A comprehensive energy station leakage detection method fusing infrared thermal imaging and gas sensing data

Abstract

This invention relates to the field of hydrogen energy station safety technology, and particularly to a comprehensive method for detecting leaks in energy stations that integrates infrared thermal imaging and gas sensing data. The method utilizes convolutional neural networks to deeply extract the spatial features of infrared thermal imaging data and long short-term memory networks to independently model the temporal dynamic features of gas sensor data. The probability outputs of the two independent models are converted into the basic probability assignment function in Dempster-Shafer evidence theory through an uncertainty quantification method based on information entropy. Evidence reasoning and fusion are performed at the decision-making level, which not only integrates multi-source information but also explicitly models and handles the uncertainties of each piece of evidence. This application not only overcomes the shortcomings of insufficient reliability of existing single-sensor technologies and the poor fault tolerance and high model coupling of low-level data fusion methods, but also solves the problem of performance degradation of existing systems in complex environments.

Description

Technical Field

[0001] This invention relates to the field of hydrogen energy station safety technology, and in particular to a comprehensive method for detecting leaks in energy stations that integrates infrared thermal imaging and gas sensing data. Background Technology

[0002] Hydrogen energy, as a clean and efficient secondary energy source, is increasingly widely used in transportation, energy storage, and distributed energy supply. As a critical infrastructure for hydrogen storage, refueling, and supply, the operational safety of hydrogen energy stations is paramount. Hydrogen has characteristics such as small molecular weight, rapid diffusion, high concealment of leaks, and a wide explosive limit range (4%–75%). Once a leak occurs, it can easily accumulate in a short time, forming a flammable and explosive atmosphere, posing a serious threat to personnel, equipment, and the environment. Therefore, developing efficient, reliable, and timely hydrogen leak detection technologies is of great significance for ensuring the safe operation of hydrogen energy stations.

[0003] Existing hydrogen power station leak detection technologies mainly rely on single-type sensors, such as hydrogen concentration sensors based on electrochemical or semiconductor principles, or thermal anomaly detection methods based on infrared imaging. However, single-sensor detection methods have significant limitations in practical engineering applications.

[0004] While gas sensor-based methods offer high sensitivity and relatively fast response, their performance is susceptible to factors such as ambient temperature, humidity, airflow disturbances, cross-gas interference, and sensor aging and drift. This can lead to false alarms and missed alarms in complex real-world conditions. Furthermore, gas sensors are typically point-based, making it difficult to comprehensively cover large or well-ventilated areas, resulting in delayed responses to early, minute leaks or leaks located far from the sensor placement point.

[0005] Infrared methods offer advantages such as wide monitoring range, rapid response, and the ability to pinpoint leak locations. However, their detection effectiveness is significantly affected by factors such as background heat sources (e.g., sunlight, equipment heating), ambient temperature fluctuations, atmospheric conditions, and imaging distance. When used alone, they struggle to distinguish between genuine leak signals and interference signals in complex thermal backgrounds, resulting in limited reliability.

[0006] To improve the reliability of hydrogen leak detection, some studies have attempted to fuse infrared and gas data. However, these methods often focus on simple threshold superposition or feature-level stitching, exhibiting limitations such as high model coupling, poor system scalability, and a sharp decline in overall performance when a single module fails. More importantly, existing fusion methods generally lack explicit modeling and compensation mechanisms for dynamic environmental disturbances. In the real and variable environment of a hydrogen energy station, changes in ambient temperature, humidity, wind speed, air pressure, and other interfering gases simultaneously affect the sensing reliability of both infrared and gas sensors. Existing methods fail to incorporate these key environmental factors into the decision-making system, resulting in insufficient robustness under harsh or extreme conditions and limiting their practical value. Summary of the Invention

[0007] In a first aspect, the present invention discloses a comprehensive method for detecting leaks in energy plants that integrates infrared thermal imaging and gas sensing data, the specific method of which is as follows: Acquisition and preprocessing of infrared thermal imaging data and gas sensing data; Infrared thermal imaging data is processed using a convolutional neural network model to obtain the infrared leakage probability value; Gas sensor data is processed using a long short-term memory network model to obtain the probability value of gas leakage; Information entropy is introduced to calculate the credibility of convolutional neural network models and long short-term memory network models; combined with the leakage probability value and the corresponding credibility, the probability allocation function of the infrared model and the probability allocation function of the gas model are output. The probability assignment functions of the infrared model and the gas model were modified by taking into account environmental factors. Based on the DS evidence theory, the probability assignment function of the infrared model and the probability assignment function of the gas model are integrated to complete the leak detection.

[0008] Furthermore, the specific methods for infrared thermal imaging data acquisition and preprocessing are as follows: Infrared thermal imaging data of the hydrogen energy station to be tested is obtained using infrared thermal imaging equipment. Infrared thermal imaging data is corrected for non-uniformity and calibrated for radiometric data to obtain temperature field data. The temperature field data matrix is ​​smoothed using a Gaussian filter model, as shown in the formula: ; in, Here is the temperature field matrix. For Gaussian kernel function, The radius of the convolution window; The denoised temperature field matrix is ​​normalized to obtain the original normalized temperature field feature map. The formula is: ; in, and These are the minimum and maximum temperature values ​​in the current frame's temperature field, respectively. The temperature field is then mapped to [the specified values] using a normalization operation. interval; The formula for calculating the temperature gradient feature map is: ; in, Represents the temperature gradient; The formula for calculating the time-varying rate characteristic plot is: ; in, and These represent the normalized temperature field data at adjacent time points. Indicates time rate of change over time Fusion of original normalized temperature field feature map Temperature gradient feature map Characteristic plot of rate of change over time Construct an infrared feature map consisting of three channels; The specific methods for gas sensor data acquisition and preprocessing are as follows: For the time series of gas sensing data, time-series features are constructed, and signal rate of change features are extracted. The formula is: ; Calculate cumulative change characteristics : ; The original normalized signal, rate of change feature, and cumulative change feature are fused together to obtain the following expression: Multi-feature sequence at time step : ; in, , , ; The continuous time series is divided using a sliding time window method. For the first... A time window can be used to obtain time series samples. : ; in, Indicates the length of the time window. Indicates the sample number; The dataset consisting of all the partitioned gas feature sequence samples is represented as follows: ; in, Includes all items categorized by fixed time window The time series samples obtained after partitioning.

[0009] Furthermore, the infrared leakage probability value is obtained using the following method: The convolutional neural network model includes a feature extraction module, a feature fusion module, and a detection head. The feature extraction module is based on ResNet and uses convolutional kernels of different sizes to extract input features. The feature fusion module uses pooling to achieve multi-source feature fusion and outputs the features to the detection head. The detection head makes result predictions based on multi-scale features. The specific method for obtaining the probability value of gas leakage is as follows: Sequence samples After inputting the LSTM model, the model at each time step Receive the feature vector at the current time step by step And perform recursive state update calculations; Regarding time input vector The internal state update process of an LSTM cell is jointly controlled by the input gate, forget gate, output gate, and memory cell. After stacking multiple LSTM networks, a sequence of hidden states within a time window is formed. : ; The hidden state sequence is modeled and represented holistically through time-dimensional feature aggregation operations to form a high-level semantic feature vector under the gas mode. : ; in, This represents a time-series aggregation function.

[0010] Furthermore, the probability assignment function includes the leakage state. Non-leaking state and uncertain states The state set is as follows: ; The CNN model outputs a classification probability vector as follows: ; The LSTM model outputs a classification probability vector as follows: ; in, Indicates the confidence level of the leak. Indicates the non-leakage confidence level; The probability assignment function for the infrared model is constructed as follows: ; ; ; in, The infrared model BPA represents the confidence level of the prediction of three system states of a hydrogen energy station based on infrared image data. The probability assignment function for constructing the gas sensing model is as follows: ; ; ; in, The gas sensing model BPA represents the confidence level of the prediction of three system states of a hydrogen energy station based on gas sensing data.

[0011] Furthermore, information entropy is introduced to calculate the credibility of convolutional neural network models and long short-term memory network models. The specific method is as follows: The information entropy of the output probabilities of CNN and LSTM models is: ; ; in, The information entropy represents the output probability of a CNN model. The information entropy represents the output probability of the LSTM model; The entropy value is normalized to construct an uncertainty measure factor: ; ; in, The uncertainty measure of the prediction results of the CNN model. This represents a measure of the uncertainty of the LSTM model's prediction results. Based on the uncertainty metric, the model outputs a confidence level: ; ; in, This indicates the reliability of the CNN model's prediction results. This indicates the reliability of the LSTM model's prediction results.

[0012] Furthermore, considering environmental factors, the probability assignment functions of the infrared model and the gas model are modified. The specific methods are as follows: Real-time acquisition of environmental parameters of the hydrogen energy station under test, forming an environmental parameter vector. : ; in, For ambient temperature, For ambient humidity, Atmospheric pressure, For wind speed, The concentration of interfering gases; Establish a rule base for the impact of environmental parameters on sensor reliability, and calculate the weights using fuzzy inference. The formula is as follows: ; ; in, Infrared mode weights are affected by ambient temperature, humidity, and wind speed. When the ambient temperature is close to the target temperature range, humidity is low, and wind speed is low, It approaches 1; conversely, it decreases. As the gas mode weights, they are affected by temperature, humidity, pressure, and interference from other gases. When environmental conditions are stable and there is no cross-gas interference, It approaches 1; conversely, it decreases. The probability assignment function of the infrared model is corrected using the following formula: ; ; ; in, This indicates the corrected result; The formula for correcting the probability assignment function of the gas model is as follows: ; ; ; in, This indicates the corrected result.

[0013] Furthermore, based on DS evidence theory, the probability assignment functions of the infrared model and the gas model are fused, and the specific method is as follows: Introducing the coefficient of evidence conflict : ; The closer a value is to 0, the higher the consistency between the infrared model evidence and the gas model evidence. The closer a value is to 1, the more significant the conflict between the infrared model evidence and the gas model evidence. Based on the DS evidence theory, infrared model evidence Evidence from gas models Perform combined calculations: ; ; in, , This represents the final confidence level of the fusion of the various states of the hydrogen energy station.

[0014] Further, leak detection is completed, and the final judgment rules are as follows: ; ; in, and These are the system safety control threshold parameters.

[0015] Secondly, the present invention provides a hydrogen energy station leak detection device that integrates infrared thermal imaging and gas sensing data, the device comprising: The multi-source data acquisition module is responsible for real-time acquisition of temperature field data and gas concentration data in key areas of the hydrogen energy station. The fault diagnosis module is responsible for running the infrared leak detection CNN model, the gas leak detection LSTM model, and the decision-level fusion algorithm based on DS evidence theory. The communication and alarm response module is responsible for uploading the detection results, alarm information and system status to the monitoring center in real time, and supports local audible and visual alarms and security linkage. The power and management module provides operating power.

[0016] Due to the adoption of the above technical solutions, this application has the following beneficial effects: 1. This application utilizes convolutional neural networks to deeply extract the spatial features of infrared thermal imaging data and uses long short-term memory networks to independently model the temporal dynamic features of gas sensor data. The fusion of these two types of information can improve detection accuracy.

[0017] 2. Compared to traditional fusion methods, this application converts the probability outputs of two independent models into the basic probability assignment function in Dempster-Shafer evidence theory using an uncertainty quantification method based on information entropy. Subsequently, evidence reasoning and fusion are performed at the decision-making level, which not only integrates multi-source information but also explicitly models and handles the uncertainty of each piece of evidence.

[0018] 3. This application further introduces an environmental weighted correction mechanism. By collecting and quantifying environmental parameters in real time, it dynamically assesses the impact of environmental factors on the reliability of infrared and gas sensors and converts them into evidence weights. Before evidence fusion, these weights are used to adaptively correct the basic probability allocation function, thereby enhancing the overall robustness and decision accuracy of the system under changing environments.

[0019] 4. This application not only overcomes the shortcomings of insufficient reliability of existing single sensor technology and poor fault tolerance and high model coupling of low-level data fusion methods, but also solves the problem of performance degradation of existing systems in complex environments by introducing environmental perception and adaptive weighting mechanism, providing an innovative technical solution for realizing intelligent detection and graded early warning of hydrogen leakage in hydrogen energy stations.

[0020] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0021] The accompanying drawings of this invention are described below.

[0022] Figure 1 This is a schematic diagram of the overall process of the method described in this invention.

[0023] Figure 2 This is a schematic diagram of the overall framework of the device described in this invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Example 1: A comprehensive method for detecting leaks in energy plants that integrates infrared thermal imaging and gas sensing data, comprising the following steps: S1. Acquisition and preprocessing of infrared thermal imaging data and gas sensing data.

[0026] In step S1 of this embodiment, infrared thermal imagers and gas sensing devices are deployed in key areas of the hydrogen energy station to collect infrared thermal imaging data and gas sensing data, respectively. The gas sensing data is directly derived from the raw output of the gas sensor and is used to characterize hydrogen concentration, diffusion state, and their changing trends. The collected infrared thermal imaging data undergoes radiometric calibration, noise suppression, and background modeling to obtain an infrared input feature map; the gas sensing data undergoes outlier removal, time synchronization, and smoothing to obtain a standardized input feature sequence.

[0027] S11, Infrared Data Preprocessing Infrared data of the target area to be detected is acquired using an infrared thermal imaging device; the data is a 16-bit grayscale image with a resolution of 320*240. , This represents the pixel coordinates of the image. Each pixel represents the original temperature. First, non-uniformity correction is performed on the original infrared image to eliminate systematic errors introduced by inconsistent responses of different pixels in the infrared detector array. Then, radiometric calibration is performed to map the radiation intensity values ​​output by the sensor to the actual physical temperature values, thus realizing the conversion of infrared image data into physical temperature field data.

[0028] The obtained temperature field matrix is ​​smoothed using a Gaussian filter model, as shown in the following formula: ; in, Here is the temperature field matrix. For Gaussian kernel function, is the radius of the convolution window.

[0029] To enhance the contrast between the abnormal thermal field region of the leak and the background temperature field, the denoised temperature field matrix is ​​normalized using the following formula: ; in, and These are the minimum and maximum temperature values ​​in the current frame's temperature field, respectively. The temperature field is then mapped to [the specified values] using a normalization operation. The interval ensures that the data distribution in different scenarios maintains a consistent scale, which is beneficial to the stability of subsequent model training.

[0030] To fully characterize the spatial and temporal features of the leakage process, infrared feature maps with different physical meanings are fused and modeled to construct an infrared feature map consisting of three channels. The three channels represent the original normalized temperature field feature map. Temperature gradient feature map and time rate of change feature map.

[0031] The formulas used to calculate the temperature gradient feature map are as follows: ; in, This represents the temperature gradient.

[0032] The formula used to calculate the time-rate-of-change characteristic plot is as follows: ; in, and These represent the normalized temperature field data at adjacent time points. Indicates time The rate of change over time.

[0033] Finally, multiple infrared feature maps were combined to form an infrared input image dataset for the hydrogen energy station.

[0034] S12, Gas Sensing Data Preprocessing The gas sensor outputs a continuous time series signal data, and its raw output signal can be expressed as: ; in, Indicates the first Each gas sensing channel at time The output signal value, It indicates the number of sensing channels, and the signal form can be voltage value, current value, or digital output signal.

[0035] The gas sensing time series signal was then filtered to suppress the effects of environmental noise, electromagnetic interference, and random disturbances on signal stability. The moving average filtering expression used is as follows: ; in, The length of the sliding window. This is the output of the filtered signal.

[0036] The filtered signal sequence is normalized, and the normalization expression is as follows: ; in, and They represent the first The mean and standard deviation of the signals from each sensor channel in historical stable operating condition data are standardized to eliminate the influence of the differences in the dimensions of different sensor channels, thereby improving the stability of multi-source signal fusion modeling.

[0037] To characterize the dynamic evolution and abrupt changes during hydrogen leakage, time-series data were used to construct time-series features and extract signal rate of change features. : ; and cumulative change characteristics : ; The original normalized signal, rate of change feature, and cumulative change feature are fused together to obtain the following expression: Multi-feature sequence at time step : ; in, , , .

[0038] Then, a sliding time window method is used to divide the continuous time series. For the ... A time window can be used to obtain time series samples. : ; in, Indicates the length of the time window. Indicates the sample number.

[0039] The dataset consisting of all the partitioned gas feature sequence samples can be represented as: ; in, Includes all items categorized by fixed time window The time series samples obtained after partitioning.

[0040] S2. Use a convolutional neural network model to process infrared thermal imaging data and obtain the infrared leakage probability value.

[0041] In step S2 of this embodiment, an infrared leakage detection model based on a convolutional neural network (CNN) is constructed. The processed infrared thermal imaging data is used as the model input. The spatial features of the temperature distribution are extracted through multi-layer convolution operations. After feature mapping and nonlinear transformation, the infrared determination result or infrared leakage probability value of hydrogen leakage at the current moment is output as infrared detection evidence information.

[0042] The infrared leakage detection model for hydrogen energy stations consists of three parts: a feature extraction module, a feature fusion module, and a detection head. The feature extraction module, based on ResNet, uses convolutional kernels of different sizes to extract input features, specifically four sizes: 1*1, 3*3, 5*5, and 7*7. The extracted features are then fed into the feature fusion network. The feature fusion module uses pooling to fuse multi-source features and further extracts gas temperature field sensing features, finally outputting the features to the detection head. The detection head predicts the result based on multi-scale features, filtering out unreasonable predictions by eliminating non-maximum values ​​to obtain the final infrared detection prediction result.

[0043] S3. Use a long short-term memory network model to process gas sensing data and obtain the probability value of gas leakage.

[0044] In step S3 of this embodiment, a gas leak detection model based on a Long Short-Term Memory (LSTM) network is constructed. The gas sensing data is divided into time series according to a preset time window and input into the LSTM model. By modeling the time series of gas concentration change trends and diffusion behavior, the gas judgment result or gas leak probability value of hydrogen leakage in the current state is output as gas detection evidence information.

[0045] The completed gas sensing time series samples The input is a Long Short-Term Memory (LSTM) network model for dynamic feature learning. Because a sliding time window is used to construct the time series samples, there is overlap between samples from different time windows on the time axis, meaning that feature vectors from some time points will appear in multiple samples. This mechanism enhances the model's ability to learn continuous-time evolution patterns, enabling it to learn the dynamic changes in the gas diffusion process from different temporal contexts.

[0046] Sequence samples After inputting the LSTM model, the model at each time step Receive the feature vector at the current time step by step And perform recursive state update calculations.

[0047] Regarding time input vector The internal state update process of an LSTM cell is jointly controlled by the input gate, forget gate, output gate, and memory unit. Its state update mechanism is expressed as follows: ; ; ; ; ; ; in, , , These represent the state vectors of the forget gate, input gate, and output gate, respectively. Indicates the state of the memory cell. This indicates that the hidden layer outputs a feature vector. These are the network weight parameters. , , , For bias parameters, This is an element-wise inner product operation.

[0048] Through the aforementioned recursive modeling mechanism, the LSTM network can model the dynamic changes of gas feature sequences over time, automatically learning the time-dependent characteristics, cumulative effect characteristics, and abrupt response characteristics inherent in gas diffusion, thus achieving a mapping transformation from structured gas feature input to high-dimensional temporal semantic feature expression. After stacking multiple layers of LSTM networks, a hidden state sequence within the time window is formed. : ; The hidden state sequence is modeled and represented holistically through time-dimensional feature aggregation operations to form a high-level semantic feature vector under the gas mode. : ; in, This represents a time-series aggregation function, used to comprehensively represent and model dynamic feature information within a time window.

[0049] During the model training phase, the gas feature time series samples output by the gas data processing module are used as the training set input to the LSTM network. The network weight parameters are iteratively optimized through the error backpropagation algorithm, enabling the model to gradually learn the dynamic change law of the gas environment under normal operating conditions. During the prediction phase, the gas feature time series collected in real time and constructed by the data processing module are input into the trained LSTM model, and the corresponding hydrogen energy station leakage prediction results based on gas sensing data are output.

[0050] S4. Introduce information entropy to calculate the credibility of the convolutional neural network model and the long short-term memory network model; combine the leakage probability value and the corresponding credibility to output the probability allocation function of the infrared model and the probability allocation function of the gas model.

[0051] In step S4 of this embodiment, in order to achieve a unified judgment output of infrared thermal imaging detection results and gas sensing time-series detection results in the multimodal intelligent sensing system, this invention constructs a decision-level fusion module. By introducing an evidence theory framework, the output results of multiple models are mapped from a probabilistic expression form to an evidence expression form, thereby realizing the fusion modeling of multimodal information at the decision level.

[0052] The state set of the hydrogen energy station system is constructed as follows: ; in, Indicates "leakage status". Indicates "non-leaking state". This indicates an "uncertain state" and is used to describe the basic states that the system may be in.

[0053] The CNN model outputs a classification probability vector as follows: ; The LSTM model outputs a classification probability vector as follows: ; in, Indicates the confidence level of the leak. Indicates the non-leakage confidence level.

[0054] To construct the evidence representation, uncertainty is introduced. This method uses information entropy to construct uncertainty weights. The information entropy of the output probabilities of the two models is as follows: ; ; in, The information entropy represents the output probability of a CNN model. The information entropy represents the output probability of the LSTM model.

[0055] The entropy value was then normalized to construct an uncertainty measure factor: ; ; in, The uncertainty measure of the prediction results of the CNN model. This represents the uncertainty measure of the prediction results of the LSTM model.

[0056] Based on the uncertainty metric, the model outputs a confidence level. ; ; in, This indicates the reliability of the CNN model's prediction results. This indicates the reliability of the LSTM model's prediction results.

[0057] Constructing the basic probability assignment function (BPA) for the infrared model: ; ; ; in, The infrared model BPA represents the confidence level of the prediction of three system states of a hydrogen energy station based on infrared image data.

[0058] Constructing the basic probability assignment function (BPA) for the gas sensing model: ; ; ; in, The gas sensing model BPA represents the confidence level of the prediction of three system states of a hydrogen energy station based on gas sensing data.

[0059] This enables the model output to be transformed from a probability space to an entropy-based evidence space, allowing both the model's predictive credibility and uncertainty information to participate in decision modeling simultaneously.

[0060] S5. Modify the probability assignment functions of the infrared model and the gas model by taking into account environmental factors.

[0061] In step S5 of this embodiment, to improve the robustness of the system's decision-making in complex environments, environmental factors (such as external temperature, humidity, wind speed, atmospheric pressure, and concentrations of other interfering gases) are introduced to dynamically weight and correct the quality of evidence. By constructing an environmental impact quantification model, real-time environmental parameters are converted into evidence weights, which are then adjusted during the BPA calculation process, thereby enhancing the adaptability of the fusion decision to environmental interference.

[0062] Real-time collection of environmental parameters in key areas of the hydrogen energy station to construct an environmental parameter vector. : ; in, ambient temperature , For ambient humidity , atmospheric pressure , Wind speed , The concentration of other interfering gases (such as methane and carbon monoxide). .

[0063] Because the infrared detection mode and the gas detection mode have different sensitivities to environmental factors, it is necessary to construct separate environmental impact weighting functions for each. and Based on sensor physical characteristics and engineering experience, a rule base for the influence of environmental parameters on sensor reliability is established, and weights are calculated using fuzzy inference. The calculation formula is as follows: ; ; in, The infrared mode weights are mainly affected by ambient temperature, humidity, and wind speed. When the ambient temperature is close to the target temperature range, humidity is low, and wind speed is low... It approaches 1; conversely, it decreases. The gas mode weights are mainly affected by temperature, humidity, pressure, and interference from other gases. When environmental conditions are stable and there is no cross-gas interference, It approaches 1; conversely, it decreases.

[0064] The weighted BPA of the corrected infrared data is: ; ; ; in, This indicates the corrected result.

[0065] The weighted BPA of the corrected gas data is: ; ; ; in, This indicates the corrected result.

[0066] Environmental weight This reflects the overall reliability of the sensor mode under current environmental conditions. Under ideal environmental conditions, The evidence before and after the revision remained essentially unchanged; when environmental interference was severe, As the confidence level decreases, the level of trust in the certain state (leaked / unleaked) in the corresponding evidence also decreases, with more trust being allocated to the uncertain state. This reflects the system's cautious decision-making in this environment.

[0067] S6. Based on DS evidence theory, the probability assignment function of the infrared model and the probability assignment function of the gas model are integrated to complete the leak detection.

[0068] In step S6 of this embodiment, an evidence space is constructed by outputting the predicted probability, and external factors are converted into weights for correction. Finally, the result is input into a decision-level fusion model built based on evidence theory to obtain a comprehensive prediction result of hydrogen leakage under the current state. When the fusion prediction result exceeds a preset threshold, the system determines that there is a risk of hydrogen leakage and triggers corresponding alarm prompts and safety linkage responses; when the fusion prediction result does not exceed the preset threshold, the system determines that the current state is a safe operating state.

[0069] To characterize the consistency and conflict among multiple sources of evidence, an evidence conflict coefficient is introduced. : ; The closer a value is to 0, the higher the consistency between infrared image evidence and gas sensing evidence. The closer a value is to 1, the more significant the conflict between infrared image evidence and gas sensing evidence.

[0070] Based on the DS evidence theory, infrared image evidence Evidence from gas sensing Perform combined calculations: ; ; in, , This represents the final confidence level of the fusion of the various states of the hydrogen energy station.

[0071] After obtaining the final fusion confidence result, the final decision rule of the system is constructed as follows: ; ; in, and These are the system safety control threshold parameters.

[0072] Example 2: A hydrogen energy station leak detection device that integrates infrared thermal imaging and gas sensing data, such as Figure 2 As shown, the device includes: a multi-source data acquisition module, a fault diagnosis module, a communication and alarm response module, and a power supply and management module.

[0073] P1, Multi-source data acquisition module This module consists of an infrared thermal imaging unit and a gas sensing unit, responsible for real-time acquisition of temperature field data and gas concentration data in key areas of the hydrogen energy station. The infrared thermal imaging unit uses a high-resolution infrared thermal imager (such as the FLIRA series or HIKVISIONDS-2TP series), with a resolution of 320×240 or higher pixels, supporting 16-bit temperature data output, and a frame rate of at least 25fps. This unit has built-in radiometric calibration and nonlinear correction functions, achieving a temperature measurement accuracy of ±0.5℃ within the range of −20℃ to +150℃, adapting to complex indoor and outdoor lighting and temperature difference environments. The gas sensing unit uses a high-sensitivity hydrogen sensor (such as the FIGAROTGS2611 series or MEMBRAPORH2-S-100), supporting detection ranges of 0−1000ppm or wider, with a response time of less than 10 seconds, and features strong anti-interference capabilities and good long-term stability. The unit can also be optionally equipped with a temperature and humidity sensor (such as the SensirionSHT35) for environmental compensation, improving the reliability of gas detection.

[0074] This module can employ a high-speed analog-to-digital converter (such as the ADIAD4020), which features 8-channel synchronous sampling capability, 16-bit resolution, and built-in analog front-end and overvoltage protection. Its synchronous sampling characteristics ensure strict alignment of the infrared temperature field data and multi-channel gas sensor data in time steps, providing a precise data foundation for subsequent time-series-based fusion analysis.

[0075] P2, Fault Diagnosis Module This module is the core computing unit, responsible for running the infrared leak detection CNN model, the gas leak detection LSTM model, and the decision-level fusion algorithm based on DS evidence theory. It employs a high-performance embedded AI computing platform (such as NVIDIA Jetson Nano / Xavier NX or Huawei Atlas200DK), integrating a multi-core CPU and GPU, supporting TensorRT accelerated inference, and capable of processing infrared image sequences and gas time-series data in real time.

[0076] P3, Communication and Alarm Response Module This module is responsible for uploading detection results, alarm information, and system status to the monitoring center in real time, and supports local audible and visual alarms and security linkage. It adopts an industrial-grade wireless communication and microcontroller integrated chip solution, such as the ESP32 series. This chip integrates Wi-Fi and Bluetooth dual-mode radio frequency, and has a built-in low-power microprocessor and rich peripheral interfaces. Its integrated wireless communication function provides a stable and flexible data upload channel, ensuring that fused alarm information can be delivered to the remote monitoring platform in real time. Simultaneously, its built-in MCU can directly process logic judgments and drive GPIO ports to control audible and visual alarms or manage ventilation equipment through relays, achieving integrated "sensing-decision-control," simplifying system design and improving the speed and reliability of linkage response.

[0077] P4, Power and Management Module To adapt to the industrial environment of hydrogen energy stations, the device adopts a wide-voltage input power supply design and has a built-in power management chip (such as TITPS54560) to achieve efficient voltage regulation and overload protection. It can also be equipped with a battery or solar power supply interface to ensure continuous operation in the event of a sudden power outage.

[0078] In addition, the device supports remote firmware upgrades, fault self-diagnosis, and log export functions. It can flexibly configure detection thresholds, alarm levels, and linkage strategies according to actual scenarios, enabling all-weather, intelligent monitoring and rapid handling of leakage risks at hydrogen energy stations.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An integrated energy site leak detection method that fuses infrared thermography and gas sensing data, characterized by, The specific method is as follows: Acquisition and preprocessing of infrared thermal imaging data and gas sensing data; Infrared thermal imaging data is processed using a convolutional neural network model to obtain the infrared leakage probability value; Gas sensor data is processed using a long short-term memory network model to obtain the probability value of gas leakage; Information entropy is introduced to calculate the credibility of convolutional neural network models and long short-term memory network models; combined with the leakage probability value and the corresponding credibility, the probability allocation function of the infrared model and the probability allocation function of the gas model are output. The probability assignment functions of the infrared model and the gas model were modified by taking into account environmental factors. Based on the DS evidence theory, the probability assignment function of the infrared model and the probability assignment function of the gas model are integrated to complete the leak detection.

2. The integrated energy field station leak detection method that fuses infrared thermography with gas sensor data of claim 1, wherein, The specific methods for infrared thermal imaging data acquisition and preprocessing are as follows: Infrared thermal imaging data of the hydrogen energy station to be tested is obtained using infrared thermal imaging equipment. Infrared thermal imaging data is corrected for non-uniformity and calibrated for radiometric data to obtain temperature field data. The temperature field data matrix is ​​smoothed using a Gaussian filter model, as shown in the formula: ； wherein, is a temperature field matrix, is a Gaussian kernel function, is a convolution window radius; The denoised temperature field matrix is ​​normalized to obtain the original normalized temperature field feature map. The formula is: ； in, and These are the minimum and maximum temperature values ​​in the current frame's temperature field, respectively. The temperature field is then mapped to [the specified values] using a normalization operation. interval; The formula for calculating the temperature gradient feature map is: ； in, Represents the temperature gradient; The formula for calculating the time-varying rate characteristic plot is: ； in, and These represent the normalized temperature field data at adjacent time points. Indicates time rate of change over time Fusion of original normalized temperature field feature map Temperature gradient feature map Characteristic plot of rate of change over time Construct an infrared feature map consisting of three channels; The specific methods for gas sensor data acquisition and preprocessing are as follows: For the time series of gas sensing data, time-series features are constructed, and signal rate of change features are extracted. The formula is: ； Calculate cumulative change characteristics : ； The original normalized signal, rate of change feature, and cumulative change feature are fused together to obtain the following expression: Multi-feature sequence at time step : ； in, , , ; The continuous time series is divided using a sliding time window method. For the first... A time window can be used to obtain time series samples. : ； in, Indicates the length of the time window. Indicates the sample number; The dataset consisting of all the partitioned gas feature sequence samples is represented as follows: ； in, Includes all items categorized by fixed time window The time series samples obtained after partitioning.

3. The integrated energy station leak detection method fusing infrared thermal imaging and gas sensing data as described in claim 2, characterized in that, The specific method for obtaining the infrared leakage probability value is as follows: The convolutional neural network model includes a feature extraction module, a feature fusion module, and a detection head. The feature extraction module is based on ResNet and uses convolutional kernels of different sizes to extract input features. The feature fusion module uses pooling to achieve multi-source feature fusion and outputs the features to the detection head. The detection head makes result predictions based on multi-scale features. The specific method for obtaining the probability value of gas leakage is as follows: Sequence samples After inputting the LSTM model, the model at each time step Receive the feature vector at the current time step by step And perform recursive state update calculations; Regarding time input vector The internal state update process of an LSTM cell is jointly controlled by the input gate, forget gate, output gate, and memory cell. After stacking multiple LSTM networks, a sequence of hidden states within a time window is formed. : ； The hidden state sequence is modeled and represented holistically through time-dimensional feature aggregation operations to form a high-level semantic feature vector under the gas mode. : ； in, This represents a time-series aggregation function.

4. The integrated energy station leak detection method fusing infrared thermal imaging and gas sensing data as described in claim 1, characterized in that, The probability assignment function includes leakage states. Non-leaking state and uncertain states The state set is as follows: ； The CNN model outputs a classification probability vector as follows: ； The LSTM model outputs a classification probability vector as follows: ； in, Indicates the confidence level of the leak. Indicates the non-leakage confidence level; The probability assignment function for the infrared model is constructed as follows: ； ； ； in, The infrared model BPA represents the confidence level of the prediction of three system states of a hydrogen energy station based on infrared image data. The probability assignment function for constructing the gas sensing model is as follows: ； ； ； in, The gas sensing model BPA represents the confidence level of the prediction of three system states of a hydrogen energy station based on gas sensing data.

5. The integrated energy station leak detection method fusing infrared thermal imaging and gas sensing data as described in claim 4, characterized in that, The credibility of convolutional neural network models and long short-term memory network models is calculated using information entropy, and the specific method is as follows: The information entropy of the output probabilities of CNN and LSTM models is: ； ； in, The information entropy represents the output probability of a CNN model. The information entropy represents the output probability of the LSTM model; The entropy value is normalized to construct an uncertainty measure factor: ； ； in, The uncertainty measure of the prediction results of the CNN model. This represents a measure of the uncertainty of the LSTM model's prediction results. Based on the uncertainty metric, the model outputs a confidence level: ； ； in, This indicates the reliability of the CNN model's prediction results. This indicates the reliability of the LSTM model's prediction results.

6. The integrated energy station leak detection method fusing infrared thermal imaging and gas sensing data as described in claim 4, characterized in that, Taking environmental factors into account, the probability assignment functions of the infrared model and the gas model are modified. The specific methods are as follows: Real-time acquisition of environmental parameters of the hydrogen energy station under test, forming an environmental parameter vector. : ； in, For ambient temperature, For ambient humidity, Atmospheric pressure, For wind speed, The concentration of interfering gases; Establish a rule base for the impact of environmental parameters on sensor reliability, and calculate the weights using fuzzy inference. The formula is as follows: ； ； in, Infrared mode weights are affected by ambient temperature, humidity, and wind speed. When the ambient temperature is close to the target temperature range, humidity is low, and wind speed is low, It approaches 1; conversely, it decreases. As the gas mode weights, they are affected by temperature, humidity, pressure, and interference from other gases. When environmental conditions are stable and there is no cross-gas interference, It approaches 1; conversely, it decreases. The probability assignment function of the infrared model is corrected using the following formula: ； ； ； in, This indicates the corrected result; The formula for correcting the probability assignment function of the gas model is as follows: ； ； ； in, This indicates the corrected result.

7. The integrated energy station leak detection method fusing infrared thermal imaging and gas sensing data as described in claim 1, characterized in that, Based on DS evidence theory, the probability assignment functions of the infrared model and the gas model are integrated. The specific method is as follows: Introducing the coefficient of evidence conflict : ； The closer a value is to 0, the higher the consistency between the infrared model evidence and the gas model evidence. The closer a value is to 1, the more significant the conflict between the infrared model evidence and the gas model evidence. Based on the DS evidence theory, infrared model evidence Evidence from gas models Perform combined calculations: ； ； in, , This represents the final confidence level of the fusion of the various states of the hydrogen energy station.

8. The integrated energy station leak detection method fusing infrared thermal imaging and gas sensing data as described in claim 7, characterized in that, Leak detection is complete, and the final judgment rules are as follows: ； ； in, and These are the system safety control threshold parameters.

9. A hydrogen energy station leak detection device integrating infrared thermal imaging and gas sensing data, characterized in that, The device includes: The multi-source data acquisition module is responsible for real-time acquisition of temperature field data and gas concentration data in key areas of the hydrogen energy station. The fault diagnosis module is responsible for running the infrared leak detection CNN model, the gas leak detection LSTM model, and the decision-level fusion algorithm based on DS evidence theory. The communication and alarm response module is responsible for uploading the detection results, alarm information and system status to the monitoring center in real time, and supports local audible and visual alarms and security linkage. The power and management module provides operating power.

Images