Sensor electromagnetic pulse coupling signal separation and reconstruction method

Abstract

The application discloses a sensor electromagnetic pulse coupling signal separation and reconstruction method and belongs to the field of engine electromagnetic research. The method is characterized in that: the electromagnetic pulse is identified through continuous wavelet transform and a convolutional neural network, a deep clustering-based electromagnetic pulse coupling signal separation method is designed based on deep learning, feature representation and clustering allocation are jointly optimized, the electromagnetic pulse coupling signal under a multi-source signal is subjected to signal separation, then a basis pursuit algorithm is used to reconstruct the signal, the Griffin-Lim algorithm and the Wiener filter are combined to iteratively optimize the reconstructed signal, the electromagnetic pulse coupling signal is effectively separated, the electromagnetic pulse interference is eliminated, and the fault tolerance is achieved.

Description

Technical Field

[0001] This invention relates to the field of engine electromagnetic research technology, and in particular to a method for separating and reconstructing electromagnetic pulse coupling signals from sensors. Background Technology

[0002] In various modern mechanical equipment and intelligent vehicles, the engine electronic control system acts as the "brain" of the equipment, connecting to various sensors via a cable network. When the equipment is in a complex electromagnetic environment, the engine sensor signals are easily affected by electromagnetic pulse interference, causing the engine to malfunction. Existing electromagnetic pulse protection is divided into hardware protection and software protection. Hardware protection targets the coupling path, taking both active and passive protection approaches. Passive protection measures include physical shielding, grounding, and redundancy design, while active protection measures include designing protective circuits and redundant system designs to improve the safety and reliability of the electronic control system.

[0003] Despite hardware protection, electromagnetic pulses can still enter sensors at sensitive frequencies via wire harness coupling, causing abnormal sensor signals and affecting normal engine operation. Therefore, in addition to hardware protection, it is necessary to study active fault-tolerant methods for engine sensor signals. Since software protection methods mainly design anti-interference algorithms based on the characteristics of electromagnetic interference sources, their protection is highly targeted and cannot be applied to complex environments. Therefore, it is of great significance to study universal fault-tolerant methods for engine electromagnetic shocks. Summary of the Invention

[0004] This invention provides a method for separating and reconstructing electromagnetic pulse coupling signals from sensors, which achieves effective separation of electromagnetic pulse coupling signals, eliminates electromagnetic pulse interference, and achieves fault tolerance.

[0005] This invention provides a method for separating and reconstructing electromagnetic pulse coupling signals from a sensor, comprising the following steps:

[0006] Collect engine sensing signals;

[0007] When the engine sensing signal is subject to electromagnetic pulse interference, continuous wavelet transform and convolutional neural network are used to identify the electromagnetic pulse interference signal, and the feature representation and clustering assignment are jointly optimized. The electromagnetic pulse coupling signal under multi-source signal is separated by deep clustering.

[0008] The engine sensing signal is reconstructed using the basis pursuit algorithm. The reconstructed engine sensing signal is then optimized by combining the Griffin-Lim algorithm and Wiener filtering. Electromagnetic pulse interference signals in the engine sensing signal are then separated and eliminated.

[0009] Optionally, in one embodiment of the present invention, the identification of electromagnetic pulse interference signals is performed using continuous wavelet transform and convolutional neural networks, including:

[0010] The one-dimensional engine sensing signal timing signal has a unified signal length and a fixed number of sampling points;

[0011] The one-dimensional engine sensing signal time series signal is converted into a two-dimensional time-frequency image, and bilinear interpolation is used to unify the size of the time-frequency image;

[0012] Image enhancement is performed on the two-dimensional time-frequency images, and the images are divided into training and testing sets.

[0013] Construct a network model, train the network model in stages using the training set, select the cross-entropy loss function in multi-classification tasks, and select network model parameters based on the convergence trend of the cross-entropy loss.

[0014] The test set is input into the network model. By identifying the characteristics of sensor malfunction signals, it is determined whether the engine sensing signal is affected by electromagnetic interference. The network model parameters are updated and optimized by combining the classification results and judgment accuracy of sensor malfunctions.

[0015] Optionally, in one embodiment of the present invention, the network model is a convolutional neural network.

[0016] Optionally, in one embodiment of the present invention, signal separation of electromagnetic pulse coupling signals under multi-source signals is performed by deep clustering, specifically including:

[0017] Convert the electromagnetic pulse coupling signal from the time domain signal x(t) to its time-frequency domain representation x. CWT (f,t);

[0018] Based on the trained network model, signal features are extracted, and the center points of each single source are selected. According to the formula Calculate the Euclidean distance between a single source point and its cluster center. By setting a distance threshold U, datasets whose distance to the cluster center is less than the threshold U are classified as having the same label. Where S... i (t) Let C be the Euclidean distance. p Let k be the feature vector and k be the number of feature points.

[0019] Optionally, in one embodiment of the present invention, the engine sensing signal is reconstructed using a basis pursuit algorithm, specifically including:

[0020] Establish a sparse representation model, define the observation matrix, and initialize the basis matrix Λ, the perception matrix A, and the sparse coefficient vector α;

[0021] According to the formula Minimize the L1 norm of the sparse coefficient vector to calculate the sparsity of the model under constraints, where Φ is the dictionary matrix, ε is the tolerance threshold of the reconstructed signal, y is the observed signal, |||1 is the L1 norm, used to represent the measure of sparsity, and |||2 is the constraint condition for the reconstruction error.

[0022] We use linear programming to optimize the sparse coefficient vector and obtain the sparse coefficient vector α after signal reconstruction. * Then, according to the formula x=Φα * The engine sensing signal was reconstructed.

[0023] The sensor electromagnetic pulse coupling signal separation and reconstruction method of this invention, based on hardware protection and combined with electromagnetic pulse software protection method, can effectively identify whether the sensor signal is affected by electromagnetic pulse interference, effectively separate the electromagnetic pulse coupling signal according to the separation and reconstruction method, and reconstruct the original sensor signal, thereby realizing active electromagnetic pulse fault tolerance of the sensor signal.

[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0026] Figure 1 This is a flowchart of a sensor electromagnetic pulse coupling signal separation and reconstruction method according to an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the sensor electromagnetic pulse coupling signal separation and reconstruction method according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the neural network model used in an embodiment of the present invention;

[0029] Figure 4 This is a flowchart illustrating the training process of the neural network model used in an embodiment of the present invention. Detailed Implementation

[0030] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0031] Figure 1This is a flowchart of a sensor electromagnetic pulse coupling signal separation and reconstruction method provided according to an embodiment of the present invention.

[0032] like Figure 1 As shown, the method for separating and reconstructing the electromagnetic pulse coupling signal of this sensor includes the following steps:

[0033] In step S101, engine sensing signals are acquired.

[0034] The engine sensing signal is collected by sensors. It is then determined whether the engine sensing signal is affected by electromagnetic pulse interference. If there is no interference, the engine sensing signal is directly input into the engine digital controller. Otherwise, if electromagnetic interference exists, the engine sensing signal needs to be processed by signal separation to remove electromagnetic pulse interference signals before inputting it into the digital controller.

[0035] In step S102, when electromagnetic pulse interference exists in the engine sensing signal, the electromagnetic pulse interference signal is identified by continuous wavelet transform (CWT) and convolutional neural network, and the feature representation and clustering assignment are jointly optimized. The electromagnetic pulse coupling signal under multi-source signal is separated by deep clustering.

[0036] In an embodiment of the present invention, the electromagnetic pulse (EMP) identification method identifies whether engine sensor signals are affected by electromagnetic shock. Since the original signal cannot reflect the complete EMP coupling waveform under the influence of an EMP, the CWT method is used to convert the EMP coupling signal into a time-frequency image. This obtains the time-frequency domain image features of the EMP signal, and a convolutional neural network model is constructed. The network model is trained using the time-frequency domain image features as input. Through a multi-layer network, the video image features of the EMP coupling signal at different levels of refinement are extracted, achieving effective identification of EMP interference.

[0037] The CWT method is a powerful time-frequency analysis method used to analyze non-stationary signals, i.e., signals whose frequency components change over time. Compared with the traditional Fourier transform, the wavelet transform can provide localized information in both time and frequency, better capturing the transient characteristics of the signal.

[0038] Optionally, in one embodiment of the present invention, the identification of electromagnetic pulse interference signals is performed using continuous wavelet transform and convolutional neural networks, including:

[0039] Acquire one-dimensional time-series signals, with uniform signal length and a fixed number of sampling points;

[0040] One-dimensional sensor signals are converted into two-dimensional time-frequency images, and bilinear interpolation is used to unify the size of the time-frequency image.

[0041] Image enhancement is performed on the two-dimensional time-frequency images, and the images are divided into training and testing sets.

[0042] Construct a network model, train the network model in stages using the training set, select the cross-entropy loss function in multi-classification tasks, and select network model parameters based on the convergence trend of the cross-entropy loss.

[0043] The test set is input into the network model. By identifying the characteristics of sensor malfunction signals, it is determined whether the sensor is subject to electromagnetic interference. The network model parameters are updated and optimized by combining the classification results and judgment accuracy of sensor malfunctions.

[0044] Preferably, the network model is a convolutional neural network (CNN).

[0045] In embodiments of the present invention, the signal is transformed from the time domain to the time-frequency domain by using a time-frequency analysis method, which visualizes the frequency components of the signal at different time periods and reduces the difficulty of data partitioning.

[0046] Optionally, in one embodiment of the present invention, signal separation of electromagnetic pulse coupling signals under multi-source signals is performed by deep clustering, specifically including:

[0047] Convert the electromagnetic pulse coupling signal from the time domain signal x(t) to its time-frequency domain representation x. CWT (f,t);

[0048] Based on the trained network model, signal features are extracted, and the center points of each single source are selected. According to the formula Calculate the Euclidean distance between a single source point and its cluster center. By setting a distance threshold U, datasets whose distance to the cluster center is less than the threshold U are classified as having the same label. Where S... i (t) Let C be the Euclidean distance. p Let k be the feature vector and k be the number of feature points.

[0049] In embodiments of this invention, clustering assignment is commonly used in data analysis to group a set of data points so that each cluster contains similar objects, while the data points in different groups have significant differences. Sparse representation refers to representing a signal or image as a sparse linear combination, that is, using as few basis functions as possible to represent the signal or image. Sparse representation can reconstruct signals or images with fewer non-zero coefficients, reducing the amount of data required for storage and transmission. Deep clustering combines deep learning techniques with traditional clustering methods. It uses deep neural networks to learn high-level representations of data, clusters image features, and uses traditional clustering methods to divide feature regions, achieving the separation of electromagnetic pulse signals.

[0050] Based on the principles of clustering assignment and sparse representation, this paper utilizes deep clustering and basis pursuit algorithms to separate electromagnetic pulse coupling signals and reconstruct source signals, achieving effective separation of electromagnetic pulse signals from sensor input signals. Deep clustering is employed to classify electromagnetic pulse time-frequency images and separate electromagnetic pulse signals. K-means clustering minimizes the distance from each feature to its corresponding cluster center. The clustering effect is evaluated based on the silhouette coefficient, and the cluster centers are updated according to the evaluation results, thereby minimizing the loss function.

[0051] In step S103, the engine sensing signal is reconstructed using the basis pursuit algorithm. The reconstructed engine sensing signal is then optimized by combining the Griffin-Lim algorithm and Wiener filtering. Electromagnetic pulse interference signals in the engine sensing signal are separated and eliminated.

[0052] Optionally, in one embodiment of the present invention, the engine sensing signal is reconstructed using a basis pursuit algorithm, specifically including:

[0053] Establish a sparse representation model, define the observation matrix, and initialize the basis matrix Λ, the perception matrix A, and the sparse coefficient vector α;

[0054] According to the formula Minimize the L1 norm of the sparse coefficient vector to calculate the sparsity of the model under constraints, where Φ is the dictionary matrix, ε is the tolerance threshold of the reconstructed signal, y is the observed signal, |||1 is the L1 norm, used to represent the measure of sparsity, and |||2 is the constraint condition for the reconstruction error.

[0055] We use linear programming to optimize the sparse coefficient vector and obtain the sparse coefficient vector α after signal reconstruction. * Then, according to the formula x=Φα * The engine sensing signal was reconstructed.

[0056] In embodiments of this invention, a basis pursuit algorithm is used to represent the signal by selecting a suitable dictionary basis and initializing an empty sparse coefficient vector. Then, an iterative method is used to progressively select the dictionary element most similar to the current residual, and the sparse coefficients are updated according to the selected basis, while the residual is also updated. The algorithm iterates continuously until the change in the sparse coefficients is less than a certain threshold, and finally outputs a sparse signal representation.

[0057] Basis pursuit (BPT) is an optimization algorithm for signal processing and image reconstruction. Its core idea is to reconstruct the original signal by sparsely representing it. BPT leverages the sparsity of signals, treating electromagnetic pulse signals as a sparse representation under a certain basis. Using the electromagnetic pulse signal features obtained through Continuous Wound Theorem (CWT) as a dictionary matrix, the minimum L1 norm can be effectively solved using appropriate optimization methods to obtain the optimal sparse coefficient vector α. * .

[0058] The sensor electromagnetic pulse coupling signal separation and reconstruction method of the present invention will be described in detail below through specific embodiments.

[0059] Figure 2 This is a schematic diagram of the sensor electromagnetic pulse coupling signal separation and reconstruction method proposed in this invention. It consists of two parts: electromagnetic pulse identification and electromagnetic pulse coupling signal separation and sensor input signal reconstruction. The electromagnetic pulse identification method uses a combination of CWT and convolutional neural networks, while the electromagnetic pulse coupling signal separation and sensor input signal reconstruction methods are implemented based on deep clustering and source signal reconstruction as theoretical foundations.

[0060] Specifically, the electromagnetic pulse identification method consists of four parts: time-frequency processing, model building, model training, and electromagnetic pulse identification, as described below:

[0061] Step 1: Use CWT to convert the one-dimensional sensor signal into a corresponding two-dimensional time-frequency image. Select a suitable wavelet basis function ψ(t), calculate the wavelet coefficients W(a,b), and use the wavelet coefficient amplitude |W(a,b)| as color to discretize the scale a and translation parameter b, generating the time-frequency image x. CWT (f,t).

[0062] Step two, as Figure 3 As shown, a convolutional neural network model is constructed, which can extract features from the EMP-sensitive frequency band and quickly detect whether the sensor input signal of the digital controller is affected by EMP. Utilizing the residual skip structure design, the feature images generated by convolutional layer 7 and pooling layer 4 are upsampled 4 times and 2 times respectively. These are then combined with the feature image generated by pooling layer 3, and the combined result is upsampled 8 times to obtain the output result.

[0063] Step 3, as Figure 4 As shown, the two-dimensional time-frequency images are randomly divided into training and test sets according to a certain ratio. The training set is used to update the network model parameters, and the test set is used to validate the trained model. After labeling the fault classification labels on the time-frequency image training set, the constructed CNN model is trained using the training set. The model parameters are randomly initialized, and the cross-entropy loss function is selected as the objective function. The optimal network model structure parameters are saved based on the convergence trend of the cross-entropy loss.

[0064] Step four: After the convolutional neural network model completes the parameter update, the test set is used as the model input to test the accuracy of the constructed convolutional neural network model in recognizing electromagnetic pulse coupling signals.

[0065] Specifically, the electromagnetic pulse coupling signal separation method is the second part of the sensor input signal fault tolerance, which includes two steps: deep clustering and source signal reconstruction.

[0066] First, the time-frequency domain signal of the electromagnetic pulse coupling signal is extracted, and then STFT is performed on the multi-channel observation signal.

[0067]

[0068] In the formula, x(m,n) is the discrete-time domain signal, ω(nt) is the window function, m is the time frame index, f is the frequency index, and N is the frame length. By normalizing the logarithmic power spectrum to a uniform scale, the normalization formula is:

[0069]

[0070] By combining bidirectional LSTM with CNN to form a deep clustering network, the generated feature representation is as follows:

[0071]

[0072] Optimization is performed using a clustering loss function to generate an ideal binary mask. Based on this, a time-frequency mask is generated. The mask Q is defined by the softmax function:

[0073] Q = Soft max(V) T W c )

[0074] Deep clustering is performed using K-means clustering, and the specific steps are described below:

[0075] Step 1, based on the mask M k (f,t), create with x TF The same shape of mask matrix columns are used, and each time frequency point is assigned a value according to its cluster label to indicate the cluster class to which each time frequency point belongs.

[0076] Step 2: Extract the time-frequency components x corresponding to different clusters using time-frequency masks. k,TF (f,t)=M k (f,t)·x TF (f,t);

[0077] Step 3, according to the formula Calculate the Euclidean distance between a single source point and the cluster center. By setting a distance threshold U, the datasets whose distance from the cluster center is less than the threshold U are classified as having the same label.

[0078] The source signal reconstruction is performed using the basis pursuit method, and the specific steps are described below:

[0079] Step 1: Extract the sparse coefficient vector α of the observed signal using a sparse representation method;

[0080] Step 2, convert the time-frequency signal x TF (t,f) is represented as x TF (t,f)=Φα, where Φ is the dictionary matrix. The sparse coefficients are obtained by minimizing the objective function:

[0081]

[0082] In the formula, S is the current observed signal, and λ is the normalization parameter of the control coefficient.

[0083] Step 3: Use the inverse short-time Fourier transform to convert the time-frequency graph into a time-frequency signal x. (n) =iSTFT(S (n) init );

[0084] Step four: Introduce Wiener filtering to recover the distorted sensor input signal by minimizing the mean square error. The expression is:

[0085]

[0086] In the formula, S xy (f,t) is the cross spectrum of the signal and the observed signal, S xx (f,t) is the auto-spectrum of the signal, and N0 is the noise power spectrum;

[0087] Step 5, according to the formula Obtain the time-frequency representation after Wiener filtering;

[0088] Step 6: Initialize the complex time-frequency graph after Wiener filtering, expressed by the formula: The time-frequency image is updated in real time using the generated signal, and the phase information is also updated:

[0089]

[0090] In the formula, It is the time-frequency signal after Wiener filtering;

[0091] Step 7: Repeat steps 3 through 6 until the convergence criterion is met, and an approximate signal is obtained.

[0092] In the process of separating electromagnetic pulse coupling signals and reconstructing input signals, source points of the electromagnetic pulse coupling signals in the time-frequency domain are extracted. Then, wavelet coefficient matrices are generated based on wavelet transform to represent the time-frequency image. The time-frequency image is then used as the input to the CNN network model to identify the electromagnetic pulse coupling signals. K-means clustering is then used for feature clustering, and different time-frequency components are extracted based on the time-frequency mask. Finally, the reconstructed signal is optimized by combining the basis pursuit algorithm, the Griffin-Lim algorithm, and Wiener filtering. The estimated source signal components can be obtained from the reconstructed signal. Finally, the separated component signals are transformed to the time domain using inverse short-time Fourier transform, thus completing the separation of the electromagnetic pulse coupling signals.

[0093] The sensor electromagnetic pulse coupling signal separation and reconstruction method proposed in this invention identifies electromagnetic pulses through continuous wavelet transform and convolutional neural networks. Based on deep learning, a deep clustering-based electromagnetic pulse coupling signal separation method is designed. Joint optimization of feature representation and cluster assignment is used to separate electromagnetic pulse coupling signals under multi-source signals. Then, the basis pursuit algorithm is used to reconstruct the signal, and the Griffin-Lim algorithm and Wiener filtering are combined to iteratively optimize the reconstructed signal, achieving effective separation of electromagnetic pulse coupling signals, eliminating electromagnetic pulse interference, and achieving fault tolerance.

[0094] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0095] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0096] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

Claims

1. A method for separating and reconstructing electromagnetic pulse coupling signals from a sensor, characterized in that, Includes the following steps: Collect engine sensing signals; When the engine sensing signal is subject to electromagnetic pulse interference, continuous wavelet transform and convolutional neural network are used to identify the electromagnetic pulse interference signal, and the feature representation and clustering assignment are jointly optimized. The electromagnetic pulse coupling signal under multi-source signal is separated by deep clustering. The engine sensing signal is reconstructed using the basis pursuit algorithm. The engine sensing signal is then reconstructed by combining the Griffin-Lim algorithm and Wiener filtering iterative optimization. Electromagnetic pulse interference signals in the engine sensing signal are then separated and eliminated. Electromagnetic pulse interference signals are identified using continuous wavelet transform and convolutional neural networks, including: The one-dimensional engine sensing signal timing signal has a unified signal length and a fixed number of sampling points; The one-dimensional engine sensing signal time series signal is converted into a two-dimensional time-frequency image, and bilinear interpolation is used to unify the size of the time-frequency image; Image enhancement is performed on the two-dimensional time-frequency images, and the images are divided into training and testing sets. Construct a network model, train the network model in stages using the training set, select the cross-entropy loss function in multi-classification tasks, and select network model parameters based on the convergence trend of the cross-entropy loss. The test set is input into the network model. By identifying the characteristics of sensor malfunction signals, it is determined whether the engine sensing signal is affected by electromagnetic interference. The network model parameters are updated and optimized by combining the classification results and judgment accuracy of sensor malfunctions. The engine sensing signal is reconstructed using the basis pursuit algorithm, specifically including: Establish a sparse representation model, define the observation matrix, and initialize the basis matrix. Perception Matrix Sparse coefficient vector ; According to the formula ( Minimize the L1 norm of the sparse coefficient vector to calculate the sparsity of the model under constraints, where, It is a dictionary matrix. To reconstruct the tolerance threshold of the signal, For observing signals, The L1 norm is used to represent a measure of sparsity. These are constraints imposed to address reconstruction errors. The sparse coefficient vector after signal reconstruction is obtained by optimizing the sparse coefficient vector using linear programming. Then, according to the formula The engine sensing signal was reconstructed.

2. The method according to claim 1, characterized in that, The network model is a convolutional neural network.

3. The method according to claim 1, characterized in that, Signal separation of electromagnetic pulse coupling signals from multiple sources is achieved through deep clustering, specifically including: Electromagnetic pulse coupling signal from time domain signal Convert to time-frequency domain representation ; Based on the trained network model, signal features are extracted, and the center points of each single source are selected. According to the formula Calculate the Euclidean distance between a single source point and the cluster center by setting a distance threshold. Those whose distance from the cluster center is less than a threshold The datasets are classified under the same label, where, For Euclidean distance, For feature vectors, This represents the number of feature points.

Images