Small sample incremental fault diagnosis method based on prototype adaptive network

Abstract

The invention provides a small sample incremental fault diagnosis method based on a prototype adaptive network, and the method comprises the steps: S1, obtaining and processing data at a basic stage, preprocessing the data, and dividing the data into a pseudo basic data set and a pseudo incremental data set; s2, training a pseudo feature extraction network, training the PMCNN through a pseudo basic data set in a pseudo basic stage, and providing feature embedding for the PAN; s3, PAN training: training PAN in a pseudo-increment stage through a pseudo-increment data set to update the preliminary prototype representation; s4, training a feature extraction network, and training the MCNN by using the complete basic data set; s5, feature extraction in an increment stage: extracting features of a new fault category through MCNN and solving prototype representation of the new fault category; and S6, updating and classifying the prototype representation, updating the prototype representation by using PAN, and performing fault classification based on cosine similarity. According to the method, the fault diagnosis accuracy can be remarkably improved under the small sample condition, meanwhile, the memory ability of the deep learning model for old categories is kept, and the method is suitable for complex industrial fault diagnosis scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment fault diagnosis, in particular to a fault prediction and health management technology based on deep learning, and specifically to a small sample incremental fault diagnosis method based on a prototype adaptive network. Background Art

[0002] The health status of rotating machinery equipment is crucial to the safe and stable operation of industrial systems. Due to problems such as frequent start-stop, overload operation, and component wear and failure, the operating reliability of the equipment faces severe challenges. In order to achieve real-time monitoring and prediction of the health status of equipment, current research has gradually focused on fault diagnosis methods based on incremental learning. However, traditional deep learning models are prone to the problem of "catastrophic forgetting" in incremental learning, that is, the model loses the ability to recognize the learned categories when learning new fault categories, resulting in a decrease in overall diagnostic accuracy. Therefore, it is particularly important to develop a small sample incremental fault diagnosis method that can take into account both new and old fault categories in an incremental learning environment.

[0003] In addition, traditional fault diagnosis methods usually rely on a large amount of full life cycle data for fault prediction and remaining life assessment. However, in practical applications, due to frequent replacement and difficulty in data collection, it is often difficult to obtain sufficient full life cycle data for prediction of rotating machinery equipment. Especially in the case of new fault categories, the acquisition of relevant samples is even more limited.

[0004] Therefore, in order to accurately identify new fault categories under small sample conditions and maintain the memory of old categories, a fault diagnosis method based on small sample incremental learning is urgently needed to improve the prediction accuracy and diagnostic robustness of the equipment. Summary of the invention

[0005] In view of this, the present invention provides a small sample incremental fault diagnosis method based on a prototype adaptive network to at least solve the above problems.

[0006] The present invention provides a small sample incremental fault diagnosis method based on a prototype adaptive network, which includes a basic stage and an incremental stage. The basic stage includes a pseudo-basic stage and a pseudo-incremental stage, and specifically includes the following steps:

[0007] S1. Data acquisition and processing in the basic stage: In the basic stage, the basic data set containing multiple fault types is preprocessed, and the basic data set is divided into a pseudo basic data set and a pseudo incremental data set. The pseudo basic data set and the pseudo incremental data set are used for training in the pseudo basic stage and the pseudo incremental stage.

[0008] S2, pseudo feature extraction network training: in the pseudo basic stage, the pseudo multi-scale convolutional neural network PMCNN is trained through the pseudo basic data set, and the parameters of the pseudo multi-scale convolutional neural network PMCNN are frozen; S3, prototype adaptive network PAN training: in the pseudo incremental stage, the prototype adaptive network PAN is trained through the pseudo incremental data set and the pseudo multi-scale convolutional neural network PMCNN, and the prototype representation of the fault category is expanded through small samples; S4, feature extraction network training: the multi-scale convolutional neural network MCNN is trained through the complete basic data set, and the parameters of the multi-scale convolutional neural network MCNN are frozen to extract more comprehensive fault features; S5, feature extraction in the incremental stage: in the incremental stage, small sample data of the new fault category is obtained, and the fault features of the new fault category are extracted through the trained multi-scale convolutional neural network MCNN, and the prototype representation of the new fault category is solved; S6, prototype representation update and classification: the prototype representation of the new fault category is updated through the trained prototype adaptive network PAN, and the fault classification is performed by calculating the cosine similarity between the query sample of the new fault category and the prototype representation of the new fault category.

[0009] In one implementation, in step S1, the acquired basic data set containing multiple fault types is preprocessed, and the basic data set is divided into a pseudo basic data set and a pseudo incremental data set, including: performing Fourier transform on the basic data set to obtain a frequency domain signal, and reshaping the one-dimensional frequency domain signal of the frequency domain signal into a two-dimensional frequency domain signal; dividing the basic data set into a pseudo basic data set and a pseudo incremental data set in a 1:1 ratio, wherein the pseudo basic data set is used to train a pseudo multi-scale convolutional neural network PMCNN, and the pseudo incremental data set is used to train a prototype adaptive network PAN.

[0010] In another implementation, the pseudo multi-scale convolutional neural network PMCNN in step S2 and the multi-scale convolutional neural network MCNN in step S4 both adopt a multi-scale convolutional structure to simultaneously extract fault features of different scales.

[0011] In another implementation, the training process of the prototype adaptive network PAN in step S3 is as follows: a self-attention mechanism is introduced, and the prototype representation is dynamically updated through the self-attention mechanism, wherein, in the pseudo-incremental stage, the prototype adaptive network PAN is trained using the pseudo-incremental data set, the pseudo-incremental data is divided into a support set and a query set, and the support set and the query set are respectively input into the pseudo multi-scale convolutional neural network PMCNN to obtain the support set embedding E s and queryset embedding E q ; Embed the support set E s Calculate the mean and get the new category prototype representation w in the pseudo-incremental stage new , input the pseudo basic data into the pseudo multi-scale convolutional neural network PMCNN to obtain the old category prototype representation wold ; The new category prototype is represented by w old and the old category prototype representation w new The replication expansion dimensions are: and Where L is the number of samples in the query set, N old is the number of categories in the pseudo-base stage, N new is the number of categories in the pseudo-incremental stage, D is the number of dimensions of each embedding; the new category prototype is represented by w old , the old category prototype represents w new , query set embedding E q The prototype adaptive network PAN is trained as input. The principle of the prototype adaptive network PAN dynamically updating the prototype representation through the self-attention mechanism is as follows:

[0012]

[0013] Q=W Q X,K=W K X,V=W V X

[0014] Where X is the input vector, X={w old , w new , E q}, Q is the query vector, K is the key vector, V is the value vector, Q, K, V are respectively passed through three trainable linear layers W Q (·),W K (·) and W V (·) get, d k is the dimension of the key vector K, W Q is the weight matrix that transforms the input vector into the query vector, W K is the weight matrix that converts the input vector into the key vector, W V is the weight matrix that converts the input vector into a value vector;

[0015] The input vector X is updated by the following formula to obtain the updated X′:

[0016] X′=Ψ(Attention(Q,K,V))+X

[0017] Among them, Ψ(·) is a trainable linear layer, and pseudo incremental data is used to simulate the incremental stage training W Q (·),W K (·),W V (·) and Ψ(·) four linear layers. The trained prototype adaptive network PAN is obtained by multiple training of the four linear layers. The trained prototype adaptive network PAN can adaptively update the prototype representation of the new fault category according to the new and old prototype representations and feature embedding.

[0018] In another implementation, in step S6, fault classification is performed by calculating the cosine similarity between the query sample of the new fault category and the prototype representation of the new fault category, and the formula is as follows:

[0019]

[0020] Among them, w i represents the prototype of the i-th category, q represents the feature vector of the query sample, and the category with the highest similarity is used as the fault classification result, so as to achieve accurate diagnosis of new fault categories.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] (1) The present invention extracts fault features of different scales through a multi-scale convolutional neural network (MCNN) to effectively capture the feature information of different levels of industrial equipment in a complex environment and improve the adaptability of the deep learning model to various working conditions. After training, the MCNN will freeze its parameters, providing a stable feature extraction basis for the subsequent incremental learning stage, thereby avoiding the forgetting of old category knowledge when learning new categories, and improving the diagnostic accuracy of the model and the robustness of incremental learning.

[0023] (2) The present invention artificially divides the pseudo-basic stage and the pseudo-incremental stage to simulate in advance the learning order of different fault categories under small sample conditions. The pseudo-incremental stage training PAN simulates the gradual expansion of the category prototype representation, thereby improving the adaptability and generalization ability of the model under different working conditions and when new categories appear. The prototype adaptive network can efficiently learn and update the prototype representation of the fault category when the new category samples are limited, reducing the model's dependence on large amounts of data.

[0024] (3) Compared with the existing fault diagnosis methods, the present invention generates clear fault category prototypes by comparing fault samples with fault category prototypes, so that non-professionals can also intuitively understand the diagnosis results. At the same time, based on the similarity calculation method of the prototype, the model can effectively classify new faults while keeping the memory of the old categories. Professionals can timely identify potential faults and make maintenance decisions based on the trend of prototype changes, thereby effectively ensuring the operational stability of the system and reducing the possibility of faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the present invention;

[0026] Figure 2 It is a feature extraction network structure diagram of the present invention;

[0027] Figure 3This is a diagram of the prototype adaptive network structure of the present invention;

[0028] Figure 4 This is the confusion matrix result of each method in the CRWU dataset of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] See also Figure 1 The present invention provides a small sample incremental fault diagnosis method based on a prototype adaptive network, which includes a basic stage and an incremental stage. The basic stage includes a pseudo-basic stage and a pseudo-incremental stage. The main steps are as follows:

[0031] S1. Data acquisition and processing in the basic stage: In the basic stage, the basic data set containing multiple fault types is preprocessed, and the basic data set is divided into a pseudo basic data set and a pseudo incremental data set. The pseudo basic data set and the pseudo incremental data set are used for training in the pseudo basic stage and the pseudo incremental stage.

[0032] S2, pseudo feature extraction network training: In the pseudo base stage, the pseudo multi-scale convolutional neural network PMCNN is trained through the pseudo base data set, and the parameters of the pseudo multi-scale convolutional neural network PMCNN are frozen;

[0033] S3, prototype adaptive network PAN training: In the pseudo-incremental stage, the prototype adaptive network PAN is trained through the pseudo-incremental dataset and the pseudo multi-scale convolutional neural network PMCNN, and the prototype representation of the fault category is expanded through small samples;

[0034] S4, feature extraction network training: train the multi-scale convolutional neural network MCNN through the complete basic data set, and freeze the parameters of the multi-scale convolutional neural network MCNN to extract more comprehensive fault features;

[0035] S5. Feature extraction in the incremental stage: In the incremental stage, small sample data of new fault categories are obtained, and the fault features of the new fault categories are extracted through the trained multi-scale convolutional neural network MCNN, and the prototype representation of the new fault categories is solved;

[0036] S6. Prototype representation update and classification: The prototype representation of the new fault category is updated through the trained prototype adaptive network PAN, and the fault classification is performed by calculating the cosine similarity between the query sample of the new fault category and the prototype representation of the new fault category.

[0037] It should be understood that the full name of the pseudo-multiscale convolutional neural network PMCNN (Pseudo-Multiscale Convolutional Neural Network, PMCNN), the full name of the prototype adaptive network PAN (Prototype Adaptive Network, PAN), and the full name of the multiscale convolutional neural network MCNN (Multiscale Convolutional Neural Network, MCNN).

[0038] Optionally, in step S1, the acquired basic data set containing multiple fault types is preprocessed, and the basic data set is divided into a pseudo basic data set and a pseudo incremental data set, including: performing Fourier transform on the basic data set to obtain a frequency domain signal, and reshaping the one-dimensional frequency domain signal of the frequency domain signal into a two-dimensional frequency domain signal; dividing the basic data set into a pseudo basic data set and a pseudo incremental data set in a 1:1 ratio, wherein the pseudo basic data set is used to train a pseudo multi-scale convolutional neural network PMCNN, and the pseudo incremental data set is used to train a prototype adaptive network PAN.

[0039] It should be understood that performing Fourier transform on the basic data set to obtain the frequency domain signal and reshaping the one-dimensional frequency domain signal into a two-dimensional frequency domain signal can more effectively extract information in different frequency bands and improve the expressiveness of fault characteristics.

[0040] Optionally, the pseudo multi-scale convolutional neural network PMCNN in step S2 and the multi-scale convolutional neural network MCNN in step S4 both adopt a multi-scale convolutional structure to simultaneously extract fault features of different scales.

[0041] Specifically, the two networks can be configured with convolution kernels of different sizes, such as 1×1, 3×3, and 5×5, to extract fault features of different scales. Through multi-scale convolution, the model can capture more comprehensive frequency and time domain features, which helps to adapt to the changes in fault features under different working conditions. In addition, parameters are frozen after training to ensure stability in the incremental stage.

[0042] It should be understood that step S2 of pseudo feature extraction network training is to train the pseudo multi-scale convolutional neural network PMCNN, and the training process of the pseudo multi-scale convolutional neural network PMCNN is:

[0043] In the pseudo feature extraction network, multi-scale convolution operations are introduced, and convolution kernels of different sizes (1×1, 3×3, and 5×5) are used to extract fault features of different scales, ensuring the deep learning model's sensitivity to features and robustness to environmental changes, thereby improving the accuracy and robustness of fault feature extraction.

[0044] It should also be understood that step S4 of feature extraction network training is to train the multi-scale convolutional neural network MCNN, and the training process of the multi-scale convolutional neural network MCNN is:

[0045] In the feature extraction network, multi-scale convolution operation is introduced, and convolution kernels of different sizes (1×1, 3×3 and 5×5) are used to extract fault features of different scales and reliable fault features. The model parameters are frozen after training to ensure the stability of the network in the feature extraction of new fault samples in the incremental stage and keep it consistent with the pseudo feature extraction network structure.

[0046] By adopting the pseudo multi-scale convolutional neural network PMCNN in the pseudo feature extraction network training stage and the multi-scale convolutional neural network MCNN in the feature extraction network training stage, the feature extraction effects of the real stage and the pseudo stage can be kept close.

[0047] This multi-scale convolution method can effectively extract feature information at different levels, improve the model's generalization ability and recognition accuracy for small sample data, and thus enhance its adaptability to new fault categories.

[0048] Optionally, the training process of the prototype adaptive network PAN in step S3 is: introducing a self-attention mechanism, dynamically updating the prototype representation through the self-attention mechanism, wherein, in the pseudo-incremental stage, the prototype adaptive network PAN is trained using the pseudo-incremental data set, the pseudo-incremental data is divided into a support set and a query set, and the support set and the query set are respectively input into the pseudo multi-scale convolutional neural network PMCNN to obtain the support set embedding E s and queryset embedding E q ; Embed the support set E s Calculate the mean and get the new category prototype representation w in the pseudo-incremental stage new , input the pseudo basic data into the pseudo multi-scale convolutional neural network PMCNN to obtain the old category prototype representation w old ; The new category prototype is represented by w old and the old category prototype representation w new The replication expansion dimensions are: and Where L is the number of samples in the query set, N old is the number of categories in the pseudo-base stage, N new is the number of categories in the pseudo-incremental stage, D is the number of dimensions of each embedding; the new category prototype is represented by w old , the old category prototype represents w new , query set embedding E q The prototype adaptive network PAN is trained as input. The principle of the prototype adaptive network PAN dynamically updating the prototype representation through the self-attention mechanism is as follows:

[0049]

[0050] Q=W Q X,K=W K X, V = W V X

[0051] Where x is the input vector, X = {w old , w new , E q}, Q is the query vector, K is the key vector, V is the value vector, Q, K, V are respectively passed through three trainable linear layers W Q (·),W K (·) and W V (·) get, d k is the dimension of the key vector K, W Q is the weight matrix that transforms the input vector into the query vector, W K is the weight matrix that converts the input vector into the key vector, W V is the weight matrix that converts the input vector into a value vector;

[0052] The input vector X is updated by the following formula to obtain the updated X′:

[0053] X′=Ψ(Attention(Q,K,V))+X

[0054] Among them, Ψ(·) is a trainable linear layer, and the updated X′ is obtained, X′={w′ old , w′ new , E′ q}, and then split X′ to get the updated prototype representation W′={w′ old , w′ new} and the updated embedding E′ q The prototype adaptive network PAN is obtained by training four linear layers multiple times.

[0055] Specifically, pseudo incremental data is used to simulate the incremental phase training W Q (·),W K (·),W V (·) and Ψ(·) four linear layers. The trained prototype adaptive network PAN is obtained by multiple training of the four linear layers. The trained prototype adaptive network PAN can adaptively update the prototype representation of the new fault category according to the new and old prototype representations and feature embedding.

[0056] Optionally, in step S6, fault classification is performed by calculating the cosine similarity between the query sample of the new fault category and the prototype representation of the new fault category, and the formula is as follows:

[0057]

[0058] Among them, w i represents the prototype of the i-th category, q represents the feature vector of the query sample, and the category with the highest similarity is used as the fault classification result, so as to achieve accurate diagnosis of new fault categories.

[0059] It should be understood that the higher the cosine similarity, the closer the query sample is to the fault category. The model predicts the fault category of the query sample based on the prototype with the highest similarity, thereby realizing the identification of new faults in the incremental stage.

[0060] It should be understood that the method of the present invention first proposes PAN and combines it with a multi-scale convolutional neural network for incremental fault diagnosis of industrial equipment. The method extracts multi-scale features through MCNN, which can fully capture and adapt to fault features under different working conditions, and ensure the stability of feature extraction by freezing MCNN parameters. Combined with the self-attention mechanism of PAN, the present invention dynamically updates the prototype representation of fault categories in the incremental stage, so that the model can achieve efficient learning and recognition of new fault categories under small sample conditions, effectively improving the adaptability and diagnostic robustness of the model.

[0061] In addition, the present invention also classifies new fault categories by cosine similarity, so that the model can complete incremental diagnosis without forgetting old categories, which helps professionals monitor the changing trends of fault prototypes, discover potential problems in advance and make maintenance decisions, thereby effectively ensuring the operational stability of the system and reducing the risk of equipment failure.

[0062] Specifically, the scheme of the present invention is further described according to the following examples:

[0063] Replay Figure 1 The present invention includes two parts: basic stage and incremental stage, and the main steps are as follows:

[0064] S1. Data acquisition and processing in the basic stage: The present invention selects a fault data set from the CWRU bearing data set for training the small sample incremental fault diagnosis method. The data set is collected by Case Western Reserve University and contains bearing data under various fault states. It is widely used in the field of equipment fault diagnosis and state prediction. The experiment uses data collected by the drive end accelerometer (DE) running at a speed of 1797 rpm, as well as the normal operation data of the equipment to form a data set. The data used include three typical fault types with different fault diameters (0.007 inches, 0.014 inches and 0.021 inches): inner race fault (Inner Race Fault, IRF), rolling element fault (Ball Fault, BF) and outer race fault (Outer Race Fault, ORF), and the data sampling frequency is 48kHz. The specific information is shown in Table 1.

[0065] Table 1 Introduction to the CRWU bearing dataset

[0066]

[0067] The data set used in the basic stage is the data of stage 0 in Table 1, and the incremental stage includes three stages with gradually increasing faults, corresponding to stages 1, 2, and 3 in Table 1.

[0068] To improve the effect of feature extraction, the data preprocessing process first performs Fourier transform on the original data to convert the one-dimensional time domain data into frequency domain data, and then reshapes the one-dimensional data into a two-dimensional format to better capture the spatial distribution and frequency characteristics of the fault features.

[0069] The dataset of the basic stage consists of normal data and three types of fault data of 0.007 inches. In order to meet the training requirements of the pseudo-basic stage and pseudo-incremental stage, the dataset is further divided: normal data and 0.007-inch inner race fault (IRF) are used for the dataset of the pseudo-basic stage; 0.007-inch rolling element fault (BF) is used for the dataset of pseudo-incremental stage 1; 0.007-inch outer race fault (ORF) is used for the dataset of pseudo-incremental stage 2.

[0070] S2. Pseudo feature extraction network training: In the pseudo basic stage, the present invention uses the pseudo basic data set to train PMCNN. The network structure of PMCNN adopts multi-scale convolution to enhance the extraction effect of fault features. The network structure of the feature extraction network and the size of the multi-scale convolution kernel are as follows: Figure 2 shown.

[0071] The specific training of the pseudo feature extraction network, PMCNN, is as follows: the training batch size is 128, the number of training times is 50, and the learning rate is 0.05. The parameters can capture both fine-grained and coarse-grained fault feature information at different scales through 1×1, 3×3 and 5×5 convolution kernels. After the PMCNN network training is completed, the parameters are frozen to ensure stable feature embedding for the fault category in the training of PAN.

[0072] S3. PAN training: In the pseudo-incremental stage, the present invention uses a pseudo-incremental dataset to train PAN, which aims to expand the prototype representation of fault categories through adaptive updates.

[0073] Specifically, the pseudo-incremental data is divided into a support set and a query set, and the support set embedding E can be obtained through the feature extraction network, namely PAN. s and queryset embedding E q . Let the support set be embedded in E s Taking the average value can get the prototype representation w of the new category in the pseudo-incremental stage new , the old category prototype representation w can be obtained through the original network old , w old and w new The replication expansion dimensions are: and Where L is the number of samples in the query set, N old is the number of categories in the pseudo-base stage, N new is the number of categories in the pseudo-incremental phase, D is the number of dimensions of each embedding, and w old 、w new 、E q As input, we train PAN. The training process is as follows Figure 3 .

[0074] The input of the model is Through three trainable linear layers W Q (·),W K (·) and W V (·) to obtain the query vector Q, key vector K and value vector V respectively, and then pass them through a trainable linear layer Ψ(·) to get the updated X′:

[0075]

[0076]

[0077] We get X′={w′ old , w′ new , E′ q}, then split X′ to get the updated prototype representation W′={w′ old , w′new} and the updated embedding E′ q .

[0078] Among them, W Q (·),W K (·),W V (·) and Ψ(·) are trainable, and the specific training parameters are: training batch size is 128, number of training times is 50, and learning rate is 0.01.

[0079] S4. Feature extraction network training: In the last step of the basic stage, the MCNN is trained using the complete basic dataset. The network structure and convolution kernel size are as follows Figure 2 As shown in Figure 2, compared with PMCNN, this network uses more data and can further optimize multi-scale feature extraction to obtain more comprehensive fault features. After training, the MCNN parameters are frozen for feature extraction of new fault categories in the incremental stage.

[0080] The specific training parameters are: training batch size is 128, training times is 50 times, and learning rate is 0.05.

[0081] S5. Feature extraction in the incremental stage: In the incremental stage, small sample data of new fault categories are obtained and input into the trained MCNN for feature extraction. MCNN can ensure the comprehensiveness and consistency of extracted features, thereby providing high-quality feature embedding for prototype generation of new fault categories. After obtaining the feature embedding of a certain fault type, the prototype representation of the fault type can be obtained by averaging the feature embedding of the fault type.

[0082] 6. Prototype representation update and classification: Use the trained PAN network to update the prototype representation of the new fault type and the original fault type in the incremental stage.

[0083] The query samples of the new fault category are classified by cosine similarity calculation. The specific calculation formula is as follows:

[0084]

[0085] Among them, w i represents the prototype of the i-th type of fault, q represents the query sample feature embedding, and the category with the highest similarity is used as the basis for judging the fault category, thereby achieving accurate identification of new fault types in the incremental stage.

[0086] Based on the above steps, the method of the present invention can effectively maintain the memory of old categories in the incremental learning scenario, and accurately identify new fault categories using only a small number of samples, and has high application value and robustness.

[0087] The comparison of the prediction results of the method MCNN-PAN of the present invention and fine-tuning FT, iCARL, and LLB is shown in Table 2.

[0088] Table 2 Comparison of performance indicators between the proposed method and small sample incremental learning

[0089]

[0090] The evaluation indicators used in the present invention process the accuracy of each stage, as well as the average accuracy AA and the performance degradation rate PD, which are defined as:

[0091]

[0092] PD=A 0 -A I-1

[0093] Among them, A i It represents the accuracy of stage i. The higher the AA or the smaller the PD, the better the performance of the method.

[0094] It can be seen from Table 2 that all methods have good diagnostic accuracy in the basic stage. As the stage increases, the diagnostic accuracy will continue to decrease. The diagnostic accuracy of the method proposed in the present invention is better than other small sample incremental learning methods. Therefore, the scheme of the present invention has high diagnostic accuracy.

[0095] Figure 4 The comparison results of the confusion matrix between the method of the present invention and other methods are shown. The method of the present invention adopts MCNN and introduces PAN. MCNN can extract features of different scales and effectively capture the diversity and complexity of fault features, thereby improving the comprehensiveness of feature extraction. PAN adaptively adjusts the prototype representation so that the model can not only maintain a stable memory of old knowledge in the incremental learning stage, but also flexibly update the prototype representation to adapt to the changes of new categories. This improvement enables the model to achieve a better balance between stability and plasticity, so that it performs better in identifying new fault categories.

[0096] Based on this, the present invention uses multiple fault data in the basic stage to train MCNN and PAN offline to learn the representation and evolution rules of various fault features; and applies the trained MCNN and PAN to real-time fault data in the incremental stage to dynamically update the prototype representation of the fault category and achieve accurate diagnosis of new fault categories. As an incremental fault diagnosis scheme, the method of the present invention can keep the memory of old categories under small sample conditions and identify new faults in a timely manner, thereby helping staff to quickly perceive changes in equipment status, provide support for maintenance decisions, significantly reduce fault risks, and has high engineering application value and practical significance.

[0097] It should be noted that the present invention may be a method, a system, an apparatus and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0098] Computer readable storage medium can be a tangible device that can hold and store instructions used by an instruction execution device. Computer readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, for example, a punch card or a convex structure in a groove on which instructions are stored, and any suitable combination thereof. The computer readable storage medium used here is not interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated by a waveguide or other transmission medium (for example, a light pulse by an optical fiber cable), or an electrical signal transmitted by a wire.

[0099] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0100] The computer program instructions for performing the operation of the present invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, Python, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed entirely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be personalized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present invention.

[0101] Various aspects of the present invention are described herein with reference to the flow charts and / or block diagrams of the methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each box of the flow chart and / or block diagram and the combination of each box in the flow chart and / or block diagram can be implemented by computer-readable program instructions.

[0102] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0103] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0104] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a part of a module, a program segment or an instruction, and a part of the module, a program segment or an instruction contains one or more executable instructions for realizing the specified logical function. In some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions. It is well known to those skilled in the art that it is equivalent to implement it by hardware, implement it by software, and implement it by combining software and hardware.

[0105] Embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the marketplace, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein. The scope of the present invention is defined by the appended claims.

Claims

1. A small sample incremental fault diagnosis method based on a prototype adaptive network, comprising a basic stage and an incremental stage, wherein the basic stage comprises a pseudo-basic stage and a pseudo-incremental stage, and is characterized in that: The specific steps include: S1. Data acquisition and processing in the basic stage: In the basic stage, the basic data set containing multiple fault types is preprocessed, and the basic data set is divided into a pseudo basic data set and a pseudo incremental data set. The pseudo basic data set and the pseudo incremental data set are used for training in the pseudo basic stage and the pseudo incremental stage. S2, pseudo feature extraction network training: In the pseudo base stage, the pseudo multi-scale convolutional neural network PMCNN is trained through the pseudo base data set, and the parameters of the pseudo multi-scale convolutional neural network PMCNN are frozen; S3, prototype adaptive network PAN training: In the pseudo-incremental stage, the prototype adaptive network PAN is trained through the pseudo-incremental dataset and the pseudo multi-scale convolutional neural network PMCNN, and the prototype representation of the fault category is expanded through small samples; S4, feature extraction network training: train the multi-scale convolutional neural network MCNN through the complete basic data set, and freeze the parameters of the multi-scale convolutional neural network MCNN to extract more comprehensive fault features; S5. Feature extraction in the incremental stage: In the incremental stage, small sample data of new fault categories are obtained, and the fault features of the new fault categories are extracted through the trained multi-scale convolutional neural network MCNN, and the prototype representation of the new fault categories is solved; S6. Prototype representation update and classification: The prototype representation of the new fault category is updated through the trained prototype adaptive network PAN, and the fault classification is performed by calculating the cosine similarity between the query sample of the new fault category and the prototype representation of the new fault category.

2. The method according to claim 1, characterized in that In step S1, the basic data set containing multiple fault types is preprocessed, and the basic data set is divided into a pseudo basic data set and a pseudo incremental data set, including: Perform Fourier transform on the basic data set to obtain a frequency domain signal, and reshape the one-dimensional frequency domain signal into a two-dimensional frequency domain signal; The basic dataset is divided into a pseudo basic dataset and a pseudo incremental dataset in a 1:1 ratio. The pseudo basic dataset is used to train the pseudo multi-scale convolutional neural network PMCNN, and the pseudo incremental dataset is used to train the prototype adaptive network PAN.

3. The method according to claim 1, characterized in that Both the pseudo multi-scale convolutional neural network PMCNN in step S2 and the multi-scale convolutional neural network MCNN in step S4 adopt a multi-scale convolutional structure to simultaneously extract fault features of different scales.

4. The method according to claim 1, characterized in that: The training process of the prototype adaptive network PAN in step S3 is: The self-attention mechanism is introduced to dynamically update the prototype representation. In the pseudo-incremental stage, the pseudo-incremental data set is used to train the prototype adaptive network PAN, the pseudo-incremental data is divided into a support set and a query set, and the support set and the query set are respectively input into the pseudo multi-scale convolutional neural network PMCNN to obtain the support set embedding E s and queryset embedding E q ; Embed the support set E s Calculate the mean and get the new category prototype representation w in the pseudo-incremental stage new , input the pseudo basic data into the pseudo multi-scale convolutional neural network PMCNN to obtain the old category prototype representation w old ; The new category prototype is represented by w old and the old category prototype representation w new The replication expansion dimensions are: and Where L is the number of samples in the query set, N old is the number of categories in the pseudo-base stage, N new is the number of categories in the pseudo-incremental phase, and D is the number of dimensions of each embedding; The new category prototype is represented by w old , the old category prototype represents w new , query set embedding E q The prototype adaptive network PAN is trained as input. The principle of the prototype adaptive network PAN dynamically updating the prototype representation through the self-attention mechanism is as follows: Q=W Q X,K=W K X,V=W V X Where X is the input vector, X={w old , w new , E q }, Q is the query vector, K is the key vector, V is the value vector, Q, K, V are respectively passed through three trainable linear layers W Q (·),W K (·) and W V (·) get, d k is the dimension of the key vector K, W Q is the weight matrix that transforms the input vector into the query vector, W K is the weight matrix that converts the input vector into the key vector, W V is the weight matrix that converts the input vector into a value vector; The input vector X is updated by the following formula to obtain the updated X′: X′=Ψ(Attention(Q,K,V))+X Among them, Ψ(·) is a trainable linear layer, and pseudo incremental data is used to simulate the incremental stage training W Q (·),W K (·),W V (·) and Ψ(·) four linear layers. The trained prototype adaptive network PAN is obtained by multiple training of the four linear layers. The trained prototype adaptive network PAN can adaptively update the prototype representation of the new fault category according to the new and old prototype representations and feature embedding.

5. The method according to claim 1, characterized in that In step S6, the fault classification is performed by calculating the cosine similarity between the query sample of the new fault category and the prototype representation of the new fault category. The formula is as follows: Among them, w i represents the prototype of the i-th category, q represents the feature vector of the query sample, and the category with the highest similarity is used as the fault classification result, so as to achieve accurate diagnosis of new fault categories.

Images