A vehicle-mounted hydrogen system leakage diagnosis method, system, device and storage medium

Abstract

The application discloses a vehicle-mounted hydrogen system leakage diagnosis method, system, equipment and storage medium, and relates to the field of hydrogen leakage diagnosis. The method comprises the following steps: acquiring actual gas pressure data in a hydrogen cylinder of a fuel cell vehicle; constructing a weighted fractional order Fourier recurrence graph and a time domain recurrence graph based on the actual gas pressure data according to a Mahalanobis distance; performing lower triangular conversion and fusion on the weighted fractional order Fourier recurrence graph and the time domain recurrence graph to obtain a fused recurrence graph; and performing diagnosis on the fused recurrence graph by using a hydrogen leakage diagnosis network to obtain a hydrogen leakage diagnosis result. The application can simultaneously realize rapid and accurate identification of trace hydrogen leakage and large hydrogen leakage.

Description

Technical Field

[0001] This invention relates to the field of hydrogen leak diagnosis, and in particular to a method, system, device and storage medium for diagnosing leaks in vehicle-mounted hydrogen systems. Background Technology

[0002] Hydrogen is colorless, odorless, flammable, and highly permeable, making it extremely prone to seepage and leakage during use. A hydrogen leak in a hydrogen system can lead to combustion, explosion, or other dangerous situations, posing a significant safety threat to the vehicle and its passengers. Therefore, accurate monitoring of hydrogen leaks in onboard hydrogen systems helps prevent dangerous accidents and improves the reliability, safety, and market acceptance of fuel cell vehicles.

[0003] Currently, the primary method for monitoring hydrogen system leaks in fuel cell vehicles is sensor detection. However, sensor detection is highly susceptible to the number and installation location of sensors. Even with high-precision hydrogen concentration sensors, leak detection can be slow if the sensors are located far from the leak point, hindering rapid diagnostic results. Furthermore, when objects obstruct the view between the leak point and the sensor, the sensor cannot promptly acquire hydrogen diffusion concentration information, leading to excessively long diagnostic times and even inaccurate leak diagnoses. For diagnosing minute leaks, single-sensor detection cannot provide accurate judgments, resulting in low reliability.

[0004] In summary, existing methods for diagnosing leaks in on-board hydrogen systems cannot quickly, efficiently, and accurately identify leaks in hydrogen systems, nor can they simultaneously identify both minor and major leaks. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device and storage medium for diagnosing leaks in vehicle-mounted hydrogen systems, which can simultaneously achieve rapid and accurate identification of both trace and large-scale hydrogen leaks.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for diagnosing leaks in an on-board hydrogen system includes:

[0008] Obtain actual gas pressure data inside the hydrogen tank of a fuel cell vehicle;

[0009] Based on the Mahalanobis distance, a weighted fractional Fourier recursive graph and a time-domain recursive graph are constructed respectively using the actual gas pressure data.

[0010] The weighted fractional Fourier recursive graph and the time-domain recursive graph are subjected to a lower triangular transformation and fused to obtain a fused recursive graph;

[0011] Based on the fusion recursive graph, a hydrogen leak diagnosis network is used to perform diagnosis, and the hydrogen leak diagnosis results are obtained.

[0012] Optionally, a weighted fractional Fourier recurrence graph and a time-domain recurrence graph are constructed based on the Mahalanobis distance and the actual gas pressure data, specifically including:

[0013] The actual gas pressure data is separated to obtain a fixed-length non-overlapping continuous time-series signal segment;

[0014] The non-overlapping continuous time-series signal segments in each actual gas pressure data are used to extract the weighted fractional Fourier pressure signal segments.

[0015] Based on Mahalanobis distance, a weighted fractional Fourier recursive graph is constructed from the weighted fractional Fourier pressure signal segments;

[0016] A time-domain recursive graph is constructed based on the Mahalanobis distance from the non-overlapping continuous time-series signal segments.

[0017] Optionally, a weighted fractional Fourier recursive graph is constructed based on the Mahalanobis distance according to the weighted fractional Fourier pressure signal segments, specifically including:

[0018] The embedding dimension and delay time are selected based on the weighted fractional Fourier stress signal segment using the CAO method and the mutual information method;

[0019] Spatial reconstruction is performed based on the embedding dimension and delay time to obtain the reconstructed phase space;

[0020] The distance is calculated based on the covariance matrix of the actual gas pressure data and the reconstructed phase space, using Mahalanobis distance.

[0021] Calculate the recursive value based on the distance;

[0022] Construct a weighted fractional Fourier recursion graph based on the recursive values.

[0023] Optionally, the weighted fractional Fourier recursive graph and the time-domain recursive graph are subjected to a lower triangular transformation and fused to obtain a fused recursive graph, specifically including:

[0024] The weighted fractional Fourier recursive graph and the time-domain recursive graph are transformed by a lower triangular transformation to obtain the weighted fractional Fourier domain recursive graph.

[0025] The time-domain recursive graph is transformed into a lower triangular graph to obtain a time-domain lower triangular recursive graph.

[0026] The weighted fractional Fourier domain recursive graph and the time-domain lower triangular recursive graph are fused using the Laplace pyramid image fusion method to obtain a fused recursive graph.

[0027] Optionally, the hydrogen leak diagnosis network is a trained VGG-16 neural network.

[0028] The present invention also provides an on-board hydrogen system leak diagnosis system, comprising:

[0029] The acquisition module is used to acquire actual gas pressure data inside the hydrogen tank of a fuel cell vehicle;

[0030] The module is used to construct a weighted fractional Fourier recursive graph and a time-domain recursive graph based on the actual gas pressure data using Mahalanobis distance.

[0031] The lower triangular transformation and fusion module is used to perform lower triangular transformation and fusion on the weighted fractional Fourier recursive graph and the time-domain recursive graph to obtain a fused recursive graph;

[0032] The diagnostic module is used to perform a diagnosis based on the fused recursive graph using a hydrogen leak diagnostic network to obtain a hydrogen leak diagnostic result.

[0033] Optionally, the building module specifically includes:

[0034] A separation unit is used to separate the actual gas pressure data to obtain a fixed-length non-overlapping continuous time-series signal segment;

[0035] The extraction unit is used to extract the weighted fractional Fourier pressure signal segment from the non-overlapping continuous time-series signal segments in each actual gas pressure data through weighted fractional Fourier transform.

[0036] The first construction unit is used to construct a weighted fractional Fourier recursive graph based on the Mahalanobis distance according to the weighted fractional Fourier pressure signal segment;

[0037] The second construction unit is used to construct a time-domain recursive graph based on the Mahalanobis distance from the non-overlapping continuous time-series signal segments.

[0038] Optionally, the first building unit specifically includes:

[0039] Select sub-units to select the embedding dimension and delay time based on the weighted fractional Fourier stress signal segment using the CAO method and the mutual information method;

[0040] The spatial reconstruction subunit is used to perform spatial reconstruction based on the embedding dimension and delay time to obtain the reconstructed phase space;

[0041] The distance calculation subunit is used to calculate the distance based on the covariance matrix of the actual gas pressure data and the reconstructed phase space, using the Mahalanobis distance.

[0042] A recursive value calculation subunit is used to calculate a recursive value based on the distance;

[0043] A recursion graph construction subunit is used to construct a weighted fractional Fourier recursion graph based on the recursion values.

[0044] The present invention also provides an electronic device, comprising:

[0045] One or more processors;

[0046] A storage device on which one or more programs are stored;

[0047] When the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0048] The present invention also provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described thereon.

[0049] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] This invention utilizes weighted fractional Fourier transform to process pressure data. On the one hand, the weighting mechanism can emphasize specific frequency components in the signal, and on the other hand, it can extract more significant leakage features and deep frequency domain features to identify fractional-order minute leakage signals in vehicle hydrogen systems. This is beneficial for leakage type identification and machine learning. By using Mahalanobis distance to process the recurrence graph, the overall adaptability is improved, thereby achieving rapid and accurate identification of hydrogen leaks. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 Schematic diagram of a method for diagnosing leaks in an onboard hydrogen system;

[0053] Figure 2 A flowchart of the overall method for diagnosing leaks in automotive hydrogen systems;

[0054] Figure 3 Flowchart for selecting the optimal transformation order;

[0055] Figure 4 A schematic diagram for constructing a recursive graph of the weighted fractional Fourier domain;

[0056] Figure 5 A schematic diagram for constructing a time-domain recursive graph;

[0057] Figure 6A schematic diagram of the fusion process of the Laplace pyramid;

[0058] Figure 7 This is a diagram of the VGG-16 neural network model architecture.

[0059] Figure 8 To identify the accuracy change curve;

[0060] Figure 9 The flowchart shows the method for diagnosing leaks in an on-board hydrogen system provided by this invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The purpose of this invention is to provide a method, system, device and storage medium for diagnosing leaks in vehicle-mounted hydrogen systems, which can simultaneously achieve rapid and accurate identification of both trace and large-scale hydrogen leaks.

[0063] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] like Figure 1 , Figure 2 and Figure 9 As shown, the present invention provides a method for diagnosing leaks in an on-board hydrogen system, comprising:

[0065] Step 101: Obtain the actual gas pressure data inside the hydrogen tank of the fuel cell vehicle.

[0066] Step 102: Based on the Mahalanobis distance, construct a weighted fractional Fourier recursive graph and a time-domain recursive graph according to the actual gas pressure data.

[0067] Step 102 specifically includes: separating the actual gas pressure data to obtain non-overlapping continuous time-series signal segments of fixed length; extracting weighted fractional Fourier pressure signal segments from the non-overlapping continuous time-series signal segments in each actual gas pressure data using weighted fractional Fourier transform; constructing a weighted fractional Fourier recursive graph based on the weighted fractional Fourier pressure signal segments using Mahalanobis distance; and constructing a time-domain recursive graph based on the non-overlapping continuous time-series signal segments using Mahalanobis distance.

[0068] Constructing a weighted fractional Fourier recursive graph based on Mahalanobis distance from the weighted fractional Fourier pressure signal segments specifically includes: selecting the embedding dimension and delay time using the CAO method and mutual information method based on the weighted fractional Fourier pressure signal segments; performing spatial reconstruction based on the embedding dimension and delay time to obtain the reconstructed phase space; calculating the distance based on Mahalanobis distance using the covariance matrix of the actual gas pressure data and the reconstructed phase space; calculating the recursive value based on the distance; and constructing the weighted fractional Fourier recursive graph based on the recursive value.

[0069] In practical applications, such as Figure 4 and Figure 5 The actual gas pressure data is separated into non-overlapping continuous time-domain signal segments of fixed length N to obtain the actual gas pressure data inside the hydrogen tank of a fuel cell vehicle. The pressure data is then separated into non-overlapping continuous time-domain signal segments X = {x1, x2, ... x...} of fixed length N. N-1 ,x N}

[0070] Transform one-dimensional time series data into two-dimensional images.

[0071] For each non-overlapping continuous time-domain signal segment of the pressure signal, a weighted fractional Fourier domain recursion graph based on Mahalanobis distance and a time-domain recursion graph based on Mahalanobis distance are constructed respectively.

[0072] Step 1: As Figure 3 As shown, each non-overlapping continuous time-domain signal segment of the pressure signal is subjected to a weighted fractional Fourier transform to extract the weighted fractional Fourier domain pressure signal segment at the optimal transform order. The specific operation steps are as follows:

[0073] (1) Determine the candidate set W of the weighting function j = {w0(α), w1(α), w2(α), w3(α)}, where j = 1, 2, ..., n; the candidate set of orders N = {N1, N2, ..., N k-1 , N k}, where k = 1, 2, ..., n. W j Let N represent the j-th weight combination. k This represents the k-th order.

[0074] (2) Set the minimum performance metric to positive infinity and initialize the optimal order N. best For the first order N1, the optimal weight combination W best Let W1 be the first weight combination.

[0075] (3) For each order N k In each weight combination W jNext, a weighted fractional Fourier transform is performed to obtain a new dataset X. jk .

[0076] The formula for the four-term weighted fractional Fourier transform is:

[0077]

[0078] Where: parameter α is the order of the fractional Fourier transform, g(x) is the pressure signal segment to be transformed; G(x) = F 1 [g(x)], that is, when the transformation order is 1, the fractional Fourier transform F α It degenerates into a regular Fourier transform; This represents the transformed weighted fractional Fourier domain pressure signal segment; W represents the fractional-order transform, w l The weighting coefficients correspond to the weighted function, with l = 0, 1, 2, 3 representing the 0th, 1st, 2nd, and 3rd Fourier transforms of the original signal, respectively. The weighting coefficients are defined as follows:

[0079]

[0080] (4) Calculate the new dataset X under all combinations of orders and weights. jk The performance metric mean square error (MSE) is taken as the minimum performance metric. min The MSE min The corresponding optimal order N k And the optimal weight combination W j Then the optimal order N best =N k Optimal weight combination W best =W j .

[0081] (5) Output the weighted fractional Fourier domain pressure signal segment at the optimal transformation order.

[0082] Step 2: Based on each weighted Fourier domain pressure signal segment after transformation, select the appropriate embedding dimension m and delay time τ using the CAO method and mutual information method. The specific operation steps are as follows:

[0083] (1) The embedding dimension m is calculated using the nearest neighbor error method. The specific calculation process is as follows:

[0084] Given that the nth vector in the m-dimensional phase space and the (m+1)-dimensional space are respectively L m (n) and L m+ (n), the nearest neighbors of the two in their respective phase spaces are respectively and The Euclidean distance d(n,m) is defined as follows:

[0085]

[0086] If d(n,m) is greater than the determined threshold ε, then point L m (n) is identified as an adjacency error. All points in the space are calculated in the same way, and the value of the embedding dimension m is continuously increased. When the number of adjacency errors gradually decreases and approaches 0, it is the required value of m.

[0087] (2) The delay time τ is calculated using the mutual information method. The specific calculation process is as follows:

[0088]

[0089] Where: S is the mutual information value; P i P represents the probability that the magnitude of the sequence falls within the i-th segment. j P is the probability that the magnitude of the sequence falls within the j-th segment. ij (t) represents two points in a sequence with an interval of Δt, and the joint probability that the magnitude of one point falls in the i-th segment and the magnitude of the other point falls in the j-th segment.

[0090] Based on this, the mutual information value is obtained, and the mutual information curve is obtained. The delay time corresponding to the first local minimum on the curve is the value.

[0091] Step 3: Reconstruct the phase space based on the appropriate embedding dimension m and delay time τ to obtain the reconstructed phase space x. i ={u i ,u i+τ, ,…u i+(m-1)τ}, where i=1,2,…,n-(m-1)τ.

[0092] Where: u i This is the reconstructed stress data segment.

[0093] Step 4: Calculate the covariance matrix C of the pressure data, and then calculate the inverse matrix Cinverse of the covariance matrix. -1 If the covariance matrix is ​​singular, then the pseudo-inverse is calculated through regularization.

[0094] Step 5: Based on the reconstructed phase space, calculate the i-th point x in the reconstructed phase space using Mahalanobis distance. i and point x of j j Distance D ij .

[0095]

[0096] Among them: i=1,2,…,n-(m-1)τ, j=1,2,…,n-(m-1)τ; To reconstruct the phase space from the origin (0,0) to the i-th point xi ; To reconstruct the phase space from the origin (0,0) to the j-th point x j The vector.

[0097] Step 6: Based on the distance D ij Calculate the recursive value R ij (ε).

[0098] R ij (ε)=H(ε-D ij )

[0099] Where: recursive value R ij This forms an N×N two-dimensional texture matrix, D ij For the i-th point x in the reconstructed phase space mentioned above i and point x of j j The distance, ε is the distance threshold such that R ij ∈{0,1}, if D ij If ≤ε, then R ij =1, if D ij If ≥ε, then R ij =0; H represents the Heaviside unit function, in the following form:

[0100]

[0101] Where: r represents ε and D ij The difference.

[0102] Step 7: Based on the recursive values, construct a weighted fractional Fourier domain recursion graph, as follows: Figure 4 As shown.

[0103] Step 8: Similar to steps 2, 3, 4, 5, 6, and 7, based on the non-overlapping continuous time-domain signal segments of each pressure signal, a suitable embedding dimension m and delay time τ are selected using the CAO method and mutual information method, and then the phase space is reconstructed to obtain the reconstructed phase space; based on the reconstructed phase space, the i-th point x in the reconstructed phase space is calculated using Mahalanobis distance. i and point x of j j The distance; based on the distance, calculate the recursive value; based on the recursive value, construct a time-domain recursive graph, such as... Figure 5 As shown.

[0104] Step 103: Perform a lower triangular transformation and fusion on the weighted fractional Fourier recursive graph and the time-domain recursive graph to obtain a fused recursive graph.

[0105] Step 103 specifically includes: performing a lower triangular transformation on the weighted fractional Fourier recursive graph and the time-domain recursive graph to obtain a weighted fractional Fourier domain recursive graph; performing a lower triangular transformation on the time-domain recursive graph to obtain a time-domain lower triangular recursive graph; and fusing the weighted fractional Fourier domain recursive graph and the time-domain lower triangular recursive graph using the Laplace pyramid image fusion method to obtain a fused recursive graph.

[0106] In practical applications, the weighted fractional Fourier domain recursive graph and the time domain recursive graph are respectively converted into a weighted fractional Fourier domain recursive graph lower triangular recursive graph and a time domain lower triangular recursive graph, and a fused recursive graph is obtained by image fusion using Laplace pyramid.

[0107] Step 1: As Figure 6 As shown, based on the property that the recursion graph is symmetrical along the diagonal, the weighted fractional Fourier domain recursion graph and the time domain recursion graph are respectively converted into a weighted fractional Fourier domain recursion graph lower triangular recursion graph and a time domain lower triangular recursion graph.

[0108] Step 2: As Figure 6 As shown, the weighted fractional Fourier domain lower triangular recursive graph and the time domain lower triangular recursive graph are decomposed into Gaussian pyramids to obtain two Gaussian pyramids. For each layer of the Gaussian pyramid, the difference between corresponding layers is calculated, and the Laplacian pyramids of the two are obtained. The Laplacian pyramids of the two are then weighted and averaged with preset weights to obtain a new Laplacian pyramid. Finally, the new Laplacian pyramid is inversely transformed, and the image is reconstructed by upsampling at each level.

[0109] Step 104: A hydrogen leak diagnosis network is used to diagnose the leak based on the fused recursive graph, yielding a hydrogen leak diagnosis result. The hydrogen leak diagnosis network is a trained VGG-16 neural network. Deep texture features are extracted and identified using the VGG-16 neural network based on the fused recursive graph to obtain the hydrogen leak diagnosis result.

[0110] Step 1: Improve the lower triangular fusion recursive graph into a fusion recursive graph image with 224 pixels × 224 pixels × 3 vectors.

[0111] Step 2: As Figure 7As shown, a VGG-16 neural network leakage diagnosis model is constructed. This model contains 16 weighted layers, including convolutional layers, pooling layers, fully connected layers, and a SoftMax classification layer. Specifically, the input is a fused recursive image of 224 pixels × 224 pixels × 3 vectors. In the convolutional layer, a 3×3 filter is used, with 2 or 3 filters stacked consecutively to form a convolutional sequence, mimicking a larger field of view. The stride is 1, and boundary padding is used to maintain the dimensionality of the data before and after convolution. In the pooling layer, a 2×2 pooling window with a stride of 2 is used to reduce the size of the feature image after convolution and to ensure the model's translation invariance. The fully connected layer consists of three consecutive fully connected layers with 4096, 4096, and 1000 channels respectively. Finally, a SoftMax classifier with 1000 labels outputs the leakage type classification.

[0112] Step 2: Train the VGG-16 neural network leak diagnosis model. First, the training sets for normal operation, minor leaks, and major leaks are fused into a recursive graph and used as network input to the VGG-16 neural network leak diagnosis model to complete the training of the leak diagnosis model.

[0113] Step 3: Input the target set fusion recursive graph into the VGG-16 neural network leak diagnosis model. By transferring the parameters of the already trained leak diagnosis model through transfer learning, and then fine-tuning the parameters according to the task set fusion recursive graph, the VGG-16 neural network leak diagnosis model is further trained. During the training process, deep texture features are extracted and identified, and finally the leak diagnosis results are output.

[0114] This invention enables rapid and accurate identification of three scenarios simultaneously: normal operation, minor leakage, and major leakage, thereby increasing the safety and reliability of on-board hydrogen systems.

[0115] The present invention also provides an on-board hydrogen system leak diagnosis system, comprising:

[0116] The acquisition module is used to acquire actual gas pressure data inside the hydrogen tank of a fuel cell vehicle.

[0117] The module is used to construct a weighted fractional Fourier recursive graph and a time-domain recursive graph based on the actual gas pressure data using Mahalanobis distance.

[0118] The lower triangular transformation and fusion module is used to perform lower triangular transformation and fusion on the weighted fractional Fourier recursive graph and the time-domain recursive graph to obtain a fused recursive graph.

[0119] The diagnostic module is used to perform a diagnosis based on the fused recursive graph using a hydrogen leak diagnostic network to obtain a hydrogen leak diagnostic result.

[0120] As an optional implementation, the construction module specifically includes: a separation unit for separating the actual gas pressure data to obtain a fixed-length non-overlapping continuous time-series signal segment; an extraction unit for extracting a weighted fractional Fourier pressure signal segment from the non-overlapping continuous time-series signal segment in each actual gas pressure data using a weighted fractional Fourier transform; a first construction unit for constructing a weighted fractional Fourier recursive graph based on Mahalanobis distance according to the weighted fractional Fourier pressure signal segment; and a second construction unit for constructing a time-domain recursive graph based on Mahalanobis distance according to the non-overlapping continuous time-series signal segment.

[0121] As an optional implementation, the first construction unit specifically includes: a selection subunit, used to select the embedding dimension and delay time based on the weighted fractional Fourier pressure signal segment using the CAO method and the mutual information method; a spatial reconstruction subunit, used to perform spatial reconstruction based on the embedding dimension and delay time to obtain a reconstructed phase space; a distance calculation subunit, used to calculate the distance based on the covariance matrix of the actual gas pressure data and the reconstructed phase space using Mahalanobis distance; a recursion value calculation subunit, used to calculate the recursion value based on the distance; and a recursion graph construction subunit, used to construct a weighted fractional Fourier recursion graph based on the recursion value.

[0122] The present invention also provides an electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method described herein.

[0123] The present invention also provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described thereon.

[0124] This invention has the following advantages:

[0125] 1. Using weighted fractional Fourier transform to process pressure signals can, on the one hand, emphasize specific frequency components in the signal through the weighting mechanism, and on the other hand, extract more significant leakage features and deep frequency domain features, and identify fractional-order micro-leakage signals in vehicle hydrogen systems. This is beneficial for leakage type identification and machine learning.

[0126] 2. Transforming one-dimensional time series into two-dimensional images for machine learning has become a mainstream approach, with key methods including Gram angle field (GAF), Markov transform field (MTF), and recursive graph (RP). However, GAF and MTF are computationally cumbersome, while recursive graph (RP) is not only computationally simple but also an effective visualization tool for multi-scale feature extraction of time series, capturing hidden dynamics and deep texture features.

[0127] 3. This invention improves the recurrence relation (RP) plotting based on Mahalanobis distance. First, since leakage causes pressure data to have a different covariance structure than data under normal operation, Mahalanobis distance can capture state patterns that are significantly different from the normal state, accurately identifying abnormal pressure changes caused by leakage. Second, because Mahalanobis distance is adaptive and can automatically adjust the covariance matrix, it can provide better adaptability for on-board hydrogen systems under different operating conditions.

[0128] 4. The Laplacian pyramid image fusion technique was used to fuse the temporal recurrent graph and the weighted fractional Fourier domain recurrent graph of the pressure data. On the one hand, by fusing temporal and weighted fractional Fourier domain features, the features of the recurrent graph were enhanced, enabling the neural network to analyze the data under multi-scale information and capture weak information about leakage events. On the other hand, due to the integration of multi-scale information, it is more sensitive to small changes in the data, amplifying the weak pressure disturbances caused by minor leaks, thereby increasing robustness, facilitating subsequent feature extraction and classification, and improving diagnostic accuracy.

[0129] 5. Convolutional Neural Networks (CNNs) are multi-layered, non-fully connected neural networks that simulate the structure of the human brain. Through supervised deep learning, they can directly identify visual patterns from raw images, making them excellent tools for fault diagnosis. The VGG-16 CNN used in this invention is characterized by its ability to extract more minute features from the input domain through the combination and stacking of 3×3 filters. Therefore, using the VGG-16 CNN to diagnose leak types from collected pressure data in vehicle-mounted hydrogen cylinders can not only identify a large number of leaks but also more specifically diagnose minute leak types, thereby improving diagnostic accuracy.

[0130] 6. Transfer learning is used to transfer the parameters of convolutional and pooling layers in the training model to the leak diagnosis model. This eliminates the need for retraining of the convolutional layers in the leak diagnosis model. By determining the parameters of the convolutional layers, the training speed is improved, and the problems caused by insufficient data or excessive training time can be effectively compensated for.

[0131] 7. Figure 8 The accuracy is obtained after 100 rounds of training using the VGG-16 neural network leak diagnosis model with pressure data input. It can be seen that by the 80th training iteration, the accuracy of both the training and validation sets reached 100%, and the accuracy curves thereafter gradually flattened out, indicating that the model had reached a stable convergence state. This demonstrates that the VGG-16 leak diagnosis model of this invention achieved a satisfactory leak diagnosis effect on this pressure data.

[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0133] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for diagnosing leaks in an on-board hydrogen system, characterized in that, include: Obtain actual gas pressure data inside the hydrogen tank of a fuel cell vehicle; Based on Mahalanobis distance, a weighted fractional Fourier recursive graph and a time-domain recursive graph are constructed from the actual gas pressure data. Specifically, this includes: separating the actual gas pressure data to obtain fixed-length non-overlapping continuous time-series signal segments; extracting weighted fractional Fourier pressure signal segments from the non-overlapping continuous time-series signal segments in each actual gas pressure data set using weighted fractional Fourier transform; constructing a weighted fractional Fourier recursive graph based on the weighted fractional Fourier pressure signal segments using Mahalanobis distance, specifically including: selecting the embedding dimension and delay time using the CAO method and mutual information method based on the weighted fractional Fourier pressure signal segments; performing spatial reconstruction based on the embedding dimension and delay time to obtain a reconstructed phase space; calculating the distance based on the covariance matrix of the actual gas pressure data and the reconstructed phase space using Mahalanobis distance; calculating the recursive value based on the distance; constructing a weighted fractional Fourier recursive graph based on the recursive value; and constructing a time-domain recursive graph based on the non-overlapping continuous time-series signal segments using Mahalanobis distance. The weighted fractional Fourier recursive graph and the time-domain recursive graph are subjected to a lower triangular transformation and fused to obtain a fused recursive graph; Based on the fusion recursive graph, a hydrogen leak diagnosis network is used to perform diagnosis, and the hydrogen leak diagnosis results are obtained.

2. The method for diagnosing leaks in an on-board hydrogen system according to claim 1, characterized in that, The weighted fractional Fourier recursive graph and the time-domain recursive graph are subjected to a lower triangular transformation and fused to obtain a fused recursive graph, specifically including: The weighted fractional Fourier recursive graph and the time-domain recursive graph are transformed by a lower triangular transformation to obtain the weighted fractional Fourier domain recursive graph. The time-domain recursive graph is transformed into a lower triangular graph to obtain a time-domain lower triangular recursive graph. The weighted fractional Fourier domain recursive graph and the time-domain lower triangular recursive graph are fused using the Laplace pyramid image fusion method to obtain a fused recursive graph.

3. The method for diagnosing leaks in an on-board hydrogen system according to claim 1, characterized in that, The hydrogen leak diagnosis network is a trained VGG-16 neural network.

4. A vehicle-mounted hydrogen system leak diagnosis system, characterized in that, include: The acquisition module is used to acquire actual gas pressure data inside the hydrogen tank of a fuel cell vehicle; A construction module is used to construct a weighted fractional Fourier recursive graph and a time-domain recursive graph based on the actual gas pressure data using Mahalanobis distance. The construction module specifically includes: a separation unit for separating the actual gas pressure data to obtain fixed-length non-overlapping continuous time-series signal segments; an extraction unit for extracting weighted fractional Fourier pressure signal segments from the non-overlapping continuous time-series signal segments in each actual gas pressure data set using weighted fractional Fourier transform; a first construction unit for constructing a weighted fractional Fourier recursive graph based on the weighted fractional Fourier pressure signal segments using Mahalanobis distance; and a second construction unit for constructing a weighted fractional Fourier recursive graph based on the Mahalanobis distance and the... The method describes constructing a time-domain recursive graph from non-overlapping continuous time-series signal segments. The first construction unit specifically includes: a selection subunit, used to select the embedding dimension and delay time based on the weighted fractional Fourier pressure signal segments using the CAO method and mutual information method; a spatial reconstruction subunit, used to perform spatial reconstruction based on the embedding dimension and delay time to obtain a reconstructed phase space; a distance calculation subunit, used to calculate the distance based on the covariance matrix of the actual gas pressure data and the reconstructed phase space using Mahalanobis distance; a recursive value calculation subunit, used to calculate the recursive value based on the distance; and a recursive graph construction subunit, used to construct a weighted fractional Fourier recursive graph based on the recursive value. The lower triangular transformation and fusion module is used to perform lower triangular transformation and fusion on the weighted fractional Fourier recursive graph and the time-domain recursive graph to obtain a fused recursive graph; The diagnostic module is used to perform diagnosis based on the fused recursive graph using a hydrogen leak diagnostic network to obtain hydrogen leak diagnostic results.

5. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 3.

6. A computer storage medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 3.

Images