Equipment state prediction method and evaluation method

Abstract

The invention provides an equipment state prediction method and evaluation method, and the method comprises the steps: collecting the historical state data of equipment, carrying out the normalization processing of the historical state data, and obtaining time series data; training the TimeGAN-LSTM model on the basis of the time sequence data and an anti-overfitting strategy, and performing joint optimization on the TimeGAN-LSTM model according to weights on the basis of an adversarial loss, a supervision loss and an embedding loss; migrating the supervision network into an LSTM prediction network to obtain a state prediction model; and inputting the state data at the current moment into the state prediction model to obtain state data at a plurality of continuous moments in the future. The invention provides an equipment state prediction method and an evaluation method, which are used for solving the problems that a traditional state prediction method is difficult to process non-linear, non-stationary and time dependence characteristics in time sequence data and TimeGAN is not fully applied in the field of state prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of equipment state prediction, and in particular to an equipment state prediction method and an evaluation method. BACKGROUND

[0002] With the improvement of industrial automation and intelligence level, the demand for real-time monitoring and prediction analysis of equipment state is increasing. Traditional state prediction methods are mostly based on statistical models or machine learning algorithms, but it is often difficult to handle the nonlinear, non-stationary and time-dependent characteristics of time series data.

[0003] In the field of time series data analysis, traditional prediction models such as autoregressive integrated moving average (ARIMA) model, long short-term memory network (LSTM) etc. Although they perform well in some linear, stationary time series prediction tasks, they face many challenges when dealing with complex, nonlinear, non-stationary time series data. With the advent of the big data era, these complex time series data become more and more common in fields such as financial markets, energy management, intelligent transportation, weather forecasting, biomedical signal processing, etc. Time series data in these fields usually have the following characteristics: first, large amount of data, second, nonlinear, non-stationary data, and third, complex interdependence between time series data.

[0004] Under such circumstances, traditional prediction models often fail to capture the deep features and complex dynamics of the data, resulting in a decline in prediction performance. In order to overcome these limitations, researchers have begun to explore the application of deep learning techniques in time series prediction. As a new deep learning method, generative adversarial network (GAN) has achieved remarkable results in image generation, style transfer, etc. The core idea of GAN is to generate data close to the real data distribution through the adversarial learning of generator and discriminator.

[0005] However, direct application of traditional GAN to time series data generation and prediction faces two major challenges: one is the time order of time series data, i.e. the generated data needs to maintain the continuity and correlation in time; the second is the mode collapse problem that may occur in the training process of GAN, resulting in lack of diversity of generated data.

[0006] TimeGAN is a generative adversarial network designed specifically for time series data, such as Figure 1As shown, it introduces embedding network and recovery network on the basis of retaining the original structure of GAN, effectively solves the mode collapse problem of traditional GAN in time series generation, and the generated data not only maintains the time sequence continuity, but also has diversity. In this way, TimeGAN can not only learn the latent distribution of time series data, but also has strong time series data generation capability, so it is often used for generating data, but its application in state prediction field is not sufficient. SUMMARY

[0007] The application provides a device state prediction method and an evaluation method to solve the problems that traditional state prediction methods are difficult to handle the nonlinear, non-stationary and time-dependent characteristics in time series data and TimeGAN is not sufficient in the application in the state prediction field.

[0008] In a first aspect, the application provides a device state prediction method, comprising:

[0009] Collecting historical state data of the device, normalizing the historical state data to obtain time series data;

[0010] Training a TimeGAN-LSTM model based on the time series data and an anti-overfitting strategy, and jointly optimizing the TimeGAN-LSTM model according to weights of an adversarial loss, a supervision loss and an embedding loss, wherein the TimeGAN-LSTM model comprises an auto-encoding network with a multi-layer LSTM structure, an adversarial network and a supervision network;

[0011] Migrating the supervision network to an LSTM prediction network to obtain a state prediction model, wherein the LSTM prediction network comprises an input layer of the supervision network, an LSTM network and a fully connected layer, the LSTM network and the fully connected layer belong to a trained LSTM prediction model, and the LSTM prediction model is trained based on historical state data of the entire life cycle of the device;

[0012] Inputting state data of a current time into the state prediction model to obtain state data of a plurality of future continuous times.

[0013] Optionally, the historical state data is collected by a plurality of types of sensors.

[0014] Optionally, the normalization is realized by Z-Score standardization.

[0015] Optionally, before the historical state data is normalized, the historical state data is subjected to data cleaning, and the data cleaning comprises:

[0016] The missing data is repaired by removing abnormal values through a sliding window, and then using linear interpolation or mean filling method.

[0017] Optionally, the anti-overfitting strategy comprises early stopping.

[0018] Optionally, the TimeGAN-LSTM model is trained based on the time series data and the anti-overfitting strategy, and the TimeGAN-LSTM model is jointly optimized according to the adversarial loss, the supervised loss and the embedding loss according to weights, comprising:

[0019] The time series data is divided into a training set and a validation set;

[0020] The training set is input into an embedding network and a recovery network for training;

[0021] When the adversarial loss, the supervised loss and the embedding loss of the validation set do not decrease continuously for multiple rounds, the training is stopped.

[0022] Optionally, the migration of the supervised network to the LSTM prediction network comprises:

[0023] The auto-encoding network and the adversarial network are frozen;

[0024] The input layer of the supervised network is frozen;

[0025] The multi-layer LSTM structure of the supervised network is replaced by the LSTM network;

[0026] The output layer of the supervised network is replaced by the fully connected layer.

[0027] The method can more accurately capture the complex dynamics and nonlinear characteristics of historical state data by collecting historical state data of equipment, combining the processing advantages of TimeGAN for time series and the prediction advantages of LSTM for time series, establishing a state prediction model, accurately predicting state data at future continuous time points according to state data at the current time, and solving the problems that traditional state prediction methods are difficult to handle nonlinear, non-stationary and time-dependent characteristics in time series data and TimeGAN is not fully applied in the field of state prediction. At the same time, unlike the traditional mode of first training a TimeGAN model for data enhancement, then training an LSTM model for prediction using the enhanced data, the method first trains a TimeGAN-LSTM model, then migrates the supervised network to an LSTM prediction network, and finally uses the state prediction model to obtain a state prediction result, which has the following beneficial effects:

[0028] (1) Integrated model architecture: In the traditional method, the TimeGAN model and the LSTM model are two independent models, which are trained respectively. In the present method, a tightly coupled TimeGAN-LSTM model joint training framework is constructed. The original TimeGAN model is adopted with a multi-layer LSTM structure, and the supervision network also has a certain prediction function, so that the data generation and prediction tasks are deeply integrated at the model structure level.

[0029] (2) Task-driven joint training and optimization mechanism: In the traditional method, the training target of the TimeGAN model is only to generate realistic data, and the optimization is completely irrelevant to the downstream prediction task. In the present method, the three loss functions of the adversarial loss, the supervision loss and the embedding loss are simultaneously optimized in the whole training process. The supervision loss is used to directly measure the prediction error of the supervision network. Through joint optimization, the goal of the generation network is not only to deceive the discriminator network, but also to generate data that can help to reduce the prediction error, realizing the qualitative change from blind data enhancement to task-driven data generation, so as to ensure that the generated data is most beneficial to the final prediction task.

[0030] (3) Efficient prediction model based on network migration: In the traditional method, after the TimeGAN model generates enhanced data, a LSTM prediction model is randomly initialized and trained from scratch. In the present method, after the training is completed, the trained supervision network is directly migrated to the final LSTM prediction network, which improves the performance and stability of the final state prediction model.

[0031] In a second aspect, the present application provides a device state evaluation method based on the state data of the future continuous multiple time obtained by any of the above methods, comprising:

[0032] determining whether the state data of each time in the future continuous multiple time is within the normal threshold range;

[0033] If it belongs, the device state at this time is healthy;

[0034] If it does not belong, determine whether the state data of each time is within the fault threshold range;

[0035] If it belongs, the device state at this time is failure;

[0036] If it does not belong, the device state at this time is degraded.

[0037] The present method evaluates the device state of the future continuous multiple time by threshold judgment on the predicted state data, reduces the unexpected downtime, prolongs the service life of the device, and optimizes the allocation of maintenance resources.

[0038] In a third aspect, the present application provides an electronic device comprising:

[0039] a memory for storing program instructions;

[0040] a processor for invoking the program instructions stored in the memory to perform any of the above methods according to the obtained program.

[0041] In a fourth aspect, the present application provides a computer-readable non-volatile storage medium comprising computer-readable instructions, which, when read and executed by a computer, cause the computer to perform any of the above methods. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 a structure diagram of the TimeGAN model provided for the background of the present application;

[0043] Figure 2 a flowchart of a device state prediction method provided for embodiment 1 of the present application;

[0044] Figure 3 a structure diagram of the TimeGAN-LSTM model provided for embodiment 1 of the present application;

[0045] Figure 4 a PCA diagram provided for embodiment 1 of the present application;

[0046] Figure 5 a t-SNE diagram provided for embodiment 1 of the present application;

[0047] Figure 6 a structure diagram of the LSTM prediction model provided for embodiment 1 of the present application;

[0048] Figure 7 a training loss value curve diagram provided for embodiment 1 of the present application;

[0049] Figure 8 a prediction result diagram provided for embodiment 1 of the present application;

[0050] Figure 9 a structure diagram of a device state prediction device provided for embodiment 2 of the present application;

[0051] Figure 10 a flowchart of a device state evaluation method provided for embodiment 3 of the present application. DETAILED DESCRIPTION

[0052] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments. It is to be understood that the drawings are very simplified and are not drawn to scale, and are only used to facilitate, clarify and assist in the description of the embodiments of the present application.

[0053] It should be noted that, in order to clearly illustrate the content of the present application, the present application is characterized by multiple embodiments to further illustrate different implementations of the present application, wherein the multiple embodiments are enumerated rather than exhaustive. In addition, for the sake of simplicity, the content mentioned in the previous embodiments is often omitted in the subsequent embodiments, therefore, the content not mentioned in the subsequent embodiments can be correspondingly referred to the previous embodiments.

[0054] Embodiment 1

[0055] Please refer to Figure 2 , Figure 2 The device state prediction method provided by the present embodiment is shown. In the present embodiment, the method comprises:

[0056] Step 102, collecting historical state data of the device, and performing normalization processing on the historical state data to obtain time series data;

[0057] In the present embodiment, the historical state data is collected by multiple types of sensors. The multiple types of sensors (such as vibration, temperature, current sensors) collect device state data in real time, and the sampling frequency can be set to 1 kHz to 10 kHz according to the characteristics of the device, so as to ensure that the data covers the full working condition of the device.

[0058] In the present embodiment, the existing aero-engine CMAPSS data set can be used as historical state data. The sensors on the engine are collected in real time in the data set, and there are a total of 21 sensors. The sensor name table is shown in Table 1:

[0059] Table 1

[0060] Serial Number Sensor Name Serial Number Sensor Name 1 Fan Inlet Total Temperature 2 Low Pressure Compressor Outlet Total Temperature 3 High Pressure Compressor Outlet Total Temperature 4 Turbine Outlet Total Temperature 5 Fan Inlet Pressure 6 Bypass Pipe Total Pressure 7 High Pressure Compressor Outlet Total Pressure 8 High Pressure Fan Speed 9 High Pressure Rotor Speed 10 Engine Pressure Ratio 11 High Pressure Compressor Outlet Static Pressure 12 Engine Fuel Gas Ratio 13 Low Pressure Rotor Corrected Speed 14 High Pressure Rotor Corrected Speed 15 Bypass Ratio 16 Combustor Combustion Air Ratio 17 Fuel Enthalpy 18 Required Fan Speed 19 Fan Corrected Speed 20 High Pressure Compressor Coolant Release 21 Low Pressure Compressor Coolant Release

[0061] The data in the data set is screened, and it is found that sensors 1, 5, 10, 16, 18 and 19 are all constant values, which are not related to the degradation of the aero-engine, so these six sensors are removed from the data set, and the remaining sensor values are normalized.

[0062] In the present embodiment, the normalization processing is realized by Z-Score standardization.

[0063] It should be noted that the above-mentioned aero-engine CMAPSS data set is a standard data. For non-standard data, data cleaning is also required before normalization processing.

[0064] Specifically, the data cleaning includes:

[0065] The abnormal values (sensor noise or signal loss, etc.) are removed through a sliding window, and then the missing data is repaired by using a linear interpolation or mean filling method.

[0066] In step 104, the TimeGAN-LSTM model is trained based on time series data and an anti-overfitting strategy, and the TimeGAN-LSTM model is jointly optimized based on an adversarial loss, a supervised loss and an embedding loss according to weights, the TimeGAN-LSTM model including an auto-encoding network adopting a multi-layer LSTM structure, an adversarial network and a supervised network;

[0067] Specifically, referring to Figure 3 , the TimeGAN-LSTM model includes:

[0068] An embedding network (Encoder): a multi-layer LSTM structure is adopted to map the original time series data to a low-dimensional latent space to obtain a latent vector;

[0069] A recovery network (Recovery): a multi-layer LSTM structure is adopted to restore the latent vector into time series data;

[0070] A generation network (Generator): a multi-layer LSTM structure is adopted to input random noise to generate time series data.

[0071] A discriminator network (Discriminator): a multi-layer LSTM structure is adopted, and the output layer is a classifier to determine whether the data is time series data generated by the generation network.

[0072] A supervised network (Supervisor): a multi-layer LSTM structure is adopted to predict the next time latent vector, and the loss function (i.e., the supervised loss) is the mean square error (MSE).

[0073] In this embodiment, the multi-layer LSTM structure is a double-layer LSTM structure.

[0074] In this embodiment, in the TimeGAN-LSTM model, the number of nodes of each network is 15, the number of layers of each network is 4, the dimension of the hidden layer is 128, and the number of iterations is 10,000.

[0075] The TimeGAN-LSTM model uses LSTM in TimeGAN to force the data output by the generation network to satisfy time continuity, and through the generation network of the LSTM structure, random noise is fused to improve the diversity and authenticity of the generated data.

[0076] In this embodiment, the anti-overfitting strategy includes an early stopping method.

[0077] Further, step 104 includes:

[0078] Step 1041, dividing the time series data into a training set and a validation set;

[0079] Step 1042, inputting the training set into the embedding network and the recovery network for training;

[0080] Step 1043, stopping training when the adversarial loss, the supervised loss and the embedding loss of the validation set do not decrease for a plurality of consecutive rounds.

[0081] In this embodiment, the training parameters are set as follows:

[0082] Optimizer: Adam (learning rate 0.001, decay rate 0.9);

[0083] Batch size: 128;

[0084] Training period (Epochs): 10000.

[0085] Specifically, the adversarial loss is the output of the adversarial network between the generation network and the discriminator network; the supervised loss is the difference between the predicted next time latent vector of the multi-layer LSTM structure and the real next time latent vector, which is output by the supervision network; and the embedding loss is the reconstruction error of the auto-encoding network, which is output by the recovery network.

[0086] By introducing the joint optimization of the adversarial loss, the supervised loss and the embedding loss, the TimeGAN-LSTM model can dynamically adjust the time sequence logic of the generated data, ensure the time coherence of the generated data, and realize the synchronous training of the auto-encoding network, the adversarial network and the supervision network.

[0087] In this embodiment, the weights of the adversarial loss, the supervised loss and the embedding loss are 1.0:0.5:0.2.

[0088] In this embodiment, the plurality of rounds is 10 rounds.

[0089] Referring to Figure 4 and Figure 5 , the distribution of the generated data (Synthetic) and the original data (Original) is visualized through the PCA graph and the t-SNE graph, and the overall training result of the TimeGAN-LSTM model is obtained.

[0090] Step 106, migrating the supervision network to an LSTM prediction network to obtain a state prediction model, the LSTM prediction network including an input layer of the supervision network, an LSTM network and a fully connected layer, the LSTM network and the fully connected layer belonging to the trained LSTM prediction model, and the LSTM prediction model being trained based on historical state data of the equipment full life cycle.

[0091] In this embodiment, step 106 includes:

[0092] Step 1061, freezing the auto-encoding network and the adversarial network;

[0093] Step 1062, freezing the input layer of the supervised network;

[0094] Step 1063, replacing the multi-layer LSTM structure of the supervised network with an LSTM network;

[0095] Step 1064, replacing the output layer of the supervised network with a fully connected layer.

[0096] Specifically, referring to Figure 6 , the LSTM prediction model includes an input layer, an LSTM network, and an output layer. The number of nodes in each layer is 1-M-1, where M is the propagation length of the LSTM network.

[0097] In this embodiment, M of the LSTM prediction model is 200, the input during training is the normalized value of the current time sensor in the above-mentioned aero-engine CMAPSS data set, and the output is the normalized value of the next time sensor.

[0098] The training loss value curve of the LSTM prediction model is shown in Figure 7 , and the prediction result is shown in Figure 8 .

[0099] Step 108, inputting the state data of the current time into the state prediction model to obtain the state data of the future continuous multiple time points.

[0100] The state data of the current time is extracted into a latent vector by the embedding network, and the LSTM prediction network outputs the latent vector of the future continuous multiple time points based on the latent vector. The latent vector of the future continuous multiple time points is output to the recovery network to obtain the state data of the future continuous multiple time points.

[0101] In this embodiment, the mean absolute error (MAE) is used to measure the prediction result of the state prediction model, and the MAE of the prediction result is 0.1143.

[0102] The method collects historical state data of the equipment, combines the processing advantage of TimeGAN on time series and the prediction advantage of LSTM on time series, more accurately captures the complex dynamics and nonlinear characteristics of the historical state data, establishes a state prediction model, accurately predicts the state data of future continuous time according to the state data of the current time, solves the problems that the traditional state prediction method is difficult to process the nonlinear, non-stationary and time-dependent characteristics in the time series data and TimeGAN is not sufficient in the field of state prediction; at the same time, unlike the traditional mode of training the TimeGAN model first, then enhancing the data, and then training the LSTM model with the enhanced data for prediction, the method trains the TimeGAN-LSTM model first, then migrates the supervised network to the LSTM prediction network, and finally uses the state prediction model to obtain the state prediction result, the beneficial effects are as follows:

[0103] (1) Integrated model architecture: In the traditional method, the TimeGAN model and the LSTM model are two independent models, which are trained respectively. In the present method, a tightly coupled TimeGAN-LSTM model joint training framework is constructed. The original TimeGAN model is adopted with a multi-layer LSTM structure, and the supervised network also has a certain prediction function, so that the data generation and prediction tasks are deeply integrated at the model structure level.

[0104] (2) Task-driven joint training and optimization mechanism: In the traditional method, the training target of the TimeGAN model is only to generate realistic data, and its optimization has nothing to do with the downstream prediction task. In the present method, the three loss functions of adversarial loss, supervised loss and embedding loss are optimized simultaneously in the overall training process. The supervised loss is used to directly measure the prediction error of the supervised network. Through joint optimization, the goal of the generation network is no longer to deceive the discriminator network, but to generate data that can deceive the discriminator network and help to reduce the prediction error, realizing the qualitative change from blind data enhancement to task-driven data generation, to ensure that the generated data is most beneficial to the final prediction task.

[0105] (3) Efficient prediction model based on network migration: In the traditional method, after generating enhanced data using the TimeGAN model, a LSTM prediction model needs to be randomly initialized and trained from scratch. In the present method, after training is completed, the trained supervised network is directly migrated to the final LSTM prediction network, improving the performance and stability of the final state prediction model.

[0106] Embodiment 2

[0107] Please refer to Figure 9 , Figure 9 A device state prediction apparatus provided by the present embodiment is shown. In the present embodiment, the apparatus comprises:

[0108] a processing module configured to collect historical state data of the equipment, and normalize the historical state data to obtain time series data;

[0109] a training module configured to train a TimeGAN-LSTM model based on the time series data and an anti-overfitting strategy, and jointly optimize the TimeGAN-LSTM model based on an adversarial loss, a supervision loss and an embedding loss according to weights, the TimeGAN-LSTM model comprising an auto-encoding network with a multi-layer LSTM structure, an adversarial network and a supervision network;

[0110] a migration module configured to migrate the supervision network into an LSTM prediction network to obtain a state prediction model, the LSTM prediction network comprising an input layer of the supervision network, an LSTM network and a fully connected layer, the LSTM network and the fully connected layer belonging to the trained LSTM prediction model, and the LSTM prediction model being trained based on historical state data of the entire life cycle of the equipment;

[0111] a prediction module configured to input state data at a current time into the state prediction model to obtain state data at a plurality of continuous future times.

[0112] Embodiment 3

[0113] Please refer to Figure 10 , Figure 10 which shows a device state evaluation method provided by the embodiment. In the embodiment, the method is based on the state data at a plurality of continuous future times obtained in Embodiment 1, and comprises:

[0114] determining whether the state data at each time of the state data at the plurality of continuous future times belongs to a normal threshold range;

[0115] if yes, the device state at the time is healthy;

[0116] if no, determining whether the state data at each time belongs to a failure threshold range;

[0117] if yes, the device state at the time is failure;

[0118] if no, the device state at the time is degradation.

[0119] The method evaluates the device state at the plurality of continuous future times by threshold judgment on the predicted state data, reduces unexpected downtime, prolongs the service life of the equipment, and optimizes the allocation of maintenance resources.

[0120] Embodiment 4

[0121] The embodiment of the present application provides an electronic device, comprising:

[0122] a memory for storing program instructions;

[0123] a processor for invoking the program instructions stored in the memory to implement the above method.

[0124] Embodiment 5

[0125] The embodiment of the present application provides a computer readable nonvolatile storage medium, which comprises computer readable instructions, when the computer reads and executes the computer readable instructions, the computer executes the above method.

[0126] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer readable storage media containing computer usable program code.

[0127] In addition, for the above system specific embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts refer to the part of the method embodiment. Moreover, it should be noted that in each module of the system of the present application, the components therein are logically divided according to the functions to be implemented, but the present application is not limited thereto, and each component can be re-divided or combined as needed.

[0128] Each embodiment in the present specification is described in a progressive manner, and the same and similar parts between each embodiment can be referred to each other, and each embodiment mainly describes the differences between other embodiments.

[0129] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims can be performed in a different order than the order in which they are recited and still achieve desirable results. In addition, the specific order or sequential order of processes depicted in the figures is not necessarily the only order in which the processes can be carried out, and in some embodiments, multitasking and parallel processing can be advantageous.

[0130] The above only describes the embodiments of the present application and is not intended to limit the present application. The present application can have various changes and modifications for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A method for predicting equipment status, characterized in that, include: Collect historical status data of the equipment, and normalize the historical status data to obtain time series data; The TimeGAN-LSTM model is trained based on the time series data and the anti-overfitting strategy, and the TimeGAN-LSTM model is jointly optimized based on the adversarial loss, supervision loss and embedding loss according to the weights. The TimeGAN-LSTM model includes an autoencoder network, an adversarial network and a supervision network with a multi-layer LSTM structure. The supervisory network is transferred to an LSTM prediction network to obtain a state prediction model. The LSTM prediction network includes the input layer of the supervisory network, an LSTM network, and a fully connected layer. The LSTM network and the fully connected layer belong to the trained LSTM prediction model. The LSTM prediction model is trained based on historical state data of the entire life cycle of the device. The current state data is input into the state prediction model to obtain state data for multiple consecutive future times.

2. The method according to claim 1, characterized in that, The historical status data is collected through multiple types of sensors.

3. The method according to claim 1, characterized in that, The normalization process is achieved through Z-Score standardization.

4. The method according to claim 1, characterized in that, Before normalizing the historical state data, the historical state data undergoes data cleaning, which includes: Outliers are removed by using a sliding window, and then missing data is repaired using linear interpolation or mean fill.

5. The method according to claim 1, characterized in that, The overfitting prevention strategy includes early stopping.

6. The method according to claim 5, characterized in that, The process of training the TimeGAN-LSTM model based on the time series data and the anti-overfitting strategy, and jointly optimizing the TimeGAN-LSTM model according to the weights of adversarial loss, supervised loss, and embedding loss, includes: The time series data is divided into a training set and a validation set; The training set is input into the embedding network and the recovery network for training; Training is stopped when the adversarial loss, supervised loss, and embedding loss of the validation set do not decrease for several consecutive rounds.

7. The method according to claim 1, characterized in that, The process of migrating the supervised network to an LSTM prediction network includes: Freeze the autoencoder network and the adversarial network; Freeze the input layer of the supervisory network; Replace the multi-layer LSTM structure of the supervision network with the LSTM network; Replace the output layer of the supervisory network with the fully connected layer.

8. A method for assessing equipment condition, based on condition data obtained at multiple consecutive future time points using the method described in any one of claims 1-6, characterized in that, include: Determine whether the state data at each of the future consecutive time periods falls within the normal threshold range; If so, the device is in a healthy state at that moment; If it does not belong to the fault threshold range, determine whether the state data at each moment falls within the fault threshold range. If so, the device status at that moment is faulty; If it does not belong to the category, the device status at that moment is downgraded.

9. An electronic device, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke program instructions stored in the memory and execute the method of any one of claims 1-7 according to the obtained program.

10. A computer-readable non-volatile storage medium, characterized in that, It includes computer-readable instructions that, when read and executed by a computer, cause the computer to perform the method of any one of claims 1-7.

Images