Multi-dimensional acoustic feature fusion small sample learning pipeline leakage identification method

Abstract

The present application relates to the technical field of heat supply pipe network, in particular to a multi-dimensional acoustic feature fusion small sample learning pipeline leakage identification method, comprising the following steps: collecting sound pressure data based on the main line of the heat supply pipe network, analyzing signal amplitude and frequency characteristics by sliding window, constructing a multi-dimensional acoustic feature set, screening and aggregating leakage and normal state samples, comparing the differences between new data and prototypes, quantifying feature distance, and determining the leakage state according to the minimum attribution principle. The present application establishes multi-angle behavior expression, generates typical category prototypes through feature induction, combines multi-category feature center aggregation and dynamic measurement, improves the high reliable discrimination ability of the leakage state, compared with the traditional technology, the multi-dimensional feature fully represents the leakage signal, effectively enhances the adaptability to environmental noise, sudden interference and complex pipe network structure, reduces the misjudgment, and the small sample learning technology greatly reduces the dependence on a large number of labeled data, significantly reduces the data collection and labeling cost.

Description

Technical Field

[0001] This invention relates to the field of heating pipeline technology, and in particular to a small-sample learning method for pipeline leak identification by fusing multi-dimensional acoustic features. Background Technology

[0002] Heating pipeline networks involve technologies related to the layout, operation, maintenance, and fault diagnosis of heating pipelines in urban centralized heating systems. These include the pipeline layout structure, heat medium transport efficiency, heat loss control, operational safety management, and fault detection and identification. During the heating process, the pipeline system permeates the urban underground space, bearing the task of transporting high-temperature and high-pressure heat media. Leaks and other faults not only waste energy but also pose operational risks to the pipeline network. Therefore, research on leak detection and condition monitoring technologies is of great significance. Traditional pipeline leak identification methods involve using point sensors installed in the heating network to acquire changes in pipe surface temperature and pressure, or using simple acoustic sensors to detect characteristic acoustic signals generated by leaks, and then judging the presence of a leak based on manual experience or feature template matching.

[0003] Existing pipeline leak detection technologies mainly rely on point sensors or acoustic detection. However, acoustic signals are easily affected by multipath superposition and external noise in underground environments, resulting in unstable leak signal characteristics that are easily masked by background interference. Thresholding methods are susceptible to environmental or sudden noise interference, leading to a high false positive rate. Artificial intelligence-based methods rely on a large amount of labeled data, but pipeline leak data is difficult to obtain, and model performance is insufficient in small sample scenarios. Traditional signal processing methods (such as Fourier transform and wavelet transform) are difficult to cope with feature interference and distortion caused by reflection, refraction, and superposition in complex pipeline structures, resulting in poor adaptability and low accuracy in leak feature extraction and identification, which is difficult to meet practical needs. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a small-sample learning method for pipeline leak identification by fusing multi-dimensional acoustic features. The technical solution is as follows:

[0005] On the one hand, a few-sample learning method for pipeline leak identification by fusing multi-dimensional acoustic features is provided, including the following steps:

[0006] S1: Based on the main pipeline of the heating network, analyze the acquired signals of the acoustic sensors, distinguish the target frequency band and noise interference, digitize the sampled signals through the analog-to-digital conversion unit, and arrange them in sequence according to the acquisition time to obtain the pipeline sound pressure data sequence;

[0007] S2: Based on the pipeline sound pressure data sequence, a sliding window process is used to analyze the signal amplitude changes in each window, calculate the amplitude variation trend, summarize the energy change parameters, and combine the frequency characteristics to obtain a multi-dimensional acoustic feature set;

[0008] S3: Based on the multidimensional acoustic feature set, filter leakage and normal structure features, input various types of samples into the recognition unit, adjust the fusion order, merge the features of samples of the same category, and obtain a set of category prototype vectors;

[0009] S4: Based on the set of category prototype vectors, compare the structural differences between the unlabeled data and the prototypes, analyze the category classification through feature correspondence, quantify the differences of key features, and obtain the feature distance measurement parameters;

[0010] S5: Based on the aforementioned feature distance measurement parameters, determine the direction of difference between the sample and the leaking and normal prototypes, read the structural deviation data, classify the sample leakage status according to the least attribution principle, and obtain the pipeline leakage determination result.

[0011] On the other hand, the pipeline sound pressure data sequence includes time stamps, sound pressure amplitude, and channel number; the multidimensional acoustic feature set includes sound intensity features, spectral features, power spectrum features, and acoustic signature features; the category prototype vector set includes leakage category prototypes and normal category prototypes; the feature distance metric parameters include leakage category distance and normal category distance; and the pipeline leakage determination result includes the determination category and confidence score.

[0012] On the other hand, the specific steps for obtaining the pipeline sound pressure data sequence are as follows:

[0013] S101: Based on the main line of the heating pipeline network, analyze the output signal of the acoustic sensor, construct frequency comparison conditions according to the boundary parameters of the operating frequency band, calculate the frequency distribution of the complete signal, locate the frequency segment corresponding to the operating frequency band, and obtain the frequency band aggregated acoustic segment set.

[0014] S102: Based on the frequency band aggregated acoustic segment set, analyze the continuous change pattern of the segments on the time axis, compare the sudden fluctuations of the envelope with the morphological differences of the continuous segments, identify segments with continuous characteristics, remove interfering segments, and recombine the remaining segments to obtain a continuous and effective segment sequence.

[0015] S103: Based on the continuous effective segment sequence, adjust the analog signal structure of each sampling point, complete the analog-to-digital conversion, determine the time position of each sampling point, and integrate the corresponding channel identifier and digital signal to obtain the pipe sound pressure data sequence.

[0016] On the other hand, the specific steps for obtaining the multidimensional acoustic feature set are as follows:

[0017] S201: Based on the pipe sound pressure data sequence, analyze the continuous sound pressure amplitude and time label, determine the fluctuation trend of the signal amplitude curve in each time period, compare the amplitude trend of adjacent window segments, identify the signal interval with continuous amplitude change, optimize the amplitude change trajectory in the data frame, and obtain the amplitude dynamic change trajectory set.

[0018] S202: Based on the aforementioned dynamic amplitude change trajectory set, calculate the amplitude fluctuation direction of each curve, determine the change characteristics of the start and end points, analyze the curve distribution of increasing, decreasing and stable modes, identify curve sequences that conform to the amplitude change law, summarize the change mode categories, and obtain the dynamic trend induction parameter set.

[0019] S203: Based on the dynamic trend induction parameter set, integrate the frequency distribution data of each time period, analyze the correlation between each parameter, adjust the arrangement order of each feature parameter in the data sequence, map and fuse the acoustic parameters, and obtain a multidimensional acoustic feature set.

[0020] On the other hand, the specific steps for obtaining the set of category prototype vectors are as follows:

[0021] S301: Based on the multidimensional acoustic feature set, analyze the sound intensity parameters and spectrum parameters in each data frame, determine the distribution of leakage scene and normal scene in the feature space, compare the change pattern of feature combination under the two types of scenes, identify feature groups with discriminative ability, and obtain the scene feature expression structure.

[0022] S302: Based on the scene feature expression structure, determine the dense distribution area of ​​each type of sample in the multidimensional space, analyze the spatial distance between sound intensity, frequency, power spectrum and voiceprint features, identify spatially similar parameters and merge them to obtain a set of clustering feature parameters;

[0023] S303: Based on the clustering feature parameter set, calculate the spatial center position of each group of feature parameters, optimize the feature classification method, adjust the arrangement of data frames and the channel mapping order, summarize the parameter structure with aggregation, and obtain the category prototype vector set.

[0024] On the other hand, the specific steps for obtaining the feature distance metric parameters are as follows:

[0025] S401: Based on the set of prototype vectors of the categories, analyze the spatial distribution of the multidimensional acoustic feature set of the unlabeled data on each feature parameter, compare the variation law of acoustic features among the prototypes, determine the structural differences between each feature parameter, and obtain a spatial feature comparison array.

[0026] S402: Based on the spatial feature comparison array, calculate the spatial correspondence between unlabeled data and prototypes of each category in the feature dimension, identify the structural differences of feature parameters, determine the spatial similarity between each group of parameters and the category to which they belong, and obtain heterogeneous parameter indication groups;

[0027] S403: Based on the heterogeneous parameter indication group, determine the category discrimination ability of each feature parameter, quantify the spatial distance performance of each parameter in prototype attribution discrimination, organize the distance data between unlabeled data and prototypes, optimize the arrangement order of feature parameters, and obtain feature distance measurement parameters.

[0028] On the other hand, the specific steps for obtaining the pipeline leakage determination result are as follows:

[0029] S501: Based on the feature distance measurement parameters, analyze the spatial relationship between the sample to be judged and the leakage behavior prototype and the normal behavior prototype, determine the distance direction between each prototype, compare the attribution trend of the sample features in space, map the spatial offset direction with the category judgment relationship, and obtain offset attribution judgment data.

[0030] S502: Based on the offset attribution determination data, calculate the degree of deviation between the sample features and each prototype in the feature space, compare the spatial trajectory with the category attribution conditions, filter the parameter paths that are consistent with the attribution relationship, and organize the spatial attribution data of each category to obtain the attribution path parameter set.

[0031] S503: Based on the attribution path parameter set, determine the category label corresponding to the attribution direction, compare the mapping between the sample space determination path and the category label, organize the label results and recognition order of the sample determination, and obtain the pipeline leakage determination result.

[0032] On the other hand, the external acoustic sensor refers to a sensor installed outside the heating network pipeline to collect information on the flow of the medium inside the pipeline, leakage, or environmental noise, and the analog-to-digital conversion unit refers to the circuit part that converts the analog signals collected by the sensor into digital signals.

[0033] On the other hand, the sliding window processing refers to the process of using a fixed-length time window for long-term acoustic signals, with the window gradually sliding and the features analyzed independently in each window. The signal amplitude change refers to the amplitude of the acoustic signal at each sampling point, reflecting the intensity of the sound wave.

[0034] On the other hand, the leakage state and normal structure refer to the acoustic signal feature categories corresponding to the pipeline being in a leakage condition or a normal condition, respectively, and the structural difference refers to the numerical distance or distribution difference between the multidimensional acoustic feature set of the newly acquired signal and the existing behavioral prototype in each feature dimension.

[0035] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0036] By integrating multi-dimensional features such as sound intensity, spectrum, power spectrum, and acoustic signature, a multi-angle behavioral expression is established. Typical category prototypes are generated through feature induction. Combined with multi-category feature center aggregation and dynamic measurement, the ability to reliably identify leakage status is improved. Compared with traditional technologies, multi-dimensional features comprehensively characterize leakage signals, effectively enhance adaptability to environmental noise, sudden interference, and complex pipeline structures, and reduce misjudgments. The few-sample learning technology significantly reduces the dependence on a large amount of labeled data, significantly reducing data collection and labeling costs. It can achieve rapid location and automated classification of leakage status in continuous signal monitoring, complex environments, and few-sample scenarios, with better overall performance. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is the overall flowchart of the present invention;

[0039] Figure 2 This is a technical roadmap for the present invention;

[0040] Figure 3 This is a flowchart of steps S1 of the present invention;

[0041] Figure 4 This is a flowchart of steps S2 of the present invention;

[0042] Figure 5 This is a flowchart of steps S3 of the present invention;

[0043] Figure 6 This is a flowchart of step S4 of the present invention;

[0044] Figure 7 This is a flowchart of step S5 of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0046] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0047] This invention provides a few-sample learning method for pipeline leak identification by fusing multi-dimensional acoustic features, such as... Figure 1As shown, it includes the following steps:

[0048] S1: Based on the main line of the heating pipeline network, analyze the continuously acquired signals of the acoustic sensors, including external wall type and insertion type. Separate the target frequency band generated during equipment operation, determine the non-operational noise interference in the signal sequence, use the analog-to-digital conversion unit to digitize each sampling point in sequence, gradually integrate the digital signals, and arrange them in order according to the acquisition time to obtain the pipeline sound pressure data sequence.

[0049] S2: Based on the pipe sound pressure data sequence, optimize the sliding window processing flow, analyze the signal amplitude variation curve of each window segment over time, calculate the variation trend of signal amplitude in each segment, summarize the correlation parameters of energy variation over time, combine the frequency characteristics within each segment, and integrate all acoustic parameters through feature mapping to obtain a multidimensional acoustic feature set.

[0050] S3: Based on the multidimensional acoustic feature set, filter the feature performance in the leakage scenario, distinguish the structural features of the leakage state and the normal state, input each group of samples into the embedded recognition unit, adjust the fusion order of structural features, aggregate the parameters of similar behaviors in turn, merge the multidimensional acoustic features of samples of the same category, and obtain the category prototype vector set.

[0051] S4: Based on the set of category prototype vectors, compare the structural differences between the multidimensional acoustic feature set of unlabeled data and each prototype. Analyze the category classification through the correspondence of structural features, quantify the feature differences in the structural discrimination process, organize the key parameters that reflect the differences of each prototype, and obtain the feature distance measurement parameters.

[0052] S5: Based on the feature distance metric parameter, determine the direction of difference between the sample and the leakage behavior prototype and the normal behavior prototype, read the data corresponding to the structural deviation, and classify the current sample as a leakage state or a normal state according to the principle of minimum deviation, so as to obtain the pipeline leakage judgment result.

[0053] The pipeline sound pressure data sequence includes time stamps, sound pressure amplitude, and channel number. The multidimensional acoustic feature set includes sound intensity features, spectral features, power spectrum features, and acoustic signature features. The category prototype vector set includes leakage category prototypes and normal category prototypes. The feature distance measurement parameters include leakage category distance and normal category distance. The pipeline leakage determination result includes the determination category and confidence score.

[0054] like Figure 2 As shown, a passive noise leakage identification method suitable for complex pipe networks is constructed by combining "multi-dimensional signal feature integration" with "small sample model training," including:

[0055] Multi-dimensional anti-interference feature extraction and fusion;

[0056] To address the interference characteristics of leakage signals in complex pipeline networks, complementary features are extracted from multiple dimensions:

[0057] Temporal characteristics: The root mean square value of the signal strength is traversed by a short window of 50-150ms to capture the changes in audio intensity between the leaked audio and the normal operating audio.

[0058] Spectral characteristics: The time-frequency distribution of the signal is obtained through short-time Fourier transform, capturing the dynamic frequency changes of the leakage signal over time;

[0059] Power spectrum characteristics: Frequency-energy distribution is extracted by power spectrum estimation, highlighting the energy accumulation characteristics of the leakage-related frequency band;

[0060] Voiceprint features: Based on Mel frequency cepstral coefficients (MFCC) or voiceprint feature extractor, it simulates the characteristics of human ear perception, enhances the characterization of low-frequency anti-attenuation leakage signals, and suppresses high-frequency environmental noise interference.

[0061] Feature fusion: The four types of features are integrated into a 4-channel tensor (the first channel is the spectral feature matrix, the second channel is the power spectral feature matrix, the third channel is the acoustic signature feature matrix, and the fourth channel is the acoustic intensity matrix), forming a comprehensive feature matrix covering time, frequency, energy, and perception dimensions, which enhances the anti-interference ability against reflection, refraction, and superposition interference.

[0062] Small sample model training mechanism;

[0063] By mining the similarities (common features of similar leakage signals) and differences (distinguishing features between leakage and non-leakage, and different leakage intensities) between samples, a small-sample learning model is constructed.

[0064] By using a small number of labeled samples (e.g., only a few dozen samples per scenario), the model can quickly learn the discriminative features of the leaked signal through an episodic training method of "support set-query set", reducing the dependence on large-scale labeled data.

[0065] End-to-end identification process;

[0066] Original noise signal preprocessing (denoising, normalization) → multi-dimensional feature extraction → feature fusion → few-sample model inference → output leakage identification result (leakage or not), realizing automated identification from signal input to result output.

[0067] Specifically, it includes four stages: signal acquisition and preprocessing, multi-dimensional feature extraction and fusion, few-sample CNN model training, and model inference and result output, as detailed below:

[0068] Phase 1: Signal acquisition and preprocessing (providing high-quality raw data for feature extraction);

[0069] S1A: Pipeline noise signal acquisition;

[0070] A passive noise monitoring method is adopted, in which noise acquisition devices (such as microphones, piezoelectric sensors, etc.) are deployed in the heating pipeline network to collect the original acoustic signals (including leakage signals, normal operation noise, environmental interference noise, etc.) during pipeline operation.

[0071] Sampling parameters: The sampling frequency is set according to the main frequency range of the pipeline leakage signal (usually 2kHz-20kHz) to ensure coverage of the main frequency components of the leakage signal; the sampling duration is divided according to a fixed time window (e.g., 1-5 seconds for each signal segment) to facilitate subsequent frame processing.

[0072] S1B: Signal denoising preprocessing;

[0073] For environmental noise (such as traffic noise, equipment vibration) and sudden noise (such as instantaneous impact sound) in the original signal, adaptive filtering (such as the LMS algorithm) or wavelet thresholding denoising is used:

[0074] Wavelet thresholding denoising: Select an appropriate wavelet basis (such as db4) and decomposition level (3-5 levels) to perform multi-scale decomposition on the signal, and suppress high-frequency noise components through a threshold function (such as soft threshold) while retaining the effective components of low-frequency leakage signals.

[0075] Objective: To initially filter out interference unrelated to pipeline leaks and reduce noise impact during subsequent feature extraction.

[0076] S1C: Signal normalization processing;

[0077] The denoised signal is then subjected to numerical scaling, using Z-score normalization:

[0078] Calculate the mean (μ) and standard deviation (σ) of the signal, and map the signal value to a distribution with a mean of 0 and a standard deviation of 1 using the formula x_norm=(x-μ) / σ.

[0079] Objective: To eliminate amplitude differences in signals from different sensors and at different times (such as signal strength changes caused by pipeline pressure fluctuations) and ensure the stability of feature extraction.

[0080] Phase 2: Multi-dimensional feature extraction and fusion (constructing comprehensive features resistant to interference);

[0081] S2A: Sound intensity feature extraction;

[0082] Use a short window of 50-150ms (e.g., 100ms) and a frame overlap rate of 50%-75% (e.g., 60%, to preserve continuous information between frames and reduce feature breaks caused by frame division) to slide through the signal.

[0083] The root mean square (RMS) value of the signal strength within each window is calculated, which reflects the average energy level of the sound intensity within the window and suppresses transient impulse noise interference. The sound intensity feature matrix is ​​output with the shape (T, 1) (T is the number of time frames, which is determined by the total signal duration, window length, and overlap rate. For example, a 10-second signal is calculated with a 100ms window and 60% overlap, so T≈246). The matrix elements represent the root mean square value of the signal strength of a "certain time frame".

[0084] S2B: Spectral Feature Extraction (Time-Frequency Dimension);

[0085] The normalized signal is processed using Short-Time Fourier Transform (STFT):

[0086] Frame segmentation parameters: frame length 20-50ms (e.g., 32ms), frame overlap rate 50%-75%, window function selected is Hanning window (to suppress spectral leakage).

[0087] Calculate the spectrum for each time frame and output the spectrum feature matrix with the shape (T, F) (T is the number of time frames and F is the number of frequency points, such as T=100 and F=128). The matrix elements represent the signal amplitude of "a certain time frame - a certain frequency point".

[0088] Function: To capture the dynamic frequency changes of leakage signals over time.

[0089] S2C: Power Spectrum Feature Extraction (Frequency-Energy Dimension);

[0090] The power spectrum of the signal is estimated using the Welch method:

[0091] The signal is divided into frames (with the same frame length and overlap rate as STFT), and the Fourier transform result of each frame is calculated. The amplitude squared is then averaged (averaging the amplitude squared at corresponding points within adjacent K frames) to obtain the power spectrum feature matrix, which has the shape (T, F) (aligned with the time frame and frequency point of the spectrum feature). The matrix elements represent the energy value of "a certain time frame - a certain frequency point".

[0092] Function: To highlight the concentrated energy frequency band of the leakage signal (such as the inherent resonant frequency band of pipeline leakage), and the energy characteristics remain relatively stable even if the signal is reflected and superimposed.

[0093] S2D: Voiceprint Feature Extraction (Perception Dimension);

[0094] Voiceprint features extracted based on Mel-frequency cepstral coefficients (MFCC):

[0095] Preprocessing: Pre-emphasis (boosting high-frequency components), framing (same as STFT), and windowing of the signal;

[0096] Mel filtering: Maps the spectrum to the Mel frequency axis using 20-40 Mel filter banks (simulating the human ear's sensitivity to low frequencies);

[0097] Cepstral Transform: After taking the logarithm of the Mel spectrum, perform a Discrete Cosine Transform (DCT), take the first 12-16 order coefficients (including the DC component), and obtain the MFCC characteristic matrix with the shape (T, M) (M is the order of the Mel coefficients, such as 40).

[0098] If the time frame (T) of the MFCC is inconsistent with the spectrum / power spectrum, the time dimension is unified by interpolation (e.g., the T of the MFCC is adjusted to be the same as the spectrum characteristics).

[0099] Function: Enhances the characterization of low-frequency leakage signals (which attenuate slowly in pipelines) and suppresses high-frequency environmental noise interference.

[0100] S2E: Multi-feature fusion (constructing multi-channel tensors);

[0101] The sound intensity matrix, spectral feature matrix, power spectrum feature matrix, and MFCC feature matrix after unifying the time dimension are concatenated according to the "channel dimension" to form a multi-channel tensor with the shape (T, F, 4) (the four channels correspond to the four types of features respectively).

[0102] Function: Integrates complementary information from time, frequency, energy, and perception dimensions to provide "multi-angle feature input" that is resistant to interference for subsequent CNNs.

[0103] Phase 3: Training of the few-sample CNN model (building a discriminative model based on a small number of samples);

[0104] S3A: Small sample dataset partitioning;

[0105] From the labeled samples (including "leaking" and "normal operation"), the support set and query set are divided according to the "episodic task" mechanism:

[0106] k samples are randomly selected from each class (k=5-20, such as 10) to form a support set (used to build class prototypes).

[0107] q samples (q=5-20, e.g., 10) are randomly selected from the remaining samples of each class to form a query set (used for model validation and optimization).

[0108] Repeatedly divide the data to generate multiple episodic tasks (e.g., 1000 groups) to cover different sample combination scenarios.

[0109] Example: 50 leaked samples + 50 normal samples → For each task, take 10 leaked samples + 10 normal samples as the support set, and 10 leaked samples + 10 normal samples as the query set.

[0110] S3B: Construction of CNN Feature Extraction Network;

[0111] Network structure (adapted for multi-channel tensor input):

[0112] Input layer: Receives multichannel tensors of shape (T, F, 4) (e.g., 100, 128, 4).

[0113] Convolutional layer 1: 32 3×3 convolutional kernels, stride 1, activation function ReLU, output feature map (98, 126, 32) (edge ​​reduction 2).

[0114] Pooling layer 1: 2×2 max pooling, step size 2, output (49, 63, 32).

[0115] Convolutional layer 2: 64 3×3 convolutional kernels, stride 1, ReLU activation, output (47, 61, 64).

[0116] Pooling layer 2: 2×2 max pooling, step size 2, output (23, 30, 64).

[0117] Fully connected layer: Flatten the pooled features (23×30×64=44160 dimensions) and map them into 256-dimensional feature vectors through the fully connected layer.

[0118] Function: To compress high-dimensional multi-channel features into low-dimensional discriminant vectors while preserving the core patterns of the leaked signal (such as cross-channel feature correlation).

[0119] S3C: Few-sample training based on prototype networks;

[0120] Taking the Prototypical Network as an example, training it in conjunction with a CNN:

[0121] Building category prototypes: Supports inputting concentrated samples into the CNN, obtaining feature vectors, and then calculating the mean of each class vector as the "prototype" (e.g., leak prototype = mean of 10 leaked sample vectors, normal prototype = mean of 10 normal sample vectors).

[0122] Calculate the query set distance: Input the samples in the query set into the CNN to obtain feature vectors, and calculate the Euclidean distance with the leaked prototype and the normal prototype (the smaller the distance, the higher the probability of belonging to that class).

[0123] Loss function optimization: The cross-entropy loss function is adopted, with the goal of minimizing the difference between the "true class of the query sample and the distance determination result". The convolutional kernel weights and fully connected layer parameters of the CNN are updated through backpropagation.

[0124] Multi-round iterative training: Repeat steps S2E-S3B (changing the episodic task) for 500-2000 rounds until the loss function converges (e.g., the validation set loss is stable).

[0125] Function: To enable CNNs to learn to extract common features of "leaked / normal signals" from a small number of samples, reducing their dependence on large-scale labeled data.

[0126] Phase 4: Model Inference and Result Output (Achieving Automated Leakage Monitoring)

[0127] S4A: New signal feature processing;

[0128] For the newly acquired pipeline noise signal from the sensor, repeat steps S1B (denoising), S1C (normalization), S2A-S2D (extracting sound intensity / spectrum / power spectrum / MFCC features), and S2E (fusion into a multi-channel tensor) to generate input features consistent with the training data format.

[0129] S5A: Leakage Identification and Result Output;

[0130] The multi-channel tensor of the new signal is input into the trained CNN to obtain a 256-dimensional feature vector.

[0131] Calculate the Euclidean distance between this vector and the "leaked prototype" and "normal prototype" saved during the training phase;

[0132] Judgment rule: If the distance to the leaking prototype is smaller, output "Leak"; otherwise output "Normal operation" to complete automated monitoring.

[0133] In S1, external acoustic sensors refer to sensors (such as microphones, piezoelectric sensors, etc.) installed on the outside of the heating network pipeline to collect data on the flow of the medium inside the pipeline, leakage, or environmental noise; equipment refers to the relevant hardware facilities in the pipeline system, including sensors, data acquisition units, signal conditioning and sampling equipment, etc.; target frequency band refers to the key frequency range determined based on the characteristics of the leakage acoustic signal for subsequent analysis, usually covering the part of the frequency band where the leakage signal energy is most concentrated; signal sequence refers to the digital acoustic signal (i.e., the time series of sound pressure data) formed by the sensor and analog-to-digital conversion, arranged in the sampling order; non-operational noise interference refers to other noise sources that are not generated by the normal operation or leakage of the pipeline, such as traffic, equipment vibration, personnel activity, and other external noise; analog-to-digital conversion unit refers to the circuit part that converts the analog signal collected by the sensor into a digital signal, commonly an ADC (analog-to-digital converter); digitization refers to the entire process of converting the signal from an analog quantity into a discrete digital quantity, that is, the operation of converting the analog signal into a digital form after sampling and quantization.

[0134] In S2, the sliding window processing flow refers to the process of using a fixed-length time window for long-term acoustic signals, with the window gradually sliding and features analyzed independently within each window. It is commonly used to extract short-term statistics, such as the root mean square (RMS). Signal amplitude refers to the amplitude (or voltage, etc.) of the acoustic signal at each sampling point, reflecting the intensity of the sound wave. The variation curve refers to the continuous trend of the signal amplitude over time on the time axis; this curve reveals the signal's periodicity, abrupt changes, and trends. The variation trend refers to the change pattern of a certain statistic (such as amplitude or energy) with the time window, such as increase, decrease, or abrupt change. The correlation parameters of energy over time describe the acoustic signal energy and its related characteristics (such as RMS, power spectrum, etc.) within each analysis window. Frequency characteristics refer to the distribution of frequency components obtained through frequency domain analysis such as Fourier transform within each time window, used to characterize the spectral characteristics of leakage or noise. Feature mapping methods refer to the way to map and integrate different types and dimensions of original acoustic parameters (such as amplitude, frequency, energy, MFCC, etc.) into a unified feature format, including steps such as normalization and tensor splicing.

[0135] In S3, "feature performance" refers to the characteristic differences and typical performance of acoustic signals under different scenarios (leakage, normal) (such as energy level, frequency distribution, etc.); "leakage state" and "normal state" refer to the acoustic signal feature categories corresponding to the pipeline being in a leaking or normal operating condition, respectively, which are the labels (classification basis) of the samples; "each group of samples" refers to the collection of signal features with known labels (leakage / normal), which are usually used to train models or extract prototypes; "embedded recognition unit" refers to the hardware device (such as microcontroller, ARM module, etc.) deployed on the pipeline network or acquisition node that can process and classify acoustic features in real time; "structural feature fusion order" refers to the specific process order of fusion and sorting of feature parameters from different sources and of different types during the model or prototype construction process; "parameters of the same behavior" refers to the acoustic feature parameters extracted from all samples under the same category (leakage or normal), which are used to merge and aggregate to form the typical features of this category; "multidimensional acoustic features" refers to acoustic features in multiple aspects such as time domain, frequency domain, energy domain, and perception domain, which are feature representations fused from multiple channels and attributes.

[0136] In S4, structural difference refers to the numerical distance or distribution difference between the multidimensional acoustic feature set of the newly acquired signal and the existing behavioral prototypes (leaking, normal) in each feature dimension; classification refers to the classification of the new signal sample as either leaking or normal through feature comparison analysis; structural discrimination process refers to the entire process of comparing the features of the unlabeled signal with the features of the behavioral prototype, including feature input, distance calculation, and classification determination; key parameters of each prototype difference refer to the quantification of the difference between the new signal and the leaking / normal prototype in the core feature dimension, such as Euclidean distance, cosine distance, etc.

[0137] In S5, the direction of difference refers to the bias of the new sample in the feature space between it and the two types of prototypes, i.e., whether it is closer to the leaked or normal prototype. The structural deviation refers to the distance or deviation of the new sample from the leaked prototype and the normal prototype in the feature space, which is used as a basis for classification. The least-attribution principle means that when making a judgment, it is given priority to be assigned to the prototype category with the closest distance (the smallest feature deviation), which is the standard method for small-sample learning / prototype network discrimination.

[0138] like Figure 3 As shown, the specific steps for obtaining the pipeline sound pressure data sequence are as follows:

[0139] S101: Based on the main line of the heating pipeline, analyze the continuous signal output by the acoustic sensor continuously or periodically, construct frequency comparison conditions according to the boundary parameters of the operating frequency band, calculate the frequency distribution of the complete signal, and locate the frequency segment corresponding to the operating frequency band to obtain the frequency band aggregated acoustic segment set.

[0140] Based on the structural characteristics of the main heating pipeline, acoustic sensors are deployed on typical straight sections or structurally simple main pipeline sections. Raw continuous acoustic signals are collected at a sampling frequency of 20,000 times per second, with a sampling duration of 60 seconds to obtain high-density sampling data covering the entire time period. Then, frequency boundary ranges are set according to known equipment operating frequency parameters; for example, a lower frequency limit of 850 Hz and an upper frequency limit of 1300 Hz can be set as criteria for selecting target frequency bands. After reading the signal, all frequency components are extracted using frequency domain analysis methods, and each frequency point is judged to determine whether it falls within the target frequency range. If a frequency point is between the set upper and lower limits, then that frequency is... The time interval where the point is located is considered a qualified segment and temporarily stored. At the same time, all segments that meet the frequency determination conditions are arranged in chronological order to generate the most basic set of frequency band aggregated acoustic segments. The time length of each segment in this set does not exceed 100 milliseconds, and adjacent segments can overlap by 30 milliseconds to enhance the temporal continuity. If the frequency points extracted in a certain time period are concentrated in the range of 900 Hz, 1050 Hz, and 1280 Hz, the corresponding three time periods are all included in this set. In actual deployment, this operation can be performed for each sensor separately, and then the multi-channel segment sets are superimposed and integrated into a segment set with a unified time index for subsequent time series analysis tasks.

[0141] S102: Based on the frequency band aggregated acoustic segment set, analyze the continuous change pattern of the segments on the time axis, compare the sudden fluctuations of the envelope with the morphological differences of the continuous segments, identify segments with continuous characteristics, remove interfering segments, and recombine the retained segments to obtain a continuous and effective segment sequence.

[0142] After sorting by timestamp in ascending order, a 500-millisecond sliding analysis window is set, with each slide moving 100 milliseconds. Within each sliding window, the envelope change trend of all segments is extracted, and the rate of change of the envelope over time is recorded. Sudden increases or decreases are identified by calculating the increase in the rate of change between adjacent points. This rate is compared with the average rate of change of all segments. If the envelope rate of a segment is higher than 1.5 times the average rate, the segment is considered to have sudden characteristics; otherwise, if the envelope rate changes continuously and slightly, it is considered a continuous segment. Sudden segments and continuous segments are classified separately. After marking, based on the shortest time length screening criterion, the minimum duration of a segment is set to 200 milliseconds. If a segment is less than 200 milliseconds in length, it is considered invalid and is discarded. In analyzing a set of segments, if three consecutive segments appear with durations of 120 milliseconds, 180 milliseconds, and 400 milliseconds respectively, and the rate of change of the first segment is twice the average rate, the second segment is 30% of it, and the third segment is one-tenth of it, then only the third segment is retained. The first two segments are considered as sudden interference or discontinuous signals and are discarded. All retained consecutive segments are merged and connected according to time to form a new time segment sequence, which is used for the next stage of signal digitization and channel mapping processing.

[0143] S103: Based on the continuous effective segment sequence, adjust the analog signal structure of each sampling point, complete the analog-to-digital conversion, determine the time position of each sampling point, and integrate the corresponding channel identifier and digital signal to obtain the pipe sound pressure data sequence;

[0144] The retained continuous time segments are subjected to analog-to-digital conversion. After reading the analog voltage value of each sampling point, it is converted into a discrete digital signal value. The analog-to-digital converter is set to a 12-bit wide device, supporting a maximum voltage input of 3.3 volts. Each voltage value is rounded down by step and converted into a digital value, then encapsulated into a three-element information group with time position, channel number, and digital value. Each time position is inverted by the time index through the sampling frequency. For example, a frequency of 20,000 times per second corresponds to sampling once every 50 microseconds. The channel number is generated by mapping the sensor hardware address or deployment number. All converted information groups are sorted according to time sequence, numbered and classified by channel, and merged into a complete sound pressure data sequence set. For example, if a sampling point is time point 1001, channel number is channel 3, and the output after analog-to-digital conversion is 2000, then this point constitutes a three-element structure of [1001, 3, 2000]. All structure points constitute a complete time sequence data organized by channel.

[0145] like Figure 4 As shown, the specific steps for obtaining the multidimensional acoustic feature set are as follows:

[0146] S201: Based on the pipe sound pressure data sequence, analyze the continuous sound pressure amplitude and time label, determine the fluctuation trend of the signal amplitude curve in each time period, compare the amplitude trend of adjacent window segments, identify the signal interval with continuous amplitude change, optimize the amplitude change trajectory in the data frame, and obtain the set of dynamic amplitude change trajectories.

[0147] Based on the recorded timestamps of continuous sampling points and their corresponding sound pressure amplitude values, each sound pressure value and its corresponding timestamp are first read in the order of sampling time. Amplitude analysis is performed by setting every 500 sampling points as a window segment. The difference between the maximum and minimum values ​​of the amplitude values ​​within each window segment is calculated as the amplitude range of that segment. Simultaneously, the amplitude difference between the first and last points is recorded as the trend scalar for that segment. After completing the amplitude statistics for each window segment, the trend scalar differences between adjacent window segments are compared. If the difference is less than a set comparison threshold (e.g., a change of less than 10%), it is determined that the two window segments have a continuous amplitude trend; otherwise, if the difference is greater than the threshold, it is considered a trend break. Furthermore, for each of the two adjacent window segments, the sampling point with the largest sound pressure amplitude is determined, and the relative position of this maximum amplitude point within the window segment is obtained using the starting position of each window segment as a reference. By comparing the relative position difference of the maximum amplitude points in two adjacent window segments on the time axis, the similarity in their temporal distribution is determined. When this position difference is less than one-fifth of the sampling length of a single window segment, and the amplitude change directions of the two window segments are consistent, the adjacent window segments are considered to have a high degree of similarity in their temporal structure, thus confirming that they belong to the same continuous amplitude change interval. If the above conditions are not met, the position is determined to be a trend breakpoint. After marking multiple continuous amplitude change intervals in the entire data segment using this method, the sampling point trajectories in each interval are plotted as curves with time as the horizontal axis and amplitude as the vertical axis. For points with frequent spikes in the curve, five consecutive points in that segment are selected. The average amplitude value is used to replace the original value of the current peak point. This smoothing process optimizes the amplitude trend in each data frame. For example, if the sound pressure amplitude in a certain window segment drops steadily from a higher value to a lower value, and the sound pressure amplitude in the next adjacent window segment also shows the same downward trend, and the amplitude difference between the two window segments at the junction is small, when the amplitude difference at the junction accounts for only a small proportion of the overall amplitude change range of the previous window segment, is significantly lower than the preset continuity judgment threshold, and the change direction of the two window segments is consistent, it is determined that the adjacent window segments have continuity in amplitude change, and they are classified into the same continuous amplitude change interval, forming a set of dynamic amplitude change trajectories.

[0148] S202: Based on the dynamic change trajectory set of amplitude, calculate the amplitude fluctuation direction of each curve, determine the change characteristics of the start and end points, analyze the curve distribution of increasing, decreasing and stationary modes, identify the curve sequence that conforms to the amplitude change law, summarize the change mode category, and obtain the dynamic trend induction parameter set.

[0149] For each trajectory, its overall variation characteristics are analyzed sequentially. First, the amplitude changes between the starting and ending positions of the trajectory are compared: when the ending amplitude is significantly higher than the starting amplitude, the trajectory is marked as increasing; when the ending amplitude is significantly lower than the starting amplitude, the trajectory is marked as decreasing; when the difference between the starting and ending amplitudes is small, and the entire trajectory does not show a continuous upward or downward trend on the time axis, but fluctuates around a certain amplitude range, the trajectory is marked as a stable pattern. After determining the direction or pattern, the amplitude of the middle segment of the trajectory is further compared with the average amplitude of the beginning and end to analyze whether there are obvious mid-segment fluctuations. The distribution of local maximum and minimum values ​​within the trajectory is also used to determine its variation pattern: if there are many local extreme values ​​and they are scattered, it is considered a fluctuating change; if there are few local extreme values ​​and the direction of change is consistent, it is considered a monotonic change. Based on this, increasing, decreasing, and stationary trajectories are categorized and their proportion in all trajectories is statistically analyzed. If the proportion of a certain type in all trajectories exceeds a preset threshold, that type is marked as the main change pattern. For increasing and decreasing trajectories, the change intensity range is further divided according to the change span between their initial and final amplitudes, and the trajectories are classified in combination with the change direction and change intensity. At the same time, stationary trajectories are statistically analyzed as a separate change pattern, and their quantity and proportion in the total sample are recorded. Finally, the direction category, change pattern, and distribution statistics of various trajectories are integrated to generate a dynamic trend induction parameter set.

[0150] S203: Based on the dynamic trend induction parameter group, integrate the frequency distribution data of each time period, analyze the correlation between each parameter, adjust the arrangement order of each feature parameter in the data sequence, map and fuse acoustic parameters, and obtain a multi-dimensional acoustic feature set;

[0151] Frequency distribution data for each time period is read and mapped one-to-one with the dynamic trend parameters within that time period. First, a unified time axis is established based on time labels, and each frequency point in the frequency distribution matrix is ​​assigned to its corresponding amplitude trend interval. Then, the energy value within each frequency segment is correlated with its corresponding trend direction. If the number of times the energy value rises and amplitude increases within a frequency segment matches a set benchmark (e.g., 70%), then that frequency segment is considered to have a significant coupling with the increasing trend. Next, the distribution density of frequency segments is calculated across all trend intervals. If a frequency segment appears more than twice the median value in a certain trend category, it is designated as a key frequency segment for that trend. The coupling of all frequency segments is cross-compared, and the average number of correspondences between each frequency segment and different trend categories is recorded in the results. This index is used as a reference value for the degree of trend-frequency coupling. Subsequently, the arrangement order of feature parameters in the sequence is adjusted according to the priority of trend direction, change intensity, and frequency concentration to form a trend-dominant arrangement scheme. Frequency concentration is used to characterize the acoustic signal's frequency distribution within the trend. The determination of whether the main frequency components within a certain time period are concentrated within a limited frequency range is as follows: Analyze the frequency distribution results within each time period, identify the main frequency range with a high energy proportion, and determine whether the main frequency range is concentrated in a relatively continuous and narrow frequency segment. When the signal energy is mainly concentrated in a few adjacent frequency segments, and the concentrated segment appears continuously in multiple time periods, the time period is determined to have a high frequency concentration. When the signal energy is dispersed in multiple non-adjacent frequency segments, or the concentrated segment is unstable in time, its frequency concentration is determined to be low. During the arrangement of feature parameters, feature parameters with high frequency concentration and consistent with the trend direction and change intensity are preferentially retained to enhance the ability of the feature sequence to represent leakage behavior. After all trend parameters are aligned with frequency parameters, the sound intensity value, frequency amplitude, frequency energy, and sensing coefficient of each category are merged and spliced ​​according to the trend category as the index. The result is mapped into a unified multi-channel tensor structure, where each time frame contains composite information of four acoustic feature types, and a multi-dimensional acoustic feature set is output.

[0152] like Figure 5 As shown, the specific steps for obtaining the set of category prototype vectors are as follows:

[0153] S301: Based on a multi-dimensional acoustic feature set, analyze the sound intensity parameters and spectral parameters in each data frame, determine the distribution of leakage scenarios and normal scenarios in the feature space, compare the change patterns of feature combinations under the two scenarios, identify feature groups with discriminative capabilities, and obtain the scene feature expression structure.

[0154] The sound intensity and spectral parameters of each frame are read sequentially and grouped according to the data frame number and scene label. All leak-labeled and normal-labeled samples are classified and statistically analyzed. The sound intensity values ​​and spectral amplitude distributions of all frames are extracted for each scene category. Frames with intra-frame sound intensity variations exceeding 5 or spectral frequency shifts exceeding 300 Hz are defined as abnormal fluctuation frames. The frequency of these frames is first counted in both the leak and normal groups. Frames with a frequency exceeding 30% in both categories are considered indistinguishable and are removed. For the remaining frames, the co-current trend between sound intensity and spectral peak values ​​is calculated. If the two parameters show a positive correlation and the absolute value of the correlation coefficient exceeds 0.6, then the frame is considered to have a positive correlation. Feature combinations exhibit structural consistency. These combinations are grouped into two-dimensional feature sets according to their parameters. The mean differences between the corresponding feature sets in the leak and normal samples are compared group by group. If the mean difference exceeds 50% of the overall sample standard deviation, it is marked as a feature combination with discriminative power. For example, in the leak scenario, the average sound intensity of a certain feature set is 32, and the main peak of the spectrum is at 1020 Hz. In the normal scenario, the average sound intensity of the corresponding group is only 24, and the main peak of the spectrum is concentrated at 880 Hz. The difference reaches 8 and 140 Hz, which is significantly higher than the upper and lower limits of the standard deviation interval. This combination is recorded as a valid feature combination. All feature combinations that meet the discrimination threshold are arranged in the order of combination and summarized into a scene feature expression structure.

[0155] S302: Based on the scene feature expression structure, determine the dense distribution area of ​​each type of sample in the multi-dimensional space, analyze the spatial distance between sound intensity, frequency, power spectrum and voiceprint features, identify spatially similar parameters and merge the structures to obtain a set of clustering feature parameters;

[0156] All feature combination data corresponding to leakage and normal samples were extracted separately. Coordinate mapping was performed according to four parameter dimensions: sound intensity, frequency, power spectrum, and acoustic signature. Each frame's feature vector was represented as a four-dimensional spatial point. A spatial distribution map of each sample type was constructed sequentially, scanning the density distribution of points in four-dimensional space for each sample type. The coordinate axes were divided at 0.5 intervals. If the number of samples in a certain region exceeded 10% of the total number of samples of that type, it was identified as a locally dense area for that type. All points were sorted from high to low density values, and the top five dense areas were taken as representative areas. Pairwise comparisons were performed on the parameter dimension distribution within each dense area to analyze the relative changes of different acoustic parameters within that dense area. Specifically, this included comparing the sound intensity parameter and the frequency parameter under the same conditions... The numerical distribution differences within dense regions reflect the degree of proximity between the two in the feature space. For the sound intensity parameter and the power spectrum parameter, the difference between their corresponding values ​​within the same time period is calculated to characterize the synchronicity between sound intensity changes and energy distribution changes; this difference is the magnitude of the difference between sound intensity and power spectrum. For the frequency parameter and the voiceprint parameter, the fluctuation of voiceprint features during corresponding frequency changes is analyzed. By comparing the dispersion of voiceprint feature values ​​at different time points, the stability of its frequency variation is reflected; this dispersion is used to characterize the variance of frequency and voiceprint variation. Through the above pairwise comparison methods, parameter dimensions with similar trends and numerical distributions within the same dense region are identified and used as spatially similar feature parameter combinations.

[0157] S303: Based on the clustering feature parameter set, calculate the spatial center position of each group of feature parameters, optimize the feature classification method, adjust the arrangement of data frames and channel mapping order, summarize the parameter structure with aggregation, and obtain the category prototype vector set;

[0158] Read the parameter combination data within each cluster feature parameter set. For all samples in the same group, statistically analyze the sound intensity, frequency, power spectrum, and voiceprint features sequentially. Calculate the average value of all samples for each parameter dimension as the spatial center position of that dimension. Simultaneously, calculate the variance of samples of the same class in that dimension. If the variance is less than 20% of the average value of all samples in that class, the parameter is considered a stable dimension for that class. After marking the stability of all dimensions, prioritize stable parameters and place parameters with high variability later. Update and rearrange the data frame structure according to this order, adjusting the channel arrangement within the data frame. The columnar approach prioritizes stable parameter dimensions in the horizontal structure of the data frame, while the vertical structure remains unchanged according to the time series. The updated data frame is mapped to the new channel structure in order of parameter stability. At the same time, the parameter vectors of all samples in each class are merged by grouping by category label. If there are N samples in a certain class, the corresponding dimension values ​​of the N samples are added together and averaged to output a set of center vectors representing the class. A set of vectors is output for each class, with leaked and normal classes output separately. All center vectors are organized into a set of category prototype vectors for subsequent sample classification and similarity calculation operations.

[0159] like Figure 6 As shown, the specific steps for obtaining the feature distance metric parameters are as follows:

[0160] S401: Based on the set of category prototype vectors, analyze the spatial distribution of the multidimensional acoustic feature set of unlabeled data on each feature parameter, compare the variation law of acoustic features among each prototype, judge the structural differences between each feature parameter, and obtain a spatial feature comparison array.

[0161] The multidimensional acoustic feature set of the sample to be judged is read, and the spatial distribution differences of the two prototypes in each feature dimension are compared. First, each unlabeled sample is expanded into a channel tensor according to the frame structure, and its sound intensity, spectrum, power spectrum, and acoustic signature are extracted as vector representations. The difference between the sample vector value and the vectors of the two prototypes in the same dimension is calculated. The sign of the difference is determined and the absolute value is sorted for each dimension. At the same time, the baseline value range of the dimension in the leaky and normal prototypes is recorded. If the current sample parameter exceeds the upper and lower boundaries of the middle region of the two prototype values ​​by more than 20%, the dimension is marked as a high-discrepancy feature dimension. Then, all dimensions are compared one by one. During the comparison process, if a certain If a parameter shows an increasing trend in the leaked prototypes but a decreasing trend in the current sample, the direction of the parameter difference is marked as a negative deviation. If the trends are consistent, it is marked as a positive deviation. Then, the total number of difference directions and the average difference magnitude of each dimension are compared. If the number of a certain type of difference direction exceeds two-thirds of the total number of dimensions and its average difference magnitude ranks in the top 30% of all dimensions, then this type of difference is judged as a significant feature structure difference. The high-difference parameters are combined into a comparison sequence to form a structural record table representing the status of all parameter differences between the current sample and the two prototypes. This record table is expanded in the order of time frame number, and the results of all frames are integrated to form a spatial feature comparison array.

[0162] S402: Based on the spatial feature comparison array, calculate the spatial correspondence between unlabeled data and prototypes of each category in the feature dimension, identify the structural differences of feature parameters, determine the spatial similarity between each group of parameters and the category to which they belong, and obtain heterogeneous parameter indicator groups;

[0163] The values ​​of the currently unlabeled samples for each parameter are read dimension by dimension, and simultaneously compared with the standard values ​​of the same parameter in the leaked prototype and the normal prototype. Within the same dimension, it is determined whether the current sample value is close to any prototype parameter. If the difference is less than 80% of the amplitude difference of the prototype in that parameter dimension, the sample value is considered to have a high correspondence with the prototype space in that dimension. Based on this criterion, the attribution direction is identified for each dimension. Dimensions that are not similar to either type of prototype are marked as uncertain attribution dimensions. All uncertain dimensions are sorted in descending order of difference. Then, from the remaining high similarity dimensions, those with differences between the midpoint and the deviation direction of the two prototypes are selected. Consistent parameters are used as attribution directional parameters. If the proportion of such parameters exceeds half of all stable parameters, then the parameters in this group are assigned to the corresponding directional group of a certain prototype. This identification process is repeated in all time frames, and all results are summarized. According to indicators such as directional parameters, attribution trend direction, and difference magnitude, heterogeneous parameter indication groups are constructed for all unlabeled samples. For example, if the current sample sound intensity value is 28, the leakage prototype is 35, and the normal prototype is 22, and the current value is closer to the center value of the leakage class, then this dimension is assigned to the leakage directional type. If this judgment logic applies to multiple key parameters, then the parameters form the heterogeneous parameter indication group of the current frame.

[0164] S403: Based on the heterogeneous parameter indicator group, determine the category discrimination ability of each feature parameter, quantify the spatial distance performance of each parameter in prototype attribution discrimination, organize the distance data between unlabeled data and prototypes, optimize the arrangement order of feature parameters, and obtain feature distance measurement parameters.

[0165] Read all feature dimensions from the heterogeneous parameter indicator group and count the frequency of each feature dimension being classified as belonging to the same category in all unlabeled samples. Simultaneously, compare the value range of this feature dimension in the corresponding category prototype with the value of this dimension in the unlabeled samples to analyze the deviation of the unlabeled samples from the category prototype in this dimension. Summarize the deviations of all unlabeled samples to obtain the overall difference performance of this feature dimension within that category. Use this overall difference performance as the classification performance index of the feature dimension to reflect its effectiveness in distinguishing different categories. Based on this, rank and filter the discriminative power of each feature dimension. For each dimension, if its difference value is in the top 25% of the same type of parameter, it is considered a parameter with high discriminative power. After summarizing the discriminative power indicators of all dimensions, generate a parameter weight. The reordering table is used to number all dimensions from strongest to weakest discriminative power. Then, the channel arrangement order of the original data frame is read, and the parameter channels in the data frame are rearranged according to the number, so that the dimensions with high discriminative power are arranged at the beginning of the frame structure. The same parameter rearrangement operation is performed on all unlabeled samples. On this basis, the spatial distance between each frame sample vector and the two types of prototype vectors is calculated. For each parameter channel, the distance difference value between the leaked prototype and the normal prototype is recorded, and the average distance and maximum distance difference of all frames are calculated. If the distance difference of a certain parameter channel is consistently above the top third of all channels, it is set as a key dimension of feature distance and its difference index is marked. A distance data table of all parameters between each unlabeled sample and the prototype is constructed. This data table is arranged from smallest to largest distance index and summarized as the feature distance metric parameters corresponding to the sample.

[0166] like Figure 7 As shown, the specific steps for obtaining the pipeline leakage determination result are as follows:

[0167] S501: Based on the feature distance metric parameter, analyze the spatial relationship between the sample to be judged and the leakage behavior prototype and the normal behavior prototype, determine the distance direction between each prototype, compare the attribution trend of the sample features in space, map the spatial offset direction with the category judgment relationship, and obtain offset attribution judgment data.

[0168] Based on the spatial distance data of unlabeled samples and leakage and normal prototypes across various acoustic feature dimensions, the distance values ​​for each parameter dimension are first extracted and divided into two vector sets for leakage and normal prototypes respectively. Then, within each dimension, the direction of the distance difference between the current sample and the two prototypes is compared. If the sample's distance from the leakage prototype in the current dimension is less than its distance from the normal prototype, the parameter is classified as belonging to the leakage class; otherwise, it is classified as belonging to the normal class. After each dimension independently completes this classification, the direction labels for all dimensions are compiled into a direction voting sequence. Based on this, the frequency of occurrence of two directions in all parameters is counted, and it is determined which direction accounts for more than 50% of the total number of dimensions. For example, the leakage direction label count is 14 dimensions while the normal direction count is 10. The dimensions indicate that the overall offset direction tends to leak the prototype. Further, the differences of all dimensions with consistent directions are summed, and the differences of dimensions with opposite directions are subtracted. If the weighted difference is positive, the overall offset trend of the sample is attributed to the leak direction; otherwise, it is attributed to the normal direction. Simultaneously, the proportion of all dimensions with consistent directions among all dimensions is calculated as the confidence index of the attribution trend. If this proportion exceeds 0.6, it is marked as a sample with strong offset attribution; if it is below 0.4, it is marked as a sample with weak offset attribution. Based on the above comparison, the determination direction, attribution category, confidence value, and other information for each time frame are combined into a structured record table. Each row in this table corresponds to a sample frame, and each column records the directional trend, main attribution category, sum of differences, directional consistency ratio, and other determination criteria, constituting the offset attribution determination data.

[0169] S502: Based on the offset attribution determination data, calculate the degree of deviation between the sample features and each prototype in the feature space, compare the spatial trajectory with the category attribution conditions, filter the parameter paths that are consistent with the attribution relationship, and organize the spatial attribution data of each category to obtain the attribution path parameter set;

[0170] Based on the directional trend and category attribution of each time frame, the specific values ​​of acoustic feature parameters in each frame are read, and each parameter is compared within the parameter boundary range of the leaking prototype and the normal prototype. The spatial deviation of the sample feature vector at the center point of the two prototypes is calculated. After calculating the distance value in each dimension, it is uniformly normalized. The distance values ​​of all dimensions are sorted by prototype label to generate a two-dimensional deviation vector matrix. This matrix records the relative deviation direction and deviation magnitude of each dimension relative to each prototype. Then, all dimensional paths consistent with the attribution direction are marked by traversal. The dimensional paths are used as the set of parameter paths that match the attribution judgment. For dimensional paths with contradictory directions, an exclusion action is performed. After exclusion, the remaining parameters are retained. The frequency of occurrence of parameters in the remaining paths is calculated, and the top five most frequent path combinations are marked as the core path set. For example, if the sound intensity, the main frequency of the spectrum, and the third dimension of the voiceprint are all determined to be the leakage direction and have the same affiliation in most frames, then the paths of these three are combined into a high consistency affiliation path. The structure of all frame path combinations is statistically analyzed to form a ternary relation table with parameter name as index, affiliation category as label, and path combination frequency as value. This table is used to summarize which path combinations in the sample most stably point to the same category. After summarizing the affiliation path data of all frames, the affiliation matching times, the proportion of consistent directions, the maximum deviation difference, and other indicators of various parameter paths are sorted out, and the output constitutes the affiliation path parameter set.

[0171] S503: Based on the attribution path parameter set, determine the category label corresponding to the attribution direction, compare the mapping between the sample space determination path and the category label, organize the label results and recognition order of the sample determination, and obtain the pipeline leakage determination result.

[0172] The system reads all parameter path combinations and their corresponding category directions from the attribution path parameter set, sequentially determines the category label pointed to by each parameter path, and performs a vote count on all path pointing results. If the number of times a leak path is pointed to exceeds half of the total number of path combinations, the overall sample label is classified as leaking; otherwise, it is classified as normal. Subsequently, the path combination and its label mapping relationship for all time frames of the sample are compared. For each frame, the consistency between its path pointing category and label is recorded. If the consistency ratio is higher than 0.75, it is marked as a high-matching frame; otherwise, it is a low-matching frame. After sorting the path matching degree of all frames, a path matching table for each frame sample is constructed. Each record contains information such as path sequence, attribution label, voting direction, and consistency score. All records are then arranged in frame order to form a time-ordered judgment trajectory list. The overall sample label is determined based on the voting label ratio of all frames. For example, if 15 out of 20 frames are classified as leaking, the sample label output is leaking. At the same time, the 15 frames are calculated as 0.75, which is used as the confidence score of the label result. The label result is bound to the sample number and output to obtain the complete pipeline leak determination result.

[0173] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A few-sample learning method for pipeline leak identification by fusing multi-dimensional acoustic features, characterized in that, The method includes: S1: Based on the main pipeline of the heating network, analyze the acquired signals of the acoustic sensors, distinguish the target frequency band and noise interference, digitize the sampled signals through the analog-to-digital conversion unit, and arrange them in sequence according to the acquisition time to obtain the pipeline sound pressure data sequence; S2: Based on the pipeline sound pressure data sequence, a sliding window process is used to analyze the signal amplitude changes in each window, calculate the amplitude variation trend, summarize the energy change parameters, and combine the frequency characteristics to obtain a multi-dimensional acoustic feature set; S3: Based on the multidimensional acoustic feature set, filter leakage and normal structure features, input various types of samples into the recognition unit, adjust the fusion order, merge the features of samples of the same category, and obtain a set of category prototype vectors; S4: Based on the set of category prototype vectors, compare the structural differences between the unlabeled data and the prototypes, analyze the category classification through feature correspondence, quantify the differences of key features, and obtain the feature distance measurement parameters; S5: Based on the aforementioned feature distance measurement parameters, determine the direction of difference between the sample and the leaking and normal prototypes, read the structural deviation data, classify the sample leakage status according to the least attribution principle, and obtain the pipeline leakage determination result.

2. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The pipeline sound pressure data sequence includes time stamps, sound pressure amplitude, and channel number; the multidimensional acoustic feature set includes sound intensity features, spectral features, power spectrum features, and acoustic signature features; the category prototype vector set includes leakage category prototypes and normal category prototypes; the feature distance metric parameters include leakage category distance and normal category distance; and the pipeline leakage determination result includes the determination category and confidence score.

3. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The specific steps for obtaining the pipeline acoustic pressure data sequence are as follows: S101: Based on the main line of the heating pipeline network, analyze the output signal of the acoustic sensor, construct frequency comparison conditions according to the boundary parameters of the operating frequency band, calculate the frequency distribution of the complete signal, locate the frequency segment corresponding to the operating frequency band, and obtain the frequency band aggregated acoustic segment set. S102: Based on the frequency band aggregated acoustic segment set, analyze the continuous change pattern of the segments on the time axis, compare the sudden fluctuations of the envelope with the morphological differences of the continuous segments, identify segments with continuous characteristics, remove interfering segments, and recombine the remaining segments to obtain a continuous and effective segment sequence. S103: Based on the continuous effective segment sequence, adjust the analog signal structure of each sampling point, complete the analog-to-digital conversion, determine the time position of each sampling point, and integrate the corresponding channel identifier and digital signal to obtain the pipe sound pressure data sequence.

4. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The specific steps for obtaining the multidimensional acoustic feature set are as follows: S201: Based on the pipe sound pressure data sequence, analyze the continuous sound pressure amplitude and time label, determine the fluctuation trend of the signal amplitude curve in each time period, compare the amplitude trend of adjacent window segments, identify the signal interval with continuous amplitude change, optimize the amplitude change trajectory in the data frame, and obtain the amplitude dynamic change trajectory set. S202: Based on the aforementioned dynamic amplitude change trajectory set, calculate the amplitude fluctuation direction of each curve, determine the change characteristics of the start and end points, analyze the curve distribution of increasing, decreasing and stable modes, identify curve sequences that conform to the amplitude change law, summarize the change mode categories, and obtain the dynamic trend induction parameter set. S203: Based on the dynamic trend induction parameter set, integrate the frequency distribution data of each time period, analyze the correlation between each parameter, adjust the arrangement order of each feature parameter in the data sequence, map and fuse the acoustic parameters, and obtain a multidimensional acoustic feature set.

5. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on small sample learning according to claim 1, characterized in that, The specific steps for obtaining the set of category prototype vectors are as follows: S301: Based on the multidimensional acoustic feature set, analyze the sound intensity parameters and spectrum parameters in each data frame, determine the distribution of leakage scene and normal scene in the feature space, compare the change pattern of feature combination under the two types of scenes, identify feature groups with discriminative ability, and obtain the scene feature expression structure. S302: Based on the scene feature expression structure, determine the dense distribution area of ​​each type of sample in the multidimensional space, analyze the spatial distance between sound intensity, frequency, power spectrum and voiceprint features, identify spatially similar parameters and merge them to obtain a set of clustering feature parameters; S303: Based on the clustering feature parameter set, calculate the spatial center position of each group of feature parameters, optimize the feature classification method, adjust the arrangement of data frames and the channel mapping order, summarize the parameter structure with aggregation, and obtain the category prototype vector set.

6. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The specific steps for obtaining the feature distance metric parameters are as follows: S401: Based on the set of prototype vectors of the categories, analyze the spatial distribution of the multidimensional acoustic feature set of the unlabeled data on each feature parameter, compare the variation law of acoustic features among the prototypes, determine the structural differences between each feature parameter, and obtain a spatial feature comparison array. S402: Based on the spatial feature comparison array, calculate the spatial correspondence between unlabeled data and prototypes of each category in the feature dimension, identify the structural differences of feature parameters, determine the spatial similarity between each group of parameters and the category to which they belong, and obtain heterogeneous parameter indication groups; S403: Based on the heterogeneous parameter indication group, determine the category discrimination ability of each feature parameter, quantify the spatial distance performance of each parameter in prototype attribution discrimination, organize the distance data between unlabeled data and prototypes, optimize the arrangement order of feature parameters, and obtain feature distance measurement parameters.

7. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The specific steps for obtaining the pipeline leakage determination result are as follows: S501: Based on the feature distance measurement parameters, analyze the spatial relationship between the sample to be judged and the leakage behavior prototype and the normal behavior prototype, determine the distance direction between each prototype, compare the attribution trend of the sample features in space, map the spatial offset direction with the category judgment relationship, and obtain offset attribution judgment data. S502: Based on the offset attribution determination data, calculate the degree of deviation between the sample features and each prototype in the feature space, compare the spatial trajectory with the category attribution conditions, filter the parameter paths that are consistent with the attribution relationship, and organize the spatial attribution data of each category to obtain the attribution path parameter set. S503: Based on the set of attribution path parameters, determine the category label corresponding to the attribution direction, compare the mapping between the sample space determination path and the category label, organize the label results and recognition order of the sample determination, and obtain the pipeline leakage determination result.

8. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The acoustic sensor includes an external acoustic sensor; the external acoustic sensor refers to a sensor installed outside the heating network pipeline to collect the flow of the medium inside the pipeline, leakage or environmental noise, and the analog-to-digital conversion unit refers to the circuit part that converts the analog signals collected by the sensor into digital signals.

9. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The sliding window processing refers to the process of using a fixed-length time window for long-term acoustic signals, gradually sliding the window, and independently analyzing the features in each window. The signal amplitude change refers to the amplitude of the acoustic signal at each sampling point, reflecting the intensity of the sound wave.

10. The multi-dimensional acoustic feature fusion method for pipeline leak identification based on few-sample learning according to claim 1, characterized in that, The leakage state and normal structure refer to the acoustic signal feature categories corresponding to the pipeline being in a leakage condition or a normal condition, respectively. The structural difference refers to the numerical distance or distribution difference between the multidimensional acoustic feature set of the newly acquired signal and the existing behavioral prototype in each feature dimension.

Images