Equipment fault detection method and system based on time sequence comparative learning

Through the timing comparison learning method, the equipment fault detection technology combined with CNN model and support vector machine is used to solve the problem of low accuracy in equipment fault classification diagnosis, and efficient fault detection in complex environments is achieved, and the adaptability of the model is improved. and accuracy.

CN119046845APending Publication Date: 2024-11-29BEIJING NOSI SPACETIME TECH CO LTD

Patent Information

Application Number
CN202411248730.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-06
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

In the classification diagnosis of equipment failures, traditional statistical analysis and model-based methods are difficult to capture complex features, and the time series data lacks labeling, resulting in low accuracy of classification methods and difficult to respond flexibly in complex and changing production environments.

Method used

Using a time-series comparison learning method, feature representations are extracted through the CNN model, and equipment failure detection is carried out in combination with the support vector machine. Positive and negative sample pairs are generated using self-supervised learning and data augmentation to train the model to improve generalization ability and detection performance.

Benefits of technology

It realizes the accuracy and adaptability of equipment fault detection in the absence of label data, can effectively diagnose faults in complex and changeable environments, and improves the generalization ability and detection performance of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046845A_ABST
    Figure CN119046845A_ABST
Patent Text Reader

Abstract

The invention discloses an equipment fault detection method and system based on time sequence comparative learning, and belongs to the technical field of equipment fault detection. According to the method, the problem of low accuracy of equipment fault detection by a traditional time series data classification method due to lack of industrial time series data with labels at present is solved. According to the method, positive and negative sample pairs are generated based on data collected by a sensor to perform model training, a large amount of labeled data does not need to be depended on, more diversified features can be learned by the model through training on non-labeled data, and classifier training of downstream tasks is performed by using a small amount of labeled data, so that the training efficiency is improved. Therefore, equipment fault detection and diagnosis are realized, and the generalization ability of the model is improved. And the selection of the positive and negative sample pairs comprises a normal working state and different abnormal states, so that the detection performance of the model on different fault forms of the equipment is improved, and the fault detection accuracy is ensured. The method can be applied to equipment fault detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment fault detection, and in particular relates to an equipment fault detection method and system based on time series contrast learning. Background Art

[0002] The existing technologies for equipment fault classification and diagnosis in industry are often traditional statistical analysis or model-based methods. However, traditional statistical analysis methods mainly rely on a large number of simple statistical laws, which are difficult to capture complex potential features, and are sensitive to noise and outliers. They are easily affected by abnormal data, resulting in inaccurate equipment fault classification results; model-based methods analyze based on the physical characteristics and operating principles of the system. Although they do not rely on a large amount of data, for systems with complex environments or dynamic changes, the model may be difficult to adapt, and the model needs to be frequently adjusted and updated, making it difficult to flexibly respond to complex and changing production environments.

[0003] Sensors collect data continuously over a period of time. Each data point has a clear timestamp, which can provide high-frequency, fine-grained equipment status information. The current data point usually has a certain correlation with the previous data point. Analyzing the continuously collected data can capture the real-time changes in the system status. Therefore, the time series data analysis method can be used to analyze and process the sensor data, discover potential abnormal information, and flexibly apply it in dynamically changing scenarios to achieve equipment fault diagnosis. However, the labeling cost of time series data is high, and the data often lacks labels, especially industrial time series data, which lacks labels for different fault types. This makes the traditional time series data classification method difficult to implement and has low accuracy, making it difficult to achieve accurate and efficient equipment fault diagnosis. Summary of the invention

[0004] The purpose of the present invention is to solve the problem that the accuracy of equipment fault detection by traditional time series data classification methods is low due to the lack of labeled industrial time series data, and to propose an equipment fault detection method and system based on time series comparative learning.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] According to one aspect of the present invention, a device fault detection method based on time series contrast learning comprises the following steps:

[0007] Step 1: Use sensors to collect time series data when the equipment to be detected is running, and pre-process the data collected by the sensors to obtain pre-processed data;

[0008] Step 2: Use the preprocessed data as the input of the CNN model, and output the feature representation of the preprocessed data through the CNN model;

[0009] The CNN model is a CNN model trained based on a contrastive learning method;

[0010] Step 3: Use the feature representation output in step 2 as the input of the support vector machine, and output the equipment fault detection result through the support vector machine.

[0011] According to another aspect of the present invention, a device fault detection system based on time series contrast learning includes a time series data acquisition module to be detected, a first neural network module and a second neural network module, wherein:

[0012] The detection time series data acquisition module is used to collect the time series data of the device to be detected when it is running, pre-process the acquired time series data, and send the pre-processed data to the first neural network module;

[0013] A CNN model is deployed in the first neural network module, and the CNN model is used to extract feature representations of the preprocessed data and send the extracted feature representations to the second neural network module;

[0014] The training process of the CNN model is:

[0015] Step 1: Collect a time series data set of the same type of equipment as the equipment to be detected during historical operation, and the collected time series data set includes labeled time series data and unlabeled time series data; then perform data enhancement on the collected time series data;

[0016] Step 2: construct positive sample pairs and negative sample pairs based on the enhanced time series dataset and self-supervised learning method;

[0017] Step 3: The CNN model is trained based on the constructed positive sample pairs and negative sample pairs. The loss function used in training is And use the gradient descent method for back propagation and CNN model parameter update;

[0018] Loss Function for:

[0019]

[0020] in, is the feature representation of the i-th anchor sample pair extracted by the CNN model, is the feature representation of the i-th positive sample pair extracted by the CNN model, is the feature representation of the i-th negative sample extracted by the CNN model, is the square of the L2 norm, α is a hyperparameter, [·] + ReLU operation, the number of anchor sample pairs, positive sample pairs and negative sample pairs are all N;

[0021] Until the loss function Stop training when convergence and obtain the trained CNN model;

[0022] A support vector machine is deployed in the second neural network module, and the support vector machine is used to output the device fault detection result according to the feature representation;

[0023] The training process of support vector machine is:

[0024] Step (1), passing the labeled time series data collected in step 1 through the trained CNN model, and outputting the feature representation of each labeled time series data through the CNN model;

[0025] Step (2): The feature representations of the M labeled time series data are recorded as h1, h2, ..., h M , using h1,h2,…,h M Composition feature matrix m is the dimension of each feature representation;

[0026] The classification labels of M labeled time series data are recorded as y1, y2, ..., y M , using y1,y2,…,y M The label vector y = [y1, y2, ..., y N ];

[0027] Step (3), use the feature matrix H and the label vector y to train the support vector machine, calculate the classification loss according to the output h(H) of the support vector machine after the feature matrix H and the label vector y, and stop training when the classification loss converges to obtain a trained support vector machine.

[0028] The beneficial effects of the present invention are:

[0029] Contrastive learning generates positive and negative sample pairs based on the data collected by sensors for model training. It does not need to rely on a large amount of labeled data. By training on unlabeled data, the model can learn more diverse features, and then use a small amount of labeled data to train the classifier of downstream tasks, thereby realizing equipment fault detection and diagnosis and improving the generalization ability of the model. Moreover, the selection of positive and negative sample pairs includes normal working conditions and different abnormal conditions, thereby improving the model's detection performance for different fault forms of equipment, ensuring the accuracy of fault detection, and being able to adapt to complex and changing actual production environments, and having good performance for scenes with complex or dynamically changing environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of a device fault detection method based on time series contrast learning of the present invention. DETAILED DESCRIPTION

[0031] Specific implementation method 1: Combination Figure 1 The present embodiment describes a device fault detection method based on time series contrast learning, and the method specifically comprises the following steps:

[0032] Step 1: Use sensors to collect time series data when the equipment to be detected is running, and pre-process the data collected by the sensors to obtain pre-processed data;

[0033] Step 2: Use the preprocessed data as the input of the CNN model, and output the feature representation of the preprocessed data through the CNN model;

[0034] The CNN model is a CNN model trained based on the contrastive learning method; the principle of contrastive learning is to bring positive sample pairs (similar samples) closer in the embedding space and push negative sample pairs (dissimilar samples) farther away;

[0035] Step 3: Use the feature representation output in step 2 as the input of a support vector machine (SVM), and output the equipment fault detection result through the support vector machine.

[0036] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that the sensor is a temperature sensor, a current sensor, a vibration sensor, a pressure sensor or a radiation sensor.

[0037] The other steps and parameters are the same as those in the first embodiment.

[0038] The collected data is first temporarily stored on the collection device to prevent data loss when the transmission is interrupted, and the data is transmitted to the central server or cloud storage for long-term storage and backup. The corresponding sensor can be selected for data collection according to the actual device to be detected.

[0039] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the sensor is connected to a PLC (programmable logic controller) and a DCS (distributed control system), and the PLC and DCS are used to control the sensor to collect and pre-process data.

[0040] The other steps and parameters are the same as those in the first or second embodiment.

[0041] Specific implementation method 4: This implementation method is different from any one of specific implementation methods 1 to 3 in that the data collected by the sensor is preprocessed, and the preprocessing method includes data filtering and data cleaning.

[0042] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0043] By filtering and cleaning the collected time series data, you can remove erroneous, duplicate or invalid data, improve the accuracy and reliability of the data, and fill in missing data to ensure data integrity, thereby improving the accuracy of the analysis.

[0044] Data filtering and cleaning operations can be performed based on rule verification and replacement methods. According to business needs and data characteristics, a series of data verification rules can be defined. Data verification rules include but are not limited to the following methods:

[0045] Range calibration: Temperature data should be between T1-T2℃.

[0046] Logic check: The pressure value should not fluctuate suddenly and drastically.

[0047] Time interval verification: The time interval between data points should remain consistent.

[0048] Invalid data can be identified and processed through predefined rules. Data that does not conform to the rules can be deleted or processed using methods such as mean replacement and linear interpolation to ensure data quality and consistency. Noise can also be removed through filtering and other methods to improve data quality.

[0049] Specific implementation method 5: This implementation method is different from the specific implementation methods 1 to 4 in that the training process of the CNN model is:

[0050] Step 1: Collect a time series data set of the same type of equipment as the equipment to be detected during historical operation, and the collected time series data set includes labeled time series data and unlabeled time series data; then perform data enhancement on the collected time series data;

[0051] Step 2: construct positive sample pairs and negative sample pairs based on the enhanced time series dataset and self-supervised learning method;

[0052] Step 3: The CNN model is trained based on the constructed positive sample pairs and negative sample pairs. The loss function used in training is And use the gradient descent method for back propagation and CNN model parameter update;

[0053] Until the loss function Stop training when convergence and obtain the trained CNN model.

[0054] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0055] The specific working process of the CNN model is introduced as follows:

[0056] Input layer: Anchor, Positive, and Negative sample input.

[0057] Input three types of sample pairs: anchor sample pairs (anchor), positive sample pairs (positive), and negative sample pairs (negative).

[0058] Convolutional layer: Design multiple layers of one-dimensional convolutional layers to extract local features of time series data.

[0059] Each layer of convolution operation can be expressed as:

[0060] Conv1D(x)=σ(W*x+b)

[0061] Among them, W is the convolution kernel, * represents the convolution operation, b is the bias, and σ is the activation function (such as ReLU).

[0062] Pooling Layer: Add a pooling layer, such as Max Pooling, to reduce the dimension of the feature map and extract the main features.

[0063] The pooling operation can be expressed as:

[0064] MaxPooling1D(x)=max(x)

[0065] Fully Connected Layer: Flattens the feature map and connects it to the fully connected layer to output a fixed-dimensional feature representation.

[0066] The operation of the fully connected layer can be expressed as:

[0067] Dense(x)=σ(Wx+b)

[0068] Specific implementation example 6: This implementation example is different from any one of specific implementation examples 1 to 5 in that the enhancement method used for data enhancement of the collected time series data is time masking, adding noise and time scaling.

[0069] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0070] Data augmentation can increase the robustness of the model, and the model can maintain stable performance when facing different noises, deformations or interferences.

[0071] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that the loss function for:

[0072]

[0073] in, is the feature representation of the i-th anchor sample pair extracted by the CNN model (that is, the concatenation of the feature representations of the two samples in the sample pair extracted by the CNN model), is the feature representation of the i-th positive sample pair extracted by the CNN model, is the feature representation of the i-th negative sample extracted by the CNN model, is the square of the L2 norm, α is a hyperparameter, [·] + ReLU operation represents the ReLU operation, and the number of anchor sample pairs, positive sample pairs, and negative sample pairs are all N.

[0074] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0075] The square of the L2 norm is used to measure the Euclidean distance between feature vectors; the hyperparameter is used to control the distance interval between positive and negative sample pairs. It is expected that the distance between the anchor sample pair and the negative sample pair is at least α larger than the distance between the anchor sample pair and the positive sample pair; the ReLU operation takes the maximum value between the value in the brackets and 0 to ensure that the loss value is non-negative.

[0076] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that the training process of the support vector machine is as follows:

[0077] Step (1), passing the labeled time series data collected in step 1 through the trained CNN model, and outputting the feature representation of each labeled time series data (that is, each time segment data) through the CNN model;

[0078] Step (2): record the feature representations of M labeled time series data (that is, M time segment data) as h1, h2, ..., h M , using h1,h2,…,h M Composition feature matrix m is the dimension of each feature representation;

[0079] The classification labels of M labeled time series data are recorded as y1, y2, ..., y M , the label is the normal state of the device and the actual fault type of the device, using y1, y2, …, y M The label vector y = [y1, y2, ..., y N ];

[0080] Step (3), use the feature matrix H and the label vector y to train the support vector machine, calculate the classification loss according to the output g(H) of the support vector machine after the feature matrix H and the label vector y, and stop training when the classification loss converges to obtain a trained support vector machine.

[0081] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0082] Specific implementation method nine: This implementation method is different from any one of specific implementation methods one to eight in that the method also includes step four, when detecting the time series data of K consecutive time segments of the operation of the equipment to be detected, if the detection results are all fault states, a fault warning is triggered to notify the staff to conduct equipment inspection.

[0083] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0084] Feedback the fault diagnosis and maintenance results to the system to update the model and rule base. By learning new fault data, the classification model and diagnosis method are continuously improved to improve the accuracy and reliability of fault detection and diagnosis.

[0085] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that the value of K is 5.

[0086] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0087] By setting a suitable K value, both the timeliness and accuracy of fault detection can be ensured.

[0088] Specific implementation eleven: This implementation is a device fault detection system based on time series contrast learning, the system includes a time series data acquisition module to be detected, a first neural network module and a second neural network module, wherein:

[0089] The detection time series data acquisition module is used to collect the time series data of the device to be detected when it is running, pre-process the acquired time series data, and send the pre-processed data to the first neural network module;

[0090] A CNN model is deployed in the first neural network module, and the CNN model is used to extract feature representations of the preprocessed data and send the extracted feature representations to the second neural network module;

[0091] The training process of the CNN model is:

[0092] Step 1: Collect a time series data set of the same type of equipment as the equipment to be detected during historical operation, and the collected time series data set includes labeled time series data and unlabeled time series data; then perform data enhancement on the collected time series data;

[0093] Step 2: construct positive sample pairs and negative sample pairs based on the enhanced time series dataset and self-supervised learning method;

[0094] Step 3: The CNN model is trained based on the constructed positive sample pairs and negative sample pairs. The loss function used in training is And use the gradient descent method for back propagation and CNN model parameter update;

[0095] Loss Function for:

[0096]

[0097] in, is the feature representation of the i-th anchor sample pair extracted by the CNN model, is the feature representation of the i-th positive sample pair extracted by the CNN model, is the feature representation of the i-th negative sample extracted by the CNN model, is the square of the L2 norm, α is a hyperparameter, [·] + ReLU operation, the number of anchor sample pairs, positive sample pairs and negative sample pairs are all N;

[0098] Until the loss function Stop training when convergence and obtain the trained CNN model;

[0099] A support vector machine is deployed in the second neural network module, and the support vector machine is used to output the device fault detection result according to the feature representation;

[0100] The training process of support vector machine is:

[0101] Step (1), passing the labeled time series data collected in step 1 through the trained CNN model, and outputting the feature representation of each labeled time series data through the CNN model;

[0102] Step (2): The feature representations of the M labeled time series data are recorded as h1, h2, ..., h M , using h1,h2,…,h M Composition feature matrix m is the dimension of each feature representation;

[0103] The classification labels of M labeled time series data are recorded as y1, y2, ..., y M , using y1,y2,…,y M The label vector y = [y1, y2, ..., y N ];

[0104] Step (3), use the feature matrix H and the label vector y to train the support vector machine, calculate the classification loss according to the output g(H) of the support vector machine after the feature matrix H and the label vector y, and stop training when the classification loss converges to obtain a trained support vector machine.

[0105] Example

[0106] This embodiment proposes a motor bearing fault detection method based on time series contrast learning, which specifically includes the following steps:

[0107] Step 1: Connect the vibration sensor to the PLC and DCS, use the PLC and DCS to control the vibration sensor to collect the timing data of the motor bearing to be tested during operation, and pre-process the data collected by the vibration sensor to obtain the pre-processed data;

[0108] Preprocessing methods include data filtering and data cleaning;

[0109] Step 2: Use the preprocessed data as the input of the CNN model, and output the feature representation of the preprocessed data through the CNN model;

[0110] The CNN model is a CNN model trained based on the contrastive learning method; the training process of the CNN model is:

[0111] Step 1: Collect time series data sets of other motor bearings during historical operation, and the collected time series data sets include labeled time series data and unlabeled time series data; and then perform data enhancement on the collected time series data;

[0112] The enhancement methods used for data enhancement are time masking, adding noise and time scaling;

[0113] Step 2: construct positive sample pairs and negative sample pairs based on the enhanced time series dataset and self-supervised learning method;

[0114] Step 3: The CNN model is trained based on the constructed positive sample pairs and negative sample pairs. The loss function used in training is And use the gradient descent method for back propagation and CNN model parameter update;

[0115] Loss Function for:

[0116]

[0117] in, is the feature representation of the i-th anchor sample pair extracted by the CNN model, is the feature representation of the i-th positive sample pair extracted by the CNN model, is the feature representation of the i-th negative sample extracted by the CNN model, is the square of the L2 norm, α is a hyperparameter, [·] + ReLU operation represents the ReLU operation, and the number of anchor sample pairs, positive sample pairs, and negative sample pairs are all N.

[0118] Until the loss function Stop training when convergence and obtain the trained CNN model.

[0119] Step 3: Use the feature representation output in step 2 as the input of the support vector machine, and output the motor bearing fault detection result through the support vector machine.

[0120] The training process of support vector machine is:

[0121] Step (1), passing the labeled time series data collected in step 1 through the trained CNN model, and outputting the feature representation of each labeled time series data through the CNN model;

[0122] Step (2): The feature representations of the M labeled time series data are recorded as h1, h2, ..., h M , using h1,h2,…,h M Composition feature matrix m is the dimension of each feature representation;

[0123] The classification labels of M labeled time series data are recorded as y1, y2, ..., y M , using y1,y2,…,y M The label vector y = [y1, y2, ..., y N ];

[0124] Step (3), use the feature matrix H and the label vector y to train the support vector machine, calculate the classification loss according to the output g(H) of the support vector machine after the feature matrix H and the label vector y, and stop training when the classification loss converges to obtain a trained support vector machine.

[0125] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A device fault detection method based on time series contrast learning, characterized in that: The method specifically comprises the following steps: Step 1: Use sensors to collect time series data when the equipment to be tested is running, and pre-process the data collected by the sensors to obtain pre-processed data; Step 2: Use the preprocessed data as the input of the CNN model, and output the feature representation of the preprocessed data through the CNN model; The CNN model is a CNN model trained based on a contrastive learning method; Step 3: Use the feature representation output in step 2 as the input of the support vector machine, and output the equipment fault detection result through the support vector machine.

2. According to the method of equipment fault detection based on time series contrast learning in claim 1, it is characterized in that: The sensor is a temperature sensor, a current sensor, a vibration sensor, a pressure sensor or a radiation sensor.

3. The device fault detection method based on time series contrast learning according to claim 1 is characterized in that: The sensor is connected to the PLC and DCS, and the PLC and DCS are used to control the sensor to collect and pre-process data.

4. The device fault detection method based on time series contrast learning according to claim 1 is characterized in that: The data collected by the sensor is preprocessed, and the preprocessing method includes data filtering and data cleaning.

5. The device fault detection method based on time series contrast learning according to claim 1 is characterized in that: The training process of the CNN model is: Step 1: Collect a time series data set of the same type of equipment as the equipment to be detected during historical operation, and the collected time series data set includes labeled time series data and unlabeled time series data; then perform data enhancement on the collected time series data; Step 2: construct positive sample pairs and negative sample pairs based on the enhanced time series dataset and self-supervised learning method; Step 3: The CNN model is trained based on the constructed positive sample pairs and negative sample pairs. The loss function used in training is And use the gradient descent method for back propagation and CNN model parameter update; Until the loss function Stop training when convergence and obtain the trained CNN model.

6. The device fault detection method based on time series contrast learning according to claim 5 is characterized in that: The enhancement methods used for data enhancement of the collected time series data are time masking, adding noise and time scaling.

7. The device fault detection method based on time series contrast learning according to claim 5 is characterized in that: The loss function for: in, is the feature representation of the i-th anchor sample pair extracted by the CNN model, is the feature representation of the i-th positive sample pair extracted by the CNN model, is the feature representation of the i-th negative sample extracted by the CNN model, is the square of the L2 norm, α is a hyperparameter, [·] + ReLU operation represents the ReLU operation, and the number of anchor sample pairs, positive sample pairs, and negative sample pairs are all N.

8. The device fault detection method based on time series contrast learning according to claim 5 is characterized in that: The training process of the support vector machine is: Step (1), passing the labeled time series data collected in step 1 through the trained CNN model, and outputting the feature representation of each labeled time series data through the CNN model; Step (2): The feature representations of the M labeled time series data are recorded as h1, h2, ..., h M , using h1,h2,…,h M Composition feature matrix m is the dimension of each feature representation; The classification labels of M labeled time series data are recorded as y1, y2, ..., y M , using y1,y2,…,y M The label vector y = [y1, y2, ..., y N ]; Step (3), use the feature matrix H and the label vector y to train the support vector machine, calculate the classification loss according to the output g(H) of the support vector machine after the feature matrix H and the label vector y, and stop training when the classification loss converges to obtain a trained support vector machine.

9. The device fault detection method based on time series contrast learning according to claim 1 is characterized in that: The method further includes step 4, when testing the time series data of K consecutive time segments of the operation of the equipment to be tested, if the test results are all fault states, triggering a fault warning and notifying the staff to conduct equipment inspection.

10. A device fault detection system based on time series contrast learning, characterized in that: The system comprises a time series data acquisition module to be detected, a first neural network module and a second neural network module, wherein: The detection time series data acquisition module is used to collect the time series data of the device to be detected when it is running, pre-process the acquired time series data, and send the pre-processed data to the first neural network module; A CNN model is deployed in the first neural network module, and the CNN model is used to extract feature representations of the preprocessed data and send the extracted feature representations to the second neural network module; The training process of the CNN model is: Step 1: Collect a time series data set of the same type of equipment as the equipment to be detected during historical operation, and the collected time series data set includes labeled time series data and unlabeled time series data; then perform data enhancement on the collected time series data; Step 2: construct positive sample pairs and negative sample pairs based on the enhanced time series dataset and self-supervised learning method; Step 3: The CNN model is trained based on the constructed positive sample pairs and negative sample pairs. The loss function used in training is And use the gradient descent method for back propagation and CNN model parameter update; Loss Function for: in, is the feature representation of the i-th anchor sample pair extracted by the CNN model, is the feature representation of the i-th positive sample pair extracted by the CNN model, is the feature representation of the i-th negative sample extracted by the CNN model, is the square of the L2 norm, α is a hyperparameter, [·] + ReLU operation, the number of anchor sample pairs, positive sample pairs and negative sample pairs are all N; Until the loss function Stop training when convergence and obtain the trained CNN model; A support vector machine is deployed in the second neural network module, and the support vector machine is used to output the device fault detection result according to the feature representation; The training process of support vector machine is: Step (1), passing the labeled time series data collected in step 1 through the trained CNN model, and outputting the feature representation of each labeled time series data through the CNN model; Step (2): The feature representations of the M labeled time series data are recorded as h1, h2, ..., h M , using h1,h2,…,h M Composition feature matrix m is the dimension of each feature representation; The classification labels of M labeled time series data are recorded as y1, y2, ..., y M , using y1,y2,…,y M The label vector y = [y1, y2, ..., y N ]; Step (3), use the feature matrix H and the label vector y to train the support vector machine, calculate the classification loss according to the output g(H) of the support vector machine after the feature matrix H and the label vector y, and stop training when the classification loss converges to obtain a trained support vector machine.

Citation Information

Patent Citations

  • Rotating equipment small sample fault diagnosis method based on semi-supervised contrast learning

    CN117113198A

  • Industrial scene-oriented extremely few sample fault diagnosis method

    CN117349595A

  • Wind generating set bearing fault diagnosis method based on semi-supervised time sequence learning

    CN117686229A

  • Equipment fault real-time prediction method and device based on time sequence data and electronic equipment

    CN118503816A

Cited By

  • Substation equipment anomaly detection method and system based on data enhancement, computer readable storage medium and terminal equipment

    CN119885005A