Remaining life prediction method of electric drive system based on few-sample data enhancement and migration
By enhancing and transfer learning of the fault data of the electric drive system, the problem of poor fault prediction accuracy with few sample data is solved, and more efficient fault prediction and system reliability are achieved.
Patent Information
- Application Number
- CN202410452105.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-16
- Publication Date
- 2025-02-07
- Estimated Expiration
- 2044-04-16
AI Technical Summary
In the fault test of electric drive systems, the sample size of the fault data is small, resulting in poor accuracy of the deep learning model and unable to meet the actual application needs.
Using the method of enhanced and migration of few samples, the feature label calibration and time sliding window data enhancement of the fault data of the electric drive system in the fault state, combined with the pre-trained LSTM network model for transfer learning, and a robust fault is trained. Predictive model.
Effectively using less data for training improves the identification and prediction accuracy of potential faults of motor assembly and improves the reliability and performance of the entire drive system.
Smart Images
Figure CN118052154B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of fault testing, and in particular relates to a method for predicting the remaining life of an electric drive system by enhancing and migrating a small number of sample data. Background Art
[0002] Various types of failures that occur during the testing of the drive motor system and the electric drive assembly it contains will cause the project to stagnate, especially during the project development stage, which will greatly affect production. The electric drive assembly fault prediction model is a machine learning-based fault prediction method that is used to identify and predict possible faults in advance. This model usually relies on data-driven and reduces the risk of sudden failures.
[0003] However, in the actual testing process, the sample size of fault data is relatively small, and the accuracy obtained by training with deep learning models is poor and cannot meet the needs of actual applications. Therefore, how to train a robust model under effective data conditions is still a difficult problem that needs to be solved urgently. Summary of the invention
[0004] In view of this, the present invention aims to propose a method for predicting the remaining life of an electric drive system with a small number of sample data enhancement and migration, in order to solve at least one of the above-mentioned technical problems.
[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:
[0006] The first aspect of the present invention provides a method for predicting the remaining life of an electric drive system by enhancing and migrating a small number of sample data, comprising:
[0007] Collect the operating data of the electric drive system under normal working conditions, and use the operating data to pre-train the LSTM network model;
[0008] Collect fault data of the electric drive system in a fault state, calibrate the feature labels of the fault data, and perform data enhancement processing on the small sample fault data through a time sliding window;
[0009] Perform transfer learning on the pre-trained LSTM network model to obtain a fault prediction model, and use the enhanced fault data to train the fault prediction model;
[0010] Use the trained fault prediction model to perform fault prediction and output the predicted fault curve.
[0011] Furthermore, the process of collecting the operating data of the electric drive system under normal working conditions includes: selecting a plurality of drive devices of different powers in the electric drive system as prototypes, and collecting the operating data of the plurality of prototypes under the same working conditions and normal working conditions; wherein the operating data includes output current, temperature, and housing vibration data;
[0012] The collected data are normalized and preprocessed after unifying the time scale.
[0013] Furthermore, the LSTM network model includes an encoder and a decoder, a repeated vector layer is provided between the encoder and the decoder, and a time-distributed connection layer is provided at the output end of the decoder;
[0014] The encoder calculates the activation vectors of the forget gate, input gate, output gate, and cell state, forward propagates through the LSTM unit, obtains the hidden state of each time step, and uses the hidden state as the context to represent the cell state;
[0015] The decoder receives the hidden state corresponding to the last time step as the initial state, uses the same LSTM unit as the encoder, calculates the hidden state and cell state of the decoder and generates a reconstructed output sequence.
[0016] Furthermore, the process of calibrating the characteristic labels of the fault data includes: setting label points in the fault data based on the time of fault occurrence, calculating the absolute value of the difference between the time of each label point and the time of fault occurrence, and obtaining the remaining time;
[0017] The remaining time is input into the LSTM network model as a feature label, and the fault data input into the LSTM network model is associated with time points.
[0018] Further, after collecting the fault data of the electric drive system in the fault state, the same normalization preprocessing operation as the operating data is performed on the fault data;
[0019] The process of data enhancement processing for small-sample fault data through a time sliding window includes: inputting preprocessed fault data, setting the window size and moving step of the time sliding window, and controlling the time sliding window to extract multiple samples from the sequence of fault data according to the window size and step, and adding time series context relationships to the samples.
[0020] Furthermore, the process of performing transfer learning on the pre-trained LSTM network model to obtain a fault prediction model includes:
[0021] The encoder part in the pre-trained LSTM network model is retained, multiple LSTM layers and a regression output layer are added on the basis of the encoder part, and the Relu function is used as the activation function to obtain the fault prediction model.
[0022] Furthermore, after outputting the predicted fault curve, the predicted fault curve is compared with the real fault curve to evaluate the effectiveness of the model.
[0023] A second aspect of the present invention provides an electronic device, comprising a processor and a memory that is communicatively connected to the processor and is used to store executable instructions of the processor, characterized in that the processor is used to execute the remaining life prediction method of the electric drive system with few-sample data enhancement and migration as described in any of the first aspects above.
[0024] The third aspect of the present invention provides a server, characterized in that it includes at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, and the instructions are executed by the processor so that the at least one processor executes the remaining life prediction method of the electric drive system with few-sample data enhancement and migration as described in any one of the first aspects.
[0025] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, characterized in that when the computer program is executed by a processor, it implements the remaining life prediction method of the electric drive system with few-sample data enhancement and migration as described in any one of the first aspects.
[0026] Compared with the prior art, the electric drive system remaining life prediction method with few-sample data enhancement and migration described in the present invention has the following beneficial effects:
[0027] It can efficiently use less data for training and can more accurately identify and predict potential faults of motor assemblies compared to traditional models;
[0028] By accurately predicting potential failures of the electric drive assembly, it helps to improve the reliability and performance of the entire drive system, which is especially important for ensuring safe operation and meeting high standards of performance requirements.
[0029] The method involved in the present invention uses a large amount of normal operation data as a source domain to train the model, which can provide a reference for fault prediction in other component fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:
[0031] Figure 1 A schematic flow chart of a method for predicting the remaining life of an electric drive system using a small number of sample data enhancement and migration according to an embodiment of the present invention;
[0032] Figure 2 This is a schematic diagram of the LSTM network model structure according to an embodiment of the present invention;
[0033] Figure 3This is a schematic diagram of the fault prediction curve of a prototype machine and the actual fault life curve of the three LSTM models described in an embodiment of the present invention. DETAILED DESCRIPTION
[0034] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0035] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.
[0036] Embodiment 1:
[0037] like Figure 1 As shown: A method for predicting the remaining life of an electric drive system based on few-sample data enhancement and migration, including:
[0038] S1. Collect the operating data of the electric drive system under normal working conditions, and use the operating data to pre-train the LSTM network model;
[0039] S2. Collect fault data of the electric drive system in a fault state, perform feature label calibration on the fault data, and perform data enhancement processing on the small sample fault data through a time sliding window;
[0040] S3. Perform transfer learning on the pre-trained LSTM network model to obtain a fault prediction model, and use the enhanced fault data to train the fault prediction model;
[0041] S4. Use the trained fault prediction model to perform fault prediction and output a predicted fault curve.
[0042] The principle and training objectives of pre-training the LSTM network model using running data are shown in the following formula:
[0043] The encoder hidden state is:
[0044]
[0045] The decoder hidden state is:
[0046]
[0047] The autoencoder loss function is:
[0048]
[0049] in, It is the normal working data of the source domain. is the hidden state of the encoder at time t-1, is the hidden state of the decoder at time t-1.
[0050] The objective function of the encoder is:
[0051]
[0052] For each source domain Calculate the square of the reconstruction error between the encoder and decoder in the LSTM autoencoder And make it minimal.
[0053] The process of collecting the operating data of the electric drive system under normal working conditions includes:
[0054] Five drive devices with different powers in the electric drive system were selected as prototypes, and the operating data of the five prototypes under the same working conditions and normal working conditions were collected; the operating data included output current, temperature, and housing vibration data; the electric drive system electrical signal, temperature signal, and mechanical signal were used as characteristic data to fully reflect the operating status of the system;
[0055] The collected data are normalized and preprocessed after unifying the time scale to ensure consistent time standards during analysis; the formula for processing the data by the normalization preprocessing method (Min-Max-scaler) is:
[0056] Among them, x is the original record point of a feature of the sample, min(x) is the minimum value in the data set, max(x) is the maximum value in the data set, and X norm is the normalized value.
[0057] The LSTM network model includes an encoder and a decoder. A repeated vector layer is provided between the encoder and the decoder, and a time-distributed connection layer is provided at the output end of the decoder.
[0058] The encoder calculates the forget gate f t , input gate i t , output gate o t And cell state c t The activation vector is forward propagated through the LSTM unit to obtain the hidden state h at each time step t , and use the hidden state as the context to represent the cell state;
[0059] The decoder receives the hidden state corresponding to the last time step as the initial state, uses the same LSTM unit as the encoder, calculates the hidden state and cell state of the decoder and generates the reconstructed output sequence x t .
[0060] The specific process includes: using the Autoencoder network structure to embed into the LSTM network, learning the normal data set, extracting important features, and characterizing the normal input data of different machines;
[0061] Assume there is a sequence input X = (x1, x1...x t ), where x t is the input vector at time step t, and the encoder forward propagates through the LSTM unit to calculate the hidden state h at each time step t :
[0062] The mathematical explanation formula of the automatic encoder and decoder algorithm is as follows:
[0063] f t =σ(W f ·[h t-1 , x t ]+b f );
[0064] i t =σ(W i ·[h t-1 , x t ]+b i );
[0065] o t =σ(W o ·[h t-1 , x t ]+b o );
[0066]
[0067] h t =o t ☉tanh(c t );
[0068] Among them, f t ,i t , o t are the activation vectors of the forget gate, input gate, and output gate, σ represents the sigmoid function, ⊙ represents the Hadamard product, W and b are the weight matrix and bias vector. The final state of the encoder h t is used as context to represent the cell state c t .
[0069] The decoder part is used to reconstruct the input sequence X. It receives the final hidden state h from the encoder. t as the initial state and try to reconstruct the original input X;
[0070] Using a similar LSTM cell as the encoder, but with the initial hidden state and cell state initialized from the context provided by the encoder, for each time step t the decoder’s hidden state is and cell status is updated, similar to the way the encoder is updated, using the decoder's hidden state to generate the reconstructed output sequence x t , the specific formula is as follows:
[0071]
[0072] The ultimate goal is to make the reconstructed sequence as close as possible to the original acquired data X;
[0073] Specifically, the structure of the encoder selects two LSTM layers, the first LSTM layer selects 64 units, the second LSTM layer selects 32 units, and the activation function selects the Relu function;
[0074] Add a Repeatvector layer between the encoder and decoder to repeat the encoder's final output multiple times in order to keep the input within each step of the decoder;
[0075] For the decoder structure, two LSTM layers are selected. The first LSTM layer selects 32 units, the second LSTM layer selects 64 units, and the activation function also selects the Relu function. Finally, the TimeDistributed connection layer is used as the output layer to complete the construction of the autoencoder network.
[0076] In the process of pre-training the LSTM network model using running data, the Adam optimizer is used as the optimizer, the adaptive learning rate method is used, β is selected as 0.04, epoch is selected as 50, and batch-size is selected as 32; the mean square error (MSE) is used as the training error function, and the obtained parameter weights are prepared for transfer learning.
[0077] The process of calibrating the feature labels of fault data includes: setting label points in the fault data based on the time of fault occurrence, calculating the absolute value of the difference between the time of each label point and the time of fault occurrence, and obtaining the remaining time; inputting the remaining time into the LSTM network model as a feature label, and associating the fault data input into the LSTM network model with time points.
[0078] After collecting the fault data of the electric drive system in a fault state, the same normalization preprocessing operation as the operating data is performed on the fault data; the process of performing data enhancement processing on the few-sample fault data through a time sliding window includes: inputting the preprocessed fault data, setting the window size and moving step of the time sliding window, and controlling the time sliding window according to the window size and step to extract multiple samples in the sequence of fault data, and adding time series context relationships to the samples.
[0079] The specific setting method of the time sliding window method is as follows:
[0080] Let the original input be X = (x1, x2, ..., x N ), where N is the total length of the sequence, and a window size W and step size S are defined, then M samples can be extracted. The definition of each sample is as follows: i =(x (i-1)S+1 , x (i-1)S+2 , ..., x (i-1)S+W ); wherein, i=1, 2, ..., M; in this embodiment, the window size W is set to 3, the step size S of the window movement is set to 2, and the extracted data retains the contextual relationship of the time series.
[0081] The process of performing transfer learning on the pre-trained LSTM network model to obtain a fault prediction model includes:
[0082] The encoder part of the pre-trained LSTM network model is retained to ensure that it is exactly the same as the new model in the architecture of the shared layer and the weights are correctly matched. Two LSTM layers are added on the basis of the encoder part. The number of nodes in the first LSTM layer is selected as 16 and the number of nodes in the second LSTM layer is selected as 64. A regression output layer is added and the Relu function is used as the activation function to obtain the fault prediction model.
[0083] The target domain deep learning model iterative update target of the encoder's transfer learning is as follows:
[0084]
[0085] The adaptive learning rate is defined as:
[0086] Where β is a hyperparameter for adjusting the learning rate.
[0087] The model was trained using the Adam optimizer as the gradient descent method and MSE as the error function with 1000 epoch iterations.
[0088] After outputting the predicted failure curve, the predicted failure curve is compared with the actual failure curve to evaluate the effectiveness of the model.
[0089] Embodiment 2:
[0090] An electronic device comprises a processor and a memory connected to the processor for storing instructions executable by the processor, wherein the processor is used to execute the method for predicting the remaining life of an electric drive system by enhancing and migrating a small number of sample data as described in any one of the above-mentioned embodiments.
[0091] Embodiment three:
[0092] A server, characterized in that it includes at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor so that the at least one processor executes the remaining life prediction method for an electric drive system with few-sample data enhancement and migration as described in any one of the first embodiments.
[0093] Embodiment 4:
[0094] A computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the method for predicting the remaining life of an electric drive system using a small number of sample data enhancement and migration as described in any one of the first embodiments is implemented.
[0095] Those of ordinary skill in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0096] In the several embodiments provided in the present application, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the division of the units described above is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The above-mentioned units may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.
[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.
[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting the remaining life of an electric drive system based on few-sample data enhancement and migration, characterized in that: include: Collect the operating data of the electric drive system under normal working conditions and use the operating data to pre-train the LSTM network model, where the operating data includes output current, temperature, and housing vibration data; The process of collecting fault data of the electric drive system in a fault state and calibrating the fault data with feature labels includes: setting label points in the fault data based on the time of fault occurrence, calculating the absolute value of the difference between the time of each label point and the time of fault occurrence, obtaining the remaining time, inputting the remaining time as a feature label into the LSTM network model, and associating the fault data input into the LSTM network model with time points; And perform data enhancement processing on the small sample fault data through time sliding window; Perform transfer learning on the pre-trained LSTM network model to obtain the remaining life prediction model, and use the enhanced fault data to train the remaining life prediction model; Use the trained remaining life prediction model to predict the remaining life and output the predicted remaining life curve; The LSTM network model includes an encoder and a decoder. A repeated vector layer is provided between the encoder and the decoder, and a time-distributed connection layer is provided at the output end of the decoder. The encoder calculates the activation vectors of the forget gate, input gate, output gate, and cell state, forward propagates through the LSTM unit, obtains the hidden state of each time step, and uses the hidden state as the context to represent the cell state; The decoder receives the hidden state corresponding to the last time step as the initial state, uses the same LSTM unit as the encoder, calculates the hidden state and cell state of the decoder and generates a reconstructed output sequence ; The specific process includes: using the Autoencoder network structure to embed into the LSTM network, learning the normal data set, extracting important features, and characterizing the normal input data of different machines; Among them, the sequence input is X=(x1,x2...x t ),in is the input vector at time step t, and the encoder forward propagates through the LSTM unit to calculate the hidden state at each time step : The mathematical explanation formula of the automatic encoder and decoder algorithm is as follows: ； ； ； ； ； ； in, , , are the activation vectors of the forget gate, input gate, and output gate, respectively. represents the sigmoid function, represents the Hadamard product, W and b are the weight matrix and bias vector, and the final hidden state of the encoder Used as context to represent cell state ; The decoder part is used to reconstruct the sequence input X, which receives the final hidden state from the encoder. As the initial state, and try to reconstruct the sequence input X; Uses a similar LSTM unit as the encoder, but the initial hidden state and cell state are initialized by the context provided by the encoder. For each time step t, the decoder hidden state is and cell status is updated, similar to the way the encoder is updated, using the decoder’s hidden state to generate the reconstructed output sequence , the specific formula is as follows: ； ； The ultimate goal is to make the reconstructed sequence as close to the sequence input as possible; Specifically, the structure of the encoder selects two LSTM layers, the first LSTM layer selects 64 units, the second LSTM layer selects 32 units, and the activation function selects the Relu function; Add a Repeatvector layer between the encoder and decoder to repeat the encoder's final output multiple times in order to keep the input within each step of the decoder; For the decoder structure, two LSTM layers are selected. The first LSTM layer has 32 units and the second LSTM layer has 64 units. The activation function also uses the Relu function. Finally, the TimeDistributed connection layer is used as the output layer to complete the construction of the autoencoder network. In the process of pre-training the LSTM network model using running data, the Adam optimizer is used as the optimizer, the adaptive learning rate is used, β is selected as 0.04, epoch is selected as 50, and batch-size is selected as 32; the mean square error is used as the training error function, and the obtained parameter weights are prepared for transfer learning; The principle and training objectives of pre-training the LSTM network model using running data are shown in the following formula: The encoder hidden state is: ； The decoder hidden state is: ； The autoencoder loss function is: ； in, It is the normal working data of the source domain. is the hidden state of the encoder at time t-1, is the hidden state of the decoder at time t-1; The objective function of the encoder is: ； For each source domain , calculate the square of the reconstruction error between the encoder and decoder in the LSTM autoencoder , and minimize it; After collecting the fault data of the electric drive system in the fault state, the fault data is subjected to the same normalization preprocessing operation as the running data; the process of performing data enhancement processing on the few-sample fault data through the time sliding window includes: inputting the preprocessed fault data, setting the window size and moving step of the time sliding window, and controlling the time sliding window to extract multiple samples from the sequence of fault data according to the window size and step, and adding the context relationship of the time series to the samples; The specific setting method of the time sliding window method is as follows: Let the sequence input be , where N is the total length of the sequence, defining a window size W and a step size S, then M samples can be extracted. , the definition of each sample is as follows: ; Where i=1, 2, ..., M; Where the window size W is set to 3, the step size S of the window movement is set to 2, and the extracted data retains the contextual relationship of the time series; The process of performing transfer learning on the pre-trained LSTM network model to obtain a fault prediction model includes: The encoder part of the pre-trained LSTM network model is retained to ensure that it is exactly the same as the new model in the architecture of the shared layer and the weights are correctly matched. Two LSTM layers are added on the basis of the encoder part. The number of nodes in the first LSTM layer is 16, and the number of nodes in the second LSTM layer is 64. A regression output layer is added and the Relu function is used as the activation function to obtain the fault prediction model. The target domain deep learning model iterative update target of the encoder's transfer learning is as follows: ； The adaptive learning rate is defined as: ; Where β is a hyperparameter for adjusting the learning rate; The target domain deep learning model iterative update target of the encoder's transfer learning is as follows: ； The adaptive learning rate is defined as: ; Where β is a hyperparameter for adjusting the learning rate; The model was trained using the Adam optimizer as the gradient descent method and MSE as the error function with 1000 epoch iterations. After outputting the predicted remaining life curve, the predicted remaining life curve is compared with the actual remaining life curve to evaluate the effectiveness of the model.
2. The method for predicting the remaining life of an electric drive system using a small number of sample data enhancement and migration according to claim 1, characterized in that: The process of collecting the operation data of the electric drive system under normal working conditions includes: selecting a plurality of drive devices with different powers in the electric drive system as prototypes, and collecting the operation data of the plurality of prototypes under the same working conditions and normal working conditions; The collected data are normalized and preprocessed after unifying the time scale.
3. An electronic device, comprising a processor and a memory connected to the processor for storing instructions executable by the processor, characterized in that: The processor is used to execute the electric drive system remaining life prediction method with few-sample data enhancement and migration as described in any one of claims 1-2 above.
4. A server, characterized in that: It includes at least one processor and a memory communicatively connected to the processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the processor so that the at least one processor executes the remaining life prediction method of the electric drive system with few-sample data enhancement and migration as described in any one of claims 1-2.
5. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for predicting the remaining life of an electric drive system by enhancing and migrating small sample data as described in any one of claims 1-2 is implemented.