Transfer adaptation method based on similarity explainable weight distribution for small sample

Abstract

The application discloses a transfer adaptation method based on similarity explainable weight distribution for small samples, and relates to the technical field of rotating machinery predictive maintenance. The method comprises the following steps: obtaining source domain time series and target domain time series of a rotating machinery; performing feature extraction on time domain features and / or frequency domain features of the source domain time series by using an initial feature extraction network to obtain hidden layer feature vectors output by each network layer; calculating the similarity between the hidden layer feature vectors and real values, and distributing weights to the initial feature extraction network by using an initial explainable weight distribution network based on the similarity; predicting the service life of the rotating machinery according to the feature extraction network after weight distribution and the initial explainable weight distribution network, and measuring the distribution difference between the source domain time series and the target domain time series; performing weight updating based on the predicted service life and the distribution difference, and outputting a preset service life prediction model. The application can improve the explainability of the service life prediction model of the rotating machinery.

Description

Technical Field

[0001] This application relates to the field of predictive maintenance technology for rotating machinery, and in particular to a transfer adaptation method based on similarity-interpretable weight allocation for small samples. Background Technology

[0002] Industrial equipment is increasingly trending towards larger scale, automation, and intelligence. Rotating machinery such as bearings and gears are widely used in aerospace, transportation, and other fields, and their operating conditions are directly related to the lifespan of various mechanical equipment. Therefore, lifespan prediction for rotating machinery such as bearings has significant practical implications.

[0003] In recent years, deep transfer learning has been applied in the field of health management, encompassing anomaly detection, fault diagnosis, predictive maintenance, digital twins, and data augmentation. The application of transfer learning in health management not only effectively addresses the problems of data scarcity and model generalization ability, but also improves the accuracy and efficiency of fault prediction, anomaly detection, fault diagnosis, and health assessment. This makes transfer learning an important technology in the field of health management, contributing to the realization of smarter maintenance and management strategies.

[0004] Currently, the weight allocation in traditional transfer learning methods is typically a black-box operation. However, this black-box approach lacks interpretability, making it difficult to understand the decision-making process of rotating machinery life prediction models and hindering the improvement of model prediction stability. Summary of the Invention

[0005] In view of this, this application provides a transfer adaptation method based on similarity-interpretable weight allocation for small samples. The main purpose is to improve the interpretability and decision transparency of the rotating machinery life prediction model, thereby improving the prediction stability of the model.

[0006] According to a first aspect of this application, a transfer adaptation method based on similarity-interpretable weight allocation for small samples is provided, the method comprising:

[0007] Obtain the source domain time series and target domain time series of rotating machinery, as well as an initial lifetime prediction model, wherein the initial lifetime prediction model includes an initial feature extraction network and an initial interpretable weight allocation network;

[0008] Extract the time-domain features and / or frequency-domain features corresponding to the source domain time series, and use the initial feature extraction network to extract features from the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window;

[0009] Calculate the similarity between the hidden layer feature vector and the corresponding true value, and based on the similarity, use the initial interpretable weight allocation network to assign weights to the initial feature extraction network to obtain the weighted feature extraction network;

[0010] Based on the feature extraction network after weight allocation and the initial interpretable weight allocation network, the lifespan of the rotating machinery is predicted, and the distribution difference between the source domain time series and the target domain time series is measured.

[0011] Based on the predicted lifespan of the rotating machinery and the distribution differences, the weights of the feature extraction network after weight allocation and the initial interpretable weight allocation network are updated.

[0012] Repeat the weight allocation and weight update process until the preset conditions are met, then stop iterative updates and output the preset lifetime prediction model.

[0013] According to a second aspect of this application, a transfer adaptation apparatus based on similarity-interpretable weight allocation for small samples is provided, the apparatus comprising:

[0014] An acquisition unit is used to acquire the source domain time series and target domain time series of the rotating machinery, as well as an initial lifetime prediction model, wherein the initial lifetime prediction model includes an initial feature extraction network and an initial interpretable weight allocation network.

[0015] The extraction unit is used to extract the time-domain features and / or frequency-domain features corresponding to the source domain time series, and to use the initial feature extraction network to extract the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window.

[0016] The allocation unit is used to calculate the similarity between the hidden layer feature vector and the corresponding true value, and based on the similarity, to allocate weights to the initial feature extraction network using the initial interpretable weight allocation network, thereby obtaining the weighted feature extraction network.

[0017] The prediction unit is used to predict the lifespan of the rotating machinery and measure the distribution difference between the source domain time series and the target domain time series based on the feature extraction network after weight allocation and the initial interpretable weight allocation network.

[0018] An update unit is used to update the weights of the weighted feature extraction network and the initial interpretable weight allocation network based on the predicted lifespan of the rotating machinery and the distribution differences.

[0019] The output unit is used to repeat the process of weight allocation and weight update until the preset conditions are met, at which point the iterative update stops and the preset lifetime prediction model is output.

[0020] According to a third aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described transfer adaptation method for small samples based on similarity-interpretable weight allocation.

[0021] According to a fourth aspect of this application, an electronic device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described transfer adaptation method for small samples based on similarity-interpretable weight allocation.

[0022] By employing the aforementioned technical solutions, this application provides a transfer adaptation method for small samples based on similarity-interpretable weight allocation. By calculating the similarity between the hidden feature vectors of each network layer and the true values, weights can be assigned to the initial feature extraction network using this similarity. This ensures that each weight has a clear calculation basis, thereby improving the interpretability and decision transparency of the lifespan prediction model, and relatively enhancing the model's predictive stability. Simultaneously, by introducing transfer learning and similarity calculation, this application can obtain more useful information from the source domain time series, thereby enhancing the learning effect on small sample data in the target domain, reducing the model's dependence on large-scale data, and improving the model's generalization ability.

[0023] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 The illustration shows a flowchart of a similarity-interpretable weight allocation-based transfer adaptation method for small samples provided in an embodiment of this application.

[0026] Figure 2 This paper illustrates a flowchart of another similarity-interpretable weight allocation-based transfer adaptation method for small samples provided in an embodiment of this application.

[0027] Figure 3This illustration shows a schematic diagram of the uncertainty quantification results provided in an embodiment of this application;

[0028] Figure 4 This paper illustrates a schematic diagram of the overall process of model training provided in an embodiment of this application.

[0029] Figure 5 This illustration shows a structural schematic diagram of a similarity-based interpretable weight allocation transfer adaptation device for small samples, provided in an embodiment of this application. Detailed Implementation

[0030] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0031] Traditional transfer learning methods typically employ a black-box approach to weight allocation. However, this black-box approach lacks interpretability, making it difficult to understand the decision-making process of rotating machinery life prediction models and hindering the improvement of model prediction stability.

[0032] To address the aforementioned problems, embodiments of the present invention provide a transfer adaptation method based on similarity-interpretable weight allocation for small samples, such as... Figure 1 As shown, the method includes:

[0033] Step 101: Obtain the source domain time series and target domain time series of the rotating machinery, as well as the initial life prediction model.

[0034] The initial lifetime prediction model includes an initial feature extraction network and an initial interpretable weight allocation network. The source domain time series and target domain time series contain measurement data collected in chronological order, which may specifically include one or more of the following: vibration signals, temperature signals, sound wave signals, and current signals. Rotating machinery may specifically include gears, engines, bearings, etc.

[0035] This invention is primarily applicable to scenarios where a rotating machinery life prediction model is trained based on a small sample size. The executing entity of this invention is a device or equipment capable of training a rotating machinery life prediction model based on a small sample size, specifically, it can be located on a server side.

[0036] Specifically, embodiments of the present invention can apply two well-known rotating machinery life cycle datasets, XJTU-SY and IEEE PHM Challenge 2012, using one dataset as the source domain time series for training and the other dataset as the target domain time series for testing.

[0037] Step 102: Extract the time-domain features and / or frequency-domain features corresponding to the source domain time series, and use the initial feature extraction network to extract the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window.

[0038] The time-domain features specifically include the maximum value, mean, absolute mean, root mean square, skewness, peak factor, and waveform factor of the measurement data collected in chronological order. The frequency-domain features are obtained by decomposing the signal into different sub-bands using Discrete Wavelet Transform (DWT) and calculating the energy of each sub-band, thus obtaining a time-frequency resolution table representing different frequency domain capabilities. The initial feature extraction network can be a Long Short-Term Memory (LSTM) neural network. It should be noted that this embodiment of the invention does not specifically limit the number of network layers and neurons in the LSTM neural network; these can be set according to actual business needs. For example, the LSTM neural network can be set to include three network layers, a total of 160 neurons, a learning rate of 0.002, and 1000 iterations.

[0039] In this embodiment of the invention, after extracting the time-domain features and / or frequency-domain features corresponding to the source domain time series, the time-domain features and / or frequency-domain features can be input into the LSTM neural network for feature extraction to obtain the hidden layer feature vectors of the three network layers under the corresponding time window.

[0040] In some embodiments, feature filtering can also be performed on time-domain features and / or frequency-domain features to reduce computational load while ensuring the training effect of the model. For this process, the method includes: filtering time-domain features and / or frequency-domain features that meet target characteristics from the time-domain features and / or frequency-domain features to obtain the filtered time-domain features and / or frequency-domain features, wherein the target characteristics include monotonicity, trend and predictability.

[0041] Since the extracted time-domain and / or frequency-domain features need to satisfy properties such as monotonicity, trend, and predictability, this embodiment of the invention employs a suitable algorithm to further select time-domain and / or frequency-domain features that simultaneously satisfy monotonicity, trend, and predictability. Specifically, for monotonicity screening, gradient analysis can be used, i.e., calculating the gradient change of the feature over time. If the signs of the gradients are consistent, it indicates that the feature has monotonicity. In addition, monotonic function approximation can be used, i.e., fitting the data using methods such as linear regression or exponential function approximation. If the fitting effect is good, it indicates that the feature may have monotonicity. For trend screening, time series decomposition can be used, i.e., decomposing the time series of time-domain and / or frequency-domain features into trend, seasonal, and random noise components, and extracting the trend component. In addition, weighted moving average can be used, which places more emphasis on the trend of recent data. For predictability screening, autocorrelation coefficients can be used to check whether features exhibit predictable periodic or trend patterns. ARIMA models can also be employed, using a sub-regression integral moving average model to predict time-domain features and analyzing the model's fit to determine predictability. Spectral analysis can also be used to analyze the stability of frequency-domain features, such as whether the main frequency components remain stable over time. This invention, by combining these various screening methods, can extract time-domain and / or frequency-domain features that simultaneously satisfy monotonicity, trend, and predictability.

[0042] Meanwhile, the initial feature extraction network is used to extract features from the selected time-domain and / or frequency-domain features, resulting in the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window.

[0043] Specifically, the extracted time-domain features and / or frequency-domain features that meet the above characteristics are input into the initial LSTM neural network for feature extraction, thereby obtaining the hidden layer feature vectors output by each layer of the LSTM neural network.

[0044] Step 103: Calculate the similarity between the hidden layer feature vector and the corresponding true value, and based on the similarity, use the initial interpretable weight allocation network to assign weights to the initial feature extraction network to obtain the weighted feature extraction network.

[0045] The initial interpretable weight allocation network can be a single-layer attention network. It should be noted that the number of layers in the attention network is not limited to a single layer; it can be set according to actual business needs, and this embodiment of the invention does not impose specific limitations here.

[0046] In this embodiment of the invention, after obtaining the hidden layer feature vectors output by each network layer in the initial LSTM neural network, the similarity between the hidden layer feature vector and the true value under the corresponding time window of each network layer is calculated. Specifically, the similarity between the hidden layer feature vector and the true value can be measured by calculating the cosine similarity, Euclidean distance, etc. between the hidden layer feature vector and the true value.

[0047] The similarity calculation result is then fed into a single-layer attention network, which assigns weights to the initial LSTM neural network based on this result. Specifically, for larger similarities, the single-layer attention network assigns higher weights; conversely, for smaller similarities, it assigns lower weights. This explains the weight allocation process within the LSTM neural network, making the calculation basis for each weight clear, thereby improving the decision-making transparency of the lifespan prediction model and relatively enhancing its predictive stability.

[0048] Step 104: Based on the feature extraction network after weight allocation and the initial interpretable weight allocation network, predict the lifespan of the rotating machinery and measure the distribution difference between the source domain time series and the target domain time series.

[0049] In this embodiment of the invention, after weight allocation, the lifespan of the rotating machinery is predicted using the weighted feature extraction network and the initial interpretable weight allocation network, and the distribution difference between the source domain time series and the target domain time series is measured. Specifically, step 104 includes: inputting the source domain time series into the weighted feature extraction network and the initial interpretable weight allocation network for feature extraction and lifespan prediction, obtaining the first feature vector output by each network layer of the weighted feature extraction network, the second feature vector output by the initial interpretable weight allocation network, and the lifespan of the rotating machinery; simultaneously inputting the target domain time series into... The weighted feature extraction network and the initial interpretable weight allocation network perform feature extraction to obtain the third feature vector output by each network layer of the weighted feature extraction network and the fourth feature vector output by the initial interpretable weight allocation network; calculate the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector; based on the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector, measure the distribution difference between the source domain time series and the target domain time series.

[0050] Specifically, after extracting temporal and / or frequency domain features from the source and target time series and selecting those that meet the corresponding characteristic requirements, the selected temporal and / or frequency domain features corresponding to the source time series are input into a weighted LSTM neural network and a single-layer attention network for feature extraction and lifetime prediction. This yields the first feature vector output by each layer of the weighted LSTM neural network, the second feature vector output by the single-layer attention network, and the predicted lifetime of the rotating machinery. Simultaneously, the selected temporal and / or frequency domain features corresponding to the target time series are input into a weighted LSTM neural network and a single-layer attention network for feature extraction. This yields the third feature vector output by each layer of the weighted LSTM neural network and the fourth feature vector output by the single-layer attention network.

[0051] Furthermore, the maximum mean difference (MMD) between the first and third feature vectors output by each layer of the LSTM neural network after weight allocation is calculated to obtain the maximum mean difference for each layer of the LSTM neural network after weight allocation. Simultaneously, the maximum mean difference (MMD) between the second and fourth feature vectors output by the single-layer attention network is calculated. Finally, the maximum mean differences for each layer of the LSTM neural network after weight allocation are summed, and then summed with the maximum mean difference for the single-layer attention network to obtain the distribution difference between the source domain time series and the target domain time series.

[0052] It should be noted that, in this embodiment of the invention, when measuring the distribution difference between the source domain time series and the target domain time series, only the maximum mean difference corresponding to a single-layer attention network may be considered.

[0053] Step 105: Based on the predicted lifespan of the rotating machinery and the distribution differences, update the weights of the feature extraction network after weight allocation and the initial interpretable weight allocation network.

[0054] In this embodiment of the invention, after determining the predicted lifespan of the rotating machinery and the distribution difference between the source domain time series and the target domain time series, a loss function is constructed to update the model weights. Specifically, step 105 includes: calculating the lifespan loss based on the predicted lifespan and the actual lifespan of the rotating machinery; constructing a loss function based on the lifespan loss and the distribution difference; and updating the weights of the weighted feature extraction network and the initial interpretable weight allocation network based on the loss function. The specific formula for the loss function L′ is as follows:

[0055] L′=L1+λL MMD

[0056] Where L1 is the lifetime loss, L MMD λ represents the distribution difference, and can be set according to actual business needs. Based on this loss function L′, the weights in the LSTM neural network and the single-layer attention network after weight allocation are updated. Each weight allocation and loss function weight update completes one iteration.

[0057] Step 106: Repeat the weight allocation and weight update process until the preset conditions are met, then stop the iterative update and output the preset lifetime prediction model.

[0058] In this embodiment of the invention, the above-described weight allocation and weight update process is repeated until the loss function reaches its minimum value or the preset number of iterations is reached. Then, the iteration is stopped, and the finally updated LSTM neural network and single-layer attention network are output. The finally updated LSTM neural network and single-layer attention network are determined as the preset lifetime prediction model.

[0059] This invention provides a transfer learning method for small-sample, similarity-based, interpretable weight allocation. By calculating the similarity between the hidden feature vectors of each network layer and the true values, weights can be assigned to the initial feature extraction network based on this similarity. This ensures that each weight has a clear calculation basis, thereby improving the interpretability and decision transparency of the lifespan prediction model, and relatively enhancing the model's predictive stability. Simultaneously, by introducing transfer learning and similarity calculation, this invention can obtain more useful information from the source domain time series, thus enhancing the learning effect on small-sample data in the target domain, reducing the model's dependence on large-scale data, and improving the model's generalization ability.

[0060] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the implementation of this embodiment, this embodiment also provides another transfer adaptation method based on similarity-interpretable weight allocation for small samples, such as... Figure 2 As shown, the method includes:

[0061] Step 201: Obtain the source domain time series and target domain time series of the rotating machinery, as well as the initial life prediction model.

[0062] The initial lifetime prediction model includes an initial feature extraction network, an initial interpretable weight allocation network, and an initial uncertainty quantification network. The initial uncertainty quantification network can be a single-layer Bayesian neural network. It should be noted that the number of layers in the Bayesian neural network is not limited to a single layer; it can be set according to actual business needs, and this embodiment of the invention does not impose specific limitations.

[0063] In this embodiment of the invention, in order to train a preset life prediction model for rotating machinery, it is necessary to collect the source domain time series and the target domain time series of the rotating machinery.

[0064] Step 202: Extract the time-domain features and / or frequency-domain features corresponding to the source domain time series, and use the initial feature extraction network to extract the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window.

[0065] In this embodiment of the invention, the specific process of extracting the hidden layer feature vectors output by each network layer under the corresponding time window is exactly the same as step 102, and will not be repeated here.

[0066] Step 203: Calculate the similarity between the hidden layer feature vector and the corresponding true value, and based on the similarity, use the initial interpretable weight allocation network to assign weights to the initial feature extraction network to obtain the weighted feature extraction network.

[0067] In the embodiments of the present invention, when assigning weights to the LSTM neural network, a single-layer attention network will assign higher weights to those with greater similarity, and conversely, a single-layer attention network will assign lower weights to those with less similarity. This can explain the weight assignment process within the LSTM neural network, making the calculation basis of each weight clear, thereby improving the decision transparency of the life prediction model and relatively improving the prediction stability of the model.

[0068] Step 204: Based on the feature extraction network after weight allocation and the initial interpretable weight allocation network, predict the lifespan of the rotating machinery and measure the distribution difference between the source domain time series and the target domain time series.

[0069] In this embodiment of the invention, the process of predicting the lifespan of rotating machinery and measuring the distribution differences between the source domain time series and the target domain time series is exactly the same as step 104, and will not be repeated here.

[0070] Step 205: Use the initial uncertainty quantization network to perform uncertainty quantization on the predicted lifetime to obtain the uncertainty quantization result of the predicted lifetime.

[0071] The initial uncertainty quantification network can be a single-layer Bayesian neural network. It should be noted that the number of layers in the Bayesian neural network is not limited to a single layer; it can be set according to actual business needs, and this embodiment of the invention does not impose specific limitations on this.

[0072] In the embodiments of the present invention, the similarity-based weight allocation strategy in Embodiment 1 still contains uncertainties, namely prediction uncertainty and cognitive uncertainty. To overcome this deficiency, the embodiments of the present invention employ a Bayesian neural network to treat the weights as random variables for uncertainty quantification. This allows the lifespan prediction model to output a prediction result along with the probability distribution of that result, thereby improving the reliability of the model's decision. Specifically, step 205 includes: using the single-layer Bayesian neural network, performing uncertainty quantification on the predicted lifespan using variational inference to obtain the probability distribution of the predicted lifespan; and determining the uncertainty quantification result based on the probability distribution.

[0073] Specifically, a Bayesian neural network is used to treat the model weights as random variables, and variational inference is used for approximation. Forward propagation is employed to quantify the uncertainty of the model output, and a probability distribution is used to describe the confidence interval of the lifespan prediction results, thereby enhancing the reliability of the lifespan prediction model's decisions. Figure 3 As shown, the slanted straight line represents the actual lifespan of the rotating machinery, the dotted broken line represents the predicted lifespan of the rotating machinery, and the shaded area of ​​the broken line represents the probability interval.

[0074] The embodiments of the present invention use a Bayesian neural network to quantify the uncertainty of the weights. By outputting the probability distribution of the lifetime prediction results, the confidence interval of the lifetime prediction results can be given, thereby enhancing the reliability of the model's decision-making.

[0075] Step 206: Based on the predicted lifespan of the rotating machinery, the distribution differences, and the uncertainty quantification results, update the weights of the weighted feature extraction network, the initial interpretable weight allocation network, and the initial uncertainty quantification network.

[0076] In this embodiment of the invention, after determining the predicted lifetime, distribution difference, and uncertainty quantification results, a loss function is constructed to update the model weights. Specifically, step 206 includes: calculating the lifetime loss based on the predicted lifetime and the actual lifetime of the rotating machinery; calculating the divergence loss based on the uncertainty quantification results and the actual uncertainty quantification results of the rotating machinery; constructing a loss function based on the lifetime loss, the divergence loss, and the distribution difference; and updating the weights of the weighted feature extraction network, the initial interpretable weight allocation network, and the initial uncertainty quantification network based on the loss function. The specific formula for the loss function L is as follows:

[0077] L=L1+K1+λL MMD

[0078] Where L1 is the lifetime loss, K1 is the divergence loss, and L MMD λ represents the distribution difference, and can be set according to actual business needs. Based on this loss function L, the weights in the LSTM neural network, single-layer attention network, and single-layer Bayesian neural network after weight allocation are updated. Each weight allocation and loss function weight update completes one iteration.

[0079] Step 207: Repeat the weight allocation and weight update process until the preset conditions are met, then stop the iterative update and output the preset lifetime prediction model.

[0080] In this embodiment of the invention, the above-described weight allocation and weight update process is repeated until the loss function reaches its minimum value or the preset number of iterations is reached. Then, the iteration stops, and the finally updated LSTM neural network, single-layer attention network, and single-layer Bayesian neural network are output. These finally updated LSTM neural network, single-layer attention network, and single-layer Bayesian neural network are determined as the preset lifetime prediction model. The overall training process of the lifetime prediction model is as follows: Figure 4 As shown.

[0081] Furthermore, after training is completed, the feature extraction network and interpretable weight allocation network in the trained preset life prediction model are used to predict the life of the rotating machinery.

[0082] This invention provides a similarity-based interpretable weight allocation transfer adaptation method for small samples. By calculating the similarity between the hidden feature vectors of each network layer and the true values, weights can be assigned to the initial feature extraction network using similarity. This ensures that each weight has a clear calculation basis, thereby improving the interpretability and decision transparency of the lifespan prediction model, and relatively enhancing the model's predictive stability. Simultaneously, by introducing transfer learning and similarity calculation, this invention can obtain more useful information from the source domain time series, thus enhancing the learning effect on small sample data in the target domain, reducing the model's dependence on large-scale data, and improving the model's generalization ability. Furthermore, by employing an uncertainty quantification network to quantify the uncertainty of the weights, this invention can output the probability distribution of the lifespan prediction results, thereby providing a confidence interval for the lifespan prediction results and enhancing the trustworthiness of the model's decisions. Further, this invention utilizes a pre-trained model and domain adaptation methods to ensure the accuracy of feature extraction and the rationality of weight allocation.

[0083] Furthermore, as Figure 1 and Figure 2 The specific implementation of the method shown in this embodiment provides a transfer adaptation device based on similarity-interpretable weight allocation for small samples, such as... Figure 5As shown, the device includes: an acquisition unit 31, an extraction unit 32, an allocation unit 33, a prediction unit 34, an update unit 35, and an output unit 36.

[0084] The acquisition unit 31 can be used to acquire the source domain time series and target domain time series of the rotating machinery, as well as the initial life prediction model, wherein the initial life prediction model includes an initial feature extraction network and an initial interpretable weight allocation network.

[0085] The extraction unit 32 can be used to extract the time-domain features and / or frequency-domain features corresponding to the source domain time series, and use the initial feature extraction network to extract the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window.

[0086] The allocation unit 33 can be used to calculate the similarity between the hidden layer feature vector and the corresponding true value, and based on the similarity, use the initial interpretable weight allocation network to allocate weights to the initial feature extraction network to obtain the weighted feature extraction network.

[0087] The prediction unit 34 can be used to predict the lifespan of the rotating machinery and measure the distribution difference between the source domain time series and the target domain time series based on the feature extraction network after weight allocation and the initial interpretable weight allocation network.

[0088] The update unit 35 can be used to update the weights of the feature extraction network and the initial interpretable weight allocation network based on the predicted lifespan of the rotating machinery and the distribution differences.

[0089] The output unit 36 ​​can be used to repeat the weight allocation and weight update process until the preset conditions are met, at which point the iterative update stops and the preset lifetime prediction model is output.

[0090] In some embodiments, the apparatus further includes a screening unit.

[0091] The filtering unit can be used to filter out time-domain features and / or frequency-domain features that satisfy the target characteristics from the time-domain features and / or frequency-domain features, and obtain the filtered time-domain features and / or frequency-domain features, wherein the target characteristics include monotonicity, trend and predictability.

[0092] The extraction unit 32 can be specifically used to extract features from the filtered time-domain features and / or frequency-domain features using the initial feature extraction network, so as to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window.

[0093] In some embodiments, when the initial feature extraction network is a long short-term memory neural network including three network layers, the extraction unit 32 can also be specifically used to input the time-domain features and / or the frequency-domain features into the long short-term memory neural network for feature extraction, so as to obtain the hidden layer feature vectors output by the three network layers under the corresponding time windows.

[0094] In some embodiments, when the initial interpretable weight allocation network is a single-layer attention network, the allocation unit 33 may be specifically used to allocate weights to the initial feature extraction network based on the similarity using the single-layer attention network, thereby obtaining a feature extraction network with weight allocation.

[0095] In some embodiments, the prediction unit 34 can also be used to predict the lifespan of the rotating machinery using the feature extraction network and the interpretable weight allocation network in the trained preset lifespan prediction model.

[0096] In some embodiments, the prediction unit 34 may be specifically configured to input the source domain time series into the weighted feature extraction network and the initial interpretable weight allocation network for feature extraction and lifetime prediction, obtaining a first feature vector output by each network layer of the weighted feature extraction network, a second feature vector output by the initial interpretable weight allocation network, and the lifetime of the rotating machinery; simultaneously input the target domain time series into the weighted feature extraction network and the initial interpretable weight allocation network for feature extraction, obtaining a third feature vector output by each network layer of the weighted feature extraction network, and a fourth feature vector output by the initial interpretable weight allocation network; calculate the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector; and measure the distribution difference between the source domain time series and the target domain time series based on the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector.

[0097] In some embodiments, when the initial lifetime prediction model further includes an initial uncertainty quantization network, the apparatus further includes a quantization unit.

[0098] The quantization unit can be used to perform uncertainty quantization on the predicted lifetime using the initial uncertainty quantization network to obtain the uncertainty quantization result of the predicted lifetime.

[0099] In some embodiments, when the initial uncertainty quantization network is a single-layer Bayesian neural network, the quantization unit can be specifically used to use the single-layer Bayesian neural network to perform uncertainty quantization on the predicted lifetime using variational inference, thereby obtaining the probability distribution of the predicted lifetime; and to determine the uncertainty quantization result based on the probability distribution.

[0100] In some embodiments, the updating unit 35 may be specifically used to update the weights of the weighted feature extraction network, the initial interpretable weight allocation network, and the initial uncertainty quantification network based on the predicted lifespan of the rotating machinery, the distribution difference, and the uncertainty quantification result.

[0101] In some embodiments, the update unit 35 may also be specifically used to calculate the life loss based on the predicted life and the actual life of the rotating machinery; calculate the divergence loss based on the uncertainty quantization result and the actual uncertainty quantization result of the rotating machinery; construct a loss function based on the life loss, the divergence loss and the distribution difference; and update the weights of the feature extraction network after weight allocation, the initial interpretable weight allocation network and the initial uncertainty quantization network based on the loss function.

[0102] It should be noted that other corresponding descriptions of the functional units involved in the similarity-based interpretable weight allocation transfer adaptation device for small samples provided in this embodiment of the invention can be found in the following references. Figure 1 and Figure 2 The corresponding descriptions in [the document] will not be repeated here.

[0103] Based on the above, Figure 1 and Figure 2 Accordingly, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method. Figure 1 and Figure 2 The example shown is a similarity-based, interpretable weight allocation-based transfer adaptation method for small samples.

[0104] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause an electronic device (such as a personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0105] Based on the above, Figure 1 and Figure 2 The method shown, and Figure 5To achieve the above objectives, the present application also provides an electronic device, specifically a personal computer, tablet computer, server, or other network device, as shown in the virtual device embodiment. This device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figure 1 and Figure 2 The example shown is a similarity-based, interpretable weight allocation-based transfer adaptation method for small samples.

[0106] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0107] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0108] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0110] This invention, through calculating the similarity between the hidden feature vectors of each network layer and the true values, can assign weights to the initial feature extraction network using this similarity. This ensures that each weight has a clear calculation basis, thereby improving the interpretability and decision transparency of the lifespan prediction model, and relatively enhancing the model's predictive stability. Simultaneously, by introducing transfer learning and similarity calculation, this invention can obtain more useful information from the source domain time series, thus enhancing the learning effect on small sample data in the target domain, reducing the model's dependence on large-scale data, and improving the model's generalization ability. Furthermore, by employing an uncertainty quantification network to quantify the uncertainty of the weights, this invention can output the probability distribution of the lifespan prediction results, thereby providing a confidence interval for the lifespan prediction results and enhancing the reliability of the model's decisions. Further, this invention utilizes a pre-trained model and domain adaptation methods to ensure the accuracy of feature extraction and the rationality of weight allocation.

[0111] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0112] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A transfer adaptation method based on similarity-interpretable weight allocation for small samples, characterized in that, include: Obtain the source domain time series and target domain time series of rotating machinery, as well as an initial lifetime prediction model, wherein the initial lifetime prediction model includes an initial feature extraction network and an initial interpretable weight allocation network; Extract the time-domain features and / or frequency-domain features corresponding to the source domain time series, and use the initial feature extraction network to extract features from the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window; Calculate the similarity between the hidden layer feature vector and the corresponding true value, and based on the similarity, use the initial interpretable weight allocation network to assign weights to the initial feature extraction network to obtain the weighted feature extraction network; Based on the feature extraction network after weight allocation and the initial interpretable weight allocation network, the lifespan of the rotating machinery is predicted, and the distribution difference between the source domain time series and the target domain time series is measured. Based on the predicted lifespan of the rotating machinery and the distribution differences, the weights of the feature extraction network after weight allocation and the initial interpretable weight allocation network are updated. Repeat the process of weight allocation and weight update until the preset conditions are met, then stop iterative updates and output the preset lifetime prediction model; When the initial interpretable weight allocation network is a single-layer attention network, based on the similarity, weights are allocated to the initial feature extraction network using the initial interpretable weight allocation network to obtain the weighted feature extraction network, including: Based on the similarity, the initial feature extraction network is assigned weights using the single-layer attention network to obtain the weighted feature extraction network. Based on the feature extraction network after weight allocation and the initial interpretable weight allocation network, the lifespan of the rotating machinery is predicted, and the distribution difference between the source domain time series and the target domain time series is measured, including: The source domain time series is input into the weighted feature extraction network and the initial interpretable weight allocation network for feature extraction and lifetime prediction, to obtain the first feature vector output by each network layer of the weighted feature extraction network, the second feature vector output by the initial interpretable weight allocation network, and the lifetime of the rotating machinery. Simultaneously, the target domain time series is input into the weighted feature extraction network and the initial interpretable weight allocation network for feature extraction, to obtain the third feature vector output by each network layer of the weighted feature extraction network and the fourth feature vector output by the initial interpretable weight allocation network. Calculate the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector; The distribution difference between the source domain time series and the target domain time series is measured based on the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector.

2. The method according to claim 1, characterized in that, After extracting the time-domain features and / or frequency-domain features corresponding to the source domain time series, the method further includes: Select time-domain features and / or frequency-domain features that satisfy the target characteristics from the time-domain features and / or frequency-domain features to obtain the selected time-domain features and / or frequency-domain features, wherein the target characteristics include monotonicity, trend and predictability; The initial feature extraction network is used to extract features from the time-domain features and / or frequency-domain features, resulting in hidden layer feature vectors output by each network layer of the initial feature extraction network within the corresponding time window, including: The initial feature extraction network is used to extract features from the filtered time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window.

3. The method according to claim 1, characterized in that, When the initial feature extraction network is a long short-term memory neural network comprising three network layers, the initial feature extraction network is used to extract features from the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window, including: The time-domain features and / or the frequency-domain features are input into the long short-term memory neural network for feature extraction, thereby obtaining the hidden layer feature vectors output by the three network layers under the corresponding time windows; and / or After outputting the preset lifetime prediction model, the method further includes: The lifespan of the rotating machinery is predicted using the feature extraction network and interpretable weight allocation network in the pre-trained preset lifespan prediction model.

4. The method according to any one of claims 1-3, characterized in that, When the initial lifetime prediction model further includes an initial uncertainty quantification network, the method further includes: The predicted lifetime is quantified using the initial uncertainty quantification network to obtain the uncertainty quantification result of the predicted lifetime.

5. The method according to claim 4, characterized in that, When the initial uncertainty quantization network is a single-layer Bayesian neural network, the uncertainty quantization of the predicted lifetime is performed using the initial uncertainty quantization network to obtain the uncertainty quantization result of the predicted lifetime, including: The single-layer Bayesian neural network is used to quantify the uncertainty of the predicted lifetime using variational inference, thereby obtaining the probability distribution of the predicted lifetime. Based on the probability distribution, the uncertainty quantification result is determined.

6. The method according to claim 4, characterized in that, Based on the predicted lifespan of the rotating machinery and the distribution differences, the weights of the feature extraction network after weight allocation and the initial interpretable weight allocation network are updated, including: Based on the predicted lifespan of the rotating machinery, the distribution differences, and the uncertainty quantification results, the weights of the feature extraction network after weight allocation, the initial interpretable weight allocation network, and the initial uncertainty quantification network are updated.

7. The method according to claim 6, characterized in that, Based on the predicted lifespan of the rotating machinery, the distribution differences, and the uncertainty quantification results, the weights of the weighted feature extraction network, the initial interpretable weight allocation network, and the initial uncertainty quantification network are updated, including: Based on the predicted lifespan and the actual lifespan of the rotating machinery, the lifespan loss is calculated; Based on the uncertainty quantification results and the actual uncertainty quantification results of the rotating machinery, the divergence loss is calculated. Based on the lifetime loss, the divergence loss, and the distribution difference, a loss function is constructed; Based on the loss function, the weights of the feature extraction network after weight allocation, the initial interpretable weight allocation network, and the initial uncertainty quantification network are updated.

8. A transfer adaptation device based on similarity-interpretable weight allocation for small samples, characterized in that, include: An acquisition unit is used to acquire the source domain time series and target domain time series of the rotating machinery, as well as an initial lifetime prediction model, wherein the initial lifetime prediction model includes an initial feature extraction network and an initial interpretable weight allocation network. The extraction unit is used to extract the time-domain features and / or frequency-domain features corresponding to the source domain time series, and to use the initial feature extraction network to extract the time-domain features and / or frequency-domain features to obtain the hidden layer feature vectors output by each network layer of the initial feature extraction network under the corresponding time window. The allocation unit is used to calculate the similarity between the hidden layer feature vector and the corresponding true value, and based on the similarity, to allocate weights to the initial feature extraction network using the initial interpretable weight allocation network, thereby obtaining the weighted feature extraction network. The prediction unit is used to predict the lifespan of the rotating machinery and measure the distribution difference between the source domain time series and the target domain time series based on the feature extraction network after weight allocation and the initial interpretable weight allocation network. An update unit is used to update the weights of the weighted feature extraction network and the initial interpretable weight allocation network based on the predicted lifespan of the rotating machinery and the distribution differences. The output unit is used to repeat the process of weight allocation and weight update until a preset condition is met, at which point the iterative update stops and the preset lifetime prediction model is output. The allocation unit is specifically used to allocate weights to the initial feature extraction network based on the similarity when the initial interpretable weight allocation network is a single-layer attention network, thereby obtaining the feature extraction network after weight allocation. The prediction unit is specifically configured to input the source domain time series into the weighted feature extraction network and the initial interpretable weight allocation network for feature extraction and lifetime prediction, obtaining a first feature vector output by each network layer of the weighted feature extraction network, a second feature vector output by the initial interpretable weight allocation network, and the lifetime of the rotating machinery; simultaneously, it inputs the target domain time series into the weighted feature extraction network and the initial interpretable weight allocation network for feature extraction, obtaining a third feature vector output by each network layer of the weighted feature extraction network, and a fourth feature vector output by the initial interpretable weight allocation network; calculates the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector; and measures the distribution difference between the source domain time series and the target domain time series based on the maximum mean difference between the first feature vector and the third feature vector, and the maximum mean difference between the second feature vector and the fourth feature vector.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images