Hydrogen leakage risk assessment model based on artificial intelligence

Abstract

The invention relates to the technical field of risk assessment, and particularly discloses an artificial intelligence-based hydrogen leakage risk assessment model, which comprises a dynamic monitoring data module, a risk feature extraction module, a leakage rate prediction and source point inversion module and a leakage evolution module, the dynamic monitoring data module collects real-time parameters and preprocesses the real-time parameters; the risk feature extraction module introduces a 1D-CNN and CNN-LSTM fusion neural network to generate a leakage risk factor sequence; the leakage rate prediction and source point inversion module predicts the leakage rate by using a sliding window and a dynamic correction mechanism, and inverts the position of a hydrogen leakage source in combination with a Gaussian plume model and sensor data; and the leakage evolution module displays a high-risk area heat map and a leakage source position. According to the method, the multi-source sensing data and the physical model are fused, and the defect that the leakage risk cannot be dynamically predicted through a traditional method is overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment, and more specifically, to a hydrogen leakage risk assessment model based on artificial intelligence. Background Art

[0002] Hydrogen is mainly used for cooling the stator coils of generators, which can effectively improve the cooling efficiency and reduce energy consumption. However, hydrogen itself has extremely strong diffusibility and extremely high flammability and explosiveness characteristics. Its minimum ignition energy is only 0.02 mJ, and the limit explosion range is 4% - 75%. Once a leakage occurs, it is extremely likely to cause major safety accidents such as fires and even explosions.

[0003] The existing public document 1 (Analysis of Hydrogen Leakage and Explosion Accidents and Risks in Hydrogen Energy Buses in Parking Lots, 2024) discloses a method for analyzing hydrogen leakage and explosion accidents and risks in hydrogen energy buses. This method uses FLACS software to simulate the hydrogen diffusion characteristics under different leakage conditions after the failure of key components (such as TPRD and hydrogen pipeline valves), analyzes the hydrogen volume fraction distribution, and studies the overpressure characteristics after the gas cloud explosion according to the ignition position. Finally, it quantifies the accident impact range and explosion consequence risks and proposes prevention and control suggestions. However, this method cannot predict potential risks in advance.

[0004] The existing public document 2 (Analysis of Consequences of Liquid Hydrogen Leakage and Quantitative Risk Assessment in Hydrogen Refueling Stations, 2023) discloses a method based on quantitative risk assessment. This method obtains the diffusion, fire, and explosion impact ranges after liquid hydrogen leakage through CFD simulation, and then combines parameters such as failure probability, meteorological data, and personnel distribution to calculate the risk values of each scenario and compares them with the internationally commonly used ALARP criterion. However, this method does not conduct dynamic risk assessment.

[0005] Therefore, there is an urgent need for a hydrogen leakage assessment model that can predict potential risks in advance, conduct dynamic risk assessment, and improve the safety of hydrogen use. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a hydrogen leakage risk assessment model based on artificial intelligence. By collecting real-time parameters, combining a 1D-CNN and CNN-LSTM fusion neural network, generating a leakage risk factor sequence, using a sliding window and a dynamic correction mechanism to predict the leakage rate, and combining the Gaussian plume model and sensor data to invert the leakage source position, it dynamically assesses the leakage risk to solve the problems raised in the above background art.

[0007] To achieve the above object, the present invention provides the following technical solutions: An artificial intelligence-based hydrogen leakage risk assessment model, including a dynamic monitoring data module, a risk feature extraction module, a leakage rate prediction and source inversion module, and a leakage evolution module; the dynamic monitoring data module collects real-time parameters and preprocesses the real-time parameters; the risk feature extraction module introduces a 1D-CNN and CNN-LSTM fusion neural network to generate a leakage risk factor sequence; the leakage rate prediction and source inversion module uses a sliding window and a dynamic correction mechanism to predict the leakage rate and combines the Gaussian plume model and sensor data to predict the location of the hydrogen leakage source; the leakage evolution module displays a heat map of high-risk areas and the location of the leakage source; The leakage rate prediction and source inversion module predicts the source location of hydrogen leakage, and the specific steps are as follows: S1, the change of hydrogen concentration in the hydrogen storage device is monitored in real time through a hydrogen concentration sensor, the air pressure change in the hydrogen storage device is monitored by a pressure sensor, the ambient wind speed value is monitored by a wind speed sensor, and the ambient temperature value is monitored by a temperature sensor; S2, according to the estimated value of the leakage rate As the source strength, combined with the ambient wind speed value and the ambient temperature value, the Gaussian plume model is used to simulate the change of the hydrogen concentration distribution in space over time; S3, after the hydrogen leakage occurs, the hydrogen concentration sensor records the real-time concentration value, compares the real-time concentration value with the preset theoretical concentration value, constructs an error function, accumulates the square difference between the actual concentration and the theoretical value at each sensor position, and uses the gradient descent method to calculate the leakage source position coordinates that minimize the error function.

[0008] As a further solution of the present invention, in S3, when using the gradient descent method to optimize the error function, calculate the gradient of the error function with respect to the leakage source position. The gradient is the rate of change of the error function in different directions and reflects the sensitivity of the error in each direction; adjust the position of the leakage source according to the opposite direction of the gradient, that is, move forward along the direction of error reduction; each time it is updated, the change amplitude of the position is controlled by the parameter of the learning rate, and the learning rate determines the adjustment step; continuously iterate and update the position of the leakage source until the gradient of the error function is less than the preset threshold, and the iteration stops.

[0009] As a further solution of the present invention, the calculation formula for the leakage rate of the leakage rate prediction and source inversion module is: , where is the time the estimated value of the hydrogen leakage rate of the system at time is the time the instantaneous flow rate of the hydrogen inlet main pipe at time is the time the instantaneous flow rate at the outlet of the system at time is the change rate of the internal hydrogen gas volume of the system, and errors are suppressed through a sliding window and a dynamic correction mechanism.

[0010] As a further solution of the present invention, the dynamic correction mechanism combines the statistical characteristics of the sliding window with the first derivative trend analysis to judge the authenticity of leakage, including the average leakage rate , the leakage alarm threshold and the trend threshold .

[0011] As a further solution of the present invention, the real-time parameters collected by the dynamic monitoring data module include the instantaneous flow rate of the hydrogen inlet main pipe , the instantaneous flow rate at the outlet of the hydrogen discharge system , the pressure in the stator cooling area , the pressure in the end cover area , the casing temperature , the jacking oil pressure , the bearing vibration amplitude value and the shaft vibration amplitude value .

[0012] As a further solution of the present invention, the preprocessing steps of the dynamic monitoring data module include time axis alignment, outlier removal and normalization processing, and convert the data into a structured vector sequence.

[0013] As a further solution of the present invention, the 1D-CNN layer of the risk feature extraction module scans the local short-term change trend of the feature vector sequence through a convolution kernel, and inputs it into the CNN-LSTM fusion neural network after being processed by a non-linear activation function and a max pooling layer.

[0014] As a further solution of the present invention, the structure of the CNN-LSTM fusion neural network includes a convolutional layer, an LSTM layer and a fully connected layer. The CNN-LSTM fusion neural network outputs the leakage risk factor at the current time point , and generates a risk factor sequence , where represents the length of the sliding window.

[0015] As a further solution of the present invention, the leakage evolution module displays the hydrogen concentration distribution through a color gradient heat map, and marks the leakage source position and the diffusion path.

[0016] Technical effects and advantages of an artificial intelligence-based hydrogen leakage risk assessment model of the present invention: By means of the dynamic monitoring data module, the present invention collects and preprocesses key parameters in real time, and combines the 1D-CNN and CNN-LSTM fusion neural networks in the risk feature extraction module to generate a leakage risk factor sequence, realizing the early prediction of potential hydrogen leakage risks. In addition, the leakage rate prediction and source inversion module uses a sliding window and a dynamic correction mechanism to predict the leakage rate in real time, and combines the Gaussian plume model and sensor data to invert the leakage source location, realizing the dynamic assessment of leakage risks. At the same time, the leakage evolution module intuitively presents the leakage diffusion dynamics and risk area prediction through the display of the high-risk area heat map and the leakage source location. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic structural diagram of an artificial intelligence-based hydrogen leakage risk assessment model of the present invention; Figure 2 is a schematic diagram of the hydrogen cloud distribution in the event of hydrogen leakage after the failure of the TPRD of a hydrogen energy bus in the prior art; Figure 3 is a schematic diagram of the steps for quantitative risk assessment of a hydrogen refueling station in the prior art; Figure 4 is the fully connected layer of the CNN-LSTM fusion neural network of the present invention; Figure 5 is a schematic diagram of the change in hydrogen leakage rate and hydrogen intake and exhaust volume of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] Refer to Figure 1 the shown schematic structural diagram. An embodiment of the present invention provides an artificial intelligence-based hydrogen leakage risk assessment model, which includes a dynamic monitoring data module, a risk feature extraction module, a leakage rate prediction and source inversion module, and a leakage evolution module; the dynamic monitoring data module collects real-time parameters and preprocesses the real-time parameters; the risk feature extraction module introduces a 1D-CNN and CNN-LSTM fusion neural network to generate a leakage risk factor sequence; the leakage rate prediction and source inversion module uses a sliding window and a dynamic correction mechanism to predict the leakage rate, and combines the Gaussian plume model and sensor data to predict the location of the hydrogen leakage source; the leakage evolution module displays the high-risk area heat map and the leakage source location.

[0020] Further, the dynamic monitoring data module collects real-time parameters and preprocesses the real-time parameters, including: the dynamic monitoring data module combines with the PLC and SCADA systems to collect parameters. The parameters include the instantaneous flow rate of the hydrogen inlet main pipe , the instantaneous flow rate at the outlet of the hydrogen discharge system , the pressure in the stator cooling area , the pressure in the end cover area , the temperature of the machine casing , the jacking oil pressure , the vibration amplitude value of the bearing and the vibration amplitude value of the shaft . The instantaneous flow rate of the hydrogen inlet main pipe is measured in real time through a thermal mass flowmeter, which reflects the gas supply volume from the external hydrogen source to the generator cavity. The instantaneous flow rate at the outlet of the hydrogen discharge system is monitored in real time by using a thermal mass flowmeter, so as to reflect the hydrogen emission volume and the venting speed; the instantaneous flow rate at the outlet of the hydrogen discharge system and the instantaneous flow rate of the hydrogen inlet main pipe can be jointly used to preliminarily judge the retention change of hydrogen: , where is the storage change amount. The pressure difference between the pressure in the regular cooling area and the pressure in the end cover area can reflect the sealing performance of the hydrogen seal system. The temperature of the machine casing is obtained through a distributed thermocouple. The fluctuation of the machine casing temperature may indicate stress abnormality caused by thermal expansion or a decrease in heat dissipation efficiency. The jacking oil pressure is used as an operating state parameter of the shafting support system, and its fluctuation may be coupled and conducted to the sealing system. The vibration amplitude value of the bearing and the vibration amplitude value of the shaft are obtained through an eddy current sensor installed on the bearing. The vibration amplitude value of the bearing and the vibration amplitude value of the shaft reflect the mechanical disturbance and unbalance state during the operation of the motor.

[0021] The dynamic monitoring data module preprocesses the above parameters. First, the data from each channel are aligned on the unified time axis to ensure that all features have the same time resolution and sampling frequency; then, the outlier points that may be caused by sensor drift, short-term jump or communication error are removed by using the Huadong window anomaly detection method; and then, the standardized change is applied to normalize all features so that their mean value is 0 and the standard deviation is 1. The preprocessed structure is passed into the risk feature extraction module in the form of a structured vector sequence , where is the full-scale working condition feature vector at the 1st second, is the full-scale working condition feature vector at the 1st second, is the full-scale operating condition feature vector at the nth second.

[0022] Furthermore, the risk feature extraction module introduces a 1D-CNN and CNN-LSTM fusion neural network to generate a leakage risk factor sequence, including: the risk feature extraction module obtains the preprocessed standardized feature vector sequence from the dynamic monitoring data module . The risk feature extraction module introduces a one-dimensional convolutional neural network (1D-CNN) layer. The one-dimensional convolutional neural network layer scans the local short-term change trends of the feature vector sequence through multiple convolutional kernels, extracts the edge features in the hydrogen system during sudden fluctuations or the initial stage of leakage, and after being processed by a non-linear activation function, is sent to the max pooling layer to compress the feature space and enhance the robustness to weak features. The extracted local features enter the CNN-LSTM fusion neural network. The window size of the input feature vector sequence at each moment is set to n = 60 seconds to form an input matrix : , where is the number of features.

[0023] Refer to Figure 4 , the CNN-LSTM fusion neural network consists of a convolutional layer structure, an LSTM layer, and a fully connected layer; the convolutional layer structure is used for the change features within the local time window of the feature vector sequence, and its structure is as follows: the input dimension is , the convolutional kernel size is 3, the number of convolutional kernels is 32, the activation function is ReLU, the pooling method is max pooling, and the output size is ; the LSTM layer is used to capture long-term time dependencies and learn non-linear leakage trends, and its structure is as follows: the input size is , the number of hidden layer units is 64, Droput = 0.3, and the output dimension is ; the fully connected layer uses the Sigmoid activation function to map the output to the interval [0,1] to obtain the leakage risk factor at the current time point . The leakage risk factor represents the probability of hydrogen leakage in the system under the time window corresponding to the input features. To form a continuous risk evolution sequence, the system evaluates each frame of the continuous input sliding window to generate a complete risk factor sequence , where represents the length of the sliding window.

[0024] In this embodiment, the CNN-LSTM fusion neural network labels the data in the corresponding time period according to the fault work order, and labels it as "safe", "warning", "suspicious leakage", and "confirmed leakage" for the supervised learning of the network. The dataset division ratio is 70% for the training set, 15% for the validation set, and 15% for the test set. The training set is preprocessed by standardization, differential stationarity, etc. and then input into the model. First, the 1D convolutional layer extracts local feature patterns in the sequence, and then enters the LSTM layer to model long-term dependencies. Then, the fully connected layer and the Sigmoid activation function output the risk factor R(t) ∈ [0, 1], and the cross-entropy loss function is used to measure the error between the model output and the true label. The calculation formula of the cross-entropy loss function is: , where is the cross-entropy loss function, is the number of categories, is the weight of the th category, is the true label of the th category, is the probability that the model outputs the th category. Table 1 shows the training parameter settings of the CNN-LSTM fusion neural network: Table 1 Training Parameter Settings

[0025] During the entire training process, the cross-validation method is used to evaluate the generalization ability of the CNN-LSTM fusion neural network. Table 2 shows the test results of the CNN-LSTM fusion neural network on the thermal power unit data: Table 2 Evaluation Metrics

[0026] The following is the Python code of the CNN-LSTM fusion neural network: def risk_prediction_model(X_input): X_cnn = Conv1D(filters=32, kernel_size=3)(X_input) X_pool = MaxPooling1D(pool_size=2)(X_cnn) X_lstm = LSTM(units=64, return_sequences=False)(X_pool) output = Dense(1, activation='sigmoid')(X_lstm) return output Perform grading according to the risk factor sequence. If , the response action is normal operation and the risk level is level 0 (safe); if , the response action is the yellow light flashing, start enhanced monitoring, and the risk level is level 1 (warning); if , the response action is the red light alarm, prompt manual confirmation, and the risk level is level 2 (suspected leakage); if , the response action is the red light remaining on, link to the hydrogen discharge operation, and the risk level is level 3 (confirmed leakage).

[0027] Furthermore, the leakage rate prediction and source point inversion module uses a sliding window and a dynamic correction mechanism to predict the leakage rate, and combines the Gaussian plume model and sensor data to invert the position of the hydrogen leakage source, including: when the leakage rate prediction and source point inversion module identifies a risk level of level 2 or above, start the inversion process. The inversion process includes: 1. Predict the leakage rate; 2. Predict the leakage position.

[0028] During the normal operation of the hydrogen-cooled generator set, the instantaneous flow rate of the hydrogen inlet main pipe should maintain a dynamic balance with the instantaneous flow rate at the outlet of the hydrogen discharge system. Once a hydrogen leakage occurs, the un-discharged part may leak out through system gaps or seal failure parts, manifested as an imbalance phenomenon. Therefore, the leakage rate is predicted by constructing a mathematical model based on mass conservation, and the calculation formula is: , where is the estimated value of the hydrogen leakage rate of the system at time , is the instantaneous flow rate of the hydrogen inlet main pipe at time , is the instantaneous flow rate at the outlet of the system at time , is the change rate of the hydrogen volume inside the system; introduce a dynamic correction mechanism, and adopt a strategy that combines the statistical characteristics of the sliding window and the first derivative trend analysis to verify the authenticity of the segment leakage in a dynamically changing environment. The system constructs a sliding window with a fixed time length (10 seconds), calculates the leakage rate estimated value within each window to obtain the average leakage rate value of this window. If continuously exceeds the preset leakage alarm threshold , it indicates that there is a trend of flow imbalance between the inlet and outlet of hydrogen during this time period; at the same time, the system further introduces the calculation of the first derivative of the leakage rate sequence. If the derivative value continuously exceeds the preset trend threshold , it is determined that the current leakage exists and has a risk of increase. The dynamic correction mechanism can effectively suppress the following three types of errors: 1. False leakage prompts caused by single-point spike errors or sensor jumps; 2. Instantaneous flow imbalance caused by operation disturbances; 3. Short-term drift of the equilibrium point caused by normal system regulation.

[0029] In this embodiment, refer to Figure 5 , in most time periods, the hydrogen intake and hydrogen discharge are relatively close or equal, indicating that the system is operating normally. However, at certain moments (such as 2025-04-02 00:00:30 and 2025-04-02 00:01:30), there are obvious differences between the hydrogen intake and hydrogen discharge, resulting in an increase in the leakage rate, which indicates that hydrogen leaks at these moments. Especially at certain time periods, such as 00:00:30, the leakage rate reaches a relatively high value (15.45 Nm³ / h), reflecting the severity of the leakage.

[0030] In this embodiment, the following is the Python code for the leakage rate estimation model: THRESHOLD_LEAK_RATE = 0.5 # Leakage rate alarm threshold (unit: Nm³ / h) TREND_SLOPE_EPSILON = 0.1 # Threshold for judging the leakage trend slope MIN_SUSTAIN_DURATION = 10 # Shortest duration criterion (unit: seconds) WINDOW_SIZE = 10 # Sliding window size (seconds) leak_rate_buffer = [] # Save the leakage rate values in the past WINDOW_SIZE seconds timestamp_buffer = [] # Corresponding timestamp list def estimate_leak_rate(Q_in, Q_out, V_H2_current=None, V_H2_previous=None, use_V=False): """ Q_in: Current hydrogen intake (Nm³ / h) Q_out: Current hydrogen discharge (Nm³ / h) V_H2_current: Estimated hydrogen gas volume inside the current system V_H2_previous: Volume at the previous time point use_V: Whether to enable the volume compensation mode (boolean value) return: The currently estimated leakage rate """ if use_V and V_H2_current is not None and V_H2_previous is not None: dV_dt = (V_H2_current - V_H2_previous) Q_leak = Q_in - Q_out - dV_dt else: Q_leak = Q_in - Q_out return max(Q_leak, 0)# Do not allow negative leakage rates def update_leak_detection(Q_in_t, Q_out_t, timestamp_t): global leak_rate_buffer, timestamp_buffer Q_leak_t = estimate_leak_rate(Q_in_t, Q_out_t) leak_rate_buffer.append(Q_leak_t) timestamp_buffer.append(timestamp_t) if len(leak_rate_buffer)>WINDOW_SIZE: leak_rate_buffer.pop(0) timestamp_buffer.pop(0) sustained_above_threshold = all(rate>THRESHOLD_LEAK_RATE for rate inleak_rate_buffer) sustained_duration = timestamp_buffer[-1] - timestamp_buffer[0]if len(timestamp_buffer)>= 2 else 0 if len(leak_rate_buffer)>= 2: slope = (leak_rate_buffer[-1] - leak_rate_buffer[0]) / sustained_duration else: slope = 0 if sustained_above_threshold and sustained_duration>= MIN_SUSTAIN_DURATION and slope>= TREND_SLOPE_EPSILON: alert_leak_detected(Q_leak_t) else: continue_monitoring(Q_leak_t) def alert_leak_detected(Q_leak_est): print(f"[ALERT] Leak occurred! Estimated leak rate: {Q_leak_est:.2f} Nm³ / h") def continue_monitoring(Q_leak_est): print(f"[INFO] No leak at present. Estimated leak rate: {Q_leak_est:.2f} Nm³ / h") The leakage rate prediction and source point inversion module predicts the source location of hydrogen leakage by collecting multi-point sensor data and combining the physical diffusion model with the data-driven method, including the following specific steps: S1. Real-time monitor the change of hydrogen concentration in the hydrogen storage device through a hydrogen concentration sensor, monitor the air pressure change in the hydrogen storage device through a pressure sensor, monitor the ambient wind speed value through a wind speed sensor, and monitor the ambient temperature value through a temperature sensor.

[0031] S2. Hydrogen leakage is affected by environmental factors, diffuses in space and forms a certain concentration distribution. According to the estimated leakage rate as the source strength, combine the ambient wind speed value and the ambient temperature value to calculate the first diffusion coefficient and the second diffusion coefficient , and then use the Gaussian plume model to simulate the change of hydrogen concentration distribution in space over time.

[0032] S3. After hydrogen leakage occurs, hydrogen concentration sensors distributed around the hydrogen storage equipment record the real-time concentration values at different positions, and compare the real-time concentration values with the preset theoretical concentration values. By constructing an error function, the sum of the squared differences between the actual concentration and the theoretical value at each sensor position is accumulated, and the gradient descent method is used to calculate the leakage source position coordinates that minimize the error function.

[0033] When using the gradient descent method to optimize the error function, calculate the gradient of the error function with respect to the leakage source position. The gradient is the rate of change of the error function in different directions, reflecting the sensitivity of the error in each direction. Adjust the position of the leakage source according to the opposite direction of the gradient, that is, move forward along the direction of decreasing error. Each time it is updated, the change amplitude of the position is controlled by the parameter of the learning rate, and the learning rate determines the adjustment step. By continuously iteratively updating the position of the leakage source until the gradient of the error function is less than the preset threshold, the iteration stops.

[0034] Furthermore, the leakage evolution module displays the high-risk area heat map and the leakage source position, including: the leakage evolution prediction module displays the high-risk area heat map, visually shows the distribution of hydrogen concentration by means of color gradient, and marks the possible positions and diffusion paths of the leakage source. According to the calculation results of the leakage flow rate and the diffusion model, the diffusion path and concentration change of the gas are rendered, providing users with accurate risk area prediction and dynamic demonstration of leakage diffusion. Users can view the leakage situation at different time points through the interactive interface at any time, and combine historical data analysis to analyze the possible evolution trend of the leakage source.

[0035] The present invention realizes the early prediction of potential hydrogen leakage risks by dynamically monitoring data module to collect and preprocess key parameters in real time, and combining the 1D-CNN and CNN-LSTM fusion neural network in the risk feature extraction module to generate a leakage risk factor sequence. In addition, the leakage rate prediction and source point inversion module uses a sliding window and a dynamic correction mechanism to predict the leakage rate in real time, and combines the Gaussian plume model and sensor data to invert the leakage source position, realizing the dynamic assessment of leakage risks. At the same time, the leakage evolution module visually presents the leakage diffusion dynamics and risk area prediction through the display of the high-risk area heat map and the leakage source position.

[0036] Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A hydrogen leakage risk assessment model based on artificial intelligence, characterized in that: It includes a dynamic monitoring data module, a risk feature extraction module, a leakage rate prediction and source point inversion module and a leakage evolution module; the dynamic monitoring data module collects real-time parameters and preprocesses the real-time parameters; the risk feature extraction module introduces 1D-CNN and CNN-LSTM fusion neural networks to generate leakage risk factor sequences; the leakage rate prediction and source point inversion module predicts the leakage rate using a sliding window and a dynamic correction mechanism, and predicts the location of the hydrogen leakage source in combination with the Gaussian plume model and sensor data; the leakage evolution module displays a high-risk area heat map and the leakage source location; the leakage rate prediction and source point inversion module predicts the source location of the hydrogen leakage, and the specific steps are as follows: S1, monitor the change of hydrogen concentration in the hydrogen storage device in real time through the hydrogen concentration sensor, monitor the change of air pressure in the hydrogen storage device through the pressure sensor, monitor the ambient wind speed value through the wind speed sensor, and monitor the ambient temperature value through the temperature sensor; S2, estimated value based on leak rate As the source strength, the Gaussian plume model is used to simulate the change of hydrogen concentration distribution in space over time in combination with the ambient wind speed and ambient temperature values; S3, after a hydrogen leak occurs, the hydrogen concentration sensor records the real-time concentration value, compares the real-time concentration value with the preset theoretical concentration value, constructs an error function, accumulates the square difference between the actual concentration and the theoretical value at each sensor position, and uses the gradient descent method to calculate the leakage source position coordinates that minimize the error function.

2. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 1 is characterized in that In S3, when the gradient descent method is used to optimize the error function, the gradient of the error function relative to the position of the leakage source is calculated. The gradient is the rate of change of the error function in different directions, reflecting the sensitivity of the error in each direction; the position of the leakage source is adjusted in the opposite direction of the gradient, that is, moving in the direction of reducing the error; each time the position is updated, the change amplitude is controlled by the parameters of the learning rate, and the learning rate determines the adjustment pace; the position of the leakage source is updated through continuous iteration until the gradient of the error function is less than the preset threshold, and the iteration stops.

3. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 1 is characterized in that ,The calculation formula of the leakage rate of the leakage rate prediction and source point inversion module is: ,in, For time Estimated value of the system hydrogen leakage rate at time, For time The instantaneous flow rate of hydrogen inlet manifold at any time, For time The instantaneous flow at the system outlet at time, is the rate of change of the hydrogen volume inside the system, and the error is suppressed through a sliding window and dynamic correction mechanism.

4. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 3, characterized in that: The dynamic correction mechanism combines the sliding window statistical characteristics with the first-order derivative trend analysis to determine the authenticity of the leak, including the average leak rate , Leakage alarm threshold and trend threshold .

5. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 1, characterized in that: The real-time parameters collected by the dynamic monitoring data module include the instantaneous flow rate of the hydrogen inlet manifold , Instantaneous flow rate at the outlet of hydrogen exhaust system , Pressure in stator cooling area , Pressure in the end cap area , Case temperature , top shaft oil pressure , Watt Amplitude Axis amplitude .

6. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 1, characterized in that: The preprocessing steps of the dynamic monitoring data module include time axis alignment, outlier removal and standardization to convert the data into a structured vector sequence.

7. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 1, characterized in that: The 1D-CNN layer of the risk feature extraction module scans the local short-term change trend of the feature vector sequence through the convolution kernel, and inputs the CNN-LSTM fusion neural network after being processed by the nonlinear activation function and the maximum pooling layer.

8. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 7, characterized in that: The structure of the CNN-LSTM fusion neural network includes a convolutional layer, an LSTM layer and a fully connected layer. The CNN-LSTM fusion neural network outputs the leakage risk factor at the current time point. , and generate the risk factor sequence ,in Indicates the length of the sliding window.

9. The artificial intelligence-based hydrogen leakage risk assessment model according to claim 1, characterized in that: The leakage evolution module displays the hydrogen concentration distribution through a color gradient heat map, marking the leakage source location and diffusion path.

Images