature importance matrix output by the model, the correlation coefficient between the abnormal item point and the feature is updated. The correlation coefficient between the abnormal item point and the feature is positively correlated with the feature importance; second, according to the "(label, confidence)" list output by the fault prediction model, the corresponding initial data state credibility coefficient is updated, that is, for the confirmed abnormal label, the initial data state credibility coefficient is adjusted upward, and the value is close to 1. If the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed, that is, for the misjudged abnormal label, the initial data state credibility coefficient is adjusted downward, and the value is [0,0.5).

[0134] The data tag correction process of this embodiment includes: If for a door fault prediction, the diagnosis result is abnormal label 2 (too large resistance throughout the journey), the data correction methods at each level are as follows: (a) Update the correlation coefficient between abnormal items and features. According to the feature importance matrix output by the algorithm model, update the correlation coefficient between abnormal items and features. The higher the importance, the higher the correlation coefficient between abnormal items and features.

[0135] The list of characteristic parameter numbers associated with abnormal point Fault2 is F(2)=[8, 9, 10, 11] The corresponding abnormal item point and feature correlation coefficient vector before updating the mark is R 2_F = [R 2_8 , R 2_9 , ..., R 2_10 , R 2_11 ] = [0.3, 0.3, 0.2, 0.2] The algorithm output feature importance matrix for this model training round P' 2_F = [P' 2_8 , P' 2_9 , ..., P' 2_10 , P'2_11 ] = [0.5, 0.45, 0.1, 0.05], Updated abnormal item point and feature correlation coefficient vector: R' 3_F = [R 3_3 ×P' 3_3 , R 3_4 ×P' 3_4 , R 3_5 ×P' 3_5 , R 3_6 ×P' 3_6 ] / (R 3_F · P' 3_F ) =[0.15,0.135,0.02,0.01] / 0.315 = [0.476, 0.429, 0.063, 0.032] The updated relationship between abnormal item points and features is shown in Table 17.

[0136] Table 17 Update table of abnormal item points and feature correlation coefficients

[0137] (b) Update the association between the initial data and the abnormal label. For the abnormal label L i ,According to the "(label,confidence)" list output by the fault prediction, the corresponding initial data and abnormal label association relationship are updated.

[0138] For example, the initial data sequence number used for the door fault prediction feature extraction is: "OrigData3,OrigData13,OrigData23,...,OrigData503", For the initial data with serial number OrigData3, the correlation relationship between the initial data and the abnormal label before the data label correction is: DF(3) = [(0,0.5)] The corresponding fault prediction confidence output is: LP(3) = [(0,0.3),(2,0.8),(3,0.5),(4,0.3),(5,0.6)] Then, the relationship between the corrected initial data and the abnormal label is: DF(3) = [(0,0.3),(2,0.8)] The updated initial data labels are shown in Table 18.

[0139] Table 18 Initial data and abnormal label association update table .

Claims

1. A method for predicting failures of key rail transit equipment, characterized in that: The following steps are involved: (1) Data collection and labeling, preliminary calibration of the initial data state credibility coefficient; (2) Based on the experience of equipment diagnosis experts, the feature set that characterizes the health status of the equipment is initialized and defined, the abnormal items are associated with the features, the correlation coefficients between the abnormal items and the features are preliminarily calibrated, and the initial feature extraction is performed; (3) Use machine learning algorithms to establish a fault prediction model; (4) Use the fault prediction model to diagnose the main line operation data and output the online prediction results. Based on the results, the data labeling is corrected, and the correlation coefficient between the abnormal item and the feature and the corresponding initial data state credibility coefficient are updated; the features are updated based on the corrected labeled data, and then the trained fault prediction model is further updated.

2. The method for predicting failures of key rail transit equipment according to claim 1, characterized in that: The data labeling in step (1) includes: firstly defining multi-source data labels, wherein the data labels include equipment operating conditions, working modes, and data states; and then performing initial data labeling according to the data labels.

3. The method for predicting failures of key rail transit equipment according to claim 1, characterized in that: The initial feature extraction in step (2) includes: setting an initial data state credibility coefficient threshold, screening data that reaches the initial data state credibility coefficient threshold from the marked data, and preliminarily calculating its feature value.

4. The method for predicting failures of key rail transit equipment according to claim 1, characterized in that: Step (3) includes: using the feature set obtained in step (2) as the input of the fault prediction model to carry out model training; and evaluating the model based on the external control factor index and the internal control factor index; if the model meets the external control factor index requirements, the model training is terminated; otherwise, the model internal control factor index is calculated, and the training sample is updated and the model update training is carried out.

5. The method for predicting failures of key rail transit equipment according to claim 4, characterized in that: The external control factor indicators include model accuracy, macro-average or micro-average precision, recall rate, and F1 score; the internal control factor indicator is model credibility.

6. The method for predicting failures of key rail transit equipment according to claim 4, characterized in that: The training sample update and model update training include: The threshold of the initial data state credibility coefficient is dynamically adjusted according to the model credibility. On the premise of ensuring the number of training samples, samples whose initial data state credibility coefficient does not reach the adjusted threshold are eliminated, and then the model update training is carried out.

7. The method for predicting failures of key rail transit equipment according to claim 1, characterized in that: The data mark correction in step (4) includes: First, the reliability of the positive prediction is calculated and compared with the set threshold. If the reliability of the positive prediction reaches the set threshold, an early warning is pushed for known abnormal types, and the data is strongly marked. If the reliability of the positive prediction does not reach the set threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, an early warning is pushed for unknown abnormal types, and the data is weakly marked. Secondly, compare the warning push result with the actual maintenance feedback result: if the warning push result is consistent with the actual maintenance feedback result, the strengthening mark is confirmed. First, according to the feature importance matrix output by the fault prediction model, the abnormal item point and feature correlation coefficient are updated; second, according to the confidence result output by the fault prediction model, the corresponding initial data state credibility coefficient is increased; if the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed and the corresponding initial data state credibility coefficient is lowered.

8. The method for predicting failures of key rail transit equipment according to claim 1, characterized in that: The feature update in step (4) includes: According to the feature importance matrix output by the fault prediction model, the abnormal item point and feature correlation coefficient are updated, and the threshold of the abnormal item point and feature correlation coefficient is set. When the model update history record or the amount of data accumulated during normal operation reaches the set sample size threshold, the features that reach the threshold of the abnormal item point and feature correlation coefficient are screened, the feature set is updated, and the corresponding feature value is obtained.

9. A rail transit key equipment fault prediction system, characterized in that: include: The data collection and labeling module is used to collect multi-source data, perform initial data labeling according to data labels, and preliminarily calibrate the initial data state credibility coefficient; Feature extraction module, used for initial feature extraction and feature update; Initial feature extraction includes initializing and defining a feature set that characterizes the health status of the equipment based on the experience of equipment diagnosis experts, associating abnormal items with features, preliminarily calibrating the correlation coefficients between abnormal items and features, and performing initial feature extraction; feature update includes updating features based on the corrected labeled data; A fault prediction model building module is used to build a fault prediction model based on a machine learning algorithm; The online prediction module is used to perform online prediction based on the established fault prediction model, perform data labeling correction based on the positive line prediction results, update the correlation coefficient between abnormal items and features and the corresponding initial data state credibility coefficient; perform feature update based on the corrected labeled data, and further update the trained fault prediction model.

10. The rail transit key equipment fault prediction system according to claim 9, characterized in that: The initial feature extraction includes setting an initial data state credibility coefficient threshold, screening data reaching the initial data state credibility coefficient threshold from the marked data, and preliminarily calculating its feature value.

11. The rail transit key equipment fault prediction system according to claim 9, characterized in that: The data mark correction based on the positive line prediction result includes: First, the reliability of the positive prediction is calculated and compared with the set threshold. If the reliability of the positive prediction reaches the set threshold, an early warning is pushed for known abnormal types, and the data is strongly marked. If the reliability of the positive prediction does not reach the set threshold, that is, it is not successfully identified as a known abnormal type or a healthy type, an early warning is pushed for unknown abnormal types, and the data is weakly marked. Secondly, compare the warning push result with the actual maintenance feedback result: if the warning push result is consistent with the actual maintenance feedback result, the strengthening mark is confirmed. First, according to the feature importance matrix output by the model, the abnormal item point and feature correlation coefficient are updated; second, according to the confidence result output by the fault prediction model, the corresponding initial data state credibility coefficient is increased; if the warning push result is inconsistent with the actual maintenance feedback result, the weak mark is confirmed and the corresponding initial data state credibility coefficient is lowered.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the method for predicting failures of key rail transit equipment according to any one of claims 1 to 8 is implemented.

13. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for predicting failures of key rail transit equipment according to any one of claims 1 to 8 is implemented.

Images