Contrastive learning based bidirectional incremental multi-class fault classification method

Abstract

The application discloses a bidirectional incremental multi-class fault classification method based on contrast learning, wherein bidirectional refers to that during the incremental updating process, both maintenance of historical information for current learning and updating of the historical information along with the current learning are included. Due to the characteristics that the distance between homogeneous class samples and the feature center of the class is relatively short, and the distance between heterogeneous class feature centers is relatively long, the application solves the problem of how to retain old knowledge in incremental fault classification through contrast learning on the basis of establishing the new and old task features through a cross-entropy loss function. Meanwhile, the feature center of the historical class is updated through random movement and weighted exponential, the class center drift problem caused by parameter updating is eliminated, the old feature space is updated, the new feature space is aligned, the incremental fault classification effect is better, and the accuracy is higher.

Description

Technical Field

[0001] This invention belongs to the field of industrial process monitoring, and in particular relates to a bidirectional incremental multi-class fault classification method based on contrastive learning. Background Technology

[0002] Incremental multi-class classification modeling is a very common problem in machine learning or deep learning classification, widely existing in various fields such as bioinformatics, smart grids, medical imaging, and fault diagnosis. Most existing classification methods are based on the assumption that observational data can be acquired in a single instance. However, real-world industrial datasets often violate this assumption, exhibiting diverse fault categories over time. For example, data-driven fault classification is a crucial component of industrial process monitoring. Due to the complexity of process industries, equipment failures present multiple scenarios and exhibit fault categories that differ from historical categories over time. In such cases, updating the classification model is necessary. However, training the model using only data with new fault categories leads to a decline in the model's historical performance; retraining the model using massive amounts of historical industrial process data wastes storage and computing resources. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a bidirectional incremental multi-class fault classification method based on contrastive learning. This method can update the model while minimizing storage and computational resource consumption, without excessively sacrificing the accuracy of historical classes. It can achieve good fault classification results for incremental multi-class problems. The specific technical solution is as follows:

[0004] A bidirectional incremental multi-class fault classification method based on contrastive learning, comprising two parts: offline training and online application;

[0005] The offline training includes:

[0006] Set the number of rounds for the classification task to t. When t = 1, execute S1.1-1-4.

[0007] S1.1: Collect offline industrial process data samples of C1 categories under the current operating conditions, including normal operating condition data and various fault data;

[0008] S1.2: Establish a neural network fault classification model. According to the stochastic gradient descent method, backpropagate to train the cross-entropy loss function to update the network parameters. The neural network fault classification model includes a feature extractor and a linear classifier. The input of the neural network fault classification model is the fault sample, and the output is the prediction confidence of each fault category.

[0009] S1.3: Based on the reservoir algorithm, operate on sample data of unknown length, assign equal probability to each sample, and update the historical database;

[0010] S1.4: Calculate the feature center point for each category and save it to the historical database.

[0011] Whenever a new fault category appears, a new round of classification task begins, i.e. when t≥2, steps S1.5~S1.8 are executed;

[0012] S1.5: Collect offline industrial process data samples of category C2 under the working conditions of the t-th round of tasks. These samples are different from the fault categories in the first round and are used as an incremental task.

[0013] S1.6: Calculate the temporary feature center point for each category;

[0014] S1.7: Construct a loss function, which is a weighted sum of the current task classification loss function, the contrastive learning loss function, and the cross-entropy loss function of historical samples in the historical database; train the loss function by backpropagation according to the stochastic gradient descent method, and randomly update the center points of historical category features during the training process by using the moving weighted average method;

[0015] S1.8: Update the historical database according to the reservoir algorithm, keeping the total number of samples in the historical database unchanged; at the same time, for each new fault category, calculate the feature center point of each category and save it into the historical database.

[0016] The online applications include:

[0017] Obtain online samples, input them into the trained neural network fault classification model, obtain the output of the feature extractor, calculate the Euclidean distance between the output of the feature extractor and the feature center points of all categories, and select the category of the feature center point corresponding to the smallest Euclidean distance as the category of the online sample.

[0018] Further, step S1.3 includes the following sub-steps:

[0019] ① Set the upper limit for storing sample data in the historical database to be [value]. The sample count is S;

[0020] ② When new sample data appears, the sample data count S is incremented by 1;

[0021] ③ Determine the sample data count S and the upper limit of the historical database storage of sample data. If the relationship, Then the current sample is stored in the historical database, with record number S; if... Then select a random integer b from 0 to S. If Then the sample is stored in the historical database, with record number b. The sample will not be stored in the historical database.

[0022] ④ Repeat steps ② and ③ until all samples in the current round have been traversed and the historical database has been updated.

[0023] Furthermore, in S1.4, the calculation formula for the center point of each category feature is as follows:

[0024]

[0025] Among them, e c x is the feature center point of category c. i Let g represent the i-th sample. t,c N represents the set of all samples of category c in task round t. c,t Represents the sample set g t,c The number of samples in the sample is φ(·), which is the feature extractor.

[0026] Furthermore, in S1.6, the calculation formula for the temporary feature center point of each category is as follows:

[0027]

[0028] Among them, e c ′ ,t x is the temporary feature center point of category c. i Let B represent the i-th sample. c,t Let |B| represent the set of batch samples of category c under task round t. c,t | represents sample set B c,t The number of samples in the sample is φ(·), which is the feature extractor.

[0029] Furthermore, the cross-entropy loss function The calculation formula is:

[0030]

[0031] Where, p i The input sample x i The prediction confidence score y for each category obtained by the neural network fault classification model i The input sample x i The corresponding label, N t This represents the number of training samples in the t-th round of the task.

[0032] Furthermore, in S1.7, the loss function The calculation formula is as follows:

[0033]

[0034] in, The classification loss function for the current task is used to enable the model to learn fault information under the current operating conditions; A contrastive learning loss function is used to achieve compactness within homogeneous classes and separation between different classes; The cross-entropy loss of historical samples in the historical database is used to allow the model to retain the learned historical knowledge; a1, a2 and α3 are the weight parameters of the current task classification loss, the contrastive learning loss and the historical task classification loss, respectively. This represents the historical sample data stored in the historical database at task round t; C 1:t-1 Let t represent all categories encountered in the previous rounds; E(a,b) represents the cosine distance obtained by normalizing the input elements a and b, scaling the cosine distance by a coefficient τ, and then using it as a power of e. The final result is used as the contrastive loss function. One of the comparison distances in the middle; e j e represents the central point of historical data features. t ' represents all temporary feature center points in the current round. Further, in S1.7, the formula for calculating the historical category feature center points is:

[0035] e c =αe c +(1-α)e′ c,t

[0036]

[0037] Where α is the weighting parameter, e′ c,t It is a feature extractor trained in the current task round, which obtains temporary feature centers for historical samples of category c in the historical database in the current task round t.

[0038] Furthermore, in S1.7, the random update specifically involves: generating a random number; if the random number is less than the random update frequency parameter r... p If the condition is met, the center points of the historical category features will be updated; otherwise, the center points of the historical category features will not be updated.

[0039] A bidirectional incremental multi-class fault classification device based on contrastive learning includes one or more processors for implementing a bidirectional incremental multi-class fault classification method based on contrastive learning.

[0040] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements a bidirectional incremental multi-class fault classification method based on contrastive learning.

[0041] The beneficial effects of this invention are as follows:

[0042] The bidirectional incremental multi-class fault classification method of this invention, in its bidirectional incremental update process, includes both maintaining historical information during the current learning phase and updating historical information as the current learning progresses. Due to the characteristics of homogeneous class samples being close to their own class feature centers and heterogeneous class feature centers being far apart, this invention addresses the problem of retaining old knowledge in incremental fault classification by using contrastive learning, based on establishing new and old task features through a cross-entropy loss function. Simultaneously, updating the feature centers of historical classes by randomly shifting the weighted exponent eliminates the class center drift problem caused by parameter updates, ensuring the old feature space is updated and aligned with the new feature space, resulting in better incremental fault classification performance and higher accuracy. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of a bidirectional incremental multi-class fault classification method based on contrastive learning, according to an embodiment of the present invention.

[0044] Figure 2 A flowchart of a benchmark process for Eastman (TE) Industries in Tennessee;

[0045] Figure 3 The image shows the 2D mapping of data extraction features from the method of this invention and two existing methods after dimensionality reduction using t-SNE. Detailed Implementation

[0046] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0047] This invention addresses the fault classification problem under incremental updates across multiple categories. It defines offline modeling and online testing datasets and first uses a cross-entropy loss function to classify or identify fault categories. In subsequent incremental updates, considering the characteristics of homogeneous class samples being close to their own category feature centers and heterogeneous class feature centers being far apart, a contrastive learning approach using cosine distance as a metric is adopted to address the issue of preserving old knowledge in incremental fault classification. Simultaneously, the feature centers of historical categories are updated by randomly shifting the weighted exponent, eliminating the category center drift problem caused by parameter updates, thus aligning the old feature space with the new feature space. Finally, online testing of fault classification is performed using cosine distance as the metric.

[0048] The method of this invention is applicable to data-driven fault classification scenarios such as bioinformatics, smart grids, medical imaging, and fault diagnosis, based on industrial process data.

[0049] On the one hand, such as Figure 1As shown, the bidirectional incremental multi-class fault classification method based on contrastive learning of the present invention includes the following steps:

[0050] S1: Offline training;

[0051] Set the number of rounds for the classification task to t. When t = 1, execute S1.1-1-4.

[0052] S1.1: Collect offline industrial process data samples of C1 categories under the current operating conditions, including normal operating condition data and various fault data;

[0053] S1.2: Establish a neural network fault classification model f, and train the cross-entropy loss function using backpropagation based on the stochastic gradient descent method. Update network parameters.

[0054] The neural network fault classification model includes a feature extractor and a linear classifier, as follows:

[0055]

[0056] Where φ(·) is the feature extractor, and w is the linear classifier. This represents the operation of concatenating φ and w; the input to the neural network fault classification model is sample x. i The output is the prediction confidence score for each fault category. The feature extractor consists of alternating layers of fully connected layers and non-linear activation function layers, with the activation function being the ReLU function. The specific number of layers in the feature extractor can vary depending on the actual application scenario. In this embodiment, for the TE process, the feature extractor consists of three alternating layers of fully connected layers and three layers of non-linear activation function ReLU layers. The linear classifier is a single fully connected layer.

[0057] The cross-entropy loss function is as follows:

[0058]

[0059] Where, p i The input sample x i The prediction confidence score y for each category obtained by the neural network fault classification model i The input sample x i The corresponding label, N t This represents the number of training samples in the t-th round of the task.

[0060] S1.3: Based on the reservoir algorithm, operate on the sample data of unknown length, assign an equal probability to each sample to be stored in the historical database, and update the historical database. This includes the following sub-steps:

[0061] ① Set the upper limit for storing sample data in the historical database to be [value]. The sample count is S;

[0062] ② When new sample data appears, the sample data count S is incremented by 1.

[0063] ③ Determine the sample data count S and the upper limit of the historical database storage of sample data. The relationship. If Then the current sample is stored in the historical database, with record number S; if... Then select a random integer b from 0 to S. If Then the sample is stored in the historical database, with record number b. The sample will not be stored in the historical database.

[0064] ④ Repeat steps ② and ③ until all samples in the current round have been traversed and the historical database has been updated.

[0065] S1.4: Calculate the feature center point of each category using the following formula. Save to historical database:

[0066]

[0067] Among them, e c x is the feature center point of category c. i Let g represent the i-th sample. t,c N represents the set of all samples of category c in task round t. c,t Represents the sample set g t,c The number of samples in the sample.

[0068] Whenever a new fault category appears, a new round of classification task begins, i.e. when t≥2, steps S1.5~S1.8 are executed;

[0069] S1.5: Collect offline industrial process data samples of category C2 under the working conditions of the t-th round of tasks. These samples are different from the fault categories in the first round and are used as an incremental task.

[0070] S1.6: Calculate the temporary feature center point for each category using the following formula:

[0071]

[0072] Among them, e′ c,t x is the temporary feature center point of category c. i Let B represent the i-th sample. c,t Let |B| represent the set of batch samples of category c under task round t. c,t | represents sample set B c,t The number of samples in the sample.

[0073] S1.7: Construct the loss function as follows Based on stochastic gradient descent, the training loss function is backpropagated. The network parameters are updated, and during training, historical category feature centers are randomly updated using a moving weighted average method.

[0074] Among them, the loss function The calculation formula is as follows:

[0075]

[0076] in, The classification loss function for the current task is used to enable the model to learn fault information under the current operating conditions; A contrastive learning loss function is used to achieve compactness within homogeneous classes and separation between different classes; The cross-entropy loss of historical samples in the historical database is used to allow the model to retain the learned historical knowledge; a1, a2 and α3 are the weight parameters of the current task classification loss, the contrastive learning loss and the historical task classification loss, respectively. This represents the historical sample data stored in the historical database at task round t; C 1:t-1 Let t represent all categories encountered in the previous rounds; E(a,b) represents the cosine distance obtained by normalizing the input elements a and b, scaling the cosine distance by a coefficient τ, and then using it as a power of e. The final result is used as the contrastive loss function. One of the comparison distances in the middle; Specifically, this involves realizing the historical data feature center e j and all temporary feature centers e in the current round t 'Away from historical data samples in historical databases' and the same category feature center e j compact.

[0077] Historical category feature center point The calculation formula is as follows:

[0078] e c =αe c +(1-α)e′ c,t

[0079]

[0080] Where α is the weighting parameter, e′ c,t These are temporary feature centers obtained by the feature extractor trained in the current task round from historical samples of category c in the historical database under the current task round; random update refers to the system generating random numbers, if the number is less than the random update frequency parameter r. pIf the historical category feature center points are updated, then the historical category feature center points will be updated; otherwise, they will not be updated.

[0081] S1.8: Update the historical database according to the reservoir algorithm, keeping the total number of samples in the historical database unchanged; at the same time, for each new fault category, calculate the feature center point of each category. Saved to the historical database. Category feature centroid. The calculation formula is the same as that for S1.4.

[0082] S2: Online Application Testing

[0083] S2.1: Obtain online samples;

[0084] S2.2: Calculate the output of the feature extractor of the neural network fault classification model after training online samples:

[0085] z online =φ(x online )

[0086] Where, x online This indicates that the sample was entered online.

[0087] S2.3: Output the sample z online and the center point e of all categories of features c Normalize and dot product, that is, calculate the sample output z. online With all category feature center points e c The Euclidean distance is used to select the category of the online sample whose category feature center point corresponds to the smallest Euclidean distance.

[0088]

[0089] On the other hand, the present invention also provides a bidirectional incremental multi-class fault classification device based on contrastive learning, comprising one or more processors for implementing the bidirectional incremental multi-class fault classification method based on contrastive learning in the above embodiments. Embodiments of the bidirectional incremental multi-class fault classification device based on contrastive learning of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar devices. The device embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the device with data processing capabilities reading the corresponding computer program instructions from non-volatile memory into memory and executing them. From a hardware perspective, a hardware structure diagram of the device with data processing capabilities where the bidirectional incremental multi-class fault classification device based on contrastive learning is located includes, in addition to the processor, memory, network interface, and non-volatile memory, other hardware components may also be included depending on the actual function of the device in the embodiments; these will not be elaborated further.

[0090] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the bidirectional incremental multi-class fault classification method based on contrastive learning described in the above embodiments.

[0091] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computing program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0092] The effectiveness of this invention is verified using a specific industrial example. The proposed method is evaluated using the Eastman (TE) industrial benchmark process in Tennessee. The TE process is an industrial simulation platform developed by Eastman Chemical Company based on actual chemical processes and has been widely used to test the effectiveness of process monitoring and fault diagnosis methods. The TE process flow is as follows: Figure 2 As shown.

[0093] The TE data has 52 process variables, and all categories consist of 21 fault categories and 1 normal category. A subset of fault categories are randomly selected as incremental fault diagnosis scenarios. The comparison methods chosen include DER (Dark Knowledge Distillation Network), iCaRL (Historical Data Replay), SGD (Stochastic Gradient Update), and this invention (Bidirectional Multi-Class Incremental Fault Classification Based on Comparative Learning).

[0094] An incremental network for different rounds was trained on the TE process training samples using a bidirectional multi-class incremental fault classification method based on contrastive learning. The network trained offline was then used to predict online samples (test set). The incremental results for 10 rounds are shown in Table 1.

[0095] Table 1 shows the classification performance of various methods on TE process data.

[0096]

[0097]

[0098] As can be seen from Table 1, the final accuracy and average accuracy of the bidirectional multi-class incremental fault classification method based on contrastive learning proposed in this invention are significantly better than the contrastive methods.

[0099] To more intuitively demonstrate the superiority of the method of this invention, a 2D graph is drawn showing the dimensionality reduction of all categories of data after feature extraction and t-SNE method after each round of training in the case of five incremental rounds. For example... Figure 3 As shown, different colors represent different categories. The SGD method is represented in the left 5 columns. Figure 3 In the table, (a), (d), (g), (k), and (n) correspond to task rounds 1, 2, 3, 4, and 5 of the SGD method, respectively; the middle 5 columns represent the DER method. Figure 3 In this context, (b), (e), (h), (l), and (o) correspond to task rounds 1, 2, 3, 4, and 5 of the DER method, respectively; in this method, the right-hand column represents the 5 columns. Figure 3 In the figure, (c), (f), (i), (m), and (p) correspond to task rounds 1, 2, 3, 4, and 5 of this method, respectively. As can be seen from the figure, the bidirectional multi-class incremental fault classification method based on contrastive learning in this invention obtains more distinct boundaries between each class, thus indicating that the classification performance of the method in this invention is improved.

[0100] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A bidirectional incremental multi-class fault classification method based on contrastive learning, characterized in that, It includes two parts: offline training and online application; The offline training includes: Set the number of rounds for the classification task to t. When t = 1, execute S1.1-1-4. S1.1: Collect offline industrial process data samples of C1 categories under the current operating conditions, including normal operating condition data and various fault data; S1.2: Establish a neural network fault classification model. According to the stochastic gradient descent method, backpropagate to train the cross-entropy loss function to update the network parameters. The neural network fault classification model includes a feature extractor and a linear classifier. The input of the neural network fault classification model is the fault sample, and the output is the prediction confidence of each fault category. S1.3: Based on the reservoir algorithm, operate on the sample data of unknown length, assign an equal probability to each sample to be stored in the historical database, and update the historical database. S1.4: Calculate the feature center point for each category and save it to the historical database. Whenever a new fault category appears, a new round of classification task begins, i.e. when t≥2, steps S1.5~S1.8 are executed; S1.5: Collect offline industrial process data samples of category C2 under the working conditions of the t-th round of tasks. These samples are different from the fault categories in the first round and are used as an incremental task. S1.6: Calculate the temporary feature center point for each category; S1.7: Construct a loss function, which is a weighted sum of the current task classification loss function, the contrastive learning loss function, and the cross-entropy loss function of historical samples in the historical database; train the loss function by backpropagation according to the stochastic gradient descent method, and randomly update the center points of historical category features during the training process by using the moving weighted average method; S1.8: Update the historical database according to the reservoir algorithm, keeping the total number of samples in the historical database unchanged; at the same time, for each new fault category, calculate the feature center point of each category and save it into the historical database. The online applications include: Obtain online samples, input them into the trained neural network fault classification model, obtain the output of the feature extractor, calculate the Euclidean distance between the output of the feature extractor and the feature center points of all categories, and select the category of the feature center point corresponding to the smallest Euclidean distance as the category of the online sample.

2. The bidirectional incremental multi-class fault classification method based on contrastive learning according to claim 1, characterized in that, Step S1.3 includes the following sub-steps: ① Set the upper limit for storing sample data in the historical database to be [value]. The sample count is S; ② When new sample data appears, the sample data count S is incremented by 1; ③ Determine the sample data count S and the upper limit of the historical database storage of sample data. If the relationship, Then the current sample is stored in the historical database, with record number S; if... Then select a random integer b from 0 to S. If Then the sample is stored in the historical database, with record number b. The sample will not be stored in the historical database. ④ Repeat steps ② and ③ until all samples in the current round have been traversed and the historical database has been updated.

3. The bidirectional incremental multi-class fault classification method based on contrastive learning according to claim 1, characterized in that, In S1.4, the calculation formula for the center point of each category feature is as follows: Among them, e c x is the feature center point of category c. i Let g represent the i-th sample. t,c N represents the set of all samples of category c in task round t. c,t Represents the sample set g t,c The number of samples in the sample is φ(·), which is the feature extractor.

4. The bidirectional incremental multi-class fault classification method based on contrastive learning according to claim 1, characterized in that, In S1.6, the calculation formula for the temporary feature center point of each category is as follows: Among them, e c ′ ,t x is the temporary feature center point of category v. i Let B represent the i-th sample. c,t Let |B| represent the set of batch samples of category v under task round t. c,t | represents sample set B c,t The number of samples in the sample is φ(·), which is the feature extractor.

5. The bidirectional incremental multi-class fault classification method based on contrastive learning according to claim 4, characterized in that, The cross-entropy loss function The calculation formula is: Where, p i The input sample x i The prediction confidence score y for each category obtained by the neural network fault classification model i The input sample x i The corresponding label, N t This represents the number of training samples in the t-th round of the task.

6. The bidirectional incremental multi-class fault classification method based on contrastive learning according to claim 5, characterized in that, In S1.7, the loss function The calculation formula is as follows: in, The classification loss function for the current task is used to enable the model to learn fault information under the current operating conditions; A contrastive learning loss function is used to achieve compactness within homogeneous classes and separation between different classes; The cross-entropy loss of historical samples in the historical database is used to allow the model to retain the learned historical knowledge; α1, α2 and α3 are the weight parameters of the current task classification loss, the contrastive learning loss and the historical task classification loss, respectively. This represents the historical sample data stored in the historical database at task round t; C 1:t-1 Let t represent all categories encountered in the previous rounds; E(a,b) represents the cosine distance obtained by normalizing the input elements a and b, scaling the cosine distance by a coefficient τ, and then using it as a power of e. The final result is used as the contrastive loss function. One of the comparison distances in the middle; e j e′ represents the central point of historical data features. t This represents the center point of all temporary features in the current round.

7. The bidirectional incremental multi-class fault classification method based on contrastive learning according to claim 6, characterized in that, In S1.7, the formula for calculating the center point of historical category features is: e c =αe c +(1-α)e′ c,t Where α is the weighting parameter, e′ c,t It is a temporary feature center obtained by the feature extractor trained in the current task round from historical samples of category c in the historical database in the current task round.

8. The bidirectional incremental multi-class fault classification method based on contrastive learning according to claim 7, characterized in that, In S1.7, the random update specifically involves: generating a random number; if the random number is less than the random update frequency parameter r... p If the condition is met, the center points of the historical category features will be updated; otherwise, the center points of the historical category features will not be updated.

9. A bidirectional incremental multi-class fault classification device based on contrastive learning, characterized in that, It includes one or more processors for implementing the bidirectional incremental multi-class fault classification method based on contrastive learning as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the bidirectional incremental multi-class fault classification method based on contrastive learning as described in any one of claims 1 to 8.

Images