A method and system for acoustic imaging localization and leakage parameter inversion of hydrogen leaks

By combining a hierarchical heterogeneous microphone array with a multipath propagation model, the problems of leakage parameter inversion and multi-physics field coupling decision-making in hydrogen leak detection were solved, realizing accurate location and parameter inversion of hydrogen leaks, and improving the intelligence and reliability of detection.

CN123111584APending Publication Date: 2026-11-17TONGJI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611594964.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing hydrogen leak detection technologies struggle to accurately determine the leakage amount and orifice size in complex environments, and lack multi-physics field coupling decision-making, resulting in insufficient reliability of detection results.

Method used

A hierarchical, heterogeneous MEMS microphone array consisting of high-frequency, mid-frequency, low-frequency components and a central reference unit is employed. Combined with frequency band discrimination priors and acoustic fingerprint features, acoustic imaging of hydrogen leaks is performed using beamforming and multipath propagation models. The leak aperture and pressure are inverted using gradient boosting regression trees, and multi-physics coupling consistency checks are conducted.

Benefits of technology

It enables precise location of hydrogen leak points and accurate inversion of leak parameters in complex environments, improving the intelligence level of hydrogen system leak detection and enhancing the credibility and reliability of detection results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123111584A_ABST
    Figure CN123111584A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for acoustic imaging localization and leakage parameter inversion of hydrogen leaks. It utilizes a layered heterogeneous MEMS microphone array composed of a central reference unit, a high-frequency detection layer, a mid-frequency detection layer, and a low-frequency detection layer to acquire acoustic signals, while simultaneously acquiring pipeline pressure, valve status, equipment operating status, and temperature and humidity information. Leak acoustic features are extracted through frequency band discrimination priors and acoustic fingerprint feature modeling. The three-dimensional space is discretized into multi-depth two-dimensional scanning surfaces, and candidate leak locations are obtained by combining beamforming, direct path modeling, reflection path modeling, and cross-plane fusion. After propagation and environmental compensation of the acoustic fingerprint, the leak aperture and leakage pressure are obtained using an XGBoost regression model. Event levels are generated and updated by fusing acoustic, spatial, parameter, and operational condition evidence.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of hydrogen energy safety detection technology, and more specifically, to a method and system for acoustic imaging localization and leakage parameter inversion of hydrogen leaks. Background Technology

[0002] With the rapid development of the hydrogen energy industry, the rapid and accurate detection of hydrogen leaks has become a crucial requirement for ensuring the safe operation of hydrogen energy systems. Hydrogen is characterized by rapid diffusion, low ignition energy, and fast combustion; if leaks are not identified and addressed promptly, they can easily lead to safety risks. Therefore, developing hydrogen leak detection and location technologies suitable for industrial environments is of great significance.

[0003] Existing hydrogen leak detection technologies mainly include gas concentration detection, pressure change detection, optical detection, and acoustic detection. Among these, acoustic detection has attracted considerable attention in recent years due to its advantages such as non-contact operation, fast response, and long-distance sensing capability. However, existing acoustic detection technologies have the following shortcomings:

[0004] Insufficient Leakage Calculation and Orifice Diameter Inversion: Current acoustic leak detection mainly focuses on "whether there is a leak," with only a few patents performing neural network-based inversion of orifice diameter and leakage amount. However, these studies offer limited explanation of the relationship between features and results. The limited research on leakage parameter inversion may be due to the fact that different combinations of physical parameters can project similar acoustic characteristics. For example, "small orifice high pressure" and "large orifice low pressure" may exhibit similar acoustic intensities. Furthermore, high-frequency sound waves are relatively fragile and easily affected by environmental factors (such as humidity), making them unsuitable for leakage parameter inversion.

[0005] Insufficient multiphysics coupling decision-making: Although current research has begun to focus on the application of multi-sensor fusion in leak diagnosis, it is still mainly used to identify whether a leak has occurred or to estimate the location of the leak. There is insufficient research on the reliability judgment of the detection results, the consistency test of the multiphysics, and the feedback correction mechanism of the results.

[0006] Patent CN106908755A discloses a method for estimating the sound source location of compressed gas leaks using a wireless acoustic sensor network. This method employs a multi-node uniform linear array microphone for sampling and reconstructs the sound source's azimuth spectrum using compressed sensing theory to obtain the sound source location. However, this method does not specifically consider the inversion of leak parameters, making it difficult to accurately assess the severity of the leak. Furthermore, it does not incorporate multi-physics information to jointly determine the reliability of the detection results. Additionally, its sound source location estimation model is primarily based on the assumption of a free field and unobstructed direct sound waves. In practical applications, this may lead to misjudgments by mistaking virtual hotspots reflected from obstacles as leak points. Moreover, this method uses a uniform array model and fails to consider the wideband, multi-scale acoustic characteristics that may arise under different operating conditions, making it difficult to handle complex situations.

[0007] Patent CN121163776A discloses a complete detection method, device, and system. The method involves acquiring signals using a microphone array, synthesizing a target sound beam using a beamforming algorithm, obtaining the energy spectrum of the beamformed signal, extracting the main lobe height and threshold height for comparison, and finally outputting a binary "yes / no" leakage judgment. However, this method lacks parameter inversion or a multi-physics coupling decision scheme, and also lacks virtual image removal and targeted microphone placement solutions. Therefore, existing technology still lacks a hydrogen leakage acoustic imaging and quantitative assessment system that can adapt to complex environments, balance broadband acquisition and spatial positioning, and perform leakage parameter inversion and multi-physics coupling decision-making. Summary of the Invention

[0008] The purpose of this invention is to address the problems mentioned in the background section by providing a method and system for acoustic imaging localization and leak parameter inversion of hydrogen leaks. This method enables the localization of leak points and accurate inversion of leak aperture and pressure within a closed-loop adaptive framework. Simultaneously, based on acoustic acquisition results and multi-physics information, the final results are evaluated for reliability and feedback is provided in a closed loop, significantly improving the intelligence level of hydrogen system leak detection and providing a foundation for subsequent consequence analysis. This overcomes the shortcomings of existing acoustic leak detection devices, such as the lack of detailed leak information inversion and the lack of multi-physics coupling to determine the rationality of detection results. Furthermore, the invention utilizes a beam-based multipath propagation model to eliminate virtual images and simulate real leak point information even when multiple obstacles interfere with the acoustic wave path.

[0009] The present invention adopts the following technical solution:

[0010] This invention provides a method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks, including:

[0011] Step S1: Acoustic signals are collected using a layered heterogeneous microelectromechanical system microphone array consisting of a central reference unit, a high-frequency detection layer, a mid-frequency detection layer, and a low-frequency detection layer. At the same time, pipeline pressure, valve status, equipment operating status, and ambient temperature and humidity information are collected. The collected signals are then conditioned at the front end, synchronized in time, and transmitted after being unified to the same time axis.

[0012] Step S2 involves preprocessing and normalizing the original frequency domain features, original time domain features, original features at each layer, and original spatial features in the offline database. The offline database includes a background database, an interference database, and a leakage database with operating condition labels. The target frequency band is divided into bands, and the band energy, inter-layer energy ratio, and coefficient of variation are calculated. Inter-class separation analysis, interference exclusion analysis, operating condition robustness analysis, and random forest importance analysis are then performed to obtain the frequency band discriminative prior and the screening results of acoustic fingerprint features. Based on the screening results, the acoustic fingerprint feature vector of the current leakage event is extracted from the acoustic signal collected in step S1.

[0013] Step S3: Discretize the continuous three-dimensional space into multiple two-dimensional scanning surfaces. Perform beamforming based on array geometric parameters, sound velocity parameters, the prior frequency band discrimination capability, and multi-channel acoustic signals to extract candidate hotspots. For each candidate hotspot, establish a direct path model with the candidate hotspot as the actual leakage point and a reflection path model with the candidate hotspot as the reflection virtual image and the mirror source method to obtain the mirror point. Compare the overall index of the two models to obtain the virtual image probability of the candidate hotspot. Based on the virtual image probability, retain, downgrade, or eliminate candidate hotspots. Perform cross-plane fusion on the candidate hotspots retained in different two-dimensional scanning surfaces to obtain the leakage candidate location, propagation distance, candidate depth layer, spatial confidence, and virtual image probability.

[0014] Step S4: The acoustic fingerprint feature vector, leakage candidate location, propagation distance, candidate depth layer, spatial confidence, virtual image probability and environmental temperature and humidity information are used for parameter inversion. After propagation and environmental compensation, the leakage aperture inversion model and leakage pressure inversion model are called respectively to obtain the leakage aperture estimate, leakage pressure estimate and corresponding confidence.

[0015] Step S5: Summarize acoustic evidence, spatial evidence, parametric evidence, and operational condition evidence, and sequentially perform single evidence validity checks and multi-physics coupling consistency checks. Output the leakage event level, and update the offline database based on the judgment results, or adjust the credibility weights of the corresponding inputs in subsequent parameter inversion.

[0016] Furthermore, the central reference unit is a central reference microphone located at the center of the array; the high-frequency detection layer is formed by four high-frequency microelectromechanical system (MEMS) microphones arranged in a compact square; the intermediate-frequency detection layer is formed by eight intermediate-frequency MEMS microphones arranged in an equidistant concentric circle array; and the low-frequency detection layer is formed by twelve low-frequency MEMS microphones arranged in an equidistant concentric circle array.

[0017] Furthermore, the front-end conditioning in step S1 includes sequential bandpass filtering, signal amplification, and anti-aliasing filtering, and optional gain control is set during signal amplification; each acoustic channel is sampled simultaneously using a unified clock source, or quasi-simultaneous sampling is used with fixed time delay compensation.

[0018] Furthermore, the target frequency band in step S2 is divided into four sub-frequency bands: 1kHz to 5kHz, 5kHz to 15kHz, 15kHz to 40kHz, and 40kHz to 80kHz. The frequency band discrimination prior is obtained by fusing and normalizing the leakage-background interlayer separation degree, the leakage-interference interlayer separation degree, the interference exclusivity, the operating condition robustness characterized by the fluctuation of the interlayer separation degree under different operating conditions, and the importance of random forest.

[0019] Furthermore, the acoustic fingerprint features are obtained by sorting the candidate feature set according to the discriminative power index and taking the top six, and then successively undergoing z-score standardization, reference benchmark normalization, background bias compensation under the same working conditions, and fingerprint weighting according to the discriminative power score.

[0020] Furthermore, in step S3, the intermediate frequency detection layer and the low frequency detection layer are first used for coarse localization, and then the high frequency detection layer is used for fine correction near the candidate hot spot.

[0021] Furthermore, the direct path model constructs a time delay consistency score based on the error between the theoretical relative time delay and the actual relative time delay obtained from the generalized cross-correlation; the reflection path model obtains a geometric score based on the geometric rationality of the path and the reliability of the reflection surface; the weighted fusion of each index yields the total index of the two models, and the comparison of the two total indices is transformed into the virtual image probability.

[0022] Furthermore, in step S4, the propagation and environmental compensation obtain the air absorption coefficient according to the air absorption model in ISO 9613-1, express the air absorption loss using the exponential decay model, and calculate the distance diffusion loss using the spherical diffusion approximation; the leakage aperture inversion model and the leakage pressure inversion model are both gradient boosting regression tree models with squared error as the loss function.

[0023] Furthermore, the multiphysics coupling consistency check in step S5 includes acoustic-operating condition coupling, acoustic-spatial feature coupling, and acoustic-interference coupling, and outputs suppression, low confidence warning, medium confidence warning, or high confidence warning.

[0024] The present invention also provides an acoustic imaging localization and leakage parameter inversion system for hydrogen leaks, including a data acquisition system, a feature modeling system, a leak localization system, a parameter inversion system, a data transmission and control system, and a comprehensive judgment and feedback system; the data transmission and control system is connected to the data acquisition system, the feature modeling system, the leak localization system, the parameter inversion system, and the comprehensive judgment and feedback system respectively.

[0025] Beneficial effects

[0026] This invention employs a layered heterogeneous MEMS microphone array consisting of high-frequency, mid-frequency, low-frequency, and central reference units, combined with frequency band discrimination priors and acoustic fingerprint features, enabling the frequency band sensitivity of different detection layers to be used together for detection and localization. Multi-scan beamforming, direct path models, and reflection path models are used to interpret candidate hotspots, reducing the impact of false hotspots caused by obstacle obstruction or reflection. Gradient boosting regression trees are used to invert leakage aperture and leakage pressure, suitable for processing structured acoustic fingerprint features. Rule-based gating and feedback updates to the database or trusted weights are performed using acoustic evidence, spatial evidence, parametric evidence, and operational condition evidence, forming a closed-loop processing chain from acquisition, localization, inversion to judgment feedback. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall structure of a hydrogen leak acoustic imaging localization and leak parameter inversion system according to an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the feature modeling system structure of the present invention.

[0029] Figure 3 This is a schematic diagram of the prior construction logic for the frequency band discrimination capability of the present invention.

[0030] Figure 4 This is a schematic diagram illustrating the construction logic of the acoustic fingerprint database in this invention.

[0031] Figure 5 This is a schematic diagram of the overall structure of the leak location system of the present invention.

[0032] Figure 6 This is a schematic diagram of the direct path model for leak location in this invention.

[0033] Figure 7 This is a schematic diagram of the leakage location reflection path model of the present invention.

[0034] Figure 8 This is a schematic diagram of the leakage parameter inversion system of the present invention.

[0035] Figure 9 This is a schematic diagram of the integrated judgment and feedback system structure of the present invention.

[0036] Figure 10 This is a plan view of the layered heterogeneous microphone array arrangement of the present invention.

[0037] Figure 11 This is a side view of the layered heterogeneous microphone array arrangement of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0039] As shown in the figure, an embodiment of the present invention discloses a method and system for acoustic imaging localization and leakage parameter inversion of hydrogen leakage, using a high-pressure hydrogen pipeline system as the implementation object. The system includes a main pipeline, a high-pressure clamp connection structure, a shut-off valve, a support column, and a nearby operating fan. The surface of the support column reflects mid-to-high frequency sound waves. The overall process of the method, its subordinate technical features, and the system composition are described below.

[0040] One embodiment of the present invention performs a leakage event processing cycle according to steps S1 to S5. The inputs of the processing cycle include multi-channel acoustic signals, pipeline pressure, valve status, equipment operating status, ambient temperature and relative humidity; intermediate data include unified time axis data frames, frequency band discrimination priors, acoustic fingerprint feature vectors, spatial focusing thermal maps of each scanning plane, candidate hot spots and their direct and reflection path evaluation results; the outputs include leakage candidate locations, propagation distances, candidate depth layers, spatial confidence, virtual image probability, leakage aperture estimates, leakage pressure estimates, corresponding confidence levels and leakage event levels.

[0041] In step S1, the hierarchical heterogeneous MEMS microphone array is positioned facing the pipeline under test, such as... Figure 10 and Figure 11 The central reference unit (red) serves as the origin. Four microphones in the high-frequency layer are arranged in a square array, with the coordinates of one microphone already given. Eight microphones in the mid-frequency layer are arranged in equidistant concentric circles, with adjacent microphones at a 45° angle. Twelve microphones in the low-frequency layer are arranged in equidistant concentric circles, with adjacent microphones at a 30° angle. Four evenly spaced concentric mounting holes are attached to the edge. The mounting substrate thickness is 4mm. The microphone dimensions are simplified here to cylinders 4mm high and 3.5mm in diameter; these dimensions can be adjusted as needed in actual design. This establishes a correspondence between the array's geometric coordinate system and the measured spatial coordinate system. The central reference unit, high-frequency detection layer, mid-frequency detection layer, and low-frequency detection layer simultaneously acquire acoustic signals, while pressure, valve, equipment operating status, and temperature and humidity sensors synchronously acquire operating condition information. Each acoustic channel undergoes bandpass filtering, amplification, and anti-aliasing filtering before analog-to-digital conversion. The time synchronization module adds timestamps under the same time base to the acoustic and operating condition data, and the edge processor encapsulates these into unified data frames according to event time windows before transmitting them to the host computer.

[0042] In step S2, background, interference, and leakage libraries are pre-established in the offline phase. The background library records background sounds under different operating conditions and environmental conditions; the interference library records interference sounds such as pipe impacts, valve actions, sudden changes in fan background, or equipment startup; and the leakage library records intrinsic leakage samples and leakage samples under actual operating conditions, with operating condition labels assigned to each sample. Frequency domain, time domain, each detection layer, and spatial features are extracted from the offline samples. After banding, normalization, and secondary feature calculation, a frequency band discriminative prior is formed; then, acoustic fingerprint features are selected from the candidate features. In the online phase, the data frames from step S1 are processed according to the same preprocessing and feature definition, and the acoustic fingerprint feature vector of the current event is output to ensure consistency in feature caliber between offline training and online detection.

[0043] In step S3, multiple two-dimensional scanning surfaces are first established based on the depth range of the measured area, the location of easily leaking components, and structures that can serve as reflective surfaces. Then, the array geometric coordinates, sound velocity, frequency band discrimination priors, and multi-channel acoustic signals are input into the beamforming module to generate spatial focusing thermal maps on each scanning surface and extract candidate hotspots. For each candidate hotspot, on the one hand, it is assumed to be a real leak point and a direct path model is established; on the other hand, it is assumed to be a reflected virtual image and a mirror sound source model is established using the candidate reflective surface. The overall index of the two models is compared to obtain the virtual image probability. After retaining, downgrading, or eliminating candidate hotspots according to their virtual image probability, the retained hotspots with a spatial distance less than a preset distance threshold within different scanning surfaces are grouped into candidate hotspot clusters. Cross-plane scoring is performed based on the number of supporting planes, positional dispersion, and virtual image probability within the cluster, outputting the optimal candidate position, depth layer, and propagation distance.

[0044] In step S4, based on the acoustic fingerprint feature vector of the current event, the propagation distance, candidate depth layer, spatial confidence, virtual image probability, and temperature and humidity information are incorporated into the inversion input. First, the features related to sound intensity are compensated using air absorption and spherical diffusion approximations. Then, the aperture inversion model and pressure inversion model, which were trained offline, are called respectively. The result evaluation module checks whether the estimated values ​​are within the coverage range of the training samples and the allowable operating conditions of the equipment, and checks whether the estimation results are consistent with the acoustic fingerprint matching degree, spatial confidence, and virtual image probability. It then outputs the aperture, pressure, and their respective confidence levels.

[0045] In step S5, the integrated judgment and feedback system uses acoustic fingerprint matching degree, sensitive frequency band deviation degree, and event duration as acoustic evidence; candidate location, hot spot concentration, main-to-sub-peak ratio, spatial credibility, and virtual image probability as spatial evidence; aperture and pressure estimates and their credibility as parametric evidence; and pipeline pressure changes, valve status, equipment operating status, and temperature and humidity as operating condition evidence. The system first checks the validity of each piece of evidence, then checks whether different physical field evidences support each other, and finally outputs suppression, low credibility warning, medium credibility warning, or high credibility warning. Events confirmed as interference can be added to the interference library, and inversion inputs that are inconsistent with operating conditions for a long time can have their credibility weight reduced in subsequent training or online estimation.

[0046] A layered heterogeneous MEMS microphone array is mounted on a rigid mounting substrate. The substrate is equipped with a center positioning reference, array orientation marks, outer mounting holes, and position reference lines for each detection layer. A central reference microphone is located at the center of the array; four high-frequency MEMS microphones are arranged in a compact square around the central reference microphone to form a high-frequency detection layer; eight mid-frequency MEMS microphones are arranged equidistantly along concentric circles outside the high-frequency detection layer to form a mid-frequency detection layer; and twelve low-frequency MEMS microphones are arranged equidistantly along concentric circles outside the mid-frequency detection layer to form a low-frequency detection layer.

[0047] The high-frequency detection layer collects sensitive information in the 20kHz to 40kHz and higher frequency bands. After obtaining coarse localization results from the mid-frequency and low-frequency detection layers, it performs local fine correction on the neighborhood of candidate hotspots. The mid-frequency detection layer is used for main space scanning and mid-frequency leakage sound wave extraction. The low-frequency detection layer is used for stable detection and auxiliary spatial constraints under complex backgrounds or partial occlusion conditions. The central reference microphone serves as a unified benchmark between different samples, channels, and detection layers, used for channel normalization and inter-layer relative energy ratio calculation. The coordinates of each microphone are written into the array geometry parameter table during system initialization for use in calculations of theoretical propagation distance, theoretical time delay, and beamforming.

[0048] Each microphone output first enters a bandpass filter unit to suppress low-frequency mechanical vibrations and excessively high-frequency noise outside the target detection frequency band; the filtered weak signal then enters an amplification unit. The amplification unit features selectable gain control, using different sampling ranges for input signals of different magnitudes, and records the channel number, range, and gain status along with the sampled data, enabling subsequent feature calculations to restore or normalize the amplitude based on the gain status. The amplified signal is then input to the analog-to-digital converter module after passing through an anti-aliasing filter. In this embodiment, an elliptic filter can be used for the bandpass filter, and a Butterworth filter can be used for the anti-aliasing filter.

[0049] Multiple acoustic channels are preferably triggered by the same clock source for simultaneous sampling. When the hardware can only perform quasi-simultaneous sampling, a fixed time delay of each channel relative to the central reference channel is pre-calibrated, and time delay compensation is performed in the data transmission and control system. Pressure, valve, equipment operating status, and temperature and humidity information use the same time base with added timestamps. The edge processor aligns the acoustic data and operating condition data to the same event time window according to the timestamps, and encapsulates the acoustic data, operating condition data, channel identifiers, gain status, timestamps, and check fields into a unified data frame.

[0050] Offline database acquisition should utilize the same microphone equipment and array setup as online detection. For each sample with a working condition label, channel gain normalization and coarse compensation are first performed. Then, the target frequency band is divided into four sub-bands: 1kHz to 5kHz, 5kHz to 15kHz, 15kHz to 40kHz, and 40kHz to 80kHz. For the k-th sample within the i-th sub-band, the energy of each detection layer is calculated, and the inter-layer energy ratio is calculated based on the energy of the central reference unit. Taking the high-frequency detection layer as an example, the inter-layer energy ratio is:

[0051]

[0052] In the formula, This represents the high-frequency detection layer energy of the k-th sample in the i-th sub-band. This represents the energy of the corresponding center reference cell, with ε being a positive number set to prevent the denominator from being too small. The energy ratios of the intermediate frequency detection layer and the low frequency detection layer relative to the center reference cell can be obtained in the same way.

[0053] To characterize the short-time stability of sub-bands, each sample is divided into multiple consecutive short time windows, and the frame energy of the t-th short time window is extracted. The coefficient of variation and the mean frame energy are calculated according to the following formula:

[0054]

[0055]

[0056] In the formula, and These represent the standard deviation and mean of the frame energy, respectively. A larger coefficient of variation indicates poorer stability of the sub-band within a continuous short time window; conversely, a smaller coefficient indicates better stability.

[0057] Using frequency band energy, interlayer energy ratio, and coefficient of variation as secondary features, the inter-class separation degree between leaked samples and background samples, and between leaked samples and interfering samples, was calculated. The degree of overlap between the distribution intervals of leaked samples and interfering samples was statistically analyzed to form the interference exclusion result. The samples were then divided into several operating condition groups according to the operating condition label, and the separation degree was repeatedly calculated within each operating condition group. The fluctuation of the separation degree between different operating condition groups was used to characterize the operating condition robustness.

[0058] Simultaneously, the frequency band energy, inter-layer energy ratio, and short-time stability of the frequency band are input into the random forest to obtain the importance of the three types of frequency band features in each detection layer. , and Its importance index is calculated according to the following formula:

[0059]

[0060] The leakage-background interlayer separation, leakage-interference interlayer separation, interference exclusivity, operational robustness, and random forest importance are fused, and the results for each sub-band are normalized to obtain the prior discriminative weights for each sub-band. During online beamforming, the focusing values ​​of each sub-band participate in the synthesis according to their corresponding prior discriminative weights.

[0061] To avoid inconsistencies in the dimensions of the evaluation indicators, each indicator is normalized before fusion, with min-max normalization being the preferred method.

[0062]

[0063] Where, x i For the index to be normalized, x min With x max These are the minimum and maximum values ​​of the index in all candidate frequency bands, respectively, and ε is a positive number to prevent the denominator from being 0.

[0064] For the i-th sub-band, let the normalized leakage-background separation be denoted as . The leakage-interference separation degree is Interference Exclusivity Random forest importance is Operating condition robustness is The overall discrimination capability of a frequency band can be expressed as:

[0065]

[0066] The sum of all weights is 1. In practice, each weight can be optimized and adjusted by the performance of the validation set.

[0067] The final prior weights for frequency band discrimination are:

[0068]

[0069] Where N f N represents the number of sub-bands in this invention. f 4

[0070] The candidate feature set includes at least the spectral peak feature, energy features of the four sub-bands, energy ratio features of the high-frequency / mid-frequency / low-frequency detection layers relative to the central reference cell, spectral shape features, temporal features, and spatial consistency features. For each candidate feature, leakage-background separation and leakage-interference separation are calculated to obtain a discriminant score, which is then combined with the operating condition robustness coefficients obtained based on different operating condition groups to form a discriminant index. All candidate features are sorted in descending order of discriminant index, and the top six features are selected as acoustic fingerprint features. To shorten the feature-by-feature evaluation time, XGBoost, LightGBM, or random forest can also be used to learn feature importance, but the final unified acoustic fingerprint vector is still composed of the selected six features.

[0071] z-score standardization is performed on the j-th selected feature of the k-th sample:

[0072]

[0073] In the formula, For the corresponding sample feature values, and These are the mean and standard deviation of the feature in the training library, respectively, with ε being a positive number set to prevent the denominator from being too small. For features related to the detection layer energy, reference normalization is then performed using the energy of the central reference unit:

[0074]

[0075] In the formula, and These represent the high-frequency detection layer energy and the central reference cell energy of the k-th sample in the i-th sub-frequency band, respectively. The same reference normalization method is applied to both the mid-frequency and low-frequency detection layers. Then, the background bias under the same operating condition is subtracted based on the sample's operating condition label.

[0076]

[0077] In the formula, The average value of the samples in the background library that are identical to the current operating condition in the i-th sub-frequency band is used. Finally, fingerprint weights are configured according to the discriminative scores of each selected feature, and the six standardized, reference-normalized, and background-compensated features are arranged sequentially to form the acoustic fingerprint feature vector of the current event, which is then transmitted to the parameter inversion system and the comprehensive judgment and feedback system.

[0078] For example, in a preferred practical example, after comprehensively evaluating the parameters of the candidate features, the six acoustic fingerprints obtained are arranged in descending order of their discriminative power score feature weights, namely: 15-40kHz band energy (discriminative power score 0.92, normalized weight 0.194), 40-80kHz band energy (discriminative power score 0.88, normalized weight 0.186), 15-40kHz high-frequency layer / center reference layer energy ratio (discriminative power score 0.82, normalized weight 0.173), 40-80kHz high-frequency layer / center reference layer energy ratio (discriminative power score 0.78, normalized weight 0.165), spectral center (discriminative power score 0.70, normalized weight 0.148), and 1540kHz band short-time variation coefficient (discriminative power score 0.64, normalized weight 0.135).

[0079] Several parallel depth layers are set according to the depth range of the space to be detected in front of the array, and local component section layers are set according to the surfaces of easily leaking components such as valves, ferrule connections, and pipe joints. The k-th two-dimensional scanning surface is represented as:

[0080]

[0081] In the formula, Let k be the reference point for the k-th two-dimensional scan surface. and Let be the in-plane basis vectors of this two-dimensional scanning surface. and These are local coordinates in the plane. For , By discretizing the grid according to the preset grid spacing, all the points to be scanned on the scanning surface can be obtained, and each point to be scanned can be transformed into a unified three-dimensional coordinate system.

[0082] At each scanning point, the relative time delay is calculated based on the propagation distance and sound speed from that point to each microphone, and the multi-channel signal is then focused with a time delay. After obtaining the focusing values ​​for each sub-band, a weighted sum is performed according to the prior frequency band discrimination to generate a spatial focusing heatmap. Preferably, a coarse scan of the entire scanning surface is first performed using a mid-frequency detection layer and a low-frequency detection layer to obtain the initial peak position; then, a local scanning area is defined with the initial peak as the center, and a higher-frequency detection layer is used for denser local scanning and position correction.

[0083] In the spatially focused heatmap, the original peak value is extracted using local neighborhood search. Weak peaks are suppressed using a background statistical threshold formed by background data under the same operating conditions. Repeating peaks with a spatial distance smaller than the suppression radius of neighboring peaks are merged or only the stronger ones are retained. For the retained candidate hotspots, the peak size, the area exceeding the hotspot boundary threshold, the ratio of the main peak to the secondary peak, the hotspot concentration reflecting the degree of energy accumulation near the peak, and the corresponding three-dimensional coordinates are recorded as inputs for subsequent direct-reflection path interpretation and cross-plane fusion.

[0084] In one embodiment of the present invention, the scanning plane spacing The determination should be based on three aspects. First, the scanning plane spacing should not exceed the actual axial resolution of beamforming. Specifically, it should be:

[0085]

[0086] in This represents the typical depth resolution scale of the system under the current frequency band and array aperture. This method of value selection ensures that at least two layers cover a resolvable scale; secondly, given the device resolution, The depth positioning error should be less than the user's expected depth. Furthermore, the interplanar spacing can be adjusted appropriately based on two different scanning strategies: coarse scanning of the entire space and fine scanning near hotspots. The number of scanning planes is determined by the depth D of the space to be measured and the interlayer spacing. Sure.

[0087] Local component cut surfaces are preferably located at: valve stems, valve covers and connecting parts, ferrule fittings, flanges and sealing surfaces, welds, pipe joints, etc.

[0088] Coarse scan preferred When performing local fine scanning near candidate hotspots, the preferred method is... During implementation, the settings should be adaptively adjusted based on the array aperture, acoustic wavelength, and target positioning accuracy. The average focus value of the same scanning point or frequency band in a background library under the same working conditions should be statistically analyzed. and standard deviation The background statistics threshold is defined as:

[0089]

[0090] In the candidate hot spot extraction stage, neighboring peak suppression is applied to repeating peaks that are spatially close: the neighboring peak suppression radius is optimally selected during the coarse localization stage. In the local high-frequency precision positioning stage, the preferred method is to select... When the spatial distance between two candidate peaks is less than the corresponding threshold, only the peak with the larger focus value is retained. For retaining hot spots, the maximum focusing power is used. Based on, and combined with background statistical thresholds Determine the hotspot boundary threshold and select the preferred value. All satisfied Furthermore, grid points connected to the main peak space constitute hotspot areas. Furthermore, based on the peak position of the hotspot Centered on, within radius The focused energy within the range is statistically analyzed, and the hotspot concentration is defined as the ratio of the focused energy in the local core region to the total focused energy in the entire hotspot region, i.e.:

[0091]

[0092] in The preferred size is 20–50 mm. The closer it is to 1, the more concentrated the hotspot energy is, and the more obvious the spatial focusing characteristics of the candidate leakage point are.

[0093] For each candidate hotspot, a direct path model and a reflection path model are established. The direct path model treats the candidate hotspot as a real leakage point, calculating the propagation distance, theoretical arrival delay, and theoretical relative delay of each channel relative to the central reference channel based on its three-dimensional coordinates and the coordinates of each microphone. The actual relative delay is obtained using the generalized cross-correlation of the signals from each channel, preferably employing PHAT weighting; the weighted absolute error between the actual relative delay and the theoretical relative delay is:

[0094]

[0095] In the formula, This represents the actual relative time delay of the m-th channel with respect to the center reference channel. This represents the theoretical relative time delay obtained from the direct path model. This represents the weight of the m-th channel. A time delay consistency score is constructed based on this error; simultaneously, it is determined whether the candidate hotspot is located within the potential leakage area of ​​easily leaking components such as valves, ferrules, and pipe fittings, and a direct-to-the-model geometric score is formed by combining this with the leakage probability of the component itself.

[0096] The reflection path model treats candidate hotspots as virtual images formed by reflection. Structural surfaces or corresponding two-dimensional layers with reflectivity are considered as candidate reflecting surfaces, represented as follows:

[0097]

[0098] In the formula, n is the unit normal vector of the candidate reflecting surface. Let be a point on the candidate reflecting surface. Let the location of the candidate hot spot be... Then its mirror image point relative to the reflecting surface is:

[0099]

[0100] Based on the distance from the mirror point to each microphone, and combined with other deep-layer hotspots, easily leaky structural points, and actual relative time delay relationships, several candidate real sound source points are generated. The propagation paths of these candidate real sound source points to each microphone are simulated via the reflecting surface. The reflection model does not directly adopt the naturally high time delay consistency due to the selection of candidate sound source points. Instead, it evaluates whether the reflection point is within the effective range of the reflecting surface, whether the incident and reflection paths are geometrically reasonable, and the reliability of the structural surface as a reflecting surface, and forms a geometric score for the reflection path accordingly.

[0101] For both the direct-path and reflection models, the number of frequency bands supporting the same candidate hotspot is counted, and the hotspot locations in different frequency bands are extracted. The weighted centers of the hotspot locations in different frequency bands are:

[0102]

[0103] In the formula, This indicates the location of the hot spot in the i-th frequency band. This represents the weight of the i-th frequency band. A positional consistency index is formed based on the positional dispersion of hotspot locations in each frequency band relative to the weighting center. The frequency band co-occurrence count and the positional consistency index are then fused into a multi-frequency band consistency result. The acoustic signal is further divided into multiple continuous time windows, and the frequency of occurrence and positional changes of candidate hotspots within each time window are statistically analyzed to form a time stability index.

[0104] The direct-reach model's time delay consistency score, direct-reach geometric score, location consistency index, and time stability index are weighted and fused to obtain the overall index of the direct-reach model. Similarly, the reflection path geometric rationality, reflection geometric score, location consistency index, and time stability index are weighted and fused to obtain the overall index of the reflection model. The two overall indices are compared and converted into virtual image probabilities. Candidate hotspots are then retained, downgraded, or eliminated according to their virtual image probabilities. Uneliminated candidate hotspots, carrying their virtual image probabilities, are included in the cross-plane fusion process.

[0105] After completing the calculations for all scan surfaces, the retained candidate hotspots from each scan surface are transformed to a unified coordinate system. Hotspots with a spatial distance less than a preset distance threshold are grouped into the same candidate hotspot cluster. A cross-plane consistency score is constructed based on the number of scan surfaces supporting the cluster and the dispersion of hotspot positions on each plane. This score is then combined with the virtual image probability of the hotspots within the cluster to form a cross-plane fusion score. Candidate hotspot clusters with a fusion score higher than a preset threshold are retained, and the optimal hotspot is selected from them. The candidate position, candidate depth layer, propagation distance, spatial confidence level, and virtual image probability are output.

[0106] For the direct path model, based on the above weighted absolute error sum (denoted as in the following formula) Construct the time delay consistency fraction, preferably using an exponential decay form:

[0107]

[0108] in For the allowable time delay residual scale, this embodiment preferably uses... The smaller the delay error, the better. The closer it is to 1, the better. The overall index of the direct path model is obtained by weighting the time delay consistency score, the direct geometric rationality score, the multi-band consistency score, and the time stability score. The calculation formula is:

[0109]

[0110] The overall index of the reflection path model is obtained by weighting the reflection path geometric rationality score, candidate reflector reliability score, multi-band consistency score, and time stability score.

[0111]

[0112] Each component is normalized to the range of 0 to 1. The weights mentioned above are preferred values ​​in this embodiment and can also be adjusted based on the validation set. To avoid making a hard decision based solely on the magnitude of the two total indicators, the score difference between the direct model and the reflection model is converted into the probability of a continuous virtual image.

[0113]

[0114] The sensitivity coefficient is preferably taken as k=8; when When the direct-access model is considered superior, candidate hotspots are retained; when At that time, it was considered that direct and reflection interpretations were competing, so the candidate hotspot was retained but its spatial credibility was reduced and it was included in subsequent verification; when When the reflection model is considered superior, the candidate hotspot is downgraded or eliminated. When a corresponding highly reliable candidate real sound source exists, it is preferred to directly determine it as a reflected virtual image. Candidate reflecting surfaces are obtained in advance from the equipment's structural geometry information, which can be extracted from the equipment's CAD model, on-site 3D point cloud, or manually entered structural dimensions. Priority is given to supporting column surfaces, walls, large equipment shells, metal baffles, pipe support plates, and other structural surfaces with large areas, hard surfaces, and significant sound reflection potential. For curved structures, a local tangent plane approximation can be used. During the detection process, further screening is conducted based on whether the candidate reflecting surface is located between the array and candidate hotspots, whether the mirror path can form an effective reflection point, and whether the propagation path conflicts with solid components. Only geometrically feasible reflecting surfaces are retained for subsequent mirror sound source calculations. After completing the direct-reflection competition evaluation for each scanning surface, candidate hotspots with a 3D spatial distance less than a preset distance threshold in different scanning surfaces are grouped into the same candidate hotspot cluster. In this embodiment, the cross-plane distance threshold is preferably set to... Construct cross-plane consistency scores based on the number of supporting scan surfaces and intra-cluster position dispersion. Then, by combining the average virtual image probability within the cluster, a cross-plane fusion score is formed:

[0115]

[0116] when The candidate hot spot cluster is retained, and the candidate hot spot with the highest fusion score or the highest spatial confidence is selected as the final candidate leakage location. At the same time, the candidate depth layer, propagation distance, spatial confidence and virtual image probability are output.

[0117] Parameter inversion is divided into offline training and online retrieval. During offline training, the acoustic fingerprint features, leak location, propagation distance, environmental conditions, leak pore size label, and leak pressure label for each sample are read from the training library. Missing values ​​are supplemented, and feature transformation is performed. Based on the acoustic frequency f, air temperature T, relative humidity RH, and atmospheric pressure p, the air absorption coefficient is obtained according to the ISO 9613-1 air absorption model.

[0118]

[0119] in , and Let be the relaxation frequencies of oxygen and nitrogen in the air, respectively, calculated according to ISO 9613-1 based on the current temperature, pressure, and water vapor content. For the propagation distance d, the sound level loss due to air absorption is calculated according to... Calculate; using reference distance Based on the baseline, the distance loss caused by spherical diffusion is calculated according to... Therefore, the total propagation loss is calculated as follows: For acoustic fingerprint features of sound pressure amplitude type, the compensated feature values ​​are calculated according to... For band-segmented energy or power-type acoustic fingerprint features, then according to Compensation is performed to convert acoustic features obtained under different propagation distances and environmental conditions to a unified reference state. The same propagation compensation method is used in both the offline training and online detection phases to ensure that the input feature aperture of the inversion model is consistent. In a preferred embodiment, the offline database contains 800 sets of valid leakage samples, with the training set, validation set, and independent test set divided into 560, 120, and 120 sets respectively, according to a ratio of 70%, 15%, and 15%. The leakage aperture covers 0.1–1.0 mm, pipeline pressure covers 5–50 MPa, array-to-leakage source propagation distance covers 0.5–3.0 m, ambient temperature covers 10–40 ℃, and relative humidity covers 30%–90% RH. Different combinations of aperture, pressure, distance, and environmental conditions are used to form sample operating conditions. Leakage orifice diameter labels are obtained from the nominal size of a standard leakage orifice plate or the actual size after microscopic measurement and calibration. Leakage pressure labels are obtained from pressure sensors that have been calibrated. Propagation distance is calculated from the actual three-dimensional coordinates of the leak source and the reference coordinates of the array center. Temperature and relative humidity labels are obtained from temperature and humidity sensors that have been calibrated. Model performance is measured using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination. Evaluation showed that the optimal implementation of the aperture inversion model on the independent test set yielded an MAE of approximately 0.06 mm and an RMSE of approximately 0.09 mm. Approximately 0.91; the MAE of the pressure inversion model is approximately 1.8 MPa, and the RMSE is approximately 2.6 MPa. Approximately 0.93. The above sample size, operating condition coverage, and error index are used to illustrate the preferred implementation method. In practical applications, adjustments can be made based on the operating pressure range of the target equipment, the expected leakage orifice diameter range, and environmental conditions.

[0120] An exponential decay model is used to represent the air absorption loss over the propagation distance, and a spherical diffusion approximation is used to calculate the distance diffusion loss. The two losses are added together to obtain the total propagation loss, which is then used to compensate for the acoustic fingerprint features related to sound intensity or frequency band energy. In the online phase, the same compensation is performed using the propagation distance corresponding to the candidate leak location and the current temperature and humidity information to ensure that the features fed into the inversion model are consistent with the offline training features.

[0121] The compensated dataset was divided into a training set (70%), a validation set (15%), and a test set (15%). The leakage aperture inversion model uses a gradient boosting regression tree with squared error as the loss function, with 300 trees, a maximum tree depth of 6, a learning rate of 0.05, a sample sampling ratio of 0.8, and a feature sampling ratio of 0.8. The leakage pressure inversion model also uses a gradient boosting regression tree with squared error as the loss function, with 400 trees, a maximum tree depth of 5, a learning rate of 0.05, a sample sampling ratio of 0.8, and a feature sampling ratio of 0.8. Specifically, the gradient boosting regression tree used in this embodiment is an XGBoost Regressor.

[0122] During online invocation, the acoustic fingerprint feature vector of the current leakage event, after propagation compensation, is input into the aperture inversion model and the pressure inversion model, respectively, to obtain the aperture estimate and pressure estimate for each credible candidate hotspot cluster. The result evaluation module first checks whether the two estimates exceed the coverage range of the training samples, then checks whether the pressure estimate exceeds the allowable operating pressure range of the equipment, and checks whether the estimation results are consistent with the spatial confidence, virtual image probability, and acoustic fingerprint matching degree. When there is an out-of-training domain, out-of-equipment operating condition, or inconsistent evidence, the corresponding confidence level is reduced, and finally the aperture estimate, pressure estimate, aperture estimate confidence level, and pressure estimate confidence level are output.

[0123] The integrated judgment and feedback system receives four types of evidence. Acoustic evidence includes whether the acoustic fingerprint was successfully extracted, the degree of deviation of the sensitive frequency band from the background, the duration of the event, the probability of acoustic leakage, and the matching degree of the acoustic fingerprint; spatial evidence includes candidate location, spatial credibility, hot spot concentration, main-to-sub-peak ratio, and virtual image probability; parametric evidence includes aperture and pressure estimates, the credibility of both, and the degree of anomaly of the estimates; and operating condition evidence includes pipeline pressure change trends, valve opening and closing status, equipment operating status, temperature, and humidity.

[0124] The validity check of single evidence is carried out as follows: when acoustic fingerprint extraction fails, the target sensitive frequency band deviates insufficiently from the background, or the event duration is extremely short, the acoustic evidence is downgraded or suppressed, among which the extremely short-duration event can be judged as transient interference; when hot spots are scattered, the main peak is not prominent, the probability of virtual images is too high, or the candidate position does not belong to the easily leaking component, the spatial evidence is downgraded or suppressed; when the aperture or pressure estimate is extremely abnormal, the inversion confidence is too low, or the two parameters conflict with each other, the parametric evidence is downgraded or suppressed.

[0125] After completing the single-evidence check, perform acoustic-condition coupling: when the probability of acoustic leakage is high, the pipeline pressure change is close to zero, and the event duration is long, the result is downgraded or enters the review stage; when the acoustic characteristics indicate high-pressure leakage and the current system operating pressure is low, the result is downgraded or enters the review stage; when the acoustic characteristics indicate high-pressure leakage and the pipeline pressure is significantly reduced, the pass priority is increased.

[0126] Perform acoustic-spatial feature coupling: when the acoustic fingerprint matching degree is high but the spatial consistency score is low, the result is downgraded or enters the review; when the high-frequency sensitive band is significantly enhanced but the probability of virtual images is high, the result is downgraded or enters the review; when the location of each detection frequency band is consistent, the hot spots are concentrated and the event duration is long, the pass priority is increased.

[0127] Perform acoustic-interference coupling: When the inversion model indicates a high probability of leakage, but the acoustic fingerprint matches the valve action template well and the valve switching action is detected synchronously, the event is judged as interference and suppressed; when the event duration is short and both the peak value and peak factor are high, the event is judged as interference and suppressed; when the current frequency band enhancement is similar to the fan background change template or the equipment start-up template, the event is suppressed.

[0128] In one embodiment of the present invention, the integrated judgment and feedback system first sets minimum validity thresholds for acoustic evidence, spatial evidence, parametric evidence, and operating condition evidence. For acoustic evidence, the acoustic leakage probability is preferably not less than 0.60, and the acoustic fingerprint matching degree is preferably not less than 0.60. Events lasting less than 0.10 s are preferentially judged as transient impacts or short-term interferences; events lasting 0.20 s or more enter the complete leakage judgment process. For spatial evidence, the hotspot concentration is preferably not less than 0.50, the main-to-sub-peak ratio is preferably not less than 3 dB, and a virtual image probability below 0.35 indicates that the direct model interpretation is dominant; between 0.35 and 0.65, the direct and reflection models are considered to be competing, reducing spatial credibility; a virtual image probability of not less than 0.65 indicates a high virtual image risk, leading to downgrading, suppression, or verification. For parametric evidence, the aperture inversion credibility and pressure inversion credibility are preferably both not less than 0.60. If the inversion results exceed the coverage range of the training samples or the allowable operating condition range of the equipment, the credibility of the parametric evidence is reduced. Regarding pressure changes, considering the different operating pressures of various pipelines and the varying noise levels of sensors, a fixed absolute pressure value is not used. Instead, a pressure change threshold is adaptively set based on normal operating pressure fluctuations and the current operating pressure. The calculation formula is as follows: ,in This represents the standard deviation of pressure fluctuations under normal and stable operating conditions. The current or baseline operating pressure; when the absolute value of the pressure change within the detection time window meets the following conditions. At the time, it is considered that there is a significant pressure change. However, for minor leaks, the absence of a significant pressure drop should not be used as a sole criterion for rejecting a leak event. It should only be used as evidence of consistency across multiple physics fields in a comprehensive judgment.

[0129] After completing the validity checks of each piece of evidence and the judgment of multiphysics coupling rules, the valid acoustic evidence will be scored. Spatial evidence scoring Parametric evidence scoring and working condition evidence scoring Normalize to 0-1, and according to A comprehensive reliability score C is obtained; acoustic and spatial evidence directly determine the existence of leak characteristics and the reliability of candidate locations, therefore they are assigned a weight of 0.30 each, parametric evidence is assigned 0.25, and operational condition evidence, as an auxiliary consistency check, is assigned 0.15. When C < 0.40, the event is judged as a suppressed event, and no formal leak warning is output; when... When a low-confidence warning is issued, it indicates that some leakage characteristics have been detected but the evidence is not yet sufficient, and the process enters a phase of continuous observation or verification; when The system outputs a reliable early warning when most evidence supports the occurrence of a leak, and outputs candidate locations, orifice sizes, and pressure estimates; when... The timely output of highly reliable early warnings indicates a high degree of consistency between acoustic, spatial, parameter, and operational evidence.

[0130] Based on the results of single-evidence checks and multiphysics coupling checks, output suppression, low-confidence warnings, medium-confidence warnings, or high-confidence warnings are generated. For events confirmed as interference by manual checks or subsequent operating conditions, their acoustic and operating condition characteristics, along with their labels, are written into the interference database; for confirmed leakage events, the corresponding characteristics and operating condition labels are written into the leakage database. If rule-based gating consistently detects a significant discrepancy between a certain type of inversion result and the operating condition evidence, it is fed back to the parameter inversion system, reducing the confidence weight of that type of input in subsequent training or online estimation.

[0131] To avoid error accumulation caused by directly writing erroneous model results back to the database, events are only allowed to be added as new samples to the leak database after being manually verified on-site, confirmed by independent hydrogen concentration sensors or other independent detection methods, or when similar events consistently achieve high-confidence warnings for at least three independent detection cycles and their spatial location and acoustic fingerprint remain stable. For interfering events such as valve actions, mechanical impacts, and sudden fan changes confirmed by manual or independent operating condition evidence, their acoustic characteristics and operating condition labels are added to the interference database. If a certain type of input shows significant inconsistency with the operating condition evidence in five or more consecutive confirmed events, a confidence weight adjustment is triggered. The weights are updated incrementally, with each adjustment preferably not exceeding 5% of the original weight, and the update formula is as follows: The update coefficient is preferably selected from the following values: And set upper and lower limits for the weights. This is to avoid abrupt changes in model weights caused by a few abnormal events. The above thresholds and weights are preferred values ​​in this embodiment, and can be adjusted based on the validation set results, pipeline operating pressure range, and on-site background noise level in actual applications.

[0132] As shown in the figure, a hydrogen leak acoustic imaging localization and leak parameter inversion system includes a data acquisition system, a feature modeling system, a leak localization system, a parameter inversion system, a data transmission and control system, and a comprehensive judgment and feedback system. The data transmission and control system is connected to the other five systems and is responsible for unified time synchronization, buffering, preprocessing, data packaging, protocol transmission, task scheduling, and feedback control.

[0133] The data acquisition system includes a hierarchical heterogeneous MEMS microphone array, pipeline pressure sensors, valve status sensors, temperature and humidity sensors, and equipment operation status sensors, as shown in the figure. A multi-channel data acquisition interface receives analog signals from each microphone after front-end conditioning; an analog-to-digital converter (ADC) converts the analog voltage signals into digital signals; analog or digital signals from the status sensors enter the unified control link through their respective interfaces. A time synchronization module adds timestamps to the data using a unified clock source, and an edge processor performs data aggregation and preliminary processing.

[0134] The feature modeling system includes a feature extraction and preprocessing module, a frequency band discrimination prior establishment module, and an acoustic fingerprint extraction module. It performs the specific processing of obtaining the frequency band discrimination prior through fusion normalization and screening and preprocessing acoustic fingerprint features in step S2. The leak location system includes a beamforming module, a virtual image removal module, and a spatial consistency evaluation module. It performs the specific processing of beamforming, virtual image removal, and spatial consistency evaluation in step S3. The parameter inversion system includes a propagation compensation module, an aperture pressure inversion module, and a result evaluation module. It performs the specific processing of propagation and environmental compensation, leak aperture and leak pressure inversion, and result evaluation in step S4. The comprehensive judgment and feedback system includes a single evidence inspection module, a multi-physics coupling rule gating module, and a comprehensive evaluation and closed-loop feedback correction module. It performs the specific processing of single evidence validity inspection, multi-physics coupling consistency inspection, comprehensive judgment, and closed-loop feedback correction in step S5.

[0135] During data packaging and communication, acoustic data, operating condition data, channel identifiers, gain status, timestamps, and check fields are encapsulated into a unified data frame. Front-end hardware can aggregate data via SPI, I2S, or high-speed serial interfaces, while the host computer communication module can send data to the back-end software host via industrial Ethernet, TCP / IP, UDP, or USB interfaces. The host computer sequentially schedules feature modeling, leak location, parameter inversion, and comprehensive judgment based on detected events as task units, and sends database update or trusted weight adjustment information back to the corresponding modules through a feedback control interface, thus forming a closed loop.

[0136] Preferably, the data transmission and control system uses an AMD Xilinx Zynq-7000 XC7Z020 SoC as the edge processor. The analog-to-digital conversion module uses four Texas Instruments ADS127L18 24-bit eight-channel synchronous sampling ADCs to form a 32-channel acquisition system, of which 25 channels are used for the hierarchical heterogeneous MEMS microphone array, and the remaining channels are reserved. All ADCs share a unified master clock, with a sampling rate of 256 kS / s. The edge terminal is configured with 1 GB DDR3 circular cache, and the raw acoustic data rate is approximately 153.6 Mbit / s based on 25 channels, which can store data at least 2 seconds before the event and 5 seconds after the event. The edge terminal communicates with the host computer via 1000BASE-T gigabit Ethernet, continuous acoustic data is transmitted via UDP, and control parameters and feedback information are transmitted via TCP / IP. Data is encapsulated in 10 ms basic frames. Each frame includes at least a frame header, protocol version, frame type, frame sequence number, timestamp, number of channels, number of sampling points, channel identifier, gain status, acoustic data, pressure, temperature and humidity, valve status, equipment operating status, and a CRC32 checksum field. The length of a single frame is approximately 192.1 kByte. The host computer uses an Intel Core i7-12700 or equivalent industrial computer with at least 32 GB of memory, running Ubuntu 22.04 LTS. Data acquisition, synchronization, latency compensation, caching, and communication are deployed at the edge, while feature modeling, leak location, parameter inversion, multiphysics determination, and database management are deployed on the host computer. Database updates and trusted weight adjustment information are transmitted back via TCP / IP, forming a closed-loop processing chain.

[0137] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks, characterized in that, include: Step S1: Acoustic signals are collected using a layered heterogeneous microelectromechanical system microphone array consisting of a central reference unit, a high-frequency detection layer, a mid-frequency detection layer, and a low-frequency detection layer. At the same time, pipeline pressure, valve status, equipment operating status, and ambient temperature and humidity information are collected. The collected signals are then conditioned at the front end, synchronized in time, and transmitted after being unified to the same time axis. Step S2 involves preprocessing and normalizing the original frequency domain features, original time domain features, original features at each layer, and original spatial features in the offline database. The offline database includes a background database, an interference database, and a leakage database with operating condition labels. The target frequency band is divided into bands, and the band energy, inter-layer energy ratio and coefficient of variation are calculated. Inter-class separation degree analysis, interference exclusion analysis, operating condition robustness analysis and random forest importance analysis are performed to obtain the frequency band discrimination prior and the screening results of acoustic fingerprint features. Based on the filtering results, the acoustic fingerprint feature vector of the current leakage event is extracted from the acoustic signal collected in step S1. Step S3: Discretize the continuous three-dimensional space into multiple two-dimensional scanning surfaces, and perform beamforming based on array geometric parameters, sound velocity parameters, the prior frequency band discrimination capability, and multi-channel acoustic signals to extract candidate hot spots; For each candidate hotspot, a direct path model with the candidate hotspot as the actual leakage point and a reflection path model with the candidate hotspot as the reflection virtual image and the mirror point obtained using the mirror sound source method are established. The virtual image probability of the candidate hotspot is obtained by comparing the total index of the two models. Candidate hotspots are retained, downgraded, or eliminated according to the virtual image probability. Furthermore, cross-plane fusion is performed on the candidate hotspots retained in different two-dimensional scanning planes to obtain the leakage candidate location, propagation distance, candidate depth layer, spatial confidence, and virtual image probability. Step S4: The acoustic fingerprint feature vector, leakage candidate location, propagation distance, candidate depth layer, spatial confidence, virtual image probability and environmental temperature and humidity information are used for parameter inversion. After propagation and environmental compensation, the leakage aperture inversion model and leakage pressure inversion model are called respectively to obtain the leakage aperture estimate, leakage pressure estimate and corresponding confidence. Step S5: Summarize acoustic evidence, spatial evidence, parametric evidence, and operational condition evidence, and sequentially perform single evidence validity checks and multi-physics coupling consistency checks. Output the leakage event level, and update the offline database based on the judgment results, or adjust the credibility weights of the corresponding inputs in subsequent parameter inversion.

2. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 1, characterized in that, The central reference unit is a central reference microphone located at the center of the array; the high-frequency detection layer consists of four high-frequency microelectromechanical system (MEMS) microphones arranged in a compact square, used to acquire sensitive information in the 20kHz to 40kHz and higher frequency bands and perform local fine correction; the mid-frequency detection layer consists of eight mid-frequency MEMS microphones arranged in an equidistant concentric circle array, used for primary localization and mid-frequency sound wave extraction; the low-frequency detection layer consists of twelve low-frequency MEMS microphones arranged in an equidistant concentric circle array, used for stable detection and auxiliary spatial constraints in complex backgrounds; the central reference microphone is used for normalization and interlayer relative energy ratio calculation.

3. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 1, characterized in that, The front-end conditioning in step S1 includes bandpass filtering, signal amplification, and anti-aliasing filtering performed sequentially. Selectable gain control is set during signal amplification, and different sampling ranges are selected according to different signal intensities and the corresponding gain states are recorded. Each acoustic channel uses a unified clock source for simultaneous sampling, or uses quasi-simultaneous sampling with fixed time delay compensation, so that the acoustic signal and operating condition information are unified to the same time axis.

4. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 1, characterized in that, In step S2, the target frequency band is divided into four sub-bands: 1kHz to 5kHz, 5kHz to 15kHz, 15kHz to 40kHz, and 40kHz to 80kHz. The coefficient of variation is calculated from the frame energy of each sub-band within multiple consecutive short time windows and is used to characterize the short-time stability of the sub-band. The larger the coefficient of variation, the worse the stability of the sub-band. The frequency band discrimination prior is obtained by fusing and normalizing the leakage-background interlayer separation, leakage-interference interlayer separation, interference exclusivity, operational robustness characterized by fluctuations in interlayer separation under different operating conditions, and random forest importance; wherein, the interlayer energy ratio of the high-frequency detection layer relative to the central reference cell is calculated according to the following formula: In the formula, This represents the high-frequency detection layer energy of the k-th sample in the i-th frequency band. This represents the energy of the corresponding central reference unit, where ε is a positive number set to prevent the denominator from being too small; the coefficient of variation is calculated according to the following formula: And the mean is calculated according to the following formula: In the formula, and These represent the standard deviation and mean of the frame energy, respectively. This represents the frame energy of the k-th sample within the t-th short time window of the ith frequency band; the random forest importance index is calculated according to the following formula: In the formula, , and The importance of frequency band energy, interlayer energy ratio, and short-time frequency band stability are respectively.

5. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 4, characterized in that, The acquisition of the acoustic fingerprint features includes: The spectral main peak features, band energy features, inter-layer relative energy ratio features, spectral shape features, temporal features, and spatial consistency features extracted in step S2 are used to form a candidate feature set. The leakage-background separation degree and leakage-interference separation degree of any candidate feature are calculated to obtain a discrimination score, which is then combined with the operating condition robustness coefficient to obtain a discrimination index. The top six candidate features are selected as acoustic fingerprint features based on the discrimination index. The acoustic fingerprint features are then subjected to z-score normalization, reference benchmark normalization based on the energy of the central reference unit, background bias compensation under the same operating condition, and fingerprint weighting based on the discrimination score. The z-score normalization is performed according to the following formula: In the formula, For the j-th feature of the k-th sample, This is the mean of the feature in the training database. The standard deviation of this feature in the training library is ε, which is a positive number set to prevent the denominator from being too small; the reference benchmark normalization is performed according to the following formula: In the formula, and These are the high-frequency detection layer energy and the central reference unit energy of the k-th sample in the i-th frequency band, respectively; the background bias compensation under the same operating condition is performed according to the following formula: In the formula, This represents the average value of the background library in the i-th frequency band under the same operating conditions.

6. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 1, characterized in that, Step S3 includes: Within each two-dimensional scanning plane, a spatial focusing heatmap is generated based on the beamforming results. First, a mid-frequency detection layer and a low-frequency detection layer are used for coarse localization, and then a high-frequency detection layer is used for fine correction near the candidate hotspot. Candidate peaks are extracted from the spatial focusing heatmap, and weak peaks below the background statistical threshold and adjacent repeating peaks are suppressed. The peak size, hotspot area, main-to-sub-peak ratio, hotspot concentration, and corresponding three-dimensional coordinates of the candidate hotspots are extracted. The k-th two-dimensional scanning plane is represented by the following formula: In the formula, Let k be the reference point for the k-th two-dimensional scan surface. and Let be the in-plane basis vectors of this two-dimensional scanning surface. and These are local coordinates in the plane.

7. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 6, characterized in that, The direct path model constructs a time delay consistency score based on the weighted absolute error between the theoretical relative time delay of each microphone and the actual relative time delay obtained from the generalized cross-correlation of signals from each channel, and constructs a geometric score based on whether the candidate hotspot is located at a location prone to leakage. The reflection path model obtains the mirror point and simulated reflection path using the mirror source method based on the candidate reflecting surface, and obtains a geometric score based on the geometric rationality of the path and the reliability of the reflecting surface. The time delay consistency score or the path geometric rationality evaluation result, the geometric score, the position consistency index, and the time stability index are weighted and fused to obtain the total indexes of the direct path model and the reflection path model, respectively, and the comparison of the two total indices is converted into the virtual image probability. The position consistency index is constructed from the position dispersion of each frequency band hotspot position relative to its weighted center. The candidate reflecting surface is expressed by the following formula: In the formula, Let be the unit normal vector of the candidate reflecting surface. Let be a point on the candidate reflective surface; the mirror image of the candidate hot spot rs relative to the candidate reflective surface is calculated according to the following formula: The weighted absolute error between the actual relative delay and the theoretical relative delay of the direct path model is calculated according to the following formula: In the formula, Δτm represents the actual relative time delay of the m-th channel relative to the central reference channel. This represents the theoretical relative time delay obtained from the direct path model. This represents the weight of the m-th channel; the weighted center of the hotspot location in different frequency bands is calculated according to the following formula: In the formula, Indicates the location of the hot spot in the i-th frequency band. This represents the weight of the i-th frequency band.

8. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 1, characterized in that, Step S4, propagation and environmental compensation, includes: obtaining the air absorption coefficient according to the air absorption model based on the sound wave frequency, air temperature, relative humidity, and atmospheric pressure; representing the air absorption loss using an exponential decay model; calculating the distance diffusion loss using the spherical diffusion approximation; and compensating for the acoustic fingerprint characteristics by adding the air absorption loss and the distance diffusion loss together. The air absorption coefficient is expressed as... In the formula, f is the frequency of the sound wave propagating in the air, T is the air temperature, RH is the relative humidity of the air, and p is the atmospheric pressure. The leakage aperture inversion model and the leakage pressure inversion model are both gradient boosting regression tree models with squared error as the loss function. The training data is divided into a training set of 70%, a validation set of 15%, and a test set of 15%. The leakage aperture inversion model has 300 trees, a maximum tree depth of 6, a learning rate of 0.05, a sample sampling ratio of 0.8, and a feature sampling ratio of 0.

8. The leakage pressure inversion model has 400 trees, a maximum tree depth of 5, a learning rate of 0.05, a sample sampling ratio of 0.8, and a feature sampling ratio of 0.

8.

9. The method for acoustic imaging localization and leakage parameter inversion of hydrogen leaks according to claim 1, characterized in that, The single evidence validity check in step S5 includes checking at least one of the following: acoustic fingerprint extraction results, deviation of the target sensitive frequency band from the background, event duration, acoustic leakage probability, candidate hot spot concentration, virtual image probability, candidate location, main-to-sub-peak ratio, degree of inversion result anomaly, and inversion credibility; the multi-physics coupling consistency check includes acoustic-operating condition coupling, acoustic-spatial feature coupling, and acoustic-interference coupling, and outputs suppression, low credibility warning, medium credibility warning, or high credibility warning.

10. A system for acoustic imaging localization and leakage parameter inversion of hydrogen leaks, characterized in that, A method for implementing any one of claims 1 to 9 includes a data acquisition system, a feature modeling system, a leak location system, a parameter inversion system, a data transmission and control system, and a comprehensive judgment and feedback system; the data transmission and control system is connected to the data acquisition system, the feature modeling system, the leak location system, the parameter inversion system, and the comprehensive judgment and feedback system respectively; the data acquisition system includes a hierarchical heterogeneous array composed of low-frequency, medium-frequency, and high-frequency microelectromechanical system microphones, as well as pipeline pressure sensors, valve status sensors, temperature and humidity sensors, and equipment operating status sensors; the feature modeling system... The system includes a feature extraction and preprocessing module, a frequency band discrimination prior establishment module, and an acoustic fingerprint extraction module; the leak location system includes a beamforming module, a virtual image removal module, and a spatial consistency assessment module; the parameter inversion system includes a propagation compensation module, an aperture pressure inversion module, and a result evaluation module; the comprehensive judgment and feedback system includes a single evidence inspection module, a multi-physics coupling rule gating module, and a comprehensive evaluation and closed-loop feedback correction module; the data transmission and control system includes a multi-channel data acquisition interface, an analog-to-digital conversion module, a serial data interface module, a time synchronization module, an edge processor, a host computer communication module, and a feedback control interface.

Citation Information

Patent Citations

  • Sound source orientation estimation method for compressed gas leakage in wireless sound sensor network

    CN106908755A

  • Gas leakage detection method and application thereof

    CN121163776A