Intelligent comprehensive management method and system for natural gas pipeline protection

Abstract

The invention relates to the technical field of artificial intelligence and data processing, and discloses an intelligent comprehensive management method and system for natural gas pipeline protection. The method comprises the steps of obtaining a natural gas pipeline sound pressure signal; performing time-frequency characteristic decomposition on the sound pressure signal of the natural gas pipeline by adopting a self-adaptive basis function decomposition method to obtain a time-frequency characteristic matrix; constructing a leakage identification model, utilizing the leakage identification model to perform cascade multi-scale convolutional coding, time-frequency feature purification, dual-path feature fusion and leakage sensitive feature enhancement on the time-frequency feature matrix in sequence, and performing dynamic classification decision according to the obtained leakage sensitive features to obtain leakage level probability distribution; and performing comprehensive management and control on the natural gas pipeline according to the leakage grade probability distribution. According to the method, the leakage detection accuracy under the class imbalance scene can be improved, and real-time comprehensive management and control of the natural gas pipeline are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent integrated management method and system for the protection of natural gas pipelines. Background Technology

[0002] Natural gas, as one of the most important clean energy sources, is widely used in urban gas transmission, industrial production, and power generation. However, with the continuous expansion of pipeline transportation and the gradual aging of pipelines, pipeline leaks are becoming increasingly prominent. Natural gas is flammable, explosive, and easily dispersed; once a leak occurs, it not only causes energy waste and economic losses but can also lead to serious safety accidents and environmental pollution. Therefore, how to achieve real-time monitoring and early leak identification of natural gas pipelines has become a key issue in energy security and public safety management.

[0003] Leak detection methods primarily rely on pressure monitoring, flow balancing, and conventional acoustic feature extraction, but these methods have certain limitations. Pressure and flow-based detection often suffers from response lag, making it difficult to detect small-aperture or early micro-leaks in a timely manner. While conventional acoustic methods offer some real-time performance, commonly used feature extraction methods such as short-time Fourier transform, wavelet transform, or Mel-frequency cepstral spectra rely on fixed time-frequency resolution, making it difficult to simultaneously capture transient pulses and broadband resonance features. In noisy environments, signals are easily masked, leading to a high rate of missed micro-leaks. Furthermore, most existing deep learning detection models employ single-scale convolutional or recurrent neural network structures, failing to adequately consider the multi-scale features of different leakage modes and lacking mapping to acoustic physics, resulting in insufficient model interpretability. Simultaneously, leakage data exhibits significant class imbalance, and fixed decision boundaries make models prone to bias towards normal operating conditions, further weakening their ability to identify micro-leaks.

[0004] The above technical solution has the following problems that still need to be addressed:

[0005] 1. Fixed-window or conventional wavelets have difficulty capturing both millisecond-level transient and sustained broadband features simultaneously, and often only focus on amplitude while ignoring phase, resulting in the dilution of key leakage pulses and harmonic structures.

[0006] 2. Single-scale convolutions or simple features are easily confused between different leakage modes, lack mapping with physical acoustic quantities, and have insufficient model interpretability and generalization.

[0007] 3. Channel attention or attention only in the time / frequency domain cannot model the joint time-frequency correlation, resulting in residual background noise and insufficient prominence of sparse cluster leakage features.

[0008] 4. Fixed Softmax temperature and single cross-entropy tend to favor the majority class, micro-leakage is ignored. In addition, feature denoising and classification are not jointly optimized, and feature distortion or residual noise is prone to occur at low signal-to-noise ratios, resulting in insufficient robustness. Summary of the Invention

[0009] To address the aforementioned shortcomings in the existing technology, this invention provides an intelligent integrated management method and system for natural gas pipeline protection.

[0010] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0011] In a first aspect, the present invention proposes an intelligent integrated management method for the protection of natural gas pipelines, comprising the following steps:

[0012] Acquire acoustic pressure signals from natural gas pipelines;

[0013] An adaptive basis function decomposition method is used to perform time-frequency feature decomposition on the acoustic pressure signal of the natural gas pipeline, and the time-frequency feature matrix is ​​obtained.

[0014] A leakage identification model is constructed. The model is then used to perform concatenated multi-scale convolutional coding, time-frequency feature purification, dual-path feature fusion, and leakage-sensitive feature enhancement on the time-frequency feature matrix. Based on the obtained leakage-sensitive features, dynamic classification decisions are made to obtain the leakage level probability distribution.

[0015] Comprehensive management and control of natural gas pipelines based on the probability distribution of leakage levels.

[0016] Secondly, the present invention proposes an intelligent integrated management system for natural gas pipeline protection, comprising:

[0017] The signal acquisition module is used to acquire the acoustic pressure signal of the natural gas pipeline;

[0018] The feature decomposition module is used to perform time-frequency feature decomposition on the acoustic pressure signal of the natural gas pipeline using an adaptive basis function decomposition method to obtain the time-frequency feature matrix.

[0019] The leakage identification module is used to build a leakage identification model. The leakage identification model is used to perform concatenated multi-scale convolutional coding, time-frequency feature purification, dual-path feature fusion and leakage-sensitive feature enhancement on the time-frequency feature matrix in sequence. Based on the obtained leakage-sensitive features, dynamic classification decision is made to obtain the leakage level probability distribution.

[0020] The integrated management and control module is used for the integrated management and control of natural gas pipelines based on the probability distribution of leakage levels.

[0021] The present invention has the following beneficial effects:

[0022] (1) The present invention adopts a parameterized adaptive basis function, sets it as a differentiable parameter and optimizes it end-to-end with gradient descent, which significantly improves the energy focusing and signal-to-noise ratio of the leakage sensitive frequency band under strong noise, and enhances the detectability of transient pulses through phase alignment.

[0023] (2) The present invention constructs a cascaded dilated convolution group that covers the receptive field from transient to broadband resonance, and establishes a mapping between the receptive field time and the acoustic propagation time, taking into account both micro-leakage and severe leakage and having physical interpretability.

[0024] (3) Based on the product weight of time-gated and frequency-gated, the present invention forms a weight matrix for time-frequency joint focusing, which suppresses time-frequency uniform noise and highlights oblique cluster harmonics and sparse leakage characteristics.

[0025] (4) This invention uses a combination of dual-path feature interaction and differentiable spectrum enhancement operator, and uses hybrid loss to train and optimize model parameters, which significantly improves micro-leakage recall and overall accuracy in class imbalance scenarios. Attached Figure Description

[0026] Figure 1 A schematic diagram of a smart integrated management method for natural gas pipeline protection;

[0027] Figure 2 This is a schematic diagram of the short-time Fourier transform time-frequency representation.

[0028] Figure 3 This is a schematic diagram of the adaptive time-frequency feature decomposition of the present invention;

[0029] Figure 4 This is a schematic diagram of the original multi-scale features;

[0030] Figure 5 This is a schematic diagram illustrating the purification characteristics;

[0031] Figure 6 A comparison chart showing the leak detection accuracy of different leak detection methods under different signal-to-noise ratio environments;

[0032] Figure 7 A comparison chart showing the detection performance of different leak detection methods for different leak types;

[0033] Figure 8 This is a comparison chart of the detection performance of different leak detection methods under different leak orifice sizes. Detailed Implementation

[0034] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0035] like Figure 1 As shown in the figure, an intelligent integrated management method for natural gas pipeline protection provided by an embodiment of the present invention includes the following steps S1 to S4:

[0036] S1. Obtain the acoustic pressure signal from the natural gas pipeline;

[0037] In an optional embodiment of the present invention, step S1 involves collecting and labeling monitoring data for the natural gas pipeline. A high-sensitivity acoustic sensor array is deployed along the natural gas pipeline to continuously collect pipeline sound pressure signals at a sampling rate of 16kHz, forming a raw sound pressure signal dataset. The data collection covers three operating conditions: normal pipeline operation, minor leakage, and severe leakage. For each operating condition, no less than 5,000 samples are collected, and each sample has 16,000 sampling points.

[0038] The labeling process employs a dual verification mechanism: firstly, the start and end times and leakage levels are marked by manual listening combined with professional voiceprint analysis software; secondly, tracer gas is injected at the point of sudden pressure change in the pipeline, and the authenticity of the leakage event is verified using a gas concentration detector. The labeling categories are divided into three categories: Category 0 (normal operating conditions), Category 1 (minor leaks: leak orifice diameter ≤ 2mm), and Category 2 (serious leaks: leak orifice diameter > 2mm).

[0039] The final training dataset contains 15,000 labeled samples, of which 60% are normal operating conditions, 25% are micro-leakage, and 15% are severe leakage, ensuring coverage of different combinations of background noise intensity and pipeline pressure.

[0040] S2. The time-frequency feature matrix of the acoustic pressure signal of the natural gas pipeline is obtained by using the adaptive basis function decomposition method.

[0041] In an optional embodiment of the present invention, step S2 involves performing video feature decomposition using an adaptive time-frequency decomposition module.

[0042] In the process of acoustic signal acquisition in natural gas pipelines, the signals acquired by sound sensors often contain a mixture of strong background noise and weak leakage signals. Leakage acoustic patterns exhibit non-stationary and sparse characteristics in the time-frequency domain. Conventional short-time Fourier transforms, using a fixed window length, cannot simultaneously capture the transient leakage pulses and the broadband characteristics of continuous leakage, leading to the loss of crucial information. This invention dynamically adjusts the time-frequency resolution through an adaptive basis function decomposition method, enhancing the energy concentration in the leakage-sensitive frequency band, thereby improving the signal-to-noise ratio. The specific steps are as follows:

[0043] 1) Constructing adaptive basis functions

[0044] By employing parameterized basis functions, the time-domain center parameter, time-domain width parameter, center frequency parameter, phase parameter, and bandwidth sharpness control parameter are dynamically adjusted to match different leakage acoustic signature characteristics. These parameters are optimized using a gradient descent-based backpropagation algorithm, enabling the basis functions to enhance the energy concentration in the leakage-sensitive frequency band, thereby improving the signal-to-noise ratio, as expressed below:

[0045]

[0046] in, Indicates the first An adaptive basis function is used to extract specific time-frequency features in the signal that are related to leakage; It represents a time variable and is used to describe the process of a signal changing over time; Represents a frequency variable, used to describe the distribution of a signal in the frequency domain; Indicates the index of the basis function. ; This is the total number of basis functions, preferably set to 64; Indicates the first Each adaptive basis function time The value at a frequency is used to extract specific time-frequency features in the signal related to leakage; its parameter set... Optimization was performed using a gradient descent-based backpropagation algorithm to match leakage characteristics; It is a natural exponential function; Indicates the first The time-domain center parameter of each basis function controls the position of the basis function on the time axis to align with the start time of the leakage event; Indicates the first The time-domain width parameter of each basis function controls the decay rate of the basis function in the time direction, thus affecting the adaptability of the time resolution. Indicates the first The center frequency parameter of each basis function controls the main energy distribution region of the basis function in the frequency domain, and is used to match the leakage sensitive frequency band; Indicates the first The phase parameters of each basis function are used to adjust the initial phase of the basis function in order to better match the phase characteristics of the leakage signal; This represents the band selection function, used to control the attenuation characteristics of the basis functions in the frequency domain and enhance the energy concentration of a specific frequency band; This is a band selection function used to control the attenuation characteristics of the basis functions in the frequency domain, thereby enhancing the energy concentration of a specific frequency band. Its calculation method is expressed as follows: ; Indicates the first The bandwidth sharpness control parameter of each basis function is used to adjust the steepness of the band selection function, affecting the adaptability of the frequency domain resolution.

[0047] 2) Perform time-frequency feature decomposition

[0048] The original sound pressure signal is multiplied and summed element-wise with the adaptive basis function to generate a time-frequency feature matrix, highlighting the time-frequency energy distribution in the leakage-sensitive frequency band, as shown below:

[0049]

[0050] in, Let be the time-frequency feature matrix, representing the time-frequency feature matrix obtained after adaptive basis function decomposition, with dimension . This is used to highlight the time-frequency energy distribution in the leakage-sensitive frequency band; The time step is the number of steps from the original sound pressure signal. The result is obtained through framing operations. For example, if the original signal length is 16,000 points, and a frame length of 256 points and a frame shift of 128 points are used, then... ; Indicates the rounding operation; This indicates element-wise multiplication; This represents the original sound pressure signal, which belongs to the real number vector space. It includes pipeline vibration noise and environmental interference components; This represents a real vector space with a dimension of 16000. That is, the original sound pressure signal is a 16000-dimensional real vector that includes pipeline vibration noise and environmental interference components.

[0051] 3) Optimize basis function parameters

[0052] The parameter set of the basis functions is optimized using a backpropagation algorithm based on gradient descent. This concentrates energy in the leakage-sensitive frequency band, thereby improving the signal-to-noise ratio;

[0053] In the specific implementation, stochastic gradient descent is adopted, the learning rate is set to 0.001, the gradient is calculated by the chain rule, the loss function is maximized during training, and iterative optimization is used to concentrate energy on the leakage-sensitive frequency band.

[0054] It should be noted that the parameter set of the adaptive basis function The system is constructed as a differentiable optimization object, and these parameters are dynamically optimized using gradient descent backpropagation. This allows the basis functions to adaptively match the non-stationary characteristics of the leakage sound pattern, particularly the phase parameter. Breaking away from the limitations of conventional time-frequency analysis that only focuses on amplitude, this method significantly enhances the coherent superposition effect of transient pulses through phase alignment, resolving the contradiction between transient pulses and continuous leakage in fixed-window-length time-frequency analysis. This is achieved when the time-domain width parameter of the basis functions... It can capture millisecond-level leakage pulses during automatic contraction, while the bandwidth sharpness control parameter of the basis function... When adjusting the frequency band sharpness, it can also cover wideband resonance, and its dynamic adaptability allows weak leakage to exhibit an energy focusing effect in a strong noise background.

[0055] This embodiment compares the effects of different time-frequency feature decomposition methods through visualization, analyzing the superiority of the method of the present invention in extracting leakage features. Time-frequency analysis decomposes a signal in the time and frequency dimensions to identify the start time and frequency band distribution of a leakage event. This experiment compares the short-time Fourier transform with the adaptive time-frequency feature decomposition method of the present invention, such as... Figure 2 and Figure 3 As shown. Figure 2 and Figure 3 The results are presented using a heatmap, with the horizontal axis representing time (in seconds) and the vertical axis representing frequency (in Hertz). The intensity of the color indicates the signal amplitude, with higher amplitude indicating more concentrated energy. The short-time Fourier transform results show that the leakage signal is somewhat ambiguous in both time and frequency, with dispersed energy, making it difficult to clearly distinguish the leakage area. However, the time-frequency diagram of the method presented in this invention shows that the energy in the leakage area (approximately 0.3 to 0.7 seconds) is significantly enhanced and concentrated in the high-frequency band (around 800 Hz and 1200 Hz), forming a clear time-frequency pattern. This indicates that the method presented in this invention can more accurately locate the time and frequency characteristics of leakage events. Experimental results show that adaptive basis function decomposition, by dynamically adjusting parameters to match the non-stationary characteristics of the leakage acoustic pattern, enhances energy focusing in sensitive frequency bands, thereby effectively suppressing background noise and preserving key information. This provides higher-quality feature input for subsequent leakage identification and improves the overall detection reliability.

[0056] S3. Construct a leakage identification model. Using the leakage identification model, perform concatenated multi-scale convolutional coding, time-frequency feature purification, dual-path feature fusion, and leakage-sensitive feature enhancement on the time-frequency feature matrix in sequence. Then, make dynamic classification decisions based on the obtained leakage-sensitive features to obtain the leakage level probability distribution.

[0057] In an optional embodiment of the present invention, step S3 involves constructing a leakage identification model for online feature extraction and identification. The specific steps are as follows:

[0058] S301. Multi-scale convolutional feature encoding is performed using cascaded multi-scale convolutional modules.

[0059] Leakage acoustic signatures exhibit multi-scale characteristics. Micro-leakage presents as narrowband transient pulses, while severe leakage includes broadband resonances. Conventional single-scale convolutions can only capture features within a fixed receptive field, making it difficult to cover different leakage modes. This invention constructs cascaded dilated convolutional groups to fuse local details and global context to capture multi-scale leakage features. The specific steps are as follows:

[0060] 1) Set multi-scale expansion rate parameters

[0061] This embodiment determines a set of dilatation rate values ​​based on the timescale range of acoustic events to cover different receptive fields, ensuring the capture of multi-scale features ranging from short transient pulses to long-term broadband resonances; the dilatation rate set is defined as follows. The expansion rate is , which corresponds to the dilation factor of the convolution kernel, where This represents the standard convolution operation, with the smallest receptive field, suitable for short-term transient features. This indicates a dilated convolution operation, where the receptive field increases progressively, making it suitable for capturing broadband resonance features over longer time periods.

[0062] In practical implementation, The corresponding time scale is approximately 24ms to 248ms. For example, assuming a sampling rate of 16kHz, a time step interval of 8ms for the time-frequency features, and a kernel size of 3, the receptive field time of dilated convolution is... , Given the time step interval, the calculation yields:

[0063] Time, experiencing the wild = ;

[0064] Time, experiencing the wild = ;

[0065] Time, experiencing the wild = ;

[0066] Time, experiencing the wild = .

[0067] 2) Calculate multi-scale convolutional features

[0068] This embodiment performs multi-scale dilated convolution on the input time-frequency feature matrix, sums the convolution results with different dilation rates, and applies batch normalization and a modified linear unit activation function to output a multi-scale feature matrix. By fusing leakage features from different receptive fields, the model's ability to capture multi-scale voiceprints is enhanced, as shown below:

[0069]

[0070] in, For the first Multi-scale feature matrices of layers represent the first layer of cascaded multi-scale convolutional modules. The feature matrix output by the multi-scale convolutional layer has a dimension of It integrates leakage characteristics under multi-scale receptive fields; This indicates that the activation function of the modified linear unit is used to enhance the expressive power of the model through nonlinear transformation; This indicates batch normalization, which is used to accelerate the model training process and improve training stability. The cascaded multi-scale convolutional module is represented by the first... In-layer expansion rate The convolution kernel weight matrix has dimensions of , used to extract features under a specific receptive field, are trainable parameters; This represents the dilation convolution operation; The number of channels output by the current layer convolution operation of the cascaded multi-scale convolution module; The number of channels is the input to the current layer convolution operation of the cascaded multi-scale convolution module. For the first layer, .

[0071] It should be noted that the total number of layers in the cascaded multi-scale convolutional modules It can be set to 3 layers, each containing a set of expansion rates. The convolution operation uses the output of the previous layer as the input of the current layer, and the input of the first layer is... .

[0072] It should also be noted that the jump expansion rate design allows each path to focus on a specific scale. Corresponding to a 24ms transient pulse, For a 248ms wideband resonance, the discretized calculation method involves summation operations. This generates a cross-scale feature complementarity effect, improving the temporal positioning accuracy of transient pulses and the bandwidth coverage of wideband resonance. Through batch normalization, synergistic enhancement is achieved, resolving the feature confusion problem between micro-leakage and severe leakage in single-scale convolution. More importantly, the receptive field time... The computational model maps the physical propagation time of sound waves to convolution operations, giving the deep learning model an interpretable acoustic physics basis.

[0073] S302. Obtain purification features using a time-frequency feature purification module.

[0074] Background noise forms a continuous distribution in the time-frequency domain, while leakage features are sparsely clustered. Conventional attention mechanisms ignore the time-frequency correlation, resulting in noise residue and an inability to effectively separate leakage features from background interference. This invention employs a dual-domain gating unit to extract salient features from both the time and frequency dimensions, and constructs a joint time-frequency weight matrix through outer product operations to suppress incoherent noise. The specific steps are as follows:

[0075] 1) Calculate the time-domain gating weights

[0076] This embodiment compresses the frequency dimension information of the multi-scale feature matrix by average pooling along the frequency axis, and generates a gated weight matrix in the time dimension through a fully connected layer, thereby highlighting the regions with significant leakage features in the time dimension, as shown below:

[0077]

[0078] in, This represents the time-domain gated weight matrix, with dimension 1. , used to highlight areas with significant leakage characteristics over time; The time step is the number of steps, and the column number is the number of columns in the multi-scale feature matrix. This represents the Sigmoid activation function, which compresses the output value to the (0,1) interval; This represents the weight matrix of the first fully connected layer, used to learn the feature importance in the time dimension; This represents the average pooling operation along the frequency axis. Term representation pairs of characteristic matrix Compression is performed on the frequency dimension to extract global information in the time dimension; It is a multi-scale feature matrix with dimension . The characterization of the cascaded multi-scale convolutional module The feature matrix output by the multi-scale convolutional layer, i.e. ; This represents the total number of layers in the cascaded multi-scale convolutional modules.

[0079] 2) Calculate the frequency domain gating weights

[0080] This embodiment compresses the time dimension information of the multi-scale feature matrix by average pooling along the time axis, and generates a gated weight matrix in the frequency dimension through a fully connected layer, emphasizing the frequency bands with significant leakage characteristics in the frequency dimension, represented as:

[0081]

[0082] in, This represents the frequency domain gated weight matrix, with dimension 1. , used to emphasize the frequency band where leakage characteristics are significant in the frequency dimension; This represents the weight matrix of the second fully connected layer, used to learn the feature importance in the frequency dimension; This represents the average pooling operation along the time axis. Term representation of the input feature matrix Compression is performed in the time dimension to extract global information in the frequency domain dimension.

[0083] 3) Generate the time-frequency joint weight matrix

[0084] In this embodiment, the time-domain gated weight matrix and the frequency-domain gated weight matrix are multiplied by an outer product to construct a time-frequency joint weight matrix, which is used to capture the joint correlation between the time and frequency dimensions, and is represented as follows:

[0085]

[0086] in, This represents the time-frequency joint weight matrix, with dimension . This is used to simultaneously model the saliency of features in both the time and frequency dimensions; This represents the outer product operation, which expands two vectors into a two-dimensional matrix to capture the joint time-frequency correlation.

[0087] 4) Apply gated weighted purification features

[0088] The time-frequency joint weight matrix is ​​multiplied element-wise with the original multi-scale feature matrix to output a purified feature matrix, thereby enhancing leakage-related features and suppressing incoherent noise, as shown below:

[0089]

[0090] in, The purified feature matrix represents the feature matrix after time-frequency purification, with dimensions of . .

[0091] It should be noted that in the time-domain gated weight matrix While obtaining the temporal location of transient events, the frequency domain gated weight matrix By enhancing the sensitive frequency band, the outer product operation of the two produces a "focusing lens" effect on the time-frequency plane, which can automatically enhance the leakage harmonic clusters that are distributed obliquely, while suppressing the time-frequency uniform distribution characteristics of background noise.

[0092] This embodiment performs a visual analysis of the time-frequency purification module's effect, using a heatmap to intuitively demonstrate the module's optimization effect on the feature matrix. For example... Figure 4 and Figure 5As shown, the original multi-scale feature map displays the original multi-scale feature matrix, with the horizontal axis representing time (in milliseconds), the vertical axis representing the frequency band index, and the color intensity representing the magnitude of the feature value. The image shows a uniform background noise distribution across the entire time-frequency plane, while the leakage feature regions marked by the two rectangles are not prominent enough. The purified feature map shows the feature matrix processed by the time-frequency purification module. Under the same coordinate system, the background noise is significantly reduced, while the leakage feature regions are significantly enhanced. This comparison verifies the working mechanism of the dual-domain gating unit. The time-domain gating weights accurately locate the time period of the leakage event, while the frequency-domain gating weights enhance the sensitive frequency band. The "focusing lens" effect formed by the outer product of these two effectively enhances the clustered leakage harmonics. The non-uniform distribution of feature intensity in the image reflects the energy differences between different leakage levels, and the color scale on the right side of the heatmap quantitatively displays the range of feature intensity variation.

[0093] S303. Dual-path fusion feature extraction based on dual-path feature interaction mechanism

[0094] Leakage acoustic signals exhibit long-range temporal dependencies caused by pressure wave propagation. However, fully connected networks disrupt the spatiotemporal structure of the signal, while recurrent neural networks are difficult to parallelize, resulting in low computational efficiency and failing to meet real-time monitoring requirements. This invention constructs a dual-path structure combining temporal convolution and frequency-domain self-attention to capture the temporal relationship and frequency band energy correlation of sound wave propagation, respectively, and fuses the dual-path outputs to enhance feature representation capabilities. The specific steps are as follows:

[0095] 1) Temporal path feature extraction

[0096] This embodiment uses a temporal convolutional network to process and refine the feature matrix, utilizing its dilated convolutional structure to capture long-range temporal dependencies. This process is represented as follows: The feature matrix of the time-domain path output is obtained. Strengthen the modeling of temporal characteristics caused by sound wave propagation; among which, This represents a temporal convolutional network, consisting of multiple layers of dilated convolutions and residual connections, used to extract long-range temporal dependency features; The feature matrix representing the output of the time-domain path has dimensions of . To enhance the modeling of temporal relationships.

[0097] It should be noted that the temporal convolutional network consists of multiple dilated convolutional layers. Each layer includes dilated convolution, weight normalization, ReLU activation, and residual connections. For example, a 4-layer temporal convolutional network with dilation rates of 1, 2, 4, and 8 for each layer and a kernel size of 3 is used to extract long-range temporal dependent features.

[0098] 2) Frequency Domain Path Feature Extraction

[0099] This embodiment combines a self-attention mechanism in the frequency domain, models the energy correlation between different frequency bands through query matrices, key matrices, and value matrices, and outputs a frequency domain enhanced feature matrix, thereby enhancing the representation capability of energy correlation between frequency bands, as shown below:

[0100]

[0101] in, To query the matrix, input the feature and weight matrices. The product is multiplied to obtain the attention weights, which are then used to calculate the attention weights, and are expressed as follows: ; The key matrix is ​​obtained by inputting the feature and weight matrices. The product is multiplied to obtain the attention weights, which are then used to calculate the attention weights, and are expressed as follows: ; The value matrix is ​​obtained by inputting the feature and weight matrix. The product obtained by multiplication is used to generate the output feature, and is represented as follows: ; To query the weight matrix of the transformation, the dimension is... , are trainable parameters; Let be the weight matrix of the transformation, with dimension . , are trainable parameters; The weight matrix for the value transformation has dimensions of . , are trainable parameters; Key matrix Transpose of; This represents the dimension of the key vector, used to scale the attention score and stabilize the training process; it is preferably set to [value missing]. This represents the normalized exponential function, used to calculate the attention weight distribution; The feature matrix representing the output of the frequency domain path has a dimension of . To enhance the modeling of energy correlation between frequency bands.

[0102] 3) Dual-path feature fusion

[0103] This embodiment combines the feature matrices output from the time-domain path and the feature matrices output from the frequency-domain path through weighted fusion and element-wise multiplication to output a dual-path fused feature matrix. This integrates temporal relationships and frequency band correlation features, enhancing the overall expressive power, as shown below:

[0104]

[0105] in, Represents the hyperbolic tangent activation function; The fusion weight matrix for the time-domain paths has dimensions of . , are trainable parameters; The fusion weight matrix for the frequency domain path has dimensions of . , are trainable parameters; This represents the dual-path fusion feature matrix, with dimension 1. .

[0106] It should be noted that the propagation delay characteristics of the pressure wave captured by the temporal convolutional network in the time domain are related to the harmonic energy modeled by the self-attention model in the frequency domain. Terms maintain feature independence by adding terms together, while The method employs element-wise multiplication to generate a cross-domain modulation effect, using the pressure wave propagation timing information learned in the time domain path as a modulation signal applied to the frequency domain features, thereby enabling the frequency band energy correlation to have time-varying characteristics and solving the defect that recurrent neural networks have difficulty modeling long-range frequency band correlations.

[0107] S304, Enhanced Leakage Sensitivity Characteristics

[0108] The main difference in characteristics between different leakage levels lies in the distribution of high-frequency harmonic energy. However, conventional feature extraction methods rely on manual design, lose phase information, and cannot adaptively enhance the leakage-sensitive frequency band, resulting in micro-leakage features being overwhelmed by noise and a low detection rate. This invention automatically learns and enhances the unique high-frequency harmonic abrupt change characteristics of leaks using a differentiable spectral enhancement operator, preserving phase information and improving the recall rate of micro-leakage. The specific steps are as follows:

[0109] 1) Calculate the high-frequency energy change

[0110] This embodiment calculates the second derivative of the dual-path fusion feature matrix along the frequency domain dimension, extracts the high-frequency energy change matrix through the high-frequency energy weight matrix and the linear rectification activation function, highlighting the high-frequency harmonic abrupt change characteristics caused by leakage, expressed as:

[0111]

[0112] in, This represents the high-frequency energy change matrix, with dimension 1. Highlighting the high-frequency sudden changes caused by the leakage; The sign for partial derivatives; This represents the feature matrix fused from the two paths. The second derivative is calculated along the frequency domain dimension to detect high-frequency energy abrupt changes. This represents the high-frequency energy weight matrix, with dimension 1. , used to scale and transform second-order differential results, are trainable parameters; This represents the linear rectification activation function, ensuring that the energy change is non-negative.

[0113] 2) Feature Enhancement Output

[0114] This embodiment obtains a leakage-sensitive feature matrix based on the high-frequency energy change matrix and the dual-path fusion feature matrix, enhancing the high-frequency abrupt change characteristics while preserving the original phase polarity, as shown below:

[0115]

[0116] in, This represents the gain coefficient, used to control the enhancement intensity of high-frequency energy changes, and is preferably set to 0.3; The sign function is used to preserve the phase polarity information of the original features. Calculate matrix The symbol for each element in the text; This represents the leakage-sensitive feature matrix, with dimension 1. .

[0117] It should be noted that, The term preserves phase polarity through a sign function, combined with the use of High-frequency mutation energy extracted from the item This produces the technical effect of "phase coherence enhancement," which enhances the energy of high-frequency harmonics while maintaining their phase relationship with the fundamental wave. This allows the unique high-frequency harmonics of micro-leakage to be manifested through phase consistency, solving the problem of harmonic structure destruction caused by phase distortion in conventional methods.

[0118] S305. Prediction and recognition based on dynamic threshold classifier.

[0119] The classification of natural gas pipeline leak states suffers from a severe class imbalance problem. The number of normal operation samples far exceeds the number of leak samples. Fixing the classification boundary causes the model to favor the majority class, resulting in the neglect of smaller sample classes, especially micro-leaks, and a decline in classification performance. This invention adaptively adjusts the decision boundary through a dynamic temperature coefficient adjustment mechanism based on feature space sample density, assigning larger gradient weights to classes with fewer samples, thereby improving the classification accuracy of smaller sample classes. The specific steps are as follows:

[0120] 1) Calculate the category-related temperature coefficient

[0121] In this embodiment, the temperature coefficient is calculated using a logarithmic function based on the number of samples in each category. Categories with fewer samples have a larger temperature coefficient, resulting in smoother decision boundaries in the probability distribution for categories with fewer samples. This is expressed as:

[0122]

[0123] in, This represents the temperature coefficient of the c-th category, used to adjust the smoothness of the decision boundary for each category in the Softmax function. This represents the scaling factor, which controls the overall range of the temperature coefficient; a value of 0.5 is preferred. This represents the total number of samples in the training dataset; This represents the number of samples in the c-th class in the training dataset; This represents a logarithmic function, with the default base being the natural constant.

[0124] 2) Calculate the category logit value

[0125] In this embodiment, the compressed leakage-sensitive feature matrix is ​​mapped to the category space through a fully connected layer to obtain the original scores for each category, as follows:

[0126]

[0127] in, To compress the leakage-sensitive feature matrix, it is done by optimizing the dimension of... Leakage sensitivity feature matrix Then, global average pooling is used to compress it into dimensions. ; This represents the weight matrix of the third fully connected layer, with dimension 1. ; This represents the bias vector of the third fully connected layer; This represents the logit value of the c-th category. The logit value refers to the original classification score.

[0128] 3) Calculate the adaptive class probability

[0129] This embodiment uses a normalized exponential function adjusted for temperature coefficient to calculate the predicted probability of each category, giving more significant gradient weights to categories with fewer samples and improving the recognition ability of small sample categories, expressed as:

[0130] in, Represents a given feature matrix The predicted probability that a sample belongs to the c-th category; Represents the natural exponential function; Indicates the first Logit values ​​for each category; Indicates the first Temperature coefficients for each category; c represents the true class label of the sample; c is the class index, ranging from 1 to... ; This represents the total number of categories.

[0131] 4) Prediction category determination

[0132] In this embodiment, the category index corresponding to the highest predicted probability is used as the predicted category label for the sample. The final classification determination of the leakage state is represented as follows:

[0133]

[0134] in, Indicates the predicted class label of the sample; This indicates that the category index corresponding to the highest probability is taken.

[0135] It should be noted that the temperature coefficient is used at the decision boundary level. Adjusting the probability space geometry, when the number of leaked samples is extremely small, the temperature coefficient Logarithmic amplification makes The relative increase in the number of terms leaves more feature space for minority class samples. As the decision boundary expands outward, micro-leakage samples can form more compact clusters in the feature space, while avoiding excessive compression of majority class samples, which significantly improves the model's sensitivity to low-frequency events.

[0136] In the acoustic classification task of natural gas pipeline leaks, the single cross-entropy loss function is insufficient in constraining the feature purification process, easily leading to incomplete denoising, distortion of leak features, or residual noise, thus affecting classification robustness. Conventional methods cannot jointly optimize feature separation and classification performance, and it is difficult to maintain the integrity of leak physical features under low signal-to-noise ratio conditions. This invention constructs a dual-supervision mechanism that combines classification loss and purification loss to form a hybrid loss function, thereby jointly optimizing the feature denoising, feature preservation, and classification decision-making process, improving the model's discrimination ability and robustness in high-noise environments. The specific steps are as follows:

[0137] 1) Constructing a hybrid loss function framework

[0138] This embodiment combines classification loss and purification loss, forming a hybrid loss function through weighted summation of loss tradeoff coefficients. This function jointly optimizes the feature denoising, feature preservation, and classification decision processes, and is expressed as follows:

[0139]

[0140] in, This represents the hybrid loss function, which is the overall optimization objective during model training. This represents the classification loss, used to optimize the model's classification performance for leakage states. Cross-entropy loss is preferred as the classification loss. This represents the purification loss, used to constrain the feature purification process and prevent feature distortion due to leakage. This represents the loss trade-off coefficient, which controls the weight of purification loss in the total loss. The preferred value is 0.3.

[0141] 2) Calculation of purification loss

[0142] This embodiment calculates the L1 norm difference between the inverse time-frequency transform of the purified feature matrix and the inverse time-frequency feature matrix of the clean sample to obtain the purification loss. This constrains the feature purification process to approximate the features of the clean sample and avoids feature leakage and distortion, as expressed in:

[0143]

[0144] in, This indicates purification losses; To represent the time-frequency characteristics of clean samples, wavelet thresholding denoising is used to denoise the original sound pressure signal. Preprocessing generation specifically involves applying wavelet transform to the original signal, applying soft thresholding to process the noise coefficient, then using inverse wavelet transform to obtain the denoised signal, and finally calculating its adaptive time-frequency characteristics. This represents the inverse time-frequency transform, which maps time-frequency characteristics back to the time-domain signal; This represents the L1 norm.

[0145] It should be noted that the purification loss is calculated using the L1 norm instead of the L2 norm because the L1 norm is more robust to outliers and helps to preserve the sparsity of leakage features, while the L2 norm can over-smooth features, potentially leading to the loss of leakage details.

[0146] In this embodiment, a staged training strategy is used to initialize the model parameters during iterative training and parameter updating of the leakage identification model. The specific steps are as follows:

[0147] First, fix the parameters of the adaptive time-frequency decomposition module, and use the Adam optimizer with a learning rate of 0.001 to pre-train the cascaded multi-scale convolution module to the dynamic threshold classifier for 100 rounds.

[0148] Then, the constraints of the adaptive time-frequency decomposition module are removed, and end-to-end joint training is performed with the hybrid loss function as the optimization objective. During the training process, 32 sound pressure signals are input in each batch, and the gradient between the model prediction result and the real label is calculated through the backpropagation algorithm, and all trainable parameters are dynamically updated.

[0149] Every 10 training epochs, the micro-leakage recall and overall accuracy are evaluated on the validation set. If the metrics do not improve for 5 consecutive epochs, the learning rate is reduced to 1 / 10 of its original value.

[0150] The training terminates after 200 rounds, and the model parameters that best perform on the validation set are saved.

[0151] This study evaluates the performance of different leak detection methods under varying signal-to-noise ratio (SNR) environments. SNR, measured in decibels, is a measure of the relative strength of effective information and noise in a signal. A lower SNR indicates stronger noise interference and greater detection difficulty. This experiment compares several methods, including short-time Fourier transform combined with convolutional neural networks, wavelet transform combined with recurrent neural networks, Mel-frequency cepstral coefficients combined with support vector machines, and the intelligent integrated management method proposed in this invention. Figure 6 As shown in the figure, the line graph represents the signal-to-noise ratio (SNR) (in decibels) on the horizontal axis and the accuracy (in percentage) on the vertical axis. Higher accuracy indicates better detection performance. The graph shows that as the SNR decreases, the accuracy of all methods declines, but the method of this invention shows the smallest decrease, especially under low SNR conditions, where its accuracy is significantly higher than other comparative methods. The experimental results demonstrate the robustness advantage of the method of this invention in strong noise environments, indicating that the adaptive time-frequency feature decomposition technology can dynamically adjust the time-frequency resolution, enhance the energy concentration of the leakage-sensitive frequency band, and thus effectively improve the SNR. The short-time Fourier transform combined with convolutional neural networks and the wavelet transform combined with recurrent neural networks perform reasonably well at medium SNRs, but their performance degrades rapidly under high noise environments. The Mel frequency cepstral coefficient combined with support vector machine method has low overall performance, indicating its limited ability to handle complex acoustic scenarios. The method of this invention, by optimizing the basis function parameters and multi-scale feature encoding, ensures high detection accuracy under various noise conditions.

[0152] In this embodiment, the detection performance of different methods for different leakage levels is compared. Leakage levels are categorized as normal operating conditions, micro-leakage (leakage aperture less than 2 mm), and severe leakage (leakage aperture greater than 2 mm). Figure 7As shown in the figure. This experiment uses the F1 score as the evaluation metric. The F1 score is a measure of the combined precision and recall, expressed as a percentage. A higher value indicates more balanced and reliable detection performance. The experiment compared the short-time Fourier transform combined with convolutional neural network method, wavelet transform combined with recurrent neural network method, Mel-frequency cepstral coefficient combined with support vector machine method, and the method of this invention. The results are presented in a grouped bar chart, with the horizontal axis representing the leakage type and the vertical axis representing the F1 score. The figure shows that for normal operating conditions, all methods have high F1 scores, but the method of this invention is close to perfect, indicating its stability in high-sample-size categories. For micro-leakage, the F1 scores of other methods drop significantly, while the method of this invention maintains a high level, highlighting the advantages of this invention in detecting small-aperture leaks. This demonstrates that the leak-sensitive feature enhancement technology and dynamic threshold classifier can effectively capture weak signals and solve the class imbalance problem. For severe leaks, the method of this invention also performs best, indicating that it has consistent performance under different leakage levels. Other methods performed poorly in micro-leakage detection, especially the Mel frequency cepstral coefficient combined with support vector machine method, reflecting the limitations of conventional feature extraction methods in handling sparse and transient leakage features.

[0153] In this embodiment, the detection performance of different methods under different leakage orifice sizes is evaluated, such as... Figure 8 As shown, the leakage orifice diameter is the size of the leakage orifice, measured in millimeters, ranging from 0.5 mm to 4 mm, covering the range of micro-leakage and severe leakage. The detection rate is the percentage of leaks correctly identified; a higher value indicates a more effective method. This experiment compared the short-time Fourier transform combined with a convolutional neural network method, the wavelet transform combined with a recurrent neural network method, the Mel-frequency cepstral coefficients combined with a support vector machine method, and the method of this invention. The results are illustrated in a line graph, with the horizontal axis representing the leakage orifice diameter (in millimeters) and the vertical axis representing the detection rate. A vertical dashed line at 2 mm marks the boundary between micro-leakage and severe leakage. The detection rate of all methods increases with increasing leakage orifice diameter, but the method of this invention maintains the highest detection rate across the entire range, especially at small orifice diameters (e.g., below 1 mm), where its detection rate is significantly higher than other methods, demonstrating the high sensitivity of this invention to micro-leakage. Other methods have lower detection rates at small orifice diameters; for example, the short-time Fourier transform combined with a convolutional neural network method has a detection rate of less than 50% at 0.5 mm, indicating its difficulty in capturing weak leakage signals. Experimental results show that the differentiable spectrum enhancement operator and the dual-path feature interaction mechanism can automatically learn and enhance the high-frequency harmonic mutation features unique to leakage, and at the same time optimize the decision boundary through a dynamic threshold classifier, thereby improving the recognition ability of small sample categories.

[0154] This embodiment embeds the trained leak detection model into the edge computing equipment of the pipeline monitoring center, receiving sound pressure signals transmitted from sound sensors along the pipeline in real time to achieve natural gas pipeline leak monitoring. The monitoring process consists of three steps:

[0155] 1) Signal preprocessing: The input 16,000-point raw sound pressure signal is normalized by mean and divided into 124 time steps by frame with a frame length of 256 points and a frame shift of 128 points.

[0156] 2) Online feature extraction and recognition, which sequentially executes adaptive time-frequency feature decomposition (S2), cascaded multi-scale convolutional coding (S301), time-frequency purification (S302), dual-path feature fusion (S303), and leakage-sensitive feature enhancement (S304).

[0157] 3) Dynamic classification decision-making: The extracted leakage sensitivity feature matrix is ​​input into a dynamic threshold classifier (S305), which outputs a real-time leakage status probability distribution. When the probability of a micro-leak or severe leak exceeds 0.85, a graded alarm is triggered. The alarm information is synchronously transmitted to the management platform, and the timestamp and location station number of the abnormal signal are automatically marked, achieving a second-level response to leakage events.

[0158] S4. Conduct comprehensive management and control of natural gas pipelines based on the probability distribution of leakage levels.

[0159] In an optional embodiment of the present invention, step S4, which constructs a three-level control system based on leakage monitoring results, specifically includes:

[0160] 1) Risk warning layer: Combining the probability output of leakage level with historical data, high-risk sections of the pipeline are visualized through GIS map to generate a leakage risk heat map;

[0161] 2) Decision support layer: For pipe sections with continuous alarms, a multi-source data verification mechanism is activated to retrieve the pressure and flow sensor data of the corresponding section and cross-verify them with the acoustic signature matrix, and automatically generate a maintenance priority score.

[0162] 3) The response execution layer automatically reduces the pressure of the upstream valve of the leaking pipe section through the pipeline pressure control module, and at the same time pushes a work order containing the location coordinates, leakage level and recommended treatment plan to the inspection terminal.

[0163] The feature matrix and handling records of all leakage events are archived into a knowledge base for regular iterative updates to the leakage identification model parameters, forming a closed-loop intelligent management system of "monitoring-diagnosis-handling-optimization".

[0164] This invention also provides an intelligent integrated management system for natural gas pipeline protection, comprising:

[0165] The signal acquisition module is used to acquire the acoustic pressure signal of the natural gas pipeline;

[0166] The feature decomposition module is used to perform time-frequency feature decomposition on the acoustic pressure signal of the natural gas pipeline using an adaptive basis function decomposition method to obtain the time-frequency feature matrix.

[0167] The leakage identification module is used to build a leakage identification model. The leakage identification model is used to perform concatenated multi-scale convolutional coding, time-frequency feature purification, dual-path feature fusion and leakage-sensitive feature enhancement on the time-frequency feature matrix in sequence. Based on the obtained leakage-sensitive features, dynamic classification decision is made to obtain the leakage level probability distribution.

[0168] The integrated management and control module is used for the integrated management and control of natural gas pipelines based on the probability distribution of leakage levels.

[0169] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0173] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. An intelligent integrated management method for natural gas pipeline protection, characterized in that, Includes the following steps: Acquire acoustic pressure signals from natural gas pipelines; An adaptive basis function decomposition method is used to perform time-frequency feature decomposition on the acoustic pressure signal of the natural gas pipeline, and the time-frequency feature matrix is ​​obtained. A leakage identification model is constructed. The model is then used to perform concatenated multi-scale convolutional coding, time-frequency feature purification, dual-path feature fusion, and leakage-sensitive feature enhancement on the time-frequency feature matrix. Based on the obtained leakage-sensitive features, dynamic classification decisions are made to obtain the leakage level probability distribution. Comprehensive management and control of natural gas pipelines based on the probability distribution of leakage levels.

2. The intelligent integrated management method for natural gas pipeline protection according to claim 1, characterized in that, The time-frequency feature matrix of the acoustic pressure signal from the natural gas pipeline is obtained by using an adaptive basis function decomposition method, including: Construct adaptive basis functions; The time-frequency feature matrix is ​​generated by performing element-wise multiplication and summation operations on the acoustic pressure signal of the natural gas pipeline and the adaptive basis function. The parameter set of the basis functions is optimized using a backpropagation algorithm based on gradient descent.

3. The intelligent integrated management method for natural gas pipeline protection according to claim 1, characterized in that, Cascaded multi-scale convolutional coding includes: Set the multi-scale expansion rate parameters; Multi-scale dilation convolution is performed on the time-frequency feature matrix, the convolution results with different dilation rates are summed, and batch normalization and modified linear unit activation functions are applied to obtain the multi-scale feature matrix.

4. The intelligent integrated management method for natural gas pipeline protection according to claim 3, characterized in that, Time-frequency feature purification includes: A dual-domain gating unit is used to extract salient features from the time and frequency dimensions respectively, and a time-frequency joint weight matrix is ​​constructed by outer product operation; The purified feature matrix is ​​obtained by multiplying the time-frequency joint weight matrix element-wise with the multi-scale feature matrix.

5. The intelligent integrated management method for natural gas pipeline protection according to claim 4, characterized in that, Dual-path feature fusion includes: A temporal convolutional network is used to process and refine the feature matrix, and a dilated convolutional structure is used to capture long-range temporal dependencies to obtain a temporal path feature matrix. By combining a self-attention mechanism in the frequency domain dimension, the energy correlation between different frequency bands is modeled through query matrix, key matrix and value matrix to obtain frequency domain path feature matrix; The time-domain path feature matrix and the frequency-domain path feature matrix are combined by weighted fusion and element-wise multiplication to obtain the dual-path fusion feature matrix.

6. The intelligent integrated management method for natural gas pipeline protection according to claim 5, characterized in that, Enhanced leakage sensitivity features include: The second derivative of the dual-path fusion feature matrix is ​​calculated along the frequency domain dimension, and the high-frequency energy change matrix is ​​extracted through the high-frequency energy weight matrix and the linear rectified activation function. The leakage-sensitive feature matrix is ​​calculated based on the high-frequency energy change matrix and the dual-path fusion feature matrix.

7. The intelligent integrated management method for natural gas pipeline protection according to claim 6, characterized in that, The leakage sensitivity feature matrix is ​​as follows: in, Represents the leakage-sensitive feature matrix. This represents the dual-path fusion feature matrix. Indicates the gain coefficient. This represents the high-frequency energy change matrix. Represents a symbolic function. , Represents the linear rectified activation function. Represents the high-frequency energy weight matrix. The sign of the partial derivative. Represents a frequency variable.

8. The intelligent integrated management method for natural gas pipeline protection according to claim 1, characterized in that, Based on the obtained leakage sensitivity characteristics, a dynamic classification decision is made to obtain the leakage level probability distribution, including: The temperature coefficient is calculated using a logarithmic function based on the number of samples in each category; The leakage-sensitive feature matrix is ​​mapped to the category space through a fully connected layer to obtain the original scores for each category; Based on the original scores of each category, the predicted probability of each category is calculated using a normalized exponential function adjusted for temperature coefficient. The category index corresponding to the highest predicted probability is selected as the predicted category label of the sample to obtain the leakage level probability distribution.

9. The intelligent integrated management method for natural gas pipeline protection according to claim 1, characterized in that, The leakage detection model uses a hybrid loss function consisting of classification loss and purification loss during training, specifically: in, For a mixed loss function, For classifying losses, To purify the losses, For loss trade-off coefficients; The purification losses are as follows: in, It is an inverse time-frequency transform. To purify the feature matrix, For clean sample time-frequency features, It is an L1 norm.

10. An intelligent integrated management system for natural gas pipeline protection, characterized in that, include: The signal acquisition module is used to acquire the acoustic pressure signal of the natural gas pipeline; The feature decomposition module is used to perform time-frequency feature decomposition on the acoustic pressure signal of the natural gas pipeline using an adaptive basis function decomposition method to obtain the time-frequency feature matrix. The leakage identification module is used to build a leakage identification model. The leakage identification model is used to perform concatenated multi-scale convolutional coding, time-frequency feature purification, dual-path feature fusion and leakage-sensitive feature enhancement on the time-frequency feature matrix in sequence. Based on the obtained leakage-sensitive features, dynamic classification decision is made to obtain the leakage level probability distribution. The integrated management and control module is used for the integrated management and control of natural gas pipelines based on the probability distribution of leakage levels.

Images