Aviation power supply conversion module health state prediction method and system

Abstract

The invention relates to the technical field of power electronic equipment fault prediction and health management, in particular to an aviation power supply module health state prediction method and system. The method aims at solving the problems that in an existing prediction method, noise interference on health indexes is large, degradation trend extraction is insufficient, and long-term prediction precision of a model is insufficient. The method comprises the steps that firstly, an output voltage signal of an aviation power supply module serves as a health state sensitive parameter, the replacement entropy of the output voltage signal is calculated to construct an initial health index, and the index can effectively represent the complexity degradation process of a system; secondly, performing adaptive decomposition on the initial health index sequence by adopting feature mode decomposition (FDM) to obtain a series of feature mode components so as to accurately extract a smooth and monotonous degradation trend; then, performing single-step prediction on each characteristic modal component by using a bidirectional long-short-term memory network (AM-BiLSTM) in which an attention mechanism is introduced; and finally, reconstructing all prediction results to obtain a final health state prediction value. According to the method, the Boost converter is taken as an embodiment, the prediction error is obviously reduced compared with that of a traditional method through experimental verification, and higher accuracy and reliability are shown.

Description

Technical Field

[0001] This invention belongs to the field of avionics equipment condition monitoring and fault prediction, specifically involving a data-driven health status prediction method for aviation power conversion modules, and particularly a fusion solution based on signal entropy characteristics, adaptive signal decomposition and advanced deep learning technology. Background Technology

[0002] The power conversion module is the energy hub of avionics equipment, and its reliability is paramount. Accurate health status prediction is central to its predictive maintenance. Existing technologies face two major bottlenecks: First, the construction of health indicators. Directly using raw values ​​or simple features of electrical signals (such as output voltage) is susceptible to fluctuations in operating conditions and noise contamination, resulting in poor monotonicity and unclear trends, making it difficult to accurately characterize the slow degradation process. Second, the limitations of prediction models. Traditional time-series models struggle to capture the complex nonlinear dynamics of the degradation process. While deep learning models such as GRU and its bidirectional variant (BiGRU) have been applied, Long Short-Term Memory (LSTM) networks and their bidirectional variant (BiLSTM) have a recognized theoretical advantage in handling longer sequence dependencies, as their carefully designed "gating" mechanism more effectively avoids the gradient vanishing problem. However, even BiLSTM struggles to adaptively distinguish the importance of historical information when faced with raw health indicators or their simple decomposition components that mix different frequency features, leading to insufficient learning of key degradation inflection points and affecting long-term prediction accuracy. Therefore, there is an urgent need in this field for a technical solution that can construct health indicators with strong characterization capabilities, adopt decomposition methods that can extract degradation trends more accurately, and combine them with prediction models that can make full use of long-term contextual information and have the ability to filter key information, so as to comprehensively improve prediction performance. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for predicting the health status of aviation power modules with high accuracy and good stability. This method constructs a sensitive health index through permutation entropy, enhances its trend through FDM decomposition, and accurately predicts each component using an AM-BiLSTM model, ultimately achieving high-performance health status prediction.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the health status of an aircraft power module, characterized by comprising the following steps: S1: Construction of Health Indicators Based on Permutation Entropy Acquire the time-series signal of the output voltage of the aviation power module during operation. The permutation entropy of the output voltage signal is calculated to construct an initial health indicator sequence. Permutation entropy can quantify the complexity and randomness of time series. Its value exhibits a regular monotonic change as the internal components of the converter (such as electrolytic capacitors and power switching transistors) age, thus sensitively and robustly characterizing the degradation of health status.

[0005] S2: Health Indicator Decomposition and Trend Extraction Based on FDM The initial health indicator sequence was analyzed using the Eigenmode Decomposition (FDM) method. The signal is decomposed. Frequency Directional Modeling (FDM) is an adaptive decomposition method based on local signal features. It can decompose a signal into a series of Feature Mode Components (FMCs) from high frequency to low frequency according to the signal's scale characteristics. High-frequency FMCs mainly contain noise and random fluctuations, while low-frequency FMCs contain the main monotonic degradation trends of the device's health status. In this way, FDM can effectively separate the pure degradation information hidden in the original noisy HI.

[0006] S3: Component Prediction and Result Reconstruction Based on AM-BiLSTM Construct a bidirectional long short-term memory network (AM-BiLSTM) model incorporating an attention mechanism. Each feature modal component obtained from step S2 is then processed. Each component is treated as an independent time series and input into the AM-BiLSTM model for single-step prediction to obtain the predicted value of each component at the next time step. The AM-BiLSTM model, through its forward and backward LSTM layers, fully captures the past and future contextual information of each component sequence, thus gaining a more comprehensive understanding of its evolutionary patterns. It also automatically calculates the importance weights of the hidden states at each time step through an attention mechanism, focusing on the most critical degradation stages for predicting the next state. Finally, all predicted components are summed and reconstructed to obtain the final health state prediction value at time t+1. : .

[0007] S4: Iterative Prediction and State Assessment The latest predicted value Using the new observational data, the model input sequence is updated, and step S3 is repeated to perform multi-step predictions and obtain the evolution trajectory of the future health status. Based on this predicted trajectory and a preset fault threshold, the remaining service life of the converter can be assessed, or an early warning can be triggered before its performance degrades significantly.

[0008] The beneficial effects of this invention are: (1) Strong health index representation: Displacement entropy can capture subtle changes in dynamic characteristics caused by component aging in the output voltage sequence, is more sensitive to early degradation, and has a certain robustness to noise.

[0009] (2) Excellent predictive model performance: The innovative framework of “FDM-AM-BiLSTM prediction-reconstruction” is adopted. The BiLSTM model has a solid theoretical foundation in dealing with long sequence dependency problems. When combined with the attention mechanism, it can grasp the complete degradation context from both directions and dynamically focus on key degradation inflection points, making the prediction of each FMC component more accurate, thereby significantly improving the accuracy of the final health status prediction value, especially for long-term prediction. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a general block diagram of the technical solution of the present invention.

[0012] Figure 2 This is a diagram of the network structure of the BiLSTM of the present invention.

[0013] Figure 3 This is a network structure diagram of the AM-BiLSTM of the present invention. Detailed Implementation Plan The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] This method is described in detail using a Boost converter undergoing a high-temperature accelerated degradation experiment as an example.

[0015] S1: Construction of Health Indicators Based on Permutation Entropy Throughout the entire lifecycle experiment of the Boost converter, its output voltage signal was continuously acquired using a data acquisition card. The sampling frequency is 10kHz. A time window with a length of 1000 data points (0.1 seconds) is set, sliding in steps of 500 points (0.05 seconds). Within each window, the permutation entropy of the output voltage signal is calculated (parameter set to: embedding dimension). Time delay Arranging the calculated permutation entropy values ​​in chronological order constitutes the initial health indicator sequence. As the equivalent series resistance of the capacitor increases and the characteristics of the switching transistor degrade, It exhibits a clear, monotonous upward trend.

[0016] S2: Health Indicator Decomposition and Trend Extraction Based on FDM The Eigenmode Decomposition (FDM) algorithm is used to... The algorithm decomposes the sequence into five feature mode components based on its length and complexity. ~ ).in, This is the highest frequency component, mainly containing measurement noise; and The mid-frequency component reflects certain fluctuations in operating conditions; and These are low-frequency components, with smooth waveforms and a strong monotonic upward trend, collectively representing the core performance degradation process of the Boost converter.

[0017] S3: Component Prediction and Result Reconstruction Based on AM-BiLSTM S31: Input Layer Step S2 decomposes all component FMCs as prediction objects to construct supervised learning samples. Each sample defines the shape (50, 1) of the input tensor, indicating that each sample contains 50 time steps, and each time step has 1 feature.

[0018] S32: Attention Layer This is a custom layer, and its internal operations are as follows: S321: Score Calculation Features of the input at 50 time steps (each) It is a 64-dimensional vector, passed through a trainable weight matrix. and bias And using the tanh activation function, the score at each time step is calculated: .

[0019] S322: Weight Normalization 50 scores The attention weights at each time step are obtained by normalization using the Softmax function. : .

[0020] S323: Attention Weight Vector Features of each time step Its corresponding attention weight Multiply and then sum to obtain the final attention weight vector. : This vector It is a 64-dimensional vector that condenses the information most relevant to the training objective from the entire input sequence.

[0021] S33: Component Prediction Layer See attached Figure 3 BiLSTM consists of a feedforward LSTM containing past data information of the input sequence and a feedback LSTM containing future data information of the input sequence. By performing a concatenation operation on the feedforward and feedback LSTMs, BiLSTM can simultaneously consider contextual information before and after the current time step and can capture more long-term dependencies.

[0022]

[0023]

[0024]

[0025] S34: Result Reconstruction Layer During online operation, the latest normalized data points for each trend component are input into their corresponding trained AM-BiLSTM models to obtain the results. The predicted values ​​are summed and then inversely normalized to reconstruct the final health status prediction at time t+1. .

[0026] S4: Iterative Prediction and State Assessment The latest predicted value Using the new observational data, the model input sequence is updated, and step S3 is repeated to perform multi-step predictions and obtain the evolution trajectory of the future health status. Based on this predicted trajectory and a preset fault threshold, the remaining service life of the converter can be assessed, or an early warning can be triggered before its performance degrades significantly.

[0027] On an independent test set, the prediction results of our method (based on FDM and AM-BiLSTM) were compared with the actual health indicator values. Simultaneously, comparisons were made with methods based on ICEEMDAN-BiGRU and methods using only AM-BiLSTM without decomposition. Performance metrics (such as root mean square error RMSE and mean absolute error MAE) show that the prediction error of the proposed method is significantly lower than the comparative methods, demonstrating the combined advantages of FDM in trend extraction and AM-BiLSTM in sequence prediction.

[0028] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0029] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the health state of an airborne power module, characterized in that, The method comprises the following steps: S1: Health index construction based on permutation entropy: Collect the output voltage time series signal of the aviation power module during operation , Calculate the permutation entropy of the output voltage signal to construct the initial health index sequence ; S2: Health index decomposition and trend extraction based on FDM: Use the feature mode decomposition (FDM) method to decompose the initial health index sequence , Get a series of feature mode components (FMCs) from high frequency to low frequency ; S3: Component prediction and result reconstruction based on AM-BiLSTM: Input each feature mode component into the AM-BiLSTM model respectively for single-step prediction to get the predicted value of each component at the next time , Add all components to reconstruct the final health state prediction value at t+1 ; S4: Iterative prediction and state evaluation: Use the prediction value as new observation data to update the model input sequence, repeat step S3 for multi-step prediction to get the evolution trajectory of future health state, evaluate the remaining service life of the conversion module according to the preset fault threshold, and trigger the warning before the performance seriously degrades.

2. The method of claim 1, wherein, In step S2, the feature mode decomposition (FDM) is an adaptive signal decomposition method based on local feature scale of a signal, and a low-frequency feature mode component obtained by decomposition is used to represent a core monotone degradation trend of the aviation power module.

3. The method of claim 1, wherein, In step S3, the AM-BiLSTM model with the attention mechanism comprises the following steps: S31: an input layer, which constructs a supervised learning sample by taking all decomposed components FMC as a prediction object; S32: an attention layer, which is used to assign different attention weights to hidden states of each time step output by the BiLSTM layer, and perform weighted summation to generate a context vector focusing on key information; S33: a component prediction layer, which adopts a BiLSTM layer to extract forward and backward long-term time sequence features from an input sequence; and S34: a result reconstruction layer, which is used to map the context vector to a predicted value.

4. According to the AM-BiLSTM model structure described in claim 3, the S32 attention layer operation is as follows: S321 Score calculation: for the features of the input at 50 time steps... (each) It is a 64-dimensional vector, passed through a trainable weight matrix. and bias And using the tanh activation function, the score at each time step is calculated: S322 Weight Normalization: Normalize the 50 scores The attention weights at each time step are obtained by normalization using the Softmax function. : S323 Attention Weight Vector: This vector combines the features at each time step. Its corresponding attention weight Multiply and then sum to obtain the final attention weight vector. : The vector It is a 64-dimensional vector that condenses the information most relevant to the training objective from the entire input sequence.

5. The method of claim 1, wherein, In step S3, the single-step prediction of the at least one low-frequency feature mode component is that a BiLSTM model with the attention mechanism is trained for each low-frequency feature mode component to be predicted.

6. The method of claim 1, wherein, The aviation power module is an aviation power module, and is preferably a Boost converter.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method according to any one of claims 1 to 5. The program is executed by the processor to implement the steps of the method according to any one of claims 1 to 5.

Images