Intelligent comprehensive management method and system for natural gas pipeline protection

Abstract

The application relates to the field of artificial intelligence and data processing technology, and discloses an intelligent comprehensive management method and system for natural gas pipeline protection. The method comprises the following steps: acquiring a sound pressure signal of a natural gas pipeline; adopting an adaptive basis function decomposition method to perform time-frequency feature decomposition on the sound pressure signal of the natural gas pipeline to obtain a time-frequency feature matrix; constructing a leakage identification model; sequentially performing cascade multi-scale convolution coding, time-frequency feature purification, double-path feature fusion and leakage sensitive feature enhancement on the time-frequency feature matrix by using the leakage identification model; and performing dynamic classification decision according to the obtained leakage sensitive feature to obtain a leakage grade probability distribution; and performing comprehensive management and control of the natural gas pipeline according to the leakage grade probability distribution. The application can improve the leakage detection accuracy in a class imbalance scene and realize real-time comprehensive management and control of the natural gas pipeline.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an intelligent comprehensive management method and system for natural gas pipeline protection. BACKGROUND

[0002] Natural gas, as one of the most important clean energy, has been widely used in urban gas transportation, industrial production and power generation. With the continuous expansion of pipeline transportation scale and the gradual aging of pipeline, pipeline leakage problem is increasingly prominent. Natural gas has the characteristics of flammability, explosiveness and easy diffusion. Once leakage occurs, not only energy waste and economic loss will be caused, but also serious safety accidents and environmental pollution may be caused. Therefore, how to realize real-time monitoring and early leakage identification of natural gas pipeline has become a key problem in energy safety and public safety management.

[0003] The leakage detection method mainly relies on pressure monitoring, flow balance and conventional acoustic feature extraction, etc. However, these methods have certain limitations. The detection based on pressure and flow often has a lagging response, and it is difficult to find small aperture or early micro leakage in time. Although the conventional acoustic method has a certain real-time performance, the commonly used feature extraction methods such as short-time Fourier transform, wavelet transform or mel frequency cepstrum depend on fixed time-frequency resolution, and it is difficult to capture both transient pulse and wideband resonance characteristics at the same time. In strong noise environment, the signal is easy to be covered, resulting in high micro leakage detection rate. In addition, most of the existing deep learning detection models use single-scale convolution or recurrent neural network structure, which cannot fully consider the multi-scale features of different leakage modes, and also lack the mapping with acoustic physical law, and the model has poor interpretability. At the same time, the leakage data has significant class imbalance, and the fixed decision boundary makes the model easily biased towards normal working conditions, further weakening the identification ability of micro leakage.

[0004] The above technical solutions still need to be further solved, which have the following problems:

[0005] 1. Fixed window or conventional wavelet is difficult to capture both millisecond transient and continuous wideband characteristics, and only amplitude is considered without phase, resulting in dilution of key leakage pulse and harmonic structure.

[0006] 2. Single-scale convolution or simple features are easily confused between different leakage modes, lack of mapping with physical acoustic quantities, and the model has poor interpretability and generalization.

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

[0008] 4. The fixed Softmax temperature is easy to deviate to the majority class with single cross-entropy, and the micro leakage is ignored, in addition, the feature denoising and the classification are not optimized jointly, the feature distortion or residual noise is easy to appear under low signal-to-noise ratio, and the robustness is insufficient. SUMMARY

[0009] In view of the above problems in the prior art, the present application provides an intelligent comprehensive management method and system for natural gas pipeline protection.

[0010] In order to achieve the above-mentioned purposes, the technical scheme adopted by the present application is as follows:

[0011] In the first aspect, the present application provides an intelligent comprehensive management method for natural gas pipeline protection, comprising the following steps:

[0012] Obtaining a sound pressure signal of the natural gas pipeline;

[0013] Using an adaptive basis function decomposition method to perform time-frequency feature decomposition on the sound pressure signal of the natural gas pipeline to obtain a time-frequency feature matrix;

[0014] Constructing a leakage identification model, using the leakage identification model to sequentially perform cascade multi-scale convolution coding, time-frequency feature purification, double-path feature fusion and leakage sensitive feature enhancement on the time-frequency feature matrix, and making a dynamic classification decision according to the obtained leakage sensitive feature to obtain a leakage level probability distribution;

[0015] According to the leakage level probability distribution, the natural gas pipeline is comprehensively managed.

[0016] In the second aspect, the present application provides an intelligent comprehensive management system for natural gas pipeline protection, comprising:

[0017] A signal acquisition module for acquiring a sound pressure signal of the natural gas pipeline;

[0018] A feature decomposition module for using an adaptive basis function decomposition method to perform time-frequency feature decomposition on the sound pressure signal of the natural gas pipeline to obtain a time-frequency feature matrix;

[0019] A leakage identification module for constructing a leakage identification model, using the leakage identification model to sequentially perform cascade multi-scale convolution coding, time-frequency feature purification, double-path feature fusion and leakage sensitive feature enhancement on the time-frequency feature matrix, and making a dynamic classification decision according to the obtained leakage sensitive feature to obtain a leakage level probability distribution;

[0020] A comprehensive management module for comprehensively managing the natural gas pipeline according to the leakage level probability distribution.

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

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

[0023] (2) The application constructs a cascade expansion convolution group, covers the receptive field from transient to wideband resonance, and establishes a mapping between the receptive field time and the acoustic propagation time, taking into account micro-leakage and severe leakage and having physical interpretability.

[0024] (3) The application is based on the outer product weight of time gating and frequency gating, forms a weight matrix of time-frequency joint focusing, suppresses time-frequency uniform noise, highlights diagonal cluster harmonics and sparse leakage characteristics.

[0025] (4) The application adopts a differentiable spectrum enhancement operator combined with a double-path feature interaction, and trains and optimizes the model parameters with a hybrid loss, significantly improving the micro-leakage recall and overall accuracy in the class imbalance scenario. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 It is a schematic diagram of an intelligent comprehensive management method for natural gas pipeline protection;

[0027] Figure 2 It is a short-time Fourier transform time-frequency diagram;

[0028] Figure 3 It is a schematic diagram of the adaptive time-frequency feature decomposition of the application;

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

[0030] Figure 5 It is a schematic diagram of the purified feature;

[0031] Figure 6 It is a comparison chart of the leakage detection accuracy of different leakage detection methods in different signal-to-noise ratio environments;

[0032] Figure 7 It is a comparison chart of the detection performance of different leakage detection methods for different leakage types;

[0033] Figure 8 It is a comparison chart of the detection performance of different leakage detection methods under different leakage apertures. DETAILED DESCRIPTION

[0034] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0035] As Figure 1 shown, the intelligent comprehensive management method for natural gas pipeline protection provided by the embodiment of the present application comprises the following steps S1 to S4:

[0036] S1, acquiring a natural gas pipeline sound pressure signal;

[0037] In an optional embodiment of the present application, step S1 performs natural gas pipeline monitoring data collection and labeling. A high-sensitivity sound sensor array is deployed along the natural gas pipeline to continuously collect pipeline sound pressure signals at a sampling rate of 16 kHz, forming an original sound pressure signal dataset. The data collection covers three working conditions of the pipeline: normal working condition, micro leakage, and serious leakage. Each working condition collects not less than 5000 samples, and each sample has 16000 sampling points.

[0038] The labeling process adopts a double verification mechanism: first, the start and end time of the leakage and the leakage level are marked by artificial listening combined with professional voiceprint analysis software; second, tracer gas is injected at the pipeline pressure mutation point, and a gas concentration detector is used to verify the authenticity of the leakage event. The labeling categories are divided into three categories: category 0 (normal working condition), category 1 (micro leakage: leakage aperture ≤ 2 mm), and category 2 (serious leakage: leakage aperture > 2 mm);

[0039] The final training dataset contains 15000 labeled samples, of which normal working condition samples account for 60%, micro leakage accounts for 25%, and serious leakage accounts for 15%, ensuring coverage of combined scenarios of different background noise intensities and pipeline pressures.

[0040] S2, performing time-frequency feature decomposition on the natural gas pipeline sound pressure signal using an adaptive basis function decomposition method to obtain a time-frequency feature matrix;

[0041] In an optional embodiment of the present application, step S2 performs time-frequency feature decomposition using an adaptive time-frequency decomposition module.

[0042] In the process of natural gas pipeline acoustic signal collection, the signal obtained by the sound sensor often contains strong background noise and weak leakage signal superimposed on each other, the leakage voiceprint shows non-stationary and sparse characteristics in time-frequency domain, and the conventional short-time Fourier transform adopts fixed window length, which cannot capture both transient leakage pulse and wide frequency characteristics of continuous leakage, resulting in loss of key information. The adaptive basis function decomposition method is adopted to dynamically adjust the time-frequency resolution, enhance the energy concentration degree of the leakage sensitive frequency band, and improve the signal-to-noise ratio, and the specific steps are as follows:

[0043] 1) Constructing adaptive basis function

[0044] By adopting the parameterized basis function, 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 voiceprint characteristics, and these parameters are optimized by the back propagation algorithm based on gradient descent, so that the basis function can enhance the energy concentration degree of the leakage sensitive frequency band, thereby improving the signal-to-noise ratio, which is expressed as:

[0045]

[0046] Wherein, Ψ k (k,t,f) represents the kth adaptive basis function for extracting specific time-frequency characteristics related to leakage in the signal; t represents the time variable for describing the process of signal change with time; f represents the frequency variable for describing the distribution of signal in the frequency domain; k represents the index of the basis function, k=1, 2,..., K n ; K n is the total number of basis functions, which is preferably set to 64; Ψ k (t,f) represents the value of the kth adaptive basis function at t time f frequency, which is used to extract specific time-frequency characteristics related to leakage in the signal, and its parameter set {τ k ,σ t，k ,f k ,φ k ,ζ k} is optimized by the back propagation algorithm based on gradient descent to match the leakage characteristics; exp(·) is the natural exponential function; τ k represents the time domain center parameter of the kth basis function, which controls the position of the basis function on the time axis to align the starting time of the leakage event; σ t,k represents the time domain width parameter of the kth basis function, which controls the decay rate of the basis function in the time direction and affects the adaptability of the time resolution; f kis the center frequency parameter of the k-th basis function, which controls the main energy distribution area of the basis function in the frequency domain, used to match the leakage sensitive frequency band; φ is the phase parameter of the k-th basis function, which is used to adjust the starting phase of the basis function to better match the phase characteristics of the leakage signal; Γ(·) is the band selection function, which is used to control the attenuation characteristics of the basis function in the frequency domain, and enhance the energy concentration of the specific frequency band; Γ(f, ζ k ) is the band selection function, which is used to control the attenuation characteristics of the basis function in the frequency domain, and enhance the energy concentration of the specific frequency band, and the calculation method is represented as Γ(f, ζ k ) = exp(-ζ k (f-f k ) 2 ); ζ k is the bandwidth sharpness control parameter of the k-th basis function, which is used to adjust the steepness of the band selection function and affect the adaptability of the frequency domain resolution.

[0047] 2) Perform time-frequency feature decomposition

[0048] Multiply and sum the original sound pressure signal with the adaptive basis function element by element to generate a time-frequency feature matrix, which highlights the time-frequency energy distribution of the leakage sensitive frequency band, represented as:

[0049]

[0050] Where X ATF is the time-frequency feature matrix, representing the time-frequency feature matrix obtained after decomposition by the adaptive basis function, with a dimension of K n ×T, used to highlight the time-frequency energy distribution of the leakage sensitive frequency band; T is the time step, obtained from the original sound pressure signal X raw by frame operation, for example, if the original signal length is 16000 points, the frame length is 256 points, and the frame shift is 128 points, then represents the rounding operation; ⊙ represents element-wise multiplication; X raw represents the original sound pressure signal, belonging to the real number vector space R 16000 , containing pipeline vibration noise and environmental interference components; R 16000 represents the real number vector space with a dimension of 16000, i.e., the original sound pressure signal is a 16000-dimensional real number vector, containing pipeline vibration noise and environmental interference components.

[0051] 3) Optimize the basis function parameters

[0052] Optimize the parameter set {τ k , σ t，k , f k , φ k , ζ k}, so as to concentrate energy on the leakage sensitive frequency band, thereby improving the signal-to-noise ratio;

[0053] In a specific implementation, a random gradient descent method is used, the learning rate is set to 0.001, the gradient is calculated by the chain rule, the signal-to-noise ratio is maximized as the loss function during training, and iterative optimization is performed to concentrate energy on the leakage sensitive frequency band.

[0054] It should be noted that the parameter set {τ k , σ t，k , f k , φ k , ζ k} of the adaptive basis function is constructed as a differentiable optimization object, and these parameters are dynamically optimized using gradient descent backpropagation, so that the basis function can adaptively match the non-stationary characteristics of the leakage voiceprint. In particular, the phase parameter φ k breaks through the limitation of conventional time-frequency analysis which only focuses on amplitude, significantly enhances the coherent superposition effect of transient pulses through phase alignment, and solves the contradiction between transient pulses and continuous leakage in fixed window length time-frequency analysis. When the time domain width parameter σ t，k of the basis function is automatically contracted, millisecond-level leakage pulses can be captured, and when the bandwidth sharpness control parameter ζ k of the basis function is adjusted, wide frequency resonance can be covered, and dynamic adaptability makes weak leakage present energy focusing effect in strong noise background.

[0055] The embodiment compares the effects of different time-frequency feature decomposition methods through visualization, and analyzes the superiority of the method in extracting leakage features. Time-frequency analysis is to decompose the signal in time and frequency dimensions to identify the start time and frequency band distribution of the leakage event. This experiment compares the short-time Fourier transform and the adaptive time-frequency feature decomposition method of the present application, as shown in Figure 2 and Figure 3 . Figure 2 and Figure 3 use heat maps to show that the horizontal axis represents time (in seconds), the vertical axis represents frequency (in hertz), and the color depth represents the signal amplitude. The higher the amplitude, the more concentrated the energy. From the results of the short-time Fourier transform, it can be seen that the leakage signal is relatively blurred in time and frequency, and the energy is dispersed, making it difficult to clearly distinguish the leakage area. The time-frequency diagram of the method of the present application shows that the energy of the leakage area (between about 0.3 seconds and 0.7 seconds) in the low frequency band (near 400 Hz and 600 Hz) is significantly enhanced and concentrated, forming a clear time-frequency pattern, indicating that the method of the present application can more accurately locate the time and frequency characteristics of the leakage event. The experimental results show that the adaptive basis function decomposition matches the non-stationary characteristics of the leakage voiceprint by dynamically adjusting the parameters, enhances the energy focusing of the sensitive frequency band, thereby effectively suppressing the background noise and retaining the key information, providing higher quality feature input for subsequent leakage recognition, and improving the overall detection reliability.

[0056] S3, construct a leakage identification model, and sequentially perform cascade multi-scale convolution coding, time-frequency feature purification, double-path feature fusion and leakage sensitive feature enhancement on the time-frequency feature matrix by using the leakage identification model, and perform dynamic classification decision according to the obtained leakage sensitive feature to obtain a leakage grade probability distribution;

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

[0058] S301, multi-scale convolution feature coding is performed by using a cascade multi-scale convolution module

[0059] Leakage voiceprints have multi-scale characteristics, micro-leakage presents narrowband transient pulses, and serious leakage contains wideband resonance. Conventional single-scale convolution can only capture fixed receptive field features, and it is difficult to cover different leakage modes. The present application captures multi-scale leakage features by constructing a cascade dilated convolution group to fuse local details and global context. The specific steps are as follows:

[0060] 1) Set a multi-scale dilation rate parameter

[0061] In this embodiment, a group of dilation rate values are determined according to the time scale range of acoustic events to cover different receptive fields and ensure that multi-scale features from short transient pulses to long wideband resonance can be captured. The dilation rate set is defined as D={1, 3, 7, 15}, and the dilation rate d corresponds to the dilation factor of the convolution kernel, where d=1 represents standard convolution operation, the receptive field is the smallest, and it is suitable for short transient features, d=3, 7, 15 represents dilated convolution operation, and the receptive field increases level by level, which is suitable for capturing longer time wideband resonance features.

[0062] In specific implementation, D={1, 3, 7, 15} corresponds to a time scale of about 24ms to 248ms. For example, assuming that the sampling rate is 16kHz and the time step interval of the time-frequency feature is 8ms, the receptive field time of the dilated convolution is [(3-1)×d+1]×Δt, and Δt is the time step interval. The calculation can obtain:

[0063] When d=1, the receptive field time is [(3-1)×1+1]×8ms=24ms;

[0064] When d=3, the receptive field time is [(3-1)×3+1]×8ms=56ms;

[0065] When d=7, the receptive field time is [(3-1)×7+1]×8ms=120ms;

[0066] When d=15, the receptive field time is [(3-1)×15+1]×8ms=248ms.

[0067] 2) Computing multi-scale convolution features

[0068] The embodiment performs multi-scale dilated convolution operation on the input time-frequency feature matrix, sums the convolution results of different dilation rates, and applies batch normalization and rectified linear unit activation function to output a multi-scale feature matrix. By fusing the leakage features under different receptive fields, the model's ability to capture multi-scale voiceprints is enhanced, which is represented as:

[0069]

[0070] wherein, is the l-th layer multi-scale feature matrix, representing the feature matrix output by the l-th layer multi-scale convolution layer of the cascaded multi-scale convolution module, with a dimension of C out x T, which fuses the leakage features under multi-scale receptive fields; ReLU(·) represents a rectified linear unit activation function, which enhances the expression ability of the model through nonlinear transformation; BN(·) represents a batch normalization operation, which is used to accelerate the model training process and improve the stability of training; represents the convolution kernel weight matrix with a dilation rate of d in the l-th layer of the cascaded multi-scale convolution module, with a dimension of C in x C out x d, which is used to extract features under a specific receptive field and is a trainable parameter; * represents a dilated convolution operation; C out is the number of channels output by the current layer convolution operation of the cascaded multi-scale convolution module; C in is the number of channels input by the current layer convolution operation of the cascaded multi-scale convolution module. For the first layer, C in = K n .

[0071] It should be noted that the total number of layers L of the cascaded multi-scale convolution module can be set to 3 layers, each layer containing convolution operations with a set of dilation rates D, and the input of the current layer is the output of the previous layer, and the input of the first layer is X ATF .

[0072] It should also be noted that the skip design of the dilation rate makes each path focus on a specific scale, d = 1 corresponds to a 24 ms transient pulse, and d = 15 corresponds to a 248 ms wideband resonance. The discrete calculation method produces a cross-scale feature complementary effect in the summation operation ∑ d∈D , the time-domain positioning accuracy of the transient pulse and the frequency band coverage ability of the wideband resonance are synergistically enhanced through batch normalization, solving the feature confusion problem of single-scale convolution between micro-leakage and severe leakage. More importantly, the calculation model of the receptive field time [(3-1) x d + 1] x Δt establishes a mapping between physical sound wave propagation time and convolution operation, making the deep learning model have an interpretable acoustic physical basis.

[0073] S302, obtain a purified feature by using a time-frequency feature purification module

[0074] The background noise forms a continuous distribution in the time-frequency domain, and the leakage features are in the form of sparse clusters. The conventional attention mechanism ignores the time-frequency correlation, resulting in residual noise and failing to effectively separate the leakage features from the background interference. The present application extracts significant features from the time and frequency dimensions by using a dual-domain gating unit, and constructs a time-frequency joint weight matrix through outer product operation to suppress non-coherent noise. The specific steps are as follows:

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

[0076] In this embodiment, the multi-scale feature matrix is averaged and pooled along the frequency axis to compress the frequency dimension information, and a gating weight matrix in the time dimension is generated through a fully connected layer to emphasize the region with significant leakage features in the time dimension, which is represented as:

[0077] G t =sig(W t ·AvgPool f (Z MSF ))

[0078] Where G t represents the time-domain gating weight matrix, with a dimension of T×1, used to emphasize the region with significant leakage features in the time dimension; T is the number of time steps, which is the number of columns of the multi-scale feature matrix; sig(·) represents the Sigmoid activation function, which compresses the output value to the interval (0, 1); W t represents the weight matrix of the first fully connected layer, used to learn the importance of features in the time dimension; AvgPool f (·) represents the average pooling operation along the frequency axis, and AvgPool f (Z MSF ) represents the compression of the feature matrix Z MSF in the frequency dimension to extract global information in the time dimension; Z MSF is a multi-scale feature matrix with a dimension of C out ×T, representing the feature matrix output by the lth multi-scale convolution layer of the cascaded multi-scale convolution module, i.e. L is the total number of cascaded multi-scale convolution modules.

[0079] 2) Calculate the frequency-domain gating weight

[0080] In this embodiment, the multi-scale feature matrix is averaged and pooled along the time axis to compress the time dimension information, and a gating weight matrix in the frequency dimension is generated through a fully connected layer to emphasize the frequency band with significant leakage features in the frequency dimension, which is represented as:

[0081] G f= sig(W f · AvgPool t (Z MSF )

[0082] where G f denotes the frequency domain gating weight matrix with dimension C out × 1, which is used to emphasize the frequency band where the leakage feature is significant; W f denotes the weight matrix of the second fully connected layer, which is used to learn the feature importance in the frequency dimension; AvgPool t (·) denotes the average pooling operation along the time axis, and AvgPool t (Z MSF ) term represents the compression of the input feature matrix Z MSF in the time dimension to extract the global information in the frequency dimension.

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

[0084] In this embodiment, the time domain gating weight matrix and the frequency domain gating weight matrix are subjected to outer product operation 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:

[0085]

[0086] where G joint denotes the time-frequency joint weight matrix with dimension T × C out , which is used to model the feature significance in the time and frequency dimensions simultaneously; denotes the outer product operation, which expands two vectors into a two-dimensional matrix, and is used to capture the joint correlation between time and frequency.

[0087] 4) Purifying the features by applying the gating weight

[0088] The time-frequency joint weight matrix is multiplied element by element with the original multi-scale feature matrix to output a purified feature matrix, which further enhances the leakage-related features and suppresses non-coherent noise, and is represented as:

[0089] Z TFP = Z MSF ⊙ G joint

[0090] where Z TFP is the purified feature matrix, representing the feature matrix after time-frequency purification, with dimension C out × T.

[0091] It should be noted that, while the time domain gating weight matrix G t obtains the time positioning of the transient event, the frequency domain gating weight matrix G fThe outer product operation of the two produces a "focusing lens" effect on the time-frequency plane, which can automatically enhance the oblique distribution of the leakage harmonic cluster while suppressing the uniform distribution of the background noise.

[0092] This embodiment performs visual analysis of the effect of the time-frequency purification module, which directly displays the optimization effect of the time-frequency purification module on the feature matrix through a heat map. As shown in Figure 4 and Figure 5 , the original multi-scale feature map displays the original multi-scale feature matrix, the horizontal coordinate represents time (unit: millisecond), the vertical coordinate represents the frequency band index, and the color intensity represents the feature value size. It can be seen from the figure that the entire time-frequency plane is uniformly distributed with background noise, and the leakage feature area marked by the two rectangular boxes is not prominent. The purified feature map displays the feature matrix processed by the time-frequency purification module. Under the same coordinate system, the background noise is significantly weakened, and the leakage feature area is significantly enhanced. The comparison verifies the working mechanism of the dual-domain gating unit, which accurately locates the leakage event time period in the time domain, strengthens the sensitive frequency band in the frequency domain, and forms a "focusing lens" effect through the outer product operation, which effectively enhances the leakage harmonic cluster. The non-uniform distribution of feature intensity in the figure reflects the energy difference of different leakage levels, and the color scale on the right side of the heat map quantitatively displays the range of feature intensity change.

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

[0094] There is a long-range temporal dependence relationship caused by pressure wave propagation in the leakage acoustic signal, but the fully connected network will destroy the spatio-temporal structure of the signal, and the recurrent neural network is difficult to parallelize and has low computational efficiency, which cannot meet the real-time monitoring demand. The present application constructs a time-domain convolution and frequency-domain self-attention dual-path structure to capture the temporal relationship and frequency band energy correlation of sound wave propagation, and fuses the dual-path output to enhance the feature expression ability, the specific steps are as follows:

[0095] 1) Time domain path feature extraction

[0096] This embodiment uses a time series convolution network to process the purified feature matrix, which uses its dilated convolution structure to capture long-range temporal dependence, which is represented as TCN(Z TFP ), to obtain the feature matrix H t of the time domain path output, which strengthens the modeling of the temporal features caused by sound wave propagation; wherein TCN(·) represents a time series convolution network composed of multiple layers of dilated convolution and residual connection, used to extract long-range time domain dependence features; H t represents the feature matrix of the time domain path output, with a dimension of C out ×T, which enhances the modeling of temporal relationships.

[0097] It should be noted that the time convolution network is composed of multiple dilated convolution layers, each layer including dilated convolution, weight normalization, ReLU activation and residual connection, for example, a time convolution network with 4 layers, each layer with dilated rate of 1, 2, 4, 8, and convolution kernel size of 3, is used to extract long-range time domain dependence features.

[0098] 2) Frequency domain path feature extraction

[0099] The embodiment combines a self-attention mechanism in the frequency domain dimension, models the energy correlation between different frequency bands through a query matrix, a key matrix and a value matrix, outputs a frequency domain enhanced feature matrix, and further enhances the representation ability of the energy correlation between frequency bands, represented as:

[0100]

[0101] Wherein, Q is a query matrix, obtained by multiplying the input feature and the weight matrix W Q , used to calculate the attention weight, represented as Q=Z TFP W Q ; K is a key matrix, obtained by multiplying the input feature and the weight matrix W K , used to calculate the attention weight, represented as K=Z TFP W K ; V is a value matrix, obtained by multiplying the input feature and the weight matrix W V , used to generate output features, represented as V=Z TFP W V ; W Q is the weight matrix of the query transformation, with dimensions of C out xd k , which is a trainable parameter; W K is the weight matrix of the transformation, with dimensions of C out xd k , which is a trainable parameter; W V is the weight matrix of the value transformation, with dimensions of C out xd v , which is a trainable parameter; K T represents the transpose of the key matrix K; d k represents the dimension of the key vector, used to scale the attention score and stabilize the training process, preferably set to Soft(·) represents a normalized exponential function, used to calculate the attention weight distribution; H f represents the feature matrix output by the frequency domain path, with dimensions of C out xT, enhancing the modeling of energy correlation between frequency bands.

[0102] 3) Dual-path feature fusion

[0103] The embodiment combines the feature matrix output by the time domain path and the feature matrix output by the frequency domain path through weighted fusion and element-by-element multiplication to output a dual-path fusion feature matrix, integrates the time sequence relationship and the frequency band correlation feature, and enhances the overall expression capability, and is expressed as:

[0104] H fus = tanh(W ct H t +W cf H f )+H t ⊙H f

[0105] Wherein, tanh(·) represents a hyperbolic tangent activation function; W ct is a fusion weight matrix of the time domain path, with a dimension of C out ×C out , which is a trainable parameter; W cf is a fusion weight matrix of the frequency domain path, with a dimension of C out ×C out , which is a trainable parameter; H fus represents a dual-path fusion feature matrix, with a dimension of C out ×T.

[0106] It should be noted that the time sequence convolutional network of the time domain path captures the pressure wave propagation delay characteristics, which are associated with the harmonic energy modeled by the self-attention of the frequency domain path. The (W ct H t +W cf H f ) term maintains feature independence by adding terms, while the H t ⊙H f term produces a cross-domain modulation effect by element-by-element multiplication, using the time sequence information learned by the time domain path as a modulation signal to act on the frequency domain features, so that the frequency band energy correlation has time-varying characteristics, solving the defect that the recurrent neural network is difficult to model long-range frequency band correlation.

[0107] S304, leakage sensitive feature enhancement

[0108] The feature difference of different leakage levels mainly reflects on the high-frequency harmonic energy distribution, but the conventional feature extraction method relies on manual design, loses the phase information, and cannot adaptively enhance the leakage sensitive frequency band, resulting in that the micro-leakage feature is submerged by noise and the detection rate is low. The present application automatically learns and enhances the high-frequency harmonic mutation characteristics unique to leakage through a differentiable spectrum enhancement operator, preserves the phase information, and improves the recall rate of micro-leakage, and the specific steps are as follows:

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

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

[0111]

[0112] Wherein, ΔE represents the high-frequency energy change matrix, with a dimension of C out ×T, highlighting the high-frequency mutation caused by leakage; is the partial derivative symbol; represents the second derivative of the dual-path fusion feature matrix H fus along the frequency domain dimension, used for detecting high-frequency energy mutation points; W e represents the high-frequency energy weight matrix, with a dimension of C out ×C out , used to scale and transform the second derivative result, which is a trainable parameter; ReLU(·) represents the linear rectification activation function, which ensures that the energy change is non-negative.

[0113] 2) Feature enhancement output

[0114] Based on the high-frequency energy change matrix and the dual-path fusion feature matrix, the embodiment obtains a leakage-sensitive feature matrix, enhances the high-frequency mutation features and preserves the original phase polarity, and is expressed as:

[0115] Z LSFE = H fus + α·ΔE·Sign(H fus )

[0116] Wherein, α represents a gain coefficient, used to control the enhancement strength of the high-frequency energy change, and is preferably set to 0.3; Sign(·) represents the sign function, used to preserve the phase polarity information of the original features, and Sign(H fus ) calculates the sign of each element in the matrix H fus ; Z LSFE represents the leakage-sensitive feature matrix, with a dimension of C out ×T.

[0117] It should be noted that the Sign(H fus ) term preserves the phase polarity through the sign function, combined with the high-frequency mutation energy ΔE extracted by the term , to produce the technical effect of "phase coherent enhancement", which enhances the high-frequency harmonic energy while maintaining its phase relationship with the fundamental wave, so that the unique high-frequency harmonics of micro-leakage are displayed through phase consistency, solving the problem of harmonic structure damage caused by phase distortion in conventional methods.

[0118] S305, prediction and identification based on a dynamic threshold classifier

[0119] There is a serious class imbalance problem in natural gas pipeline leakage state classification, and the number of normal operation samples is much more than that of leakage samples. Fixed classification boundary will make the model biased to the majority class, resulting in small sample class, especially micro leakage, being ignored, and the classification performance declining. The present application adjusts the decision boundary adaptively through a dynamic temperature coefficient adjustment mechanism based on the sample density of the feature space, allocates greater gradient weight to the class with fewer samples, and improves the classification accuracy of the small sample class. The specific steps are as follows:

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

[0121] In this embodiment, the temperature coefficient is calculated by using a logarithmic function according to the number of samples in each class. The fewer the number of samples in a class, the greater the temperature coefficient, so that the class with fewer samples obtains a smoother decision boundary in the probability distribution, which is expressed as:

[0122] τ c =βlog(1+N / N c )

[0123] Where τ c represents the temperature coefficient of the cth class, which is used to adjust the smoothness of the decision boundary of the class in the Softmax function; β represents the scaling factor, which controls the overall amplitude of the temperature coefficient, and the preferred value is 0.5; N represents the total number of samples in the training data set; N c represents the number of samples of the cth class in the training data set; log(·) represents the logarithmic function, and the default base is the natural constant.

[0124] 2) Calculate the class logit value

[0125] In this embodiment, the compressed leakage sensitive feature matrix is mapped to the class space through a fully connected layer to obtain the original score of each class, which is expressed as:

[0126] s c =W c Z LSFE-pool +b c

[0127] Where Z LSFE-pool is the compressed leakage sensitive feature matrix, which is compressed from the leakage sensitive feature matrix Z out with a dimension of C LSFE ×T by using global average pooling to a vector with a dimension of C out ×1; W c represents the weight matrix of the third fully connected layer, with a dimension of C out ×C; b c represents the bias vector of the third fully connected layer; S crepresents the logit value of the c-th class, which is the original classification score.

[0128] 3) Calculate adaptive class probability

[0129] The embodiment uses the normalized exponential function adjusted by the temperature coefficient to calculate the prediction probability of each class, so that the class with less samples obtains more significant gradient weight, and the recognition ability of the small sample class is improved, which is represented as:

[0130]

[0131] wherein p(y=c|Z LSFE-pool ) represents the prediction probability of the sample belonging to the c-th class given the feature matrix Z LSFE-pool ; exp(·) represents the natural exponential function; represents the logit value of the c-th class; * represents the temperature coefficient of the c-th class; * y is the true class label of the sample; c is the class index, which ranges from 1 to C; C is the total number of classes.

[0132] 4) Prediction class determination

[0133] The embodiment takes the class index corresponding to the maximum prediction probability as the prediction class label of the sample. The final classification determination of the leakage state is completed, which is represented as:

[0134]

[0135] wherein represents the prediction class label of the sample; represents the class index corresponding to the maximum probability.

[0136] It should be noted that the temperature coefficient τ c adjusts the probability space geometry at the decision boundary level. When the number of leakage samples is extremely small, the temperature coefficient τ c exponentially amplifies the s c / τ c term, which leaves more feature space for the minority class samples. As the decision boundary expands outward, it can form a more compact cluster in the feature space for the micro-leakage samples, while avoiding excessive compression of the majority class samples, which significantly improves the sensitivity of the model to low-frequency events.

[0137] ​In the natural gas pipeline leakage acoustic classification task, the single cross-entropy loss function is insufficient to constrain the feature purification process, which easily leads to incomplete denoising, distortion or residual noise of the leakage feature, and further affects the classification robustness, the conventional method cannot jointly optimize the feature separation and classification performance, and it is difficult to maintain the integrity of the leakage physical characteristics under low signal-to-noise ratio conditions. The present application forms a hybrid loss function by constructing a double supervision mechanism combining classification loss and purification loss to jointly optimize the feature denoising, feature preservation and classification decision process, improve the discrimination ability and robustness of the model in a strong noise environment, the specific steps are as follows:

[0138] 1) Construct a hybrid loss function framework

[0139] In this embodiment, the hybrid loss function is formed by weighting and summing the classification loss and the purification loss through the loss weighting coefficient, and the feature denoising, feature preservation and classification decision process are jointly optimized, which is represented as:

[0140] Loss=L cls +λL pur

[0141] Wherein, Loss represents the hybrid loss function, which is the total optimization target of model training; L cls represents the classification loss, which is used to optimize the classification performance of the model on the leakage state, and the cross-entropy loss is preferably used as the classification loss; L pur represents the purification loss, which is used to constrain the feature purification process to avoid distortion of the leakage feature; λ represents the loss weighting coefficient, which controls the weight of the purification loss in the total loss, and the value is preferably 0.3.

[0142] 2) Purification loss calculation

[0143] In this embodiment, the L1 norm difference between the inverse time-frequency transform of the purified feature matrix and the inverse transform of the clean sample time-frequency feature matrix is calculated to obtain the purification loss, which constrains the feature purification process to approximate the clean sample feature and avoids distortion of the leakage feature, which is represented as:

[0144]

[0145] Wherein, L pur represents the purification loss; represents the clean sample time-frequency feature, which is generated by the wavelet threshold denoising method for the original sound pressure signal X raw preprocessing, specifically by applying wavelet transform to the original signal, soft threshold processing noise coefficients, then inverse wavelet transform to obtain the denoising signal, and then calculating the adaptive time-frequency feature; F -1 (·) represents the inverse time-frequency transform, which maps the time-frequency feature back to the time domain signal; ||·||1 represents the L1 norm.

[0146] It should be noted that the calculation of the purification loss adopts the L1 norm instead of the L2 norm, because the L1 norm is more robust to outliers, which helps to retain the sparsity of the leakage features, while the L2 norm will over-smooth the features, which may lead to the loss of details in the leakage.

[0147] In the process of iterative training and parameter updating of the leakage identification model, the embodiment adopts a staged training strategy to initialize the model parameters, and the specific steps are as follows:

[0148] First, fix the parameters of the adaptive time-frequency decomposition module, and use the Adam optimizer to pre-train the cascaded multi-scale convolution module to the dynamic threshold classifier with a learning rate of 0.001 for 100 rounds.

[0149] Then, release the adaptive time-frequency decomposition module constraint, and perform end-to-end joint training with a hybrid loss function as the optimization objective. During the training process, 32 segments of sound pressure signals are inputted in each batch, the gradients of the model prediction results and the true labels are calculated through the back propagation algorithm, and all trainable parameters are dynamically updated.

[0150] Every 10 rounds of training, the micro-leakage recall rate and the overall accuracy rate are evaluated on the validation set. When the indicators do not improve for 5 consecutive rounds, the learning rate is decayed to 1 / 10 of the original value.

[0151] The training termination condition is to reach 200 rounds, and the model parameters with the optimal performance on the validation set are finally saved.

[0152] The performance of different leakage detection methods in different signal-to-noise ratio environments is evaluated. The signal-to-noise ratio is an index for measuring the relative strength of effective information and noise in a signal, and the unit is decibel. The lower the signal-to-noise ratio, the stronger the noise interference, and the more difficult the detection. In this experiment, the short-time Fourier transform combined with the convolutional neural network method, the wavelet transform combined with the recurrent neural network method, the mel-frequency cepstral coefficient combined with the support vector machine method, and the intelligent comprehensive management method proposed in the present application are compared. Figure 6As 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.

[0153] 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 7 As 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.

[0154] In this embodiment, the detection performance of different methods under different leakage orifice sizes is evaluated, such as... Figure 8As 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 coefficient 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.

[0155] 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:

[0156] 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.

[0157] 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);

[0158] 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 control platform, and the timestamp and location station number of the abnormal signal are automatically marked, achieving a second-level response to leakage events.

[0159] S4. Conduct comprehensive management and control of natural gas pipelines based on the probability distribution of leakage levels.

[0160] In an optional embodiment of the present application, step S4 constructs a three-level management and control system based on the leakage monitoring results, specifically including:

[0161] 1) Risk early warning layer, combining leakage level probability output and historical data, visualizing high-risk sections of the pipeline through GIS maps, and generating a leakage risk heat map;

[0162] 2) Decision support layer, starting a multi-source data verification mechanism for continuously alarmed pipe sections, calling pressure, flow sensor data and voiceprint feature matrix of the corresponding section for cross verification, and automatically generating maintenance priority score;

[0163] 3) Response execution layer, automatically reducing the pressure of the upstream valve of the leakage pipe section through the pipeline pressure regulation module, and pushing a work order containing positioning coordinates, leakage level and recommended disposal scheme to the inspection terminal.

[0164] The feature matrix and disposal record of all leakage events are archived in the knowledge base for regular iteration of updating the leakage identification model parameters, forming a closed-loop intelligent management of "monitoring-diagnosis-disposal-optimization".

[0165] The embodiment of the present application also provides an intelligent comprehensive management system for natural gas pipeline protection, comprising:

[0166] A signal acquisition module is configured to acquire a sound pressure signal of the natural gas pipeline.

[0167] A feature decomposition module is configured to perform time-frequency feature decomposition on the sound pressure signal of the natural gas pipeline by using an adaptive basis function decomposition method to obtain a time-frequency feature matrix.

[0168] A leakage identification module is configured to construct a leakage identification model, and sequentially perform cascade multi-scale convolution coding, time-frequency feature purification, double-path feature fusion and leakage sensitive feature enhancement on the time-frequency feature matrix by using the leakage identification model, and perform dynamic classification decision according to the obtained leakage sensitive feature to obtain a leakage level probability distribution.

[0169] A comprehensive management and control module is configured to perform comprehensive management and control of the natural gas pipeline according to the leakage level probability distribution.

[0170] The present application is described with reference to flowcharts and / or block diagrams according to the method, device (system), and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device realize the functions described in the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 an apparatus for performing the functions specified in the flowchart or multiple flows and / or blocks.

[0171] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flowchart or multiple flows and / or blocks. Figure 1 one or more processes and / or blocks Figure 1 an apparatus for performing the functions specified in the flowchart or multiple flows and / or blocks.

[0172] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart or multiple flows and / or blocks. Figure 1 one or more processes and / or blocks Figure 1 an apparatus for performing the functions specified in the flowchart or multiple flows and / or blocks.

[0173] The principles and implementations of the present application are described in the embodiments, the above description of the embodiments is only for helping to understand the method of the present application and its core idea; meanwhile, for the ordinary skilled in the art, according to the idea of the present application, the specific implementation and application range will have changes, and the above description of the present application should not be understood as the limitation of the present application.

[0174] The person skilled in the art will understand that the embodiments described herein are for helping the reader to understand the principles of the present application and should be understood as the protection scope of the present application not being limited to such specific statements and embodiments. The person skilled in the art can make various other specific modifications and combinations according to the technical inspirations disclosed in the present application without departing from the essence of the present application, and these modifications and combinations still fall within the protection scope of the present application.

Claims

1. A method for intelligent integrated management of natural gas pipeline protection, characterized in that, The method comprises the following steps: acquiring a natural gas pipeline sound pressure signal; performing time-frequency feature decomposition on the natural gas pipeline sound pressure signal by using an adaptive basis function decomposition method to obtain a time-frequency feature matrix, wherein the adaptive basis function is constructed; element-wise multiplication and summation operation is performed on the natural gas pipeline sound pressure signal and the adaptive basis function to generate the time-frequency feature matrix; a parameter set of the basis function is optimized by using a gradient descent-based back propagation algorithm; a leakage identification model is constructed, and the time-frequency feature matrix is sequentially subjected to cascade multi-scale convolution coding, time-frequency feature purification, double-path feature fusion and leakage sensitive feature enhancement by using the leakage identification model, and a leakage level probability distribution is obtained according to the obtained leakage sensitive feature; the cascade multi-scale convolution coding comprises: setting a multi-scale dilation rate parameter; performing multi-scale dilation convolution operation on the time-frequency feature matrix, summing the convolution results of different dilation rates, and adopting batch normalization and a rectified linear unit activation function to obtain a multi-scale feature matrix; performing comprehensive management and control of the natural gas pipeline according to the leakage level probability distribution.

2. The intelligent comprehensive management method for natural gas pipeline protection according to claim 1, characterized in that, The time-frequency feature purification comprises: a double-domain gating unit is used to extract significant features from the time and frequency dimensions respectively, and a time-frequency joint weight matrix is constructed by using outer product operation; the time-frequency joint weight matrix is multiplied with the multi-scale feature matrix element by element to obtain a purified feature matrix.

3. The intelligent comprehensive management method for natural gas pipeline protection according to claim 2, characterized in that, The double-path feature fusion comprises: a time series convolution network is used to process the purified feature matrix, and an expansion convolution structure is used to capture long-range time series dependence to obtain a time domain path feature matrix; a self-attention mechanism is combined in the frequency domain dimension, and the energy correlation between different frequency bands is modeled by using a query matrix, a key matrix and a value matrix to obtain a 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 a double-path fusion feature matrix.

4. The intelligent comprehensive management method for natural gas pipeline protection according to claim 3, characterized in that, The leakage sensitive feature enhancement comprises: a second derivative is calculated along the frequency domain dimension of the double-path fusion feature matrix, and a high-frequency energy change matrix is extracted by using a high-frequency energy weight matrix and a linear rectification activation function; a leakage sensitive feature matrix is calculated based on the high-frequency energy change matrix and the double-path fusion feature matrix.

5. The intelligent comprehensive management method for natural gas pipeline protection according to claim 4, characterized in that, The leakage sensitive feature matrix specifically comprises: Z LSFE = H fus + a·△E·Sign(H fus ) where Z LSFE represents a leakage sensitive feature matrix, H fus represents a dual-path fusion feature matrix, a represents a gain coefficient, ΔE represents a high-frequency energy variation matrix, Sign(·) represents a sign function, ReLU(·) represents a linear rectifier activation function, W e represents a high-frequency energy weight matrix, is a partial derivative sign, and f represents a frequency variable.

6. The intelligent comprehensive management method for natural gas pipeline protection according to claim 1, characterized in that, dynamic classification decision is made according to the obtained leakage sensitive feature to obtain a leakage level probability distribution, which comprises: a temperature coefficient is calculated by using a logarithmic function according to the number of samples of each category; the leakage sensitive feature matrix is mapped to a category space by using a fully connected layer to obtain original scores of each category; a normalized exponential function with a temperature coefficient is used to calculate the prediction probability of each category based on the original scores of each category; the category index corresponding to the maximum prediction probability is selected as the predicted category label of the sample to obtain the leakage level probability distribution.

7. The intelligent comprehensive management method for natural gas pipeline protection according to claim 1, characterized in that, The leakage identification model uses a hybrid loss function composed of a classification loss and a purification loss during training, and the hybrid loss function specifically comprises: Loss = L cls + λL pur wherein, Loss is a mixed loss function, L cls is a classification loss, L pur is a purification loss, and λ is a loss weighting coefficient; The purification loss specifically comprises: where F -1 (·) is the inverse time-frequency transform, Z TFP is the purified feature matrix, is the clean sample time-frequency feature, and ||·||1 is the L1 norm.

8. An intelligent integrated management system for natural gas pipeline protection, characterized in that, a signal acquisition module is configured to acquire a natural gas pipeline sound pressure signal; ​ The feature decomposition module is configured to perform time-frequency feature decomposition on the sound pressure signal of the natural gas pipeline by using an adaptive basis function decomposition method to obtain a time-frequency feature matrix, wherein the adaptive basis function is constructed, the sound pressure signal of the natural gas pipeline is multiplied by the adaptive basis function element by element, and summation operation is performed to generate the time-frequency feature matrix, and the parameter set of the basis function is optimized by using a back propagation algorithm based on gradient descent; The leakage identification module is configured to construct a leakage identification model, perform cascade multi-scale convolution coding, time-frequency feature purification, double-path feature fusion and leakage sensitive feature enhancement on the time-frequency feature matrix in sequence by using the leakage identification model, and perform dynamic classification decision according to the obtained leakage sensitive feature to obtain a leakage grade probability distribution, wherein the cascade multi-scale convolution coding comprises setting a multi-scale dilation rate parameter, performing multi-scale dilated convolution operation on the time-frequency feature matrix, summing the convolution results of different dilation rates, and adopting batch normalization and a rectified linear unit activation function to obtain a multi-scale feature matrix; The comprehensive management and control module is configured to perform comprehensive management and control of the natural gas pipeline according to the leakage grade probability distribution.

Images