Self-lifting diagnosis method of mechatronic transmission system based on memory subnetwork

Abstract

The present invention relates to the technical field of intelligent operation and maintenance of mechanical equipment, and specifically to a self-promoting diagnosis method for a mechatronic transmission system based on a memory subnetwork, comprising: constructing an autonomously promoted fault diagnosis framework including a health category prediction network, a stable information memory subnetwork and a fast information memory subnetwork; in the autonomously promoted learning stage, randomly extracting a fault diagnosis data set A, and randomly extracting a data set B from the learned information space; using two subnetworks to predict data set B, and selecting the best playback logits; predicting data set A and data set B based on a health category prediction network to obtain two logits; calculating the total loss based on the playback logits and the two logits of the health category prediction network, updating the weights of each network and the learned information space; performing multiple autonomously promoted learning until the training target is met. The present invention can continuously train a fault diagnosis network from a continuously added data stream to achieve autonomous promotion of the network.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of mechanical equipment, and more specifically, to a self-improving diagnosis method for an electromechanical composite transmission system based on a memory sub-network. Background Art

[0002] With the vigorous development of modern science and technology, complex mechanical equipment such as large machine tools, high-speed trains, and wind turbines has become an important pillar of social production. In the face of the continuous increase in the operating intensity of mechanical equipment and the increasing improvement of technical precision, it is particularly important to ensure the safe and reliable operation of mechanical equipment. Mechanical equipment is an integrated whole with strong coupling of multiple systems, and the normal operation of each subsystem is indispensable. Among them, the electromechanical composite transmission system undertakes the key task of efficient power transmission. Due to the complex and changeable service environment, various types of failures will inevitably occur. Once a failure occurs, it will weaken the operating performance of the equipment, and even lead to the paralysis of the overall system, causing heavy economic losses. Equipment operation and maintenance personnel hope to detect failures at the initial stage of the abnormality of the transmission system, take maintenance measures in time to reduce accidents, and lower the maintenance cost.

[0003] Benefiting from the progress of artificial intelligence and Internet of Things technologies, intelligent fault diagnosis methods based on deep learning overcome the limitations of insufficient equipment maintenance experience, can autonomously learn fault features from the vibration signals of mechanical equipment, and achieve efficient and accurate diagnosis, which are highly regarded in the field of mechanical equipment operation and maintenance. Traditional intelligent diagnosis methods rely on a complete training data set, that is, the data set covers all fault categories and provides sufficient data for each category. However, in the early stage of actual operation and maintenance, the equipment is in a healthy state for a long time, and it is difficult to construct a training data set with a complete variety of fault modes in a short period of time. Usually, the equipment data is gradually acquired over time, and the number of fault categories will also increase accordingly. After introducing new fault categories, training the model separately with the data of new fault categories may cause "catastrophic forgetting" of the model, that is, the model loses the ability to identify old fault categories; while retraining the model requires a lot of time and computing resources, which limits the application potential of intelligent fault diagnosis technology.

[0004] Continuous learning methods can well solve the above problems, but the existing methods have the following deficiencies: they cannot retain the high-precision recognition ability for old categories when learning new category samples progressively, and it is difficult to achieve the balance between network plasticity and stability; they cannot efficiently retain the learned information, that is, they often rely on storing a large number of past redundant task samples, but these samples may not fully represent the core knowledge of the task, and as the number of tasks increases, the number of replay samples will also increase sharply, resulting in storage pressure. These limitations make it difficult to apply the self-improving diagnosis method efficiently and flexibly in actual engineering. There is an urgent need for a method that can continuously train the fault diagnosis network from the continuously increasing data stream and achieve the self-improvement of the network. Summary of the Invention

[0005] In view of this, the present invention provides a self-improving diagnosis method for an electromechanical composite transmission system based on a memory sub-network, which can continuously train a fault diagnosis network from continuously increasing data streams and achieve the autonomous improvement of the network.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A self-improving diagnosis method for an electromechanical composite transmission system based on a memory sub-network, comprising the following steps:

[0008] Construct an autonomous improvement fault diagnosis framework including a health category prediction network, a stable information memory sub-network, and a fast information memory sub-network;

[0009] Enter the initial learning stage, and randomly extract a data set A from the continuously increasing fault diagnosis data stream of the electromechanical composite transmission system;

[0010] Predict the health status of the samples in the current stage data set A based on the health category prediction network;

[0011] Calculate the loss based on the logits output of the health category prediction network, and update the weights of each network and the learned information space;

[0012] Enter the autonomous improvement learning stage, randomly extract a data set A from the continuously increasing fault diagnosis data stream of the electromechanical composite transmission system; randomly extract a data set B from the learned information space, and combine the data set A and the data set B as the training set in the current stage;

[0013] Respectively use the stable information memory sub-network and the fast information memory sub-network to predict the health status of the samples in the current stage data set B, and select the one with a higher score as the final replay logits output;

[0014] Based on the health category prediction network, respectively predict the health status of the samples in the current stage data set A and data set B, and obtain two logits outputs;

[0015] Calculate the total loss based on the final replay logits output and the two logits outputs of the health category prediction network, and update the weights of each network and the learned information space;

[0016] Perform multiple autonomous improvement learning until the training target is met.

[0017] Further, in the initial learning stage, calculate the loss L based on the cross-entropy loss function, and its calculation formula is:

[0018]

[0019] Among them, represents the logits output value of any sample X in the dataset A at the initial learning stage a belonging to the health category j, represents the sample label Y a the true label value of the health category j encoded in one-hot, C represents the number of health categories, and bs represents the batch size.

[0020] Furthermore, in the initial learning stage and the self-improving learning stage, the weight update method for the health category prediction network is the same, and the update formula is:

[0021] φ←φ - η▽ φ L

[0022] Among them, η represents the learning rate of the health category prediction network; ▽ φ L is the gradient of the loss function with respect to the weights φ of the health category prediction network, representing the change direction and magnitude of the loss L with respect to the weights φ.

[0023] Furthermore, in the initial learning stage and the self-improving learning stage, the weight update method for the stable information memory sub-network and the fast information memory sub-network is the same, and the update process includes:

[0024] Generate two random numbers a and b uniformly distributed to determine whether to update the weights of the fast information memory sub-network and the stable information memory sub-network;

[0025] If a < σ p , then update the weights of the fast information memory sub-network through the EMA formula, and the update formula is:

[0026] δ←ε p δ+(1 - ε p )φ

[0027] Among them, σ p represents the upgrade rate of the fast information memory sub-network, ε p represents the forgetting rate of the fast information memory sub-network, δ represents the weights of the fast information memory sub-network, and φ represents the weights of the health category prediction network at the current stage;

[0028] If b < σ s , then update the weights of the stable information memory sub-network through the EMA formula, and the update formula is:

[0029]

[0030] Among them, σ s represents the upgrade rate of the stable information memory sub-network, ε s represents the forgetting rate of the stable information sub-network, Represents the weights of the stable information memory sub-network.

[0031] Furthermore, in the initial learning stage and the self-improving learning stage, the way of updating the learned information space is the same, and the update process includes:

[0032] For any sample X in the dataset A extracted in each learning stage a , if the learned information space has not reached the capacity M, then directly add the current sample X a to the memory bank;

[0033] If the learned information space is full, generate a random integer j in the interval [0, N], where N represents the total number of samples received so far; if j < M, replace the j-th sample in the memory bank with the current sample X a , and if j ≥ M, do not store the current sample.

[0034] Furthermore, in the self-improving learning stage, the determination method of the final replay logits output includes:

[0035] Select the learning method based on the softmax function, compare the logits output O of the stable information memory sub-network s and the logits output O of the fast information memory sub-network p , and select the one with a higher score as the final replay logits output. The comparison formula is as follows:

[0036] if Softmax(O p )(Y b ) > Softmax(O s )(Y b )

[0037] O r = O p else O r = O s

[0038] where Y b represents the true sample label, and O r represents the final replay logits output.

[0039] Furthermore, in the self-improving learning stage, calculate the total loss L based on the cross-entropy loss function and the consistency loss function. The calculation formula of the total loss L is:

[0040]

[0041] where O ADenote the set of logits values obtained by the health category prediction network for predicting dataset A during the self-improving learning phase, O B Denote the set of logits values obtained by the health category prediction network for predicting dataset B; O R Is the final replay logits value set; L CE Denote the cross-entropy loss function; L MSE Denote the consistency loss function; Denote any sample X in dataset A during the self-improving learning phase a The logits output value when the health category to which it belongs is j; Denote the logits output value when the health category to which any sample in dataset B during the self-improving learning phase belongs is j; Denote O R The logits output value when the health category in is j; Denote the sample label Y a The true label value in one-hot encoding, where C represents the number of health categories.

[0042] From the above technical solutions, it can be seen that compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention establishes a learned information space, and uses a stable information memory sub-network and a fast information memory sub-network to carry out self-improvement.

[0044] 1) The present invention uses a stable information memory sub-network and a fast information memory sub-network to form a dual information memory sub-network. The fast information sub-network simulates the rapid learning ability, and the stable information sub-network maintains long-term knowledge. According to the output scores of the dual networks, the best replay logits are selected, maximizing the network's recognition ability; comprehensively learning new sample and old sample data, calculating the consistency loss and cross-entropy loss, and updating the parameters of the health category prediction network, achieving a balance between the ability to adapt to new knowledge and the ability to retain old knowledge, while improving the recognition ability of new category faults, ensuring the accuracy of recognizing old category faults to the greatest extent.

[0045] 2) The present invention stores long-term knowledge through the learned information space, and uses the local shuffling sampling method to dynamically update the stored old task samples, efficiently integrating the core samples of the tasks, facilitating continuous review of old sample knowledge during the subsequent incremental information learning process; without requiring a large sample storage space, and continuously integrating long-term knowledge using the memory sub-network during subsequent learning, effectively avoiding catastrophic forgetting caused by continuous learning.

[0046] 3) The present invention belongs to a general self-improving diagnosis method, without making any requirements on data and tasks, and can adapt to non-independent and identically distributed data streams from dynamic environments, conforming to engineering reality and being more flexible. Description of the Drawings

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0048] Figure 1 It is a flowchart of the self-improving diagnosis method for an electro-mechanical composite drive system based on a memory sub-network provided by the present invention;

[0049] Figure 2 It is a schematic diagram of the self-improving fault diagnosis framework provided by the present invention;

[0050] Figure 3 It is a schematic diagram of the update strategy of the health category prediction network provided by the present invention;

[0051] Figure 4 It is a performance comparison diagram between the method of the present invention and the existing learning methods. Detailed Embodiments

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0053] As Figure 1 shown, the embodiments of the present invention disclose a self-improving diagnosis method for an electro-mechanical composite drive system based on a memory sub-network, including the following steps:

[0054] Construct a self-improving fault diagnosis framework including a health category prediction network, a stable information memory sub-network, and a fast information memory sub-network;

[0055] Enter the initial learning stage, and randomly extract the data set A from the continuously increasing electro-mechanical composite drive system fault diagnosis data stream;

[0056] Predict the health status of the samples in the current stage data set A based on the health category prediction network;

[0057] Calculate the loss based on the logits output of the health category prediction network, and update the weights of each network and the learned information space;

[0058] Enter the self-improving learning stage. Randomly extract dataset A from the continuously increasing fault diagnosis data stream of the electromechanical composite drive system, and randomly extract dataset B from the learned information space. Combine dataset A and dataset B as the training set for the current stage;

[0059] Use the stable information memory sub-network and the fast information memory sub-network to predict the health status of the samples in dataset B of the current stage respectively, and select the one with a higher score as the final replay logits output;

[0060] Based on the health category prediction network, predict the health status of the samples in dataset A and dataset B of the current stage respectively, and obtain two logits outputs;

[0061] Calculate the total loss based on the final replay logits output and the two logits outputs of the health category prediction network, and update the weights of each network and the learned information space;

[0062] Perform multiple self-improving learning until the training goal is met.

[0063] The following further explains the above steps.

[0064] 1. Construct the self-improving fault diagnosis framework as shown in Figure 2 , including the health category prediction network F(X, φ), the stable information memory sub-network and the fast information memory sub-network F p (X, δ); where X is the input sample of the network, and φ, δ are the learnable network parameters (i.e., weights) of F, F s , F p respectively. At the same time, set hyperparameters, including the cross-entropy loss function L CE and the consistency loss function L MSE , the learning rate η, the consistency weight ω, the training batch size bs, the number of iterations, the upgrade rate σ s and σ p , the forgetting rate ε s and ε p , and the size M of the learned information space. The above network is stacked by an input layer, a convolutional layer, and a fully connected layer without Softmax activation, and its output is the logits that have not been processed by the Softmax function, which are the "raw scores" of the model for each category.

[0065] 2. Receive the fault diagnosis data stream. Randomly obtain a dataset from the continuously increasing non-independent and identically distributed fault diagnosis data stream of the electromechanical composite drive system. Among them, non-independent and identically distributed means that the data does not meet the conditions of mutual independence or identical distribution, or neither of them is satisfied. Any data can be expressed as (X, Y), where Represents the multi-channel signal data of the drive system, l is the signal sampling length, c is the number of signal channels, and Y is the corresponding health type label.

[0066] 3. Enter the initial learning stage, specifically including:

[0067] 3.1) Dataset acquisition.

[0068] Randomly obtain dataset A from the continuously increasing non-i.i.d. mechatronic drive system fault diagnosis data stream, where any data can be expressed as (X a , Y a ). It should be noted that the learned information space is empty at this time.

[0069] 3.2) Predict the health category and obtain the output result.

[0070] Predict the health status of the samples in the current task dataset A through the health category prediction network F(X, φ), where any sample can be expressed as (X a , Y a ), the logits output is O a , O a = F(X a , φ).

[0071] 3.3) Calculate the cross-entropy loss.

[0072] In the initial learning stage, calculate the loss L based on the cross-entropy loss function, and its calculation formula is:

[0073]

[0074] Among them, represents the logits output value when the health category of any sample X a in the dataset A in the initial learning stage belongs to health category j, represents the true label value of the sample label Y a for health category j in one-hot encoding, C represents the number of health categories, and bs represents the batch size.

[0075] 3.4) Use the backpropagation algorithm to train the health category prediction network, and the update formula is:

[0076] φ ← φ - η▽ φ L

[0077] Among them, η represents the learning rate of the health category prediction network; ▽ φ L is the gradient of the loss function with respect to the weights φ of the health category prediction network, indicating the change direction and magnitude of the loss L with respect to the weights φ.

[0078] 3.5) Update the weights of the dual memory sub-network. The update process includes:

[0079] Generate two random numbers a and b that are uniformly distributed on [0, 1] to determine whether to update the weights of the fast information memory sub-network and the stable information memory sub-network;

[0080] If a < σ p , then update the weights of the fast information memory sub-network through the EMA formula. The update formula is:

[0081] δ ← ε p δ+(1 - ε p )φ

[0082] where σ p represents the upgrade rate of the fast information memory sub-network, ε p represents the forgetting rate of the fast information memory sub-network, δ represents the weights of the fast information memory sub-network, and φ represents the weights of the healthy class prediction network in the current stage;

[0083] If b < σ s , then update the weights of the stable information memory sub-network through the EMA formula. The update formula is:

[0084]

[0085] where σ s represents the upgrade rate of the stable information memory sub-network, ε s represents the forgetting rate of the stable information sub-network, represents the weights of the stable information memory sub-network.

[0086] 3.6) Update the learned information space, specifically including:

[0087] For any sample X in the dataset A extracted in each learning stage a , if the learned information space has not reached the capacity M, then directly add the current sample X a to the memory bank;

[0088] If the learned information space is full, generate a random integer j in the interval [0, N), where N represents the total number of samples received so far; if j < M, replace the j-th sample in the memory bank with the current sample X a , if j ≥ M, do not store the current sample. This method ensures that each sample in the data stream is sampled into the learned information space with equal probability.

[0089] 4. Enter the autonomous improvement learning stage, specifically including:

[0090] 4.1) Input data stream and obtain data set: Randomly obtain data set A from the continuously increasing non-i.i.d. mechatronic composite transmission system fault diagnosis data stream, where any data can be expressed as

[0091] (X a ,Y a ). Randomly extract data set B from the learned information space, where any data can be expressed as (X b ,Y b ), and Y b is the corresponding health type label. Combine the above data set samples into training set S, where S = A ∪ B.

[0092] 4.2) Predict the health type of samples in data set B and obtain the output result: Use different memory sub-networks to predict the health type of samples in data set B, where any sample can be expressed as (X b ,Y b ), and obtain the logits prediction results O of the stable information memory sub-network p and the fast information memory sub-network F s (X,δ), and O p . O p = F p (X b ,δ).

[0093] 4.3) Compare the output results of the two memory sub-networks and select the final replay logits output, specifically including:

[0094] Select the learning method based on the softmax function, compare the logits output O s of the stable information memory sub-network and the logits output O p of the fast information memory sub-network, and select the one with a higher score as the final replay logits output. The comparison formula is as follows:

[0095] if Softmax(O p )(Y b ) > Softmax(O s )(Y b )

[0096] O r = O p else O r = O s

[0097] where, Y b represents the true sample label, and O rRepresents the final replay logits output. In this way, it can be ensured that the appropriate replay logits are selected for the healthy category prediction when dealing with new and old tasks.

[0098] 4.4) Calculate the cross-entropy loss and the consistency loss.

[0099] Use the healthy category prediction network to predict dataset A and dataset B, where any sample can be represented as X a and X b , and obtain the result O a =F(X a ,φ) and O b =F(X b ,φ), and calculate the total loss L. The total loss includes two parts, namely the cross-entropy loss function L CE and the consistency loss function L MSE , and its expression is:

[0100]

[0101] where, O A represents the set of logits values for the prediction of dataset A by the healthy category prediction network in the self-improving learning stage, and O B represents the set of logits values for the prediction of dataset B by the healthy category prediction network; O R is the set of final replay logits values; L CE represents the cross-entropy loss function; L MSE represents the consistency loss function; represents the logits output value when any sample X a in dataset A belongs to the healthy category j in the self-improving learning stage; represents the logits output value when any sample in dataset B belongs to the healthy category j in the self-improving learning stage; represents the logits output value of the healthy category j in O R ; represents the true label value of the sample label Y a encoded in one-hot, and C represents the number of healthy categories.

[0102] 4.5) Use the backpropagation algorithm to update the weights of the healthy category prediction network. The specific update method is as follows:

[0103] φ←φ-η▽ φ L

[0104] where, η is the learning rate, and ▽ φ L represents the gradient of the loss function with respect to the weights φ of the healthy category prediction network, which represents the change direction and magnitude of the loss L with respect to the weights φ.

[0105] 4.6) As shown in Figure 3 , update the weights of the dual memory sub-networks. This part is the same as the initial learning stage and specifically includes:

[0106] Generate two uniformly distributed random numbers a and b to determine whether to update the weights of the fast information memory sub-network and the stable information memory sub-network. For the fast information memory, if a < σ p , then use the EMA formula to update the weights of the fast information memory sub-network:

[0107] δ ← ε p δ+(1 - ε p )φ

[0108] where ε p is the forgetting rate of the fast information sub-network.

[0109] For the long-term semantic memory, if b < σ s , then use the EMA formula of the weights of the health recognition model to update the weights of the stable information memory sub-network:

[0110]

[0111] where ε s is the forgetting rate of the stable information sub-network.

[0112] 4.7) Update the learned information space. This part is similar to the initial learning stage and specifically includes:

[0113] For each data sample X of the new task a , if the learned information space has not reached the capacity M, directly add the current sample X a to the memory bank. If the learned information space is full, generate a random integer j in the interval [0, N], where N is the total number of samples received so far. If j < M, replace the j-th sample in the memory bank with the current sample. This method ensures that each sample in the data stream is sampled into the learned information space with equal probability.

[0114] 5. Repeat the above step 4 until the training objective is met.

[0115] The present invention also takes the fault diagnosis of the electromechanical composite drive system of the rail train as an example, and verifies the effectiveness of the method of the present invention through the fault simulation data of the test bench.

[0116] In this embodiment, the test bench adopted is controlled by a frequency converter to vary the rotational speeds of different motors, and an electro-hydraulic load device is used to apply different lateral loads. The signals collected in the experiment include 18 channels, with a sampling frequency of 64 kHz, including the three-axis accelerations of the motor, gearbox, and axle box, as well as the three-phase currents of the motor. In the experiment, 16 health states of the electromechanical composite transmission system are considered, including the normal state, 4 types of axle box bearing faults (inner race fault, outer race fault, rolling element fault, and cage fault), 4 types of pinion support bearing faults, 4 types of pinion faults (crack, wear, tooth breakage, and tooth missing), and 3 types of motor faults (short circuit, rotor bending, and rotor bar breakage). Each health state includes 1000 samples, and each sample includes 1024 sampling points. The method proposed in the present invention is used for fault diagnosis of the motor, gearbox, and axle box in the experiment.

[0117] On this basis, to facilitate the comparative experiment of traditional methods, the dataset is divided into three stages according to the axle box fault, gearbox fault, and motor fault. The data in each stage is used for training and testing in a 1:1 ratio. The training set of each stage includes a small number of sample originals sampled from the training data of the previous stage and the training data of the current stage. For the test set, it includes all the test data of the previous stage and the current stage.

[0118] The method proposed in the present invention is compared with two common methods, namely classifier parameter fine-tuning and full network parameter fine-tuning. The full network parameter fine-tuning updates all the trainable parameters of the feature extraction layer and the output layer based on the parameters of the previous stage, while the classifier fine-tuning only updates the parameters of the output layer. The basic structure of the network is shown in Table 1, and the training-related parameters are shown in Table 2. The experimental results are summarized in Table 3 and Figure 4 .

[0119] Table 1. Summary Table of the Basic Structure of the Network

[0120]

[0121]

[0122] Table 2. Summary Table of Training-Related Parameters

[0123]

[0124] Table 3. Summary of the Accuracy in the Final Stage

[0125]

[0126] It can be seen from the experimental results that with the continuous learning of the method of the present invention, the diagnostic effect of the network is significantly better than the two comparative methods, and it can better achieve autonomous improvement in diagnosis.

[0127] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0128] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A self-lifting diagnosis method for a mechatronic transmission system based on a memory subnetwork, characterized in that: The following steps are involved: Construct an autonomous improvement fault diagnosis framework consisting of a health category prediction network, a stable information memory subnetwork, and a fast information memory subnetwork; Entering the initial learning stage, a data set A is randomly selected from the continuously increasing mechatronic transmission system fault diagnosis data stream; Predict the health status of samples in the current dataset A based on the health category prediction network; Calculate the loss based on the logits output of the health category prediction network and update the weights and learned information space of each network; Entering the autonomous learning stage, randomly extracting data set A from the continuously increasing mechatronic transmission system fault diagnosis data stream; Randomly extract dataset B from the learned information space, and combine dataset A and dataset B as the training set for the current stage; Use the stable information memory subnetwork and the fast information memory subnetwork to predict the health of the samples in the current stage dataset B, and select the one with the higher score as the final playback logits output; Based on the health category prediction network, the health status of the samples in the current data set A and data set B is predicted respectively, and two logits outputs are obtained; Calculate the total loss based on the final playback logits output and the two logits outputs of the healthy category prediction network, and update the weights and learned information space of each network; Conduct multiple self-improvement learning sessions until the training goals are met.

2. The self-lifting diagnosis method of the electromechanical hybrid transmission system based on the memory sub-network according to claim 1 is characterized in that: In the initial learning stage, the loss L is calculated based on the cross entropy loss function, and its calculation formula is: in, Represents any sample X in the dataset A in the initial learning phase a The logits output value of the health category j, Represents the sample label Y a The true label value of the healthy category j is encoded in one-hot, C represents the number of healthy categories, and bs represents the batch size.

3. The self-lifting diagnosis method of the electromechanical hybrid transmission system based on the memory sub-network according to claim 1 is characterized in that: In the initial learning stage and the autonomous learning stage, the weight update method of the health category prediction network is the same, and the update formula is: Where η represents the learning rate of the health category prediction network; It is the gradient of the loss function for the weight φ of the health category prediction network, indicating the direction and magnitude of the change of the loss L relative to the weight φ.

4. The self-lifting diagnosis method of the electromechanical hybrid transmission system based on the memory sub-network according to claim 1 is characterized in that: In the initial learning stage and the autonomous improvement learning stage, the weights of the stable information memory subnetwork and the fast information memory subnetwork are updated in the same way. The update process includes: Generate two uniformly distributed random numbers a and b to decide whether to update the weights of the fast information memory subnetwork and the stable information memory subnetwork; If a<σ p , then the weight of the fast information memory sub-network is updated through the EMA formula, and the update formula is: d←e p d+(1-e p )φ Among them, σ p represents the upgrade rate of the fast information memory subnetwork, ε p represents the forgetting rate of the fast information memory subnetwork, δ represents the weight of the fast information memory subnetwork, and φ represents the weight of the health category prediction network at the current stage; If b < σ s , then the weight of the stable information memory subnetwork is updated through the EMA formula, and the update formula is: Among them, σ s represents the upgrade rate of the stable information memory subnetwork, ε s represents the forgetting rate of the stable information sub-network, Represents the weight of the stable information memory sub-network.

5. The self-lifting diagnosis method of the electromechanical hybrid transmission system based on the memory sub-network according to claim 1 is characterized in that: In the initial learning stage and the self-improvement learning stage, the updated method of the learned information space is the same, and the update process includes: For any sample X in the dataset A extracted in each learning stage a If the learned information space has not reached the capacity M, then directly convert the current sample X a , added to the memory bank; If the learned information space is full, generate a random integer j in the range [0, N], where N represents the total number of samples received so far; if j < M, replace the j-th sample in the memory bank with the current sample X a If j ≥ M, do not store the current sample.

6. The self-lifting diagnosis method of the electromechanical hybrid transmission system based on the memory sub-network according to claim 1 is characterized in that: In the self-improvement learning stage, the final playback logits output is determined in the following ways: Based on the softmax function, the learning method is selected to stabilize the logits output of the information memory subnetwork. s and the output O of the logits of the fast information memory subnetwork p , select the one with a higher score as the final playback logits output, and the comparison formula is as follows: if Softmax(O p )(AND b )>Softmax(O s )(AND b ) O r =O p else O r =O s Among them, Y b represents the true sample label, O r Represents the final playback logits output.

7. The self-lifting diagnosis method of the electromechanical hybrid transmission system based on the memory sub-network according to claim 6 is characterized in that: In the autonomous learning stage, the total loss L is calculated based on the cross entropy loss function and the consistency loss function. The calculation formula of the total loss L is: Among them, O A represents the set of logits values ​​predicted by the health category prediction network for data set A during the autonomous learning phase, O B represents the collection of logits values ​​predicted by the health category prediction network for data set B; O R is the final collection of playback logits values; L CE represents the cross entropy loss function; L MSE represents the consistency loss function; Represents any sample X in the dataset A during the autonomous learning phase a The logits output value of the health category j; It represents the logits output value of the health category j of any sample in the dataset B during the autonomous learning stage; Indicates O R The logits output value of the health category j; Represents the sample label Y a The true label value encoded in one-hot, C represents the number of healthy categories.

Images