Power device health degree prediction method based on deep learning

Abstract

The invention belongs to the technical field of mechanical health prediction, and particularly discloses a power device health degree prediction method based on deep learning. The method comprises the following steps: dividing signals of detection points according to working condition types; inputting various working condition domain data into a variational auto-encoder to which domain confrontation constraints are applied, and generating feature representations irrelevant to working condition types; carrying out trend evaluation on the time sequence of all feature representations, and screening out a feature representation subset which is most relevant to the health degradation process; deducing the health degree of the detection point according to each feature representation in the feature representation subset; representing the health degree of the power device by the weighted sum of the health degrees of all the detection points; and extracting local features of the health degree of the power device through the time domain convolutional network, and obtaining a predicted value of the health degree of the power device after the long-short-term memory network captures a dynamic evolution rule based on the local features. Compared with the prior art, the method is higher in health prediction and evaluation precision of the power device in a multi-working-condition and strong-interference environment.

Description

Technical Field

[0001] This application belongs to the field of mechanical health prediction technology, and more specifically, relates to a deep learning-based method for predicting the health of power units. Background Technology

[0002] The health status of rotating equipment such as power plants is a crucial indicator of a ship's normal operation. Health status assessment and failure prediction of power plants are of great significance for the ship's health status and maintenance strategies.

[0003] With the development of neural networks and computer technology, the emergence of deep learning has provided a new direction for the health assessment and prediction of power plant functions and for fault prediction. Deep learning, based on more complex neural networks, can process large amounts of data and has excellent feature extraction capabilities. Patent document CN202211273036.8 discloses a health monitoring method for marine main propulsion diesel engines, which uses a deep learning network to construct a health detection model and achieves health prediction of marine power plants. However, this method suffers from insufficient prediction accuracy under various operating conditions and high-noise environments. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this application is to provide a deep learning-based method for predicting the health of power plants, which aims to solve the technical problem of insufficient accuracy in health assessment and fault prediction under multiple operating conditions and high noise environments.

[0005] The first aspect of this application relates to a deep learning-based method for predicting the health of power plants, comprising: The signals from the power unit detection points are divided into source operating condition domain data and target operating condition domain data according to the type of operating condition they are in. The source operating condition domain data and the target operating condition domain data are input into a variational autoencoder with applied domain adversarial constraints to generate a common feature representation that is independent of the type of operating condition. A trend assessment is performed on the time series of all common feature representations to select the subset of feature representations most relevant to the health degradation process; The health status of the detection point is derived based on the common feature representations in the feature representation subset; the overall health status of the power unit is characterized by the weighted sum of the health statuses of all detection points. The local features of the time series of the power unit's health status are extracted by a temporal convolutional network, and then the predicted value of the power unit's health status is obtained by capturing the dynamic evolution law of the time series based on the local features through a long short-term memory network.

[0006] Preferably, the variational autoencoder with applied domain adversarial constraints specifically involves: training and learning feature representations of the source and target operating condition domain data through the encoder and decoder in the variational autoencoder; classifying the feature representations for operating condition domains through a domain discriminator; setting a gradient inversion layer between the domain discriminator and the encoder for adversarial training; the gradient inversion layer normally transmits data during forward propagation, and multiplies the gradient from the domain discriminator by a negative coefficient before transmitting it back to the encoder during backward propagation, so that the variational autoencoder generates common feature representations independent of operating condition types.

[0007] Preferably, the variational autoencoder that applies domain adversarial constraints specifically includes: The encoder employs a structure combining one-dimensional convolutional layers and fully connected layers to map the original data of the source and target operating domain data to parameters of a latent Gaussian distribution, and samples the parameters to obtain the feature representation. The decoder, symmetrical to the encoder structure, adopts a structure combining fully connected layers and one-dimensional deconvolutional layers to reconstruct the feature representation and obtain reconstructed data. The domain discriminator employs a structure combining a gradient inversion layer and a fully connected layer to determine the operating domain to which the feature representation belongs.

[0008] Preferably, the variational autoencoder with applied domain adversarial constraints is trained by minimizing a comprehensive loss function, wherein the comprehensive loss function comprises a weighted sum of reconstruction loss, KL divergence loss, and domain classification loss; The reconstruction loss is used to measure the difference between the original data and the reconstructed data; The KL divergence loss is used to measure the difference between the approximate posterior distribution of the feature representation and the preset standard normal prior distribution; The domain classification loss is used to measure the difference between the operating domain to which the feature representation belongs, as determined by the domain discriminator, and the actual operating domain to which the feature representation belongs.

[0009] Preferably, the training process of the temporal convolutional network and the long short-term memory network is as follows: A historical time series of the health status of a power unit is used as a training sample and input into the temporal convolutional network. The output of the temporal convolutional network is used as the input of the long short-term memory network, and the long short-term memory network outputs a predicted value of the health of the power unit. With the goal of minimizing the error between the predicted and actual values ​​of the power unit's health, the parameters of the temporal convolutional network and the long short-term memory network are optimized using the backpropagation algorithm.

[0010] Preferably, a trend evaluation is performed on the time series of all feature representations to select the subset of feature representations most relevant to the health degradation process, specifically: The monotonicity of the time series represented by all common features is measured, wherein the monotonicity characterizes the strength of the trend of continuous rise or fall in the common feature representation; The correlation coefficients between all common feature representations and time series are used as time correlation, which characterizes the degree of linear change of the common feature representations over time. The comprehensive trend score is obtained by summing the absolute values ​​of the monotonicity and time correlation. The common feature representations with the highest preset proportion of the comprehensive trend score are taken to form the feature representation subset.

[0011] Preferably, the health status of the detection point is derived based on the common feature representations in the subset of feature representations, specifically as follows: Derive the proportion of common feature representations in the feature representation subset that exceed a preset anomaly threshold; The health status of the detection points is negatively correlated with the proportion.

[0012] Preferably, the signal at the power device detection point is specifically the vibration signal of the power device.

[0013] In a second aspect, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0014] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: (1) Ship propulsion systems often operate under multiple operating conditions and high noise environments. When extracting data features based on traditional deep learning network models, the feature distribution of the extracted features is severely offset under different operating conditions, resulting in a decrease in the generalization ability of the final model. The method in this application introduces domain adversarial constraints into the variational autoencoder. Through adversarial domain classification training, the variational autoencoder is forced to learn the common feature representations between data from different operating condition domains, suppressing the interference specific to different operating conditions, enhancing the robustness and generation ability of feature representations, thereby maintaining the consistency of feature representations under different operating conditions, and ultimately improving the accuracy of health prediction.

[0015] (2) This application evaluates the trend of the time series of feature representation, selects the key features most related to the health degradation process to deduce the health of the detection points, and then reflects the overall status of the whole machine by weighted fusion of the health of multiple detection points, so as to avoid misjudgment caused by local strong noise or failure of a single detection point.

[0016] (3) Single long short-term memory networks are prone to gradient problems in long-sequence health prediction, while traditional convolutional neural networks are difficult to capture long-term dependencies. This application combines the causal dilated convolution of temporal convolutional networks with the gating mechanism of long short-term memory networks. Temporal convolutional networks effectively extract local temporal features, and long short-term memory networks model long-term evolution patterns on this basis. The two complement each other to enhance the temporal modeling capability. Attached Figure Description

[0017] Figure 1 This is a flowchart of the power unit health prediction method provided in the embodiments of this application.

[0018] Figure 2 This is a flowchart illustrating the feature extraction and prediction processes provided in the embodiments of this application.

[0019] Figure 3a This is a waveform diagram of the vibration acceleration at the first detection point provided in the embodiments of this application. Figure 3b This is a waveform diagram of the vibration acceleration at the second detection point provided in the embodiments of this application. Figure 3c This is a waveform diagram of the vibration acceleration at the third detection point provided in the embodiments of this application. Figure 3d This is a waveform diagram of the vibration acceleration at the fourth detection point provided in the embodiments of this application.

[0020] Figure 4 This is a trend graph of the effective value of vibration acceleration at one of the detection points provided in the embodiments of this application.

[0021] Figure 5 This is a spectrum of vibration acceleration at a certain detection point on a certain day, provided in an embodiment of this application.

[0022] Figure 6 This is a spectrum of vibration acceleration at a certain detection point on another day, provided in an embodiment of this application.

[0023] Figure 7a This is a health curve diagram of the power unit provided in the embodiments of this application. Figure 7b This is a health status curve predicted from data collected 14 days in advance. Figure 7c This is a health status curve predicted from data collected 13 days in advance. Figure 7d This is a health status curve predicted from data collected 12 days in advance. Figure 7e This is a health status curve predicted from data collected 11 days in advance. Figure 7f It is a health status curve predicted from data 10 days in advance.

[0024] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] In this application, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order of objects. For example, "first operating condition domain data" and "second operating condition domain data," etc., are used to distinguish different operating condition domain data, not to describe a specific order of operating condition domain data.

[0027] In this application, the term "electrical connection" can refer to a direct circuit connection or a signal transmission via a communication protocol.

[0028] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0029] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple operating condition types refer to two or more operating condition types, multiple operating condition domain data refer to two or more operating condition domain data, etc.

[0030] First, the technical terms used in the embodiments of this application will be introduced.

[0031] Domain-Adversarial (DA) is a transfer learning method that introduces a domain discriminator to compete against a feature extractor, aiming to learn domain-invariant feature representations to reduce distributional differences between different data domains.

[0032] Variational autoencoder (VAE) is a generative model that learns the latent probability distribution of data through an encoder, samples from that distribution, and reconstructs the data through a decoder, thus combining data generation and representation learning.

[0033] Temporal Convolutional Network (TCN) is a convolutional neural network structure used for modeling time-series data. It employs causal convolution and dilated convolution to capture short- to medium-term dependencies and has high parallel computation efficiency.

[0034] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network that effectively alleviates the vanishing / exploding gradient problem by introducing input gates, forget gates, and output gates, and excels at learning long-term dependencies.

[0035] The embodiments of this application are described below with reference to the accompanying drawings.

[0036] Example 1: like Figure 1 The diagram illustrates a deep learning-based method for predicting the health of a power unit, as described in Embodiment 1 of this application. The method specifically includes the following steps: Step 1: Divide the signals from the power unit detection points into source operating condition domain data and target operating condition domain data according to the type of operating condition they are in.

[0037] Step 2: Input the source operating condition domain data and the target operating condition domain data into the variational autoencoder that applies domain adversarial constraints to generate a common feature representation that is independent of the type of operating condition.

[0038] Step 3: Perform a trend assessment on the time series of all common feature representations and select the subset of feature representations that are most relevant to the health degradation process.

[0039] Step 4: Derive the health status of the detection points based on the common feature representations in the feature representation subset; use the weighted sum of the health statuses of all detection points to characterize the overall health status of the power unit.

[0040] Step 5: Extract local features of the time series of the power unit's health status through a temporal convolutional network, and then obtain the predicted value of the power unit's health status by capturing the dynamic evolution of the time series based on the local features through a long short-term memory network.

[0041] Example 2: like Figure 2 The diagram shows the feature extraction and prediction process in Embodiment 2 of this application, which includes the following steps: Step 1: Data Acquisition and Operating Condition Identification.

[0042] Raw vibration signals are collected from multiple detection points of the power unit and then divided according to the operating conditions.

[0043] In this embodiment, vibration signals collected under stable operating conditions are classified as source domain data. Vibration signals collected under startup operating conditions are classified as target domain data.

[0044] Step 2: Extraction of cross-condition feature representation.

[0045] A domain adversarial constraint variational autoencoder (DA-VAE) model is used to learn and extract common feature representations that ignore differences in operating conditions from source and target domain data belonging to different operating condition categories. Specifically, this includes: Step 21: Construct the basic variational autoencoder (VAE) model.

[0046] This step aims to build a generative model capable of learning the latent probability distribution of the data, where the encoder output will serve as the subsequent feature representation vector. In this implementation, the encoder employs a combination of a one-dimensional convolutional neural network and a fully connected layer to map the input source and target domain data to parameters (mean) of a latent Gaussian distribution. and variance Subsequently, the parameters are sampled to obtain the feature representation of the data, and the obtained feature representation is reconstructed into the original input through a symmetric decoder. The training of the VAE model is achieved by minimizing the reconstruction loss and the KL (Kullback-Leibler, relative entropy) divergence loss, thereby ensuring that the encoder can learn the essential, compressed feature representation of the data.

[0047] Step 22: Introduce Domain Adversarial (DA) constraints.

[0048] To overcome the impact of data distribution differences across different operating conditions on feature representations, a domain adversarial training mechanism was introduced on top of the VAE model. The core of this mechanism is the addition of a domain discriminator, whose task is to determine whether the feature representation extracted by the encoder originates from the source domain or the target domain. During training, a gradient inversion layer connects the encoder and the domain discriminator. The gradient inversion layer transmits data normally during forward propagation, but during backward propagation, it multiplies the gradient from the discriminator by a negative coefficient before transmitting it back to the encoder. This forces the encoder to learn how to generate common feature representations that the domain discriminator cannot distinguish between the data source operating conditions in an "adversarial" process with the discriminator. This compels the encoder to ignore the specificity of different operating conditions and focus on extracting common feature representations across operating conditions.

[0049] Step 23: DA-VAE joint training.

[0050] This step is crucial for integrating VAEs and domain adversarial mechanisms for collaborative training. During training, source and target domain data are input simultaneously. During forward propagation, feature representations are extracted from both types of data using a shared encoder, and the reconstruction loss and KL divergence loss of the VAE are calculated. Simultaneously, the extracted feature representations are fed into the domain discriminator to calculate the domain classification loss. During backpropagation, the total loss is a weighted sum of the aforementioned losses. Through this joint optimization, the model simultaneously pursues multiple optimization objectives: The VAE variational autoencoder part is dedicated to accurately reconstructing data while maintaining potential spatial regularity; The domain adversarial component aims to obfuscate the types of operating conditions the data belongs to. Ultimately, the encoder obtained after training convergence is able to extract common feature representations across operating conditions.

[0051] Step 3: Feature representation screening and health calculation.

[0052] The extracted feature representations are evaluated and screened for trends, and the health status of each detection point and power unit is calculated based on the screened feature representations.

[0053] Step 31: Trend evaluation and screening of feature representations.

[0054] This step performs a trend assessment on the time series of all common feature representations, selecting the subset of feature representations most relevant to the health degradation process. The health degradation process exhibits both monotonicity and time correlation; therefore, the time series of common feature representations are selected based on these two indicators. Monotonicity measures the strength of the upward or downward trend of a common feature representation throughout the time series; the closer the value is to 1 or -1, the stronger the trend. Time correlation assesses the degree of linear change of a common feature representation over time using the Pearson coefficient. The absolute value of the monotonicity and the absolute value of the time correlation for each common feature representation are added together to obtain a comprehensive score. Based on this score, the highest-scoring common feature representations are selected to form the feature representation subset. These highly trending common feature representations can more stably reflect the irreversible degradation process of the power plant.

[0055] In this embodiment, the top 20% of the features with the highest scores are selected to form a subset of the feature representations.

[0056] Step 32: Calculate the health status of the detection point.

[0057] Quantify health status at each testing point level. (This refers to the process of quantifying health status at each testing point.) and The filtered feature representations are first used to set an anomaly threshold for each detection point based on historical normal data. At each time step... Statistical monitoring points health The proportion of common feature representations exceeding the anomaly threshold in the feature representation subset is negatively correlated, as shown in the following formula: ; in, This is an indicator function that returns 1 if the condition it contains is true, and 0 otherwise. For the first Each feature represents, To preset the abnormal threshold, This represents the number of feature representations in the feature representation subset.

[0058] Health It directly reflects the proportion of abnormal features at that local location; the higher the proportion, the worse the health status of that detection point.

[0059] Step 33: Health status of the power unit.

[0060] The health status of the power unit is characterized by a weighted sum of the health status of all detection points. : ; in, This represents the total number of detection points in the power unit. Indicates the detection point health Indicates the detection point Weights, weights According to the detection point Assigning susceptibility to faults, data reliability, or engineering experience. Final power unit health. The time series data constitutes a continuous health curve of the power unit. This curve can comprehensively and quantitatively reflect the evolution of the overall performance of the power unit, providing a core basis for subsequent prediction and decision-making.

[0061] Step 4: Time series modeling, health assessment and prediction, and fault prediction.

[0062] A hybrid model combining a TCN-LSTM temporal convolutional network and a long short-term memory network is used to learn the time series of power unit health status, enabling future health assessment, prediction, and fault early warning. Specifically, this includes: Step 41: Construct the TCN-LSTM hybrid network architecture.

[0063] This step aims to combine the local feature extraction capabilities of TCN with the long-term dependency modeling advantages of LSTM. First, a temporal convolutional network is constructed as the front-end feature extractor. TCN ensures that the output depends only on the current and historical inputs through causal convolutions, expands the receptive field to capture longer historical information through dilated convolutions, and stabilizes gradient propagation through residual connections. Then, the output of TCN is used as the input to the LSTM layers. LSTM, through its gating mechanism, further learns the evolution patterns and long-term dependencies across longer time periods in the health time series, building upon the short-term trends extracted by TCN. This cascaded structure effectively avoids the gradient problem that easily occurs when LSTM processes long sequences alone.

[0064] Step 42: Model training and prediction.

[0065] The constructed TCN-LSTM model is trained under supervised supervision using historical health time series data. During training, a time series data of the power unit's health is used as input, with the actual health data of one or more subsequent time steps as labels. The TCN-LSTM model parameters are optimized using backpropagation, aiming to minimize the error between the predicted and actual values. After training, the TCN-LSTM model possesses the ability to learn the evolutionary patterns of health from historical health time series data. When making health predictions, only the latest health time series data needs to be input, and the TCN-LSTM model can output predicted health values ​​for a future period. By analyzing the downward trend of the predicted health values ​​or comparing them with a set health threshold, potential faults can be predicted (early warning).

[0066] Step 43: Multi-step prediction and performance verification.

[0067] To evaluate the practicality of the TCN-LSTM model, multi-step advance prediction validation is required. Specifically, starting with historical data of varying lengths (e.g., data from 10, 12, and 14 days in advance), the model is asked to predict future health status up to the point of failure. By comparing the root mean square error (RMSE) and mean absolute error (MAE) of the predicted health status curves with the actual health status curves under different lead times, the longest lead time for effective early warning by the TCN-LSTM model can be determined. This validation directly demonstrates the feasibility and accuracy of the TCN-LSTM model in early health status and failure prediction.

[0068] Step 5: Model Validation and Engineering Application.

[0069] The complete process described above was validated on two real power unit failure cases (block failure and fuel pump shaft failure), and the prediction results were used to support maintenance decisions.

[0070] Example 3: This application's embodiment 3 specifically includes the following steps: Step 1: Construct a dataset based on actual bench test vibration data of a power unit (fuel pump shaft breakage failure). Eight vibration acceleration monitoring points were arranged at key locations on the core load-bearing structure of the power unit. During data processing, approximately three months of continuous operational data were first extracted and aggregated by day. Then, Z-score standardization (Standard Score standardization) was performed on the time-domain data of each monitoring point to eliminate dimensional differences.

[0071] To train the DA-VAE model, the data is divided into different operating domains based on the speed of the power unit or the operating stage (e.g., the stable operating segment is the source domain, and the fault segment is the target domain).

[0072] To train the TCN-LSTM prediction model, a rolling window approach is used to construct training samples and prediction labels for the health sequence of the power unit.

[0073] During the data processing phase, to compare the data characteristics before and after equipment maintenance, time-domain indicators such as the effective value of acceleration, true peak value, and total velocity are synchronously calculated from the original vibration acceleration signal, and a spectrum diagram is generated for comparative analysis. Waveform diagrams of vibration acceleration from the first, second, third, and fourth detection points are also provided. Figure 3a , Figure 3b , Figure 3c and Figure 3d ) and the trend graph of the effective value of vibration acceleration at one of the detection points ( Figure 4 As can be seen, traditional vibration indices can exhibit significant abrupt changes when a fault occurs, but lack a clear degradation trend before the fault occurs, and are significantly affected by changes in operating conditions, making them unsuitable for direct stability assessment and early warning. (Comparison of vibration acceleration spectrum diagrams) Figure 5 and Figure 6 This further reveals the significant changes in spectral components under fault conditions.

[0074] Step 2: Construct and train a domain adversarial constraint variational autoencoder model to extract generalizable common feature representations from multi-condition vibration data. The model includes a VAE and a domain discriminator.

[0075] VAE consists of an encoder and a decoder. The encoder uses a three-layer one-dimensional convolutional neural network to extract local temporal features, followed by a fully connected layer to output parameters (mean) of a Gaussian distribution in the latent space. With log variance The encoder samples the parameters to obtain feature representations. The decoder is symmetrical to the encoder, and reconstructs the feature representations back into the original signal through fully connected layers and deconvolutional layers.

[0076] The domain discriminator consists of a fully connected network containing a gradient inversion layer, used to distinguish whether the feature representation comes from the source domain or the target domain.

[0077] The model training process employs a joint optimization strategy, simultaneously inputting standardized vibration data from both the source and target domains. The VAE loss includes reconstruction loss (mean squared error) and KL divergence loss (to approximate the standard normal distribution of the latent distribution). Domain adversarial loss utilizes binary cross-entropy loss, implemented through a gradient inversion layer to force the encoder to generate domain-invariant common feature representations. The total loss is a weighted sum, optimized simultaneously through backpropagation of the encoder, decoder, and discriminator parameters.

[0078] After training, the encoder weights are fixed, and the preprocessed vibration data is input into the encoder to output its latent spatial representation, which is the required cross-condition common feature representation vector. These common feature representations will replace traditional time-domain indicators as the basic input for subsequent health assessments.

[0079] Step 3: Perform trend screening on the common feature representations extracted by DA-VAE, and calculate the overall health of the power unit based on the screening results.

[0080] First, for each common feature representation, its monotonicity (reflecting a continuous upward or downward trend) and time correlation (reflecting the linear correlation with changes over time) indices are calculated. Then, the absolute values ​​of the monotonicity and time correlation for each common feature representation are summed to obtain a comprehensive trend score. All feature representations are then sorted from highest to lowest score. Finally, the top 20% of common feature representations are selected to form the final subset of feature representations used for health assessment; these representations exhibit the strongest indicative power of degradation.

[0081] The health assessment first performs a point-level health assessment. For each vibration monitoring point, an anomaly threshold is set for each common feature representation after filtering, based on its historical normal data. At each time point, the number of common feature representations exceeding the threshold at that monitoring point is counted, and this number is divided by the total number of common feature representations to obtain the anomaly score rate for that monitoring point, which serves as the real-time health indicator for that monitoring point. Next, the overall machine health assessment is fused. Weights are assigned to each monitoring point based on its physical location and contribution to the overall machine condition. The power unit health assessment is obtained by weighted summation of the health assessments of all monitoring points.

[0082] Step 4: Use the TCN-LSTM model to predict the health of the power unit.

[0083] Historical data on the health of the power unit are input into the TCN-LSTM model, which outputs a predicted health value. Figure 7a It is a continuous time series curve of the overall health of the machine, which comprehensively reflects the overall performance degradation state of the power unit. Figures 7b-7f The figures compare the predicted health status and the actual health status 14, 13, 12, 11, and 10 days in advance. As can be seen from the figures, the TCN-LSTM model can achieve relatively accurate predictions within 12 days in advance.

[0084] By setting a health threshold, a health level exceeding the threshold is defined as a fault, thereby enabling fault prediction.

[0085] Based on the methods in the above embodiments, this application provides an electronic device, such as... Figure 8As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the methods in the above embodiments.

[0086] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0087] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0088] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0089] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0090] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0091] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0092] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0093] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the health of a power unit based on deep learning, characterized in that, include: The signals from the power unit detection points are divided into source operating condition domain data and target operating condition domain data according to the type of operating condition they are in. The source operating condition domain data and the target operating condition domain data are input into a variational autoencoder with applied domain adversarial constraints to generate a common feature representation that is independent of the type of operating condition. A trend assessment is performed on the time series of all common feature representations to select the subset of feature representations most relevant to the health degradation process; The health status of the detection point is derived based on the common feature representations in the feature representation subset; the overall health status of the power unit is characterized by the weighted sum of the health statuses of all detection points. The local features of the time series of the power unit's health status are extracted by a temporal convolutional network, and then the predicted value of the power unit's health status is obtained by capturing the dynamic evolution law of the time series based on the local features through a long short-term memory network.

2. The method for predicting the health of a power unit according to claim 1, characterized in that, The variational autoencoder with applied domain adversarial constraints specifically involves: training and learning feature representations of the source and target operating condition domain data through the encoder and decoder in the variational autoencoder; classifying the feature representations for operating condition domains through a domain discriminator; and setting a gradient inversion layer between the domain discriminator and the encoder for adversarial training. The gradient inversion layer transmits data normally during forward propagation, and multiplies the gradient from the domain discriminator by a negative coefficient before transmitting it back to the encoder during backward propagation, so that the variational autoencoder generates common feature representations that are independent of the type of operating condition.

3. The method for predicting the health of a power unit according to claim 2, characterized in that, The variational autoencoder that applies domain adversarial constraints specifically includes: The encoder employs a structure combining one-dimensional convolutional layers and fully connected layers to map the original data of the source and target operating domain data to parameters of a latent Gaussian distribution, and samples the parameters to obtain the feature representation. The decoder, symmetrical to the encoder structure, adopts a structure combining fully connected layers and one-dimensional deconvolutional layers to reconstruct the feature representation and obtain reconstructed data. The domain discriminator employs a structure combining a gradient inversion layer and a fully connected layer to determine the operating domain to which the feature representation belongs.

4. The method for predicting the health of a power unit according to claim 3, characterized in that, The variational autoencoder with applied domain adversarial constraints is trained by minimizing the comprehensive loss function, which comprises a weighted sum of reconstruction loss, KL divergence loss, and domain classification loss. The reconstruction loss is used to measure the difference between the original data and the reconstructed data; The KL divergence loss is used to measure the difference between the approximate posterior distribution of the feature representation and the preset standard normal prior distribution; The domain classification loss is used to measure the difference between the operating domain to which the feature representation belongs, as determined by the domain discriminator, and the actual operating domain to which the feature representation belongs.

5. The method for predicting the health of a power unit according to claim 1, characterized in that, The training process for the temporal convolutional network and the long short-term memory network is as follows: A historical time series of the health status of a power unit is used as a training sample and input into the temporal convolutional network. The output of the temporal convolutional network is used as the input of the long short-term memory network, and the long short-term memory network outputs a predicted value of the health of the power unit. With the goal of minimizing the error between the predicted and actual values ​​of the power unit's health, the parameters of the temporal convolutional network and the long short-term memory network are optimized using the backpropagation algorithm.

6. The method for predicting the health of a power unit according to claim 1, characterized in that, A trend evaluation was performed on the time series of all feature representations, and the subset of feature representations most relevant to the health degradation process was selected, specifically: The monotonicity of the time series represented by all common features is measured, wherein the monotonicity characterizes the strength of the trend of continuous rise or fall in the common feature representation; The correlation coefficients between all common feature representations and time series are used as time correlation, which characterizes the degree of linear change of common feature representations over time. The comprehensive trend score is obtained by summing the absolute values ​​of the monotonicity and time correlation. The common feature representations with the highest preset proportion of the comprehensive trend score are taken to form the feature representation subset.

7. The method for predicting the health of a power unit according to claim 1, characterized in that, The health status of the detection point is derived based on the common feature representations in the subset of feature representations, specifically as follows: Derive the proportion of common feature representations in the feature representation subset that exceed a preset anomaly threshold; The health status of the detection points is negatively correlated with the proportion.

8. The method for predicting the health of a power unit according to claim 1, characterized in that, The signal at the power unit detection point is specifically the vibration signal of the power unit.

9. An electronic device, characterized in that, Includes memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-8.

Images