Transfer learning method and device

Abstract

The present application discloses a transfer learning algorithm and device, which includes constructing a source model for transfer learning; the source model is constructed by a single or multiple deep learning networks; obtaining the original signal of the electromechanical equipment, and constructing an intermediate variable based on the original signal; wherein the intermediate variable has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task; determining the label corresponding to the original signal based on the fuzzy membership function of the intermediate variable; constructing a training set with the original signal and the corresponding label to train the source model to obtain a new model for transfer learning; and performing classification and regression tasks on the electromechanical equipment based on the new model for transfer learning. The transfer learning algorithm and device can solve two different types of tasks in a deep network to improve the algorithm's migration performance between different objects, different working conditions, and different tasks.

Description

Technical Field

[0001] The present application relates to the field of transfer learning, and in particular to a transfer learning method and device thereof. Background Art

[0002] In the field of PHM, the problems that need to be solved for data-driven diagnostic methods can be divided into two categories: (1) discrete classification problems, such as fault classification, aging stage division, and finite fault state identification; (2) continuous regression problems, such as remaining useful life (RUL) estimation, wear or aging trajectory prediction, and health indicator estimation.

[0003] For these two different types of fault warning tasks, the academic and industrial communities currently usually use different methods to handle them separately, that is, two different machine learning or deep learning algorithms are used for fault diagnosis and remaining life estimation respectively, without considering the association and transfer learning between the two tasks. Although two different algorithms are trained for the two types of tasks, fault diagnosis and remaining life estimation, respectively, which are highly targeted and have high algorithm accuracy, there are two significant problems: the algorithm training efficiency is low, and repeated training consumes a lot of computing resources; in actual industrial applications, there is often not enough fault experimental data or aging trajectories of the entire life cycle to meet the needs of algorithm training for two different tasks. Summary of the invention

[0004] The present application provides a transfer learning method and device, which can solve two different types of PHM tasks in a deep learning network to improve the transfer performance of the method between different objects, different working conditions and different tasks. Through a method, it is possible to perform both classification tasks and regression tasks on electromechanical equipment.

[0005] A transfer learning method provided in this application includes:

[0006] Constructing a source model for transfer learning; the source model is constructed by a single or multiple deep learning networks;

[0007] Acquire the original signal of the electromechanical device, and construct an intermediate variable based on the original signal; wherein the intermediate variable has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task;

[0008] Determining a label corresponding to the original signal based on a fuzzy membership function of the intermediate variable;

[0009] Constructing a training set with the original signal and the corresponding label to train the source model to obtain a new model for transfer learning;

[0010] The new model based on the transfer learning performs classification and regression tasks on the electromechanical devices.

[0011] In an exemplary embodiment, the construction of a source model for transfer learning includes: constructing a topological structure of the source model; determining a hyperparameter set of the source model and a value range of each parameter in the hyperparameter set; and optimizing the topological structure and hyperparameter set of the source model.

[0012] In an exemplary embodiment, when the deep learning network includes a deep convolutional neural network and a short-term long-term memory network, the topological structure of constructing the source model includes a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in series, and a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in parallel.

[0013] In an exemplary embodiment, the optimization of the topological structure and hyperparameter set of the source model includes: determining a cost function for optimization; optimizing the topological structure of the source model and the values ​​of hyperparameters in the hyperparameter set according to a preset optimization algorithm and the cost function.

[0014] In an exemplary embodiment, constructing an intermediate variable based on the original signal includes the following steps:

[0015] Step 1: extracting and selecting features from the original signal;

[0016] Step 2: Construct the first variable based on the selected features;

[0017] Step 3: Draw a probability distribution histogram of the first variable in each category of the classification task;

[0018] Step 4: judging whether the first variable meets the first preset condition according to the probability distribution histogram;

[0019] If the first preset condition is met, the first variable is used as an intermediate variable; if the first preset condition is not met, steps 2 to 4 are repeated until the first variable meets the first preset condition;

[0020] The first preset condition is that the system has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task.

[0021] In an exemplary embodiment, determining a label corresponding to the original signal based on a fuzzy membership function of the intermediate variable includes:

[0022] Constructing a fuzzy membership function of the intermediate variable based on the mean and standard deviation of the intermediate variable distributed under each category of the classification task;

[0023] A label corresponding to the original signal is determined according to the fuzzy membership function.

[0024] In an exemplary embodiment, determining a label corresponding to the original signal according to the fuzzy membership function includes:

[0025] Divide the value of the intermediate variable into N intervals at equal intervals, wherein the intervals of the intervals meet the second preset condition; N is a positive integer;

[0026] For each interval, do the following:

[0027] Count the number of samples in the interval and calculate the average membership value in the interval; determine the label of each sample in the interval according to the number of samples in the interval and the average membership value in the interval;

[0028] The label of each sample in the N intervals is determined as the label corresponding to the original signal.

[0029] In an exemplary embodiment, the label of each sample in the interval is determined according to the number of samples in the interval and the average membership value of the interval, including: assuming that the label of the interval determined according to the fuzzy membership function graph includes a first category and a second category; then the labels of the first number of samples in the interval are determined as the first category; the labels of the remaining samples after removing the first number are determined as the second category; wherein the first number is determined according to the product of the number of samples in the interval and the average membership value of the interval.

[0030] In an exemplary embodiment, the feature extraction and feature selection of the original signal includes: extracting time domain features and frequency domain features of the original signal; calculating the correlation and monotonicity of the extracted features; selecting M features whose correlation and monotonicity meet the third preset condition; wherein M is an integer greater than 1;

[0031] Constructing the first variable according to the selected features includes:

[0032] The variable obtained by linearly weighted summing the selected M features is used as the first variable.

[0033] The present application provides a transfer learning device, including a memory and a processor.

[0034] The memory is used to store a program for the transfer learning method;

[0035] The processor is used to read and execute the program for the transfer learning method, and execute the following method:

[0036] Constructing a source model for transfer learning; the source model is constructed by a single or multiple deep learning networks;

[0037] Acquire the original signal of the electromechanical device, and construct an intermediate variable based on the original signal; wherein the intermediate variable has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task;

[0038] Determining a label corresponding to the original signal based on a fuzzy membership function of the intermediate variable;

[0039] Constructing a training set with the original signal and the corresponding label to train the source model to obtain a new model for transfer learning;

[0040] The new model based on the transfer learning performs classification and regression tasks on the electromechanical devices.

[0041] In an exemplary embodiment, the construction of a source model for transfer learning includes: constructing a topological structure of the source model; determining a hyperparameter set of the source model and a value range of each parameter in the hyperparameter set; and optimizing the topological structure and hyperparameter set of the source model.

[0042] In an exemplary embodiment, when the deep learning network includes a deep convolutional neural network and a short-term long-term memory network, the topological structure of constructing the source model includes a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in series, and a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in parallel.

[0043] In an exemplary embodiment, the optimization of the topological structure and hyperparameter set of the source model includes: determining a cost function for optimization; optimizing the topological structure of the source model and the values ​​of hyperparameters in the hyperparameter set according to a preset optimization algorithm and the cost function.

[0044] In an exemplary embodiment, constructing an intermediate variable based on the original signal includes the following steps:

[0045] Step 1: extracting and selecting features from the original signal;

[0046] Step 2: Construct the first variable based on the selected features;

[0047] Step 3: Draw a probability distribution histogram of the first variable in each category of the classification task;

[0048] Step 4: judging whether the first variable meets the first preset condition according to the probability distribution histogram;

[0049] If the first preset condition is met, the first variable is used as an intermediate variable; if the first preset condition is not met, steps 2 to 4 are repeated until the first variable meets the first preset condition;

[0050] The first preset condition is that the system has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task.

[0051] In an exemplary embodiment, determining a label corresponding to the original signal based on a fuzzy membership function of the intermediate variable includes:

[0052] Constructing a fuzzy membership function of the intermediate variable based on the mean and standard deviation of the intermediate variable distributed under each category of the classification task;

[0053] A label corresponding to the original signal is determined according to the fuzzy membership function.

[0054] In an exemplary embodiment, determining a label corresponding to the original signal according to the fuzzy membership function includes:

[0055] Divide the value of the intermediate variable into N intervals at equal intervals, wherein the intervals of the intervals meet the second preset condition; N is a positive integer;

[0056] For each interval, do the following:

[0057] Count the number of samples in the interval and calculate the average membership value in the interval; determine the label of each sample in the interval according to the number of samples in the interval and the average membership value in the interval;

[0058] The label of each sample in the N intervals is determined as the label corresponding to the original signal.

[0059] In an exemplary embodiment, the label of each sample in the interval is determined according to the number of samples in the interval and the average membership value of the interval, including: assuming that the label of the interval determined according to the fuzzy membership function graph includes a first category and a second category; then the labels of the first number of samples in the interval are determined as the first category; the labels of the remaining samples after removing the first number are determined as the second category; wherein the first number is determined according to the product of the number of samples in the interval and the average membership value of the interval.

[0060] In an exemplary embodiment, the feature extraction and feature selection of the original signal includes: extracting time domain features and frequency domain features of the original signal; calculating the correlation and monotonicity of the extracted features; selecting M features whose correlation and monotonicity meet the third preset condition; wherein M is an integer greater than 1;

[0061] Constructing the first variable according to the selected features includes:

[0062] The variable obtained by linearly weighted summing the selected M features is used as the first variable.

[0063] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by implementing the present application. Other advantages of the present application can be realized and obtained by the schemes described in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide an understanding of the technical solution of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application and do not constitute a limitation on the technical solution of the present application.

[0065] Figure 1 A flow chart of the transfer learning method of an embodiment of the present application;

[0066] Figure 2 A flowchart of an example of a transfer learning method according to an embodiment of the present application;

[0067] Figure 3 A flowchart of constructing a transfer learning source model according to an embodiment of the present application;

[0068] Figure 4 A flowchart of constructing a training set label based on a fuzzy membership function according to an embodiment of the present application;

[0069] Figure 5 A probability distribution histogram of the intermediate variable HI in the embodiment of the present application;

[0070] Figure 6 A schematic diagram of a Gaussian fuzzy membership function designed based on the probability distribution of intermediate variables according to an embodiment of the present application;

[0071] Figure 7 The training and testing process of the transfer learning source model in task migration according to the embodiment of the present application;

[0072] Figure 8 Schematic diagram of a transfer learning device according to an embodiment of the present application. DETAILED DESCRIPTION

[0073] Figure 1 Flow chart of the transfer learning method of the present application embodiment, such as Figure 1 As shown, the transfer learning method of this embodiment includes steps A11-A15:

[0074] A11. Build a source model for transfer learning.

[0075] A12. Acquire original signals of electromechanical equipment, and construct intermediate variables based on the original signals;

[0076] A13. Determine a label corresponding to the original signal based on the fuzzy membership function of the intermediate variable;

[0077] A14, constructing a training set with the original signal and the corresponding label to train the source model to obtain a new model for transfer learning;

[0078] A15. Perform classification and regression tasks on electromechanical equipment based on the new model of transfer learning.

[0079] In an exemplary embodiment, the collected original signal may be, for example, acceleration.

[0080] In an exemplary embodiment, the intermediate variable has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task, wherein the output of the classification task is a discrete variable; and the output of the regression task is a continuous variable.

[0081] In an exemplary embodiment, the construction of a source model for transfer learning may include: constructing a topological structure of the source model; determining a hyperparameter set of the source model and a value range of each parameter in the hyperparameter set; and optimizing the topological structure and hyperparameter set of the source model.

[0082] In an exemplary embodiment, since structure and hyperparameters are two important factors affecting the performance of deep learning algorithms, the deep convolutional neural network CNN (Convolutional Neural Network) can capture signal features from different spatial scales, and the short-term long-term memory network LSTM (Long Short Term Memory) can extract the intrinsic relationship of input signals from different time scales. The combination of the two can theoretically enhance the network's ability to capture fault features, thereby improving the effectiveness of the fault diagnosis method. Therefore, the embodiment of the present application can combine CNN and LSTM and use them as the framework of the transfer learning source model, while using the combination mode (series, parallel) and hyperparameters of the two as optimizable parameters. In some other embodiments, other deep learning networks can also be used for combination.

[0083] When the deep learning network includes a deep convolutional neural network and a short-term long-term memory network, the topological structure of constructing the source model includes a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in series, and a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in parallel.

[0084] In an exemplary embodiment, the optimization of the topological structure and hyperparameter set of the source model may include: determining a cost function for optimization; optimizing the topological structure of the source model and the values ​​of hyperparameters in the hyperparameter set according to a preset optimization algorithm and the cost function.

[0085] In an exemplary embodiment, constructing an intermediate variable based on the original signal includes the following steps:

[0086] Step 1: extracting and selecting features from the original signal;

[0087] Step 2: Construct the first variable based on the selected features;

[0088] Step 3: Draw a probability distribution histogram of the first variable in each category of the classification task;

[0089] Step 4: judging whether the first variable meets the first preset condition according to the probability distribution histogram;

[0090] If the first preset condition is met, the first variable is used as an intermediate variable; if the first preset condition is not met, steps 2 to 4 are repeated until the first variable meets the first preset condition;

[0091] The first preset condition is that the system has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task.

[0092] In an exemplary embodiment, determining a label corresponding to the original signal based on a fuzzy membership function of the intermediate variable includes:

[0093] Constructing a fuzzy membership function of the intermediate variable based on the mean and standard deviation of the intermediate variable distributed under each category of the classification task;

[0094] A label corresponding to the original signal is determined according to the fuzzy membership function.

[0095] In an exemplary embodiment, determining a label corresponding to the original signal according to the fuzzy membership function includes:

[0096] Divide the value of the intermediate variable into N intervals at equal intervals, wherein the intervals of the intervals meet the second preset condition; N is a positive integer;

[0097] For each interval, do the following:

[0098] Count the number of samples in the interval and calculate the average membership value in the interval; determine the label of each sample in the interval according to the number of samples in the interval and the average membership value in the interval;

[0099] The label of each sample in the N intervals is determined as the label corresponding to the original signal.

[0100] In an exemplary embodiment, the second preset condition may be that the interval is as small as possible to improve the resolution of the intermediate variable, and at the same time the interval is large enough to satisfy that within the intervals of all intermediate variables, the number of samples multiplied by the membership degree corresponding to the corresponding interval is greater than or equal to 1.

[0101] In an exemplary embodiment, determining the label of each sample in the interval according to the number of samples in the interval and the average membership value of the interval includes:

[0102] Assume that the label of the interval determined according to the fuzzy membership function graph includes a first category and a second category;

[0103] The labels of the first number of samples in the interval are determined as the first category; the labels of the remaining samples after removing the first number are determined as the second category;

[0104] The first number is determined according to the product of the number of samples in the interval and the average membership value of the interval.

[0105] In an exemplary embodiment, performing feature extraction and feature selection on the original signal includes:

[0106] Extracting time domain features and frequency domain features of the original signal;

[0107] Calculate the correlation and monotonicity of the extracted features;

[0108] Select M features whose correlation and monotonicity meet the third preset condition; where M is an integer greater than 1;

[0109] Constructing the first variable according to the selected features includes:

[0110] The variable obtained by linearly weighted summing the selected M features is used as the first variable.

[0111] In an exemplary embodiment, the third preset condition may include features with high correlation and good monotonicity. Correlation is used to measure the strength and direction of the linear relationship between the extracted features and the final output (such as fault category or remaining life), and monotonicity is used to quantify the monotonicity of the trend of the extracted features over time. The stronger the strength of the linear relationship, the higher the correlation, and the higher the monotonicity, the better the monotonicity.

[0112] In an exemplary embodiment, the intermediate variable may be a universal aging index HI (Health Index).

[0113] In an exemplary embodiment, the input of the training set is the collected original signal; the output of the training set is the label corresponding to the input of the training set.

[0114] The embodiment of the present application proposes a new transfer learning framework: on the one hand, a single or multiple deep learning networks (such as deep convolutional neural networks and short-term long-term memory networks) are combined to construct a source model for transfer learning, and the two extract fault features from different angles (such as spatial scale and time scale); on the other hand, the intermediate variables (such as aging index HI) constructed based on the membership function are used as output labels, which can guide the source model to extract different (such as continuous and discrete) features. The combination of the source model structure and the intermediate variables constructed based on the membership function can effectively solve the transfer learning of diagnostic methods under different working conditions, different objects, and different tasks. It allows two completely different diagnostic tasks to be completed in one deep network, improves the training efficiency and migration performance of the diagnostic method, and saves computing resources. Through a method, it is possible to perform both classification tasks and regression tasks on electromechanical equipment.

[0115] The transfer learning method of the present application is explained below with reference to specific examples of the embodiments of the present application.

[0116] As attached Figure 2 The method flow of this application is shown in the figure. The specific implementation of this application includes three parts:

[0117] Build a source model for transfer learning;

[0118] Design training set labels based on fuzzy membership functions;

[0119] Train the source model and verify its transfer learning performance;

[0120] The following will introduce the detailed steps of each part:

[0121] 1. Build a source model for transfer learning

[0122] like Figure 3 As shown in the figure, building the source model for transfer learning consists of three steps:

[0123] Step 1.1: Design the structure of the source model

[0124] Step 1.1.1 Design the basic structures of the deep convolutional neural network (CNN) and the short-term long-term memory (LSTM) network used to build the source model for transfer learning. The structures that need to be designed for CNN include: the size and number of convolution kernels, the size and number of pooling layers, the type of activation function, and the size and number of fully connected layers; the structures that need to be designed for LSTM include: the number of LSTM neuron layers and the number of LSTM neurons in each layer, the type of activation function, and the size and number of fully connected layers.

[0125] Step 1.1.2 Design a topological structure that combines the two deep learning networks, CNN and LSTM, mainly including serial and parallel types. In the serial structure, the output of the CNN pooling layer or the fully connected layer is used as the input of the LSTM, and the output of the LSTM is used as the final output of the entire network; in the parallel structure, the output of the fully connected layer of both CNN and LSTM is sent as input to the fusion layer, and after fusion, it passes through the final fully connected layer to complete the final output.

[0126] Step 1.2: Design the hyperparameter set and range of the source model

[0127] Step 1.2.1 Select the hyperparameters that need to be optimized in the source model. Common hyperparameters include: learning rate, number of iterations, batch size, dropout method and value, activation function type, optimizer type, etc.

[0128] Step 1.2.2 determines the value range of each hyperparameter. Considering that it is difficult to converge and very time-consuming to optimize the hyperparameters within the entire continuous value range. Therefore, in this application, a discrete value domain with a reasonable value range and appropriate change step size will be designed for each hyperparameter according to the needs of the actual application scenario, and the subsequent optimization of the hyperparameters will be carried out within the corresponding discrete value domain. For example, the discrete value domain of dropout can be designed as: dropout = {0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55, 0.6, 0.65, 0.7}. The design of other hyperparameters is similar and will not be repeated here.

[0129] Step 1.3: Optimize the structure and hyperparameters of the source model

[0130] Step 1.3.1 Design a cost function for optimizing hyperparameters and source model structure. Usually, the accuracy of fault classification or the error of remaining life estimation can be selected as the cost function.

[0131] Step 1.3.2 Select the intelligent optimization algorithm and the corresponding parameters of the algorithm. Common intelligent optimization algorithms include: Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Simulated Annealing (SA), etc. Taking PSO as an example, the parameters that need to be designed include: particle population size, speed limit, position limit, inertia weight, acceleration factor, etc.

[0132] Step 1.3.3 uses the intelligent optimization algorithm selected in step 1.3.2 above to search the structure and hyperparameters of the source model within the discrete space defined in step 1.2.1, and determines the optimal model structure and hyperparameter combination based on the cost function.

[0133] 2. Design training set labels based on fuzzy membership function

[0134] like Figure 4 As shown in the figure, constructing the training set labels based on the fuzzy membership function includes five steps:

[0135] Step 2.1: Feature extraction and feature selection

[0136] Step 2.1.1 extracts features from the time domain and frequency domain of the collected original signal (such as acceleration). Common time domain features include: mean, standard deviation, maximum value, minimum value, kurtosis, skewness, crest factor, shape factor, etc.; common frequency domain indicators include: peak frequency, fault characteristic frequency, wavelet energy entropy, wavelet singular entropy, etc. The specific feature extraction method is not the focus of this application and will not be repeated here.

[0137] Step 2.2.2 For the above extracted features, use correlation and monotonicity evaluation indicators to select several features with high correlation and good monotonicity from all the extracted features for subsequent construction of intermediate variables. Among them, correlation is used to measure the strength and direction of the linear relationship between the extracted features and the final output (fault category or remaining life); monotonicity is used to quantify the monotonicity of the trend of the extracted features over time. The calculation formula for monotonicity is as follows:

[0138]

[0139]

[0140] The correlation is calculated as follows:

[0141]

[0142] Step 2.2: Construct the intermediate variable HI

[0143] For the above-mentioned indicators selected based on monotonicity and correlation, the principal component analysis PCA method is used to further reduce the dimension of the selected features, thereby constructing an intermediate variable (i.e., the above-mentioned universal aging index HI). This variable is essentially the linear weighted sum of the selected indicators. The calculation formula of HI is as follows:

[0144]

[0145] in, ω = [ ω 1 , ω 2 , …, ω n ] is the weight coefficient, x = [ x 1 , x 2 , …, x n ] The features that are selected.

[0146] Step 2.3: Analyze the probability density distribution characteristics of the intermediate variable HI

[0147] Step 2.3.1: Divide the intermediate variable HI into the same number of categories according to the number of output labels in the classification task. For example, if the fault classification label is 3 categories, divide the samples into 3 categories according to the remaining life of the samples. Then, calculate the value of the intermediate variable HI based on the input features in each category.

[0148] Step 2.3.2 Draw the probability distribution histogram of the intermediate variable HI in each category (such as Figure 5 ), and analyze the results. Check whether HI has good clustering within the class, whether it has dispersion between different classes, and whether HI can continuously transition between two adjacent classes. If the above conditions are met, proceed to the next step; if not, return to step 2.1 and repeat steps 2.1-2.3 until a suitable intermediate variable HI is found.

[0149] Step 2.4: Design fuzzy membership function

[0150] Step 2.4.1 Analyze the probability distribution histogram of the intermediate variable HI in each category and calculate the mean of the distribution of HI in each category and standard deviation , , is the number of categories of HI.

[0151] Step 2.4.2 The mean of HI in each category obtained based on the above analysis and standard deviation , design the fuzzy membership function. Common fuzzy membership functions include: triangle membership function, trapezoidal membership function, and Gaussian membership function. Taking the Gaussian membership function as an example, its formula is as follows:

[0152]

[0153] Therefore, the mean of HI in different categories can be directly and standard deviation Directly substitute into the above formula, we can get the fuzzy membership function of HI in each category (such as Figure 6 Similarly, the other two types of membership functions can be used with slight modifications, which will not be described here.

[0154] Step 2.5: Construct training set labels based on fuzzy membership function

[0155] Step 2.5.1 Discrete the intermediate variable HI into several equally spaced small intervals within the value range, that is, and maximum value Between As interval, construct an arithmetic progression. Among them, the interval length The selection of should meet two conditions: ① The interval interval should be as small as possible to improve the resolution of HI; ② The interval interval should be large enough to ensure that in all HI intervals, the number of samples multiplied by the membership degree corresponding to the corresponding interval is greater than or equal to 1.

[0156] Step 2.5.2 Count each HI interval The number of samples within , and the average membership value corresponding to the interval ,in, ; floor is the floor function.

[0157] +1

[0158] Step 2.5.3 For each interval of HI, based on the number of samples in the interval and the average membership value of the interval , complete the labeling based on the membership function. The specific process is as follows: Assume that the interval The corresponding two adjacent HI categories are A and B, then the interval samples are labeled A, and the rest are Label each sample with B. All samples in all intervals of HI are labeled with output labels according to the above process. Applying this method to the training set can obtain the label of the training set constructed based on the fuzzy membership function.

[0159] 3. Train the source model and verify its transfer learning performance

[0160] like Figure 7 As shown in the figure, the training and testing of the transfer learning source model consists of five steps:

[0161] Step 3.1: Design classification and regression tasks for transfer learning

[0162] Step 3.1.1 Design a classification task for transfer learning. The output of the classification task is a discrete variable, so the classification task can be designed as: fault classification, aging stage division, etc. Among them, the input of the classification task is the collected state signal, and the output is the state number of the fault category or aging stage.

[0163] Step 3.1.2 Design a regression task for transfer learning. The output of the regression task is a continuous variable, so the regression task can be designed as: remaining life estimation, wear prediction, etc. Among them, the input of the regression task is the collected state signal, and the output is the remaining life or wear value.

[0164] Step 3.2: Prepare training and test sets

[0165] Step 3.2.1 Prepare the training set of the transfer learning source model. Cut a certain proportion (such as 30%) of all samples as the training set. The input of the training set is the collected state signal, and the output is the label constructed based on the fuzzy membership function. The input and output data pairs together constitute the training set.

[0166] Step 3.2.2 Prepare the test set of the transfer learning source model. Collect the data for the classification task and regression task determined in step 3.1, including input data and output data, and use the input and output data pairs of the two tasks as the test data sets for the two types of tasks.

[0167] Step 3.3: Train the source model on the training set

[0168] Train the transfer learning source model determined in step 1 on the training set determined in step 3.2.1, and fine-tune the hyperparameters of the source model until the cost function of the source model on the training set converges. Finally, freeze the structure and hyperparameters of the source model.

[0169] Step 3.4: Validate the source model on the test set

[0170] Step 3.4.1 directly migrates the frozen source model to the classification task, tests it on the test set corresponding to the classification task, and calculates the classification accuracy of fault classification / aging state recognition.

[0171] Step 3.4.2 directly migrates the frozen source model to the regression task, tests it on the test set corresponding to the regression task, and calculates the error of the remaining life estimation / wear prediction.

[0172] Step 3.5: Evaluate the performance of task transfer learning of the source model

[0173] The results of the above transfer learning source model on the classification task test set and the regression task test set are compared and evaluated with the results obtained by other transfer learning methods on the same test set. The evaluation indicators include: training time of the transfer learning source model, training set accuracy, test set accuracy, algorithm stability, algorithm robustness, etc.

[0174] Compared with the prior art, the beneficial effects of this application are:

[0175] The training set labels for transfer learning source model training are constructed based on the membership theory of fuzzy logic, which are both discrete and continuous, thereby unifying the currently commonly used discrete fault classification labels and continuous remaining life into one indicator.

[0176] A new transfer learning framework is proposed: on the one hand, the deep convolutional neural network and the short-term long-term memory network are combined to construct the source model of transfer learning, which extracts fault features from different spatial scales and time scales respectively; on the other hand, the aging index HI constructed based on the membership function is used as the output label, which can guide the source model to extract continuous and discrete features. The combination of the source model structure and the new aging index can effectively solve the transfer learning of diagnostic methods between different working conditions, different objects, and different tasks.

[0177] Figure 8 The transfer learning device of the embodiment of the present application includes a memory and a processor.

[0178] The memory is used to store a program for the transfer learning method;

[0179] The processor is used to read and execute the program for the transfer learning method to execute the above-mentioned transfer learning method.

[0180] The present application describes multiple embodiments, but the description is exemplary rather than restrictive, and it is obvious to those skilled in the art that there may be more embodiments and implementations within the scope of the embodiments described in the present application. Although many possible feature combinations are shown in the drawings and discussed in the specific embodiments, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with any other feature or element in any other embodiment, or may replace any other feature or element in any other embodiment.

[0181] It will be appreciated that any feature shown and / or discussed in this application may be implemented individually or in any appropriate combination.

[0182] In addition, when describing representative embodiments, the specification may have presented the method and / or process as a specific sequence of steps. However, to the extent that the method or process does not rely on the specific order of steps described herein, the method or process should not be limited to the steps in the specific order described. As will be appreciated by those of ordinary skill in the art, other sequences of steps are also possible.

[0183] It will be appreciated by those skilled in the art that all or some of the steps, systems, and functional modules / units in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. In hardware implementations, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or transient medium). As is known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

Claims

1. A transfer learning method, characterized in that: Constructing a source model for transfer learning; the source model is constructed by a single or multiple deep learning networks; Acquire the original acceleration signal of the electromechanical device, and construct an intermediate variable based on the original acceleration signal; wherein the intermediate variable has both the discrete clusterable characteristics required by the classification task and the continuous gradual change characteristics required by the regression task; Determining a label corresponding to the original acceleration signal based on a fuzzy membership function of the intermediate variable; Constructing a training set with the original acceleration signal and the corresponding label to train the source model to obtain a new model for transfer learning; Based on the new model of transfer learning, classification tasks and regression tasks are performed on electromechanical equipment; wherein the classification tasks include fault classification, aging stage division, and limited fault state identification; the regression tasks include remaining life estimation, wear amount prediction, aging trajectory prediction, and health index estimation; The source model for transfer learning is constructed, including: Constructing a topological structure of the source model; Determine a hyperparameter set of the source model and a value range of each parameter in the hyperparameter set; Optimizing the topology and hyperparameter set of the source model; When the deep learning network includes a deep convolutional neural network and a short-term long-term memory network, the topological structure of constructing the source model includes a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in series, and a topological structure in which the deep convolutional neural network and the short-term long-term memory network are connected in parallel; Constructing an intermediate variable based on the original acceleration signal includes the following steps: Step 1: performing feature extraction and feature selection on the original acceleration signal; Step 2: Construct the first variable based on the selected features; Step 3: Draw a probability distribution histogram of the first variable in each category of the classification task; Step 4: judging whether the first variable meets the first preset condition according to the probability distribution histogram; If the first preset condition is met, the first variable is used as an intermediate variable; if the first preset condition is not met, steps 2 to 4 are repeated until the first variable meets the first preset condition; The first preset condition is that the class has both the discrete clustering characteristics required by the classification task and the continuous gradient characteristics required by the regression task; Determining a label corresponding to the original acceleration signal based on the fuzzy membership function of the intermediate variable includes: Constructing a fuzzy membership function of the intermediate variable based on the mean and standard deviation of the intermediate variable distributed under each category of the classification task; Determine a label corresponding to the original acceleration signal according to the fuzzy membership function; Determining a label corresponding to the original acceleration signal according to the fuzzy membership function includes: Divide the value of the intermediate variable into N intervals at equal intervals, wherein the intervals of the intervals meet the second preset condition; N is a positive integer; For each interval, do the following: Count the number of samples in the interval and calculate the average membership value in the interval; determine the label of each sample in the interval according to the number of samples in the interval and the average membership value in the interval; Determine the label of each sample in the N intervals as the label corresponding to the original acceleration signal; The labels of the samples in the interval are determined according to the number of samples in the interval and the average membership value of the interval, including: Assume that the label of the interval determined according to the fuzzy membership function graph includes a first category and a second category; The labels of the first number of samples in the interval are determined as the first category; the labels of the remaining samples after removing the first number are determined as the second category; The first number is determined according to the product of the number of samples in the interval and the average membership value of the interval.

2. The transfer learning method according to claim 1, characterized in that: The optimizing the topological structure and hyperparameter set of the source model includes: Determine a cost function to be optimized; The topological structure of the source model and the values ​​of the hyperparameters in the hyperparameter set are optimized according to the preset optimization algorithm and the cost function.

3. The transfer learning method according to claim 1, wherein: Performing feature extraction and feature selection on the original acceleration signal includes: Extracting time domain features and frequency domain features of the original acceleration signal; Calculate the correlation and monotonicity of the extracted features; Select M features whose correlation and monotonicity meet the third preset condition; where M is an integer greater than 1; Constructing the first variable according to the selected features includes: The variable obtained by linearly weighted summing the selected M features is used as the first variable.

4. A transfer learning device, comprising a memory and a processor, characterized in that: The memory is used to store a program for the transfer learning method; The processor is used to read and execute the program for the transfer learning method, and execute the transfer learning method as described in any one of claims 1-3.

Images