A gas leakage detection method based on multi-sensor data of underground pipe network

Abstract

The present invention relates to a gas leak detection method based on multi-sensor data from an underground pipeline network. The method comprises the following steps: acquiring detection data from multiple environmental sensors installed in the soil surrounding the underground pipeline network; performing data enhancement processing on the detection data from the multiple environmental sensors using a trained generative model to obtain time series data; and inputting the time series data into a trained neural network model to obtain gas leak detection results. The neural network model comprises a feature extraction portion for extracting deep features from the time series data to obtain deep feature data; an attention layer for denoising the deep feature data and combining the front-back spatial and channel features of the deep feature data; and an LSTM layer for processing the denoised deep feature data and outputting a gas leak detection result. The present invention can effectively improve the accuracy of gas leak detection.

Description

Technical Field

[0001] The present invention relates to the field of pipeline network safety technology, and in particular to a gas leakage detection method based on multi-element sensor data of an underground pipeline network. Background Art

[0002] In the field of pipeline safety technology, the critical importance of underground pipeline gas leak detection methods is self-evident. This is particularly true in industries such as industry and the chemical industry, where underground pipelines are used for oil and natural gas transmission. With the continuous advancement of industrialization, the coverage of underground pipelines continues to expand, increasing the likelihood of leaks. However, gas pipeline leaks present a daunting challenge. Gas pipeline leaks are categorized into three types: micropore seepage, small hole leaks, and exposed fractures. Leaks can be caused by a variety of factors, including natural corrosion, external damage, and pipeline design flaws. Once a leak occurs, the impact can far exceed the economic losses of the enterprise, potentially causing serious and irreversible damage to the surrounding soil environment, water sources, and even human health. In more serious cases, leaks can even lead to explosions, posing a significant threat to the safety of people and property in the surrounding area. Therefore, promoting the research and development and application of underground pipeline gas leak detection methods is extremely urgent and important.

[0003] Gas leak detection in underground pipelines can be done using sensor detection methods. However, sensor detection methods are limited by many factors and problems, which restrict the development of underground pipeline gas leak detection methods. The main problems are as follows: (1) The sensitivity of the sensor is limited, and it may not be able to detect small leaks in time; (2) The acquisition accuracy of the sensor is limited, and environmental interference may lead to false detection; (3) The sensor equipment is a hardware device and requires long-term and frequent system maintenance.

[0004] The current rapid development of deep learning artificial intelligence methods can make up for the above shortcomings. However, in the field of underground pipeline gas leakage detection using deep learning, there are also some problems that need to be solved urgently: neural networks have difficulty extracting features from time series data and it is difficult to obtain good prediction results; time series data often contains a large amount of noise, which affects the training of neural networks. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a gas leakage detection method based on multi-element sensor data of an underground pipe network, which can effectively improve the accuracy of gas leakage detection.

[0006] The technical solution adopted by the present invention to solve the technical problem is to provide a gas leakage detection method based on underground pipe network multi-sensor data, comprising the following steps:

[0007] Acquire detection data from various environmental sensors installed in the soil around the underground pipe network;

[0008] Using the trained generative model to perform data enhancement processing on the detection data of the multiple environmental sensors to obtain time series data;

[0009] The time series data is input into the trained neural network model to obtain the gas leakage detection result; wherein,

[0010] The neural network model includes:

[0011] A feature extraction part, used to extract deep features of the time series data to obtain deep feature data;

[0012] The attention layer is used to denoise the deep feature data and combine the front and back spatial and channel features of the deep feature data;

[0013] The LSTM layer is used to process the denoised deep feature data and output the gas leak detection results.

[0014] The multiple environmental sensors are sensors whose correlation with the gas sensor to be measured exceeds a threshold value and are found by using the RF random forest scoring system.

[0015] The RF random forest scoring system uses the Gini index or information gain to evaluate the correlation between the environmental sensor and the gas sensor to be measured. The larger the Gini index or information gain, the greater the correlation with the gas sensor to be measured.

[0016] The generation model includes:

[0017] An encoder part, for mapping the detection data into a latent space;

[0018] The decoder part is used to convert latent variables into time series data;

[0019] The generative model learns the distribution of latent variables, the parameters of the encoder part, and the parameters of the decoder part by maximizing the likelihood function.

[0020] The attention layer part includes:

[0021] A decomposition module is used to decompose the deep feature data into high-frequency subsequences, medium-frequency subsequences, low-frequency subsequences and residual sequences using complete set empirical mode decomposition;

[0022] an attention denoising module, configured to denoise the high-frequency subsequence, the medium-frequency subsequence, the low-frequency subsequence and the residual sequence;

[0023] The summation and reconstruction module is used to sum and reconstruct the denoised high-frequency subsequence, medium-frequency subsequence, low-frequency subsequence and residual sequence to obtain the denoised deep feature data.

[0024] The attention denoising module includes a channel attention submodule, a spatial attention submodule and a first wavelet threshold joint denoising module. The input feature is fused with the input feature after passing through the channel attention submodule to obtain a first fused feature. The first fused feature is passed through the first wavelet threshold joint denoising module to obtain a denoised feature. The denoised feature is fused with the denoised feature after passing through the spatial attention submodule to obtain a second fused feature. In the channel attention submodule, the input feature is processed by average pooling and maximum pooling respectively, and then passed through the second wavelet threshold joint denoising module for wavelet threshold denoising. The feature is then transferred by the shared fully connected layer, the transferred features are added, and the channel attention feature weight is obtained after passing through the sigmoid activation function.

[0025] The neural network model also includes:

[0026] The Dropout layer is located at the back end of the LSTM layer and is used to prevent the neural network model from overfitting.

[0027] The neural network model training includes the following steps:

[0028] Obtain historical detection data from various environmental sensors installed in the soil around underground pipe networks;

[0029] Taking the gas sensor to be tested among the multiple environmental sensors as a reference, marking the case where the detection data of the gas sensor to be tested is not 0 as a positive sample, and marking the case where the detection data of the gas sensor to be tested is 0 as a negative sample;

[0030] Expanding the historical detection data using the generative model, and selecting the expanded historical detection data as a training set;

[0031] The neural network model is trained using the training set to obtain a trained neural network model.

[0032] The technical solution adopted by the present invention to solve its technical problem is: to provide an electronic device, including a memory, a processor and a computer program stored in the memory and capable of running on the processor, and when the processor executes the computer program, the steps of the above-mentioned gas leakage detection method based on underground pipeline network multi-sensor data are implemented.

[0033] The technical solution adopted by the present invention to solve its technical problem is: providing a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned gas leakage detection method based on underground pipeline network multi-sensor data are implemented.

[0034] Beneficial effects

[0035] Due to the adoption of the above-mentioned technical solution, the present invention has the following advantages and positive effects compared with the prior art: the present invention measures the change data caused by gas leakage to the surrounding soil through sensors, uses artificial intelligence deep learning to predict indirect detection, and utilizes a neural network model to improve the accuracy of underground pipeline gas leakage detection, thereby reducing the consumption of manpower and material resources in actual situations; at the same time, when constructing the neural network model, by expanding the leakage gas data in the underground pipeline gas leakage dataset, the proportion of leakage data in the dataset is increased, the purpose of balancing the positive and negative samples in the dataset is achieved, and the robustness of the dataset is enhanced; a new attention module DCAM is introduced into the neural network model in the present invention, which can effectively reduce the noise in the data, comprehensively extract and fully integrate the information of each feature in the data, so as to improve the detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a technical roadmap of the gas leakage detection method according to the first embodiment of the present invention;

[0037] Figure 2 is a flow chart of the random forest scoring system in the first embodiment of the present invention;

[0038] Figure 3 is a structural diagram of a neural network model in the first embodiment of the present invention;

[0039] Figure 4 is a structural diagram of the DCAM attention module in the first embodiment of the present invention;

[0040] Figure 5 This is a structural diagram of the CEWT_CAM attention mechanism in the first embodiment of the present invention. DETAILED DESCRIPTION

[0041] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0042] The first embodiment of the present invention relates to a gas leakage detection method based on multi-sensor data of an underground pipeline network, comprising the following steps: obtaining detection data from multiple environmental sensors installed in the soil around the underground pipeline network; using a trained generative model to perform data enhancement processing on the detection data of the multiple environmental sensors to obtain time series data; and inputting the time series data into a trained neural network model to obtain gas leakage detection results.

[0043] Figure 1 This is the technical roadmap for this embodiment's gas leak detection method based on multi-sensor data from underground pipeline networks. This method relies on data preprocessing, data enhancement, and neural network model building. The preprocessing phase includes data preprocessing and multi-sensor data evaluation. The data enhancement phase employs a generative model to augment the original data with positive and negative samples, generating a new simulated dataset. This simulated dataset is then used to train predictions on the original dataset. The neural network model building phase proposes a new attention layer, DCAM, which can be used to build a predictive model for gas leaks.

[0044] Data preprocessing in this embodiment encompasses both missing value handling and label classification. Given that the historical dataset collected by the sensor contains only a small amount of missing data, this embodiment removes these missing values ​​to handle this missing data. During label classification, based on the gas sensor data to be tested, any non-zero gas sensor reading is considered a gas leak and labeled as a positive sample with a dataset label of 1. Any zero gas sensor reading is considered a non-leak and labeled as a negative sample with a dataset label of 0.

[0045] In actual situations, it is impossible to collect accurate leakage data of the gas to be detected in the underground pipeline network. Therefore, it is necessary to adopt an indirect prediction method to use other sensor data instead of the leakage gas data to be detected for prediction. This embodiment uses the RF random forest scoring system to evaluate the correlation between other sensor data and the sensor data of the leakage gas to be detected, ensuring the reliability of other sensor data, thereby improving the detection accuracy of the neural network.

[0046] The process of the RF random forest scoring system in this embodiment is as follows Figure 2 As shown in the figure, it uses the feature correlation between the remaining sensor data and the dataset label as the standard to score the sensor data features of each dimension. The higher the score, the greater the correlation with the label. The score is usually evaluated by the Gini index Gini(X) or Information Gain InformationGain(X), which is expressed as:

[0047]

[0048] Among them, X represents the feature, k represents the category, and p k is the probability that the sample on the node belongs to category k, V is all possible values ​​of feature X, S v is a subset of samples whose feature X takes the value v, Entropy(S) and Entropy(S v ) are sample set S and subset S respectively v Information entropy.

[0049] In this embodiment, the generative model used in the data enhancement stage is a time variational autoencoder (Time_VAE). The Time_VAE is used to generate the number of positive and negative samples in the gas leakage dataset after data screening, thereby balancing the positive and negative samples in the dataset and eliminating the interference of environmental factors in different data, thereby improving the detection accuracy from the perspective of dataset robustness.

[0050] Time-VAE is an extension of the Variational Autoencoder (VAE), primarily designed for processing time series data. The advantage of Time-VAE lies in its focus on modeling temporal or spatiotemporal data, capturing time-related patterns and trends and generating new data with temporal structure. By learning latent representations, it reduces the dimensionality of the data, facilitating subsequent tasks such as classification, clustering, and anomaly detection. Its uncertainty modeling capabilities provide rich uncertainty information, helping to improve the robustness of the model.

[0051] The Time-VAE model architecture typically consists of an encoder and a decoder. The encoder maps the input data into a latent space, while the decoder converts the latent variables into time series data. During training, the encoder and decoder are typically jointly optimized to minimize reconstruction error and the distribution difference of the latent variables.

[0052] The expression of the encoder part is:

[0053]

[0054] The expression of the decoder part is:

[0055]

[0056] Where x represents the input time series data, z represents the latent variable, e(x) and g(z) represent the mapping functions of the encoder and decoder parts respectively, σ 2 Represents variance.

[0057] The Time-VAE in this implementation learns the distribution of latent variables and the parameters of the encoder and decoder by maximizing the likelihood function, where the likelihood function is expressed as:

[0058] L=E q(z|x) [log p(x|z)]-βD KL (q(z|x)||p(z))

[0059] Among them, E q(z|x) (·) is the reconstruction loss, which represents the expected value of the log-likelihood of the input data x given the latent variable z. q(z|x) represents the posterior distribution of the latent variable z. log p(x|z) represents the log-likelihood of the input data x given the latent variable z. β is a tuning parameter used to balance the weight between the reconstruction loss and the KL divergence term. D KL (·) is the KL divergence (Kullback-Leibler divergence), which represents the difference between the posterior distribution q(z|x) of the latent variable and the prior distribution p(z), where p(z) represents the prior distribution of the latent variable z.

[0060] During the testing phase, the Time-VAE in this embodiment can generate new time series data by sampling latent variables from the latent space.

[0061] After generating new data through Time-VAE, the generated positive and negative samples are labeled and used for training the neural network model, and the original data is used for prediction of the neural network model to reconstruct the training data set and prediction data set.

[0062] During the neural network model building phase, the reconstructed training data set is input into the neural network model for training. The neural network model in this embodiment is as follows: Figure 3 As shown, it includes a feature extraction part, an attention layer part, and an LSTM layer part, wherein the feature extraction part can use a CNN network. After the data is input, the deep features of the data are first extracted using the CNN network. Considering that there must be a lot of noise in the data after Time_VAE, a single CNN network structure cannot combine the front and back space and channel features of the data. Therefore, the attention layer part in this embodiment is an attention mechanism module DCAM, which removes noise and combines the front and back space and channel features of the data. Finally, after processing by the LSTM layer, the gas leak detection result (i.e., binary classification output) is output. In order to avoid overfitting, the neural network model of this embodiment can also include a Dropout layer part, which is set at the back end of the LSTM layer part.

[0063] like Figure 4As shown, the attention mechanism module DCAM in this embodiment includes: a decomposition module, which is used to decompose the deep feature data into high-frequency subsequences, medium-frequency subsequences, low-frequency subsequences and residual sequences using complete set empirical mode decomposition; an attention denoising module, which is used to denoise the high-frequency subsequences, medium-frequency subsequences, low-frequency subsequences and residual sequences; and a summation and reconstruction module, which is used to sum and reconstruct the denoised high-frequency subsequences, medium-frequency subsequences, low-frequency subsequences and residual sequences to obtain the denoised deep feature data.

[0064] Since the enhanced data is generated by Time_VAE, it is impossible to determine which part of the data noise is mainly concentrated in, nor can it be determined which frequency band the main features of the data are concentrated in. Therefore, the complete set empirical mode decomposition CEEMDAN is first used to decompose the data into a high-frequency subsequence IMF1, a medium-frequency subsequence IMF2, a low-frequency subsequence IMF3 and a residual sequence RES, and each subsequence is input into the attention denoising module CEWT_CAM to obtain the denoised features of different frequency bands. Finally, the denoised deep feature data is obtained by summing and reconstructing these features, and it is used as the input of the LSTM layer.

[0065] The structure of the attention denoising module CEWT_CAM is as follows Figure 5 As shown in the figure, unlike the traditional CBAM attention mechanism, the attention denoising module CEWT_CAM adds a wavelet threshold joint denoising module CEWT between the channel attention submodule and the spatial attention submodule, as well as inside the channel attention. Its purpose is to simultaneously fuse data features and denoise data, thereby reducing the number of parameters in the network training process.

[0066] In this embodiment, the input feature is fused with the input feature after passing through the channel attention submodule to obtain a first fused feature. The first fused feature is passed through the first wavelet threshold joint denoising module to obtain a denoised feature. The denoised feature is fused with the denoised feature after passing through the spatial attention submodule to obtain a second fused feature. The second fused feature is the denoised feature data. In the channel attention submodule, the input features are processed by average pooling and maximum pooling respectively, and then pass through the second wavelet threshold joint denoising module for wavelet threshold denoising. The features are then transferred by the shared fully connected layer, the transferred features are added, and the channel attention feature weights are obtained after passing through the sigmoid activation function.

[0067] It can be seen that this implementation benefits from the data decomposition capability of the complete set empirical mode decomposition CEEMDAN, decomposing the features into several subsequences of high, medium and low frequency IMFs and residuals RES, and then independently denoising each subsequence, thereby removing the noise generated by Time_VAE data in the features while retaining the feature points of each frequency band.

[0068] The present invention is further described below by taking methane gas leakage detection as an example.

[0069] The underground pipeline methane gas leakage dataset contains four sensor data sets: system temperature (system_temperature) sensor data, water content 3 (water_content3) sensor data, soil methane temperature (soil_ch4_temp) sensor data, and soil methane (soil_ch4) sensor data.

[0070] First, the missing values ​​in the above dataset are deleted, and the soil methane (soil_ch4) sensor data is used as the benchmark for labeling. If the soil methane (soil_ch4) sensor data is not 0, it is considered to be a gas leak, which is a positive sample and the dataset label is 1. Otherwise, it is considered to be a negative sample and the dataset label is 0.

[0071] The preprocessed dataset was evaluated using the RF random forest scoring system, with scores shown in Table 1. Analysis showed that the system temperature (system_temperature), water content 3 (water_content3), and soil methane temperature (soil_ch4_temp) sensor data were highly correlated with the soil methane dataset labels, indicating that the datasets were reliable and suitable for training and prediction of neural network models.

[0072] Table 1 Random forest scoring system scoring table

[0073] Sensor Name Rating score water_content3 0.470937 system_temperature 0.306430 soil_ch4_temp 0.199622

[0074] Using Time_VAE, the data was augmented and expanded. The original data for system temperature, water content, and soil methane temperature were expanded to 9,700 positive samples and 16,485 negative samples, with a positive-to-negative ratio of 1:1.7. 9,700 positive samples and 16,485 negative samples after Time_VAE augmentation were selected for training, resulting in a trained neural network model. Given the limited number of positive samples in the original dataset, 1,040 original positive samples and 2,087 original negative samples, totaling 3,127 samples, were selected for prediction. These prediction samples were fed into the trained neural network model to obtain gas leak detection results.

[0075] To reflect the different experimental results obtained with datasets composed of different sensor and methane leakage data, we used binary classification accuracy as the evaluation criterion and conducted predictions on the original underground pipeline gas leakage dataset under the same data augmentation ratio, training set positive and negative sample ratio, prediction set positive and negative sample ratio, and network parameters. The detection accuracy is shown in Table 2, with training accuracy exceeding 80% and prediction accuracy exceeding 75%.

[0076] Table 2 Comparison of experimental results of data sets

[0077] Dataset composition Training accuracy (%) Prediction accuracy (%) Soil methane temperature 95.8 75.5 System temperature 88.3 83.0 Water content 3 89.5 80.1 System temperature, water content 3 92.9 79 Water content 3. Soil methane temperature 99.5 87.1

[0078] It is not difficult to find that the present invention measures the change data caused by gas leakage to the surrounding soil through sensors, uses artificial intelligence deep learning to predict indirect detection, and utilizes neural network models to improve the accuracy of underground pipeline gas leakage detection, thereby reducing the consumption of manpower and material resources in actual situations; at the same time, when constructing the neural network model, by expanding the leakage gas data in the underground pipeline gas leakage dataset, the proportion of leakage data in the dataset is increased, the purpose of balancing the positive and negative samples in the dataset is achieved, and the robustness of the dataset is enhanced; a new attention module DCAM is introduced into the neural network model in the present invention, which can effectively reduce the noise in the data, comprehensively extract and fully integrate the information of each feature in the data, so as to improve the detection accuracy.

[0079] A second embodiment of the present invention relates to an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned gas leakage detection method based on underground pipeline network multi-sensor data are implemented.

[0080] A third embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the above-mentioned gas leakage detection method based on multi-sensor data of an underground pipe network are implemented.

[0081] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) that contain computer-usable program code.

[0082] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0083] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction method, which is implemented in the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

Claims

1. A gas leak detection method based on underground pipe network multi-sensor data, characterized in that: The following steps are involved: Acquiring detection data from multiple environmental sensors installed in the soil surrounding the underground pipe network; the multiple environmental sensors are sensors whose correlation with the gas sensor to be measured exceeds a threshold value, as found using an RF random forest scoring system; the RF random forest scoring system uses a Gini index or information gain to evaluate the correlation between the environmental sensors and the gas sensor to be measured, where a larger Gini index or information gain indicates a greater correlation with the gas sensor to be measured; Using the trained generative model to perform data enhancement processing on the detection data of the multiple environmental sensors to obtain time series data; The time series data is input into the trained neural network model to obtain the gas leakage detection result; wherein, The neural network model includes: A feature extraction part, used to extract deep features of the time series data to obtain deep feature data; The attention layer is used to denoise the deep feature data and combine the front and back spatial and channel features of the deep feature data. The attention layer includes: A decomposition module is used to decompose the deep feature data into high-frequency subsequences, medium-frequency subsequences, low-frequency subsequences and residual sequences using complete set empirical mode decomposition; an attention denoising module, configured to denoise the high-frequency subsequence, the medium-frequency subsequence, the low-frequency subsequence and the residual sequence; The summation and reconstruction module is used to sum and reconstruct the denoised high-frequency subsequence, medium-frequency subsequence, low-frequency subsequence and residual sequence to obtain the denoised deep feature data; The LSTM layer is used to process the denoised deep feature data and output the gas leak detection results; The neural network model training includes the following steps: Obtain historical detection data from various environmental sensors installed in the soil around underground pipe networks; Taking the gas sensor to be tested among the multiple environmental sensors as a reference, marking the case where the detection data of the gas sensor to be tested is not 0 as a positive sample, and marking the case where the detection data of the gas sensor to be tested is 0 as a negative sample; Expanding the historical detection data using the generative model, and selecting the expanded historical detection data as a training set; The neural network model is trained using the training set to obtain a trained neural network model.

2. The gas leakage detection method based on underground pipe network multi-sensor data according to claim 1, characterized in that: The generation model includes: An encoder part, for mapping the detection data into a latent space; The decoder part is used to convert latent variables into time series data; The generative model learns the distribution of latent variables, the parameters of the encoder part, and the parameters of the decoder part by maximizing the likelihood function.

3. The gas leakage detection method based on underground pipe network multi-sensor data according to claim 1, characterized in that: The attention denoising module includes a channel attention submodule, a spatial attention submodule and a first wavelet threshold joint denoising module. The input feature is fused with the input feature after passing through the channel attention submodule to obtain a first fused feature. The first fused feature is passed through the first wavelet threshold joint denoising module to obtain a denoised feature. The denoised feature is fused with the denoised feature after passing through the spatial attention submodule to obtain a second fused feature. In the channel attention submodule, the input feature is processed by average pooling and maximum pooling respectively, and then passed through the second wavelet threshold joint denoising module for wavelet threshold denoising. The features are then transferred by the shared fully connected layer, the transferred features are added, and the channel attention feature weights are obtained after passing through the sigmoid activation function.

4. The gas leakage detection method based on underground pipe network multi-sensor data according to claim 1, characterized in that: The neural network model also includes: The Dropout layer is located at the back end of the LSTM layer and is used to prevent the neural network model from overfitting.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the gas leakage detection method based on underground pipe network multi-sensor data as described in any one of claims 1 to 4 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the gas leakage detection method based on underground pipe network multi-sensor data as claimed in any one of claims 1 to 4 are implemented.

Images