Fire-fighting pipeline leakage identification method and system based on multi-modal data processing

Abstract

The application discloses a fire-fighting pipeline leakage identification method and system based on multi-modal data processing, relates to the field of pipeline leakage identification, and first fuses sound-vibration waveforms and geographic coordinate data, and utilizes variational modal decomposition technology to decouple a complex broadband leakage signal into a series of narrowband intrinsic modal components. On the basis, a spatial topology construction technology based on GPS is introduced, an accurate physical path provided by geographic coordinates is used as a spatial constraint vector, a closed-loop feedback mechanism of acoustics and geographic information is constructed. By applying the spatial constraint to adaptive group velocity matching under the frequency dispersion effect, the actual group velocity of different frequency bands in the pipe propagation is dynamically calculated. Finally, combined with a multi-modal weighted fusion strategy, the accurate positioning result is obtained by comprehensively considering the corrected distance difference of each frequency band, and effective compensation and correction of the frequency dispersion error in the non-uniform medium by using the multi-modal data are realized.

Description

Technical Field

[0001] This application relates to the field of pipeline leakage identification, and more specifically, to a method and system for identifying fire pipeline leakage based on multimodal data processing. Background Technology

[0002] With the acceleration of urbanization, fire protection pipelines, as the lifeline for public safety, are expanding in coverage and becoming increasingly complex in topology. Fire protection pipelines are typically laid with a mixture of carbon steel, galvanized pipes, and some polyethylene (PE) pipes. Long-term underground or concealed operation makes it difficult to detect even minor leaks in a timely manner, resulting not only in water waste but also potentially causing serious consequences due to insufficient water pressure during a fire. In modern pipeline monitoring, with the development of IoT technology, the collected data exhibits significant multimodal characteristics, including both acoustic and vibration waveform data reflecting fluid dynamics and geographic coordinate data reflecting the spatial location of the pipeline network. How to effectively integrate this heterogeneous data to meet the precise positioning needs in complex pipe environments has become a pressing technical challenge for the industry.

[0003] However, most existing pipe leak detection schemes rely on a single acoustic signal processing technique and are generally based on idealized physical models, assuming that the propagation speed of sound waves in pipes is a constant, such as standard underwater acoustics or the universal wave velocity in steel pipes. This fixed sound velocity model ignores the non-uniformity of the physical medium in fire protection pipe networks. In reality, when sound waves propagate in bounded solid or liquid-solid coupled pipes, significant dispersion effects occur, meaning that the propagation speeds of sound waves of different frequencies are not consistent, and phase velocity and group velocity separate. After the broadband noise generated by the leak propagates over long distances, the time difference between high-frequency and low-frequency components reaching the sensor changes due to velocity differences, resulting in wave packet broadening and waveform distortion. If a single velocity is forcibly used for calculation, it will produce systematic cumulative errors that are difficult to correct. Furthermore, when multimodal data applications are involved, existing technologies suffer from severe spatiotemporal fragmentation problems. Geolocation System (GPS) data is often only used to mark the final leak location on a map and does not actually participate in the physical process of acoustic calculation. In fact, the precise Euclidean distance provided by geographic coordinates is a key constraint for the reverse calibration of unknown dispersion characteristics within the tube, but existing schemes have failed to construct such an acoustic-geographic closed-loop feedback mechanism. Traditional algorithms tend to perform crude averaging on signals across the entire frequency band, masking the true phase delay information in specific advantageous frequency bands.

[0004] Therefore, there is an urgent need for a leak identification method that can deeply integrate spatiotemporal multimodal data and perform adaptive speed matching for dispersion effects. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this application provides a method for identifying leaks in fire-fighting pipelines based on multimodal data processing. The method includes: S1: performing temporal domain alignment and spatial label binding on the raw sensor data streams collected from each node of the fire-fighting pipeline network to generate an aligned dataset containing synchronous acoustic waveforms and corresponding geographic coordinate data; S2: performing VMD-based hierarchical decomposition on the synchronous acoustic waveforms in the aligned dataset to obtain an intrinsic mode component set; S3: constructing a pipeline network spatial topology based on GPS constraints on the geographic coordinate data in the aligned dataset to obtain a spatial constraint vector; S4: performing adaptive group velocity matching under dispersion effects on the intrinsic mode component set and the spatial constraint vector to obtain an adaptive velocity matrix; S5: based on the adaptive velocity matrix, performing hierarchical cross-correlation calculation and distance transformation on the intrinsic mode component set to obtain dispersion-corrected distance difference sets for different frequency bands; S6: performing multimodal weighted fusion positioning calculation on the dispersion-corrected distance difference set according to the energy proportion of the intrinsic mode component set to obtain the final leak location result.

[0006] This application also provides a fire pipeline leakage identification system based on multimodal data processing, comprising: a sensor data stream processing module for performing temporal domain alignment and spatial label binding on the raw sensor data streams collected from each node of the fire pipeline network to generate an aligned dataset containing synchronous acoustic and vibration waveforms and corresponding geographic coordinate data; an intrinsic mode component generation module for performing VMD-based signal frequency domain hierarchical decomposition on the synchronous acoustic and vibration waveforms in the aligned dataset to obtain an intrinsic mode component set; a spatial constraint module for performing GPS-constrained pipeline spatial topology construction on the geographic coordinate data in the aligned dataset to obtain a spatial constraint vector; an adaptive group velocity matching module for performing adaptive group velocity matching under dispersion effects on the intrinsic mode component set and the spatial constraint vector to obtain an adaptive velocity matrix; a dispersion-corrected distance difference set generation module for performing hierarchical cross-correlation calculation and distance transformation on the intrinsic mode component set based on the adaptive velocity matrix to obtain dispersion-corrected distance difference sets for different frequency bands; and a positioning result generation module for performing multimodal weighted fusion positioning calculation on the dispersion-corrected distance difference set according to the energy proportion of the intrinsic mode component set to obtain the final leak location result.

[0007] Compared with existing technologies, this application provides a method and system for identifying leaks in fire-fighting pipelines based on multimodal data processing. This aims to solve the positioning error problems caused by existing technologies, which mainly rely on fixed sound velocity models that ignore dispersion effects and fail to effectively utilize geospatial constraints. First, by fusing acoustic waveforms and geographic coordinate data, variational mode decomposition (VMD) technology is used to decouple the complex broadband leak signal into a series of narrowband intrinsic mode components, precisely capturing the frequency domain characteristics of the signal at a microscopic level. Based on this, GPS-based spatial topology construction technology is introduced, using the precise physical path provided by geographic coordinates as a spatial constraint vector to construct a closed-loop feedback mechanism between acoustic and geographic information. By applying spatial constraints to adaptive group velocity matching under dispersion effects, the actual group velocity of different frequency bands propagating within the pipe is dynamically calculated, thereby correcting the wave velocity differences caused by frequency variations. Finally, combined with a multimodal weighted fusion strategy, the corrected distance differences of each frequency band are integrated to obtain accurate positioning results, achieving effective compensation and correction of dispersion errors in non-uniform media using multimodal data. Attached Figure Description

[0008] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0009] Figure 1 This is a flowchart of a fire pipeline leakage identification method based on multimodal data processing according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the data flow in the fire pipeline leakage identification method based on multimodal data processing according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of step S5 in the fire pipeline leakage identification method based on multimodal data processing according to an embodiment of this application.

[0012] Figure 4 This is a block diagram of a fire pipeline leakage identification system based on multimodal data processing according to an embodiment of this application. Detailed Implementation

[0013] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0014] In view of the shortcomings in the above-mentioned technical fields, this application proposes a method for identifying leaks in fire pipelines based on multimodal data processing. Figure 1 This is a flowchart of a fire pipeline leakage identification method based on multimodal data processing according to an embodiment of this application. Figure 2This is a schematic diagram of the data flow in the fire pipeline leakage identification method based on multimodal data processing according to an embodiment of this application. Figure 1 and Figure 2 As shown, the fire pipeline leakage identification method based on multimodal data processing according to an embodiment of this application includes: S1: performing temporal domain alignment and spatial label binding on the raw sensor data streams collected from each node of the fire pipeline network to generate an aligned dataset containing synchronous acoustic waveforms and corresponding geographic coordinate data; S2: performing VMD-based signal frequency domain hierarchical decomposition on the synchronous acoustic waveforms in the aligned dataset to obtain an intrinsic mode component set; S3: performing GPS-constrained pipeline spatial topology construction on the geographic coordinate data in the aligned dataset to obtain a spatial constraint vector; S4: performing adaptive group velocity matching under dispersion effect on the intrinsic mode component set and the spatial constraint vector to obtain an adaptive velocity matrix; S5: performing hierarchical cross-correlation calculation and distance transformation on the intrinsic mode component set based on the adaptive velocity matrix to obtain dispersion-corrected distance difference sets for different frequency bands; S6: performing multimodal weighted fusion positioning calculation on the dispersion-corrected distance difference set according to the energy proportion of the intrinsic mode component set to obtain the final leak location result.

[0015] In step S1, the raw sensor data streams collected from each node of the fire protection network are time-domain aligned and spatially labeled to generate an aligned dataset containing synchronous acoustic waveforms and corresponding geographic coordinates. It should be understood that in a distributed network monitoring scenario, because each monitoring node is physically dispersed and operates independently, its internal crystal oscillators are affected by factors such as ambient temperature and aging, resulting in frequency drift and deviations in the local time reference between different nodes. If unaligned data is directly used for cross-correlation analysis, the asynchronous clock error between nodes will overlap with the physical propagation delay of the sound waves, leading to serious deviations or even failures in the positioning calculation; simultaneously, the simple acoustic signal lacks geospatial attributes and cannot be directly mapped to the physical network topology. Therefore, rigorous temporal alignment and spatial label binding of the raw sensor data stream are performed to establish a unified spatiotemporal reference system. The timing error between nodes is eliminated through a hardware-level synchronization mechanism, and the signal is given precise geographic coordinate constraints. This generates an aligned dataset containing synchronized acoustic and vibration waveforms and corresponding geographic coordinate data, providing accurate source data support for the subsequent construction of an acoustic-geographic closed-loop feedback mechanism and adaptive group velocity correction under dispersion effects.

[0016] In one embodiment of this application, step S1 includes: S11, performing PPS edge detection and gated analog-to-digital conversion on the raw sensor data stream to obtain a digitized raw acoustic-vibration sequence and a GPS message cache to be parsed; S12, parsing the standard positioning statement in the GPS message cache to be parsed, and extracting Coordinated Universal Time and latitude and longitude coordinates from it to obtain a geographic spatiotemporal label containing spatiotemporal attributes; S13, writing the geographic spatiotemporal label as metadata into the associated structure of the digitized raw acoustic-vibration sequence to obtain an aligned dataset.

[0017] The specific processing is as follows: Step S11 is achieved through a parallel acquisition mechanism based on hardware interrupts and precise timing control. The raw sensor data stream is a collection of multi-source heterogeneous signals, physically originating from the piezoelectric ceramic vibration sensor installed on the outer wall of the fire-fighting pipeline and its integrated GNSS global navigation satellite system module. The piezoelectric sensor senses the weak elastic waves generated by fluid pressure fluctuations or leakage jets on the pipe wall in real time, converting them into a continuously changing analog voltage signal. This signal contains broadband noise characteristics generated by leakage. The GNSS module continuously outputs ASCII code message streams following the NMEA-0183 protocol through a serial port, and simultaneously outputs a high-precision second pulse signal PPS with a frequency strictly locked to 1Hz through a dedicated pin. The rising edge of this pulse signal is aligned with the whole second of Coordinated Universal Time (UTC), with the error typically controlled at the nanosecond level, serving as the absolute benchmark for achieving time-domain alignment of the data stream. In the specific implementation process, the core control unit of the monitoring terminal, namely a high-performance MCU or FPGA with a floating-point arithmetic unit, monitors the changes in the PPS signal level from the GNSS module in real time through its external interrupt pin. The controller is pre-configured with a high-priority hardware interrupt service routine, with the trigger mode set to rising edge triggering. When the voltage level of the PPS signal jumps instantaneously from a logic low level to a logic high level, the controller's interrupt detection logic captures this edge transition and determines that the current microsecond is the absolute physical zero point of the sampling period. At this time, the interrupt service routine immediately executes the gating instruction, starting the analog-to-digital converter (ADC) module through direct memory access (DMA) or hardware triggering mechanism. This gated ADC is not in free-running mode but is strictly controlled by the synchronous acquisition of the PPS signal. Once the ADC is triggered, the acquisition window opens, and the analog voltage signal from the piezoelectric sensor is discretized using preset sampling parameters. The preset sampling frequency needs to be set according to Shannon's sampling theorem and the frequency domain characteristics of the fire pipeline leakage signal. For example, considering that the effective energy of leakage noise in metal pipelines is mainly concentrated in the 500Hz to 2kHz or even wider frequency band, the sampling frequency can be set to 10240Hz, that is, 10240 points are acquired per second. Within a specified 1-second sampling period, the ADC continuously and at equal intervals quantizes analog voltage values, generating a digital sequence containing N data points, such as 10240. During this process, the first sampling point in the sequence (index n=0) strictly corresponds to the rising edge of the PPS signal, i.e., the absolute physical zero point; the absolute physical time of the subsequent nth sampling point is rigidly locked by hardware timing logic to the sum of the physical zero point time and the sampling interval n / Fs (Fs is the sampling frequency). This ensures that even if different nodes are several kilometers apart, as long as they can all receive satellite signals, the start times of their acquired digitized acoustic and vibration sequences are highly synchronized on the physical time axis. Parallel to the analog-to-digital conversion process is the buffering processing of GPS message data.While the ADC performs high-frequency voltage sampling, the controller's Universal Asynchronous Receiver / Transmitter (UART) interface continuously listens to the data stream transmitted by the GNSS module. Due to the low transmission rate of NMEA messages (baud rate of 9600bps or 115200bps) and their asynchronous nature, message arrival often lags behind the PPS pulse. Therefore, a circular buffer or double buffering mechanism is used to completely write all character data received within one second after the PPS pulse arrival into the memory area. This character data contains statements such as Recommended Positioning Information (RMC) or Positioning Information (GGA), mainly including the UTC timestamp, longitude, latitude, and positioning validity flag corresponding to that second. This step does not immediately parse these characters but encapsulates them as a raw byte stream, forming a buffer of GPS messages to be parsed. Finally, when the 1-second acquisition cycle ends—that is, before the preset number of N points are acquired or the next PPS pulse arrives—the controller packages the 10240 voltage quantization values ​​stored in the DMA buffer into a digital raw acoustic-vibration sequence; simultaneously, it locks the NMEA byte stream captured within the same time window, forming a buffer of GPS messages to be parsed. These two parts of data are linked in memory through a strict one-to-one mapping using structure pointers or timestamp indexes.

[0018] Step S12 first performs an integrity check on the cached data, using an XOR checksum algorithm to verify whether any errors occurred during transmission. In the data stream that passes the check, the parsing program retrieves specific frame header identifiers using a string matching algorithm, focusing on identifying the Recommended Minimal Positioning Information Statement ($GPRMC) or Global Positioning System Positioning Information Statement ($GPGGA) containing full-dimensional positioning information. These statements not only contain latitude and longitude data from satellite positioning but also provide high-precision Coordinated Universal Time (UTC) through satellite atomic clock synchronization. Taking the parsing of the $GPRMC statement as an example, if the valid statement extracted from the cache is $GPRMC,083559.00,A,3112.3456,N,12134.5678,E..., the parsing logic first splits the string into independent fields based on comma delimiters. 083559.00 represents the current Coordinated Universal Time (UTC) time of 08:35:59:00. Since the PPS hardware triggering mechanism in step S11 ensures that the start time of the acquisition cycle strictly corresponds to the whole-second boundary of this UTC time, this timestamp is recognized as the absolute time reference for the 0th sampling point of the acoustic-vibration sequence. The following 3112.3456,N and 12134.5678,E represent 31°12.3456′N and 121°34.5678′E, respectively. It should be noted that the original coordinate format output by the NMEA protocol is in degrees and minutes, which is not suitable for direct use in subsequent spatial topology construction and Euclidean distance calculation. Therefore, before generating geospatial labels, coordinate format standardization conversion is required. The following formula is used to convert the degrees and minutes format to a decimal format suitable for numerical calculations: In the formula, This represents the converted decimal longitude or latitude. This refers to the integer degree portion of the original data, such as 31 or 121. This refers to the fractional and decimal parts of the original data, such as 12.3456 or 34.5678. Using the example data above, the converted latitude value is approximately 31 + 12.3456 / 60 ≈ 31.20576 degrees, and the longitude value is approximately 121 + 34.5678 / 60 ≈ 121.57613 degrees. After the numerical conversion, the parsing logic encapsulates the extracted Coordinated Universal Time Stamp (08:35:59), the converted high-precision latitude and longitude coordinates (31.20576, 121.57613), and the location status flag into a structured metadata object, which is the geospatial tag containing spatiotemporal attributes.

[0019] Step S13: At the memory operation level, allocate a new contiguous memory block or instantiate a custom data structure object, which includes a file header and a data body. During implementation, key information from the geospatial-geographical tags is first written into the file header area of ​​the data structure. This establishes a strict temporal correspondence. According to the hardware triggering principle of S11, the sampling point with index 0 in the acoustic-vibration sequence array physically occurs at the rising edge of the PPS signal, which corresponds to the exact second of Coordinated Universal Time (UTC) in the tag. Therefore, writing the UTC timestamp into the file header essentially defines the zero point of the data packet's time. Subsequently, the converted decimal latitude and longitude coordinates are written as the inherent spatial attribute of this data segment. To adapt to the monitoring needs of large-scale pipeline networks, a unique device identifier (UID) also needs to be generated in the file header using a specific algorithm. The UID is composed of the hardware's MAC address, device serial number, and current UTC timestamp. For example, it can be generated as a string similar to DEV-AA:BB:CC-1704962159, ensuring that each data packet can be uniquely indexed and traced back to a specific physical node after massive amounts of data are uploaded to the cloud server. After constructing the header information, the quantized digitized raw acoustic vibration sequence from step S11 is filled into the data body area. Due to the large amount of acoustic vibration signal data, such as 10240 16-bit integers, this step involves high-speed copying of memory blocks. At this point, a complete data packet structure is assembled, and its logical structure can be represented as {UID, geospatial tag, digitized raw acoustic vibration sequence}. The final aligned dataset is a multimodal information capsule with highly standardized features, ensuring zero-latency alignment of acoustic waveforms and geographic coordinates in the temporal dimension and strong correlation in the spatial dimension.

[0020] In step S2, the synchronous acoustic waveforms in the aligned dataset are subjected to VMD-based frequency domain hierarchical decomposition to obtain the intrinsic mode component set. Correspondingly, in practical applications of fire-fighting pipeline network leakage monitoring, due to the heterogeneity of pipe materials (such as the mixing of carbon steel and PE pipes) and the complexity of fluid-pipe wall coupling, the acoustic signals generated by leakage often exhibit significant non-stationary and broadband characteristics. High-frequency jet noise and low-frequency pipe wall vibration are mixed together, and the propagation speeds of different frequency components in the medium have physical differences, i.e., dispersion effects. If the broadband mixed waveforms in the time domain are directly analyzed, it is difficult to distinguish the velocity delays under different propagation modes, leading to severe mode aliasing in the location calculation. Therefore, this application introduces variational mode decomposition (VMD) technology to finely decouple the complex broadband acoustic signal in the frequency domain into several narrowband intrinsic mode components with independent center frequencies, thereby effectively filtering out environmental background noise and separating signal components with different frequency characteristics.

[0021] In one embodiment of this application, step S2 includes: S21, performing Hilbert transform and spectrum initialization on the synchronous acoustic waveforms in the alignment dataset to obtain an initial variational parameter set, and simultaneously separating the metadata cache from the alignment dataset; S22, performing iterative optimization of the initial variational parameter set using the alternating direction multiplier method to obtain the optimized modal spectrum matrix and the final center frequency vector; S23, performing an inverse Fourier transform on the optimized modal spectrum matrix to restore each narrowband time-domain signal, and structurally recombining it with the final center frequency vector and the metadata cache to obtain the intrinsic modal component set.

[0022] The specific processing is as follows: Step S21 reads the aligned dataset, which contains unique identifiers (UIDs), geospatial labels, and a digitized raw acoustic vibration sequence. In this step, the processing logic splits the dataset into two parallel processing streams: one is to extract the digitized raw acoustic vibration sequence containing 10240 sampling points (based on the 10.24kHz sampling rate in S1), i.e., the synchronous acoustic vibration waveform, which is defined as the time-domain input signal to be decomposed, denoted as... Secondly, the geospatial-geographical label containing the Coordinated Universal Time stamp 08:35:59 and latitude / longitude coordinates (31.20576, 121.57613) is separated and stored in a separate metadata cache. This is then applied to the extracted time-domain signal. Since its essence is the acquired real physical signal, its spectrum exhibits conjugate symmetry, meaning it simultaneously contains both positive and negative frequency components. During variational mode decomposition, the negative frequency components not only carry redundant information but also interfere with the iterative estimation of the center frequency. Therefore, in practice, the Hilbert transform is first employed to construct the analytic signal. Specifically, for the original real signal... Perform Hilbert transform to obtain This transformation is equivalent to lagging all frequency components of the signal by 0.5π in phase. Using the formula... Construct an analytic signal, where The unit is the imaginary unit. Converting to the frequency domain, the one-sided spectrum of this analytic signal retains only the positive frequency portion. For example, for a signal containing leakage broadband noise... After this step, the frequency domain energy distribution will be completely concentrated in the range [0, Fs / 2], that is, from 0 to 5120 Hz, eliminating negative frequency interference. Then, the crucial parameter setting and initialization stage begins to construct the initial variational parameter set. The core of variational mode decomposition lies in solving a constrained variational problem, that is, finding... A set of intrinsic mode functions is used to minimize the sum of the estimated bandwidths of each mode, and the sum of the bandwidths of each mode equals the original signal. Here, the preset number of mode decomposition layers is used. and penalty factor These are two crucial hyperparameters. The value is determined based on the typical spectral characteristics of fire pipeline leakage signals. Since leakage signals mainly consist of fluid jet noise, pipe wall vibration, and low-frequency waves coupled with the soil, the physical modes are relatively distinct. Therefore, in this embodiment, [the following is used]. The value is set to 5 to decompose the signal into five main physical frequency bands. Penalty factor. This determines the modal bandwidth of the decomposed mode. The larger the value, the narrower the modal bandwidth of the decomposed signal. Considering the non-stationarity of fire duct signals, in order to achieve a balance between frequency separation and signal fidelity, this application will... The value was set at around 2000. This value was determined through empirical statistical analysis of a large amount of sample data or an adaptive algorithm based on entropy. and Afterwards, the algorithm needs to initialize the iterative variables for the variational problem. This includes initializing... A set of center frequencies Frequency domain mode function set and Lagrange multipliers For the center frequency The initial value is set using a uniform distribution strategy in this embodiment, that is, five frequency points are uniformly selected between 0 and the Nyquist frequency (5120Hz) as the initial guess value, for example... ={0Hz,1280Hz,2560Hz,3840Hz,5120Hz}, or initialized to 0. Frequency domain mode function The initial matrix is ​​typically set to a random small value or a zero matrix containing the noise level to avoid artificially introducing initial bias. Lagrange multipliers Similarly initialized to zero, its function is to transform the constrained variational problem into an unconstrained problem during the iteration process using the dual ascent method. Finally, the one-sided spectrum obtained above after the Hilbert transform... The set hyperparameters ( =5, =2000) and the initialized variable matrix ( , , They are collectively encapsulated as an initialization variational parameter set.

[0023] Step S22 first receives the initialization variational parameter set, then the algorithm enters the core iterative loop of the Alternating Direction Multiplier Method (ADMM). The ADMM algorithm decomposes the original constrained variational problem into a series of easier-to-solve subproblems, approximating the global optimum by alternately updating various variables. Its core objective is to minimize the sum of the estimated bandwidths of each mode component while ensuring that the sum of all recovered mode components equals the original signal. In each iteration loop, such as the transition from the nth to the (n+1th)th iteration, the mode spectrum update operation is performed first. For For each mode k in the modalities, k = 1~5, the spectrum of the other modes is preserved. Center frequency and Lagrange multipliers The new spectrum of the mode in the (n+1)th iteration is obtained by solving the extreme points of the quadratic optimization problem, which remains unchanged. This update process follows the analytical formula below: This formula is equivalent to Wiener filtering, where This represents the frequency domain value of the k-th intrinsic mode component after the (n+1)-th iteration. This is the spectrum of the original input signal (i.e., the collected acoustic and vibration signals from the fire-fighting pipeline). This represents the spectrum of the i-th mode. (Molecular part) This represents the residual signal after removing contributions from all other modes in the original signal, expressed as a Lagrange multiplier. Constraint corrections are performed. The denominator then constitutes a value with a center frequency located at... The reciprocal of the transfer function of the bandpass filter. Penalty factor. This plays a crucial role in controlling the filter bandwidth. The larger the value, the faster the denominator grows far from the center frequency, and the narrower the passband of the filter becomes, thus forcing the updated mode to... The energy is highly concentrated at the center frequency Nearby. For example, if the current processing involves the third mode representing the leaked principal energy, and its center frequency is estimated to be around 1200Hz, then this formula will extract the residual energy near 1200Hz in the original spectrum through the funnel effect, forming a new frequency domain waveform for that mode. Complete all After updating the modal spectrum, the algorithm immediately performs a center frequency update. The new center frequency... Based on the newly calculated modal spectrum The energy center of gravity is determined. Its calculation formula is: This formula essentially calculates the centroid frequency of the modal power spectrum. Physically, this means that as iterations proceed, the center frequency of the mode adaptively shifts towards the location of the highest energy concentration within that frequency band. For example, if the center frequency of the second mode is initially roughly initialized to 1280Hz, but after the first mode update, the actual energy peak in that band is found to be at 1150Hz, then the center frequency calculated using this formula... The frequency will be adjusted from 1280Hz to a value closer to 1150Hz. This ensures that VMD can automatically lock onto strong characteristic frequencies in the leakage signal, such as the specific whistling frequency of fluid flowing through a crack. After updating the modes and center frequency, the final step is to update the Lagrange multipliers. Through the formula Perform dual ascent update, where This step updates the step size. Its purpose is to enforce reconstruction constraints, ensuring that the sum of all decomposed modes can accurately reconstruct the original signal numerically, preventing information loss. These three update steps are repeated in a loop until a preset convergence condition is met. The convergence criterion is defined as the normalized mean square error of the mode spectrum changes between two consecutive iterations being less than a very small threshold. ,like ,Right now When this condition is met, the iteration terminates, and the optimized modal spectrum matrix is ​​output, which contains the frequency domain complex sequences of the five finally converged narrowband modes. and the corresponding final center frequency vector For example, the final result is The vector could be {230Hz, 850Hz, 1420Hz, 2100Hz, 3500Hz}, which locates the dominant vibration mode of a specific leakage event.

[0024] Step S23: For each mode in the optimized modal spectrum matrix Perform an inverse Fourier transform (IFFT). Since an analytic signal was constructed in S21, the result of the IFFT here is an analytic signal in complex form. Taking its real part allows us to reconstruct the real-valued narrowband time-domain signal. At this time It is no longer a chaotic broadband noise, but a noise with a specific center frequency. An amplitude-frequency modulation (AM-FM) signal whose amplitude and frequency change slowly over time. For example, the decomposed... This likely represents a fluid jet noise component with a center frequency of 1420Hz, whose waveform is clear and has a significantly higher signal-to-noise ratio than the original signal. Finally, a structured reconstruction is performed. This step addresses the aforementioned spatiotemporal fragmentation problem by reassigning spatiotemporal attributes to the processed pure physical signal. The metadata cache, temporarily separated and frozen in S21, is invoked to retrieve the device UID corresponding to this data segment, such as DEV-A-20240112, and the geospatial tag (including the UTC time stamp 08:35:59 and latitude / longitude coordinates). Then, a multi-level complex data object, namely the intrinsic modal component set, is constructed. The root node of this object contains metadata information to ensure data traceability; the leaf nodes are lists containing K=5 child elements, each consisting of a restored time-domain waveform array. and its corresponding final center frequency Composition. The final generated object structure can be represented as: {UID:"DEV-A...", tag:{time:"08:35:59", latitude:31.205, longitude:121.576}, intrinsic modal component table:[{waveform: Center frequency: 230 Hz, Waveform: Center frequency: 850 Hz, ..., Waveform: Center frequency: 3500

[0025] In step S3, the geographic coordinate data in the aligned dataset is used to construct the pipeline spatial topology based on GPS constraints to obtain the spatial constraint vector. It is understandable that the time delay reflected by acoustic signals is essentially the time it takes for a sound wave to travel a certain distance in a physical medium. However, acoustic methods alone cannot determine the exact distance the sound wave travels, leading to a mathematical dilemma of three unknowns: unknown distance, unknown sound speed, and unknown leak point. Traditional positioning methods typically assume that the pipeline is straight or rely on rough distances measured manually. This approach introduces significant geometric errors when dealing with complex urban underground fire hydrant networks with numerous right-angle bends and detours. Furthermore, GPS only obtains the straight-line distance projected onto the ground surface, while sound waves propagate strictly along the internal path of curved pipes; the physical meanings of these two are fundamentally different. Without establishing a mapping relationship from the straight line on the ground surface to the underground pipe diameter, any high-precision acoustic algorithm will deviate from the true solution due to the failure of physical constraints. Therefore, step S3 implements GPS-constrained pipeline spatial topology construction, aiming to use precise geographic coordinates as a rigid benchmark, and through spherical geometry algorithms and topology correction strategies, to restore the actual physical path length of the sound wave, transforming the originally isolated acoustic problem into a physical problem subject to strict geometric constraints, thereby obtaining a spatial constraint vector that can truly reflect the sound wave propagation distance.

[0026] In one embodiment of this application, step S3 includes: S31, parsing the latitude and longitude information of sensor nodes from the aligned dataset and performing spherical coordinate system transformation to obtain standardized radian coordinate pairs; S32, performing spherical great circle distance measurement on the standardized radian coordinate pairs to obtain the Euclidean distance on the ground surface representing the mapped straight-line distance between nodes; S33, performing topological mapping and physical path correction on the Euclidean distance on the ground surface to obtain a spatial constraint vector.

[0027] The specific processing is as follows: Step S31 first accesses the aligned dataset and extracts two sensor nodes to be analyzed for correlation, denoted as node i and node j as their geographic spatiotemporal labels. For example, the extracted node i is located at the position shown in the example of step S1 (latitude 31.20576°, longitude 121.57613°), while its adjacent node j is located at (latitude 31.20650°, longitude 121.57700°). At this point, the coordinate data is still in decimal angle format, which is easy for humans to read but not suitable for computers to perform trigonometric function calculations. Therefore, the processing logic needs to convert it to mathematically standard radians. According to the conversion formula... Calculate the radian coordinates of the two nodes respectively. Taking node i as an example, its latitude in radians... =31.20576×3.14159265 / 180≈0.54464 radians, longitude in radians =121.57613×3.14159265 / 180≈2.12189 radians. Similarly, the result for node j is... ≈0.54466, ≈2.12191. These two sets of values ​​correspond to { }and{ It is encapsulated as a structured object called a normalized radian coordinate pair.

[0028] Step S32: Since the Earth is an approximate ellipsoid, directly applying the Pythagorean theorem in plane geometry would introduce curvature errors for urban pipe networks on a scale ranging from hundreds of meters to kilometers. Using complex ellipsoidal geodesic algorithms would be computationally too expensive and unnecessary. Therefore, this embodiment uses the Haversine Formula to calculate the great circle distance between two points. This formula is particularly suitable for calculating the shortest path between two points on a sphere, and it effectively maintains numerical stability when the distance between the two points is very close (e.g., the distance between nodes in a fire duct network is typically tens to hundreds of meters). The calculation process is as follows: First, calculate the square of the sine of half the difference in latitude between the two nodes. And the product of the square of half the longitude difference and the cosine product of the two latitudes. .in , Next, calculate the intermediate variables. Great circle distance The final calculation formula is In this formula, Representing the average radius of the Earth, and considering that fire protection pipelines are usually buried in shallow layers, a value of 6371.004 kilometers is taken, which is 6371004 meters. Substituting into the previous numerical example, the result is obtained through calculation. The value is extremely small (corresponding to a short distance). By multiplying the arcsine and radius, the surface projection straight-line distance between node i and node j can be finally calculated. For example... =116.7 meters. This value is the Euclidean distance at the Earth's surface, which physically represents the theoretical shortest connection length between two sensors if the pipeline were laid straight along the Earth's surface.

[0029] Step S33 aims to address the inconsistency between surface straight lines and underground pipelines. In actual fire-fighting pipelines, right-angle bends, U-bends, and even detours are often present during installation to avoid building foundations, other pipelines, or follow road routes. If 116.7 meters is directly used as the sound propagation distance, while the actual pipeline length is 150 meters, the calculated sound velocity will be artificially amplified, leading to leak location errors. This is addressed in two scenarios. The first scenario involves a high-precision GIS (Geographic Information System) model. A pre-built pipeline GIS model is a digital three-dimensional topology network, typically containing a node table (recording the precise coordinates of valves and tees), an edge table (recording pipe segment length, diameter, and material), and a connection matrix. The node coordinates parsed in S31 are mapped to the GIS layer, identifying the corresponding positions of nodes i and j in the pipeline map. Then, the classic Dijkstra's shortest path algorithm or A* search algorithm is used to search the pipeline topology map for a series of physical pipe segments connecting these two nodes. For example, the search results show a straight-line distance of only 123.45 meters, but the actual path includes an 80-meter straight pipe section, a 90-degree bend (equivalent to 2 meters in length), and another 68-meter straight pipe section. In this case, the physical path length... =80 + 2 + 68 = 150 meters. The formula is expressed as follows: The second scenario involves situations where accurate GIS data is lacking or only CAD drawings are available. In this case, a pipe curvature correction system needs to be introduced. This coefficient is an empirical parameter greater than 1, used to estimate the additional length of the pipeline due to bends and detours. It is typically based on historical construction codes for the area or statistical data from partially surveyed pipeline sections. For example, in a regular rectangular urban road network, fire hydrants are usually laid at right angles along the roads, with a Manhattan distance to Euclidean distance ratio close to... ≈1.414; while for relatively straight water conveyance trunk lines, It may be taken as 1.05~1.10. In this embodiment, if it is determined that the two nodes are located in a complex internal building pipe network, a preset value can be used. =1.25. Therefore, the estimated physical path length is... =116.7 × 1.25 ≈ 145.88 meters. Finally, the corrected physical length obtained through any of the above paths. (e.g., 150 meters or 145.88 meters) are encapsulated as a constraint vector. This vector not only contains numerical scalars but also includes path confidence weights (the weight of GIS measurement results is greater than that of the estimated results), which clearly define the actual distance constraint of the sound wave propagating in the waveguide. The final spatial constraint vector is a matrix or list set that contains the physical distance information of all relevant pipe segments within the entire monitoring area.

[0030] In step S4, adaptive group velocity matching under dispersion effects is performed on the intrinsic mode component set and spatial constraint vector to obtain the adaptive velocity matrix. It should be understood that sound waves do not propagate in an infinitely vast free fluid medium, but are strictly confined within a finite-boundary waveguide composed of metal or polymer composite materials. In this liquid-solid coupling physical field, the elastic characteristics of the pipe wall cause the sound wave propagation to exhibit a significant dispersion effect; that is, the sound wave propagation speed is no longer a constant (e.g., 1480 m / s in standard underwater acoustics), but rather undergoes nonlinear drift with frequency changes. There is an objective physical difference between the group velocities of high-frequency and low-frequency components within the pipe. If this microscopic physical phenomenon is ignored, forcibly applying a unified sound velocity model to the modal components with different center frequencies decomposed in step S2 will lead to a misalignment in the calculation benchmarks of the arrival times of different modes, thus preventing the time-difference-based positioning results from converging to the true physical leak point. Therefore, step S4 implements adaptive group velocity matching under dispersion effect using the intrinsic mode component set and spatial constraint vector. The aim is to construct a frequency-velocity dynamic mapping relationship that conforms to the waveguide physical mechanism based on the accurate pipeline geometry parameters and material properties, and to assign each independent narrowband mode component its actual group velocity in a specific physical pipeline.

[0031] In one embodiment of this application, step S4 includes: S41, extracting modal feature frequencies from the intrinsic modal component set to obtain the modal effective frequency vector; S42, performing dispersion model parameter matching on the spatial constraint vector to obtain the dispersion characteristic parameter set; and S43, performing adaptive group velocity calculation on the modal effective frequency vector and the dispersion characteristic parameter set to obtain the adaptive velocity matrix.

[0032] The specific processing is as follows: Step S41: Although the VMD algorithm has already output the center frequencies of each mode during the iteration process... However, in non-stationary leakage signals, the actual frequency of the energy centroid often fluctuates slightly around this center frequency. To obtain frequency values ​​that better reflect the physical nature of energy transport, we iterate through each mode k in the eigenmode component set. First, for each eigenmode function in the real domain... Perform the Hilbert transform again to construct its analytic signal. The instantaneous frequency function of the mode can be obtained by differentiating the arctangent function of the ratio of the imaginary to the real part of the analytic signal. To collapse the time-varying instantaneous frequency into a scalar value that represents the overall propagation characteristics of the mode, the algorithm employs an energy-weighted averaging strategy. Using the formula... Perform the calculation. The physical meaning here is the amplitude. The frequency corresponding to a larger time point contributes more to the group velocity. For example, in step S2, the third mode... The initial center frequency is 1420Hz. After the energy weighting correction of the instantaneous frequency mentioned above, the calculated effective group frequency of the mode may be: =1418.5Hz. The calculations for all five modes are completed sequentially, ultimately generating a modal effective frequency vector containing five precise frequency values. For example, {231.2,848.6,1418.5,2095.4,3498.1} (unit: Hz). This vector accurately describes the discrete distribution characteristics of the leaked signal energy in the frequency domain.

[0033] Step S42: To calculate dispersion, the microscopic physical properties of the pipe material must be known. This requires introducing the core component pipe material physical property library. This property library is a pre-installed relational database or hash map in memory, and its specific architecture includes a material index table and a physical parameter table. The material index table establishes a mapping relationship between geographic coordinate region IDs and pipe types; the physical parameter table records in detail the elastic modulus of various pipe materials, such as 20# carbon steel, 304 stainless steel, ductile iron, PE100, etc. Poisson's ratio Pipe wall density and the density of the fluid coupled with it (usually water). and water bulk modulus During implementation, the configuration information of the current fire-fighting pipe segment is first retrieved from the attribute library based on the location information (such as the associated UID or GIS layer ID) in the spatial constraint vector. If the pipe segment is identified as a DN150 galvanized steel pipe, its standard geometric dimensions are then extracted: outer diameter... =165mm, wall thickness =4.5mm, therefore the pipe inner diameter =165-2×4.5=156mm, water bulk modulus =2.15×10 9 Pa. Simultaneously extract material properties: elastic modulus. =2.06×10 11 Pa, for the theoretical speed of sound It is directly set to the theoretical speed of sound in an infinitely large fluid medium (such as still water). For a normal temperature freshwater environment, it is usually taken as =1480 m / s. This value represents the physical upper limit velocity without elastic constraint from the pipe wall. Subsequently, two key parameters describing the dispersion characteristics are calculated: the cutoff frequency. and dispersion attenuation coefficient The cutoff frequency is usually related to the circumferential resonant frequency of the pipe. For metal pipes, ,in The longitudinal wave velocity of the pipe is typically set to 15000Hz for DN150 metal pipes. Dispersion attenuation coefficient. This is a dimensionless empirical coefficient related to the fluid-structure interaction strength, which can be preset through experimental calibration or Biot's theoretical model. For example, let... =0.12. Finally, regarding the theoretical speed of sound... =1480m / s, cutoff frequency Such as 15000Hz, pipe inner diameter =0.156m, wall thickness =0.0045m, water bulk modulus =2.15×10 9 Pa, elastic modulus =2.06×10 11 Pa and dispersion attenuation coefficient =0.12, etc., are uniformly encapsulated into a set of dispersion characteristic parameters.

[0034] Step S43 is crucial for the mathematical synthesis of frequency domain features and spatial physical features. It involves calling the frequency vector obtained in step S41. And the dispersion parameter set obtained in step S42, for each frequency in the vector The group velocity is calculated one by one using the geometric dispersion group velocity model formula. In one embodiment of this application, step S43, which involves adaptively calculating the group velocity of the modal effective frequency vector and the dispersion characteristic parameter set to obtain an adaptive velocity matrix, includes: adaptively calculating the group velocity of the modal effective frequency vector and the dispersion characteristic parameter set using the following formula:

[0035]

[0036] in, For the theoretical speed of sound, The dispersion attenuation coefficient is... Let be the effective group frequency of the k-th intrinsic mode component. This refers to the loop frequency or cutoff frequency of the pipeline. The inner diameter of the pipe. For pipe wall thickness, The bulk modulus of water For elastic modulus, For frequency The group velocity of the sound wave packet propagating inside the pipe. This formula is based on a simplified engineering physical model and is used to characterize the nonlinear attenuation of the sound velocity with frequency within a bounded pipe. Analysis of the formula shows that... The term indicates that the dispersion effect is proportional to the square of the frequency, meaning that the higher the frequency, the greater the velocity deviation. The more intense the degree; The term reflects the influence of pipe geometry; the larger the diameter-to-thickness ratio (the thinner the pipe wall), the more pronounced the drag effect of pipe wall elasticity on sound velocity. Specific calculation process: For the first mode (low frequency). =231.2Hz, the frequency is relatively low at this point. ≈0.0002, the correction term is almost zero. Calculation results ≈1266.1 m / s. This aligns with the physical understanding that low-frequency plane waves propagate at almost a reference speed. For the third mode (mid-frequency)... =1418.5Hz. Substituting into the formula, the correction term begins to have a slight effect, and the group velocity may be slightly adjusted to... ≈1266.8 m / s. For the 5th mode (high frequency), =3498.1Hz. At this time ≈0.054. Substituting this into the formula, although the change is small, the velocity value will shift definitely. For example, the calculated result would be... ≈1272.5 m / s. It is worth noting that the specific increase or decrease depends on the formula... The symbol definitions and waveguide modes are shown here, with differences illustrated by numerical variations. Calculations are performed sequentially for all five frequency points, and the results {1266.1, 1266.3, 1266.8, 1269.1, 1272.5} are encapsulated into a 5×1 numerical matrix. This matrix is ​​the adaptive velocity matrix. Each element of this matrix is ​​no longer a fuzzy empirical value, but rather precisely corresponds to the physical propagation velocity of each modal component in step S2.

[0037] In step S5, based on the adaptive velocity matrix, hierarchical cross-correlation calculation and distance transformation are performed on the intrinsic mode component set to obtain the dispersion-corrected distance difference set for different frequency bands. Correspondingly, after solving the microscopic decomposition of the signal, the macroscopic constraints of the spatial path, and the adaptive matching of the physical sound velocity, the core challenge of leak location shifts to obtaining the time difference and the final distance calculation. In complex pipeline environments, due to the superposition of multipath echoes and environmental background noise, directly performing time-domain cross-correlation on the entire waveform often produces a broad main peak, leading to unclear time delay estimation and significantly reduced accuracy. Simultaneously, traditional location methods use a coarse calculation mode of single average time delay × single average sound velocity, which masks the objective differences in the propagation behavior of signals in different frequency bands. High-frequency components are not only fast but also attenuate significantly, while low-frequency components, due to their slower speed and strong diffraction ability, often carry more stable long-distance phase information. Forcibly mixing these components in the calculation will confuse the different time delays generated by different physical velocities, causing the calculated distance difference to deviate from the true leak point. Therefore, step S5 performs hierarchical cross-correlation calculation and distance transformation, aiming to extract sharpened high-precision time delay at each independent frequency level using generalized cross-correlation and phase transformation techniques, and combine the group velocity calculated in the previous steps to independently solve the physical distance difference of each frequency band, thereby obtaining the dispersion-corrected distance difference set for different frequency bands.

[0038] Figure 3 This is a flowchart of step S5 in the fire pipeline leakage identification method based on multimodal data processing according to an embodiment of this application. Figure 3 As shown, in one embodiment of this application, step S5 includes: S51, performing cross-power spectrum sharpening and inverse time-domain transformation on the intrinsic mode component set to obtain a generalized cross-correlation function set; S52, performing peak search and subsampling-level precision fitting on the generalized cross-correlation function set to obtain a layered time delay vector; S53, calculating the dispersion-corrected distance difference between the layered time delay vector and the adaptive velocity matrix to obtain a dispersion-corrected distance difference set.

[0039] The specific processing is as follows: Step S51: For any two monitoring nodes i and j (their positions have been specified in S3), the algorithm first traverses each mode level k in the intrinsic mode component set, k=1…5. Taking the third mode (center frequency approximately 1418.5Hz) as an example, the time-domain waveform of the third mode of node i is extracted respectively. The third mode time-domain waveform of node j Subsequently, the Fast Fourier Transform (FFT) algorithm was used to transform it to the frequency domain, yielding the complex spectrum. and In the frequency domain, these two spectra contain not only amplitude information, but more importantly, phase information, and the time delay between the two signals is precisely encoded in the phase difference. In order to calculate the time delay corresponding to the phase difference, the algorithm first calculates the cross-power spectrum of the two signals. ,in This represents complex conjugation. Ideally, the inverse transform of the cross spectrum yields a cross-correlation function composed of time-delayed pulses. However, in actual fire-fighting pipelines, wall reflections and reverberation of sound waves introduce non-direct wave components into the cross spectrum, making the time-domain correlation peaks broad and less sharp. Therefore, a PHAT (phase transform) weighting technique is introduced. The processing unit applies the weighting function... The cross-power spectrum is whitened. This weighting operation essentially normalizes the amplitude of the cross-spectrum to 1, retaining only the phase information. Subsequently, the generalized cross-correlation function is calculated using the inverse Fourier transform (IFFT) formula: The output of this formula is a time-domain function. Its physical meaning is to describe the displacement of two modal signals at different times. Phase similarity. Due to PHAT sharpening, the peaks in this function become extremely sharp, similar to Dirac. This function effectively suppresses sidelobe interference caused by long reverberation within the tube. The above operation is performed sequentially on all five modal levels, ultimately generating a set of generalized cross-correlation functions containing five independent time-domain functions. .

[0040] Step S52: In traditional integer index search, the resolution of latency is limited by the sampling rate. Taking the 10.24kHz sampling rate used in this embodiment as an example, the sampling interval... ≈97.6 microseconds. If the speed of sound is 1200 m / s, then the error at one sampling point corresponds to a distance error of approximately 0.12 meters, which is clearly insufficient for precise positioning. To overcome this physical bottleneck, the algorithm implements a sub-sampling level interpolation strategy. Using the cross-correlation function of the third mode... For example, first, by traversing and searching, we find the integer index corresponding to the maximum value of the function. .exist Once determined, the algorithm does not directly... Instead of using time delay, the peak point and its two adjacent points on the left and right are selected. , , If these three points actually lie on a continuous parabola (based on the Taylor series approximation of the relevant peak apex), the true position of the vertex can be fitted using the three-point parabola interpolation formula. Let the equation of the parabola be... By solving for the coefficients using the coordinates of three points, the x-coordinate offset of the extreme point (vertices) can be derived. The final high-precision fractional time difference of arrival. For example, if the integer peak value is found to be at the 531st sampling point, the offset can be calculated using fitting. =0.34, then the actual delay is (531+0.34)×97.65μs≈51884 microseconds, or approximately 0.05188 seconds. Compared to taking only the integer 531 points corresponding to 51852 microseconds, this corrects the delay by 32 microseconds, or approximately 4 centimeters of distance accuracy. The algorithm sequentially calculates the high-precision delays of the five modes and encapsulates them into a hierarchical delay vector. .For example, ={0.05123,0.05145,0.05188,0.05210,0.05245} seconds. It can be seen that, due to the dispersion effect, there are indeed slight differences in the actual arrival times of different frequency components.

[0041] Step S53: Prior to this, step S4 yielded the precise group velocity matrix for each mode, i.e., {1266.1, 1266.3, 1266.8, 1269.1, 1272.5} m / s. In this step S52, a hierarchical high-precision time delay vector was obtained. Now, the logical frequency-velocity-time alignment operation is performed. For each mode level k, the formula is used... Calculate the distance difference for each frequency band. The core innovation of this operation lies in abandoning global variables and instead multiplying the high-frequency delay by the high-frequency speed, and the low-frequency delay by the low-frequency speed. Taking data stream as an example: Mode 1 (230Hz): =1266.1m / s × 0.05123s ≈ 64.862 meters. Third mode (1418Hz): =1266.8m / s × 0.05188s ≈ 65.721 meters. Fifth mode (3498Hz): =1272.5m / s × 0.05245s ≈ 66.742 meters. It can be clearly seen that, despite detecting the same physical leak point, due to the differences in the propagation characteristics of sound waves at different frequencies, the calculated distance differences in different frequency bands exhibit significant dispersion (the difference is close to 2 meters). This dispersion is a manifestation of the dispersion effect. If a simple averaging or using only a single velocity is employed, the calculated result might fall to 65 meters or 66 meters, introducing an error on the order of meters. These five independent physical distance differences are denoted as... The interval {64.862, 65.210, 65.721, 66.230, 66.742} is fully preserved and encapsulated as a dispersion-corrected distance difference set. Each element represents the location of the leak as seen from a physical perspective within a specific frequency band.

[0042] It is understandable that, in the process of connecting the signal processing domain and the physical space domain, the time delay accuracy of the discrete generalized cross-correlation function is limited by the sampling interval of the hardware system. For example, at common industrial sampling rates, the time axis is not only discrete but also inherently subject to quantization errors. Especially considering the extremely high propagation speed of sound waves in fire ducts (the speed of sound in metal pipes often exceeds 1000 m / s or even higher), even a one-microsecond time measurement error will be amplified by the velocity coefficient and transformed into a distance positioning deviation on the order of millimeters or even centimeters. Furthermore, due to the dispersion effect verified in the previous steps, the peak of the cross-correlation function often no longer maintains a standard symmetrical shape but instead undergoes asymmetrical broadening or becomes a distorted bell shape. In this physical context, if the traditional parabolic interpolation method, which only uses three points near the peak for quadratic curve fitting, continues to be used, its assumption based on waveform symmetry will no longer hold, thus introducing an unavoidable systematic bias. Therefore, this application introduces a subsampling delay estimation based on Sinc kernel function reconstruction and Newton iteration. It aims to leverage the mathematical completeness of the Nyquist-Shannon sampling theorem and introduce a more robust optimization algorithm from the field of numerical analysis to break the physical limitation of the sampling rate without increasing hardware costs. Through mathematical reconstruction, it extracts a nanosecond-level accurate delay that approximates the Cramer-Rao lower bound (CRLB), thereby effectively resisting waveform distortion caused by dispersion and ensuring the physical fidelity of the final layered delay vector.

[0043] Based on this, in a preferred embodiment of this application, step S5 includes: S5-1, performing cross-power spectrum sharpening and inverse time-domain transformation on the intrinsic mode component set to obtain a generalized cross-correlation function set; S5-2, performing subsampling time delay estimation on the generalized cross-correlation function set based on Sinc kernel function reconstruction and Newton iteration to obtain a layered time delay vector; S5-3, calculating the dispersion-corrected distance difference between the layered time delay vector and the adaptive velocity matrix to obtain a dispersion-corrected distance difference set. It is worth noting that the implementation processes of S5-1 and S5-3 are the same as those of S51 and S53 in the above embodiments, and will not be repeated here. In particular, the implementation process of S5-2 will be described in detail here.

[0044] In specific implementation, the system first receives a set of generalized cross-correlation functions from step S5-1, which includes... For example, five discrete cross-correlation sequences, denoted as... For each modal level k, the algorithm first performs a coarse peak search, traversing the entire discrete sequence to find the index of the sample point with the highest cross-correlation value. This point represents the integer-precision time delay estimate. Subsequently, the algorithm enters the subsampling refinement stage. To avoid the huge computational overhead of Sinc interpolation reconstruction of the entire curve, the processing scheme adopts a local window optimization strategy. A window with integer peak values ​​is set... Local window radius centered Based on experience and computational resources, we set the value to 10, meaning we only utilize 2M+1=21 sample points near the peak. The information contained within is used to reconstruct the main peak shape. This strategy is based on the theory that the continuous waveform of a band-limited signal can be completely reconstructed from its discrete samples using the Sinc function, transforming the problem of finding the true peak into finding the point where the first derivative (gradient) of the reconstruction function is zero in the continuous domain. The algorithm uses the Newton-Raphson iterative method to solve it. First, a continuous time delay variable is defined. Its initial value is set to an integer index. In the k-th iteration, it is necessary to calculate the first derivative of the objective function. and second derivative Based on the analytic derivative properties of the Sinc function, a formula for calculating the first derivative (gradient) is constructed:

[0045]

[0046] In this formula, This represents the cross-correlation magnitude of the m-th sample point within the window. This represents the current continuous time delay estimate. The term within square brackets is the analytical form of the derivative of the Sinc function. The physical meaning of this formula lies in calculating the current position using the weighted information of all points within the window. The slope at that point. Simultaneously, to determine the step size and direction of the iteration, a formula for calculating the second derivative (scalar form of the Hessian matrix) is constructed:

[0047]

[0048] This formula describes the curvature characteristics of the objective function at the current position. The auxiliary variables... The existence of the second derivative ensures that the algorithm can perceive the sharpness and asymmetry of the peaks, thereby intelligently adjusting the approximation speed. After obtaining... and Then, Newton's iterative update formula is applied:

[0049]

[0050] in The learning rate or damping factor is set to 1 to utilize the second-order convergence property. Specifically, this is used, for example, when processing the cross-correlation function of the third mode (center frequency 1418Hz) output in step S51, as a coarse search for the integer peak index. =15. The current sampling rate is 10.24kHz, and the sampling interval is... ≈97.65μs. 1. Initialization: Set... =15.0. 2. First iteration: using The 21 nearby sample points were substituted into the formula for calculation. Due to dispersion causing a tail on the right side of the waveform (skewness), the calculated first derivative g(15.0) is positive (indicating the true peak is on the right), and the second derivative h(15.0) is negative (indicating peak convexity). The update step size was then calculated. ≈0.32. Therefore... =15.32. 3. Second iteration: In The derivative is recalculated at 15.32. At this point, the gradient g(15.32) decreases rapidly, approaching 0. The calculated fine-tuning step size is approximately 0.0256. Therefore... =15.3456. 4. Convergence Judgment: Stop when the step size is less than the preset threshold (e.g., 10^{-4}). The final subsampling position is 15.3456. Finally, convert the converged floating-point index to physical time: 15.3456 × 97.65 μs ≈ 1498.5 microseconds. Compared to 1464.7 microseconds corresponding to taking only the integer 15, the correction is about 33.8 microseconds. At the sound speed of this frequency band determined in step S4 of 1266.8 m / s, this time correction corresponds to an improvement in distance accuracy of about 4.28 centimeters. This is done sequentially for all In this application, the above iterations are performed for 5 modalities, ultimately generating a high-precision hierarchical delay vector. The time delay vector obtained based on this preferred embodiment exhibits significant accuracy and physical consistency advantages in subsequent data processing. This nanosecond-level time reference, which approximates the theoretical lower bound, effectively curbs the phenomenon that tiny time errors are linearly amplified into macroscopic distance deviations under high-speed sound conditions when multiplied with the high-speed adaptive velocity matrix in step S53. This ensures that the dispersion-corrected distance difference sets calculated for different frequency bands are numerically and accurately anchored to the true physical center of gravity of the sound wave energy. More importantly, this high-fidelity time delay data, which is resistant to dispersion distortion, greatly reduces the systematic dispersion between the observations of each mode, providing a high-quality input with low variance for the multimodal weighted fusion in step S6. This significantly narrows the confidence interval of the final fusion solution, thereby compressing the ambiguity range of leak point location from the traditional meter level to the centimeter level, directly improving the hit rate of engineering excavation and greatly reducing unnecessary damage to municipal roads.

[0051] In step S6, based on the energy proportion of the intrinsic mode component set, a multimodal weighted fusion localization solution is performed on the dispersion-corrected distance difference set to obtain the final leak location result. That is, due to environmental background noise, electromagnetic interference, and the structural response of the pipeline network itself, the energy distribution of the acoustic signal generated by the leak in the frequency domain exhibits extreme non-uniformity. In the previous steps, although independent dispersion-corrected distance differences have been calculated for each mode component, these results are numerically discrete due to the differences in signal-to-noise ratio across frequency bands. Low-energy frequency bands may be dominated by random noise, and the calculated distance differences often contain huge random errors, while high-energy main frequency bands carry the true leak source location information. If the distance differences of all frequency bands are simply averaged without distinction, the errors of low-quality frequency bands will severely dilute the accuracy of high-quality frequency bands, causing the final location result to deviate from the true physical leak point. Therefore, the multimodal weighted fusion positioning solution based on energy ratio in this application aims to introduce physical energy as an objective benchmark for measuring the reliability of the solution results. By giving higher decision weights to high signal-to-noise ratio modes, the influence of noise interference and spurious modes is automatically suppressed, and the discrete frequency level results are converged into a statistically optimal physical distance difference, thereby ensuring that the final positioning coordinates are derived from the most realistic physical features of the signal.

[0052] In one embodiment of this application, step S6 includes: S61, performing energy proportion evaluation and normalization processing on the intrinsic mode component set to obtain a weighted coefficient vector; S62, performing weighted decision fusion of the weighted coefficient vector and the dispersion correction distance difference set on multi-band distance difference to obtain a fused distance difference; S63, combining the spatial constraint vector and the reference node information in the aligned dataset, performing geometric inversion and absolute coordinate mapping on the fused distance difference to obtain the final leak location result.

[0053] The specific processing is as follows: Step S61 first initiates a traversal loop, visiting each mode component k in the intrinsic mode component set one by one. In actual processing, to quantify the signal strength of each mode, the algorithm uses root mean square (RMS) energy or peak factor as evaluation metrics. Considering that leakage signals usually exhibit continuous broadband noise, RMS energy better reflects its average power level within the time window. The discretized energy integral formula is applied. ,in The number of sampling points, aligned with step S1, for example, 10240 points. This represents the amplitude value at the nth sampling point of the kth mode. This calculation yields a raw energy vector containing five energy scalars. Taking a metal pipe leak scenario as an example, if the calculated original energy value is... ={150,400,1800,300,50} (unit: square of relative voltage amplitude). It can be observed that the energy of the third mode (corresponding to around 1418Hz) is as high as 1800, several times that of the other modes. Physically, this usually means that this frequency band is the main energy concentration area of ​​fluid jet noise, with the highest signal-to-noise ratio and the most reliable phase information. The fifth mode, however, has only 50, which is highly likely to be high-frequency electromagnetic interference or quantization noise, and its calculated time delay value is often unreliable. To prevent these low-energy noise components from interfering with subsequent decisions, the algorithm introduces a nonlinear threshold filtering mechanism. First, the total energy is calculated... =150+400+1800+300+50=2700. The threshold ratio is set to 5%, which is 135. The logic unit checks each... :like <0.05× If the value is less than 135, its corresponding weight is forcibly set to zero. In the example above, the energy of the 5th mode (50 < 135) is determined to be invalid, and its corresponding weight is set to zero. It is directly locked to 0. This operation, at the algorithm level, is equivalent to building a noise firewall, automatically eliminating the negative impact of weak signals on positioning accuracy. After threshold cleaning, the energy of the remaining effective modes is normalized to generate probabilistic weights. The normalization formula is... For the effective energies {150, 400, 1800, 300} (item 5 has been removed), the total effective energy is 2650. Calculate the weights of each component: =150 / 2650≈0.0566; =400 / 2650≈0.1509; =1800 / 2650≈0.6792; =300 / 2650≈0.1132; =0, resulting in the final weighted coefficient vector. ={0.0566,0.1509,0.6792,0.1132,0}. This vector profoundly reflects the physical logic that the greater the energy, the greater the voice. The third mode dominates the subsequent distance determination with a weight of nearly 68%.

[0054] Step S62: Review the dispersion-corrected distance difference set obtained in step S5. ={64.862,65.210,65.721,66.230,66.742} meters. The difference between these values ​​(maximum approximately 1.9 meters) is caused by dispersion effects and noise interference in different frequency bands. If a simple arithmetic mean is used (i.e., all weights are 0.2), the average distance difference is approximately 65.753 meters. However, by introducing a weighted coefficient vector, the algorithm performs a dot product operation: Substituting the values ​​into a detailed derivation: 0.0566×64.862+0.1509×65.210+0.6792×65.721+0.1132×66.230+0×66.742=65.645 meters. That is, the weighted fusion result of 65.645 meters is very close to the result of the main energy mode (65.721 meters), effectively utilizing the information from modes 2 and 4 for fine-tuning. Simultaneously, the influence of mode 1 (highly affected by low-frequency vibration interference) is suppressed by low weighting, and the deviation influence of mode 5 (highly affected by electromagnetic interference) with energy below the threshold is completely shielded. This value is called the fusion distance difference, which statistically represents the unbiased estimator with the smallest variance. Physically, it represents the most reliable physical path difference between the two sensors and the leak point after considering the differences in propagation quality across frequency bands.

[0055] Step S63: Retrieve the spatial constraint vector obtained in step S3, where the core parameter is the corrected physical pipeline path length. For example, 150 meters in the S3 example, and also obtain the reference node information in the aligned dataset (such as node i as the coordinate origin). First, the algorithm performs algebraic calculations based on a one-dimensional pipe network location model. This model is based on the physical assumption that sound waves propagate along the pipe axis. Suppose that the leak point P is located between node i and node j, and the physical distance from node i is... The physical distance from node j is According to geometric relationships Meanwhile, the fusion distance difference calculated in step S5 Physically, it is defined as the path difference of the sound wave reaching the two sensors at both ends, i.e. (It is agreed here that if the leak point is close to node j, the signal will reach j first.) The value is positive or adaptive based on the sign of the time delay; if it is Di-Dj, the formula is adjusted accordingly. Solving the two linear equations simultaneously eliminates the unknown variables. The absolute physical distance of the leak point relative to the reference node i can be directly calculated: The result calculated by S62 =65.645 meters, and Substituting 150 meters into the formula: =1 / 2×(150+65.645)=107.8225 meters. This means that the leak point is located approximately 107.82 meters along the pipe from node i. After completing the algebraic solution, the absolute coordinate mapping stage begins. This is not a simple straight-line interpolation because the pipe may be curved. Using the existing GIS path data in S3, a path traversal algorithm is executed. For example, if the pipe path consists of three segments: Segment 1 (starting from node i, heading east): length 80 meters, endpoint... Section 2 (by) Turning North): Length 2 meters (bend), End Point Section 3 (by) (Continuing east): Length 68 meters, endpoint node j. The algorithm begins virtual walking: 1. Compare the target distance 107.82 meters with the first segment length 80 meters. Since 107.82 > 80, the leak is not in the first segment, and the remaining distance is 107.82 - 80 = 27.82 meters. 2. Enter the second segment, length 2 meters. Since 27.82 > 2, the leak is not at the bend either, and the remaining distance is 27.82 - 2 = 25.82 meters. 3. Enter the third segment, length 68 meters. Since 25.82 < 68, the leak is located on the third segment of the pipe, a distance from the starting point of this segment. The location is 25.82 meters in the direction of the leak. Based on the starting coordinates and direction vector of the third segment, the precise latitude and longitude of this point are calculated using geodetic formulas. For example, the final coordinates are calculated to be (latitude 31.20610°, longitude 121.57685°). Furthermore, the algorithm combines the correlation peaks (maximum cross-correlation coefficients) of each mode of the weighted coefficient vector to calculate a confidence index, such as 95%. Finally, all this information is encapsulated into a final leak location result object: {Status: "Leak detected", Latitude: 31.20610, Longitude: 121.57685, Relative distance: 107.82 meters, Confidence: 0.95}. This approximates the actual situation at the accident site as closely as possible, providing repair personnel with direct and reliable navigation information.

[0056] In summary, the fire pipeline leakage identification method based on multimodal data processing, as described in this application, aims to solve the positioning error problems caused by the use of a fixed sound velocity model that ignores dispersion effects and the failure to effectively utilize geospatial constraints in existing technologies. First, by fusing acoustic waveforms and geographic coordinate data, variational mode decomposition (VMD) technology is used to decouple the complex broadband leakage signal into a series of narrowband intrinsic mode components, precisely capturing the frequency domain characteristics of the signal at a microscopic level. Based on this, GPS-based spatial topology construction technology is introduced, using the precise physical path provided by geographic coordinates as a spatial constraint vector to construct a closed-loop feedback mechanism between acoustic and geographic information. By applying spatial constraints to adaptive group velocity matching under dispersion effects, the actual group velocity of different frequency bands propagating within the pipe is dynamically calculated, thereby correcting the wave velocity differences caused by frequency variations. Finally, combined with a multimodal weighted fusion strategy, the corrected distance differences of each frequency band are integrated to obtain accurate positioning results, achieving effective compensation and correction of dispersion errors in non-uniform media using multimodal data.

[0057] Figure 4 This is a block diagram of a fire pipeline leakage identification system based on multimodal data processing according to an embodiment of this application. Figure 4As shown, the fire pipeline leakage identification system 100 based on multimodal data processing according to an embodiment of this application includes: a sensor data stream processing module 110, used to perform temporal domain alignment and spatial label binding on the raw sensor data streams collected from each node of the fire pipeline network to generate an aligned dataset containing synchronous acoustic waveforms and corresponding geographic coordinate data; an intrinsic modal component generation module 120, used to perform VMD-based signal frequency domain hierarchical decomposition on the synchronous acoustic waveforms in the aligned dataset to obtain an intrinsic modal component set; and a spatial constraint module 130, used to perform GPS-based constraint on the geographic coordinate data in the aligned dataset. The pipeline spatial topology is constructed to obtain the spatial constraint vector; the adaptive group velocity matching module 140 is used to perform adaptive group velocity matching under the dispersion effect on the intrinsic mode component set and the spatial constraint vector to obtain the adaptive velocity matrix; the dispersion-corrected distance difference generation module 150 is used to perform hierarchical cross-correlation calculation and distance transformation on the intrinsic mode component set based on the adaptive velocity matrix to obtain the dispersion-corrected distance difference set for different frequency bands; the positioning result generation module 160 is used to perform multi-modal weighted fusion positioning calculation on the dispersion-corrected distance difference set according to the energy ratio of the intrinsic mode component set to obtain the final leak location result.

[0058] Here, those skilled in the art will understand that the specific operations of each step in the fire pipeline leakage identification system based on multimodal data processing have been referenced above. Figures 1 to 3 The description of the fire pipeline leakage identification method based on multimodal data processing is detailed here, and therefore, its repeated description will be omitted.

Claims

1. A method for identifying leaks in fire-fighting pipelines based on multimodal data processing, characterized in that, include: S1: Perform time-domain alignment and spatial label binding on the raw sensor data streams collected from each node of the fire protection network to generate an aligned dataset containing synchronous acoustic and vibration waveforms and corresponding geographic coordinate data. S2: Perform VMD-based signal frequency domain hierarchical decomposition on the synchronous acoustic vibration waveforms in the aligned dataset to obtain the intrinsic mode component set; S3: Construct a spatial topology of the pipeline network based on GPS constraints for the geographic coordinate data in the alignment dataset to obtain a spatial constraint vector, including: parsing the latitude and longitude information of the sensor nodes from the alignment dataset and performing spherical coordinate system transformation to obtain standardized radian coordinate pairs; Spherical great circle distance measurement is performed on the standardized radian coordinate pairs to obtain the Euclidean distance on the ground surface, which represents the linear distance between nodes; topological mapping and physical path correction are performed on the Euclidean distance on the ground surface to obtain the spatial constraint vector; S4: Perform adaptive group velocity matching under dispersion effect on the intrinsic mode component set and spatial constraint vector to obtain the adaptive velocity matrix; S5: Based on the adaptive velocity matrix, perform hierarchical cross-correlation calculation and range transformation on the intrinsic mode component set to obtain the dispersion-corrected range difference set for different frequency bands; S6: Based on the energy proportion of the intrinsic mode component set, perform multimodal weighted fusion localization calculation on the dispersion correction distance difference set to obtain the final leak location result.

2. The fire pipeline leakage identification method based on multimodal data processing according to claim 1, characterized in that, Step S1 includes: PPS edge detection and gated analog-to-digital conversion are performed on the raw sensor data stream to obtain the digital raw acoustic and vibration sequence and the GPS message buffer to be parsed; Parse the standard positioning statements in the GPS message cache to be parsed, and extract the Coordinated Universal Time and latitude and longitude coordinates to obtain geographic spatiotemporal labels containing spatiotemporal attributes; Geographic spatiotemporal labels are written as metadata into the associated structure of the digitized raw acoustic and vibration sequences to obtain an aligned dataset.

3. The fire pipeline leakage identification method based on multimodal data processing according to claim 1, characterized in that, Step S2 includes: Perform Hilbert transform and spectral initialization on the synchronized acoustic waveforms in the aligned dataset to obtain an initial variational parameter set, while separating the metadata cache from the aligned dataset; The initial variational parameter set is iteratively optimized using the alternating direction multiplier method to obtain the optimized modal spectrum matrix and the final center frequency vector. The optimized modal spectrum matrix is ​​subjected to inverse Fourier transform to restore each narrowband time-domain signal, and then it is restructured with the final center frequency vector and metadata cache to obtain the intrinsic mode component set.

4. The fire pipeline leakage identification method based on multimodal data processing according to claim 1, characterized in that, Step S4 includes: Modal feature frequencies are extracted from the intrinsic modal component set to obtain the effective modal frequency vector; Dispersion model parameter matching is performed on the spatial constraint vector to obtain the dispersion characteristic parameter set; Adaptive velocity matrix is ​​obtained by performing adaptive group velocity calculation on the modal effective frequency vector and the dispersion characteristic parameter set.

5. The fire pipeline leakage identification method based on multimodal data processing according to claim 4, characterized in that, The adaptive velocity matrix is ​​obtained by adaptively calculating the group velocity from the modal effective frequency vector and the dispersion characteristic parameter set, including: performing adaptive group velocity calculation on the modal effective frequency vector and the dispersion characteristic parameter set using the following formula: ； in, For the theoretical speed of sound, The dispersion attenuation coefficient is... Let be the effective group frequency of the k-th intrinsic mode component. This refers to the loop frequency or cutoff frequency of the pipeline. The inner diameter of the pipe. For pipe wall thickness, The bulk modulus of water For elastic modulus, For frequency The group velocity of the sound wave packet propagating inside the pipe.

6. The fire pipeline leakage identification method based on multimodal data processing according to claim 1, characterized in that, Step S5 includes: The set of eigenmode components is subjected to cross-power spectrum sharpening and inverse time-domain transformation to obtain the set of generalized cross-correlation functions. Peak search and subsampling-level precision fitting are performed on the set of generalized cross-correlation functions to obtain the hierarchical time delay vector; The dispersion-corrected distance difference set is obtained by calculating the dispersion-corrected distance difference between the layered time delay vector and the adaptive velocity matrix.

7. The fire pipeline leakage identification method based on multimodal data processing according to claim 1, characterized in that, Step S5 includes: The set of eigenmode components is subjected to cross-power spectrum sharpening and inverse time-domain transformation to obtain the set of generalized cross-correlation functions. Subsampling delay estimation based on Sinc kernel function reconstruction and Newton iteration is performed on the set of generalized cross-correlation functions to obtain the hierarchical delay vector; The dispersion-corrected distance difference set is obtained by calculating the dispersion-corrected distance difference between the layered time delay vector and the adaptive velocity matrix.

8. The fire pipeline leakage identification method based on multimodal data processing according to claim 1, characterized in that, Step S6 includes: The energy proportion of the intrinsic mode component set is evaluated and normalized to obtain the weighted coefficient vector; The weighted decision fusion of the weighted coefficient vector and the dispersion-corrected distance difference set is performed to obtain the fused distance difference; By combining the spatial constraint vector and the reference node information in the aligned dataset, geometric inversion and absolute coordinate mapping are performed on the fused distance difference to obtain the final leak location result.

9. A fire pipeline leakage identification system based on multimodal data processing, characterized in that, include: The sensor data stream processing module is used to perform time-domain alignment and spatial label binding on the raw sensor data streams collected from each node of the fire protection network to generate an aligned dataset containing synchronous acoustic and vibration waveforms and corresponding geographic coordinate data. The intrinsic mode component generation module is used to perform VMD-based signal frequency domain hierarchical decomposition on the synchronous acoustic vibration waveforms in the aligned dataset to obtain the intrinsic mode component set. The spatial constraint module is used to construct a spatial topology of the pipeline network based on GPS constraints from the geographic coordinate data in the alignment dataset to obtain a spatial constraint vector. This includes: parsing the latitude and longitude information of sensor nodes from the alignment dataset and performing spherical coordinate system transformation to obtain standardized radian coordinate pairs; performing spherical great circle distance measurement on the standardized radian coordinate pairs to obtain the surface Euclidean distance representing the mapped straight-line distance between nodes; and performing topological mapping and physical path correction on the surface Euclidean distance to obtain the spatial constraint vector. The adaptive group velocity matching module is used to perform adaptive group velocity matching under dispersion effects on the intrinsic mode component set and spatial constraint vector to obtain the adaptive velocity matrix. The dispersion-corrected range difference generation module is used to perform hierarchical cross-correlation calculation and range transformation on the intrinsic mode component set based on the adaptive velocity matrix to obtain the dispersion-corrected range difference set for different frequency bands. The positioning result generation module is used to perform multimodal weighted fusion positioning calculation on the dispersion correction distance difference set based on the energy ratio of the intrinsic mode component set to obtain the final leak point positioning result.

Images