A typical fault incremental monitoring method for an induction motor

Abstract

The application relates to the technical field of induction motor fault monitoring, and discloses an induction motor typical fault incremental monitoring method, which collects vibration signals and current signals of a motor, carries out multi-source signal fusion through polar coordinate mapping, and generates a fusion image; a diagnostic network composed of a feature extractor and an extensible classifier is constructed; a knowledge selection loss, a feature transmission loss, a first classifier loss and a prediction loss are constructed, and optionally combined with a prototype offset regularization loss, a second classifier classification loss and a knowledge distillation loss, so that the diagnostic network is subjected to no-example incremental training. The application does not need to store historical fault samples, realizes fine-grained feature extraction based on a patch-level perception strategy of a Transformer, effectively reduces feature overlap of mechanical faults and electrical faults, suppresses prototype drift through a prototype correction and feature generation mechanism, alleviates catastrophic forgetting, guarantees data privacy, and realizes high-precision continuous monitoring of motor faults.

Description

Technical Field

[0001] This invention relates to the field of induction motor fault monitoring technology, and in particular to a method for monitoring the incremental faults of typical induction motors. Background Technology

[0002] Induction motors, as key components converting electrical energy into mechanical energy, are fundamental to industrial systems and play a vital role in the development of a low-carbon society. Due to long-term operation and unpredictable operating conditions, motors are inevitably prone to failure. Statistics show that nearly 40%-50% of motor failures are caused by mechanical faults, while nearly 30%-40% are electrical faults. These failures not only lead to downtime and significant economic losses, but in severe cases, can also cause catastrophic safety accidents. Therefore, motor fault diagnosis and health status assessment have significant engineering implications.

[0003] Most current general-purpose fault diagnosis frameworks are developed under the closed-set learning paradigm, where modules are designed for fixed classification tasks and can only handle static data streams. However, collecting samples of all possible fault categories in advance is impractical, especially for induction motors with different fault modes and long operating cycles. Therefore, this paradigm is not suitable for practical engineering applications.

[0004] Furthermore, existing methods always require historical data or a subset thereof as examples to maintain the acquired diagnostic knowledge when dealing with constantly emerging new faults. However, in practical industrial fault diagnosis applications, long-term storage of raw data is often impractical due to data privacy policies, storage limitations, and communication costs. Moreover, replaying historical samples introduces significant computational overhead, severely limiting the applicability of these methods in real-time diagnostic systems. Therefore, it is necessary to investigate sample-free incremental learning methods to achieve continuous fault diagnosis without accessing historical data. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problem that the existing technology requires the use of historical data for fault diagnosis, which leads to high computational overhead and low diagnostic efficiency.

[0006] To address the aforementioned technical problems, this invention provides a method for monitoring the incremental faults of an induction motor, comprising: In the initial stage, current signals and vibration signals of various fault types are collected and fused to obtain multi-source signal fusion matrices of various fault types, which are then mapped to polar coordinates to obtain polar coordinate feature maps. Based on the polar coordinate feature map and its corresponding real fault type label, the cascaded feature extractor and scalable classifier are trained to obtain the initial diagnostic model; and the initial prototypes corresponding to various fault types are obtained. In the incremental phase, the polar coordinate feature maps of each sample in the current task phase are input into the Transformer module and the initial diagnostic model, respectively. Based on the patch embedding mechanism of the Transformer module, class labels and multiple patches are obtained; based on the Euclidean distance between each patch and the class label, the task relevance of each patch is obtained as a weight; the L2 norm of the difference between the embedding representations of each patch in the current task stage and the previous task stage is weighted and summed, and then added to the L2 norm of the difference between the class labels in the current task stage and the previous task stage to construct the knowledge selection loss. The feature transfer loss is constructed based on the difference between the polar coordinate feature maps of samples in the current task stage and the feature representations extracted by the feature extractor. After Gaussian augmentation of the initial prototype of the sample from the previous task stage, a linear transformation is performed to obtain the pseudo prototype of the previous task stage. The cross-entropy loss between the output of the pseudo prototype of the previous task stage and the label of the pseudo prototype of the current task stage is calculated to construct the first classifier loss. Based on the cross-entropy loss of the predicted category and the actual fault type label of each sample in the current task stage, a prediction loss is constructed. The total loss function is obtained by weighted summing of knowledge selection loss, feature transfer loss, first classifier loss and prediction loss. The initial diagnostic model is trained based on the total loss function to obtain the target diagnostic model, which is then used to diagnose samples at the current task stage and obtain the predicted fault type.

[0007] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: The incremental monitoring method for typical faults in induction motors described in this invention, by constructing a multi-source data sensing and fusion mechanism, fully integrates the information advantages of vibration and current signals, overcoming the limitation of single-signal representation capabilities and significantly improving the ability to express fault features in complex electromechanical coupling systems. Based on the multi-source signal fusion matrix, a Transformer structure is further introduced and a patch-level information sensing strategy is designed, which can evaluate the correlation between each image patch and the current fault task, achieving differentiated attention to sensitive areas of vibration and current signals; it characterizes the sensitivity differences of different signals to various faults at the local fine-grained feature level, effectively reducing feature overlap between different fault categories, improving the model's fault discrimination ability, and enhancing the adaptability to new categories without sacrificing model stability.

[0008] This invention utilizes feature propagation loss to map current features to the feature space of the previous task stage via a linear module, suppressing class prototype drift caused by feature extractor updates. It implicitly retains old knowledge without storing historical samples, mitigating catastrophic forgetting. Simultaneously, the first classifier loss provides direct supervision signals for newly emerging fault categories, optimizing the classification decision boundary and supporting dynamic expansion of the classifier output dimension to adapt to an increasing number of fault categories. Furthermore, the prediction loss enables end-to-end diagnosis from the original signal to the fault category, with low forward computation, meeting the needs of real-time industrial monitoring, and providing probabilistic output to assist decision-making.

[0009] This invention employs prototype offset regularization loss to randomly partition a sample subset and calculate the mean squared error of the offset between the diagnostic model in the current task stage and the previous task stage. This achieves a stable offset pattern that decouples model learning from the task, enhancing feature generation capabilities and generalization performance. A second classifier loss, based on the cross-entropy of pseudo-features and labels from the previous task stage, strengthens the classifier's discrimination boundary for historically occurring categories, forming a dual-path supervision with the first classifier loss to alleviate class imbalance in incremental learning. Furthermore, a knowledge distillation loss directly constrains the consistency of the feature extractor's output between the current and previous task stages. This, along with feature transfer loss and prototype offset regularization loss, constitutes a multi-level feature constraint system, further improving the robustness of knowledge retention and ensuring the accuracy and robustness of motor fault diagnosis in example-less incremental learning scenarios. Attached Figure Description

[0010] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of the steps of the incremental monitoring method for typical faults in induction motors of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the incremental monitoring method for typical faults in induction motors according to the present invention. Detailed Implementation

[0011] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0012] Reference Figure 1 The flowchart shown is a step-by-step diagram of the typical fault incremental monitoring method for induction motors of the present invention; refer to... Figure 2 The diagram shown illustrates the principle of the incremental fault monitoring method for induction motors according to the present invention. Based on Figure 1 and Figure 2 As can be seen, the specific steps of the embodiments of the present invention are shown in S101 to S109.

[0013] S101: In the initial stage, current signals and vibration signals of various fault types are collected and fused to obtain multi-source signal fusion matrices of various fault types, which are then mapped to the polar coordinate system to obtain polar coordinate feature maps.

[0014] Specifically, set and A matrix representing vibration and current signals, forming a multi-source signal matrix. , is represented as: .

[0015] Assume there exists a one-dimensional time-domain signal , It is time Amplitude at that point It is the interval time After time Amplitude at that point. Polar coordinates. Through mapping points Obtained; and These are the clockwise and counterclockwise angles of the plane of symmetry, which determine the range of the signal point in the polar coordinate system. Based on these calculations... , and , is represented as: ; in, and These are time-domain signals. The maximum and minimum values; The interval time, , For gain angle, ; It is the first in the polar coordinate system Angles of a symmetrical plane It represents the number of planes of symmetry.

[0016] For multi-source signals, the above formula can be transformed into: ; in, Represents the first in the polar coordinate system Line number The normalized signal amplitude of the column, Represents the multi-source signal fusion matrix Middle Line number The signal value of the column, and These represent the multi-source signal fusion matrix, respectively. Middle The maximum and minimum values ​​of the column. Indicates the first The initial angle of each signal source in the polar coordinate system This represents the total number of columns in the multi-source signal fusion matrix. and Represent the first and second digits in the polar coordinate system, respectively. Line number The counterclockwise and clockwise angles corresponding to the elements in the column; Indicates delay After the nth time step, the nth time step in the polar coordinate system Line number The normalized signal amplitude of the column, This represents the gain angle. Furthermore, in the polar coordinate system, the... Mapping range of column data Restricted to .

[0017] This invention, by constructing a multi-source data sensing and fusion mechanism, fully integrates the information advantages of vibration signals and current signals, overcomes the problem of limited single-signal representation capabilities, and significantly improves the ability to express fault characteristics in complex electromechanical coupling systems.

[0018] S102: Based on the polar coordinate feature map and its corresponding real fault type label, train the cascaded feature extractor and scalable classifier to obtain the initial diagnostic model; and obtain the initial prototypes corresponding to various fault types.

[0019] In this embodiment, the feature extractor adopts a deep architecture with interleaved convolutional modules, residual modules, and Transformer modules, including a first convolutional module, a first residual module, a second residual module, a Transformer module, and an scalable fully connected layer connected in sequence.

[0020] Specifically, the acquisition of the feature representation extracted by the feature extractor from the polar coordinate feature map of the sample includes: Polar coordinate feature map of the sample The input feature extractor passes through a series of convolutional modules connected in the forward propagation direction. First residual module With the second residual module Obtain the first branch features ,include: , , ; Input the polar coordinate feature map of the sample into the Transformer module to obtain the second branch features. , is represented as: ; The features from the first branch and the second branch are weighted and fused to obtain the feature representation of the polar coordinate feature map of the sample extracted by the feature extractor. , is represented as: ; Features The weight, with a value range of . .

[0021] Specifically, the fused feature representation Input scalable linear classifier Fault categories obtained from initial phase samples , is represented as: The scalable linear classifier in this embodiment is implemented using a Softmax layer.

[0022] This embodiment is based on the task phase. The training set is constructed from each sample and its corresponding real fault type label. , is represented as: Sample set tag collection .

[0023] Specifically, referring to Table 1, the structural parameters of the diagnostic model in this embodiment are defined.

[0024] Table 1 Structural parameter constraints of the diagnostic model

[0025] S103: In the incremental phase, the polar coordinate feature maps of each sample in the current task phase are input into the Transformer module and the initial diagnostic model, respectively.

[0026] S104: Based on the patch embedding mechanism of the Transformer module, obtain class labels and multiple patches; based on the Euclidean distance between each patch and the class label, obtain the task relevance of each patch as a weight; after weighted summation of the L2 norm of the difference between the embedding representations corresponding to each patch in the current task stage and the previous task stage, add it to the L2 norm of the difference between the class labels in the current task stage and the previous task stage to construct the knowledge selection loss, expressed as: ; in, This indicates the total number of patches. ; Indicates the first The task relevance of a patch is expressed as follows: , ; Indicates the first The initial weights of each patch, Indicates the current task phase Middle The embedded representation of each patch Indicates the current task phase Class marker, Represents the L2 norm; To avoid the minimum value where the denominator is 0, in this embodiment... It is assigned the minimum value 1e-8; Indicates all The maximum value in.

[0027] Since motor faults are artificially categorized into mechanical and electrical faults, different types of signals exhibit varying sensitivities to these faults. Considering that fault information contained in different types of signals with the same criteria may significantly reduce the model's fault identification capability, this is particularly evident in example-free incremental learning models. To address this issue, this embodiment leverages the patch-level representation advantages of the Transformer and designs a weighted knowledge distillation method, primarily based on the specificity of multi-source signal fusion image samples related to motor faults. Specifically, in generalized image classification tasks, there is a contradictory relationship between image background and foreground; the background may have low relevance to the classification task, but excessive focus on the background can lead to poor model classification performance, while excessive focus on the image foreground may reduce the model's consistent representation capability. Therefore, this embodiment designs a patch-level information-aware strategy, defining a metric to evaluate the relevance of each patch to the current task, which determines the patch's weight in the current and old models. Specifically, the Euclidean distance between the unique class label in the transformer and each patch is measured to represent their relevance to the task. The knowledge selection loss constructed based on task relevance maintains a stable representation of key regions in the image while also possessing the flexibility of unified image representation.

[0028] Based on the multi-source signal fusion matrix, a Transformer structure is further introduced and a patch-level information perception strategy is designed. This strategy can evaluate the correlation between each image patch and the current fault task, enabling differentiated attention to sensitive areas of vibration and current signals. It can characterize the sensitivity differences of different signals to various faults at the level of local fine-grained features, effectively reduce feature overlap between different fault categories, improve the fault discrimination ability of the model, and enhance the adaptability to new categories without sacrificing model stability.

[0029] S105: Based on the difference between the polar coordinate feature maps of samples in the current task stage and the feature representations extracted by the feature extractor, construct the feature transfer loss, expressed as: ; in, This represents a linear transformation operation. and These represent the feature extractors for the previous task stage and the current task stage, respectively. This represents the polar coordinate feature map of the l-th sample in the current task phase s. This represents the L2 norm.

[0030] S106: After Gaussian augmentation of the initial prototype of the samples from the previous task stage, a linear transformation is performed to obtain the pseudo-prototype of the previous task stage; the cross-entropy loss between the output of the pseudo-prototype of the previous task stage and the label of the pseudo-prototype of the current task stage is calculated to construct the first classifier loss, including: The actual fault type label in the previous task phase is: After Gaussian augmentation and linear transformation, the labels of the actual fault types from the previous task phase are obtained from the initial prototype of the sample. pseudo-prototype , is represented as: ; Calculate the cross-entropy loss between the output of the pseudo-prototype from the previous task stage and the scalable classifier of the current task stage, and the pseudo-prototype label from the previous task stage, and construct the first classifier loss. , is represented as: ; in, This represents a linear transformation operation. This represents the Gaussian prototype augmentation operation based on the class's global radius. This indicates that the actual fault type label in the previous task phase was... The initial prototype corresponding to the sample; Represents cross-entropy loss, This represents the scalable classifier in the current task phase s, and the pseudo-prototype label of the previous task phase. Indicates the pseudo-prototype in the previous task phase The actual fault type label corresponding to the fault type.

[0031] Since updating the model on consecutive tasks leads to changes in the model representation space, and in example-free incremental learning models, it is impossible to prevent catastrophic forgetting through sample replay, we consider modifying the model's unified representation space through prototypes. This evolution causes the old class prototypes to drift, making the retained prototypes no longer representative; this phenomenon greatly weakens the prototypes' ability to accurately represent the features of old fault categories, ultimately leading to a decline in the model's classification performance. Therefore, in this embodiment, to suppress prototype drift, prototype correction is performed by constructing a first classifier loss. Specifically, to increase the diversity of prototype data, Gaussian prototype augmentation is performed based on the class global radius to prevent overfitting due to linear changes.

[0032] S107: Construct the prediction loss based on the cross-entropy loss between the predicted category and the actual fault type label of each sample in the current task stage. , is represented as: ; in, Represents cross-entropy loss, This represents the scalable classifier in the current task phase s. This represents the feature extractor in the current task phase s. This represents the set of samples in the current task phase s. , express The Middle Polar coordinate feature map of each sample, Indicates the current task phase The set of real fault type labels in the sample.

[0033] S108: The total loss function is obtained by weighted summing of the knowledge selection loss, feature transfer loss, first classifier loss, and prediction loss, as follows: .

[0034] S109: Train the initial diagnostic model based on the total loss function to obtain the target diagnostic model, diagnose the fault samples in the current task stage, and obtain the predicted fault type.

[0035] This invention achieves fine-grained, patch-level feature perception through knowledge selection loss, suppresses prototype drift through feature propagation loss, supervises new category learning through first classifier loss, and achieves end-to-end diagnosis through prediction loss. These four elements work synergistically to enable continuous incremental monitoring of motor faults without requiring the storage of historical samples, effectively balancing the model's plasticity and stability.

[0036] Based on the above embodiments, to make the prototype offset independent of the task used to generate pseudo-features, in this embodiment of the invention, the sample set of the current task stage is equally divided into two subsets. Based on the offset between the features and the initial prototype of the fault samples in the two subsets in the current task stage and the previous task stage, a prototype offset regularization loss is constructed and added to the total loss function; the prototype offset regularization loss... , is represented as: ; in, Indicates batch size, Indicates mean square error; Represents the first subset The i-th sample in the current task stage The feature prototype offset in the model is expressed as: ; Indicates the current task phase Feature extractor in Indicates the current task phase The polar coordinate feature map of the i-th sample in the first subset; This represents the true fault type label of the i-th fault sample in the current task phase s. Indicates the current task phase Real fault type tags The corresponding initial prototype; Indicates the second subset The Middle The samples were in the previous task phase. The feature prototype offset in the model is expressed as: ; Indicates the previous task phase Feature extractor in Indicates the current task phase The second subset Polar coordinate feature map of each sample; Indicates the current task phase The Middle The true fault type label for each sample Indicates the current task phase Real fault type tags The corresponding initial prototype; the first The initial prototype of the real fault type label , is represented as: ; Indicates the current task phase The Middle The total number of class samples, Indicates the current task phase The Middle The first class of sample set One sample, .

[0037] This invention enables the model to learn a stable offset pattern independent of the task by randomly dividing the sample subset and calculating the offset of the model in the current task stage and the previous task stage; based on the prototype correction and feature generation strategy, the prototype and feature space are regularized to reduce forgetting in the model evolution process, and can effectively monitor typical faults of induction motors continuously.

[0038] Based on the above embodiments, in this embodiment of the invention, pseudo-features are constructed based on the prototype offset and initial prototype of the current task stage; based on the output of the scalable classifier of the pseudo-features through the current task stage and the cross-entropy loss of the label corresponding to the pseudo-features, a second classifier classification loss is constructed and added to the total loss function; the second classifier classification loss... , represented as: ; in, Represents cross-entropy loss, Indicates the current task phase Scalable classifiers in The pseudo-features representing the current task phase are denoted as: ; This represents the initial set of prototypes for a historical task phase. Indicates the current task phase Feature extractor in Indicates the first stage of the current task s Polar coordinate feature map of each sample, Indicates the current task phase medium sample Corresponding actual fault type label The initial prototype; This represents the set of true fault type labels for samples during the historical mission phase.

[0039] Based on the above embodiments, when new samples appear and the model is trained on them, the feature extractor is constantly updated, leading to changes in the feature space distribution. To alleviate this problem, knowledge distillation technology is used for feature matching. Specifically, in this embodiment of the invention, a knowledge distillation loss is constructed based on the L2 norm of the difference between the feature extraction representations extracted by the feature extractors corresponding to the previous task stage and the current task stage, and added to the total loss function; the knowledge distillation loss... , represented as: ; in, Indicates the current task phase Feature extractor in Indicates the first stage of the current task s Polar coordinate feature map of each sample, This represents the L2 norm.

[0040] This invention further learns stable offset patterns through prototype offset regularization loss, strengthens old category memory through second classifier classification loss, and achieves direct feature-level constraints through knowledge distillation loss. These three additional technical features, together with the aforementioned knowledge selection loss, feature transfer loss, first classifier loss, and prediction loss, form a multi-layered and multi-faceted knowledge preservation mechanism, further improving the accuracy and robustness of motor fault diagnosis in no-example incremental learning scenarios.

[0041] Based on the above embodiments, the model constructed by optimizing all loss functions collaboratively in this embodiment of the invention is expressed as follows: ;in, , and These are the preset weight parameters.

[0042] Training the feature extractor and scalable linear classifier based on the total loss function also includes: adjusting the parameters of the feature extractor through backpropagation and using a stochastic gradient descent algorithm. and the parameters of the scalable linear classifier Perform iterative optimization, adjusting the parameters of the feature extractor. For the parameters of the scalable linear classifier All are recorded as Then the iterative optimization formula is: ;in, For the first Model parameters at the next iteration For the first Model parameters at the next iteration For learning rate, For loss parameters about parameters The gradient is then calculated. The test sample is then input into the pre-trained diagnostic model to obtain the fault category information of the test sample.

[0043] The incremental monitoring method for typical faults in induction motors proposed in this invention is an interpretable, sample-free incremental learning method based on multi-source information perception. It achieves continuous diagnosis of typical motor faults without storing historical samples, effectively meeting the practical needs of data privacy and storage constraints in industrial scenarios. Simultaneously, through prototype correction and feature generation mechanisms, historical knowledge is implicitly preserved at the feature space level, significantly mitigating the problem of catastrophic forgetting and ensuring good stability and consistency of the model across different incremental stages. This invention achieves an effective balance between model stability and adaptability while ensuring data security, and balances high-precision diagnostic performance with good interpretability, demonstrating promising engineering application prospects.

[0044] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0045] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0048] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for incremental monitoring of typical faults in an induction motor, characterized in that, include: In the initial stage, current signals and vibration signals of various fault types are collected and fused to obtain multi-source signal fusion matrices of various fault types, which are then mapped to polar coordinates to obtain polar coordinate feature maps. Based on the polar coordinate feature map and its corresponding real fault type label, the cascaded feature extractor and scalable classifier are trained to obtain the initial diagnostic model; and the initial prototypes corresponding to various fault types are obtained. In the incremental phase, the polar coordinate feature maps of each sample in the current task phase are input into the Transformer module and the initial diagnostic model, respectively. Based on the patch embedding mechanism of the Transformer module, class labels and multiple patches are obtained. The task relevance of each patch is used as a weight based on the Euclidean distance between each patch and the class label. A weighted sum of the L2 norms of the differences between the embedding representations of each patch in the current task stage and the previous task stage is then added to the L2 norm of the differences between the class labels in the current task stage and the previous task stage to construct the knowledge selection loss, expressed as: ;in, This indicates the total number of patches. ; Indicates the first The task relevance of a patch is expressed as follows: , ; Indicates the first The initial weights of each patch, Indicates the current task phase The Middle The embedded representation of each patch Indicates the current task phase Class marker, Represents the L2 norm. This indicates an attempt to avoid the minimum value where the denominator is 0. Indicates all The maximum value in; The feature transfer loss is constructed based on the difference between the polar coordinate feature maps of samples in the current task stage and the feature representations extracted by the feature extractor. After Gaussian augmentation of the initial prototype of the sample from the previous task stage, a linear transformation is performed to obtain the pseudo prototype of the previous task stage. The cross-entropy loss between the output of the pseudo prototype of the previous task stage and the label of the pseudo prototype of the current task stage is calculated to construct the first classifier loss. Based on the cross-entropy loss of the predicted category and the actual fault type label of each sample in the current task stage, a prediction loss is constructed. The total loss function is obtained by weighted summing of knowledge selection loss, feature transfer loss, first classifier loss and prediction loss. The initial diagnostic model is trained based on the total loss function to obtain the target diagnostic model, which is then used to diagnose samples at the current task stage and obtain the predicted fault type.

2. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, Constructing feature transit loss , is represented as: ； in, This represents a linear transformation operation. and These represent the feature extractors for the previous task stage and the current task stage, respectively. Indicates the first stage in the current task phase s l Polar coordinate feature map of each sample, This represents the L2 norm.

3. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, Constructing the loss of the first classifier includes: The actual fault type label in the previous task phase is: After Gaussian augmentation of the initial prototype of the sample, and followed by linear transformation, the true fault type labels from the previous task phase are obtained. pseudo-prototype , is represented as: ; Calculate the cross-entropy loss between the output of the pseudo-prototype from the previous task stage and the scalable classifier of the current task stage, and the pseudo-prototype label from the previous task stage, and construct the first classifier loss. , is represented as: ； in, This represents a linear transformation operation. This represents the Gaussian prototype augmentation operation based on the class's global radius. This indicates that the actual fault type label in the previous task phase was... The initial prototype corresponding to the sample; Represents cross-entropy loss, This represents the scalable classifier in the current task phase s, and the pseudo-prototype label of the previous task phase. Indicates the pseudo-prototype in the previous task phase The actual fault type label corresponding to the fault type.

4. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, Constructing Predictive Loss ,include: ； in, Represents cross-entropy loss, This represents the scalable classifier in the current task phase s. This represents the feature extractor in the current task phase s. This represents the set of samples in the current task phase s. Indicates the current task phase The set of real fault type labels in the sample.

5. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, Also includes: The sample set of the current task stage is equally divided into two subsets. Based on the offset between the features and the initial prototype of the samples in the two subsets in the current task stage and the previous task stage, a prototype offset regularization loss is constructed and added to the total loss function. The prototype offset regularization loss , is represented as: ； in, Indicates batch size, Indicates mean square error; Represents the first subset The i-th sample in the current task stage The feature prototype offset in the model is expressed as: ; Indicates the current task phase Feature extractor in Indicates the current task phase The polar coordinate feature map of the i-th sample in the first subset; This represents the true fault type label of the i-th sample in the current task phase s. Indicates the current task phase Real fault type tags The corresponding initial prototype; Indicates the second subset The Middle The samples were in the previous task phase. The feature prototype offset in the model is expressed as: ; Indicates the previous task phase Feature extractor in Indicates the current task phase The second subset Polar coordinate feature map of each sample; Indicates the current task phase The Middle The true fault type label for each sample Indicates the current task phase Real fault type tags The corresponding initial prototype; the first The initial prototype of the real fault type label , is represented as: ; Indicates the current task phase The Middle The total number of class samples, Indicates the current task phase The Middle The first class of sample set One sample, .

6. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, Also includes: Based on the prototype offset and initial prototype of the current task stage, construct pseudo-features; based on the output of the scalable classifier of the current task stage through the pseudo-features and the cross-entropy loss of the real fault type label of the fault type corresponding to the pseudo-features, construct the classification loss of the second classifier and add it to the total loss function. The classification loss of the second classifier , is represented as: ； in, Represents cross-entropy loss, Indicates the current task phase Scalable classifiers in The pseudo-features representing the current task phase are denoted as: ; This represents the initial set of prototypes for a historical task phase. Indicates the current task phase Feature extractor in Indicates the first stage of the current task phase s Polar coordinate feature map of each sample, Indicates the current task phase medium sample Corresponding actual fault type label The initial prototype; This represents the set of true fault type labels for samples during the historical mission phase.

7. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, Also includes: Based on the L2 norm of the difference between the feature extraction representations extracted by the feature extractors corresponding to the previous task stage and the current task stage, a knowledge distillation loss is constructed and added to the total loss function; The knowledge distillation loss , is represented as: ； in, Indicates the current task phase Feature extractor in Indicates the first stage of the current task phase s Polar coordinate feature map of each sample, This represents the L2 norm.

8. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, For each fault type, the corresponding current signal and vibration signal are collected and fused to obtain the multi-source signal fusion matrix corresponding to each fault type. This matrix is ​​then mapped to a polar coordinate system to obtain a polar coordinate feature map, represented as follows: ； in, Represents the first in the polar coordinate system Line 1 The normalized signal amplitude of the column, Represents the multi-source signal fusion matrix The Middle Line 1 The signal value of the column, and These represent the multi-source signal fusion matrix, respectively. The Middle The maximum and minimum values ​​of the column, Indicates the first The initial angle of each signal source in the polar coordinate system This represents the total number of columns in the multi-source signal fusion matrix. and Represent the first and second digits in the polar coordinate system, respectively. Line 1 The counterclockwise and clockwise angles corresponding to the elements in the column; Indicates delay After the nth time step, the nth time step in the polar coordinate system Line 1 The normalized signal amplitude of the column, This indicates the gain angle.

9. The method for monitoring the incremental faults of an induction motor according to claim 1, characterized in that, The acquisition of feature representations extracted from the polar coordinate feature map of a sample by a feature extractor includes: The polar coordinate feature map of the sample is input into the feature extractor, and after passing through the convolution module, the first residual module and the second residual module connected in series along the forward propagation direction, the first branch feature is obtained. Input the polar coordinate feature map of the sample into the Transformer module to obtain the second branch features; The first branch features and the second branch features are weighted and fused to obtain the feature representation of the polar coordinate feature map of the sample extracted by the feature extractor.

Images