Multi-stage information fusion hydrogen leakage integrated learning diagnosis method, system and device

Abstract

The application discloses a kind of multi-stage information fusion hydrogen leakage integrated learning diagnosis method, system and equipment, it is related to hydrogen gas leakage diagnosis technical field, it is aimed at by threshold monitoring method sensor abnormal signal monitoring, and by Gaussian regression process realizes sensor abnormal signal recovery, provides reliable data basis for hydrogen leakage diagnosis, further obtains static characteristic information, dynamic characteristic information and recursion chart, and carries out feature layer information fusion, then the feature fusion result is input into the IGANet deep learning neural network model trained to obtain one-dimensional diagnostic result;On the other hand, obtain first derivative characteristic and amplitude spectrum data characteristic, in conjunction with support vector machine to obtain another one-dimensional diagnostic result;Combination above two-dimensional diagnostic result, by integrated learning diagnosis model, carry out decision layer information fusion, and further obtain multi-stage information fusion hydrogen leakage diagnostic result, to improve the reliability of diagnosis.

Description

Technical Field

[0001] This invention relates to the field of hydrogen leak diagnosis technology, and in particular to a multi-stage information fusion integrated learning diagnosis method, system and equipment for hydrogen leaks. Background Technology

[0002] As a pioneer in hydrogen energy applications, fuel cell vehicles have gained widespread attention due to their advantages such as high energy efficiency, zero pollution, good adaptability, and stable operation. With the deepening research and industrial application of fuel cell vehicles, higher requirements are being placed on their driving range. Under the same volume, higher pressure results in a higher hydrogen storage capacity and a longer driving range. Therefore, high-pressure gaseous hydrogen storage has become the mainstream method for on-board hydrogen storage systems, which stores hydrogen gas in containers through high-pressure compression. However, high pressure has become a core risk factor affecting the safety of on-board hydrogen storage systems, and these systems are complex in structure, operate under harsh conditions, and are prone to leakage failures.

[0003] Hydrogen is the least dense gas, possessing strong buoyancy and diffusivity. It is also highly flammable and explosive, characterized by rapid combustion and a wide explosion limit range. Therefore, it is necessary to utilize data technology to monitor hydrogen leaks in fuel cell vehicles, proactively identifying anomalies in the onboard hydrogen storage system to ensure its safe and reliable operation. However, commonly used hydrogen concentration sensors for leak detection are limited by factors such as the number of sensors, their installation location, and structural layout, making efficient diagnosis of hydrogen leak faults difficult. To address these issues, researchers have used hydrogen cylinder pressure sensor signals and machine learning methods to diagnose hydrogen leaks. However, existing methods fail to adequately consider data preprocessing, feature extraction is complex, and diagnostic reliability is poor. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-stage information fusion hydrogen leakage integrated learning diagnostic method, system, and device to obtain multi-stage information fusion hydrogen leakage diagnostic results and improve diagnostic reliability.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] In a first aspect, the present invention provides a multi-stage information fusion integrated learning diagnostic method for hydrogen leakage, comprising:

[0007] An abnormal signal threshold detection algorithm is used to detect the raw gas pressure data, and Gaussian regression is used to obtain the normal prediction data corresponding to the detected abnormal gas pressure data. The raw gas pressure data is the actual gas pressure data in the hydrogen cylinder of the on-board hydrogen storage system of a fuel cell vehicle, which is collected by a pressure sensor.

[0008] The processed gas pressure data are subjected to Gram angle field transformation, Markov transfer field transformation, and recursive graph phase space trajectory image feature extraction to obtain static feature information, dynamic feature information, and recursive graph. The static feature information, the dynamic feature information, and the recursive graph are then fused to obtain the feature fusion result. The processed gas pressure data includes normal gas pressure data from the original gas pressure data and normal predicted data corresponding to the abnormal gas pressure data.

[0009] The feature fusion result is input into the trained IGANet deep learning neural network model to obtain the first hydrogen leak diagnosis result;

[0010] Traditional feature extraction is performed on the processed gas pressure data to obtain traditional feature information. Then, using the traditional feature information and the trained support vector machine model, a second hydrogen leak diagnosis result is obtained.

[0011] By integrating the first hydrogen leak diagnosis result and the second hydrogen leak diagnosis result using the integrated learning diagnostic model, the final hydrogen leak diagnosis result is determined.

[0012] Secondly, the present invention provides a multi-stage information fusion hydrogen leakage integrated learning diagnostic system, comprising:

[0013] The abnormal data processing module is used to detect the raw gas pressure data using a sensor abnormal signal threshold detection algorithm, and to obtain the normal prediction data corresponding to the detected abnormal gas pressure data using a Gaussian regression process; the raw gas pressure data is the actual gas pressure data in the hydrogen cylinder of the on-board hydrogen storage system of the fuel cell vehicle collected by a pressure sensor.

[0014] The feature fusion result determination module is used to perform Gram angle field transformation, Markov transfer field transformation, and recursive graph phase space trajectory image feature extraction on the processed gas pressure data to obtain static feature information, dynamic feature information, and a recursive graph. The module then fuses the static feature information, the dynamic feature information, and the recursive graph to obtain the feature fusion result. The processed gas pressure data includes normal gas pressure data from the original gas pressure data and normal predicted data corresponding to the abnormal gas pressure data.

[0015] The first hydrogen leak diagnosis result determination module is used to input the feature fusion result into the trained IGANet deep learning neural network model to obtain the first hydrogen leak diagnosis result.

[0016] The second hydrogen leak diagnosis result determination module is used to perform traditional feature extraction operations on the processed gas pressure data to obtain traditional feature information, and use the traditional feature information and the trained support vector machine model to obtain the second hydrogen leak diagnosis result.

[0017] The final hydrogen leak diagnosis result determination module is used to determine the final hydrogen leak diagnosis result by integrating the first hydrogen leak diagnosis result and the second hydrogen leak diagnosis result through an integrated learning diagnosis model.

[0018] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the multi-stage information fusion hydrogen leakage integrated learning diagnostic method described in the first aspect.

[0019] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0020] 1. An abnormal signal detection method is proposed for the pressure sensor signals collected in the vehicle-mounted hydrogen storage system. Furthermore, an abnormal signal recovery method is proposed to provide a reliable data basis for subsequent hydrogen leakage fault diagnosis and improve the diagnostic accuracy.

[0021] 2. A multi-stage information fusion technique is proposed. First, during the image feature extraction process, feature layer information is weighted and fused to obtain the image feature extraction results. Then, a decision layer information weighted fusion method is used to comprehensively consider the diagnostic results of the proposed IGANet deep learning model and the diagnostic results of the traditional machine learning support vector machine to establish an ensemble learning hydrogen leak diagnosis model, thereby improving the stability, adaptability, and accuracy of hydrogen leak diagnosis.

[0022] 3. It can realize online diagnosis of hydrogen leakage in the on-board hydrogen storage system of fuel cell vehicles, and the diagnosis cost is low and the efficiency is high. Attached Figure Description

[0023] 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.

[0024] Figure 1 A flowchart illustrating the multi-stage information fusion hydrogen leakage integrated learning diagnostic method provided in this embodiment of the invention;

[0025] Figure 2 This is a schematic diagram of the overall process of the multi-stage information fusion hydrogen leakage integrated learning diagnosis method provided in the embodiments of the present invention;

[0026] Figure 3 This is a graph showing the normal operating pressure and gas leakage pressure data of the hydrogen cylinder in the vehicle-mounted hydrogen storage system provided in this embodiment of the invention.

[0027] Figure 4 This is a graph showing the gas leakage pressure data and the normal operating gas pressure data after signal recovery, provided in an embodiment of the present invention.

[0028] Figure 5 This is a diagram illustrating the adaptive feature extraction process of a Gram angle field static image provided in an embodiment of the present invention.

[0029] Figure 6 This is a diagram illustrating the dynamic image feature extraction process of the Markov transition matrix provided in an embodiment of the present invention.

[0030] Figure 7 This is a diagram illustrating the recursive phase space trajectory image feature extraction process provided in an embodiment of the present invention.

[0031] Figure 8 This is a diagram illustrating the feature layer information fusion process provided in an embodiment of the present invention.

[0032] Figure 9 This is a graph showing the changes in accuracy and loss during the training process of the IGANet deep learning neural network model provided in an embodiment of the present invention.

[0033] Figure 10 The extracted features provided in this embodiment of the invention are normalized and then plotted in a two-dimensional coordinate system based on the labels.

[0034] Figure 11 This is a support vector machine diagram for diagnosing hydrogen leakage, provided in an embodiment of the present invention. Detailed Implementation

[0035] 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.

[0036] 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.

[0037] Example 1

[0038] like Figure 1 and Figure 2As shown in the figure, this embodiment provides a multi-stage information fusion integrated learning diagnostic method for hydrogen leakage based on abnormal signal detection and recovery, including the following steps.

[0039] Step 100: The original gas pressure data is detected using a sensor abnormal signal threshold detection algorithm, and the normal prediction data corresponding to the detected abnormal gas pressure data is obtained by using a Gaussian regression process; the original gas pressure data is the actual gas pressure data in the hydrogen cylinder of the fuel cell vehicle on-board hydrogen storage system collected by a pressure sensor.

[0040] Step 200: Perform Gram angle field transformation, Markov transfer field transformation, and recursive graph phase space trajectory image feature extraction on the processed gas pressure data to obtain static feature information, dynamic feature information, and recursive graph. Then, fuse the static feature information, the dynamic feature information, and the recursive graph to obtain the feature fusion result. The processed gas pressure data includes normal gas pressure data in the original gas pressure data and normal predicted data corresponding to the abnormal gas pressure data.

[0041] Step 300: Input the feature fusion result into the trained IGANet deep learning neural network model to obtain the first hydrogen leak diagnosis result. The IGANet deep learning neural network model is a deep learning model that uses global average pooling layers to replace the three fully connected layers in the AlexNet neural network model.

[0042] Step 400: Perform traditional feature extraction on the processed gas pressure data to obtain traditional feature information, and use the traditional feature information and the trained support vector machine model to obtain the second hydrogen leak diagnosis result.

[0043] Step 500: By using an integrated learning diagnostic model, the first hydrogen leak diagnostic result and the second hydrogen leak diagnostic result are combined to determine the final hydrogen leak diagnostic result.

[0044] In one preferred embodiment, the gas pressure data includes pressure values ​​and corresponding time points; step 100 includes the following.

[0045] First, the absolute value of the pressure difference between adjacent time nodes is calculated sequentially, and it is determined whether the absolute value is greater than an empirical threshold. The adjacent time nodes include a first time node and a second time node following the first time node. If so, the adjacent time nodes corresponding to the absolute values ​​are marked in chronological order. When the second time node in the previously marked adjacent time nodes is the abnormal start point, the first time node in the currently marked adjacent time nodes is determined to be the abnormal end point. When the first time node in the previously marked adjacent time nodes is the abnormal end point, the second time node in the currently marked adjacent time nodes is determined to be the abnormal start point. Next, the gas pressure data between adjacent abnormal start points and abnormal end points is determined as abnormal gas pressure data. Then, based on the original gas pressure data and the abnormal gas pressure data, normal gas pressure data is determined. Finally, based on the normal gas pressure data, Gaussian regression is used to obtain the normal prediction data corresponding to the abnormal gas pressure data.

[0046] One example is that when using pressure sensor signals for hydrogen leak fault diagnosis, the quality of the pressure sensor signal has a significant impact on the feature extraction effect and even the final fault diagnosis result. To improve the reliability and accuracy of leak diagnosis for on-board hydrogen storage systems in fuel cell vehicles, a sensor abnormal signal threshold detection algorithm is used to preprocess the raw gas pressure data. Furthermore, for the detected abnormal gas pressure data, a Gaussian regression process is used to obtain its normal prediction data, providing a reliable data foundation for hydrogen leak diagnosis.

[0047] The first step is to obtain the normal operating pressure data and gas leakage pressure data of the hydrogen cylinder in the on-board hydrogen storage system of the fuel cell vehicle, P = [P0, P1, P2, P3...P...]. 60 ],like Figure 3 As shown, the data acquisition time span was 60 seconds, and the initial hydrogen cylinder pressure was 70 MPa.

[0048] The second step involves identifying the time points of sensor anomaly signals by using a pressure difference and threshold filtering method, specifically a sensor anomaly signal threshold detection algorithm, when a leak occurs. The specific method is as follows:

[0049]

[0050] Among them, P i P represents the pressure data corresponding to the i-th time node. i-1 This represents the pressure data corresponding to the (i-1)th time node. To determine the threshold based on experience, the [i1, i2] obtained using the above method is used, which indicates when the sensor signal becomes abnormal during the time period from time node i1 to time node i2.

[0051] The third step involves using the normal sensor signal data after anomaly detection to predict the aforementioned abnormal sensor signals using a Gaussian regression process, specifically by utilizing {1, 2, 3...i1] and [i2, i2+1...60], The input is a Gaussian regression process. After fitting, the time interval [i1, i1+1, ... i2] of the abnormal sensor signal data is used as input to obtain the non-abnormal predicted value of the sensor signal, thereby realizing the recovery of the abnormal sensor signal. Figure 4 As shown.

[0052] In one preferred embodiment, steps 200 and 300 include the following.

[0053] For pressure data recovered from abnormal sensor signals, extracting representative features and designing suitable algorithms to efficiently, accurately, and stably diagnose hydrogen leaks remains a challenge. This embodiment proposes an adaptive image feature extraction method to obtain dynamic image features of the Markov transition matrix, static image features of the Gram angle field, and recursive graph image features. Furthermore, a feature layer information weighted fusion method is used to achieve multi-dimensional image feature fusion. Using these fused image features as input, the proposed IGANet neural network is trained to obtain hydrogen leak fault diagnosis results.

[0054] The first step is to convert the sensor anomaly recovery signal (i.e., the processed gas pressure data) to the Gram angle field (e.g., ...). Figure 5 As shown, static image features are obtained.

[0055] First, normalization is applied to the sensor anomaly recovery signal X = [x1, x2, x3, ..., x...]. N Mapped to the range [0, 1]:

[0056]

[0057] Where X represents the sensor anomaly recovery signal sequence, max(X) and min(X) represent the maximum and minimum signal values, respectively, x j This represents the j-th signal collected. This represents the normalized data of the sensor's abnormal recovery signal.

[0058] Secondly, the normalized data is transformed to polar coordinates:

[0059]

[0060] Where r represents the radius coordinate. t represents angular coordinates, polar angle, or azimuth. iThis represents the signal acquisition time point, and N represents the polar coordinate transformation factor.

[0061] Then, the Gram angle and field are calculated, which can characterize the temporal correlation at different time intervals while preserving the time dependence, thus representing the static feature information of the data:

[0062]

[0063] Where I represents the unit row vector [1,1,...,1], and ~ represents the normalization operation. express The transpose of .

[0064] The second step is to convert the sensor's abnormal recovery signal to a Markov transfer field (such as...). Figure 6 (As shown). First, the sensor abnormal recovery signal is divided into bins, with the number of bins set to Q. Then, each acquired signal value will be mapped to the corresponding bin q. i For i ∈ [1, Q], the quantiles are transformed into a Q×Q adjacency weighted matrix, i.e., the Markov transition matrix W, through a first-order Markov chain along the time axis:

[0065]

[0066] Among them, w ij Indicates the number of boxes q i Transfer to sub-box q j The probability, and

[0067] Since the sensor anomaly recovery signal is distributed along the time axis, the state transition matrix is ​​given additional time information, thus enabling the construction of the Markov transition field M:

[0068]

[0069] Where, m ij This indicates that the time period k (1≤k≤N) belongs to bin q. i Each data point is transferred to a time period l (1≤l≤N) belonging to bin q. j The transition probability of data points.

[0070] The Markov transition field is determined by calculating the Markov probability transition matrix, and the dynamic transfer information of the data is represented by a two-dimensional image.

[0071] The third step is to convert the sensor anomaly recovery signal into a recursive graph (such as...). Figure 7 (As shown). The recursion graph can represent the phase space trajectory of the original data. The numerical calculation expression is as follows:

[0072]

[0073] In the formula, Let represent the point of the system at time i, ε represent the threshold, and Θ(·) represent the Heaviside function. Furthermore, for a thresholdless recursive graph RP, the matrix... From vector and The distance is calculated, which can reconstruct the time series and contain rich original data information. In this embodiment, thresholdless RP is used to obtain the recursive graph.

[0074] The fourth step is the fusion of static and dynamic image information at the feature layer. Utilizing the aforementioned static features of the Gram angle field, dynamic features of the Markov transition matrix, and recursive graph phase space trajectory features, combined with the feature layer information fusion method, the feature fusion result is obtained according to the following formula:

[0075] F if = aG + bM + cR;

[0076] In the formula, G, M, and R represent GASF, MTF, and RP features, respectively, and a, b, and c represent the weighting coefficients of the three features, which are 0.1, 1, and 10 in this case. The final feature layer information fusion result F if Includes all static correlations, dynamic transits, and phase space trajectory features of the original data, such as Figure 8 As shown.

[0077] The fifth step involves building and training the IGANet deep learning neural network model. The proposed IGANet model uses Global Average Pooling (GAP) instead of the traditional three fully connected layers in AlexNet, effectively mitigating the overfitting problem caused by an excessive number of models, reducing training parameters, improving training efficiency, and acquiring global features. Then, a fully connected layer (64×2) is used to obtain the target output size, and a SoftMax layer is combined to output the diagnostic results. Specific structural parameters are shown in Table 1. Finally, an improved deep learning network model, Improved Global Average AlexNet (IGANet), is proposed, where the pooling process expression for GAP is as follows:

[0078]

[0079] Using the feature layer information fusion result as input, the model is trained and hydrogen leak fault diagnosis is performed, yielding the fault diagnosis result X1 for the input pressure data, such as... Figure 9 As shown.

[0080] Table 1. IGANet Deep Learning Neural Network Model Construction Table

[0081]

[0082] In one preferred embodiment, step 400 includes the following.

[0083] First, the first derivative of the processed gas pressure data is calculated to obtain the first derivative feature. Second, the first derivative is subjected to a Fourier transform to obtain the amplitude spectrum data feature. Finally, the first derivative feature, amplitude spectrum data feature, and support vector machine algorithm are used to obtain the second hydrogen leak diagnosis result.

[0084] One example is the use of 2D image feature layer information fusion and deep learning methods, which exhibit good adaptive diagnostic performance. However, its adaptability to different pressure data conditions is poor. Traditional feature extraction methods such as signal transformation and time-frequency domain analysis, combined with machine learning methods like Bayesian networks and K-means clustering, offer good universality for leak diagnosis, but the process is cumbersome. Therefore, it is necessary to combine the advantages of both methods by using deep learning combined with traditional feature extraction and supervised learning methods for decision-level information fusion, constructing an ensemble learning model to obtain the final diagnostic result. Since the pressure change rate during a leak is higher than that during normal operation, the first derivative of the sensor pressure signal is used as one of the extracted features. Furthermore, Fourier transform is used to obtain the frequency domain expression of the pressure derivative and calculate the amplitude spectrum, which is then used as another extracted feature. Using these extracted features as input, a support vector machine algorithm is used for supervised learning to further diagnose hydrogen leak faults.

[0085] The first step is to acquire the characteristic data of the sensor pressure signal. For the sensor pressure signal data, obtain its first derivative and calculate the mean Θ:

[0086]

[0087] Where P i ,P i-1 For pressure signal data, Δt = 1, the characteristics of multiple sets of collected data Θ = [Θ1, Θ2, Θ3...Θ] can be obtained. N Furthermore, the frequency domain expression of the above first derivative is obtained by applying a Fourier transform:

[0088]

[0089] Where f(t) is the time-domain signal, F(ω) is the frequency-domain signal, e^(-iωt) is the integral transform kernel, dt is the integration over time, and F[] represents the Fourier transform. Next, its mean amplitude spectrum δ is extracted to obtain another feature of multiple sets of data: δ=[δ1,δ2,δ3...δ N ].

[0090] The extracted features are further processed using normalization methods:

[0091]

[0092] Where, m p m represents the normalized data. q For the original data, m qmin and m qmax These represent the minimum and maximum values ​​in the original data, respectively.

[0093] The normalized sets of feature data and leakage status labels are plotted in a two-dimensional coordinate system, as follows: Figure 10 As shown.

[0094] The second step involves supervised learning using the Support Vector Machine (SVM) machine learning algorithm on the aforementioned labeled data to obtain a trained SVM model. A multinomial kernel function is set for the SVM model, and the classification and decision boundaries for the sample are obtained through training. Ultimately, the accuracy (X2) of the SVM model in diagnosing hydrogen leak faults can be obtained. The training results are shown below. Figure 11 As shown, the dashed line represents the classification boundary, and the solid line represents the decision boundary.

[0095] In one preferred embodiment, step 500 includes the following.

[0096] An ensemble learning diagnostic model is established using a decision-level information weighted fusion method. Based on the ensemble learning diagnostic model, the first hydrogen leak diagnosis result and the second hydrogen leak diagnosis result are fused to determine the final hydrogen leak diagnosis result.

[0097] The process of establishing the ensemble learning diagnostic model is as follows: based on the training accuracy of the IGANet deep learning neural network model and the training accuracy of the support vector machine model, the ensemble learning diagnostic model is established using the decision layer information weighted fusion method.

[0098] One example is: using image feature extraction and deep learning methods versus manual feature extraction and traditional machine learning, two methods with different diagnostic properties, to obtain diagnostic results X1 respectively. * X2 * Furthermore, an ensemble learning diagnostic model is established using a weighted fusion method of decision-level information, combining the advantages of both diagnostic methods to obtain the final fault diagnosis result. Ensemble learning is a machine learning method that combines multiple weak learners (models) to construct a strong learner.

[0099] A weighted strategy is used to fuse information at the decision-making level to obtain the final ensemble learning diagnostic result, i.e., the final diagnostic result of the hydrogen leak:

[0100] Y = mX1 * +nX2 * ;

[0101] In the formula, X1 * X2 * For real-time diagnostic results, m and n are weighted fusion coefficients of decision-making level information, calculated as follows:

[0102]

[0103] In the formula, X1 and X2 are the training accuracy of the deep learning model IGANet and the training accuracy of the machine learning model SVM, respectively.

[0104] Example 2

[0105] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, a multi-stage information fusion hydrogen leakage integrated learning diagnostic system is provided below.

[0106] This embodiment provides a multi-stage information fusion hydrogen leakage integrated learning diagnostic system, including:

[0107] The abnormal data processing module is used to detect the raw gas pressure data using a sensor abnormal signal threshold detection algorithm, and to obtain the normal prediction data corresponding to the detected abnormal gas pressure data using a Gaussian regression process; the raw gas pressure data is the actual gas pressure data in the hydrogen cylinder of the fuel cell vehicle's on-board hydrogen storage system, which is collected by a pressure sensor.

[0108] The feature fusion result determination module is used to perform Gram angle field transformation, Markov transfer field transformation, and recursive graph phase space trajectory image feature extraction on the processed gas pressure data to obtain static feature information, dynamic feature information, and recursive graph. The static feature information, the dynamic feature information, and the recursive graph are then fused to obtain the feature fusion result. The processed gas pressure data includes normal gas pressure data from the original gas pressure data and normal prediction data corresponding to the abnormal gas pressure data.

[0109] The first hydrogen leak diagnosis result determination module is used to input the feature fusion result into the trained IGANet deep learning neural network model to obtain the first hydrogen leak diagnosis result.

[0110] The second hydrogen leak diagnosis result determination module is used to perform traditional feature extraction operations on the processed gas pressure data to obtain traditional feature information, and use the traditional feature information and the trained support vector machine model to obtain the second hydrogen leak diagnosis result.

[0111] The final hydrogen leak diagnosis result determination module is used to determine the final hydrogen leak diagnosis result by integrating the first hydrogen leak diagnosis result and the second hydrogen leak diagnosis result through an integrated learning diagnosis model.

[0112] Example 3

[0113] This invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to execute a multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to Embodiment 1.

[0114] Alternatively, the aforementioned electronic device may be a server.

[0115] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements a multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to Embodiment 1.

[0116] 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.

[0117] 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 multi-stage information fusion integrated learning diagnostic method for hydrogen leakage, characterized in that, include: An abnormal signal threshold detection algorithm is used to detect the raw gas pressure data, and Gaussian regression is used to obtain the normal prediction data corresponding to the detected abnormal gas pressure data. The raw gas pressure data is the actual gas pressure data in the hydrogen cylinder of the on-board hydrogen storage system of a fuel cell vehicle, which is collected by a pressure sensor. The processed gas pressure data are subjected to Gram angle field transformation, Markov transfer field transformation, and recursive graph phase space trajectory image feature extraction to obtain static feature information, dynamic feature information, and recursive graph. The static feature information, the dynamic feature information, and the recursive graph are then fused to obtain the feature fusion result. The processed gas pressure data includes normal gas pressure data from the original gas pressure data and normal predicted data corresponding to the abnormal gas pressure data. The feature fusion result is input into the trained IGANet deep learning neural network model to obtain the first hydrogen leak diagnosis result; Traditional feature extraction is performed on the processed gas pressure data to obtain traditional feature information. Then, using the traditional feature information and the trained support vector machine model, a second hydrogen leak diagnosis result is obtained. By integrating the first hydrogen leak diagnosis result and the second hydrogen leak diagnosis result using the integrated learning diagnostic model, the final hydrogen leak diagnosis result is determined.

2. The multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to claim 1, characterized in that, Gas pressure data includes pressure values ​​and corresponding time points; The process involves employing a sensor anomaly signal threshold detection algorithm to detect the raw gas pressure data, and then using a Gaussian regression process to obtain the corresponding normal prediction data for the detected abnormal gas pressure data. Specifically, this includes: The absolute value of the difference between pressure values ​​corresponding to adjacent time nodes is calculated sequentially, and it is determined whether the absolute value is greater than an empirical judgment threshold; the adjacent time nodes include a first time node and a second time node located after the first time node; If so, the adjacent time nodes corresponding to the absolute value are marked in chronological order. When the second time node in the previously marked adjacent time node is the abnormal start point, the first time node in the currently marked adjacent time node is determined to be the abnormal end point. When the first time node in the previously marked adjacent time node is the abnormal end point, the second time node in the currently marked adjacent time node is determined to be the abnormal start point. The gas pressure data between adjacent anomaly start points and anomaly termination points are defined as abnormal gas pressure data. Based on the original gas pressure data and the abnormal gas pressure data, determine the normal gas pressure data; Based on the normal gas pressure data, normal prediction data corresponding to the abnormal gas pressure data is obtained using a Gaussian regression process.

3. The multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to claim 1, characterized in that, The static feature information, the dynamic feature information, and the recursive graph are fused to obtain the feature fusion result, which specifically includes: The feature layer information fusion method is used to fuse the static feature information, the dynamic feature information, and the recursive graph to obtain the feature fusion result.

4. The multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to claim 1, characterized in that, Traditional feature extraction is performed on the processed gas pressure data to obtain traditional feature information. Then, using this traditional feature information and a trained support vector machine model, a second hydrogen leak diagnosis result is obtained, specifically including: The first derivative of the processed gas pressure data is obtained by performing a first-order derivative operation; The amplitude spectrum data characteristics are obtained by performing a Fourier transform on the first derivative; The second hydrogen leak diagnosis result was obtained by using the first derivative features, amplitude spectrum data features, and a trained support vector machine model.

5. The multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to claim 1, characterized in that, By using an ensemble learning diagnostic model, the first hydrogen leak diagnostic result and the second hydrogen leak diagnostic result are combined to determine the final hydrogen leak diagnostic result, specifically including: An ensemble learning diagnostic model is established using a decision-level information weighted fusion method. Based on the ensemble learning diagnostic model, the first hydrogen leak diagnosis result and the second hydrogen leak diagnosis result are fused to determine the final hydrogen leak diagnosis result.

6. The multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to claim 5, characterized in that, An ensemble learning diagnostic model is established using a weighted fusion method of decision-level information, specifically including: Based on the training accuracy of the IGANet deep learning neural network model and the support vector machine model, an ensemble learning diagnostic model is established using a decision-level information weighted fusion method.

7. A multi-stage information fusion hydrogen leakage integrated learning diagnostic system, characterized in that, include: The abnormal data processing module is used to detect the raw gas pressure data using a sensor abnormal signal threshold detection algorithm, and to obtain the normal prediction data corresponding to the detected abnormal gas pressure data using a Gaussian regression process; the raw gas pressure data is the actual gas pressure data in the hydrogen cylinder of the on-board hydrogen storage system of the fuel cell vehicle collected by a pressure sensor. The feature fusion result determination module is used to perform Gram angle field transformation, Markov transfer field transformation, and recursive graph phase space trajectory image feature extraction on the processed gas pressure data to obtain static feature information, dynamic feature information, and a recursive graph. The module then fuses the static feature information, the dynamic feature information, and the recursive graph to obtain the feature fusion result. The processed gas pressure data includes normal gas pressure data from the original gas pressure data and normal predicted data corresponding to the abnormal gas pressure data. The first hydrogen leak diagnosis result determination module is used to input the feature fusion result into the trained IGANet deep learning neural network model to obtain the first hydrogen leak diagnosis result. The second hydrogen leak diagnosis result determination module is used to perform traditional feature extraction operations on the processed gas pressure data to obtain traditional feature information, and use the traditional feature information and the trained support vector machine model to obtain the second hydrogen leak diagnosis result. The final hydrogen leak diagnosis result determination module is used to determine the final hydrogen leak diagnosis result by integrating the first hydrogen leak diagnosis result and the second hydrogen leak diagnosis result through an integrated learning diagnosis model.

8. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform a multi-stage information fusion hydrogen leakage integrated learning diagnostic method according to any one of claims 1 to 6.

Images