Fault diagnosis method and system for fault large-scale unit based on sharing and specific representation learning

Abstract

The invention belongs to the technical field of unit fault diagnosis, and discloses a fault diagnosis method and system for a fault large-scale unit based on sharing and specific representation learning. The method not only can be applied to multi-measuring-point fault diagnosis of a large-scale unit, but also has wide applicability. In other multi-sensor systems or complex equipment, the fault detection effect can also be improved through multi-view representation learning. The invention provides a weighted sharing-specific multi-view depth representation learning fault diagnosis method for fault diagnosis of multi-measuring-point data of a large-scale unit. The method consists of two parts, namely sharing and specifically representing a learning network and weighted multiple similarity loss and orthogonal loss. A multi-measuring-point data fault diagnosis problem of a large-scale unit is solved by sharing a specific representation learning network, and weighted multiple similarity loss and orthogonal loss are used to further improve the feature extraction capability of the network and reduce redundancy.

Description

Technical Field

[0001] The present invention belongs to the technical field of unit fault diagnosis, and in particular relates to a fault diagnosis method and system for large-scale unit faults based on shared and specific representation learning. Background Art

[0002] Prior art 1: Deep learning fault diagnosis method based on single view

[0003] In traditional single-view fault diagnosis methods, feature extraction and fault classification of vibration signals usually rely on a single data source or view. Single-view methods based on deep learning improve the accuracy of fault diagnosis by designing different network structures, such as convolutional neural networks (CNN) or long short-term memory networks (LSTM). These methods process the input vibration signal, extract the corresponding time domain or frequency domain features, and classify these features using deep learning models. However, this method faces the following problems:

[0004] Single view: Traditional methods can only use information from a single measurement point or data source and cannot process multiple sensor data at the same time, resulting in insufficient processing capabilities for complex multi-source data.

[0005] Feature redundancy: Since it is impossible to effectively distinguish features of different categories, the extracted features are prone to redundancy or excessive similarity, which affects the accuracy of the diagnosis results.

[0006] Prior Art 2: Multi-Channel Convolutional Neural Network (MC-CNN)

[0007] Multi-channel convolutional neural network (MC-CNN) is an improved fault diagnosis method that uses multiple sensors or data channels and integrates the data of these channels into the network for parallel processing. MC-CNN can process input data from multiple channels at the same time, such as multi-dimensional vibration signals, and extract features from different channels. Although MC-CNN can process multi-source data to a certain extent and improve feature extraction capabilities, its limitations are:

[0008] Insufficient distinction between shared and specific information: MC-CNN only processes data in parallel through multiple channels and fails to effectively distinguish between shared information and specific information in data from different views, resulting in confusion between shared information and specific information.

[0009] Insufficient processing of the relationship between views: MC-CNN processes the data between each channel independently, without considering the information association between different views, and cannot effectively improve the overall diagnostic accuracy.

[0010] Through the above analysis, the problems and defects of the prior art are as follows:

[0011] (1) Single view: Traditional methods can only use information from a single measurement point or data source and cannot process multiple sensor data at the same time, resulting in insufficient processing capabilities for complex multi-source data.

[0012] (2) Feature redundancy: Since it is impossible to effectively distinguish features of different categories, the extracted features are prone to redundancy or excessive similarity, which affects the accuracy of the diagnosis results.

[0013] (3) Insufficient distinction between shared and specific information: MC-CNN only processes data in parallel through multiple channels and fails to effectively distinguish between shared information and specific information in data from different views, resulting in confusion between shared information and specific information.

[0014] (4) Insufficient processing of the relationship between views: MC-CNN processes the data between each channel independently, without considering the information association between different views, and cannot effectively improve the overall diagnostic accuracy. Summary of the invention

[0015] In view of the problems existing in the prior art, the present invention provides a fault diagnosis method for large-scale faulty units based on shared and specific representation learning.

[0016] The present invention provides a fault diagnosis method for a large-scale unit with shared and characteristic representation learning, comprising:

[0017] Step 1: introduce the multi-view representation learning method and apply it to the fault diagnosis of multi-measurement point data of large units;

[0018] Four view data of X1, X2, Y1 and Y2 are constructed based on the original vibration signals collected by sensors at four measuring points located at the X-axis coupling end, X-axis non-coupling end, Y-axis coupling end and Y-axis non-coupling end;

[0019] Each view data is input into the shared representation network W1 and the specific representation network W2 to obtain shared representations G1, G2, G3, G4 and specific representations S1, S2, S3, S4 respectively; weighted orthogonal loss and similarity loss are introduced; among them, the weighted orthogonal loss ensures the independence of shared information and specific information extracted from the same view by imposing orthogonality constraints, reduces the redundancy between features, and thus improves the ability to separate fault features; while the weighted similarity loss ensures that the shared information of different views maintains consistency in fault features through adversarial training, preventing the separation effect of shared information between different views from being overly dispersed;

[0020] The shared representation G is connected with the specific representations S1, S2, S3, and S4 to obtain the final representation R, and then the final representation is input into the classification network to obtain the fault diagnosis result;

[0021] Step 2, improve the dynamics of the weighting mechanism;

[0022] In multi-measurement point data, no matter it is view X1, X2 or Y1, Y2, they are all conceptualized as a sample space, and a single sample is represented as a point distributed in a high-dimensional space; on the one hand, subsets of different view data correspond to conditional representations of the same fault, and this representation is shared and consistent in all view data; therefore, the present invention constructs a shared network W1 to extract the shared vibration signal features in each view and represents it as cross-view shared information features G k On the other hand, in each view, some segments of the vibration signal represent fault information that is different from the information shared with other views. This part of information is the specific information expression of the vibration signal on the fault. Therefore, the present invention also constructs a specific network W2 to extract the specific information features of the vibration signal of each view, and represents it as the specific information features S1, S2, S3, S4 of the view.

[0023] If X m =[x1, x2, …, x n ] T represents all samples of the mth view, where n represents the total number of samples, x i Represents a sample; WSMDRL uses the shared network W1 to i Mapping is performed to obtain its shared information mapping and a weight α is set to reduce the redundant information between the four view data to obtain the shared information mapping G k for:

[0024]

[0025] Similarly, WSMDRL uses a specific network W2 to i Mapping is performed to obtain its specific information mapping as follows:

[0026]

[0027] In order to fully utilize the complementarity of multi-view data, each view pair x i The specific mappings are retained with the shared mapping G k Connected together, it is represented as:

[0028] R=[G k , S1,…,S4] (3)

[0029] Weighted multiple similarity constraints and orthogonality constraints:

[0030] Weighted Multiple Similarity Loss

[0031] Similarity loss can ensure the similarity between shared information of different views. WSMDRL draws on the adversarial similarity constraint from the generative adversarial network to design similarity loss and constructs a weighted multiple similarity loss function to learn the shared information between different views. Specifically, WSMDRL regards the shared network as a generator G, and for the kth view of the i-th sample make Represents the shared information generated from this view, using the Sigmoid function S(x) = 1 / (1+e -x ) Compress the generated shared information to obtain the final generated shared information:

[0032] In addition, the present invention uses an M-classified classifier as the discriminator D to identify the distribution to which the shared information belongs; The probability from the kth view is:

[0033]

[0034] During the training phase, the discriminator D tries to find the source view of each shared information, that is, to maximize P i k ; Instead, the generator G tries to minimize this probability. Putting these two objectives together, we get a maximum-minimization game:

[0035]

[0036] In formula (5), N is the total number of samples, is the true label;

[0037] Since the shared information generated by different views is redundant, a similarity factor λ1 is introduced to balance the similarity loss, and the final weighted similarity loss is:

[0038]

[0039] Weighted Orthogonal Loss:

[0040] The orthogonality constraint is a constraint designed to encourage or force the parameter matrix to remain orthogonal during the training process, which can decompose the shared information and the specific information. The present invention constructs a weighted orthogonality loss function to separate the shared information and the specific information from different views, making the meaning between them clearer and more explicit, while also avoiding their mutual contamination.

[0041] If the shared features and specific features of the i-th sample of view m are compressed using the sigmoid function, they are expressed as The present invention introduces L2 norm normalization and And the shared information is orthogonal to the specific information, and the orthogonal loss of the sample is obtained as:

[0042]

[0043] In addition, in order to balance the weight of each view, the present invention introduces an orthogonal factor μ1 for each view, and finally sums the weighted orthogonal losses to obtain the total weighted orthogonal loss:

[0044]

[0045] Step 3, classification network;

[0046] The classification network consists of three fully connected layers, an activation function, and a regularization term. First, Dropout is used to deal with the overfitting problem of the model. Assume that the Dropout function is dp(x), the loss probability p is 0.5, and ReLU is selected as the activation function. For each sample x in each view i According to formula (1) and formula (2), we can get the shared mapping and the specific mapping and connect them to get the feature representation R i Finally, R i Input the classification network to obtain the fault classification result. If w represents the weight parameter and b represents the bias parameter, Representative feature representation R i After the output of the regularization and activation function, the output of the fully connected layer is:

[0047]

[0048] Furthermore, the multi-view representation learning method:

[0049] The data of multiple measuring points are regarded as multiple views, and through the shared and specific representation network, the shared information and specific information between views can be effectively distinguished while extracting the data features of each view.

[0050] Furthermore, the dynamics of the weighting mechanism is improved:

[0051] Based on multi-view representation learning, weighted multiple similarity loss and orthogonal loss are constructed to further optimize feature extraction and separation;

[0052] Similarity loss: ensures that the shared information extracted between views has high consistency and avoids the problem of excessive dispersion of shared information between different views;

[0053] Orthogonal loss: By introducing orthogonal constraints, the shared features and specific features of each view are ensured to be independent of each other, reducing redundant information and improving feature discrimination capabilities.

[0054] Another object of the present invention is to provide a fault diagnosis system for large-scale units with shared and specific representation learning, comprising:

[0055] Multi-view module, which is used to introduce multi-view representation learning methods and apply them to fault diagnosis of multi-measurement point data of large units;

[0056] The weighting module is used to weight the similarity loss and orthogonality loss based on fixed weight settings. On this basis, a dynamic weight adjustment mechanism is introduced to automatically adjust the weight parameters of similarity and orthogonality constraints according to different data characteristics and fault categories, further improving the model's adaptive ability;

[0057] Classification module, used to classify networks.

[0058] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the fault diagnosis method for large-scale fault units with shared and specific representation learning.

[0059] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the fault diagnosis method for a large-scale unit with shared and specific representation learning.

[0060] Another object of the present invention is to provide an information data processing terminal, which is used to implement the shared and specific representation learning fault large-scale unit fault diagnosis system.

[0061] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0062] The present invention proposes a weighted shared-specific multi-view deep representation learning fault diagnosis method for fault diagnosis of multi-point data of large units. The method consists of two parts, namely shared and specific representation learning networks and weighted multiple similarity loss and orthogonal loss. The problem of multi-point data fault diagnosis of large units is addressed by sharing a specific representation learning network, and weighted multiple similarity loss and orthogonal loss are used to further enhance the network's feature extraction capability and reduce redundancy.

[0063] First, compared with the traditional single-view method and MC-CNN, the present invention introduces a multi-view representation learning method for the first time and applies it to the fault diagnosis of multi-point data of large units. Unlike the existing technology that only processes a single data source or processes in parallel through multiple channels, the present invention regards the data of multiple measuring points as multiple views, and through the shared and specific representation network, while extracting the features of each view data, effectively distinguishes the shared information and specific information between views.

[0064] Distinguishing technical features: The shared representation learning network can extract the common features of multiple measurement point data and project them into the shared information subspace, reducing the differences between views; while the specific representation learning network is used to extract the unique features of each view, thus avoiding feature confusion.

[0065] Based on multi-view representation learning, the present invention constructs weighted multiple similarity losses and orthogonal losses to further optimize feature extraction and separation.

[0066] Similarity loss: ensures that the shared information extracted between views has high consistency and avoids the problem of excessive dispersion of shared information between different views.

[0067] Orthogonal loss: By introducing orthogonal constraints, the shared features and specific features of each view are ensured to be independent of each other, which reduces redundant information and improves feature discrimination ability.

[0068] Compared with the prior art, the technical solution of the present invention greatly improves the processing capability of complex multi-source data through multi-view learning and shared-specific information separation, and can effectively deal with the problem of multi-point data in large-scale unit fault diagnosis.

[0069] Analysis of the inventiveness of the present invention

[0070] 1) Innovative multi-view representation learning

[0071] Traditional fault diagnosis methods cannot fully utilize the advantages of multi-point data, but the present invention provides a unified processing framework for data from different points by introducing multi-view representation learning. The innovative mechanism of sharing and separating specific information not only improves the accuracy of feature extraction, but also significantly improves the accuracy of fault diagnosis by reducing information redundancy between different views. This innovative design provides a new fault diagnosis method for complex unit systems.

[0072] 2) Weighted Similarity and Orthogonality Constraints

[0073] Although the existing multi-view learning methods can process multi-source data, they are still insufficient in information sharing between views and separation of specific information. The present invention achieves consistency of information sharing between views and independence of information within views by introducing weighted similarity constraints and orthogonality constraints. This dual constraint mechanism not only optimizes the feature extraction process, but also avoids the interference of redundant information, ensuring the robustness and accuracy of the fault diagnosis model.

[0074] 3) Wide applicability of multi-view fault diagnosis

[0075] The present invention can not only be applied to multi-point fault diagnosis of large units, but also has wide applicability. For example, in other multi-sensor systems or complex equipment, the effect of fault detection can also be improved through multi-view representation learning. Therefore, the method of the present invention not only has targeted technical innovation, but also demonstrates its universality in different fields.

[0076] Second, the technical solution of the present invention mainly solves the following key technical problems in the prior art in industrial applications and has achieved significant results: 1. Multi-source data fusion problem

[0077] In the fault diagnosis of large units (such as generators, chemical equipment, etc.), data often comes from multiple different sensor measurement points, such as temperature, vibration, pressure, etc. These measurement point data have different physical meanings and data distributions. Traditional fault diagnosis methods can usually only process single or limited measurement point data, and it is difficult to fully and effectively integrate these heterogeneous data, resulting in poor diagnostic results.

[0078] The present invention introduces a multi-view representation learning method, which treats the data of multiple measurement points as different perspectives, and fully captures the correlation and fault characteristics between different data sources through unified modeling and feature fusion. Compared with traditional single-view data processing, the present invention can comprehensively analyze multi-dimensional data and significantly improve the accuracy and robustness of fault diagnosis.

[0079] 2. Limitations of the Fixed Weight Mechanism

[0080] In traditional fault diagnosis methods, the loss function of the model is usually based on fixed weight settings and cannot be flexibly adjusted according to specific data characteristics and fault types. This fixed weight setting easily leads to the model being unable to fully exert its diagnostic capabilities when facing different types of faults, affecting the model's adaptability and generalization capabilities.

[0081] The present invention proposes a dynamic weighting mechanism, which can automatically optimize the weight parameters according to data characteristics and fault categories by adaptively adjusting the weights of similarity loss and orthogonality loss. This dynamic weighting mechanism enables the model to be flexibly adjusted under different fault scenarios, improves the adaptive ability, and ensures the diagnostic accuracy and efficiency of the model in complex environments.

[0082] 3. Strong reliance on labeled data

[0083] Many traditional fault diagnosis methods rely on a large amount of labeled data for model training. However, labeled data is often difficult to obtain in practical applications, especially in the field of large-scale unit fault diagnosis. Due to the scarcity of fault events and the limited number of labeled samples, the application effect and promotion of fault diagnosis models are restricted.

[0084] The present invention introduces a self-supervised learning mechanism to effectively use unlabeled data for feature learning and model pre-training, reducing the reliance on a large amount of labeled data. Combined with contrastive learning and other technologies, the present invention can learn the normal working status of the equipment from unlabeled historical operation data and obtain better diagnostic performance with a small amount of labeled data. This progress greatly reduces the cost of data annotation and improves the generalization ability of the model and the feasibility of industrial application.

[0085] 4. Existing diagnostic efficiency is low

[0086] In traditional fault diagnosis methods, data processing efficiency is low, especially in scenarios of multi-source data fusion and large-scale data analysis, the response speed of the diagnostic model often cannot meet the needs of real-time monitoring and fault prediction. This leads to delayed detection of equipment failures and inability to respond and take emergency measures in a timely manner.

[0087] The present invention can efficiently process large-scale multi-source data through the combination of multi-view learning and self-supervised learning, and the dynamic weighting mechanism improves the computational efficiency and adaptability of the model. In an industrial environment, the present invention can quickly predict and diagnose faults, achieve real-time monitoring and rapid response to large-scale unit faults, and significantly shorten the diagnosis and response time.

[0088] In summary, the technical progress of the present invention in industrial applications is mainly reflected in the following aspects:

[0089] 1. Improved data fusion capabilities: Through multi-view representation learning, the problem of multi-source data fusion is effectively solved, and the accuracy of fault diagnosis is significantly improved.

[0090] 2. Enhanced adaptability: Through the dynamic weighting mechanism, the model can be flexibly adjusted according to different fault types, improving the robustness and adaptability of diagnosis.

[0091] 3. Reduce dependence on labeled data: The self-supervised learning mechanism reduces the dependence on a large amount of labeled data, improves the generalization ability of the model, and reduces the data labeling cost in industrial applications.

[0092] 4. Significantly improved diagnostic efficiency: The efficiency of the diagnostic system is optimized, enabling real-time fault monitoring and prediction, meeting the requirements of modern industry for rapid response.

[0093] These technological advances make the present invention have broad application prospects in the field of large-scale unit fault diagnosis, which can effectively improve the efficiency of equipment operation and maintenance, reduce the risk of fault downtime, and provide strong technical support for the development of industrial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 It is a flow chart of a fault diagnosis method for a large-scale unit with shared and specific representation learning provided by an embodiment of the present invention.

[0095] Figure 2 It is a structural block diagram of a fault diagnosis system for large-scale units with shared and specific representation learning provided by an embodiment of the present invention.

[0096] Figure 3 These are three different types of fault scene diagrams provided by embodiments of the present invention.

[0097] Figure 4 This is a scene diagram of a broken blade failure provided by an embodiment of the present invention.

[0098] Figure 5 4 are time domain waveform diagrams of four state types provided in an embodiment of the present invention.

[0099] Figure 6 It is a confusion matrix diagram of the present application method and the comparative method provided in the embodiment of the present invention.

[0100] Figure 7 It is a framework diagram of a WSMDRL fault diagnosis network model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0101] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0102] like Figure 1 As shown, a fault diagnosis method for a large-scale unit with shared and specific representation learning provided by an embodiment of the present invention includes the following steps:

[0103] S101, introduces a multi-view representation learning method and applies it to fault diagnosis of multi-point data of large units;

[0104] According to the original vibration signals collected by the sensors at the four measuring points located at the X-axis coupling end, the X-axis non-coupling end, the Y-axis coupling end, and the Y-axis non-coupling end, four-view data of X1, X2, Y1, and Y2 are constructed;

[0105] Each view data is input into the shared representation network W1 and the specific representation network W2 to obtain shared representations G1, G2, G3, G4 and specific representations S1, S2, S3, S4 respectively; weighted orthogonal loss and similarity loss are introduced; among them, the weighted orthogonal loss ensures the independence of shared information and specific information extracted from the same view by imposing orthogonality constraints, reduces the redundancy between features, and thus improves the ability to separate fault features; while the weighted similarity loss ensures that the shared information of different views maintains consistency in fault features through adversarial training, preventing the separation effect of shared information between different views from being overly dispersed;

[0106] The shared representation G is connected with the specific representations S1, S2, S3, and S4 to obtain the final representation R, and then the final representation is input into the classification network to obtain the fault diagnosis result;

[0107] S102, improve the dynamics of the weighting mechanism;

[0108] In multi-measurement point data, no matter it is view X1, X2 or Y1, Y2, they are all conceptualized as a sample space, and a single sample is represented as a point distributed in a high-dimensional space; on the one hand, subsets of different view data correspond to conditional representations of the same fault, and this representation is shared and consistent in all view data; therefore, the present invention constructs a shared network W1 to extract the shared vibration signal features in each view and represents it as cross-view shared information features G k On the other hand, in each view, some segments of the vibration signal represent fault information that is different from the information shared with other views. This part of information is the specific information expression of the vibration signal on the fault. Therefore, the present invention also constructs a specific network W2 to extract the specific information features of the vibration signal of each view, and represents it as the specific information features S1, S2, S3, S4 of the view.

[0109] If X m =[x1, x2, …, x n ] T represents all samples of the mth view, where n represents the total number of samples, x i Represents a sample; WSMDRL uses the shared network W1 to i Mapping is performed to obtain its shared information mapping and a weight α is set to reduce the redundant information between the four view data to obtain the shared information mapping G k for:

[0110]

[0111] Similarly, WSMDRL uses a specific network W2 to i Mapping is performed to obtain its specific information mapping as follows:

[0112]

[0113] In order to fully utilize the complementarity of multi-view data, each view pair x i The specific mappings are retained with the shared mapping G k Connected together, it is represented as:

[0114] R=[G k , S1,…,S4] (3)

[0115] Weighted multiple similarity constraints and orthogonality constraints:

[0116] Weighted Multiple Similarity Loss

[0117] Similarity loss can ensure the similarity between shared information of different views. WSMDRL draws on the adversarial similarity constraint from the generative adversarial network to design similarity loss and constructs a weighted multiple similarity loss function to learn the shared information between different views. Specifically, WSMDRL regards the shared network as a generator G. For the k-th view of the ,-th sample make Represents the shared information generated from this view, using the Sigmoid function S(x) = 1 / (1+e -x ) Compress the generated shared information to obtain the final generated shared information:

[0118] In addition, the present invention uses an M-classified classifier as the discriminator D to identify the distribution to which the shared information belongs; The probability from the kth view is:

[0119]

[0120] During the training phase, the discriminator D tries to find the source view of each shared information, that is, to maximize P i k ; Instead, the generator G tries to minimize this probability. Putting these two goals together, we get a maximum-minimization game:

[0121]

[0122] In formula (5), N is the total number of samples, is the true label;

[0123] Since the shared information generated by different views is redundant, a similarity factor λ1 is introduced to balance the similarity loss, and the final weighted similarity loss is:

[0124]

[0125] Weighted Orthogonal Loss:

[0126] The orthogonality constraint is a constraint designed to encourage or force the parameter matrix to remain orthogonal during the training process, which can decompose the shared information and the specific information. The present invention constructs a weighted orthogonality loss function to separate the shared information and the specific information from different views, making the meaning between them clearer and more explicit, while also avoiding their mutual contamination.

[0127] If the shared features and specific features of the i-th sample of view m are compressed using the sigmoid function, they are expressed as The present invention introduces L2 norm normalization and And the shared information is orthogonal to the specific information, and the orthogonal loss of the sample is obtained as:

[0128]

[0129] In addition, in order to balance the weight of each view, the present invention introduces an orthogonal factor μ for each view, and finally sums the weighted orthogonal losses to obtain the total weighted orthogonal loss:

[0130]

[0131] S103, classification network;

[0132] The classification network consists of three fully connected layers, an activation function, and a regularization term. First, Dropout is used to deal with the overfitting problem of the model. Assume that the Dropout function is dp(x), the loss probability p is 0.5, and ReLU is selected as the activation function. For each sample x in each view i According to formula (1) and formula (2), we can get the shared mapping and the specific mapping and connect them to get the feature representation R i Finally, R i Input the classification network to obtain the fault classification result. If w represents the weight parameter and b represents the bias parameter, Representative feature representation R i After the output of the regularization and activation function, the output of the fully connected layer is:

[0133]

[0134] In S102, the method introduces a dynamic weighting mechanism to optimize the adaptability of the model. Traditional weighting mechanisms usually use fixed weights to balance similarity loss and orthogonality loss. However, fixed weights cannot adapt to the feature differences of different fault data. Therefore, this method automatically adjusts the weight parameters of similarity and orthogonality constraints according to the data characteristics and fault categories, so that the model can better adapt to complex data distribution. The dynamic weight mechanism can adaptively adjust the various loss constraints of the model according to specific fault scenarios, thereby improving the diagnostic performance of the model under different fault conditions.

[0135] This fault diagnosis method combines multi-view representation learning, dynamic weighting mechanism and self-supervised learning to build an efficient and adaptable model architecture. Multi-view learning can fully capture the multi-dimensional characteristics of data, the dynamic weighting mechanism enables the model to adjust adaptively according to data characteristics, and self-supervised learning uses unlabeled data to enhance the model's feature learning ability, ultimately achieving accurate diagnosis and rapid response to large unit faults.

[0136] The multi-view representation learning method provided by the embodiment of the present invention:

[0137] The data of multiple measuring points are regarded as multiple views, and through the shared and specific representation network, the shared information and specific information between views can be effectively distinguished while extracting the data features of each view.

[0138] The embodiment of the present invention provides an improved dynamic weighting mechanism:

[0139] Based on multi-view representation learning, weighted multiple similarity loss and orthogonal loss are constructed to further optimize feature extraction and separation;

[0140] Similarity loss: ensures that the shared information extracted between views has high consistency and avoids the problem of excessive dispersion of shared information between different views;

[0141] Orthogonal loss: By introducing orthogonal constraints, the shared features and specific features of each view are ensured to be independent of each other, reducing redundant information and improving feature discrimination capabilities.

[0142] like Figure 2 As shown, a fault diagnosis system for a large-scale unit with shared and specific representation learning provided by an embodiment of the present invention includes:

[0143] Multi-view module, which is used to introduce multi-view representation learning methods and apply them to fault diagnosis of multi-measurement point data of large units;

[0144] The weighting module is used to weight the similarity loss and orthogonality loss based on fixed weight settings. On this basis, a dynamic weight adjustment mechanism is introduced to automatically adjust the weight parameters of similarity and orthogonality constraints according to different data characteristics and fault categories, further improving the model's adaptive ability;

[0145] Classification module, used for classification networks.

[0146] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the fault diagnosis method for large-scale fault units with shared and specific representation learning.

[0147] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the fault diagnosis method for a large-scale unit with shared and specific representation learning.

[0148] Another object of the present invention is to provide an information data processing terminal, which is used to implement the shared and specific representation learning fault large-scale unit fault diagnosis system.

[0149] The fault diagnosis method system of large-scale unit faults based on shared and specific representation learning includes:

[0150] The present invention is specifically implemented:

[0151] Two directions that the present invention can optimize

[0152] 1. Improve the dynamics of the weighting mechanism

[0153] The current weighted similarity loss and orthogonality loss are based on fixed weight settings. Although they can achieve feature separation between different views, in a dynamic and complex environment, the fixed weight settings cannot adapt to the diverse fault types and scenarios. Future optimization directions can introduce a dynamic weight adjustment mechanism based on this, automatically adjust the weight parameters of similarity and orthogonality constraints according to different data features and fault categories, and further improve the model's adaptive ability.

[0154] The present invention solves the shortcomings of existing fault diagnosis methods in processing multi-point data by introducing multi-view representation learning, shared and specific information separation mechanism, and weighted similarity and orthogonal constraints. Compared with traditional methods, the present invention demonstrates higher diagnostic accuracy and robustness under multi-source complex data, and provides a new technical path for multi-view fault diagnosis. At the same time, in the future, the applicability and scalability of the present invention can be further improved through the optimization direction of dynamic weighting mechanism and self-supervised learning.

[0155] like Figure 7 As shown in the figure, a large-scale unit fault diagnosis method based on multi-view deep representation learning is proposed. The specific steps are shown in the figure and the detailed description is as follows:

[0156] Step 1: The present invention constructs four-view data of X1, X2, Y1, and Y2 based on the original vibration signals collected by four measuring point sensors located at the X-axis coupling end, the X-axis non-coupling end, the Y-axis coupling end, and the Y-axis non-coupling end.

[0157] Step 2: Input each view data into the shared representation network W1 and the specific representation network W2 to obtain shared representations G1, G2, G3, G4 and specific representations S1, S2, S3, S4 respectively. At the same time, the present invention introduces weighted orthogonal loss and similarity loss (see 1.1.2 for details) to improve performance. Among them, the weighted orthogonal loss ensures the independence of shared information and specific information extracted from the same view by imposing orthogonality constraints, reduces the redundancy between features, and thus improves the ability to separate fault features. The weighted similarity loss ensures that the shared information of different views remains consistent in fault features through adversarial training, and prevents the separation effect of shared information between different views from being overly dispersed. Through this dual loss design, the model of the present invention can effectively distinguish and capture the common and unique parts of fault features in different views when extracting multi-view data features, thereby significantly improving the accuracy and robustness of fault diagnosis.

[0158] Step 3: Connect the shared representation G with the specific representations S1, S2, S3, and S4 to obtain the final representation R, and then input the final representation into the classification network to obtain the fault diagnosis result.

[0159] Experimental verification

[0160] 1. Experiment on Enterprise Large-Scale Unit Dataset

[0161] 1.1 Introduction to Enterprise Large Unit Dataset

[0162] Based on the vibration signals collected by sensors at four measuring points of the large-scale units of Shenyang Blower Enterprise, this application constructs a four-view dataset of a large petrochemical unit. The unit includes equipment such as steam turbines, compressors, and wind turbines, and its specific fault categories cover seven types of faults such as broken blades, oil film vortex, scaling, rotor imbalance, and normal categories. Compared with the data collected during laboratory operation, one of the significant advantages of the dataset constructed by this application is that the data collection is derived from the actual operation of the enterprise's large-scale units, which is more in line with the actual characteristics of industrial production data. In addition, the dataset constructed by this application contains four measuring points, and therefore has more comprehensive information.

[0163] In order to screen and diagnose faults of large units at an early stage, the 30 days before each fault occurs are used as the starting time for data download. However, the deadline for data download needs to be adjusted according to the damage degree and severity of different faults. For minor fault types such as scaling, rotor misalignment, and rotor rubbing that do not require immediate shutdown, in order to further understand and learn the subsequent changes and characteristics of their vibration signals, consider setting the last 3 days after each fault as the deadline for data download. Figure 3 (a), (b), and (c) show the on-site pictures of three types of faults: scaling, rotor misalignment, and rotor rubbing. In addition, for serious fault types that require immediate shutdown, such as broken blade faults, the shutdown time is used as the deadline for data download due to its severe impact and the great harm it brings. Figure 4 On-site pictures of broken blade failure are shown.

[0164] This application uses the vibration signals collected by sensors at four measuring points of the same unit as a view, stores the collected vibration signals as time domain signals, and finally obtains a four-view fault data set. Specifically, each view contains seven fault categories, including rotor rubbing fault, rotor imbalance fault, oil film vortex fault, broken blade fault, rotor misalignment fault, scaling fault, and surge fault, as well as eight normal categories. Table 3.2 describes the data set in detail, where the fault degree A represents good operating status, B represents the need for maintenance but no immediate shutdown, and C represents the need for immediate shutdown for maintenance. The total number of samples is the sum of the number of samples at the four measuring points.

[0165] Table 3.2 Dataset introduction

[0166]

[0167] In order to more intuitively display the vibration signals of large-scale equipment, Figure 5 The time domain signal schematic diagrams of normal class, oil film vortex fault, broken blade fault and rotor imbalance fault are shown respectively.

[0168] 1.2 Experimental setup for large-scale enterprise units

[0169] Table 3.3 shows the training sample and test sample settings for a certain view data. Although numerous sensors can capture a large amount of operating data of large petrochemical units, it is worth noting that the actual situation is that a very small part of the time the unit operates under abnormal conditions. In other words, only a small part of the huge data set is fault data. To this end, this application uses two different training and testing ratio sets (4:6 and 5:5) to evaluate the fault diagnosis performance of the proposed method under different training sample conditions. In each view data set, 3,400 samples are set for each specific category, and the total number of samples of eight categories in four views is 108,800.

[0170] Table 3.3 Experimental sample settings

[0171]

[0172]

[0173] (3) Results and analysis

[0174] In order to avoid the particularity and contingency of the fault diagnosis results of the proposed method, this application conducted ten experiments on each experimental method under different ratios of training set to test set, and used the average classification accuracy of the ten experimental results for analysis. The following Table 3.4 shows the classification accuracy and F1 score of the method proposed in this application and the comparison methods under different ratios of training set to test set.

[0175] Table 3.4 Experimental results

[0176]

[0177] Table 3.4 (Continued) Experimental results

[0178]

[0179]

[0180] like Figure 6 The confusion matrix can intuitively show the diagnostic accuracy and sensitivity of the method proposed in this application and the comparative method DRSN for each fault category. In general, both methods have shown good performance in diagnosing most types of faults. However, since the broken blade fault is instantaneous in actual operation, it is challenging to make an early prediction of this fault. Therefore, there is still room for further improvement in the fault diagnosis results of each method for the broken blade type. It is worth noting that compared with the comparative method, the method proposed in this application has improved the accuracy of oil film turbulence fault diagnosis by 10%, demonstrating its good diagnostic performance.

[0181] Example 1: Fault diagnosis of large generator sets based on multi-view representation learning

[0182] 1. Background: A large unit in a power plant has multiple measurement point data, including temperature, vibration, pressure and other monitoring parameters. Traditional fault diagnosis methods have difficulty extracting deep fault features from these multi-dimensional data, resulting in low accuracy of fault diagnosis. Therefore, the multi-view representation learning method is used to uniformly model these multi-measurement point data to improve the efficiency and accuracy of fault diagnosis.

[0183] 2. Steps:

[0184] Multi-view representation learning application: The data of multiple measurement points of the generator set are input as different perspectives to build a multi-view representation learning model. The model extracts and represents multi-dimensional features of temperature, vibration, pressure and other data to capture the correlation between various data sources.

[0185] Dynamic weighting mechanism: According to the characteristics of the data at different measuring points of the generator set, a dynamic weight adjustment mechanism is adopted to automatically adjust the weights of the similarity loss and orthogonality loss, so that the model can flexibly adapt to different feature distributions under different types of faults (such as overheating faults and mechanical faults).

[0186] Self-supervised learning: Use unlabeled historical operation data to pre-train the model and use contrastive learning to learn the normal operation mode of the equipment. The model after self-supervised learning can obtain deep features in advance and perform effective fault diagnosis even when there is less labeled data.

[0187] 3. Effect: By applying this method, the model can significantly improve the diagnostic accuracy of generator set faults, especially in the case of multiple data dimensions and few labeled samples, the accuracy of fault detection has increased by 20%. The system can also maintain high efficiency in a variety of fault scenarios, and the response speed has increased by 30%.

[0188] Example 2: Chemical Equipment Fault Diagnosis Based on Multi-View Representation Learning

[0189] 1. Background: A large reactor equipment in a chemical plant is equipped with a variety of sensors to monitor the operating status of the equipment, including temperature, flow rate, pressure, and vibration. The traditional single data source fault diagnosis method has difficulty processing these heterogeneous data sources, resulting in many failures to identify potential equipment failures in a timely manner, which brings risks to production. Therefore, a fault diagnosis method based on multi-view representation learning is used to analyze these multi-point data to improve the efficiency of fault identification.

[0190] 2. Steps:

[0191] Multi-view representation learning application: Feature extraction is performed on data sources such as temperature, flow rate, pressure and vibration of chemical equipment respectively, which are used as multi-view inputs. The multi-view representation learning model is used to uniformly perform data fusion and feature representation to fully capture the relationship between different parameters.

[0192] Dynamic weighting mechanism: A dynamic weight adjustment mechanism is introduced to dynamically weight the similarity and orthogonal constraints of the model according to the importance of different equipment parameters, ensuring that the model can adapt to different fault scenarios, especially temperature anomalies and flow rate anomalies in the reactor.

[0193] Self-supervised learning: Self-supervised learning is applied for pre-training on the historical unlabeled data of the equipment. Through comparative learning methods, the model learns the characteristic distribution of the equipment under normal conditions, laying the foundation for subsequent fine-tuning of fault diagnosis.

[0194] 3. Effect: By applying this method, the accuracy of chemical equipment fault diagnosis is improved by 15%, and an early warning can be issued 48 hours before the fault occurs, reducing the risk of equipment downtime due to faults. At the same time, through self-supervised learning, the dependence on labeled data is reduced, and the overall response speed of the diagnostic system is improved by 25%. It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. A person of ordinary skill in the art can understand that the above-mentioned devices and methods can be implemented using computer executable instructions and / or included in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as ultra-large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0195] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A fault diagnosis method for large-scale units based on shared and specific representation learning, characterized in that: The following steps are involved: Step 1: introduce the multi-view representation learning method and apply it to the fault diagnosis of multi-measurement point data of large units; Step 2, improve the dynamics of the weighting mechanism; Weighted similarity loss and orthogonality loss are based on fixed weight settings. On this basis, a dynamic weight adjustment mechanism is introduced to automatically adjust the weight parameters of similarity and orthogonality constraints according to different data characteristics and fault categories, further improving the model's adaptability. Step 3: Introduce a self-supervised learning mechanism; combine self-supervised learning or contrastive learning methods to perform feature learning and pre-training through unlabeled data.

2. The fault diagnosis method for large-scale faulty units based on shared and specific representation learning as claimed in claim 1, characterized in that: The multi-view representation learning method: The data of multiple measuring points are regarded as multiple views, and through the shared and specific representation network, the shared information and specific information between views can be effectively distinguished while extracting the data features of each view.

3. The fault diagnosis method for large-scale faulty units based on shared and specific representation learning as claimed in claim 1, characterized in that: The improved weighting mechanism is dynamic: Based on multi-view representation learning, weighted multiple similarity loss and orthogonal loss are constructed to further optimize feature extraction and separation; Similarity loss: ensures that the shared information extracted between views has high consistency and avoids the problem of excessive dispersion of shared information between different views; Orthogonal loss: By introducing orthogonal constraints, the shared features and specific features of each view are ensured to be independent of each other, reducing redundant information and improving feature discrimination capabilities.

4. The fault diagnosis method for large-scale faulty units based on shared and specific representation learning as claimed in claim 1, characterized in that: The multi-view representation learning method captures the potential correlation between data sources by extracting features and jointly modeling data from different measurement points, enhances the contribution of different data sources in fault feature representation, and performs fault diagnosis based on multi-view features to improve the comprehensiveness and accuracy of fault detection.

5. The fault diagnosis method for large-scale faulty units based on shared and specific representation learning as claimed in claim 1, characterized in that: The dynamic weight adjustment mechanism introduces an attention mechanism or a dynamic adjustment strategy based on gradient information to adjust the weights of similarity loss and orthogonal loss in real time according to the weight contribution of different data features and changes in fault modes, so as to enhance the adaptability and diagnostic performance of the model.

6. The fault diagnosis method for large-scale faulty units based on shared and specific representation learning as claimed in claim 1, characterized in that: The self-supervised learning mechanism adopts contrastive learning technology to learn the normal and abnormal modes of device operation through an enhancement strategy of unlabeled data, thereby improving the model's feature extraction ability in an unlabeled environment and pre-training subsequent labeled training to reduce dependence on a large amount of labeled data.

7. A fault diagnosis system for large-scale units using shared and specific representation learning that implements the fault diagnosis method for large-scale units using shared and specific representation learning as claimed in any one of claims 1 to 6, characterized in that: The shared and specific representation learning fault large unit fault diagnosis system comprises: Multi-view module, which is used to introduce multi-view representation learning methods and apply them to fault diagnosis of multi-measurement point data of large units; The weighting module is used to weight the similarity loss and orthogonality loss based on fixed weight settings. On this basis, a dynamic weight adjustment mechanism is introduced to automatically adjust the weight parameters of similarity and orthogonality constraints according to different data characteristics and fault categories, further improving the model's adaptive ability; Classification module, used to classify networks.

8. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the fault diagnosis method for a large-scale unit with shared and specific representation learning as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the fault diagnosis method for a large-scale unit with shared and specific representation learning as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the fault diagnosis system for large-scale units with shared and specific representation learning as described in claim 7.

Images