Ship pressure sensor virtual connection fault identification method
Through the deep learning network model, the virtual contact fault signal of ship pressure sensor is preprocessed and feature extracted, which solves the problems of low efficiency and poor recognition generalization in traditional methods, and realizes efficient fault recognition and accurate virtual contact fault detection. Suitable for real-time applications in complex environments.
Patent Information
- Application Number
- CN202311366930.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-20
- Publication Date
- 2024-12-27
AI Technical Summary
Ship pressure sensors are easily disturbed in complex environments, resulting in unstable signal recording, traditional fault detection methods are inefficient and poor generalization of identification, especially in variable working conditions, it is difficult to process complex structures or highly nonlinear data, and the calculation time is long and Serious information loss.
Using deep learning network model, the fault feature model is constructed by pre-processing, feature extraction and measurement learning of the virtual fault signal of the ship pressure sensor, and the fault feature model is realized to identify virtual faults online, using Fourier transform, ResNet and autoencoder technologies. Perform signal processing and feature extraction, and combine multi-layer perceptron and classifier for fault classification.
It effectively improves the efficiency of ship pressure sensor virtual contact fault detection, improves recognition accuracy, is suitable for real-time applications, can better handle complex structures and nonlinear data, and reduces information loss.
Smart Images
Figure CN119202879A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of ship pressure sensor fault identification, in particular to a ship pressure sensor virtual connection fault identification method. Background Art
[0002] Due to the complex operating environment of ship pressure sensors, the detection system is easily affected by many external factors such as wave resistance and cabin equipment operation noise when collecting data, which leads to more interference in the signals collected by the ship pressure sensors, affecting the stability of the equipment signal recording. In addition, the data collected by ship pressure sensors has the characteristics of large data base, many feature types and high randomness. If the traditional signal analysis-based method is used for ship pressure sensor virtual connection fault detection, there are disadvantages of low detection efficiency and poor recognition generalization.
[0003] The 19th issue of Machine Tool & Hydraulics in 2022 published an article titled "Fault Diagnosis of Rolling Bearings under Variable Conditions Based on SHO-VMD Decomposition and Multi-Feature Parameters" by Liu Wei, Liang Tao, Li Tao and Jiang Wen. The article mentioned that when the fan is running under various working conditions, it is difficult to diagnose the fault using the signals detected by the bearing vibration monitoring system, and the bearing fault diagnosis studied in a large number of literatures is mostly carried out at a constant speed. For rolling bearings running under variable conditions, a feature extraction method based on SHO-VMD decomposition and multi-feature parameter fusion is proposed, and t-SNE dimension reduction visualization is used to extract the fault information and speed change information of the vibration signal. The decomposition effect of the variational mode decomposition (VMD) method depends on the number of decompositions and the value of the penalty factor. The selfish herd optimization algorithm (SHO) is used to optimize the parameters, decompose the vibration signal into some intrinsic mode components, and then extract the characteristic parameters of each component. Multi-feature quantity fusion is performed based on singular value features, energy entropy, and sample entropy features. t-SNE dimension reduction is used to extract bearing fault information and speed change information. The experimental results show that the proposed method can effectively extract bearing fault and speed information. The disadvantage of this method is that the method based on SHO-VMD decomposition and multi-feature parameter fusion usually requires a lot of calculations, including feature dimension reduction, variational mode decomposition, and feature parameter extraction. This may lead to a long calculation time and is not suitable for real-time applications. In addition, this method cannot handle data with complex structures or high nonlinearity well. When calculating feature parameters, some associated features may be ignored, resulting in information loss. Summary of the invention
[0004] In order to solve the above technical problems, the present invention proposes a method for identifying virtual connection faults of ship pressure sensors. By analyzing the characteristics of the virtual connection fault signals of the ship pressure sensors, a deep learning network model with the fault characteristics as input parameters is constructed to realize online identification of virtual connection faults of ship pressure sensors.
[0005] The present invention provides a method for identifying a virtual connection fault of a ship pressure sensor, comprising collecting pressure data recorded by the ship pressure sensor, and further comprising the following steps:
[0006] Step 1: preprocessing the pressure data;
[0007] Step 2: Extract ship pressure characteristics;
[0008] Step 3: Conduct ship pressure measurement learning;
[0009] Step 4: Identify the fault condition.
[0010] Preferably, the step 1 comprises the following sub-steps:
[0011] Step 11: Use data cleaning to filter out erroneous information;
[0012] Step 12: Use a high-pass filter to filter out low-frequency interference;
[0013] Step 13: Use filtering to smooth out high frequency interference.
[0014] In any of the above schemes, it is preferred that the step 12 includes filtering the low-frequency noise that occurs during the recording of the ship pressure signal, eliminating it using a second-order high-pass filter, and suppressing the noise signal below 20 Hz during the recording process, wherein the transfer function of the second-order high-pass filter is The formula is
[0015]
[0016] Where K is the filter gain, ξ is the damping coefficient, ω is the system angular frequency, and s is the complex frequency.
[0017] In any of the above schemes, preferably, step 13 includes setting a fixed-size signal window, performing polynomial fitting on the signal data, repeatedly moving the sampling interval, performing pressure sensor signal translation calculation, and smoothing the ship pressure sensor signal. The calculation formula is
[0018]
[0019] Where y(i) represents the original input pressure signal, represents the convolution function, m is the number of given pressure signal data, and i represents the i-th pressure signal being processed.
[0020] In any of the above schemes, preferably, step 2 comprises the following sub-steps:
[0021] Step 21: Input pressure sensor signal X∈R L×D A Fourier transform is performed, effectively converting the input signal into a frequency domain representation X Q =F(X), where R represents the real number domain, L represents the first L time points, D represents the feature dimension, and F(X) represents the Fourier transform function;
[0022] Step 22: Use the nonlinear feature encoder based on ResNet to perform nonlinear fully connected layer calculation on the Fourier transformed data;
[0023] Step 23: Divide the encoded features into two branches and input them into the classifier and decoder respectively;
[0024] Step 24: Use the softmax function to map the output of the model to the probability distribution of different categories.
[0025] In any of the above schemes, preferably, step 22 also includes adopting an autoencoder structure, using a feature decoder based on a deconvolution operation, generating a feature data estimate of the frequency domain pressure signal, and using a mean square error as a reconstruction loss metric.
[0026] In any of the above solutions, it is preferred that when training the multi-task model, the loss L of the classification task is cls and the loss L for the reconstruction task dec Combine them to get the loss function L, the formula is
[0027] L=α·L cls +β·L dec
[0028] Among them, α and β are hyperparameters.
[0029] In any of the above solutions, preferably, step 23 includes designing a multi-layer perceptron, the input of which is the hidden state of the encoder, including three fully connected layers, and using a sigmoid activation function after each layer.
[0030] In any of the above schemes, preferably, step 3 includes the following sub-steps:
[0031] Step 31: The pre-processed and Fourier-transformed pressure signal and the virtual connection fault signal are respectively input into the metric learning model, and the extracted signal features are embedded into the nearest neighbor position through the convolutional neural network based on bilinear interpolation calculation to obtain the feature vectors of different sample pressure signals;
[0032] Step 32: Perform spatial distance calculation on the pressure feature vector output by the encoder to achieve similarity comparison of similar signal samples;
[0033] Step 33: Perform global regularization loss on the input pressure frequency domain signal feature vector, and perform iterative optimization model training in combination with back propagation.
[0034] In any of the above solutions, preferably, the calculation formula of the spatial distance is:
[0035] d=||f(x j )-f(y j )|| 2
[0036] Where d represents the signal sample x j The characteristic f(x j ) and sample y j The characteristic f(y j ), j represents the jth sample feature.
[0037] In any of the above schemes, preferably, the calculation formula of the loss function is:
[0038]
[0039] In the formula, d represents the relative distance between the two samples, y represents whether the two signal samples match, y = 1 represents that the two samples are similar, y = 0 represents that the two samples are not similar, and Y mar Represents the set threshold, N represents the total number of input sample data, and n represents the nth input sample data.
[0040] In any of the above schemes, preferably, step 4 includes the following sub-steps:
[0041] Step 41: The output result of the multilayer perceptron is input into the classifier network to classify the normal ship pressure signal and the virtual connection fault signal;
[0042] Step 42: Calculate the similarity between the input pressure signals to identify similar frequency domain virtual connection fault signals;
[0043] Step 43: By combining the output results of the classifier network with the similar fault results calculated by the measurement model, the recognition accuracy of the present invention is comprehensively improved, and finally the recognition of the ship virtual connection fault signal is realized.
[0044] The present invention proposes a method for identifying virtual connection faults of ship pressure sensors, which can effectively improve the efficiency of virtual connection fault detection of ship pressure sensors and has more advantages than traditional feature selection and extraction methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 The present invention is a flow chart of a preferred embodiment of a method for identifying a virtual connection fault of a ship pressure sensor according to the present invention.
[0046] Figure 2 The present invention is a flow chart of another preferred embodiment of the method for identifying a virtual connection fault of a ship pressure sensor according to the present invention.
[0047] Figure 3 The present invention is a training flow chart of an embodiment of a ship pressure virtual connection fault autoencoder according to a ship pressure sensor virtual connection fault identification method of the present invention.
[0048] Figure 4 The present invention is a training flow chart of an embodiment of a ship pressure virtual connection fault metric learning model according to the ship pressure sensor virtual connection fault identification method of the present invention.
[0049] Figure 5 The diagram is a schematic diagram of recorded data of a ship pressure sensor according to an embodiment of the ship pressure sensor virtual connection fault identification method of the present invention.
[0050] Figure 6 The diagram is a schematic diagram of recorded data of a ship pressure sensor according to an embodiment of the ship pressure sensor virtual connection fault identification method of the present invention. DETAILED DESCRIPTION
[0051] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.
[0052] Embodiment 1
[0053] like Figure 1 As shown, step 100 is executed to collect the pressure data recorded by the ship pressure sensor.
[0054] Execute step 110 to pre-process the pressure data, including the following sub-steps:
[0055] Execute step 111 to filter out error information using data cleaning.
[0056] Execute step 112, use a high-pass filter to filter out low-frequency interference, use a second-order high-pass filter to eliminate low-frequency noise filtering that occurs during the recording of the ship pressure signal, and suppress noise signals below 20 Hz during the recording process, where the transfer function of the second-order high-pass filter is The formula is
[0057]
[0058] Where K is the filter gain, ξ is the damping coefficient, ω is the system angular frequency, and s is the complex frequency.
[0059] Execute step 113, use filtering to smooth out high-frequency interference, set a fixed-size signal window, perform polynomial fitting on the signal data, repeatedly move the sampling interval, perform pressure sensor signal translation calculation, and smooth the ship pressure sensor signal. The calculation formula is
[0060]
[0061] Where y(i) represents the original input pressure signal, represents the convolution function, m is the number of given pressure signal data, and i represents the i-th pressure signal being processed.
[0062] Execute step 120 to extract the ship pressure characteristics, including the following sub-steps:
[0063] Execute step 121 to input the pressure sensor signal X∈R L×D A Fourier transform is performed, effectively converting the input signal into a frequency domain representation X Q =F(X), where R represents the real number domain, L represents the first L time points, D represents the feature dimension, and F(X) represents the Fourier transform function;
[0064] Execute step 122, use a nonlinear feature encoder based on ResNet to perform nonlinear fully connected layer calculation on the Fourier transformed data, use an autoencoder structure, use a feature decoder based on deconvolution operation, generate feature data estimation of the frequency domain pressure signal, and use mean square error as a reconstruction loss metric;
[0065] When training a multi-task model, the loss L of the classification task is cls and the loss L for the reconstruction task dec Combine them to get the loss function L, the formula is
[0066] L=α·L cls +β·L dec
[0067] Among them, α and β are hyperparameters.
[0068] Execute step 123, divide the encoded features into two branches, input them into the classifier and the decoder respectively, design a multi-layer perceptron, the input of the multi-layer perceptron is the hidden state of the encoder, includes three fully connected layers, and uses a sigmoid activation function after each layer.
[0069] Execute step 124 and use the softmax function to map the output of the model to the probability distribution of different categories.
[0070] Execute step 130 to perform ship pressure measurement learning, including the following sub-steps:
[0071] Execute step 131, the pressure signal and the virtual connection fault signal that have been preprocessed and Fourier transformed are respectively input into the metric learning model, and the extracted signal features are embedded into the nearest neighbor position through a convolutional neural network based on bilinear interpolation calculation to obtain feature vectors of different sample pressure signals;
[0072] Execute step 132 to calculate the spatial distance of the pressure feature vector output by the encoder to achieve similarity comparison of similar signal samples. The calculation formula of the spatial distance is:
[0073] d=||f(x j )-f(y j )|| 2
[0074] Where d represents the signal sample x j The characteristic f(x j ) and sample y j The characteristic f(y j ), j represents the jth sample feature.
[0075] Execute step 133, perform global regularization loss on the input pressure frequency domain signal feature vector, and perform iterative optimization model training in combination with back propagation. The calculation formula of the loss function is:
[0076]
[0077] Where d represents the relative distance between the two samples, y represents whether the two signal samples match, y = 1 represents that the two samples are similar, y = 0 represents that the two samples are not similar, and Y mar Represents the set threshold, N represents the total number of input sample data, and n represents the nth input sample data.
[0078] Execute step 140 to identify the fault state, including the following sub-steps:
[0079] Step 41: The output result of the multilayer perceptron is input into the classifier network to classify the normal ship pressure signal and the virtual connection fault signal;
[0080] Step 42: Calculate the similarity between the input pressure signals to identify similar frequency domain virtual connection fault signals;
[0081] Step 43: By combining the output results of the classifier network with the similar fault results calculated by the measurement model, the recognition accuracy of the present invention is comprehensively improved, and finally the recognition of the ship virtual connection fault signal is realized.
[0082] The present invention analyzes the characteristics of the virtual connection fault signal of the ship pressure sensor, constructs a deep learning network model with the fault characteristics as input parameters, and realizes online identification of the virtual connection fault of the ship pressure sensor.
[0083] Embodiment 2
[0084] The present invention provides a deep learning modeling method based on fault feature extraction, analyzes the recorded data of the ship pressure sensor, and realizes sensor virtual connection fault detection through feature extraction and similarity calculation. The present invention adopts the following technical solutions:
[0085] The present invention adopts a feature extraction model based on frequency domain transformation and autoencoder to extract features from signals collected by a ship pressure sensor, and designs a pressure sensor virtual connection fault detection method based on similarity measurement and nonlinear classifier to achieve online identification of pressure sensor virtual connection faults.
[0086] The specific detection method is as follows: first, the pressure data recorded by the ship pressure sensor is collected and preprocessed; then, the preprocessed ship pressure sensor recorded data and the virtual connection fault data are transformed by Fourier frequency domain features; then, the signal features are extracted from the frequency domain signal by designing and training the autoencoder; finally, based on the extracted features, the nonlinear classifier and the metric learning model are combined to realize the online identification of the virtual connection fault of the ship pressure sensor.
[0087] like Figure 2 The ship pressure sensor fault identification algorithm shown in the figure is based on frequency domain feature extraction, and its main contents include:
[0088] Data acquisition and preprocessing module;
[0089] Ship pressure feature extraction module;
[0090] Ship pressure measurement learning module;
[0091] Fault status identification module.
[0092] In the signal acquisition process, ship pressure sensors are used for data acquisition. By collecting and recording ship pressure signals, one-dimensional ship pressure signal processing is performed through the signal monitoring system, and the acquisition of ship pressure sensor signals is achieved through export and storage operations. Figure 5 and Figure 6 In addition, during the data preprocessing process, there will be interference from external factors, and there will be ship pressure signal data contaminated by random noise. Therefore, before analyzing the signal, it is necessary to eliminate the noise caused by the environment, instruments and human factors to provide more accurate information for subsequent analysis.
[0093] Step 1: Use data cleaning to filter out error information
[0094] The present invention performs data preprocessing operations on the signals recorded by the pressure sensor. By performing data cleaning and enhancement operations on the signal data recorded by the ship pressure sensor, errors, redundancy and missing information in the original signal data are effectively removed, ensuring the validity and availability of the original data.
[0095] Step 2: Use a high-pass filter to filter out low-frequency interference
[0096] The present invention uses a second-order high-pass filter to eliminate the low-frequency noise that occurs during the recording process of the ship pressure signal, and suppresses the noise signal below 20Hz during the recording process, wherein the transfer function of the second-order high-pass filter is The formula is as follows:
[0097]
[0098] Where K is the filter gain, ξ is the damping coefficient, ω is the system angular frequency, and s is the complex frequency.
[0099] Step 3: Use filtering to smooth out high-frequency interference
[0100] The smoothing filter method is a peak-preserving smoothing method that suppresses high-frequency noise. By setting a fixed-size signal window, the signal data is fitted with a polynomial, and the sampling interval is repeatedly moved based on this to achieve the pressure sensor signal translation calculation. Among them, the smoothed ship pressure sensor signal The calculation is as follows:
[0101]
[0102] Where y(i) represents the original input pressure signal, Represents the convolution function, and m is the number of given pressure signal data. By taking the high frequency vicinity of the high frequency signal as the sampling point and fitting the signal data polynomial, the shape of the original input pressure signal peak, that is, the signal height, width and position remain unchanged, while effectively reducing the noise in the signal, the signal-to-noise ratio of the data is effectively improved.
[0103] like Figure 3 As shown in the figure, in the ship pressure feature extraction module, this study adopted a multi-task autoencoder based on frequency domain transformation, because abnormal signals usually show different characteristics from normal signals in the frequency domain, which makes the extraction of frequency domain information very beneficial for subsequent fault detection work.
[0104] Step 4: First, the input pressure sensor signal X∈R L×DA Fourier transform is performed, effectively converting the input signal into a frequency domain representation X Q =F(X).
[0105] Step 5: Use a ResNet-based nonlinear feature encoder to perform nonlinear fully connected layer calculations on the Fourier transformed data, which helps capture the key information of the input pressure sequence.
[0106] Step 6: Subsequently, the encoded features are divided into two branches and input into the classifier and decoder respectively. First, for the classification task, a multi-layer perceptron (MLP) is designed, which accepts the hidden state of the encoder as input. It includes three fully connected layers and uses a sigmoid activation function after each layer to learn task-related feature representations in a targeted manner.
[0107] Step 7: Finally, use the softmax function to map the output of the model to the probability distribution of different categories.
[0108] Step 8: Secondly, for the reconstruction task, the present invention adopts an autoencoder structure and uses a feature decoder based on deconvolution operation to generate feature data estimation of the frequency domain pressure signal. In order to ensure that the model can effectively restore the original data, the present invention uses mean square error (MSE) as the reconstruction loss metric.
[0109] Step 9: When training a multi-task model, consider the loss functions of the two tasks: the loss L of the classification task cls and the loss L for the reconstruction task dec To this end, a multi-task loss function is defined, which organically combines the cross entropy loss of the classification task and the mean square error (MSE) loss of the reconstruction task:
[0110] L=α·L cls +β·L dec
[0111] Among them, α and β are hyperparameters used to adjust the weights of the two task losses.
[0112] like Figure 4 As shown in the figure, in the ship pressure signal metric learning module, the pressure frequency domain signal after Fourier transform processing is used as the input sample, and the loss function is designed according to the sample distribution in the metric learning training process; then, on this basis, the back propagation algorithm is combined to realize the gradient update of the ship virtual connection fault module, realize the mapping of pressure sample features to the embedding space, and combine the similarity metric function to realize the identification of similar virtual connection faults in the ship pressure signal. The ship pressure metric learning steps are as follows:
[0113] Step 10: First, the preprocessed and Fourier transformed pressure signal and virtual connection fault signal are input into the metric learning model respectively. The extracted signal features are embedded into the nearest neighbor position through the convolutional neural network based on bilinear interpolation calculation to obtain the feature vectors of different sample pressure signals.
[0114] Step 11: Then, perform spatial distance calculation on the pressure feature vector output by the encoder to achieve similarity comparison of similar signal samples. The calculation formula is as follows:
[0115] d=||f(x j )-f(y j )|| 2
[0116] Where d represents the signal sample x j The characteristic f(x j ) and sample y j The characteristic f(y j ) relative distance between them.
[0117] Step 12: During the training process, the input pressure frequency domain signal feature vector is subjected to global regularization loss, and the model training is iteratively optimized in combination with back propagation. The distance between samples of the same category is as small as possible by optimizing the feature vector for the spatially distributed positive and negative sample pressure features. The loss function is calculated as follows:
[0118]
[0119] In the formula, y represents whether the two signal samples match, y = 1 represents that the two samples are similar, y = 0 represents that the two samples are not similar, Y mar Represents setting threshold.
[0120] Step 13: During the testing phase, the similarity between the ship pressure signal and the fault signal is calculated in real time, and the recognition result of the classification network and the output of the signal restoration module are combined to comprehensively realize the accurate recognition of the fault signal.
[0121] In the ship fault state recognition module, combined with the calculation results of the classifier and the pressure measurement similarity, a fault recognition module is used to realize virtual connection fault detection. The implementation steps are as follows:
[0122] Step 14: By inputting the output results of the multi-layer perceptron into the classifier network, the classification of normal ship pressure signals and virtual connection fault signals is achieved.
[0123] Step 15: Then, the similarity between the input pressure signals is calculated to realize the identification of similar frequency domain virtual connection fault signals.
[0124] Step 16: By combining the output results of the classifier network with the similar fault results calculated by the measurement model, the recognition accuracy of the present invention is comprehensively improved, and finally the recognition of the ship virtual connection fault signal is realized.
[0125] The present invention adopts a pressure sensor fault identification method based on deep feature similarity calculation, which can effectively improve the efficiency of ship pressure sensor virtual connection fault detection and has more advantages than traditional feature selection and extraction methods.
[0126] Embodiment 3
[0127] The present invention will use three indicators, namely, accuracy P, recall R and precision A, to measure the effect of the proposed model. The calculation formulas of each indicator are as follows.
[0128] P=TP / (TP+FP)
[0129] R=TP / (TP+FN)
[0130] A=(TP+TN) / (TP+FP+TN+FN)
[0131] Among them, TP indicates that the true category of the signal sample and the predicted result are both positive; FP indicates that the true category of the sample is negative, and the final predicted result is positive; FN indicates that the true category of the sample is positive, and the final predicted result is negative; TN indicates that the true category of the sample and the final predicted result are both negative.
[0132] In order to better understand the present invention, the above is described in detail in conjunction with the specific embodiments of the present invention, but it is not intended to limit the present invention. Any simple modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention. Each embodiment in this specification focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referenced to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
Claims
1. A method for identifying a virtual connection fault of a ship pressure sensor, comprising collecting pressure data recorded by the ship pressure sensor, characterized in that: The following steps are also included: Step 1: preprocessing the pressure data; Step 2: Extract ship pressure characteristics; Step 3: Conduct ship pressure measurement learning; Step 4: Identify the fault condition.
2. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 1, characterized in that: The step 1 includes the following sub-steps: Step 11: Use data cleaning to filter out erroneous information; Step 12: Use a high-pass filter to filter out low-frequency interference; Step 13: Use filtering to smooth out high frequency interference.
3. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 2, characterized in that: The step 12 includes filtering the low-frequency noise that occurs during the recording of the ship pressure signal, eliminating it using a second-order high-pass filter, and suppressing the noise signal below 20 Hz during the recording process, wherein the transfer function of the second-order high-pass filter is The formula is Where K is the filter gain, ξ is the damping coefficient, ω is the system angular frequency, and s is the complex frequency.
4. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 3, characterized in that: The step 13 includes setting a fixed-size signal window, performing polynomial fitting on the signal data, repeatedly moving the sampling interval, performing pressure sensor signal translation calculation, and smoothing the ship pressure sensor signal. The calculation formula is Where y(i) represents the original input pressure signal, represents the convolution function, m is the number of given pressure signal data, and i represents the i-th pressure signal being processed.
5. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 4, characterized in that: The step 2 includes the following sub-steps: Step 21: Input pressure sensor signal X∈R L×D A Fourier transform is performed, effectively converting the input signal into a frequency domain representation X Q =F(X), where R represents the real number domain, L represents the first L time points, D represents the feature dimension, and F(X) represents the Fourier transform function; Step 22: Use the ResNet-based nonlinear feature encoder to perform nonlinear fully connected layer calculation on the Fourier transformed data; Step 23: Divide the encoded features into two branches and input them into the classifier and decoder respectively; Step 24: Use the softmax function to map the output of the model to the probability distribution of different categories.
6. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 5, characterized in that: The step 22 also includes adopting an autoencoder structure, using a feature decoder based on a deconvolution operation, generating a feature data estimate of the frequency domain pressure signal, and using a mean square error as a reconstruction loss metric.
7. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 6, characterized in that: When training a multi-task model, the loss L of the classification task is cls and the loss L for the reconstruction task dec Combine them to get the loss function L, the formula is L=α·L cls +β·L dec Among them, α and β are hyperparameters.
8. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 7, characterized in that: The step 23 includes designing a multi-layer perceptron, the input of which is the hidden state of the encoder, including three fully connected layers, and using a sigmoid activation function after each layer.
9. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 8, characterized in that: The step 3 includes the following sub-steps: Step 31: The pre-processed and Fourier-transformed pressure signal and the virtual connection fault signal are respectively input into the metric learning model, and the extracted signal features are embedded into the nearest neighbor position through the convolutional neural network based on bilinear interpolation calculation to obtain the feature vectors of different sample pressure signals; Step 32: Perform spatial distance calculation on the pressure feature vector output by the encoder to achieve similarity comparison of similar signal samples; Step 33: Perform global regularization loss on the input pressure frequency domain signal feature vector, and perform iterative optimization model training in combination with back propagation.
10. The method for identifying a virtual connection fault of a ship pressure sensor according to claim 9, characterized in that: The calculation formula of the spatial distance is: d=||f(x j )-f(y j )|| 2 Where d represents the signal sample x j The characteristic f(x j ) and sample y j The characteristic f(y j ), j represents the jth sample feature.