A gas leakage identification and positioning method and system suitable for mobile monitoring

Abstract

The application discloses a kind of gas leakage identification positioning method and system suitable for mobile monitoring, belong to deep learning technical field.The method includes: through the microphone array on robot real-time acquisition sound signal on inspection path, input time attention module and frequency domain space attention module, obtain frequency domain space feature sequence after frequency domain space attention enhancement and time feature sequence after time attention enhancement, fusion obtains latent space feature;Based on latent space feature, multi-task gas leakage monitoring is carried out, and prediction classification result, distance prediction value and angle prediction value of gas leakage source are obtained;Effectively capture dynamic acoustic change related to robot position movement, improve identification and positioning accuracy.

Description

Technical Field

[0001] This invention belongs to the field of deep learning, and more specifically, relates to a gas leak identification and location method and system suitable for mobile monitoring. Background Technology

[0002] Gas leak identification and location are key technologies used to determine whether a gas leak has occurred and to pinpoint the leak source. Gas pipeline leaks can lead to serious safety accidents, environmental pollution, and economic risks; therefore, gas leak identification and location technologies play an irreplaceable role in ensuring the safety of gas pipeline transportation. Efficient and accurate gas leak identification and location technologies can detect potential hazards early during pipeline operation and maintenance, shorten accident handling time, and reduce safety risks and economic costs. Furthermore, due to the complexity of the environment and the inefficiency of manual monitoring, stable gas leak identification and location technologies in dynamic inspection scenarios are a critical requirement for applications such as long-distance pipelines and urban gas pipelines. Therefore, developing high-precision, low-false-alarm-rate, and environmentally resistant real-time gas leak identification and location methods has significant academic and engineering application value.

[0003] Currently, gas leak monitoring is typically limited to fixed methods, i.e., fixed installation of sensor equipment for leak identification and location. For example, Chinese invention patent CN118346929A proposes a gas leak detection algorithm based on fiber optic sensors. This algorithm first denoises and decomposes the time-domain signal collected by a distributed fiber optic vibration sensing system, then superimposes, reconstructs, and smooths it to obtain a moving standard deviation. The comparison of this standard deviation with an alarm threshold determines whether a pipeline is leaking. However, this method requires sensors and supporting equipment to be installed along the pipeline, which presents significant difficulties for installation on existing pipelines, limiting identification and location. Chinese invention patent CN118758497A proposes a gas leak monitoring method based on optical equipment. This method uses an infrared camera to collect infrared monitoring video of the monitored location, and then uses image processing and deep learning methods to determine the location data of the leak point. However, infrared image-based methods suffer from disadvantages such as weak environmental adaptability, reliance on the infrared characteristics of gases, and high hardware deployment costs.

[0004] Acoustic methods for detecting gas leaks are widely used due to their strong environmental adaptability and fast response speed. With the development of deep learning technology, a recent research trend is to use neural networks to establish high-dimensional leak features from the audio data collected by acoustic sensors, reducing the impact of noise and improving the accuracy of identification and location. For example, Chinese invention patent CN121010816A proposes a method for multi-beam detection, location, and classification of gas leaks based on deep learning. This method first corrects the orientation of the multi-beam data collected by acoustic sensors and then classifies the leaks using a deep learning model. Chinese invention patent CN120747694A proposes a gas leak early warning method that integrates fault tree analysis and deep learning. This method first extracts the leak features from the acoustic signal and then identifies the leak area using optical images. However, current acoustic methods are mostly designed for fixed monitoring scenarios, and it is difficult to optimize real-time identification and processing efficiency and location accuracy in a coordinated manner. Therefore, they are difficult to adapt to the needs of continuous mobile inspections, and achieving real-time continuous gas leak identification and location remains a major challenge.

[0005] To address the need for continuous and flexible monitoring, the use of robots equipped with acoustic sensors for pipeline inspection has become a recent hot topic. For example, Chinese invention patent CN221348838U proposes a pipeline gas leak detection robot that uses a drone equipped with acoustic and gas detection components to analyze and locate gas leaks. Currently, inspection schemes combining acoustic arrays with robots often directly identify and locate leaks using beamforming algorithms or multimodal physical methods. Even in some methods that incorporate deep learning for leak detection, the robot merely serves as a sensor mount; the neural network does not fully utilize the dynamically changing sound source information acquired by the robot during inspection in real time. Therefore, current acoustic inspection robots still have significant room for improvement in terms of leak identification accuracy and location precision. Summary of the Invention

[0006] In view of the shortcomings of related technologies, the purpose of this invention is to provide a gas leak identification and location method and system suitable for mobile monitoring, which aims to solve the problem that the existing technology does not make full use of the dynamically changing sound source information in the spatiotemporal range in real time, and there is still a lot of room for improvement in the accuracy of leak identification and location.

[0007] To achieve the above objectives, in a first aspect, the present invention provides a gas leak identification and location method suitable for mobile monitoring, comprising: Acquire multi-channel sound signals along the inspection route; Perform gas leak identification and location operations based on sound signals: The sound signal is converted to the frequency domain, and the phase and amplitude components of each frequency domain point are combined to obtain the combined feature. The merged feature is used as input feature A. Input feature A is processed by convolution through the first convolutional layer. The result and input feature A are simultaneously input into the frequency domain spatial attention module for fusion to obtain frequency domain spatial feature E. After performing average pooling along the frequency dimension on the merged features, they are convolved in a second convolutional layer to obtain feature P. Feature P is then input into the temporal attention module to obtain temporal feature Q. The temporal attention module performs a linear transformation on each element of feature P and adds a corresponding bias term to each transformation result. After normalization, an attention weight matrix is ​​obtained. Feature P is then weighted element-wise based on the attention weight matrix and fused with feature P through a residual connection to obtain temporal feature Q. Time characteristics Perform a dimensional transformation to make its dimension the same as that of the frequency domain spatial feature E to obtain the time feature R; sum the elements of the time feature R and the frequency domain spatial feature E to obtain the latent space feature. Multi-task gas leak monitoring is performed based on latent space characteristics, and prediction classification results, distance prediction values, and angle prediction values ​​of gas leak sources are obtained.

[0008] Optionally, the frequency domain spatial attention module and the temporal attention module are trained in the following manner: Obtain the training set; the training set includes: multi-channel sound signal samples collected along the inspection path under gas leak scenarios and non-gas leak scenarios respectively; Perform gas leak identification and localization operations on each sound signal sample in the training set to obtain the corresponding prediction classification results, the distance prediction value of the gas leak source, and the angle prediction value. The frequency domain spatial attention module and the temporal attention module are trained by minimizing the weighted sum of the classification loss, distance prediction loss and angle prediction loss corresponding to each sound signal sample. Among them, classification loss represents the difference loss between the predicted classification result and the true classification result; distance prediction loss represents the difference loss between the predicted distance value and the true distance value; and angle prediction loss represents the difference loss between the predicted angle value and the true angle value.

[0009] Optionally, the step of using the merged feature as input feature A, performing convolution processing on input feature A through a first convolutional layer, and simultaneously inputting the result and input feature A into the frequency domain spatial attention module for fusion to obtain frequency domain spatial feature E includes: The merged feature is used as input feature A. Input feature A is passed through the first convolutional layer and three different convolutional kernels to obtain three new dimensionality-reduced feature maps B, K, and D respectively. Input feature A and feature maps B, K, and D are simultaneously input into the frequency domain spatial attention module. Wherein, B, K, D∈RC×W, C is the number of channels, and W is the feature length. Multiply the transpose of feature map B with feature map K by matrix multiplication, then pass the matrix through a softmax layer to obtain the spatial matrix S, expressed as:

[0010] Where S∈RW×W, This represents the attention influence weight of the i-th position on the j-th position in the frequency domain space. This represents the i-th column of feature map B. This represents the j-th column of feature map K; Multiply the feature map D and the spatial matrix S by matrix multiplication, and then multiply the result by the scale parameter. And perform a residual connection with the input feature A to obtain the frequency domain spatial feature E after spatial attention enhancement, which is expressed as:

[0011] in, It is a learnable parameter with an initial value of 0; Let A and E be the input features at the j-th position in the frequency domain space. j It is a weighted sum of each location and input feature A, which includes the interdependencies between different locations from a full-frequency domain perspective.

[0012] Optionally, the weights of the feature points at the t-th temporal position on the feature map corresponding to the C-th channel in the attention weight matrix. The expression is:

[0013] in, ∈R C×T , ∈R T , Let M represent the feature of the C-th channel in feature P, where the dimension of feature P is C×T; M∈R T×T M is a learnable time transformation matrix used to transform the features of the time series in order to learn the temporal dependencies and dynamic evolution patterns between different time frames. The learnable channel bias represents the degree of preference of the C-th channel for the signal response at different time points, i.e., different physical spatial locations. The expression for the time characteristic Q is:

[0014] Where, Q∈R C×T , ∈[0, 1] is a learnable fusion factor that represents the channel’s dependence on the temporal attention mechanism and reflects the effectiveness of temporal modeling for the channel during training.

[0015] Optionally, the three-head network structure includes a classification head, a distance regression head, and an angle regression head; The classification head is used to obtain the predicted classification result based on the latent space characteristics; the distance regression head is used to obtain the distance prediction value of the gas leak source based on the latent space characteristics; and the angle regression head is used to obtain the angle prediction value of the gas leak source based on the latent space characteristics.

[0016] Optionally, the step of collecting gas leak data and non-gas leak data separately via the robot along the inspection path includes: The robot's microphone array adopts a multi-arm spiral arrangement structure with C' microphones, and the internal acquisition card is used to record sound signals; A gas leak sound source is set up on the inspection path to simulate a gas leak scenario. The robot moves along the inspection path at a uniform speed and continuously collects the leak audio signal. At each moment, the robot's initial position is used as the origin. The angle and distance between the robot and the gas leak sound source at the current moment are calculated and normalized, and the label is set to 1 to obtain the gas leak data. To simulate a non-gas leak scenario, non-leaking sound sources of the same intensity are set up along the inspection path. The robot moves along the inspection path at a uniform speed and continuously collects non-leaking audio signals. At each moment, the robot's initial position is used as the origin. The angle and distance between the robot and the gas leak sound source at the current moment are calculated and normalized, and the label is set to 0 to obtain non-gas leak data.

[0017] Optionally, during the robot's gas leak inspection process, after each preset sampling cycle matching the input of the time attention module is completed, the audio features of the current location and the predicted distance d are recorded. n Angle prediction value θ n The corresponding timestamps are stored in the spatiotemporal feature-location result sequence library and updated, while historical data that exceeds the preset time limit is automatically removed.

[0018] Optionally, the predicted classification result is either a gas leak or a non-gas leak; after obtaining a predicted classification result of a gas leak, the method further includes: Based on the distance prediction value d and the angle prediction value θ, as well as the historical distance d k and from a historical perspective θ k The robot's direction of travel, speed of travel, and microphone sampling frequency are controlled. The process of controlling the robot's direction and speed to correct the positioning results includes: adjusting the predicted distance d from the current location to the gas leak source. n and angle prediction value θ n Distance from history d1~d n-1 and from a historical perspective θ1~θ n-1 Perform calculations and correct the positioning results, including: If the distance to the predicted value d n and angle prediction value θ n If at least one of the changes exceeds the corresponding threshold, the localization is considered not to have converged. This is determined by analyzing historical distances d1~d... n-1 and from a historical perspective θ1~θ n-1 The changing trend is used to fit and correct the current direction of travel in order to correct positioning deviation; If the distance prediction value d is output continuously for a preset number of times n and angle prediction value θ n If the changes are all less than their respective set thresholds, it is determined that the positioning has converged, and the robot will automatically accelerate and move towards the leak source according to the current travel path. Controlling the robot's speed includes: using historical angles θ k Real-time analysis of the dynamic trend of the positioning results, if θ k If the fluctuation amplitude is lower than the set angle threshold, it is determined that the robot has been aligned with the leak source and automatically adjusts the inspection speed to move closer to the leak source. Controlling the robot's microphone sampling frequency includes: through historical distance d k Real-time analysis of the dynamic trend of positioning results: if d k If the sampling frequency shows a continuous decreasing trend, it is determined that the robot is approaching the leak source, and the sampling frequency is increased to accelerate the localization convergence.

[0019] Secondly, the present invention also provides a gas leak identification and location system suitable for mobile monitoring, comprising: a robot and a microphone array and a control module mounted on the robot; The robot moves forward at a uniform speed along the inspection path, collects sound signals in real time through a microphone array, and uploads the sound signal data to the control module for processing. The control module is used to execute the gas leak identification and location method for mobile monitoring as described in any one of the first aspects, and output the prediction classification result, the distance prediction value and the angle prediction value of the gas leak source.

[0020] Compared with the prior art, the above-described technical solutions conceived in this invention can achieve the following beneficial effects: 1. This invention provides a gas leak identification and localization method suitable for mobile monitoring. It processes the real-time sound signals acquired by an acoustic inspection robot using a time-attention module and a frequency-domain spatial attention module to obtain a frequency-domain spatial feature sequence enhanced by frequency-domain spatial attention and a time-domain feature sequence enhanced by time-attention. These are then fused to obtain latent space features. Based on these latent space features, multi-task gas leak monitoring is performed to obtain the predicted classification results, distance prediction values, and angle prediction values ​​of the gas leak source at each moment. This enables real-time, high-precision monitoring and localization of gas leaks, effectively solving the problems of high false alarm rates, large localization errors, poor anti-interference capabilities, and inconvenient equipment installation in existing gas leak detection methods. It also addresses the issue that traditional robot leak inspection methods cannot fully extract the dynamically changing sound source information features acquired by the robot inspection in real time, leaving significant room for improvement in leak identification accuracy and localization precision. This method improves both identification and localization accuracy.

[0021] 2. The present invention provides a gas leak identification and location method suitable for mobile monitoring. During the gas leak inspection process of the robot, after completing one sampling, the audio features of the current location and the distance prediction value d are obtained. n Angle prediction value θ n The corresponding timestamp is stored in the spatiotemporal feature-localization result sequence library and updated. During the robot's movement, the robot can obtain its current position and adjust its direction of movement through the spatiotemporal feature-localization result sequence library, thereby accurately approaching the leak source and ultimately facing the gas leak source directly. This enables more accurate positioning and further visual detection, and timely accident handling. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of a gas leak identification and location method applicable to mobile monitoring in an embodiment of the present invention; Figure 2 This is a schematic diagram of a microphone array and a robot equipped with a microphone array in an embodiment of the present invention; Figure 3 This is a block diagram of a robot system for gas leak monitoring and location in an embodiment of the present invention; Figure 4 This is an algorithm flowchart of a gas leak identification and location method applicable to mobile monitoring in an embodiment of the present invention; Figure 5 This is a flowchart of the frequency domain spatial attention module in an embodiment of the present invention; Figure 6 This is a flowchart of the time attention module in an embodiment of the present invention; Figure 7 This is a flowchart illustrating the precise positioning of the robot in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.

[0025] Example 1 like Figure 1 As shown, a gas leak identification and location method suitable for mobile monitoring includes: The robot collects multi-channel sound signals along the inspection path in real time using a microphone array on its device. Perform gas leak identification and location operations based on sound signals: The sound signal is converted to the frequency domain, and the phase and amplitude components of each frequency domain point are combined to obtain the combined feature. The merged feature is used as input feature A. Input feature A is processed by convolution through the first convolutional layer. The result and input feature A are simultaneously input into the frequency domain spatial attention module for fusion to obtain frequency domain spatial feature E. After performing average pooling along the frequency dimension on the merged features, they are convolved in a second convolutional layer to obtain feature P. Feature P is then input into the temporal attention module to obtain the temporal feature Q enhanced by temporal attention. The temporal attention module performs a linear transformation on each element of feature P and adds a corresponding bias term to each transformation result. After normalization, an attention weight matrix is ​​obtained. Feature P is then weighted element-wise based on the attention weight matrix and fused with feature P through a residual connection to obtain the temporal feature Q. Time characteristics Perform a dimensional transformation to make its dimension the same as that of the frequency domain spatial feature E to obtain the time feature R; sum the elements of the time feature R and the frequency domain spatial feature E to obtain the latent space feature. Multi-task gas leak monitoring is performed based on latent space characteristics, and prediction classification results, distance prediction values, and angle prediction values ​​of gas leak sources are obtained.

[0026] To address the shortcomings of existing technologies, this invention provides a method for inspecting pipeline gas leaks using a robot equipped with a microphone array, combined with deep learning methods to identify and locate leaks. By equipping the robot with an acoustic array, leak audio data can be collected using more flexible monitoring methods. A gas leak identification and localization model is constructed based on a fusion of a one-dimensional residual network and time and frequency domain spatial dual attention modules. This model is highly adaptable to the robot's mobile monitoring scenarios. The frequency domain spatial attention module and the time attention module can fully extract the correlation information between frequency domain positions and the response intensity of each channel to signals at different time points, effectively capturing dynamic acoustic changes related to robot position movement, thereby reducing the false alarm rate and improving localization accuracy.

[0027] In this embodiment of the invention, a robot equipped with a microphone array is used to inspect for gas leaks in pipelines in order to collect audio information. The arrangement of the microphone array and the robot equipped with the microphone array can be found in [reference needed]. Figure 2 The robot system block diagram can be referenced. Figure 3 The robot system includes a drive module and an audio signal acquisition module. The drive module's controller controls the robot's inspection route and speed, while the motors propel the robot forward. The audio signal acquisition module includes a microphone array, power supply, industrial control computer, and display. The microphone array employs a multi-arm spiral arrangement of 30 microphones, with an internal acquisition card for recording sound signals. Simultaneously, a camera at the center of the microphone array captures real-time images of the robot's frontal monitoring. The display connects to the industrial control computer, showing the human-machine interface for backend personnel to inspect the pipeline conditions.

[0028] The training process of the gas leak identification and localization model includes: collecting the dataset; data processing and feature extraction; inputting the extracted features into the model and outputting the prediction results; comparing the predictions and the true labels according to the selected loss function, and then calculating the gradient of each parameter from back to front to update the network parameters.

[0029] The dataset collection includes: collecting gas leak data and non-gas leak data separately by the robot along the inspection path, and setting the label of leak data as 1 and the label of non-leak data as 0 to form training data; the training data also includes the true distance and true angle values ​​of the gas leak source obtained at each location.

[0030] The robot moves at a uniform speed along an inspection path in a real gas leak scenario, continuously collecting leak audio signals. During the robot's movement, each microphone channel of the microphone array collects leak audio at a sampling rate of 50 kHz. Since the robot's inspection speed is slow, the distance traveled within 1000 points (0.02 s) is very small and can be considered fixed; therefore, 1000 points are taken as one moment to record the leak source location. At each moment, with the robot's initial position as the origin, the angle and distance of the sound source at that moment are calculated and normalized, and the leak label is set to 1. Through this process, a series of gas leak data can be obtained.

[0031] With other environmental conditions unchanged, the experiment was repeated by replacing the gas leak sound source with a non-leakage sound source of the same intensity, and the leak label was set to 0 to obtain another set of non-gas leak data.

[0032] Furthermore, each microphone captures one channel of audio signal; therefore, the collected 30-channel audio signal sequence dataset is used as the training dataset. Each sample in the training dataset has C'=30 channels, and each channel microphone in each sample contains 5000 data points, corresponding to a 0.1s audio signal segment, obtained through a sliding window with a step size of 500 data points.

[0033] Data processing and feature extraction specifically include: Each sample in the training data is transformed to the frequency domain by Fast Fourier Transform, and the phase and amplitude components of each frequency domain point are merged to obtain a multi-channel merged feature; the merged feature is a one-dimensional feature sequence in C' channels, and the feature sequence length is W; the dimension of the merged feature is C'×W, and the number of FFT points W is 512.

[0034] For the frequency-domain spatial attention module, the feature extraction output is a feature sequence with C channels and dimensions C'×W. For the temporal attention module, each sample consists of T consecutive time frames; the spectral features of each frame are averaged and pooled along the frequency dimension to obtain the response intensity of each channel at different time points. The input to the temporal attention module is a tensor of dimension C×T. The extracted features are then input to the frequency-domain spatial attention module and the temporal attention module, which are arranged in parallel, respectively. For the overall algorithm flow, and details of the frequency-domain spatial attention module and the temporal attention module, please refer to [references to be inserted here]. Figure 4 , Figure 5 , Figure 6 .

[0035] Specifically, for the frequency domain spatial attention module, the merged feature is used as input feature A. Input feature A is passed through the first convolutional layer and three new feature maps B, K, and D are obtained through three different convolutional kernels. Input feature A and feature maps B, K, and D are simultaneously input into the frequency domain spatial attention module. Wherein, B, K, D∈RC×W, C is the number of channels, and W is the feature length.

[0036] Multiply the transpose of feature map B with feature map K by matrix multiplication, then pass the matrix through a softmax layer to obtain the spatial matrix S, expressed as:

[0037] Where S∈RW×W, This represents the attention influence weight of the i-th position on the j-th position in the frequency domain space. This represents the i-th column of feature map B. This represents the j-th column of feature map K.

[0038] The spatial matrix represents the relationship between two spatial location features in different frequency domains within a single channel. Since each frequency domain feature is inherently affected by the spatial propagation of sound waves from the leaking sound source, the similarity between locations in this matrix implicitly encodes the spatial correlation related to the physical location of the sound source.

[0039] Multiply the feature map D and the spatial matrix S by matrix multiplication, and then multiply the result by the scale parameter. After dimensionality upscaling, a residual connection is performed with the input feature A to obtain a frequency domain spatial feature sequence E of dimension C'×W, expressed as:

[0040] in, It is a learnable parameter with an initial value of 0; Let A and E be the input features at the j-th position in the frequency domain space. j It is a weighted sum of each position and input feature A, which includes the interdependencies between different positions from the perspective of the full frequency domain; E∈RC'×W.

[0041] Multiplying the feature map D and the spatial matrix S allows each frequency domain spatial location to aggregate contextual information from all other locations, effectively modeling how gas leak signals propagate in the array and the responses of different frequency domain spatial locations to the same leak sound source.

[0042] Adding the above product and the input feature A element-wise is equivalent to a residual connection. This residual connection preserves the original spatial structure while integrating the global correlation between frequency domain locations. Since these frequency domain features are formed by the spatial distribution of acoustic energy received by the array, the result of this step implicitly encodes the physical location of the gas leak source, including its distance and angle.

[0043] The weights of the feature points at the t-th temporal position on the feature map corresponding to the C-th channel in the attention weight matrix. The expression is:

[0044] in, ∈R C×T , ∈R T , This represents the feature of the C-th channel in feature P; M∈R T×T M is a learnable time transformation matrix used to transform the features of the time series in order to learn the temporal dependencies and dynamic evolution patterns between different time frames. The learnable channel bias represents the degree of preference of the C-th channel for the signal response at different time points, i.e., different physical spatial locations. The expression for the time characteristic Q is:

[0045] Where, Q∈R C×T , ∈[0, 1] is a learnable fusion factor that represents the channel’s dependence on the temporal attention mechanism and reflects the effectiveness of temporal modeling for the channel during training.

[0046] For the temporal attention module, after average pooling and convolutional layer processing, we obtain the C×T dimension feature P; A linear transformation is performed on the time series of each channel, and a unique bias term is added to each channel to obtain the attention weight matrix. ∈R C×T :

[0047] in, ∈R T Represents the time series of the Cth channel; M∈R T × T It is a learnable time transformation matrix that represents the mutual influence between different time points. It is used to perform feature transformation on time series in order to learn the temporal dependencies and dynamic evolution patterns between different time frames. The learnable channel bias represents the preference of the C-th channel for the signal response at different time points, i.e., different physical spatial locations.

[0048] The calculated attention weight matrix is ​​used to element-wise weight the original time series of each channel. Finally, the features P are integrated through residual connections, and the output is obtained after dimensionality increase. ∈R C'×T :

[0049] in, ∈[0, 1] is a learnable fusion factor that represents the channel’s dependence on the temporal attention mechanism and reflects the effectiveness of temporal modeling for the channel during training.

[0050] Furthermore, the time features Linear interpolation is used to align the feature dimensions, resulting in a time feature sequence R of dimension C'×W. The feature execution elements of the frequency domain spatial attention module and the time attention module are then summed to obtain the fused latent space features. These latent space features are input into a three-head network structure for multi-task gas leak measurement. The classification head outputs a "leakage" or "non-leakage" classification result through a fully connected layer and a softmax activation layer; the distance regression head and angle regression head output the predicted distance and angle of the gas leak source. The three-head network structure includes a classification head, a distance regression head, and an angle regression head. To achieve collaborative optimization across tasks, the three-head network is trained using a weighted mixed loss function. The classification head uses a binary cross-entropy loss function L1, the distance regression head uses a mean squared error loss function L2, and the angle regression head uses a mean squared error loss function L3. The total weighted mixed loss function is expressed as:

[0051] Where λ1, λ2, and λ3 are equilibrium parameters, 0 < λ1, λ2, and λ3 < 1, and λ1 + λ2 + λ3 = 1.

[0052] The three-head network structure is trained using a weighted hybrid loss function, and the network parameters are updated to obtain the trained gas leak identification and localization model.

[0053] Furthermore, robots equipped with microphone arrays were deployed on-site to inspect for gas leaks. The robots moved at a uniform speed along a pre-planned inspection path, continuously collecting sound signals and uploading the data to an onboard computer. The computer analyzed and processed the audio data in real time. The audio data was then used to determine the leak source using a pre-trained offline gas leak identification and location model.

[0054] Once a leak is detected, the robot will first report the leak and alert backend staff. Then, it will proceed to the leak localization phase. During this phase, the leak audio data, along with the distance prediction value d and angle prediction value θ output by the model, will be transmitted in real-time to the controller of the drive module to control the robot's direction, speed, and microphone sampling frequency, further refining the location of the leak source.

[0055] During leak localization, the robot does not merely analyze the current leak audio data, but integrates and analyzes dynamic historical data collected during dynamic inspections, maintaining a "spatiotemporal feature-localization result sequence library." The dual-attention network can fuse historical and current acoustic features across time dimensions, and utilize the results and trends of historical distance d and historical angle θ to optimize the current localization result, achieving dynamic iterative localization.

[0056] During the robot's gas leak inspection process, after each preset sampling cycle matching the input of the time attention module, the audio features of the current location and the predicted distance d are recorded. n Angle prediction value θ n The corresponding timestamps are stored in the spatiotemporal feature-localization result sequence library and updated, while historical data exceeding the preset time limit is automatically removed. The robot system's computer maintains the "spatiotemporal feature-localization result sequence library" in real time to ensure the real-time performance and effectiveness of localization calculations. When the network outputs the distance d of the current location... n and angle θ n Then, the predicted distance d from the current location to the gas leak source, output by the gas leak identification and location model, is used. n and angle prediction value θ n Distance from history d1~d n-1 and from a historical perspective θ1~θ n-1 Perform calculations and correct the positioning results, including: If the distance to the predicted value d n and angle prediction value θ n If at least one of the changes exceeds the corresponding threshold, the localization is considered not to have converged. This is determined by analyzing historical distances d1~d... n-1 and from a historical perspective θ1~θ n-1 The changing trend is used to fit and correct the current direction of travel in order to correct positioning deviation; If the distance prediction value d is output continuously for a preset number of times n and angle prediction value θ n If the changes are all less than their respective set thresholds, it is determined that the positioning has converged, and the robot will automatically accelerate and move towards the leak source according to the current travel path. Controlling the robot's speed includes: using historical angles θ k Real-time analysis of the dynamic trend of the positioning results, if θk If the fluctuation amplitude is lower than the set angle threshold, it is determined that the robot has been aligned with the leak source and automatically adjusts the inspection speed to move closer to the leak source. Controlling the robot's microphone sampling frequency includes: through historical distance d k Real-time analysis of the dynamic trend of positioning results: if d k If the sampling frequency shows a continuous decreasing trend, it is determined that the robot is approaching the leak source, and the sampling frequency is increased to accelerate the localization convergence.

[0057] Therefore, by adjusting its orientation and distance from the leak source, the robot can ultimately directly face the gas leak source, achieving more precise positioning and further visual inspection, and enabling timely accident handling. The process of controlling the robot's precise positioning using the "spatiotemporal feature-positioning result sequence library" is as follows: Figure 7 As shown.

[0058] This invention provides a gas leak identification and localization method and system suitable for mobile monitoring. It constructs a gas leak identification and localization model based on a fusion of a one-dimensional residual network, a temporal attention module, and a frequency-domain spatial attention module. Based on the sound signals acquired in real-time by an acoustic inspection robot, it performs real-time, high-precision monitoring and localization of gas leaks. This effectively solves the problems of high false alarm rates, large localization errors, poor anti-interference capabilities, and inconvenient equipment installation in existing gas leak detection methods. It also addresses the issue that traditional robot leak inspection methods cannot fully extract the dynamically changing sound source information features acquired by the robot inspection in real time, thus improving the accuracy of identification and localization.

[0059] Example 2 This invention also provides a gas leak identification and location system suitable for mobile monitoring, comprising: a robot and a microphone array and control module mounted on the robot; The robot moves forward at a uniform speed along the inspection path, collects sound signals in real time through a microphone array, and uploads the sound signal data to the control module for processing. The control module is used to execute the gas leak identification and location method for mobile monitoring as described in any one of Embodiment 1, and outputs the prediction classification result, the distance prediction value and the angle prediction value of the gas leak source.

[0060] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and locating gas leaks suitable for mobile monitoring, characterized in that, include: Acquire multi-channel sound signals along the inspection route; Perform gas leak identification and location operations based on sound signals: The sound signal is converted to the frequency domain, and the phase and amplitude components of each frequency domain point are combined to obtain the combined feature. The merged feature is used as input feature A. Input feature A is processed by convolution through the first convolutional layer. The result and input feature A are simultaneously input into the frequency domain spatial attention module for fusion to obtain frequency domain spatial feature E. After performing average pooling along the frequency dimension on the merged features, they are then convolved in the second convolutional layer to obtain feature P. The feature P is input into the temporal attention module to obtain the temporal feature Q. The temporal attention module is used to perform a linear transformation on each element of the feature P and add a corresponding bias term to each transformation result. After normalization, the attention weight matrix is ​​obtained. The feature P is weighted element by element based on the attention weight matrix and fused with the feature P through residual connection to obtain the temporal feature Q. Time characteristics Perform a dimensional transformation to make its dimension the same as that of the frequency domain spatial feature E to obtain the time feature R; sum the elements of the time feature R and the frequency domain spatial feature E to obtain the latent space feature. Multi-task gas leak monitoring is performed based on latent space characteristics, and prediction classification results, distance prediction values, and angle prediction values ​​of gas leak sources are obtained.

2. The gas leak identification and location method for mobile monitoring as described in claim 1, characterized in that, The frequency domain spatial attention module and the temporal attention module are trained in the following way: Obtain the training set; the training set includes: multi-channel sound signal samples collected along the inspection path in gas leak scenarios and non-gas leak scenarios respectively; Perform gas leak identification and localization operations on each sound signal sample in the training set to obtain the corresponding prediction classification results, the distance prediction value of the gas leak source, and the angle prediction value. The frequency domain spatial attention module and the temporal attention module are trained by minimizing the weighted sum of the classification loss, distance prediction loss and angle prediction loss corresponding to each sound signal sample. Among them, classification loss represents the difference loss between the predicted classification result and the true classification result; distance prediction loss represents the difference loss between the predicted distance value and the true distance value; and angle prediction loss represents the difference loss between the predicted angle value and the true angle value.

3. The gas leak identification and location method for mobile monitoring as described in claim 1, characterized in that, The process involves using the merged feature as input feature A, performing convolution processing on input feature A through a first convolutional layer, and simultaneously inputting the result and input feature A into a frequency domain spatial attention module for fusion to obtain frequency domain spatial feature E, including: The merged feature is used as input feature A. Input feature A is passed through the first convolutional layer and three different convolutional kernels to obtain three new dimensionality-reduced feature maps B, K, and D respectively. Input feature A and feature maps B, K, and D are simultaneously input into the frequency domain spatial attention module. Wherein, B, K, D∈RC×W, C is the number of channels, and W is the feature length. Multiply the transpose of feature map B with feature map K by matrix multiplication, then pass the matrix through a softmax layer to obtain the spatial matrix S, expressed as: Where S∈RW×W, This represents the attention influence weight of the i-th position on the j-th position in the frequency domain space. This represents the i-th column of feature map B. This represents the j-th column of feature map K; Multiply the feature map D and the spatial matrix S by matrix multiplication, and then multiply the result by the scale parameter. And input feature A for residual connection to obtain the frequency domain spatial feature E after spatial attention enhancement, the expression is: in, It is a learnable parameter with an initial value of 0; Let A and E be the input features at the j-th position in the frequency domain space. j It is a weighted sum of each location and input feature A, which includes the interdependencies between different locations from a full-frequency domain perspective.

4. The gas leak identification and location method for mobile monitoring as described in claim 1, characterized in that, The weights of the feature points at the t-th temporal position on the feature map corresponding to the C-th channel in the attention weight matrix. The expression is: in, ∈R C×T , ∈R T , Let M represent the feature of the C-th channel in feature P, where the dimension of feature P is C×T; M∈R T×T M is a learnable time transformation matrix used to transform the features of the time series in order to learn the temporal dependencies and dynamic evolution patterns between different time frames. The learnable channel bias represents the degree of preference of the C-th channel for the signal response at different time points, i.e., different physical spatial locations. The expression for the time characteristic Q is: Where, Q∈R C×T , ∈[0, 1] is a learnable fusion factor that represents the channel’s dependence on the temporal attention mechanism and reflects the effectiveness of temporal modeling for the channel during training.

5. The gas leak identification and location method for mobile monitoring as described in claim 1, characterized in that, The three-head network structure includes a classification head, a distance regression head, and an angular regression head; The classification head is used to obtain the predicted classification result based on the latent space characteristics; the distance regression head is used to obtain the distance prediction value of the gas leak source based on the latent space characteristics; and the angle regression head is used to obtain the angle prediction value of the gas leak source based on the latent space characteristics.

6. The gas leak identification and location method for mobile monitoring as described in claim 1, characterized in that, The process of collecting gas leak data and non-gas leak data separately via a robot along the inspection path includes: The robot's microphone array adopts a multi-arm spiral arrangement structure with C' microphones, and the internal acquisition card is used to record sound signals; A gas leak sound source is set up on the inspection path to simulate a gas leak scenario. The robot moves along the inspection path at a uniform speed and continuously collects the leak audio signal. At each moment, the robot's initial position is used as the origin. The angle and distance between the robot and the gas leak sound source at the current moment are calculated and normalized, and the label is set to 1 to obtain the gas leak data. To simulate a non-gas leak scenario, non-leaking sound sources of the same intensity are set up along the inspection path. The robot moves along the inspection path at a uniform speed and continuously collects non-leaking audio signals. At each moment, the robot's initial position is used as the origin. The angle and distance between the robot and the gas leak sound source at the current moment are calculated and normalized, and the label is set to 0 to obtain non-gas leak data.

7. The gas leak identification and location method for mobile monitoring as described in claim 1, characterized in that, During the robot's gas leak inspection process, after each preset sampling cycle matching the input of the time attention module, the audio features of the current location and the predicted distance d are recorded. n Angle prediction value θ n The corresponding timestamps are stored in the spatiotemporal feature-location result sequence library and updated, while historical data that exceeds the preset time limit is automatically removed.

8. The gas leak identification and location method for mobile monitoring as described in claim 1, characterized in that, The predicted classification result is either a gas leak or a non-gas leak; after obtaining a predicted classification result of gas leak, the following is also included: Based on the distance prediction value d and the angle prediction value θ, as well as the historical distance d k and from a historical perspective θ k The robot's direction of travel, speed of travel, and microphone sampling frequency are controlled. The process of controlling the robot's direction and speed to correct the positioning results includes: adjusting the predicted distance d from the current location to the gas leak source. n and angle prediction value θ n Distance from history d1~d n-1 and from a historical perspective θ1~θ n-1 Perform calculations and correct the positioning results, including: If the distance to the predicted value d n and angle prediction value θ n If at least one of the changes exceeds the corresponding threshold, the localization is considered not to have converged. This is determined by analyzing historical distances d1~d... n-1 and from a historical perspective θ1~θ n-1 The changing trend is used to fit and correct the current direction of travel in order to correct positioning deviation; If the distance prediction value d is output continuously for a preset number of times n and angle prediction value θ n If the changes are all less than their respective set thresholds, it is determined that the positioning has converged, and the robot will automatically accelerate and move towards the leak source according to the current travel path. Controlling the robot's speed includes: using historical angles θ k Real-time analysis of the dynamic trend of the positioning results, if θ k If the fluctuation amplitude is lower than the set angle threshold, it is determined that the robot has been aligned with the leak source and automatically adjusts the inspection speed to move closer to the leak source. Controlling the robot's microphone sampling frequency includes: through historical distance d k Real-time analysis of the dynamic trend of positioning results: if d k If the sampling frequency shows a continuous decreasing trend, it is determined that the robot is approaching the leak source, and the sampling frequency is increased to accelerate the localization convergence.

9. A gas leak identification and location system suitable for mobile monitoring, characterized in that, include: The robot, along with the microphone array and control module mounted on it; The robot moves forward at a uniform speed along the inspection path, collects sound signals in real time through a microphone array, and uploads the sound signal data to the control module for processing. The control module is used to execute the gas leak identification and location method for mobile monitoring as described in any one of claims 1-8, and output the prediction classification result, the distance prediction value and the angle prediction value of the gas leak source.

Images