Health degree assessment method for multi-mode autonomous embedded control system

Abstract

The invention relates to the technical field of health evaluation management of complex measurement and control systems, and discloses a health degree evaluation method of a multi-mode autonomous embedded control system, and the evaluation method comprises the following steps: (1) obtaining physical quantities capable of reflecting the operation state of the system to form a data set; (2) constructing a high-dimensional feature vector through abnormal point elimination and evaluation starting point prediction; (3) further extracting features by using the Jinguo convolutional neural network; (4) training the constructed encoding and decoding mulberry short-time memory neural network based on the attention mechanism; (5) performing health degree evaluation based on the trained model; (6) performing fitting regression on the health degree index of the system; and (7) improving the real-time performance, introducing position coding and improving a codec structure, and carrying out health degree real-time evaluation. According to the method provided by the invention, offline and real-time health degree evaluation is realized, and the method has relatively good practicability and evaluation accuracy.

Description

Technical Field

[0001] The patent of this invention belongs to the field of fault prediction and health management technology, and particularly relates to the construction and health prediction of health indicators of an embedded control system with multiple working modes. Background Art

[0002] With the rapid development of modern weapons and equipment, the complexity, integration and intelligence of equipment are constantly improving, and PHM technology has emerged. For example, the Chinese invention patent with the announcement number CN113602526A discloses a "verification test method and system for aircraft electromechanical fault prediction and health management system", which uses data simulation and splicing cycles to inject fault modes and realize the verification test process of electromechanical PHM system. This method of simulated fault injection cannot truly reflect the complex fault modes in the real working environment, lacks dynamic environmental factors and has limited mapping capabilities for real system performance, and has certain defects.

[0003] The internal components of embedded systems are complex and closely interrelated, and their operating status directly affects the overall performance and service quality of system equipment. In practical applications, the health status of embedded systems not only determines whether the equipment can perform tasks efficiently and stably, but is also closely related to the safety, reliability and life of the system. Therefore, how to effectively monitor the state changes during the operation of the system and build a scientific health assessment mechanism has become an important direction in the current research field. By conducting in-depth analysis of the real-time monitoring data collected during the operation of the system, key health indicators reflecting the health status of the system can be extracted, thereby achieving an accurate assessment of the health status of the embedded system. This method can not only provide support for the early identification of potential faults, but also provide a scientific basis for optimizing maintenance strategies and extending the service life of equipment. However, the existing technology still has a lot of room for improvement in the scientific nature of indicator construction and the accuracy of health assessment. Therefore, in-depth research and optimization in this field has important theoretical value and application prospects. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art such as insufficient authenticity and complexity of fault simulation injection data, insufficient fault mode coverage, lack of dynamic environmental factors, and poor real-time performance of the testing process, the present invention provides a health trend indicator constructed based on monitoring data, performs a health assessment prediction algorithm based on a bidirectional long short-term memory network, and provides an optimization method for real-time health assessment.

[0005] The technical solution adopted by the present invention to solve its technical problem is: a health assessment method for a multi-mode autonomous embedded control system, in which the voltage, current, temperature and other physical quantities that can reflect the health status of the system under different working modes and different health states have different change trends. A health index is constructed by analyzing the trend change characteristics, and then the health of the embedded system is predicted based on a neural network. The method mainly includes the following steps:

[0006] S1: First, the key node physical quantities that can reflect the internal health status of the embedded system are obtained through the data acquisition module inside the embedded system.

[0007] S2: Then remove random fluctuations and noise interference, determine the prediction starting point, and construct high-dimensional features through time-frequency domain feature analysis.

[0008] S3: The high-dimensional features after screening are input into the bidirectional long short-term memory neural network (LSTM) based on the attention mechanism to obtain the degradation trend of the health indicators of the system during long-term power-on operation. The remaining life prediction value reflecting the health of the embedded control system is obtained by fitting it.

[0009] S4: The neural network consists of an encoder-decoder. The input end of the encoder-decoder structure is a single-layer bidirectional long short-term memory neural network, and the position encoding is compressed. The physical parameters of the system at different stages are extracted, and the encoder is replaced by a convolutional feature extractor based on contrastive learning.

[0010] Furthermore, in step S1, a data acquisition module inside the multi-mode embedded system acquires physical quantities such as voltage, current, and temperature that can reflect the health of the system.

[0011] Furthermore, in step S2, the least squares support vector machine (LSSVM) is first used to perform outlier detection based on local time series prediction (Predictive Sliding Local Outlier Correction, PSLOC) on the monitored physical quantities that can reflect the health status of the system to remove random fluctuations and noise interference. And the prediction starting point is detected based on the state change rate (State Change Rate, SCR). Since the physical parameters of the system cannot directly reflect the degradation trend, the time domain and frequency domain features are extracted, and through monotonicity and trend calculations, several features suitable for health assessment are screened out to construct health indicators.

[0012] Furthermore, in step S3, the network model is composed of two long short-term memory neural networks, which serve as encoder and decoder respectively. An attention mechanism is introduced into the encoder-decoder architecture to focus on important parts and reduce attention to irrelevant parts.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] The present invention considers both time domain and frequency domain characteristics to construct health indicators, effectively covering health modes under different modes and working conditions, and provides a way for health assessment and management of complex embedded measurement and control systems.

[0015] The present invention adopts a bidirectional long short-term memory neural network with an added attention mechanism for health assessment, which effectively improves the ability to process long time series under long-term power-on conditions of embedded systems.

[0016] The present invention introduces an attention mechanism into the encoding and decoding structure, which effectively enhances the ability to process long time series.

[0017] The present invention adopts a dense and sparse extraction strategy for the monitored physical parameters, optimizes position coding and introduces a convolutional feature extractor to complete the health assessment of the last collected data before the next data collection is completed, effectively improving the real-time performance of the health assessment method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0019] Figure 1 A schematic diagram of a process of health assessment prediction starting point and high-dimensional feature construction in the health assessment of a multi-mode autonomous embedded system of the present invention;

[0020] Figure 2 Schematic diagram of encoder-decoder based on attention mechanism;

[0021] Figure 3 This is a diagram of the health assessment framework;

[0022] Figure 4 Schematic diagram of continuous feature and predictive feature extraction based on contrastive learning. DETAILED DESCRIPTION

[0023] The invention will be described in detail below in conjunction with the accompanying drawings and specific implementations. Obviously, the implementation described is only a part of the embodiments of the present invention, but not all of them.

[0024] Example 1

[0025] S1: The multi-mode autonomous embedded system deploys sensors that can monitor voltage, current, and temperature, which are integrated into the data acquisition module. The required data can be obtained through the data transmission and storage module. The data acquisition time for each time is 0.1s.

[0026] S20: Figure 1Shown is a schematic diagram of the high-dimensional feature screening process designed by the present invention.

[0027] S21: Eliminate outliers through local prediction sliding window. Introduce Lagrangian operator to optimize LSSVM problem and get the final LSSVM regression function as The embedded system may be affected by random noise during operation, which has a great impact on health assessment and should be removed. The PSLOC algorithm is used to remove such abnormal points, using the formula:

[0028]

[0029] The local outlier factor of point p can be calculated. When the value of this formula is greater than 1, it means that the point is sparse relative to the surrounding points and is more likely to be an outlier. To remove outliers such as random fluctuations and noise interference in the monitoring signal, first perform LSSVM regression calculation on the divided sliding window signal, and set the sliding window range to 8% of the laboratory data; then perform multi-step prediction on the window signal, and the prediction step length is 4% of the overall experimental data; secondly, compare the difference between the predicted value and the measured data with the threshold to remove outliers.

[0030] S22: Determine the starting point of the health assessment prediction based on the state change rate. When the state change rate of the sliding window is greater than the threshold, this window is used as the starting point of the first health assessment prediction. The formula for the state change rate is s = |y t+a -y t | / a, where y is the feature value, x is the time, and a is the sliding window. The threshold is obtained by averaging the state change rates of the first m windows.

[0031] S23: Feature screening. According to the monotonic change of health change trend and the trend of health degree changing over time, the peak value, variance, root amplitude, root mean square value, kurtosis value and peak factor are selected as the feature input of the health assessment algorithm.

[0032] S30: Figure 2 The figure shows the encoder-decoder structure based on the attention mechanism for health assessment. The encoder and decoder are each an LSTM. The encoder maps the variable-length original sequence to the fixed-length global information c, and the decoder maps the global information c back to the variable-length target sequence. The attention mechanism is added to the network to increase the network's attention to important parts and weaken the attention to irrelevant parts. After the encoder outputs the global information c, it is first processed by the attention gate a, and then decoded after obtaining information of different importance. Given the original sequence, the two networks are jointly trained to make the conditional probability P(y t |y t-1 ,y t-2 ,…,y1,c)=g(ht ,y t-1 ,c)maximize.

[0033] The specific steps of health assessment are:

[0034] S32: Select appropriate high-dimensional features and perform preprocessing. Use the PSLOC algorithm to remove abnormal points in the features, detect the prediction starting point, and construct a comprehensive evaluation standard based on the monotonicity and trend of the features to select the features.

[0035] S33: Further feature extraction through convolutional neural network. The time series during the period of embedded system health degradation is long, with a length of 1370 in the time dimension, which is difficult for recurrent neural network to process. Therefore, further feature extraction is performed through convolutional neural network, and the convolution kernel step distance of the convolutional neural network is selected as 5 and the convolution kernel size is 6. After further feature extraction through the convolution kernel, the data length is roughly shortened to 274.

[0036] S34: Training of encoder-decoder model based on attention mechanism. First, the running time in the collected embedded system running data is normalized. Then, the high-dimensional features extracted by convolutional neural network are input into the encoder-decoder model based on attention mechanism, and the ratio of the current health level to the initial health level HI is output. t =RUL t / RUL0. The number of hidden units of the bidirectional LSTM in the encoder and decoder is 200, the attention calculation adopts the splicing method, the loss function adopts the mean square error, and the Adam optimizer is used in training.

[0037] S35: Fitting regression is performed on the health index of the embedded system. The change of the health of the system is not a simple linear function relationship with time. Therefore, the coordinates of the intersection of the health status index of the embedded system and the coordinate axis and the coordinates of the test point are subtracted to be the current health of the embedded system.

[0038] S36: The average relative error Er and the practicality score Score of the health evaluation are used to quantify the accuracy of the health evaluation. When the estimated health level is worse than the actual health level, the practicality score is higher, otherwise it is lower, because the economic loss caused by statistical maintenance before the embedded system fails is less than replacing it after the failure. Table 1 shows the health evaluation results of the algorithm.

[0039] Table 1 Health evaluation results before optimization

[0040]

[0041] S40: In the present invention, considering the storage resources and computing power of the multi-mode embedded system, the proposed health assessment method is improved to improve the real-time performance of the health assessment.

[0042] S41: Introduce position coding. Introduce position coding to ensure that the temporal position information will not be lost with sampling. When the position t of the feature vector is an even number, the position coding vector is P t,2i = sin(t / 10000 2i / D ), when the position of the feature vector is odd, the position encoding vector is P t,2i+1 =cos(t / 10000 2i / D ), D is the dimension of position encoding.

[0043] S42: During the operation of the system, the health trend is relatively slow, so sparse sampling is performed on the high-dimensional feature vector. The last ten percent of the interval close to the system life limit has a more drastic change, so dense sampling is performed during this period.

[0044] S43: Improvement of encoder. The convolutional neural network feature extractor is used to replace the recurrent neural network LSTM in the encoder. The extracted high-dimensional feature vector should have continuity and predictability. The convolutional neural network feature extractor based on contrastive learning is trained to identify whether a feature sequence is a continuous sequence to ensure continuity. The metric learning loss is used as the training label. The features extracted by the convolutional neural network are used to establish an autoregressive model f k (x t+k ,c t ), extract the global information c, and use the global information to predict the subsequent high-dimensional feature sequence, and the metric learning loss of predictive feature extraction

[0045] S43: Decoder improvement. After the encoder is improved, the amount of decoder input data in the health assessment method is greatly reduced, which will lead to overfitting of the decoder during model training. Therefore, the input temporary regression method and weight temporary regression are used to regularize and restrict the decoder for optimization, and some neurons are randomly deleted during the training process to avoid the risk of overfitting.

[0046] S44: The operation process of the embedded system is divided into 6 stages from the prediction starting point to complete degradation. The average relative error and practicality score of the health assessment results of each stage are calculated. The results are shown in Table 2. The real-time performance of the health assessment algorithm before and after optimization is compared, and its running time is compared. The optimized algorithm includes a parallel predictive feature extractor and a continuous feature extractor, as well as a decoder. The running time is the feature extractor running time plus the decoder running time. The average running time of the health assessment algorithm before and after the real-time improvement is shown in Table 3. It can be seen that the real-time effect after optimization is better. After optimization, the health assessment time is 27ms, and the running speed is increased by 27 times, which is less than 0.1s of the data acquisition time.

[0047] Table 2 Optimized health evaluation results

[0048]

[0049] Table 3 Average running time of health assessment algorithm before and after real-time improvement

[0050]

Claims

1. A health assessment method for a multi-mode autonomous embedded control system, characterized in that it includes the following steps: (1) Obtain a data set composed of physical quantities that can reflect the operating status of the system; (2) Screening and constructing high-dimensional feature vectors; (3) Further extract features through convolutional neural network; (4) Training and constructing the encoder-decoder structure based on the attention mechanism; (5) Perform health assessment based on the trained model; (6) Fitting and regressing the health indicators of the system; (7) To improve real-time performance, position coding is introduced and the codec structure is improved to perform real-time health assessment.

2. The health assessment method of a multi-mode autonomous embedded control system according to claim 1 is characterized in that: The outlier detection method based on local time series prediction is used to eliminate outliers and determine the starting point of the evaluation based on the state change rate.

3. The health assessment method of a multi-mode autonomous embedded control system according to claim 1 is characterized in that: Indicators with good monotonicity and trend are selected to construct the input feature vector, and the features are further compressed through convolutional neural networks.

4. The health assessment method of a multi-mode autonomous embedded control system according to claim 1, characterized in that: Both the encoder and decoder are a long short-term memory network, and an attention mechanism is introduced between the encoder and decoder.

5. The health assessment method of a multi-mode autonomous embedded control system according to claim 1, characterized in that: The health evaluation index output by the codec structure is the ratio of the current health status to the health status when the system starts running.

6. The health assessment method of a multi-mode autonomous embedded control system according to claim 1, characterized in that: Position encoding is used for sparse and dense extraction of data. The encoder is replaced by a convolutional network extractor. The encoder is trained to extract continuous and predictive features based on contrastive learning.

7. The health assessment method of a multi-mode autonomous embedded control system according to claim 1, characterized in that: Regularization restrictions are introduced to the encoder, and some neurons are deleted by input and weight withdrawal to avoid overfitting.

8. The health assessment method of a multi-mode autonomous embedded control system according to claim 2, characterized in that: The least squares support vector machine is used to make local predictions on the signal processed by the sliding window, and the abnormal points are eliminated by comparing with the set threshold to determine the window as the starting evaluation point.

9. The health assessment method of a multi-mode autonomous embedded control system according to claim 5, characterized in that: It performs fitting regression on the health index of the system, and regards the difference between the coordinates of the intersection of the health status index of the embedded system and the coordinate axis and the coordinates of the point to be tested as the current health of the embedded system.

10. The health assessment method of a multi-mode autonomous embedded control system according to claim 6, characterized in that: The continuous feature extractor and the predictive feature extractor are parallel branches, and both are trained simultaneously. The metric learning loss is used as the training label during the continuous feature extraction training. The extracted predictive features are used to extract global information through the established autoregressive model.

Images