A device life prediction method and device based on multi-source domain hierarchical migration

Abstract

The present invention discloses a method and device for predicting equipment life based on multi-source domain hierarchical migration, which relates to the field of aerospace technology. The method comprises: using a large number of equipment samples of the same model to pre-train a prediction model, and identifying the universal degradation characteristics of the equipment through the first few layers of the model; then using highly transferable samples to fine-tune the remaining layers of the model to identify the individual degradation characteristics of the target equipment; and finally, in the fine-tuning iterative process, assigning unequal weights to different sequence positions of the samples to increase the weight of samples in the highly transferable sequence interval.

Description

Technical Field

[0001] The present invention relates to the field of aerospace technology, and in particular to a method and device for predicting equipment life based on multi-source domain hierarchical migration. Background Art

[0002] Accurate prediction of the remaining life of equipment is crucial for its safe use and reasonable and accurate maintenance. The key to improving the accuracy of equipment remaining life prediction lies in building a prediction model that can accurately recognize the performance degradation patterns of the target equipment. In practice, for the remaining life prediction model, it is usually required to estimate the remaining life as early as possible before a failure. However, the target equipment has a short historical operating data, and its performance degradation patterns have not yet been clearly revealed, making it difficult to build an effective prediction model. To address the above difficulties, this patent introduces transfer learning, a cutting-edge technology in the field of machine learning, into equipment remaining life prediction to achieve the mutual borrowing of information between different devices of the same model.

[0003] Transfer learning is currently being widely applied in various fields. A multi-source transfer learning network (MSTLN) framework addresses the issues of different fault distributions between the target and source domains, as well as the imbalance in the number of healthy and faulty samples. However, most current transfer learning research focuses on categorical data and classification tasks, with relatively little research applied to condition assessment and life prediction. In recent years, a transfer learning algorithm based on a bidirectional long short-term memory (BLSTM) recurrent neural network has been proposed for remaining useful life (RUL) estimation. However, this transfer prediction method primarily transfers a single type of information, with relatively few hybrid transfer methods. Furthermore, transfer learning research on samples from multiple source domains is limited. Existing transfer prediction methods primarily rely on direct application or minor improvements of classification transfer learning methods, with few specifically designed approaches tailored to the characteristics of time series data and prediction tasks. Existing methods do not address the fusion of samples from source domains with varying degrees of transferability. Summary of the Invention

[0004] The technical problem solved by the solution provided by the embodiment of the present invention is how to improve the accuracy of prediction of the remaining useful life of equipment.

[0005] According to an embodiment of the present invention, a device life prediction method based on multi-source domain hierarchical migration is provided, including:

[0006] Obtaining M devices of the same model as the target device and first device information of each device of the same model, and obtaining N similar devices of the same model as the target device and having a high similarity and second device information of each similar device of the same model;

[0007] Construct a DLSTM-FC prediction network model that includes, from top to bottom, 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters;

[0008] Pre-training the DLSTM-FC prediction network model using the M devices of the same model and the first device information of each device of the same model to obtain a pre-trained DLSTM-FC prediction network model;

[0009] Fine-tune the pre-trained DLSTM-FC prediction network model using the N similar devices of the same model and the second device information of each similar device of the same model to obtain a fine-tuned DLSTM-FC prediction network model;

[0010] Obtain device degradation sensitive parameter information of the target device, and input the device degradation sensitive parameter information of the target device into the DLSTM-FC prediction network model after fine-tuning and training to predict the remaining service life information of the target device.

[0011] According to an embodiment of the present invention, a device life prediction apparatus based on multi-source domain hierarchical migration is provided, comprising:

[0012] an acquisition module, configured to acquire M devices of the same model as the target device and first device information of each device of the same model, and to acquire N similar devices of the same model as the target device and having a high degree of similarity and second device information of each similar device of the same model;

[0013] A construction module is used to construct a DLSTM-FC prediction network model that includes, from top to bottom, five layers of LSTM networks with different model structure parameters and two layers of FC networks with different model structure parameters;

[0014] A training module is configured to pre-train the DLSTM-FC prediction network model using the M devices of the same model and the first device information of each device of the same model to obtain a pre-trained DLSTM-FC prediction network model; and to fine-tune the pre-trained DLSTM-FC prediction network model using the N similar devices of the same model and the second device information of each similar device of the same model to obtain a fine-tuned DLSTM-FC prediction network model;

[0015] The prediction module is used to obtain the device degradation sensitive parameter information of the target device, and input the device degradation sensitive parameter information of the target device into the DLSTM-FC prediction network model after the fine-tuning training to predict the remaining service life information of the target device.

[0016] The solution provided by the embodiments of the present invention firstly utilizes a large number of low-transferability samples and a small number of high-transferability samples to achieve accurate predictions for target devices. Secondly, it can fully exploit the common and individual characteristics of sample degradation processes using historical operating data of the same device model, enabling multi-source domain fusion migration predictions for samples at different levels. This significantly enhances the "personalization" of the prediction model and improves prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to understand the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0018] Figure 1 This is a flow chart of a device life prediction method based on multi-source domain hierarchical migration provided by an embodiment of the present invention;

[0019] Figure 2 Schematic diagram of a device life prediction apparatus based on multi-source domain hierarchical migration provided by an embodiment of the present invention;

[0020] Figure 3 Schematic diagram of the DLSTM prediction network structure provided by an embodiment of the present invention;

[0021] Figure 4 This is a flowchart of a method for predicting the remaining useful life of equipment based on multi-source domain fusion deep LSTM model migration provided by an embodiment of the present invention;

[0022] Figure 5 1 is a schematic diagram of target engine life prediction results provided by an embodiment of the present invention;

[0023] Figure 6 Schematic diagram of target engine life prediction error and score statistical distribution provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments described below are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0025] The present invention addresses the problem of time series prediction under conditions of different transferability distributions of source domain samples and proposes a multi-source domain, multi-level migration method architecture. For a specific target device, the transferability corresponding to different individuals in the source domain and different position sequences in the individuals is different. Effective migration requires the use of a large amount of low-transferability data and a small amount of high-transferability data to achieve accurate prediction. The present invention proposes a multi-level migration strategy to solve this problem. Devices of the same model have the same structure, which leads to certain degradation characteristics between different individuals being similar or even shared. These characteristics can be called "universal decay characteristics." However, manufacturing quality deviations, initial failure levels, and performance degradation rates of different individuals will lead to different decay characteristics between individuals. These personalized degradation characteristics caused by individual differences in devices are called "individual decay characteristics." The universal decay characteristics and individual decay characteristics in the decay feature set of the target device will be identified one by one.

[0026] First, a prediction model is pre-trained using a large number of samples of the same device model. The first four layers of the model identify universal degradation characteristics of the devices. Highly transferable samples are then used to fine-tune the remaining three layers of the model to identify individual degradation characteristics of the target devices. Finally, during the fine-tuning iterations, unequal weights are assigned to samples at different sequence positions to increase the weight of samples within highly transferable sequence intervals. Furthermore, based on the Long Short-Term Memory (LSTM) network architecture's proven performance in long-term time series, the present invention constructs a device life prediction model structure based on the LSTM network model to accurately predict the remaining life of the device.

[0027] Figure 1 This is a flow chart of a device life prediction method based on multi-source domain level migration provided by an embodiment of the present invention. Figure 1 Shown, including:

[0028] Step S101: Acquire M devices of the same model as the target device and first device information of each device of the same model, and acquire N similar devices of the same model as the target device and having a high degree of similarity and second device information of each similar device of the same model;

[0029] Specifically, the first device information of each device of the same model includes: device degradation sensitive parameter information and device remaining service life information of each device of the same model; the second device information of each similar device of the same model includes device degradation sensitive parameter information and device remaining service life information of each similar device of the same model.

[0030] Specifically, obtaining N similar devices of the same model and having a high degree of similarity to the target device includes: obtaining the degradation period length and multidimensional monitoring parameters related to degradation of the target device, and simultaneously obtaining the multidimensional monitoring parameters related to degradation of each of the M devices of the same model; extracting first degradation main information from the multidimensional monitoring parameters of the target device and second degradation main information from the multidimensional monitoring parameters of each of the devices of the same model using a principal component analysis method; calculating a cosine distance weight and a bulldozer distance weight based on the degradation period length of the target device, and calculating a similarity value between the target device and each of the devices of the same model using the first degradation main information, the second degradation main information, the cosine distance weight, and the bulldozer distance weight; and using the similarity value between the target device and each of the devices of the same model, selecting N devices of the same model similar to the target device from the M devices of the same model. The degradation period length of the target device is proportional to the cosine distance weight and inversely proportional to the bulldozer distance weight.

[0031] Step S102: constructing a DLSTM-FC prediction network model that includes, from top to bottom, 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters;

[0032] Step S103: pre-training the DLSTM-FC prediction network model using the M devices of the same model and the first device information of each device of the same model to obtain a pre-trained DLSTM-FC prediction network model;

[0033] Step S104: fine-tuning the pre-trained DLSTM-FC prediction network model using the N similar devices of the same model and the second device information of each similar device of the same model to obtain a fine-tuned DLSTM-FC prediction network model;

[0034] Step S105: Obtain device degradation sensitive parameter information of the target device, and input the device degradation sensitive parameter information of the target device into the fine-tuned and trained DLSTM-FC prediction network model to predict the remaining service life information of the target device;

[0035] Among them, LSTM refers to long short-term memory; FC refers to a fully connected layer; DLSTM-FC refers to a deeper long short-term memory-fully connected layer; M and N are both positive integers, and M is greater than N.

[0036] Furthermore, the construction of a DLSTM-FC prediction network model that includes, from top to bottom, 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters includes: constructing an initial DLSTM-FC prediction network model that includes, from top to bottom, 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters; obtaining a DLSTM-FC prediction network model by randomly initializing the weight value of each network layer in the initial DLSTM-FC prediction network model and setting the training state of the weight value of each network layer to a trainable state.

[0037] Furthermore, the DLSTM-FC prediction network model is pre-trained using the M devices of the same model and the first device information of each device of the same model to obtain the pre-trained DLSTM-FC prediction network model, including: pre-training the weight value of each network layer in the DLSTM-FC prediction network model by using the device degradation sensitive parameter information and the device remaining service life information of each device of the same model as the input and output of the DLSTM-FC prediction network model respectively, and obtaining the weight value of each network layer in the DLSTM-FC prediction network model; randomly initializing the weight values of the last three network layers in the DLSTM-FC prediction network model, and at the same time setting the training status of the weight values of the first four network layers in the DLSTM-FC prediction network model to a non-trainable state, and setting the training status of the weight values of the last three network layers to a trainable state, to obtain the pre-trained DLSTM-FC prediction network model.

[0038] Furthermore, the pre-trained DLSTM-FC prediction network model is fine-tuned using the N similar devices of the same model and the second device information of each similar device of the same model to obtain the fine-tuned DLSTM-FC prediction network model, including: calculating the similarity value between each similar device of the same model and the target device, and configuring a sample weight value for each similar device of the same model according to the similarity value between each similar device of the same model and the target device; and fine-tuning the weight values of the last three network layers in the pre-trained DLSTM-FC prediction network model by using the device degradation sensitive parameter information and device remaining service life information of each similar device of the same model with configured sample weight values as the input and output of the pre-trained DLSTM-FC prediction network model respectively, to obtain the fine-tuned DLSTM-FC prediction network model.

[0039] The sample weight value is proportional to the similarity value.

[0040] Figure 2FIG. 1 is a schematic diagram of a device life prediction apparatus based on multi-source domain level migration provided by an embodiment of the present invention. Figure 2 As shown, it includes: an acquisition module 201, which is used to obtain M devices of the same model as the target device model and the first device information of each device of the same model, and obtain N similar devices of the same model with the same model and high similarity as the target device model and the second device information of each similar device of the same model; a construction module 202, which is used to construct a DLSTM-FC prediction network model that includes 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters from top to bottom; a training module 203, which is used to pre-train the DLSTM-FC prediction network model using the M devices of the same model and the first device information of each device of the same model to obtain a pre-trained DLSTM-FC prediction network model. network model; and fine-tuning the pre-trained DLSTM-FC prediction network model using the N similar devices of the same model and the second device information of each similar device of the same model to obtain the fine-tuned DLSTM-FC prediction network model; a prediction module 204 is used to obtain the device degradation sensitive parameter information of the target device, and input the device degradation sensitive parameter information of the target device into the fine-tuned DLSTM-FC prediction network model to predict the remaining service life information of the target device; wherein, LSTM refers to long short-term memory; FC refers to a fully connected layer; DLSTM-FC refers to a deeper long short-term memory-fully connected layer; wherein, M and N are both positive integers, and M is greater than N.

[0041] Among them, the first device information of each device of the same model includes: device degradation sensitive parameter information and device remaining service life information of each device of the same model; the second device information of each similar device of the same model includes device degradation sensitive parameter information and device remaining service life information of each similar device of the same model.

[0042] Devices of the same model share the same structure, resulting in similar or even shared degradation characteristics between different devices. These characteristics can be referred to as "universal degradation characteristics." However, variations in manufacturing quality, initial failure levels, and performance degradation rates among different devices can lead to variations in degradation characteristics between devices. These personalized degradation characteristics, resulting from individual differences in devices, are referred to as "individual degradation characteristics." Accurate device life prediction requires efficient utilization of the degradation information contained in each sample in the source domain. However, due to differences in individual states, the degradation process varies between different devices. Therefore, only a small number of samples in the source domain share a high degree of similarity with the target device, i.e., they share similar individual degradation characteristics.

[0043] During equipment operation, sensors installed in the equipment or its environment collect various parameter signals in real time or intermittently. These monitoring parameter signals collectively monitor the equipment's operating status. As equipment performance degrades, some of these monitoring parameter signals undergo relatively significant changes. These monitoring parameters are referred to as degradation-sensitive parameters. This invention uses equipment degradation-sensitive parameter data (DP) as input and implements equipment remaining life prediction based on transfer learning methods.

[0044] As an improved recurrent neural network, LSTM can effectively address the vanishing and exploding gradient problems of RNNs in long-period signal processing. It is suitable for analyzing problems related to time series. This embodiment of the present invention uses the LSTM neural network as the core computing layer of this prediction method.

[0045] A traditional LSTM model consists of only one LSTM hidden layer, connected to the output layer via a standard feedforward network. This is because it uses only one recurrent layer to process and represent the temporal information contained in a time series. To achieve higher-level learning and abstract representation of time series, a recurrent layer is added to the traditional LSTM model. This adds an additional layer to extract more abstract features of the time series over time. The sequence output features of one LSTM layer serve as the input to another LSTM layer, stacking multiple LSTM networks to construct a deep long-short-term memory model. This creates a deeper LSTM model, called Deep-LSTM (hereinafter referred to as DLSTM).

[0046] like Figure 3 As shown in Figure 1, the DLSTM architecture can be viewed as an LSTM model consisting of multiple LSTM layers, each containing multiple memory cells operating at different time scales. The first LSTM layer provides a sequence of outputs, rather than a single value, to the second LSTM layer. Specifically, each input time step corresponds to an output in the first LSTM layer, rather than all input time steps corresponding to only the last output. Leveraging deep learning techniques, the stacked recurrent hidden layers make the LSTM model deeper and more stable, making it a better solution for sequence prediction problems.

[0047] Based on the DLSTM network, the present invention designs a model structure for device migration prediction. The prediction model with a 7-layer neural network structure is determined, and the number of frozen layers is set to 4. Combining the strong temporal feature learning ability of the LSTM network and the powerful fitting ability of the fully connected layer (FC), the 7 hidden layers are decomposed into 5 LSTM layers and 2 FC layers to learn the complex mapping relationship from device "DP" to "RUL", and the DLSTM prediction model structure is constructed as follows: Figure 3 shown.

[0048] Multi-source domain sample fusion model prediction

[0049] The present invention applies the design of migration prediction based on the DLSTM-FC network structure, which enables the prediction model to more accurately grasp the information that needs to be learned, and improves the "personalization" of the model and the prediction accuracy. In order to maximize the use of the degradation information carried by source domain samples with different migration levels, the present invention proposes a multi-source domain fusion multi-level migration strategy. The method flow is as follows: Figure 4 shown.

[0050] (1) Model construction

[0051] Based on the designed prediction model structure, a 5-layer LSTM network and a 2-layer FC network with sequential input and output connections are built in sequence, and the model structure parameters are set to complete the construction of the DLSTM-FC device prediction model. A random initial weight W is set for the prediction model.

[0052] (2) Training a general model

[0053] Based on the available DP data of all devices of the same model, the DP of each device is sequentially concatenated and integrated as the input for model training. The RUL value corresponding to the DP sequence is used as the output, and the sample weights of each device are set to a constant value to pre-train the entire model. The pre-trained model can be used to identify the common degradation characteristics of devices;

[0054] (3) Model initialization

[0055] Migrate general model structure parameters: Keep the pre-trained model structure parameters unchanged, initialize the weights of the last three layers of the model to random values, and set the weights of the first four layers of the model (w1-w4) to the "untrainable" state;

[0056] Randomly initialize the weight parameters of the remaining layers of the model: Randomly initialize the last three layers of the prediction model and set the weights of the last three layers (w5-w7) to the "trainable" state.

[0057] (4) Model fine-tuning

[0058] Adjust the weights of fine-tuning training samples: Select DP data from devices of the same model with high similarity to the target device as training input for model fine-tuning, and assign different sample weights to each device based on the degree of similarity. The sample weight value is directly proportional to the similarity.

[0059] Fine-tune the remaining model weight parameters: The prediction model's trainable layers are trained using the DP data of devices with a high degree of similarity to the target device as input and the RUL values corresponding to the DP as output. This training iteration updates the weights of the last three layers of the prediction model, which can be used to identify individual device degradation characteristics.

[0060] (5) Prediction of remaining life of target samples

[0061] The trained prediction model uses the target device DP as input to predict the target device's remaining useful life. The predicted results are compared with the actual remaining useful life to evaluate the prediction model's performance.

[0062] Example

[0063] This paper intends to use the Data Challenge dataset from the 2008 PHM International Conference for method research and validation. This data was generated using C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) simulation. The dataset contains simulation data for 100 aircraft engines. Through parameter screening, eight degradation-related gas path parameters were selected for case analysis.

[0064] Based on the characteristics of CMAPSS data and the model calculation cost, the life prediction model parameter settings are shown in Table 1. The model input is the normalized DP sample and its weight, and the output is the normalized RUL label. The number of neurons in the input layer of the DLSTM neural network model is consistent with the input sample window length and is set to 16. Seven hidden layers are used to learn the engine degradation characteristics layer by layer. The number of neurons contained in each hidden layer is 14, 12, 10, 8, 6, 4, and 2 respectively. The activation function of the LSTM layer is tanh, and the model output layer consists of 1 neuron mapping the remaining number of cycles. To reduce the overfitting phenomenon of the model, the L2 regularization term coefficient is set to 0.01 in the input layer of the model. At the same time, a dropout layer is set after the input layer and the first hidden layer, with a parameter set to 0.1. The loss function used by the model is the mean square error function, and the training optimizer is Adam.

[0065] Table 1 DLSTM prediction model structure parameter settings

[0066]

[0067] The parameter settings of the DLSTM model pre-training process are shown in Table 2. The time series window length Win of the LSTM model lstm The weights of the model's input layer and the first four LSTM layers are frozen after pre-training. To ensure both training time and effectiveness, the training batch size is set to 960, which is used to train the training data in batches. The number of training iterations is defined as 50.

[0068] Table 2 DLSTM prediction model training parameter settings

[0069] Parameter name L2 regularization coefficient epoch batch-size Parameter value 0.01 50 960

[0070] The above life prediction model is applied to perform life prediction on 100 test engines in the CMAPSS dataset. Figure 5 The life prediction results for all tested engines are presented. The results show that the actual lifespans of the target engines range from 141 to 341 cycles. While the predicted lifespans for most target engines are very close to the actual values, some target engines have actual lifespans that are significantly larger or smaller, which increases the difficulty of the prediction process and leads to some deviations in the prediction results.

[0071] The statistical results of prediction error and prediction score indicators are as follows Figure 6 As shown, the results show that the prediction errors are generally normally distributed. The absolute value of the prediction errors for 56% of the test engines is less than 10, and the prediction errors for most engines are concentrated within the acceptable range of [-10, 13]. The acceptance rate of the prediction results is 63%. The prediction scores for 81% of the engines are less than 5. This prediction result shows that the proposed method has accurate prediction results for the 100 target engines in the CMAPSS dataset.

[0072] According to the solution provided by the embodiment of the present invention, a DLSTM+FC network structure is designed to improve the model depth and ensure that the prediction model has the ability to independently mine "universal decay characteristics" and "individual decay characteristics"; the secondary training model migration strategy makes full use of the decay information in source domain samples with different transferability, improves the prediction model's ability to learn decay characteristics, and ensures the model's prediction accuracy; and makes full use of data from other devices of the same model in the device prediction process, thereby improving prediction efficiency and accuracy.

[0073] Although the present invention has been described in detail above, it is not limited thereto, and those skilled in the art can make various modifications based on the principles of the present invention. Therefore, any modifications made based on the principles of the present invention should be understood to fall within the scope of protection of the present invention.

Claims

1. A device life prediction method based on multi-source domain hierarchical migration, characterized in that: include: Acquiring M devices of the same model as the target device and first device information of each device of the same model, and acquiring N similar devices of the same model as the target device and having a high similarity and second device information of each similar device of the same model, which includes: Obtain the degradation period length and multi-dimensional monitoring parameters related to degradation of the target device, and simultaneously obtain the multi-dimensional monitoring parameters related to degradation of each of M devices of the same model; Extracting first degradation main information from the multidimensional monitoring parameters of the target device by a principal component analysis method, and extracting second degradation main information from the multidimensional monitoring parameters of each device of the same model; Calculating a cosine distance weight and a bulldozer distance weight according to the degradation period length of the target device, and calculating a similarity value between the target device and each of the devices of the same model using the first degradation main information, the second degradation main information, the cosine distance weight, and the bulldozer distance weight; Using the similarity value between the target device and each of the devices of the same model, N devices of the same model that are similar to the target device are screened out from the M devices of the same model; Construct a DLSTM-FC prediction network model that includes, from top to bottom, 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters; Pre-training the DLSTM-FC prediction network model using the M devices of the same model and the first device information of each device of the same model to obtain a pre-trained DLSTM-FC prediction network model; Fine-tune the pre-trained DLSTM-FC prediction network model using the N similar devices of the same model and the second device information of each similar device of the same model to obtain a fine-tuned DLSTM-FC prediction network model; Obtaining device degradation sensitive parameter information of the target device, and inputting the device degradation sensitive parameter information of the target device into the fine-tuned and trained DLSTM-FC prediction network model to predict the remaining service life information of the target device; Among them, LSTM refers to long short-term memory; FC refers to fully connected layer; DLSTM-FC refers to a deeper long short-term memory-fully connected layer; Wherein, both M and N are positive integers, and M is greater than N.

2. The method according to claim 1, characterized in that The first device information of each device of the same model includes: device degradation sensitive parameter information and device remaining service life information of each device of the same model; the second device information of each similar device of the same model includes device degradation sensitive parameter information and device remaining service life information of each similar device of the same model.

3. The method according to claim 2, characterized in that The DLSTM-FC prediction network model constructed from top to bottom includes five layers of LSTM networks with different model structure parameters and two layers of FC networks with different model structure parameters. Construct an initial DLSTM-FC prediction network model that includes, from top to bottom, 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters; The DLSTM-FC prediction network model is obtained by randomly initializing the weight value of each network layer in the initial DLSTM-FC prediction network model and setting the training state of the weight value of each network layer to a trainable state.

4. The method according to claim 3, characterized in that Pre-training the DLSTM-FC prediction network model using the M devices of the same model and the first device information of each device of the same model to obtain the pre-trained DLSTM-FC prediction network model includes: By using the device degradation sensitive parameter information and the device remaining service life information of each device of the same model as the input and output of the DLSTM-FC prediction network model, respectively, pre-training the weight value of each network layer in the DLSTM-FC prediction network model, and obtaining the weight value of each network layer in the DLSTM-FC prediction network model; The pre-trained DLSTM-FC prediction network model is obtained by randomly initializing the weight values of the last three network layers in the DLSTM-FC prediction network model, setting the training status of the weight values of the first four network layers in the DLSTM-FC prediction network model to a non-trainable state, and setting the training status of the weight values of the last three network layers to a trainable state.

5. The method according to claim 4, characterized in that Fine-tuning the pre-trained DLSTM-FC prediction network model using the N similar devices of the same model and the second device information of each similar device of the same model to obtain the fine-tuned DLSTM-FC prediction network model includes: Calculating a similarity value between each of the similar devices of the same model and the target device, and configuring a sample weight value for each of the similar devices of the same model according to the similarity value between each of the similar devices of the same model and the target device; By using the device degradation sensitive parameter information and device remaining service life information of each similar device of the same model with configured sample weight values as the input and output of the pre-trained DLSTM-FC prediction network model, the weight values of the last three network layers in the pre-trained DLSTM-FC prediction network model are fine-tuned to obtain the fine-tuned DLSTM-FC prediction network model.

6. The method according to claim 5, characterized in that The sample weight value is proportional to the similarity value.

7. The method according to claim 1, characterized in that The length of the degradation period of the target equipment is proportional to the cosine distance weight and inversely proportional to the bulldozer distance weight.

8. A device life prediction device based on multi-source domain hierarchical migration, characterized in that: include: an acquisition module, configured to acquire M devices of the same model as the target device and first device information of each device of the same model, and to acquire N similar devices of the same model as the target device and having a high degree of similarity and second device information of each similar device of the same model; A construction module is used to construct a DLSTM-FC prediction network model that includes, from top to bottom, five layers of LSTM networks with different model structure parameters and two layers of FC networks with different model structure parameters; A training module is configured to pre-train the DLSTM-FC prediction network model using the M devices of the same model and the first device information of each device of the same model to obtain a pre-trained DLSTM-FC prediction network model; and to fine-tune the pre-trained DLSTM-FC prediction network model using the N similar devices of the same model and the second device information of each similar device of the same model to obtain a fine-tuned DLSTM-FC prediction network model; A prediction module, configured to obtain device degradation sensitive parameter information of a target device, and input the device degradation sensitive parameter information of the target device into the fine-tuned and trained DLSTM-FC prediction network model to predict the remaining service life information of the target device; The acquisition module is specifically configured to acquire the degradation period length and multidimensional monitoring parameters related to degradation of the target device, and simultaneously acquire the multidimensional monitoring parameters related to degradation of each device of the same model among M devices of the same model; extract first main degradation information from the multidimensional monitoring parameters of the target device and extract second main degradation information from the multidimensional monitoring parameters of each device of the same model using a principal component analysis method; calculate a cosine distance weight and a bulldozer distance weight based on the degradation period length of the target device, and calculate a similarity value between the target device and each device of the same model using the first main degradation information, the second main degradation information, the cosine distance weight, and the bulldozer distance weight; and screen out N devices of the same model that are similar to the target device from the M devices of the same model using the similarity value between the target device and each device of the same model; Among them, LSTM refers to long short-term memory; FC refers to fully connected layer; DLSTM-FC refers to a deeper long short-term memory-fully connected layer; Wherein, M and N are both positive integers, and M is greater than N.

9. The device according to claim 8, characterized in that The first device information of each device of the same model includes: device degradation sensitive parameter information and device remaining service life information of each device of the same model; the second device information of each similar device of the same model includes device degradation sensitive parameter information and device remaining service life information of each similar device of the same model.

Images