Self-lifting method for health state identification network of electromechanical composite transmission system

Abstract

The invention belongs to the field of intelligent operation and maintenance of mechanical equipment, and discloses a self-lifting method for a health state identification network of an electromechanical composite transmission system, and the method comprises the following steps: S1, training the health state identification network through employing a data set # imgabs0 #, and obtaining a weight parameter # imgabs1 #; s2, freezing a weight matrix of a feature extraction layer, and obtaining a far-end memory sample # imgabs3 # based on a data set # imgabs2 #; s3, performing partition conversion on a weight matrix # imgabs6 # of the linear layer # imgabs5 # by utilizing a far-end memory sample # imgabs4 # to obtain a reconstructed active region # imgabs7 # and a reconstructed dormant region # imgabs8 #; s4, freezing and reconstructing an active region # imgabs9 #, updating and reconstructing a dormant region # imgabs11 # by using a data set # imgabs10 #, and obtaining a weight matrix # imgabs13 # of a linear layer # imgabs12 #; obtaining a far-end memory sample # imgabs15 # based on the data set # imgabs14 #; and S5, continuously repeating the steps S3-S4 to obtain a self-lifting health state identification network. According to the invention, the problem of disastrous forgetting can be improved, and the problem that few sample data is easy to over-fit is avoided, so that the health state identification network has good identification accuracy.

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-improvement method for a health status identification network of an electromechanical composite transmission system. Background Art

[0002] With the development of modern industry, electromechanical transmission systems have greatly improved energy efficiency and enhanced the power performance of equipment by integrating mechanical transmission and electric drive technology. Compared with other transmission systems, electromechanical transmission systems have complex internal structures and are more prone to various faults, which seriously reduce the safety and reliability of equipment during service. Therefore, in order to avoid equipment failure and downtime, health status identification is a key step in equipment operation and maintenance. Its accuracy and efficiency will directly affect the safe and stable operation of the equipment, and then affect the economic benefits. Traditional health status identification methods rely on manual experience and regular maintenance, which is inefficient, making it difficult for a limited number of operation and maintenance personnel to handle a large group of transmission equipment. With the development of artificial intelligence technology, data-driven intelligent diagnosis methods have effectively improved diagnostic accuracy and reduced operation and maintenance costs by integrating advanced computer and signal processing technologies, and are highly respected in academic research and industrial production.

[0003] Intelligent health status identification methods require a large amount of training data to learn health status characteristics, and the diagnostic effect after network training is closely related to the training data. In the actual production process, the transmission system is in a normal state for a long time, and rarely fails. It is usually difficult to obtain sufficient and complete health status category sample data. When the number of training samples is insufficient, the network model is prone to overfitting, poor generalization ability, and cannot accurately distinguish all health states; when the training sample categories are incomplete, new health status category data can only be obtained in stages, and self-enhanced learning is prone to forgetting the previously learned health status characteristics, resulting in a decrease in the diagnostic performance of the old health state, and re-learning will bring serious waste of computing resources. At present, scholars' research is often limited to one of the above situations. For the problem of few samples, methods such as data enhancement and transfer learning are proposed; for self-enhanced learning, replay methods and regularization methods are proposed. However, these methods still face challenges in health status identification, such as the difficulty in ensuring the quality of generated data, the field difference problem of transfer learning, the need for a large amount of storage space for replay methods, and the increase in model complexity by dynamic architecture methods. Current research rarely considers the above two situations at the same time, and it is very difficult to carry out self-enhanced learning under the constraint of few samples. In summary, there is currently a lack of an effective self-improvement method for health status recognition networks that can learn new health status categories from a small amount of data while retaining old knowledge, overcoming problems such as catastrophic forgetting and overfitting, and achieving good recognition accuracy.

[0004] Therefore, how to provide a self-improvement method for the health status identification network of an electromechanical hybrid transmission system, which can improve the catastrophic forgetting problem, avoid the problem of overfitting of small sample data, and make the health status identification network have good identification accuracy is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, an object of the present invention is to provide a self-improvement method for a health status identification network of an electromechanical hybrid transmission system.

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

[0007] A self-improvement method for a health status identification network of an electromechanical hybrid transmission system comprises the following steps:

[0008] S1: Using the Phase 1 Composite Drivetrain Health Status Dataset Train the health status recognition network to obtain the weight parameters of the health status recognition network ; Wherein, the first stage composite transmission system health status data set The health status recognition network includes sample data of several health status categories; the health status recognition network is composed of a feature extraction layer and a linear layer. Stacked, ; The weight parameter Including the linear layer The weight matrix and the weight matrix of the feature extraction layer;

[0009] S2: Freeze the weight matrix of the feature extraction layer and extract the composite transmission health status dataset from the first stage Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the first stage remote memory sample ;

[0010] S3: Using the i-1 stage remote memory sample For the linear layer The weight matrix Perform partition conversion to obtain the reconstruction activation area and refactoring the dormant area ;in, ;

[0011] S4: Freeze the reconstructed activation area , and using the phase i composite transmission system health status dataset Update the reconstruction sleep area , obtain the linear layer The weight matrix ; Among them, the health status categories included in the health status data set of the composite transmission system at each stage do not overlap;

[0012] From the i-th stage composite transmission system health status dataset A typical sample is selected from each health status category sample data in the memory storage space to obtain the remote memory sample of the i-th stage ;

[0013] S5: Repeat S3-S4 continuously to obtain a self-improving health status identification network.

[0014] Preferably, S1 specifically comprises the following steps:

[0015] S11: The first stage composite transmission system health status data set Input into the health status recognition network, perform forward propagation to predict the health status category and calculate the cross entropy loss;

[0016] S12: Using the cross entropy loss to perform back propagation update to obtain the weight parameters of the health status identification network .

[0017] Preferably, the feature extraction layer is formed by stacking multiple convolution-pooling layers.

[0018] Preferably, S3 specifically includes the following steps:

[0019] S31: The i-1th stage remote memory sample Input to the feature extraction layer to obtain the first-order feature matrix ;

[0020] S32: Using the first-order moment feature matrix Calculate the second-order moment characteristic matrix ;

[0021] S33: Using the second-order moment feature matrix The linear layer The weight matrix Active Zone and sleeping area ;

[0022] S34: for the activation area and the dormant area Convert to obtain the reconstructed activation area and the reconstructed sleep area .

[0023] Preferably, the second-order moment feature matrix The calculation formula is:

[0024] ;

[0025] in, Represents the remote memory sample of the i-1th stage The number of samples included; T indicates transpose.

[0026] Preferably, S33 is implemented based on the following formula:

[0027] ;

[0028] in, In turn, the size is The orthogonal matrix of size is The important value diagonal matrix of size is The orthogonal matrix of Represents the partition process calculation function; represents the total number of important values, Sequentially indicate The output and input dimensions of express The kth column of express The kth important value of ; express The kth row of ; Z represents the hyperparameter.

[0029] Preferably, S34 is implemented based on the following formula:

[0030] ;

[0031] in, represents a low-rank high-redundancy matrix; express The k-th dimension value of ; Indicates selection The last Z column of Indicates selection The last Z important values ​​of ; Indicates selection The last Z rows.

[0032] Preferably, S4 specifically comprises the following steps:

[0033] S41: Freeze the reconstruction activation area , and using the phase i composite transmission system health status dataset Perform back propagation to update the low-rank high-redundancy matrix and low-rank high-redundancy matrix , obtain a low-rank high-redundancy matrix and low-rank high-redundancy matrix ;

[0034] S42: Based on the low-rank high-redundancy matrix , the low-rank high-redundancy matrix and the reconstructed activation region , obtain the linear layer The weight matrix .

[0035] Preferably, the linear layer The weight matrix The calculation formula is:

[0036] .

[0037] Preferably, a K-means or K-medoids clustering method is used to extract the health status data set of the composite transmission system in the first stage. Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the first stage remote memory sample ;

[0038] Using K-means or K-medoids clustering method, the health status data set of the composite transmission system in the i-th stage is obtained. Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the remote memory sample of the i-th stage .

[0039] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a self-improvement method for the health status identification network of an electromechanical composite transmission system, which realizes the self-improvement learning of the health status identification of the electromechanical composite transmission system under the constraint of few samples, improves catastrophic forgetting, and avoids the problem of easy overfitting of few-sample data. The present invention divides the linear layer, freezes the activation area of ​​the feature extraction layer and the linear layer, retains important knowledge components, and ensures that the original task capabilities of the model will not be lost when training new tasks. The present invention transforms the dormant area of ​​the linear layer and restates it as two low-rank high-redundancy matrices, using only fewer parameters to adapt to the few-sample tasks in the new stage, significantly reducing the computational complexity and the risk of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0041] Figure 1 The present invention provides a flow chart of a self-improvement method for a health status identification network of an electromechanical hybrid transmission system. DETAILED DESCRIPTION

[0042] 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.

[0043] like Figure 1 As shown, the embodiment of the present invention discloses a self-lifting method for a health status identification network of an electromechanical hybrid transmission system, comprising the following steps:

[0044] S1: Using the Phase 1 Composite Drivetrain Health Status Dataset Train the health status recognition network to obtain the weight parameters of the health status recognition network ; Wherein, the first stage composite transmission system health status data set The health status recognition network includes sample data of several health status categories; the health status recognition network is composed of a feature extraction layer and a linear layer. Stacked, ; The weight parameter Including the linear layer The weight matrix and the weight matrix of the feature extraction layer;

[0045] It can be understood that: each sample data includes characteristic data and a health status category label, and the characteristic data is multi-source sensor signal data of the composite transmission system.

[0046] In one embodiment, S1 specifically includes the following steps:

[0047] S11: The first stage composite transmission system health status data set Input into the health status recognition network, perform forward propagation to predict the health status category and calculate the cross entropy loss;

[0048] Specific:

[0049] Cross Entropy Loss ;

[0050] in, Indicates that the feature data Input into the health status identification network to obtain the logist output value; express Corresponding true health status category label One-hot encoding value of; bs represents the batch size; Represents the health status dataset of the composite transmission system in stage 1 The number of health status categories included.

[0051] S12: Using the cross entropy loss to perform back propagation update (update the weight parameters of the health status identification network) to obtain the weight parameters of the health status identification network (Weight parameter when the maximum number of iterations is reached).

[0052] The specific update method is: ;

[0053] in, represents the learning rate, Represents the gradient of the cross entropy loss to the weight parameter, specifically indicating the direction and magnitude of the change of the cross entropy loss relative to the weight parameter. Repeat the above steps until the maximum number of iterations is reached.

[0054] In one embodiment, the feature extraction layer is formed by stacking multiple convolution-pooling layers.

[0055] In one embodiment, the health status recognition network includes a first convolution-pooling layer, a second convolution-pooling layer, a third convolution-pooling layer, a linear layer connected in sequence. , Linear Layer ; The pooling layer is the maximum pooling layer.

[0056] S2: Freeze the weight matrix of the feature extraction layer and extract the composite transmission health status dataset from the first stage Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the first stage remote memory sample ;

[0057] In one embodiment, a K-means or K-medoids clustering method is used to extract the first phase composite transmission system health status data set. Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the first stage remote memory sample ;

[0058] The first stage of remote memory sample Included There are samples of different categories, and only one sample of each category is retained.

[0059] S3: Using the i-1 stage remote memory sample For the linear layer The weight matrix Perform partition conversion to obtain the reconstruction activation area and refactoring the dormant area ;in, ;

[0060] In one embodiment, S3 specifically includes the following steps:

[0061] S31: The i-1th stage remote memory sample Input to the feature extraction layer to obtain (linear layer Before) first-order moment characteristic matrix ;

[0062] S32: Using the first-order moment feature matrix Calculate the second-order moment characteristic matrix ;

[0063] In one embodiment, the second-order moment feature matrix The calculation formula is:

[0064] ;

[0065] in, Represents the remote memory sample of the i-1th stage The number of samples included; T indicates transpose.

[0066] S33: Using the second-order moment feature matrix The linear layer The weight matrix Active Zone and sleeping area ;

[0067] It should be noted that: ; ; Represents the feature dimension extracted by the feature extraction layer, Represents the remote memory sample of stage i-1 The number of samples included; R represents a real number;

[0068] In one embodiment, S33 is implemented based on the following formula:

[0069] ;

[0070] in, In turn, the size is An orthogonal matrix (a vector describing the direction of the output space) of size The important value diagonal matrix of (the important value diagonal matrix describing the weights of the linear layer) is An orthogonal matrix of (a vector describing the characteristics of the input space); Represents the partition process calculation function; represents the total number of important values, Sequentially indicate ( ) and the input dimensions (i.e. The size is ); express The kth column of express The kth important value of ); express The kth row of ; Z represents the hyperparameter (needs to be selected manually).

[0071] In one embodiment, assuming

[0072]

[0073] Decomposition to obtain:

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] S34: for the activation area and the dormant area Convert to obtain the reconstructed activation area and the reconstructed sleep area .

[0079] In one embodiment, S34 is implemented based on the following formula:

[0080] ;

[0081] in, represents a low-rank high-redundancy matrix; express The k-th dimension value of ; Indicates selection The last Z column of Indicates selection The last Z important values ​​of ; Indicates selection The last Z rows.

[0082] S4: Freeze the reconstructed activation area , and using the phase i composite transmission system health status dataset Update the reconstruction sleep area , obtain the linear layer The weight matrix ; Among them, the health status categories included in the health status data set of the composite transmission system at each stage do not overlap;

[0083] In one embodiment, S4 specifically includes the following steps:

[0084] S41: Freeze the reconstruction activation area , and using the phase i composite transmission system health status dataset Perform back propagation to update the low-rank high-redundancy matrix and low-rank high-redundancy matrix , obtain a low-rank high-redundancy matrix and low-rank high-redundancy matrix ;

[0085] S42: Based on the low-rank high-redundancy matrix , the low-rank high-redundancy matrix and the reconstructed activation region , obtain the linear layer The weight matrix .

[0086] In one embodiment, the linear layer The weight matrix The calculation formula is:

[0087] .

[0088] From the i-th stage composite transmission system health status dataset A typical sample is selected from each health status category sample data in the memory storage space to obtain the remote memory sample of the i-th stage ;

[0089] It is understandable that:

[0090] Health status dataset of composite transmission system in phase i Includes sample data for several health status categories; The included health status categories did not overlap;

[0091] Remote memory sample at stage i The categories included are The sum of the included categories.

[0092] In one embodiment, a K-means or K-medoids clustering method is used to cluster the health status data set of the composite transmission system in the i-th stage. Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the remote memory sample of the i-th stage .

[0093] S5: Repeat S3-S4 continuously to obtain a self-improving health status identification network.

[0094] Finally, in order to verify the advancement and effectiveness of the method of the present invention, the present invention uses an urban rail train bogie test bench to conduct relevant experiments. The test bench uses a three-phase asynchronous motor to apply different speeds to simulate different running speeds during the train's travel, and uses hydraulic equipment to simulate the load to correspond to the number of passengers in different time periods. The experiment simulates three major devices including motors, gear boxes and axle boxes in the electromechanical composite transmission system, and collects 17 health states of the electromechanical composite transmission system, including normal state, 4 pinion support bearing faults, 4 pinion faults (cracks, wear, broken teeth and missing teeth), 4 axle box bearing faults (inner ring fault, outer ring fault, rolling element fault and cage fault), 4 motor faults (motor bearing fault, short circuit, rotor bending and rotor broken bars), and the sampling signal includes 18 channels with a frequency of 64kHz. In the experiment, a large number of samples are provided in the first stage, and each health state in the subsequent stage includes 10 sample data, and the length of each sample is 3200.

[0095] The present invention uses the normal state, four pinion support bearing faults, and four pinion faults (cracks, wear, broken teeth, and missing teeth) as the first stage data set; uses four axle box bearing faults (inner ring fault, outer ring fault, rolling element fault, and cage fault) as the second stage data set; and uses four motor faults (motor bearing fault, short circuit, rotor bending, and rotor bar breakage) as the third stage data set;

[0096] Then, the present invention is compared with the three existing advanced self-improvement methods of full network fine-tuning, classifier fine-tuning and replay logit. The four methods share the same health status identification network structure and training-related parameters. 10 health status category identification experiments were carried out, and the health status identification accuracy of the three major devices was recorded. The structure of the health status identification network is shown in Table 1, the training-related parameters are shown in Table 2, and the network health status identification accuracy is shown in Table 3. It can be seen from the experimental results that the health status identification effect of the method of the present invention is significantly better than the other three methods. The method of the present invention effectively avoids catastrophic forgetting, better adapts to the few-sample learning problem in the new stage, and achieves higher health status identification accuracy.

[0097] Table 1. Structure of health status recognition network

[0098]

[0099] Table 2. Training related parameters

[0100]

[0101] Table 3. Network health status identification accuracy

[0102]

[0103] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0104] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may 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-improvement method for a health status identification network of an electromechanical hybrid transmission system, characterized in that: The following steps are involved: S1: Using the Phase 1 Composite Drivetrain Health Status Dataset Train the health status recognition network to obtain the weight parameters of the health status recognition network ; Wherein, the first stage composite transmission system health status data set The health status recognition network includes sample data of several health status categories; the health status recognition network is composed of a feature extraction layer and a linear layer. Stacked, ; The weight parameter Including the linear layer The weight matrix and the weight matrix of the feature extraction layer; S2: Freeze the weight matrix of the feature extraction layer and extract the composite transmission health status dataset from the first stage Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the first stage remote memory sample ; S3: Using the i-1 stage remote memory sample For the linear layer The weight matrix Perform partition conversion to obtain the reconstruction activation area and refactoring the dormant area ;in, ; S4: Freeze the reconstructed activation area , and using the phase i composite transmission system health status dataset Update the reconstruction sleep area , obtain the linear layer The weight matrix ; Among them, the health status categories included in the health status data set of the composite transmission system at each stage do not overlap; From the i-th stage composite transmission system health status dataset A typical sample is selected from each health status category sample data in the memory storage space to obtain the remote memory sample of the i-th stage ; S5: Repeat S3-S4 continuously to obtain a self-improving health status identification network.

2. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 1 is characterized in that: S1 specifically includes the following steps: S11: The first stage composite transmission system health status data set Input into the health status recognition network, perform forward propagation to predict the health status category and calculate the cross entropy loss; S12: Using the cross entropy loss to perform back propagation update to obtain the weight parameters of the health status identification network .

3. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 1 is characterized in that: The feature extraction layer is formed by stacking multiple convolution-pooling layers.

4. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 1 is characterized in that: S3 specifically includes the following steps: S31: The i-1th stage remote memory sample Input to the feature extraction layer to obtain the first-order feature matrix ; S32: Using the first-order moment feature matrix Calculate the second-order moment characteristic matrix ; S33: Using the second-order moment feature matrix The linear layer The weight matrix Active Zone and sleeping area ; S34: for the activation area and the dormant area Convert to obtain the reconstructed activation area and the reconstructed sleep area .

5. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 4 is characterized in that: The second-order moment characteristic matrix The calculation formula is: ； in, Represents the remote memory sample of the i-1th stage The number of samples included; T indicates transpose.

6. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 4 is characterized in that: S33 is implemented based on the following formula: ； in, In turn, the size is The orthogonal matrix of size is The important value diagonal matrix of size is The orthogonal matrix of Represents the partition process calculation function; represents the total number of important values, Sequentially indicate The output and input dimensions of express The kth column of express The kth important value of ; express The kth row of ; Z represents the hyperparameter.

7. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 4 is characterized in that: S34 is implemented based on the following formula: ； in, represents a low-rank high-redundancy matrix; express The k-th dimension value of ; Indicates selection The last Z column of Indicates selection The last Z important values ​​of ; Indicates selection The last Z rows.

8. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 7 is characterized in that: S4 specifically includes the following steps: S41: Freeze the reconstruction activation area , and using the phase i composite transmission system health status dataset Perform back propagation to update the low-rank high-redundancy matrix and low-rank high-redundancy matrices , obtain a low-rank high-redundancy matrix and low-rank high-redundancy matrices ; S42: Based on the low-rank high-redundancy matrix , the low-rank high-redundancy matrix and the reconstructed activation region , obtain the linear layer The weight matrix .

9. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 8, characterized in that: The linear layer The weight matrix The calculation formula is: 。 10. The self-improvement method of the health status identification network of the electromechanical hybrid transmission system according to claim 1, characterized in that: Using K-means or K-medoids clustering method from the first phase composite transmission system health status dataset Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the first stage remote memory sample ; Using K-means or K-medoids clustering method, the health status data set of the composite transmission system in the i-th stage is obtained. Select a typical sample from each health status category in the sample data and store it in the memory storage space to obtain the remote memory sample of the i-th stage .

Images