CN113536682B - Electric hydraulic steering engine parameter degradation time sequence extrapolation prediction method based on secondary self-coding fusion mechanism - Google Patents

Abstract

The present invention provides a time series extrapolation prediction method for electro-hydraulic steering gear parameter degradation based on a secondary self-encoding fusion mechanism. The method includes: obtaining fault prediction data of the electro-hydraulic steering gear; comprehensively preprocessing the fault data, To obtain the training data set and the test data set; construct a time series extrapolation predictor: the time series extrapolation predictor includes a convolutional neural network autoencoder, an artificial time domain feature extractor based on expert knowledge, and a binary SAE-based Secondary autoencoder; the temporal extrapolation predictor fuses the training data set to obtain fusion features, and the secondary autoencoder performs secondary coding on the fused features, and then combines the secondary coding features with Label data establishes a mapping relationship; comprehensively trains the convolutional neural network primary autoencoder and the temporal extrapolation predictor to obtain a trained temporal extrapolation predictor; and uses the trained temporal extrapolation prediction model to There is data to predict.

Description

A kind of electro-hydraulic steering gear parameter degradation timing sequence based on quadratic self-encoding fusion mechanism extrapolation forecasting method

Technical field

The present invention relates to the prediction of the degradation trend of electro-hydraulic steering gear, and in particular to a time series extrapolation prediction of the degradation of electro-hydraulic steering gear parameters.

Background technique

The electro-hydraulic steering gear system is a complex electromechanical integration system and a high-precision position servo system, which has an important impact on the attitude control of the aircraft. With the continuous development of science and technology, advanced aircraft widely adopt fully digital servo steering systems with fast speed, high precision and large power-to-weight ratio. Contemporary engineering applications place higher requirements on the reliability of steering gears. The prediction of the degradation process of key parameters of the steering gear is an important aspect of the reliability research of the steering gear. Accurately predicting the future time series of key parameters of the steering gear and grasping the trend of parameter changes are of great significance for rationally arranging maintenance plans, improving flight quality, ensuring flight safety, and reducing life cycle costs. Traditional time series extrapolation forecasting methods usually use the time series decomposition strategy to forecast separately by decomposing the time series into trend items, seasonal items, residual items, etc., and finally fuse the forecast results to obtain the time series extrapolation forecast sequence of parameters. However, for complex electromechanical systems such as electro-hydraulic steering gear, the degradation process often exhibits nonlinearity, which makes the time series of its degradation parameters often difficult to effectively decompose according to traditional methods, which brings problems to the future timing prediction of key parameters of the steering gear. Very difficult.

In order to solve this problem, a time series extrapolation prediction method for electro-hydraulic steering gear parameter degradation based on the secondary autoencoding fusion mechanism of artificial features and convolutional features is proposed. This method combines artificial time domain features and convolutional depth features, and achieves feature fusion through a secondary autoencoding mechanism. It can directly map the temporal dependencies and changing trends of the original parameters to the hidden layer depth features, avoiding the sequence decomposition in traditional methods. The problem provides a more practical method for the extrapolation prediction problem of the degradation time series of key parameters of electro-hydraulic steering gear.

Contents of the invention

In order to solve the problems existing in the existing technology, the present invention proposes a time series extrapolation prediction method for electro-hydraulic steering gear parameter degradation based on a secondary autoencoding fusion mechanism.

According to one aspect of the present invention, a time series extrapolation prediction method for electro-hydraulic steering gear parameter degradation based on a quadratic auto-encoding fusion mechanism is provided. The method includes: obtaining fault prediction data of the electro-hydraulic steering gear; Perform comprehensive preprocessing to obtain training data sets and test data sets; construct a time series extrapolation predictor: the time series extrapolation predictor includes a convolutional neural network one-time autoencoder and an artificial time domain feature extractor based on expert knowledge, And a quadratic autoencoder based on SAE; the temporal extrapolation predictor fuses the training data set to obtain fusion features, and the quadratic autoencoder performs secondary encoding on the fusion features, and then Establish a mapping relationship between the secondary encoding features and the label data; comprehensively train the convolutional neural network primary autoencoder and the temporal extrapolation predictor to obtain a trained temporal extrapolation predictor; and utilize the trained temporal extrapolation Predictive models make predictions on existing data.

Preferably, the time series extrapolation prediction model takes the original training data set as input, first performs manual feature extraction based on an artificial time domain feature extractor, and uses a pre-trained convolutional neural network feature extraction model to extract the original training data Convolutional feature extraction is performed on the set, and then the convolutional features and artificial time domain features are feature fused, and the training data label S _trainy is output as the time series extrapolation prediction model to complete the training of the extrapolation predictor model.

Preferably, when predicting the existing data, for the input data with length w, the predicted data length is Ww, and the data segments with length 2w-W in the existing data are intercepted and spliced with the predicted data, so as to As the input of a new round of prediction, the prediction is completed by continuously iterating until the artificially preset prediction length L _p is reached.

Preferably, the verification set data obtained through comprehensive preprocessing is sent to the time series extrapolation prediction model, and combined with the corresponding prediction indicators, the prediction performance evaluation of the model is completed.

Preferably, the comprehensive preprocessing step includes performing sliding window cutting on key parameter time series data. The key parameter time series data is X, X={x ₁ , x ₂ , _... Window cutting generates the corresponding sample data set. When the window width is W and the step size is s, the number of samples generated by cutting is:

Then the corresponding data set is generated as {S ₁ , S ₂ ,...S _sn }, for each sample Si, nor in {S _{1, nor} , S _{2, nor} ,...S _{sn, nor }} Take the data of length w as the training data, and take the data of length Ww as the prediction data corresponding to this training data.

Preferably _{, the} construction of the convolutional neural network primary _autoencoder includes converting its data _format into Convert it into a three-dimensional data format (sn, w, 1), input the constructed three-dimensional training data set into the primary autoencoder and repeatedly perform the forward propagation and back propagation iterative calculation processes to evaluate the constructed primary autoencoder model. The model parameters of the convolution layer, pooling layer, and fully connected layer are continuously adjusted to complete the pre-training of the model, where {S _{1, nor} , S _{2, nor} ,...S _{sn, nor} } are normalized Processed sample data set, sn is the number of samples, w is the data length of each sample, and 1 is the number of channels.

Preferably, the one-shot autoencoding model includes multiple convolutional layers, multiple pooling layers and a Flatten fully connected layer. The fully connected layer uses features extracted from multiple layers of stacked convolutional layers and pooling to perform feature extraction. Recognition, using sofimax regression on the fully connected layer, the output of the sofimax function is

where k represents the number of output layer network nodes.

Preferably, in the SAE-based secondary autoencoder, the secondary encoder and decoder are further pre-trained with a two-dimensional fusion feature matrix; a two-dimensional fusion feature matrix is used as the stacked secondary autoencoder. input and output, select the appropriate loss function and number of iterations, complete the iterative calculation process of forward propagation and back propagation, so that the model continuously reconstructs its own input, and finally extracts it from the stacked quadratic autoencoder model that has completed pre-training. The encoding layer serves as an available quadratic autoencoding model.

Preferably, the deep fusion features are subjected to secondary auto-encoding based on the pre-trained secondary autoencoder model, thereby obtaining the secondary encoding feature set {F′ ₁ , F′ ₂ , .., F′ _sn }.

This summary merely serves as an introduction to the subject matter fully described in the detailed description and drawings. The Summary of the Invention should not be deemed to describe necessary technical features, nor should it be used to determine the scope of the claims. Furthermore, it is to be understood that the foregoing summary and the following detailed description are exemplary and explanatory only and are not necessarily limiting of claimed subject matter.

Description of the drawings

Various embodiments or examples ("examples") of the present disclosure are disclosed in the following detailed description and accompanying drawings. It is not necessary that the drawings be drawn to scale. In general, the operations of the disclosed methods may be performed in any order, unless otherwise specified in the claims. In the attached picture:

Figure 1 shows a flow chart of a feature extraction method based on the fusion of artificial features and convolutional features based on deep neural networks according to the present invention;

Figure 1A shows a time series extrapolation prediction method based on the fusion features output by the method shown in Figure 1;

Figure 2 shows a schematic diagram of a method for obtaining electro-hydraulic steering gear failure prediction data according to the present invention;

Figure 3 shows a structural diagram of a one-time autoencoding model based on a convolutional neural network according to the present invention;

Figure 4 shows the operation flow chart of the one-time autoencoding and artificial time domain feature extraction shown in Figure 1;

Figure 5 shows a schematic structural diagram of the SAE-based quadratic autoencoder shown in Figure 1;

Figure 6 shows a schematic diagram of the raw data of the feedback angle;

Figure 7A shows the maximum value of the artificial time domain feature obtained based on the flowchart shown in Figure 4;

Figure 7B shows the standard deviation of the artificial time domain features obtained based on the flowchart shown in Figure 4;

Figure 8 shows the prediction results obtained according to the time series extrapolation prediction method of the present invention;

Figure 9A shows all prediction data;

Figure 9B is a partially enlarged view of part of the prediction data.

Detailed ways

Before one or more embodiments of the present disclosure are explained in detail, it is to be understood that the embodiments are not limited to the construction details of their specific applications, steps or methods set forth in the following embodiments or drawings.

The time series extrapolation prediction method disclosed by the present invention is illustrated in the method flow chart shown in Figure 1 and Figure 1A. Figure 1 shows feature extraction based on the fusion of artificial features and convolutional features based on deep neural networks according to the present invention. Method flow chart, Figure 1A shows the temporal extrapolation prediction method based on the fused features output by the method shown in Figure 1 . The feature extraction method shown in Figure 1 includes multiple steps: Step 1: Obtain the fault prediction data of the electro-hydraulic steering gear; Step 2: Comprehensive preprocessing of the fault data; Step 3: Perform convolutional neural network-based Feature extraction; Step 4: Perform artificial time domain feature extraction based on expert knowledge from the training data set; Step 5: Feature splicing of artificial features extracted based on empirical knowledge and high-dimensional hidden layer features extracted based on the CNN feature extraction model; Step 6 : Perform deep feature fusion based on stacked autoencoders to obtain secondary encoding feature values. The secondary encoding feature values obtained in step 6 are sent to the time series extrapolation prediction model training module shown in Figure 1A, and step 7 is performed: time series extrapolation prediction model training. The step 7 specifically includes: constructing a time series extrapolation prediction model ;train a time series extrapolation predictor model; and make forecasts using the extrapolation forecast period. The electro-hydraulic steering gear parameter degradation time series extrapolation prediction method based on the quadratic autoencoding fusion mechanism will be described in detail below with reference to Figure 1 and Figure 1A.

1. Obtain fault prediction data of electro-hydraulic steering gear

Limited by the actual test conditions and actual use environment, it is widely difficult to obtain real product failure data. Simulation analysis is one of the main means to solve the problem of data shortage at home and abroad. Extensive research results are based on simulation models to inject faults and obtain corresponding fault data. Therefore, in order to obtain the fault prediction data of the electro-hydraulic steering gear, the present invention needs to use Simulink software to conduct structured modeling of the electro-hydraulic steering gear and perform fault simulation. In operation, using this Simulink model for fault injection can maximize the acquisition of data close to the real fault situation and achieve fault prediction model verification.

In response to the steering gear system fault prediction requirements, the steering gear system structure was first modeled, and then Simulink simulation software was used to build a steering gear control system simulation model for generating simulation data. Based on the steering gear control system simulation model, appropriate fault injection points of the steering gear system simulation model are selected for fault injection and simulation signals are collected for the development and verification of the steering gear control system fault prediction model. The method of obtaining electro-hydraulic steering gear failure prediction data is shown in Figure 2.

First, according to the steering gear system fault prediction requirements, the steering gear control system model is structured. The structured steering gear system mainly includes the energy system and the position servo system, as shown in the steering gear system structure analysis module in Figure 2. Its key components mainly include: power amplifier combination, DC motor, electro-hydraulic servo valve, hydraulic variable Pump, actuator, control mechanism, feedback potentiometer, high-pressure safety valve, low-pressure safety valve, oil filter, mailbox and other components.

Based on the structured processing of the steering gear control system model, Simulink simulation software is used to build a steering gear control system simulation model, and the fault injection point is determined based on this simulation model. The fault injection point can be selected based on fault prediction needs and historical fault data. It is usually injected on each component of the steering gear. In this application, for example, the feedback amplification coefficient can be selected as the fault injection point, and fault injection can be performed on the feedback potentiometer. Finally, fault simulation and signal collection are performed. The data collected by the signal can be various control instructions and status signals of the steering gear control system, which are the key parameter timing data of the electric steering gear in this application, including, for example: control instructions, unified clock, displacement signal, and feedback angle. These selected key parameter time series data constitute the historical time series data of the parameters to be predicted. The feedback angle can effectively characterize the health status of the steering gear system. Therefore, the feedback angle can be selected as the subsequent predicted parameter.

2. Comprehensive preprocessing of fault data

The data obtained by the steering gear fault prediction data acquisition unit, such as the feedback angle signal, is sent to the fault data preprocessing unit for comprehensive processing to obtain a training data set and a test data set. For details, refer to the comprehensive data preprocessing module shown in Figure 1 , which includes:

Step 1. Perform sliding window cutting on the key parameter time series data to construct a sample data set;

The time series data of any key parameter of the electric steering gear collected by the sensor is X, X = {x ₁ , x ₂ , _... When the window width is W and the step size is s, the number of samples generated by cutting is:

Then the corresponding data set is generated as {S ₁ , S ₂ ,...S _sn }, for each sample Si, nor in {S _{1, nor} , S _{2, nor} ,...S _{sn, nor }} Take the data of length w as the training data, and take the data of length Ww as the prediction data corresponding to this training data.

Step 2. Perform maximum and minimum normalization processing on the training data set;

In order to improve the data expression ability and speed up the convergence speed of subsequent model training, the training data set needs to be normalized, mainly by scaling the amplitude of the original parameters through the maximum and minimum value normalization method to complete the linearization of the data. Transform. For a single sample data S _i ={x ₁ , x ₂ ,...x _w }, through the formula:

Implement normalization processing to obtain a normalized sample data set {S _{1, nor} , S _{2, nor} ,...S _{sn, nor} }.

Step 3. Construct training data set and test data set;

Select the first r% of all data as the training data set, and the remaining data as the test data set to verify the model prediction performance. Generally speaking, r generally takes 60-80, preferably 70.

3. Feature extraction based on convolutional neural network

The training data set obtained after processing by the comprehensive data preprocessing module is sent to the convolutional neural network primary autoencoder and the artificial time domain feature extraction module based on expert knowledge to obtain convolutional features and artificial special time domain features. The feature extraction step specifically includes: the construction of a one-shot autoencoder model based on a convolutional neural network, as shown in Figure 3; and the pre-training of the convolutional one-shot autoencoder model as shown in Figure 4 and the use of a convolutional encoder for convolution. Product feature extraction.

First, a training data set is used to construct a one-pass autoencoding model based on a convolutional neural network (CNN), and the training data set is used to pre-train the model. Since the two-dimensional convolutional neural network requires the format of the input data to be three-dimensional data, a training data set needs to be constructed. The training set is constructed to meet the input requirements of the two-dimensional one-shot autoencoding model. The method is to convert the data format of the training data set S _train = {S _{1, nor} , S _{2, nor} ,...S _{n, nor} }. is (sn, w, 1), where sn is the number of samples, w is the data length of each sample, and 1 is the number of channels. Input the constructed training sample data set into the one-pass autoencoding model based on the convolutional neural network (CNN) shown in Figure 3.

Convolutional Neural Network is a multi-layer supervised learning neural network. The convolution layer and pool sampling layer of the hidden layer are the core parts of realizing the feature extraction function of the convolutional neural network. CNN is a neural network specially designed to process data with similar network structure. It extracts features from original data by imitating the operating mechanism of biological vision. Different CNN layers have the characteristics of weight sharing, which effectively reduces the complexity of the network. Avoid over-fitting problems caused by too little data and avoid the complexity of data reconstruction when extracting multi-dimensional data features. As shown in Figure 3, the deep convolutional neural network of the present invention includes multiple convolutional layers, multiple pooling layers and a Flatten fully connected layer.

Convolution layer: The convolution process with nonlinear activation can be described as:

in, is the output of the nth convolution kernel in the rth convolutional layer,/> is the m-th output feature vector in the r-1th convolution layer, * represents the convolution operation,/> Represents the weight and bias of the n-th convolution kernel in the r-th convolution layer respectively, and ReLU represents the nonlinear activation function.

Pooling layer: By adding a pooling layer, the spatial dimension of the convolutional features can be reduced and overfitting can be avoided. The max pooling layer is the most commonly used pooling layer. It only takes the most important part of the input (the highest value), which can be expressed as

in is the feature obtained by the convolutional layer,/> is the output of the pooling layer, and l represents the length of the pooling operation area.

Fully connected layer: Features extracted using multi-layer stacked convolutional layers and pooling are finally input into the fully connected layer for feature recognition. Softmax regression is usually used on the top fully connected layer. Define the output of the softmax function as

where k represents the number of output layer network nodes.

The convolution layer uses a certain number of convolution kernels to extract different features of the input data in the time domain. The pooling layer can effectively reduce the size of the parameter matrix, thereby reducing the number of parameters in the final connection layer and adding the pool The fully connected layer can speed up the calculation and prevent the model from over-fitting. Finally, the fully connected layer is used to map the feature parameters in the high-dimensional hidden layer to the original input data, thereby training the feature extraction capability of the model.

Secondly, select the appropriate number of iterations and loss function, and input the constructed three-dimensional training data set into the feature extraction model to repeatedly perform the forward propagation and back propagation iterative calculation processes; during this process, the convolution layer and the pooling layer are , The model parameters of the fully connected layer are continuously adjusted to complete the pre-training of the model.

Third, take out the two convolutional layers, two pooling layers and one fully connected layer of the pre-trained model, retain their weight parameters, and build them into a trained deep convolutional neural network one-pass autoencoding model.

Finally, based on the pre-trained convolutional neural network one-time autoencoding model, the convolution feature extraction is performed on the training data set {S _{1, nor} , S _{2, nor} ,...S _{sn, nor} } to obtain the convolution feature set. {F _1,CNN ,F _2,CNN ,...,F _sn,CNN }.

4. Extract artificial time domain features based on expert knowledge from the training data set

As shown in the right diagram of Figure 4, time domain feature extraction based on expert knowledge is performed on the cut training data set S _train ={S _{1, nor} , S _{2, nor} ,...S _{n, nor} }. Specifically, it includes sliding window cutting of the normalized training data, extracting different time domain data for each cut sample, and normalizing the time domain features.

Perform sliding window cutting on the normalized sample data: the window length is w′, the step size is 1, and for the sample _Si = {x ₁ , x ₂ ,...x _w }, ww′+1 samples can be cut , the length of each sample is w′, that is, {S′ ₁ , S′ ₂ ,...S′ _ww′+1 }.

For each sample S′ _i , eight time domain features are extracted: maximum value, standard deviation, variance, waveform factor, root mean square, impulse index, margin factor, and peak factor: for the time domain features extracted from window data S′ _i is F _i ={f ₁ , f ₂ ,..., f ₈ }, so the artificial features extracted for sample S _i are {F ₁ , F ₂ ,...F _ww′+1 }, using the maximum polar The small value normalization method normalizes artificial features. For details, please refer to step 2 in comprehensive data preprocessing.

5. Feature splicing of artificial features extracted based on expert knowledge and high-dimensional hidden layer features extracted based on CNN feature extraction model

Continuing to refer to Figures 1 and 4, after the CNN feature extraction model extracts high-dimensional hidden layer features and the artificial feature extraction module extracts artificial time-domain features, the high-dimensional hidden layer features and artificial time-domain features are feature spliced. The feature matrix extracted by the CNN feature extraction model is M _CNN , and its shape and size are Where n _f is the number of convolution kernels, s _f is the convolution kernel step size, and f is the convolution kernel size. The artificial time domain feature matrix extracted by the artificial feature extraction module is M _manual , and its shape and size are (ww′+1, 8). Flatten the two feature sizes M _CNN and M _manual respectively, and splice them along the column direction. The size of the resulting fused feature is:

Perform CNN feature extraction and artificial feature extraction and perform feature fusion for each sample S _{i, nor} in the training data set S _train = {S _{1, nor} , S _{2, nor} ,...S _{n, nor} }. Let the dimension of the fusion feature be n _merge , then the training data set can be reorganized into a two-dimensional fusion feature matrix of (n, n _merge ), and this fusion feature matrix can be used as the input of the subsequent SAE coding model.

6. Perform deep feature fusion based on stacked autoencoders (SAE)

Refer to Figure 5, which is a schematic structural diagram of the SAE-based quadratic autoencoder shown in Figure 1. In this quadratic autoencoder, deep feature fusion based on stacked autoencoders is performed, which specifically includes building quadratic autoencoders and decoders, training quadratic autoencoders and decoders, and using stacked quadratic autoencoders. Deep feature fusion.

First, a stacked quadratic autoencoder and decoder model is constructed. The model structure is shown in Figure 5. The number of encoding layers is the same as the number of decoding layers, which enables the model to have better secondary encoding capabilities for deep features.

Secondly, use the two-dimensional fusion feature matrix obtained in step (5) to pre-train the quadratic autoencoder model, use the two-dimensional fusion feature matrix as the input and output of the stacked quadratic autoencoder model, and select the appropriate loss function and The number of iterations, completes the iterative calculation process of forward propagation and back propagation, so that the model continuously reconstructs its own input, and finally extracts the coding layer from the stacked quadratic autoencoder model that has completed pre-training as a usable quadratic autoencoder model. .

Finally, secondary autoencoding is performed on the deep fusion features based on the pre-trained secondary autoencoder model, thereby obtaining the secondary encoding feature set {F′ ₁ , F′ ₂ , .., F′ _sn }.

7. Carry out time series extrapolation prediction model training

The block diagram shown in Figure 1A shows the execution steps and methods of time series extrapolation prediction model training, which specifically include building a time series extrapolation prediction model, training the extrapolation predictor model and using the extrapolation prediction period for prediction.

Step 7.1: Use the convolutional neural network primary autoencoder obtained during the feature extraction process based on the convolutional neural network and the stacked quadratic autoencoder model obtained during the SAE-based secondary autoencoder process to construct the time series extrapolation Predictor, this extrapolation predictor fuses CNN features and artificial features and performs secondary coding on the depth features, and then establishes a mapping relationship between the secondary coding features and the label data to complete the extrapolation prediction.

Step 7.2: Comprehensive training of CNN convolutional feature extractor and temporal extrapolation predictor. The time series extrapolation prediction model takes the original input data as input, first performs manual feature extraction, uses the pre-trained CNN feature extraction model to perform CNN convolution feature extraction on the original data, and then characterizes the CNN convolution features and artificial time domain features. Fusion, and the training data label S _trainy is used as the output of the time series extrapolation prediction model to complete the training of the extrapolation predictor model.

Step 7.3. Use the trained time series extrapolation prediction model to predict the existing data. For the input data with length w, the predicted data length is Ww. Intercept the data segment with length 2w-W in the existing data and the predicted data. Carry out splicing and use this as the input for a new round of prediction, and continue to iterate until the artificially preset prediction length L _p is reached, then the prediction ends.

Step 7.4: Send the verification set data obtained by the comprehensive data preprocessing unit into the time series extrapolation prediction model, and combine it with the corresponding prediction indicators to complete the prediction performance evaluation of the model.

[Example of feature extraction based on feature fusion]

An important work of the present invention is to innovatively design a feature extraction method based on the fusion of artificial features and convolutional features based on deep neural networks. This method directly affects the degradation trend prediction and health assessment of the hydraulic actuator system of the extrapolation prediction model. Based on this, we use the "actuator internal leakage" fault of the steering gear system and select the feedback angle data collected at the "flow injection point" measuring point as an example.

The structured model of the electro-hydraulic steering gear is shown in Figure 2. Its fault prediction is set to the "actuator internal leakage" fault, and its data is feedback angle time domain data. After obtaining the feedback angle data, preprocess the data. In this case, the window length is selected as 9000 and the step size is 1. For the data of each window, the first 6000 length of data is used as the input of the sequence extrapolation prediction model based on the convolutional neural network, and the last 3000 length of data is used as the window. The label data is the prediction data. For all normalized feedback angle data, the first 70% of the data is selected as the training data set, and the remaining 30% of the data is used as the verification data set to verify the model prediction performance.

1. After obtaining the training data, perform feature extraction based on convolutional neural network

Considering the parameter characteristics of the steering gear feedback angle data, the convolutional neural network is used to extract sample features from the normalized sample data. Continuing to refer to Figure 1, Figure 3 and Figure 4, the convolution layer uses a certain number of convolution kernels to extract different features of the input data in the time domain. The pooling layer can effectively reduce the size of the parameter matrix, thereby reducing the final The number of parameters in the connection layer. Adding a pooling layer can speed up calculations and prevent model overfitting. Use two layers of convolutional layers to map the original data to a high-dimensional implicit space to learn the nonlinear characteristics of the data, and then combine The flattening layer and the fully connected layer remap the high-dimensional sample features to the original input data to learn the key features of the original samples. The module that maps the original data samples to the low-dimensional feature space is selected as the encoder of the model. After extraction and screening, The module of the feature reconstruction sample serves as the decoder of the model. The model structure parameters selected by this invention are shown in Table 1

Table 1 Parameters of linear autoencoder model based on convolutional neural network

Select the appropriate number of iterations and loss function, input the constructed three-dimensional training data set into the feature extraction model, and repeatedly perform the forward propagation and back propagation iterative calculation processes to calculate the model parameters of the convolution layer, pooling layer, and fully connected layer. Continuously adjust and complete the pre-training of the model. Take out the two convolutional layers, two pooling layers and one fully connected layer of the pre-trained model, retain their weight parameters, and construct it as a CNN feature extraction model.

2. Extract time domain features based on expert knowledge from the cut training data set.

Specifically, the window length is 3000 and the step size is 3000, then the training data of each window is cut and manual feature extraction is performed on the data of each sub-window. The feature extraction results are shown in Figure 7A and Figure 7B.

3. Feature splicing of artificial features extracted based on expert knowledge and high-dimensional hidden layer features extracted based on CNN feature extraction model

4. Perform deep feature fusion based on stacked autoencoders

Use the two-dimensional fusion feature matrix to pre-train the quadratic autoencoder model, use the two-dimensional fusion feature matrix as the input and output of the stacked quadratic autoencoder model, select the appropriate loss function and number of iterations, and complete the forward propagation and inverse The iterative calculation process of forward propagation enables the model to continuously reconstruct its own input, and finally extract the coding layer from the stacked quadratic autoencoder model that has completed pre-training as a usable quadratic autoencoder model.

5. Time series extrapolation prediction model training

The pre-trained convolutional neural network first-order autoencoder and stacked quadratic autoencoder models are used to construct a temporal extrapolation predictor. The extrapolation predictor fuses CNN features with artificial features and performs a second step on deep features. coding, and then establish a mapping relationship between the secondary coding features and the label data to complete the extrapolation prediction. The time series extrapolation prediction model takes the original input data as input, first performs manual feature extraction, and uses the pre-trained CNN Tit model to predict the original The data is extracted by CNN, and then the CNN features and artificial features are feature fused, and the training data labels are used as model output to complete the model training. Use the trained prediction model to predict the existing data, splice the extrapolated data with the original data as the input for a new round of prediction, and continue iterating until the preset prediction length is reached, then the prediction ends and the prediction result As shown in Figure 8, the comparison results between the prediction results and the real labels are shown in Figure 9A and Figure 9B.

Although the invention has been described with reference to the embodiments shown in the drawings, equivalent or alternative means may be used without departing from the scope of the claims. The components described and illustrated herein are merely examples of systems/devices and methods that may be used to implement embodiments of the present disclosure, and may be substituted with other devices and components without departing from the scope of the claims.

Claims (9)

1. A time series extrapolation prediction method for electro-hydraulic steering gear parameter degradation based on the quadratic auto-encoding fusion mechanism, including: Obtain failure prediction data of electro-hydraulic steering gear; Perform comprehensive preprocessing on the fault data to obtain a training data set and a test data set; Construct a time series extrapolation predictor characterized by: The time series extrapolation predictor includes a convolutional neural network primary autoencoder, an artificial time domain feature extractor based on expert knowledge, and a secondary autoencoder based on SAE; The time series extrapolation predictor fuses the training data set to obtain fusion features, and the secondary autoencoder performs secondary coding on the fused features, and then establishes a mapping relationship between the secondary coding features and the label data. ; Comprehensive training of the convolutional neural network primary autoencoder and the temporal extrapolation predictor to obtain a trained temporal extrapolation predictor; and Use the trained time series extrapolation prediction model to predict existing data. 2. The electro-hydraulic steering gear parameter degradation time series extrapolation prediction method according to claim 1, characterized in that the time series extrapolation prediction model takes the original training data set as input and is first based on an artificial time domain feature extractor. Manual feature extraction, using the pre-trained convolutional neural network feature extraction model to extract convolutional features from the original training data set, and then feature fusion of the convolutional features and artificial time domain features, and use the training data label S _trainy as The time series extrapolation prediction model output is used to complete the training of the extrapolation predictor model. 3. The electro-hydraulic steering gear parameter degradation time series extrapolation prediction method according to claim 1, characterized in that when predicting the existing data, for the input data with a length of w, the predicted data length is Ww, Intercept the data segment with a length of 2w-W in the existing data and splice it with the prediction data, and use this as the input of a new round of prediction. Iterate continuously until the artificially preset prediction length L _p is reached, then the prediction ends. 4. The electro-hydraulic steering gear parameter degradation time series extrapolation prediction method according to claim 1, characterized in that the verification set data obtained through comprehensive preprocessing is sent to the time series extrapolation prediction model, combined with the corresponding prediction indicators. , complete the prediction performance evaluation of the model. 5. The electro-hydraulic steering gear parameter degradation time series extrapolation prediction method according to claim 1, characterized in that the comprehensive preprocessing step includes sliding window cutting of key parameter time series data, and the key parameter time series data is X , X={x ₁ , x ₂ ,...x _N }, perform sliding window cutting on Then the corresponding data set is generated as {S ₁ , S ₂ ,...S _sn }, for each sample Si, nor in {S _{1, nor} , S _{2, nor} ,...S _{sn, nor }} Take the data of length w as the training data, and take the data of length Ww as the prediction data corresponding to this training data. 6. The electro-hydraulic steering gear parameter degradation time series extrapolation prediction method according to claim 1, characterized in that the construction of the convolutional neural network one-time autoencoder includes based on the training data set S _train = {S _{1 ,nor} ,S _2,nor ,...S _n,nor }, convert its data format into a three-dimensional data format (sn, w, 1), input the constructed three-dimensional training data set once and execute the autoencoder repeatedly. The iterative calculation process of forward propagation and back propagation is used to continuously adjust the model parameters of the convolution layer, pooling layer, and fully connected layer of the constructed primary autoencoding model to complete the pre-training of the model, where {S _{1 , nor} , S _{2, or} ,...S _{sn, nor} } is the normalized sample data set, sn is the number of samples, w is the data length of each sample, and 1 is the number of channels. 7. The electro-hydraulic steering gear parameter degradation timing extrapolation prediction method according to claim 6, characterized in that the one-time autoencoding model includes multiple convolutional layers, multiple pooling layers and a Flatten fully connected layer, The fully connected layer uses features extracted by multi-layer stacked convolutional layers and pooling for feature recognition. Softmax regression is used on the fully connected layer. The output of the softmax function is where k represents the number of output layer network nodes. 8. The electro-hydraulic steering gear parameter degradation timing extrapolation prediction method according to claim 1, characterized in that in the SAE-based secondary autoencoder, the secondary encoding is further performed with a two-dimensional fusion feature matrix. The decoder and decoder are pre-trained; the two-dimensional fusion feature matrix is used as the input and output of the secondary autoencoder, the appropriate loss function and the number of iterations are selected, and the forward propagation and back propagation iterative calculation processes are completed to make the model continuously Reconstruct its own input, and finally extract the coding layer from the pre-trained quadratic autoencoder model as a usable quadratic autoencoder model. 9. The electro-hydraulic steering gear parameter degradation timing extrapolation prediction method according to claim 8, characterized in that, based on the quadratic autoencoder model obtained by pre-training, the deep fusion features are quadratic auto-encoded to obtain the quadratic autoencoder. Encoding feature set {F′ ₁ , F′ ₂ , ..., F′ _sn }.