Method and system for detecting leakage of water supply network by fusing audio and electromagnetic water meter multi-parameter data

Abstract

The application discloses a kind of fusion audio and electromagnetic water meter multi-parameter data's water supply pipe network leak detection method and system.The method includes collecting the pipeline radiation noise audio of monitoring node, flow, pipeline pressure and temperature;Extract acoustic features including mel-frequency cepstral coefficient, sub-band energy ratio, frequency centroid and frequency dispersion;Temperature normalization compensation is carried out to obtain equivalent acoustic features under reference temperature;Obtain leakage confidence, noise matching degree, water power measurement abnormal index;The inverse value of leakage confidence, water power measurement abnormal index and noise matching degree is weighted fusion, and leakage event is judged.The application significantly reduces the false alarm rate through audio detection and multi-parameter fusion verification of flow and pressure, combined with temperature compensation and double-branch neural network feature extraction, realizes adaptive detection for different pipe materials, and is suitable for leakage detection and early warning of various pipeline facilities such as urban water supply pipe network.

Description

Technical Field

[0001] This invention relates to the field of water supply network leakage detection and audio signal processing technology, and in particular to a method and system for detecting water supply network leakage by integrating audio and electromagnetic water meter multi-parameter data. Background Technology

[0002] Leakage in water supply networks is a major source of water waste and operational safety hazards. The average leakage rate of urban water supply networks in my country exceeds 15%, which not only causes a large amount of water waste and significant economic losses to water supply companies, but also may lead to safety accidents such as pipe bursts, affecting residents' normal water use.

[0003] Existing leak detection methods mainly include manual listening rod inspection, correlator method, and flow zoning method, but they have the following shortcomings: Poor adaptability to different pipe materials: The acoustic characteristics of leakage vary greatly among different pipe materials (metal and non-metal). Leakage sound from metal pipes (stainless steel, ductile iron) has a significant high-frequency component and a long propagation distance, with a typical frequency range of 300–2500 Hz; leakage sound from plastic pipes (PVC, composite materials) attenuates quickly and has a lower frequency, with a typical frequency range of 40–850 Hz. Traditional single algorithms cannot adapt to pipe networks with multiple materials, resulting in low accuracy in identifying leaks from different pipe materials.

[0004] Severe environmental interference leads to a high false alarm rate: background noise such as traffic noise, water pump operation noise, and user water usage noise overlap with leakage noise in the time and frequency domains, making it difficult for traditional acoustic detection methods to effectively distinguish them. This results in a false alarm rate that is generally between 8% and 15%, especially during peak water usage periods, which seriously affects operation and maintenance efficiency.

[0005] Limitations of single physical quantity and insensitivity to minute leaks: Existing methods mostly rely on a single physical quantity—methods that rely solely on audio detection are severely affected by environmental noise; methods that rely solely on flow and pressure detection are insensitive to minute leaks (orifice diameter < 2 mm) because minute leaks hardly cause measurable flow changes and pressure drops in the early stages, making early warning impossible.

[0006] Temperature affects acoustic characteristics, and cross-seasonal comparisons fail: Temperature changes not only affect sound velocity but also cause drift in the acoustic characteristics of leaks themselves. Experiments show that for the same leak source, the first dimension coefficient of MFCC (Melbourne frequency cepstral coefficients) can differ by as much as 0.43 at 5℃ and 25℃. Because current technologies do not compensate for temperature-induced characteristic drift, winter and summer characteristic data cannot be directly compared, causing trend analysis methods such as week-to-year difference to fail across seasons, resulting in a misjudgment rate as high as 15%.

[0007] Fixed threshold judgment and lack of dynamic comparison mechanism: Existing technologies mostly use fixed thresholds to judge leakage, which cannot adapt to the normal fluctuations in the operation of the pipeline network (such as the periodic changes in water consumption with the seasons and weekdays / weekends). This results in the coexistence of problems such as missed reports when the threshold is set too high and false reports when it is set too low. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems of poor pipe material adaptability, serious environmental interference, limitation of single physical quantity, lack of consideration of temperature influence, and lack of dynamic comparison mechanism in the existing technology, and to provide a method and system for leak detection of water supply network that integrates audio and electromagnetic water meter multi-parameter data, so as to achieve leak detection and graded early warning with low false alarm rate and high sensitivity.

[0009] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for detecting leaks in water supply networks that integrates audio and electromagnetic water meter multi-parameter data, characterized in that it includes: The system collects audio data of pipeline radiation noise, flow rate, pipeline pressure, and temperature at monitoring nodes to obtain a set of audio frames. Flow rate, pressure, and temperature are collected at a set period. A short-time Fourier transform is performed on each audio frame to obtain the amplitude spectrum. Acoustic features are extracted based on the amplitude spectrum. The acoustic features include at least Mel frequency cepstral coefficients, subband energy ratio, frequency centroid, and frequency dispersion. The acoustic features are temperature-normalized and compensated based on the collected temperature to obtain the equivalent acoustic features at the reference temperature. Based on the equivalent acoustic features, the leakage confidence level is obtained through a dual-branch neural network model, and the noise matching degree is obtained by comparing it with a pre-constructed noise feature library; at the same time, hydraulic measurement anomaly indicators are calculated based on the flow rate and pressure. The leakage confidence level, the hydraulic measurement anomaly index, and the inverse value of the noise matching degree are weighted and fused according to a set weight to obtain a comprehensive leakage judgment value. When the comprehensive leakage judgment value exceeds a set threshold, it is judged as a leakage event.

[0010] Secondly, the present invention provides a water supply network leakage detection system that integrates audio and electromagnetic water meter multi-parameter data, characterized in that it includes: Multiple multi-parameter intelligent monitoring terminals, each terminal integrating a sound sensor, an electromagnetic water meter, a pressure sensor, and a temperature sensor. Each sensor is connected to a communication module via a bus. The reporting cycle of the communication module is configurable and is used to collect pipeline radiation noise audio, flow rate, pipeline pressure, and temperature of the monitoring nodes and upload them according to the reporting cycle. A cloud platform or edge computing gateway is used to receive data uploaded by each monitoring terminal and execute the method described in any of the above descriptions.

[0011] The beneficial effects of this invention are as follows: Compared with existing technologies, this invention utilizes a three-path parallel fusion architecture of "static noise comparison - dynamic longitudinal comparison - hydraulic measurement verification" to cross-verify audio detection with flow and pressure, effectively filtering environmental noise interference and significantly reducing the false alarm rate; it achieves adaptive detection for different pipe materials by finely dividing the leakage frequency range of different pipe materials; it eliminates cross-seasonal feature drift through temperature normalization compensation, ensuring the accuracy of trend analysis; it extracts deep features from audio time-spectrum maps and acoustic feature sequences through a dual-branch neural network, enabling early identification of minor leaks; it adapts to fluctuations in normal pipeline network operation through dynamic baseline, avoiding false alarms and missed alarms caused by fixed thresholds; it supports configurable reporting cycles from 1 minute to 24 hours, adapting to different power supply scenarios; and it does not limit the specific sampling frequency, ensuring compatibility with future hardware upgrades. Compared with existing technologies, this invention significantly reduces the false alarm rate, greatly improves the accuracy of leak identification for different pipe materials, and significantly reduces cross-seasonal feature differences. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0013] Figure 1 This is a data processing flowchart of the present invention.

[0014] Figure 2 This is a comparison chart of the audio frequency spectrum of different pipe materials leaking according to the present invention.

[0015] Figure 3 This is a schematic diagram of the feature drift before and after temperature compensation according to the present invention.

[0016] Figure 4 This is a schematic diagram of the 7-day cycle trend profile and three-level early warning system of the present invention.

[0017] Figure 5 This is a schematic diagram of the overall system architecture of the present invention. Detailed Implementation

[0018] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0019] like Figure 5 As shown, this invention provides a water supply network leak detection system that integrates audio and electromagnetic water meter multi-parameter data, comprising: Multiple multi-parameter intelligent monitoring terminals are integrated, each combining a high-sensitivity sound sensor, an electromagnetic water meter, a pressure sensor, and a temperature sensor. All sensors are connected to the same communication module via a RS-485 bus. The communication module has a built-in high-precision real-time clock (RTC) that adds a unique timestamp to each data record. The reporting cycle of the communication module is configurable, and can be set to different values ​​such as 1 minute, 5 minutes, 1 hour, 6 hours, and 24 hours depending on the on-site power supply mode (battery-powered, solar-powered, or AC-powered) and communication network conditions. Longer cycles (e.g., 6–24 hours) are used in battery-powered scenarios to reduce power consumption; shorter cycles (e.g., 1 minute) are used in AC-powered scenarios to achieve near real-time monitoring.

[0020] The cloud platform or edge computing gateway is used to receive data reported by each monitoring terminal and execute the following methods.

[0021] like Figures 1 to 4 As shown, this invention provides a method for detecting leaks in water supply networks that integrates audio and electromagnetic water meter multi-parameter data. The method includes the following steps: S1. Multi-parameter data acquisition and configurable reporting.

[0022] Deploy multi-parameter intelligent monitoring terminals at each monitoring node to collect data: The sound sensors are all of the same model, with a fixed hardware sampling rate, but different specifications can be selected according to the application scenario. The selected sampling rate must be no less than twice the highest effective frequency component of the leak signal being measured, and is used to collect pipeline radiated noise. The noise data acquisition frequency is configured according to the application scenario as continuous uninterrupted sampling or timed sampling (e.g., once a day).

[0023] The electromagnetic water meter collects the instantaneous flow rate Q (m³ / h) and cumulative flow rate every 5 minutes.

[0024] The pressure sensor collects the pipeline pressure P (MPa) every 5 minutes.

[0025] The temperature sensor collects the pipe wall temperature T (°C) every 5 minutes.

[0026] The audio is processed by frame-by-frame windowing using a sliding time window mechanism to obtain a set of audio frames; flow rate, pressure, and temperature are collected at set intervals.

[0027] The reported data packet includes: audio files since the last report (if the noise is continuously collected, it includes the complete audio time period; if it is collected at regular intervals, it includes audio segments for a specified time period); and time-series records of flow rate, pressure, and temperature since the last report (one record every 5 minutes, the number depending on the reporting cycle). The reported data is received by the cloud platform or edge computing gateway and then subjected to subsequent batch processing and analysis. For major leaks, the system can be configured with a pressure drop threshold (e.g., 0.1 MPa / second) to trigger an emergency report.

[0028] S2, Data Preprocessing.

[0029] Audio data preprocessing and sliding window framing.

[0030] After receiving the audio file, the platform first decodes it into a continuous PCM data stream. The sampling rate of the terminal's sound sensor is selected during deployment based on the required frequency range (e.g., if the upper limit of the frequency of interest is 2500Hz, the sampling rate will not be lower than 5000Hz). The system uses a sliding time window mechanism to process the audio in frames. Each time window has a fixed length of 2 seconds, and adjacent windows overlap by 50% (i.e., each slide is 1 second long) to ensure that no events are missed.

[0031] Within each time window, the audio signal is divided into frames and windowed: the frame length is the number of sampling points corresponding to 50 milliseconds (the specific number of points = sampling rate × 0.05); the frame shift is 50% of the frame length, i.e., 25 milliseconds; the window function is the Hanning window.

[0032] This yields the set of audio frames corresponding to the time window. The number of frames n = 80 frames (the theoretical value is 80 when the window duration is 2 seconds and the frame shift is 25 milliseconds. The actual calculation may be slightly different due to boundary rounding, but 80 frames are taken as an engineering approximation).

[0033] Each It is a windowed short-time signal vector, and subsequent feature extraction, noise comparison, and model inference are all based on this frame set. After processing a time window, the system immediately releases the original audio data of that window, retaining only the extracted feature vector for historical comparison and trend analysis.

[0034] Hydraulic data preprocessing.

[0035] Flow rate, pressure, and temperature data are collected every 5 minutes and reported centrally according to the configured reporting cycle. After receiving a batch of data, the platform extracts hydraulic characteristics as follows: Minimum nighttime traffic volume statistics. When processing the data reported the previous day, extract all traffic records from 2:00 AM to 4:00 AM (24 points in 5-minute intervals), and take the minimum value among them as the minimum nighttime traffic volume for that day. At the same time, record the pressure value corresponding to the minimum value. .

[0036] Flow-pressure ratio anomaly detection. Calculation of flow-pressure ratio. Calculate the average daily hours during the nighttime period (2:00 AM to 4:00 AM). Value, and compared with the baseline (mean + 2) for the same historical period (same week type). The baseline was calculated based on data from the past 90 days of leak-free periods in that pipe section. If there were 3 consecutive days... If the value exceeds 1.2 times the baseline, it is considered an anomaly in hydraulic measurement, and the following output is generated: Increment.

[0037] S3. Multi-dimensional acoustic feature extraction and feature set construction.

[0038] Perform a short-time Fourier transform (STFT) on each frame of the signal to obtain the amplitude spectrum. ( (The frequency is in Hz). The frequency resolution of the STFT depends on the number of sampling points corresponding to the frame length, but the effective frequency range of the final analysis is limited by the Nyquist frequency of the sampling rate (half of the sampling rate). Since this invention does not limit the specific sampling rate, the following feature calculations are all based on the actual frequency (Hz), and will be converted according to the sampling rate in actual implementation.

[0039] The following features are calculated based on the amplitude spectrum to form a standardized feature set. : (I) Calculation of Mel-frequency cepstral coefficients (MFCC). The calculation of MFCC involves the following six steps: Step 1: Pre-emphasis. For each frame of signal... ( Pre-emphasis is applied to frames (e.g., 0, ..., L-1, where L is the frame length) to compensate for high-frequency attenuation. Where 0.97 is the pre-weighting coefficient, When =0, =0.

[0040] Step 2: Windowing. Multiply the pre-emphasized signal by the Hanning window function. : Step 3: Fast Fourier Transform (FFT). Perform an FFT on the windowed signal to obtain the amplitude spectrum. k = 0, 1, ..., N-1, where N is the number of FFT points (taken as an integer power of 2 greater than L, e.g., 512). Frequency resolution. ,in Sampling rate (in Hz).

[0041] Step 4: Mel filter bank energy calculation. Design M triangular filters (M=24), with the center frequencies of each filter evenly spaced on the Mel scale. The conversion relationship between Mel frequencies and ordinary frequencies is as follows: (Positive conversion) (Reverse conversion) Frequency response of the m-th filter (m=0,1,…,M-1) Defined as: Lower limit frequency Center frequency upper limit frequency ,in .

[0042] For each FFT frequency point k, its corresponding frequency .like exist[ , If the interval is ], then ;like exist[ , If the interval is ], then Otherwise, it is 0.

[0043] The output energy of the m-th filter is: .

[0044] Step 5: Logarithmic compression. Take the natural logarithm of the energy for each filter: in, It is a very small positive number (e.g.) This prevents negative infinity from occurring when taking the logarithm.

[0045] Step 6: Discrete Cosine Transform (DCT). For Perform DCT transformation to obtain MFCC coefficients. : The final output MFCC is to (Total 13 dimensions). Often overlooked in actual use. (Representing energy), only retaining to Or keep all of them.

[0046] Calculation example (simplified numerical demonstration): Suppose that after a certain frame of signal undergoes FFT, only two frequency components remain: 300Hz, amplitude 10; 1000Hz, amplitude 5. Sampling rate 5000Hz, FFT points N=512, frequency resolution ≈9.7656Hz. Corresponding FFT index: ≈300 / 9.7656≈30.7, rounded down to 31; ≈1000 / 9.7656≈102.4, rounded down to 102. Let... 10, 5, and the rest are 0.

[0047] Design M=24 Mel filters, covering 0–2500Hz. Calculate... . The center frequency of the m=8th filter. Convert to Hz: The frequency response of this filter covers approximately 300–650 Hz. The center frequency of the m=15th filter is approximately 1230 Hz, covering approximately 950–1600 Hz.

[0048] Calculate filter energy: (Assuming 300Hz is located near the center of the filter's passband) ≈100.

[0049] ,the remaining It is approximately 0.

[0050] Take the logarithm: , ,the remaining (In actual processing, a lower limit will be set, such as -50).

[0051] right Performing DCT, and extracting the first 3 dimensions of coefficients as an example: The main contribution comes from the cosine terms at m=8 and m=15, with the specific values ​​depending on the actual calculation.

[0052] In practical applications, usually for to Use after normalization.

[0053] Leakage signal and background noise to Differences in: Based on 200 sets of experimental samples (100 sets of PVC pipe leakage signals and 100 sets of normal water supply background noise, sampling rate 5000Hz, frame length 50ms), 13-dimensional MFCC coefficients were extracted from each frame. Statistical analysis revealed that... to These six dimensions exhibit separable distribution differences between the leakage signal and the background noise, as shown in Table 1. to Table of differences between leakage signal and background noise: Intra-class distance: The average Euclidean distance between leaked samples on the 6-dimensional vector is 0.35, while that between background samples is 0.28.

[0054] Inter-class distance: The average Euclidean distance between the leak and the background is 0.85, with a minimum of 0.62 and a maximum of 1.15.

[0055] Physical cause: Leakage noise (especially from plastic pipes) has broadband turbulent noise characteristics. Its spectral envelope is relatively flat in the mid-to-low frequency range (200-850Hz), resulting in even-order coefficients of MFCC ( , , The odd coefficient is relatively high. , , The MFCC coefficient is relatively low. The spectral envelope of normal water flow sound and most environmental noise attenuates faster in the mid-to-high frequency range, and the overall MFCC coefficient is closer to zero.

[0056] Classification threshold: Based on the above 200 sets of experimental data (100 sets of PVC pipe leakage signals and 100 sets of normal water supply background noise), the following conclusions are drawn: if the Euclidean distance between the MFCC vector of the frame under test and the center of the leakage template is <0.6, it is judged as a leakage; if the distance between the vector and the center of the background template is <0.4, it is judged as background; if it is in between, it enters the subsequent model.

[0057] The statistical results prove to Effectiveness in distinguishing between leakage and background noise.

[0058] (ii) Calculation of subband energy ratio vector.

[0059] Based on the statistical analysis of 200 sets of laboratory leakage sound samples from five types of pipes (stainless steel, ductile iron, composite materials, PVC, and UPVC), the frequency distribution of leakage energy exhibits the following patterns: Stainless steel pipe leakage: The energy peak occurs in the range of 850 to 1500 Hz, and the energy in this frequency band accounts for ≥40% of the total energy.

[0060] Leakage in ductile iron pipes: The energy peak occurs in the 500-1200Hz frequency range, and the energy in this frequency range accounts for ≥35%.

[0061] PVC pipe leakage: The energy peak occurs in the range of 200-850Hz, and the energy in this frequency band accounts for ≥50%.

[0062] UPVC pipe (deep burial): The energy peak appears in the range of 25-275Hz, and the energy in this frequency band accounts for ≥60%.

[0063] Accordingly, the effective frequency range is divided into four sub-bands, with fixed frequency values ​​(in Hz) as sub-band boundaries: Sub-band 1 is 25–200 Hz, sub-band 2 is 200–850 Hz, sub-band 3 is 850–1500 Hz, and sub-band 4 is 1500–2500 Hz. It should be noted that in practical use, if the sampling rate is below 5000 Hz (Nyquist frequency below 2500 Hz), sub-band 4 will be truncated. Therefore, a sampling rate of at least 5000 Hz is recommended. If a higher sampling rate is used, frequencies above sub-band 4 can be ignored or the sub-band range can be extended, but the core algorithm of this invention remains unchanged.

[0064] like Figure 2 As shown, the left figure is the time spectrum of a stainless steel pipe leak, with its energy concentrated in the high-frequency band of 800-2000Hz, corresponding to sub-bands 3 and 4 of this invention; the right figure is the time spectrum of a PVC pipe leak, with its energy concentrated in the mid-low frequency band of 200-800Hz, corresponding to sub-band 2 of this invention. This difference verifies the rationality of the frequency boundaries (25Hz, 200Hz, 850Hz, 1500Hz, 2500Hz) for sub-band division in this invention.

[0065] For each frame, calculate the sum of squared amplitudes of all frequency points within each subband, then divide by the total sum of squared amplitudes across the entire frequency band (from 25Hz to the upper limit of sampling rate / 2, but not exceeding 2500Hz) to obtain four ratios. , , , This constitutes the subband energy ratio vector. After normalization, The sum of all elements is 1.

[0066] (III) Calculation of frequency centroid.

[0067] Calculate the weighted average of the frequencies: The summation ranges from 25Hz to the smaller of the sampling rate / 2 and 2500Hz. This value represents the center frequency location of the signal energy concentration, in Hz. Based on statistics from 200 experimental samples: When metal pipes (stainless steel, ductile iron) leak, C ≥ 950 Hz (sample mean 1050 Hz, standard deviation 100 Hz).

[0068] When plastic pipes (PVC, composite materials) leak, C≤620Hz (sample mean 510Hz, standard deviation 80Hz).

[0069] The C value of normal water flow and environmental noise is between 150 and 400 Hz.

[0070] (iv) Calculation of frequency distribution dispersion.

[0071] Calculate the weighted standard deviation of the frequency distribution: The summation range is the same as above. This value reflects the degree of energy concentration or dispersion along the frequency axis. Based on statistics from 200 experimental samples: The sound of pipe leakage (especially plastic pipes) is S≥280Hz (sample mean 320Hz, standard deviation 60Hz).

[0072] The noise of rotating machinery such as water pumps is S≤120Hz (sample mean 85Hz, standard deviation 20Hz).

[0073] The transient noise passing by the vehicle has a frequency S≥400Hz, but the duration is short (<0.5 seconds).

[0074] Comprehensive judgment rule for frequency centroid and frequency dispersion (based on experimental data): If C > 950 Hz and S > 250 Hz: it is determined to be a metal pipe leak.

[0075] If C < 620Hz and S > 280Hz: it is determined that the plastic pipe is leaking.

[0076] If S < 150Hz: it is determined to be narrowband mechanical noise and is not considered leakage.

[0077] Other situations: Proceed to the next model for further evaluation.

[0078] S4, Temperature and Pressure Compensation.

[0079] Temperature correction for the speed of sound. The theoretical formula for the change of the speed of sound in water with temperature is: The formula has an error of less than 0.1 m / s within the temperature range of 0–40℃. The equivalent sound velocity in actual pipelines also needs to consider the influence of pipe material. Based on field measurements, the correction factor for stainless steel pipes… =1.12, ductile iron pipe =1.05, composite material pipe =0.92, PVC pipe =0.85. Therefore: Temperature normalization compensation. Temperature variations cause spectral drift. To eliminate this effect, a temperature-feature mapping table is pre-constructed. The temperature-feature mapping table is calibrated experimentally: audio from the same leakage source is collected at 5℃, 10℃, 15℃, 20℃, 25℃, and 30℃, and the shift of the feature at each temperature relative to 5℃ is calculated. Linear interpolation is then used to obtain the compensation value at any temperature. See Table 2 for the temperature-feature mapping table: The Δ values ​​in the table represent the offset relative to 5℃. The compensation vector at the current temperature T is obtained through linear interpolation. The reference temperature is set to 20℃, and the actual compensation amount is: Historical feature storage and weekly trend analysis both use [this method / technology]. .

[0080] Pressure compensation for audio intensity. Pressure compensation is performed before acoustic feature extraction, compensating for the amplitude spectrum energy of the original audio frame; subsequent calculations of Mel frequency cepstral coefficients, subband energy ratios, frequency centroids, and frequency dispersion are all based on the compensated amplitude spectrum energy. According to Bernoulli's equation, the jet velocity... ∝ sound pressure level ∝20 .

[0081] Mathematical proof of the compensation principle: According to Bernoulli's equation, the ejection velocity V of the fluid at the leak orifice satisfies: ∝ in, This refers to the pressure inside the pipeline.

[0082] sound pressure It is proportional to the square of the fluid jet velocity (determined by the radiation efficiency of hydrodynamic noise): ∝ sound pressure level The definition of is: Given the median design pressure of the pipeline network (taken as 0.4 MPa), the relationship between sound pressure level and jet velocity is as follows: ∝ ∝ Audio frame energy It is proportional to the square of the sound pressure (the energy density of a sound wave is proportional to the square of the sound pressure): ∝ ∝ ∝ In actual leakage sound sources, the sound energy mainly originates from turbulent fluctuations, and its radiation efficiency is related to... Therefore, the measured audio frame energy is approximately proportional to the square of the jet velocity: ∝ ∝ Based on the above derivation, within the range of pipeline pressure variation, the audio frame energy is approximately proportional to the square root of the pipeline pressure: ∝ The pressure compensation factor was determined experimentally to be: in, The energy of the original audio frame. The median design pressure for the pipeline network is 0.4 MPa. This is the pressure currently being collected.

[0083] S5, Noise Feature Library and Three-Path Comparison Detection Noise feature library construction (static benchmark).

[0084] (I) Collect 100 samples of each of the six typical interference scenarios in advance: steady-state operation of water pump, power frequency of transformer (50Hz and its harmonics), passing of vehicles, footsteps, knocking sounds, and normal water use (toilet flushing, faucet running).

[0085] For all samples in each class, extract the 13-dimensional MFCC coefficients for each frame. ~ ) and 4-dimensional subband energy ratio ( ~ ), concatenated to form a 17-dimensional feature vector Calculate the mean of the frame-level feature vectors for all samples within each class. and standard deviation Stored in the feature library, each type of noise is represented by its mean vector. As a template.

[0086] (II) Cosine Similarity Calculation. During detection, feature vectors are extracted from the current audio frame. (17-dimensional), calculation Cosine similarity with the t-th type noise template in the feature library: in: The cosine similarity value ranges from [-1, 1]. The closer the value is to 1, the more consistent the directions of the two vectors are.

[0087] (III) Noise Matching Degree The determination is as follows: Calculate the cosine similarity between the current frame and all 6 noise templates, and take the maximum value: like If the value is ≥0.85, the current frame is considered to be dominated by noise of the corresponding category; otherwise... =0 (indicating no known noise was matched).

[0088] (iv) Calculation Example: Assuming the current frame extracts a 17-dimensional feature vector (Normalized to a mean of 0 and a standard deviation of 1): =[0.52,-0.33,0.18,0.04,-0.47,0.21,-0.11,0.06,-0.03,0.09,-0.22,0.15,-0.08,0.31,0.42,0.19,0.07] Templates for "Steady-state operation of water pumps" in the feature library (Based on statistics from 100 samples): =[0.48,-0.38,0.15,0.02,-0.51,0.19,-0.14,0.05,-0.04,0.08,-0.24,0.12,-0.10,0.28,0.45,0.17,0.05] "Vehicle passing" templates in the feature library for: =[2.11,1.52,-0.67,-0.31,0.08,-0.74,-0.52,0.33,0.21,-0.18,0.42,-0.33,0.09,-0.85,-1.12,-0.53,-0.27] calculate and Cosine similarity: First, calculate the dot product: · =0.52×0.48+(-0.33)×(-0.38)+0.18×0.15+0.04×0.02+(-0.47)×(-0.51)+0.21×0.19+(-0.11)×(-0.14)+0.06×0.05+(-0.03)×(-0.04)+0.09×0.08+(-0.22)×(-0.24)+0.15×0.12+(-0.08)×(-0.10)+0.31×0.28+0.42×0.45+0.19×0.17+0.07×0.05 After calculating and summing each item, the dot product is 1.0996.

[0089] calculate Modulus length: || ||≈1.0458.

[0090] calculate The modulus length is: ||μ_pump||≈1.0575.

[0091] Cosine similarity: 1.0996 / (1.0458×1.0575)=1.0996 / 1.1060≈0.994.

[0092] Similarly, calculate and Cosine similarity: ≈0.45.

[0093] Take the maximum value: =max(0.994,0.45,…)=0.994. Since 0.994>0.85, it is determined that the current frame is dominated by "steady-state operation of the water pump" noise. =0.994.

[0094] If the current frame contains a clean leak signal, its cosine similarity with all noisy templates will be low (e.g., the maximum similarity is only 0.32). =0.

[0095] Dynamic longitudinal comparison (comparison of historical feature values).

[0096] The system maintains daily feature records for each terminal for the past 30 days (including daily average MFCC, average subband energy ratio, frequency centroid C, and frequency dispersion S). Historical audio feature baselines are constructed for both weekdays (Monday to Friday) and weekends (Saturday and Sunday) using a 7-day cycle. The average values ​​of each feature are calculated for both weekdays (Monday to Friday) and weekends (Saturday and Sunday). and standard deviation .

[0097] Calculation steps: Daily Feature Record Maintenance: Each terminal generates a daily feature record every day (after processing the data for that day), which includes the feature averages of all valid audio frames for that day: daily average values ​​of each dimension of MFCC (13 dimensions in total), daily average values ​​of each dimension of subband energy ratio (4 dimensions in total), daily average value of frequency centroid C, and daily average value of frequency dispersion S.

[0098] Each record includes a weekday type label (Monday, Tuesday, etc.). The system stores daily characteristics (day, Sunday) and date for each terminal. It also updates records on a rolling basis, deleting older records older than 30 days.

[0099] Baseline calculation period and weekday grouping: Baselines are constructed using a 7-day cycle, specifically for weekdays (Monday to Friday) and weekends (Saturday and Sunday). If the weekday type is Monday to Friday, the "weekday baseline" is used; if it is Saturday or Sunday, the "weekend baseline" is used.

[0100] For the current detection date, filter out all records with the same weekday type as the current detection date from the daily feature records of the past 30 days. For example, if the current date is Tuesday, then filter out all Tuesday records from the past 30 days (approximately 4-5 records in total).

[0101] Mean and standard deviation calculation: Suppose there are N records of the same week type selected (e.g., N = 4 Tuesdays). For a certain feature X (e.g., the first dimension of MFCC), its record value is... , ,…, ,but: Calculate for all features (MFCC 13-dimensional, subband energy ratio 4-dimensional, C, S) separately. and This yields the baseline vector.

[0102] Special case handling: If there are fewer than three records of the same week type in the past 30 days (e.g., due to a new terminal installation or data interruption), a global baseline (all historical records, regardless of week type) or merging of adjacent week types is used (e.g., if Tuesday is insufficient, Wednesday is used to make up the difference). Standard deviation To prevent zero values, a lower limit of 0.01 is set (to avoid subsequent division by zero).

[0103] For the current frame, calculate the deviation vector: (Element-by-element) Pick The maximum value of each element is denoted as Based on experimental data, when there is no leakage... The 95th percentile is 1.8, and a threshold of 2.0 is set: if A value greater than 2.0 indicates a significant anomaly in the current frame.

[0104] Hydraulic cross-validation.

[0105] Based on the nighttime flow data during the water company's 211-day leak-free period, a system was established. Baseline. No leakage. The mean is The standard deviation is Define the baseline upper limit as... The lower limit of the baseline is Similarly, regarding pressure Establish a baseline.

[0106] Hydraulic measurement anomaly indicators Calculation rules: Extract the minimum traffic volume within a set time period in the early morning of each day as the minimum traffic volume for that night. And record the corresponding pressure value. ; Calculate the flow-pressure ratio And the moving average of R over a set number of days (e.g., 3 days) is used as one of the criteria for judgment.

[0107] If for two consecutive days >( And corresponding <( ),but Increase by 0.5; If the 3-day moving average of R exceeds the historical baseline (the ratio of Q to P at the same time point) ( 1.2 times that, then Add another 0.5.

[0108] The maximum value is 1.0.

[0109] S6, a dual-branch neural network model.

[0110] This invention employs a dual-branch hybrid neural network model to extract deep features from audio frames and output leakage confidence scores. The model utilizes both the time-frequency graph (spatial features) of the audio and inter-frame temporal features to improve detection robustness. It also includes the probability distribution of pipe type and estimation of leakage intensity.

[0111] Input parameters. The model has two input branches: Branch A (CNN branch) input: time-spectral graph matrix ,in =80 (the number of frames within each 2-second time window, with a frame shift of 25 milliseconds). =128 (STFT frequency points, corresponding to 0~2500Hz, logarithmic amplitude). This input characterizes the frequency-time distribution pattern of the leakage signal.

[0112] Branch B (LSTM branch) input: feature sequence ,in =80, =19 (13-dimensional MFCC coefficients, 4-dimensional subband energy ratio, 1-dimensional frequency centroid C, 1-dimensional frequency dispersion S).

[0113] Network structure: Branch A (CNN feature extraction network): Input layer (80×128); Convolutional layer 1: (32 3×3 convolutional kernels, stride 1, padding is same, followed by batch normalization layer and ReLU activation function). Pooling layer 1 (2×2 max pooling, step size 2); Convolutional layer 2 (64 3×3 convolutional kernels, stride 1, padding is same, followed by batch normalization layer and ReLU activation function). Pooling layer 2 (2×2 max pooling, step size 2); Convolutional layer 3 (128 3×3 convolutional kernels, stride 1, padding is same, followed by batch normalization layer and ReLU activation function). Global average pooling layer; outputs a 128-dimensional vector. .

[0114] Branch B (LSTM temporal feature extraction network): Input layer (80×19); LSTM layer 1 (128 hidden units, output sequence, dropout rate 0.3); LSTM layer 2 (64 hidden units, only returning the hidden state of the last time step, dropout rate 0.3); Output: 64-dimensional vector .

[0115] Fusion layer and output layer: Will (128 dimensions) and (64-dimensional) concatenation yields a 192-dimensional joint feature vector. ; Fully connected layer 1 (128 neurons, followed by ReLU activation function, dropout rate 0.5). Fully connected layer 2 (64 neurons, followed by ReLU activation function, dropout rate 0.5); Then, three parallel output heads are connected: Leaked confidence head: (single neuron, followed by a Sigmoid activation function), output ∈[0,1].

[0116] Pipe classification header: 5 neurons (corresponding to stainless steel, ductile iron, composite materials, PVC, and others), followed by a Softmax activation function, outputting a probability distribution. ; Leakage intensity head: single neuron, followed by a linear activation function (output range 0-1, linearly scaled after sigmoid function), output intensity (relative value of leakage aperture); Output: The model output is a triple: ( , ,intensity)=Model(S,V).

[0117] : Confidence level of leakage occurrence (0-1).

[0118] : A probability vector of length 5, corresponding to 5 types of pipe materials.

[0119] intensity: Normalized leakage intensity (0~1, mapped to physical leakage aperture), used for subsequent trend analysis.

[0120] Training methods: Dataset Construction: Positive samples were collected from leakage audio in the laboratory and on-site. A total of 200 samples were collected for each of the 5 types of pipes (each sample lasting 2 seconds), for a total of 1000 samples. The training / validation / test sets were divided in a 7:2:1 ratio. The negative samples were obtained from audio recordings of leak-free operating conditions (normal water supply, water pump operation, vehicle passage, etc.), totaling 1000 samples, which were divided into equal proportions.

[0121] Data augmentation: Time stretching (speed change ±10%), Gaussian white noise (SNR=20dB), and volume scaling (gain 0.8~1.2) are applied to the training samples to enhance the model's generalization ability.

[0122] Loss function: The total loss is the sum of the three losses: Leakage confidence binary cross-entropy loss (positive sample label 1, negative sample label 0) Pipe material classification cross-entropy loss Leakage intensity mean square error loss (calculated only for positive samples, negative samples are ignored) Weight =1.0, =0.5, =0.2 (optimized by grid search).

[0123] Optimizers and hyperparameters: Optimizer: Adam, initial learning rate 0.001, decaying by 0.9 every 10 epochs.

[0124] Batch size: 64.

[0125] Training cycles: 50 epochs (early stop method, stop if the loss does not decrease for 5 consecutive cycles).

[0126] Regularization: L2 weight decay ( The dropout rate is 0.5%.

[0127] Training results: On the independent test set (200 samples), the model achieved the following performance: Leakage detection accuracy: 94.5%; Leakage confidence level AUC: 0.97; Pipe classification accuracy: 92.3%; Mean absolute error (MAE) of intensity regression: 0.08 (normalized units); Inference time: <5 milliseconds / sample (GPU), <30 milliseconds / sample (CPU).

[0128] Threshold selection: Take A threshold of ≥0.5 is considered a leak, corresponding to a false alarm rate of 4.2% and a false negative rate of 3.8% on the test set. In the actual system, to reduce the false alarm rate, the threshold was increased to 0.7, at which point the false alarm rate dropped to 2.1% and the false negative rate to 4.3%. This threshold was selected through ROC curve analysis.

[0129] S7, Three-path weighted fusion and leakage detection Fusion weights =0.6、 =0.3、 =0.1, obtained through grid search optimization with 132 test samples (AUC=0.96). Overall leakage determination value: in, This represents the original leakage confidence level output by the dual-branch neural network model in Section 4. When A value ≥0.7 is considered a leakage event. This threshold selection resulted in a false positive rate of 2.1% and a false negative rate of 4.3% on the test set.

[0130] S8. Criteria for Leakage Expansion Trend and Differentiation between Incremental / Existing Leakage For each monitoring node, daily feature records for a set number of past days are maintained. Using a weekly cycle, historical baselines for audio features are constructed for both weekdays and weekends. The deviation vector of the current feature relative to the historical baseline of the same weekly type is calculated, and the maximum value of each element of the deviation vector is denoted as... .

[0131] In this embodiment, the criteria for determining leakage expansion are based on a three-year follow-up analysis of actual leakage cases from a water company (a total of 87 leakage points, ranging from minor leaks to pipe bursts). Specifically, by analyzing the changes in acoustic and hydraulic characteristics of each leakage point during the leakage development process, the following four quantitative conditions are extracted to determine whether the leakage is in an expanding state. In this invention, the "sensitive sub-band" corresponding to a certain pipe material type refers to the frequency sub-band with the highest leakage energy concentration according to the aforementioned leakage sound frequency distribution law of that pipe material. Specifically, based on the aforementioned leakage frequency distribution law of each pipe material, for example, the sensitive sub-band for stainless steel pipes is sub-band 3 (850~1500Hz), and the sensitive sub-band for PVC pipes is sub-band 2 (200~850Hz). Once the pipe material type is determined by the terminal through the dual-branch neural network, the corresponding sensitive sub-band is determined.

[0132] A leak is considered to be in an expanding state when any one of the following conditions is met: (a) Multiple consecutive time windows Strictly incremental; (b) The slope of the linear regression of the sensitive subband energy ratio sequence within a set time exceeds a set first threshold; (c) The year-on-year difference in leakage confidence over multiple consecutive periods is positive and the cumulative absolute value exceeds the set second threshold; (d) Perform linear regression on the time series of minimum nighttime flow and the corresponding pressure. The slope of the time series of minimum nighttime flow is positive and exceeds the set third threshold, while the slope of the time series of corresponding pressure is negative and the absolute value of both exceeds the set fourth threshold.

[0133] The values ​​of the first to fourth thresholds in the above judgment criteria are determined based on the tracking and regression analysis of 87 actual leakage cases of a water company: the first threshold is the judgment threshold for the energy ratio slope of the sensitive sub-band, which is 0.05 / day; the second threshold is the judgment threshold for the cumulative absolute value of the weekly year-on-year difference, which is 0.15; the third threshold is the judgment threshold for the minimum flow slope at night, which is +0.10m³ / h / day; and the fourth threshold is the judgment threshold for the corresponding pressure slope, which is -0.01MPa / day.

[0134] Specifically, see Table 3: Leakage Expansion Judgment Criteria: Distinguishing between incremental leakage and the expansion of existing leakage: when When the leakage confidence level jumps from below the fifth threshold to above the sixth threshold and remains above the seventh threshold for multiple time windows, while the average year-on-year difference in leakage confidence over the most recent periods is less than a set value, it is considered incremental leakage. Incremental leakage occurs when the year-on-year difference in leakage confidence is positive for multiple consecutive periods and accumulates to exceed a set value, while the slope of the sensitive sub-band energy ratio exceeds a set first threshold. When fluctuations are within the set range, it is determined that the stock leakage is expanding.

[0135] Wherein, the "fifth threshold" corresponds to The lower limit for the jump is 1.5; the "sixth threshold" corresponds to... The upper limit for the jump is 2.5; the "seventh threshold" corresponds to... The lower limit, which is consistently not lower than the specified value, is 2.0. This value is based on a statistical analysis of 87 actual leakage cases from a water company, reflecting the characteristic differences between the expansion of new leaks and existing leaks.

[0136] Leakage progression rate calculation: in , , and These are the indicator functions for the above-mentioned leakage expansion determination conditions. If the conditions of the criterion are met, then... The value is 1; if the criterion condition is not met, then... The value is 0. The formula determines the weights based on regression analysis of the aforementioned 87 cases.

[0137] The specific rules for distinguishing between incremental leakage and stock leakage are as follows: like It jumped from below 1.5 to above 2.5, and remained there for the next 10 windows. It has consistently remained above 2.0, and for the past three weeks... If the average value is less than 0.05 (i.e. there was no significant abnormality last week), it is judged as an incremental leakage (new leakage).

[0138] like For three consecutive weeks, the value has been positive and the cumulative value is >0.15. >0.05 / day If the overall trend is upward and the fluctuation range within a single week does not exceed 0.5, it is determined that the leakage of existing stock is expanding.

[0139] S9, Multi-parameter trend warning (with a 7-day cycle).

[0140] like Figure 4 The image shown illustrates a 7-day trend profile and a three-level early warning system. The horizontal axis represents dates (4 consecutive weeks), and the vertical axis represents the overall leakage confidence level. and leakage progress rate The diagram marks the trigger points for Level 3, Level 2, and Level 1 warnings, and distinguishes between incremental leakage and the expansion of existing leakage. This diagram visually illustrates the complete process of this invention from leakage detection to trend warning.

[0141] Daily Feature Card Generation: The system generates a daily feature card at 0:00 every day (or after processing the day's data), which includes: Average leakage confidence level on that day (Average of all time windows); Maximum leakage confidence level of the day ; Average daily leakage intensity (The average frame energy after pressure compensation is normalized to 0-1). Percentage of the day's leakage duration ( (Number of time windows ≥ 0.7 / Total number of time windows throughout the day) The type of leak and pipe material determined on the same day (by majority vote). Minimum flow rate during the night and corresponding pressure ; Daily average flow pressure ratio ; Average temperature of the day .

[0142] Weekly trend profile generation: Every Monday (or the first processing day of the week), the system compiles statistics on the 7 daily feature cards for that week: This week's leak confidence mean sequence, intensity mean sequence, and duration percentage sequence; Year-on-year difference of each series relative to last week ; For the 7 days of this week Perform linear regression to obtain the slope ; Leakage Progress Rate Same as the calculation formula above.

[0143] Table 4. Three-level rolling early warning mechanism: Once the warning is triggered, the system pushes a message to the maintenance personnel through the cloud platform. The message includes: warning level, estimated leak location (location coordinates and confidence radius), estimated pipe type, leak intensity, leak progress rate and trend curve, and indicates whether the event is an incremental leak or an expansion of an existing leak.

[0144] S10, Model closed-loop self-learning.

[0145] After maintenance feedback data (actual leak diameter, location, and pipe material) is entered into the system, the following updates are triggered: Add the feature vectors of the false alarm frames to the corresponding categories in the noise feature library, and recalculate the mean of that category. .

[0146] Missed frames (weak leaks) are added to the training set, and the dual-branch neural network is incrementally trained every quarter (learning rate 0.001, training for 5 epochs).

[0147] The temperature-feature mapping table (Table 2) is recalibrated every six months based on newly accumulated cross-seasonal data.

[0148] Refit the pressure compensation coefficient (using the least squares method).

[0149] In summary, compared with the prior art, the present invention has the following beneficial effects: Multi-material adaptive: By refining the sub-band division of leakage frequency range for different pipe materials and classifying the output of pipe material type, the accuracy of identifying leakage of different pipe materials is 92.3% (test set of 132 samples), which is 37% higher than the traditional single model method.

[0150] Multi-parameter fusion reduces false alarms: Audio detection is cross-validated with traffic and stress, resulting in a false alarm rate of 2.1%, lower than the 8.4% of single audio detection.

[0151] Temperature compensation eliminates seasonal drift: After temperature normalization, the cross-seasonal MFCC characteristic difference decreased from 0.43 to 0.05, and the week-on-year difference misclassification rate decreased from 15% to 2%.

[0152] Clear criteria for determining the extent of leakage: The median early warning time for four quantitative indicators in 87 real-world cases was 5.2 days.

[0153] Specific application examples: Example 1: Low power consumption scenario (reported once a day, noise collected once a day at set times).

[0154] In a rural water supply network, the terminals are battery-powered and configured to report data once a day (data from the previous 24 hours is reported at 4:00 AM). Noise data collection is also configured to be performed once a day, with each collection session lasting 2 hours (0:00 AM - 2:00 AM). Flow rate, pressure, and temperature are collected every 5 minutes and also reported daily.

[0155] After receiving data at 4:05 AM each day, the system performs a sliding window analysis on two hours of audio (window length 2 seconds, 50% overlap). It was found that the energy ratio of the 200-850Hz subband of terminal #08 increased slowly over several consecutive days: 8% on day 1, 11% on day 3, and 15% on day 7. The week-on-week year-on-year difference was also calculated. =+0.12 (compared to the same day last week).

[0156] Hydraulic data: Minimum flow rate at night The flow rate increased from the baseline of 1.1 m³ / h to 1.6 m³ / h (deviation 45%), and the pressure decreased from 0.43 MPa to 0.39 MPa (decrease 9%). The abnormal hydraulic measurement conditions were met for two consecutive days. =1.0. Noise matching degree =0.1 (quiet nighttime environment). Output of the dual-branch neural network. =0.85. After fusion =0.6×0.85+0.3×1.0+0.1×0.9=0.90, which exceeds 0.7, triggering a leak detection.

[0157] Because this is a stock leak (which has existed for several days) and the subband energy ratio rises at a steeper slope... =0.09 / day > 0.05, the nighttime flow / pressure combination trend also meets the criteria, leakage progress rate calculation: indicator Satisfy (0.30)+ indicator The condition (0.20) = 0.50 is met, thus reaching the level 2 warning threshold.

[0158] The system notification reads: "There is a continuous leak in the PVC pipe at terminal #08, and it is expanding (Level 2 warning). It is recommended to manually listen for verification within 3 days." After the maintenance personnel arrived, they confirmed that there was a minor crack in the socket of the PVC branch pipe. After repair, the flow rate returned to normal.

[0159] Example 2: Mains power scenario (reporting once per minute, noise continuously collected).

[0160] In a city's main pipeline network, the terminals are powered by municipal electricity and configured to report data every minute, with continuous noise acquisition. The system processes audio streams in near real-time.

[0161] At 9:23:15 a.m. on a certain day, terminal #12 detected a sudden increase in the subband energy ratio from 850 to 2500 Hz from 8% to 45% within its audio window (2 seconds). The dual-branch neural network output... =0.96. Meanwhile, the current feature deviates from the baseline modulus. It jumped from 1.2 to 4.5.

[0162] Because this was a sudden, incremental leak (during the same period last week). All are below 0.3), and the week-on-week year-on-year difference has not yet turned significantly positive (because there was no such event last week), but the indicators... ( (Incrementing through 3 consecutive windows) to quickly satisfy: Window 1 4.5, Window 2 4.8, Window 3 5.1.

[0163] Hydraulic data: Pressure data reported every 1 minute showed that around 9:23, the pressure suddenly dropped from 0.52 MPa to 0.43 MPa (a decrease of 17%), which is consistent with the audio detection. The pressure drop was assigned a value of 1.0 (even though the nighttime flow method did not meet the requirement of two consecutive days, the pressure drop was strong evidence).

[0164] After fusion =0.6×0.96+0.3×1.0+0.1×0.9=0.576+0.30+0.09=0.966, triggering a leak event determination. A Level 1 warning is triggered. The system immediately pushes: "New incremental leak, Level 1 warning, please check immediately. Pipe type: Stainless steel (confidence level 92%)." Repair personnel arrive at the scene, find the pipeline ruptured during construction, and promptly shut off the valves for repair.

[0165] Example 3: Hybrid configuration (continuous noise acquisition + daily hydraulic reporting).

[0166] The terminal is configured for continuous noise acquisition, with audio files uploaded every 6 hours; hydraulic data is acquired every 5 minutes but reported once a day (or reported every 6 hours along with the audio).

[0167] A minor leak occurred at 1:30 AM one day. Audio was continuously collected, and an audio file containing the audio from that time period was uploaded 6 hours later (at 7:30 AM). During platform processing, the sliding window detected the leak around 1:30 AM. =0.78, and the subband energy ratio increased from 5% to 14%.

[0168] Hydraulic data has not yet been reported for the day (it is expected that the data for the entire day yesterday will be reported at 4:00 AM, but it may not have been received by the time the audio is processed), therefore The current value is 0. However, the system adjusts the value based on audio characteristics and week-to-week difference. (Compared to the same time last week) It is classified as a Level III warning.

[0169] After receiving the hydraulic data at 4:05 AM the following day, it was discovered that the previous night's... A 22% increase from baseline and a 6% decrease in pressure meet the criteria for hydraulic measurement anomalies. Updated to 1.0, recalculated. The reading rose to 0.85, escalating to a Level 2 alert. The system sent a supplementary notification: "Hydraulic data confirms leak; please arrange for verification." After locating the leak, maintenance personnel confirmed it was a loose PE pipe fitting, tightened it, and repaired the problem.

[0170] Example 4: Verification of temperature compensation effect.

[0171] like Figure 3 As shown, the left figure shows the drift of the winter feature vector relative to the summer baseline without compensation. Due to the acoustic feature shift caused by temperature changes, the winter features deviate significantly from the summer baseline. Direct week-to-week difference will result in misjudgment. The right figure shows the feature distribution after temperature normalization compensation according to the present invention. The winter features basically overlap with the summer baseline, and the cross-seasonal feature difference is reduced from 0.43 to below 0.05, proving the effectiveness of the temperature compensation method.

[0172] Cross-seasonal testing was conducted at the same PVC pipe leak point (2mm orifice): Summer (water temperature 25℃, ambient temperature 28℃): 1 hour of leakage audio was collected, and MFCC features were extracted. Using the 20℃ baseline features as a reference, the original MFCC coefficients for the first dimension were 2.35, and for the second dimension, they were -1.22. In winter (water temperature 5℃, ambient temperature 3℃): at the same leak point and location, leak audio recordings of the same duration were collected. Without compensation, the original MFCC first-dimensional coefficient decreased to 1.92 (a decrease of 0.43), and the second-dimensional coefficient increased to -0.85 (an increase of 0.37). Directly using the original features for week-to-week year-to-year difference analysis would incorrectly determine a reduced leak. (Negative value).

[0173] Apply the temperature normalization module: query the temperature-feature mapping table (Table 2), and the compensation vector at 25℃. =(+0.08,-0.05,…), compensation vector at 5℃ =(-0.35,+0.28,…). After compensation, the equivalent summer characteristics are... =(2.27,-1.17,…), Winter equivalent characteristics =(2.27,-1.13,…), the difference between the two at a 20℃ baseline is less than 0.05, with an overlap of over 95%. Weekly year-on-year difference. It accurately reflects the true state of the leak (no expansion), avoiding misjudgment.

[0174] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting leaks in water supply networks that integrates audio and electromagnetic water meter multi-parameter data, characterized in that, include: The system collects audio data of pipeline radiated noise, flow rate, pipeline pressure, and temperature at monitoring nodes to obtain a set of audio frames. Flow rate, pressure, and temperature are collected at a set period. A short-time Fourier transform is performed on each audio frame to obtain the amplitude spectrum. Acoustic features are extracted based on the amplitude spectrum. The acoustic features include at least Mel frequency cepstral coefficients, subband energy ratio, frequency centroid, and frequency dispersion. The acoustic features are temperature-normalized and compensated based on the collected temperature to obtain the equivalent acoustic features at the reference temperature. Based on the equivalent acoustic features, the leakage confidence level is obtained through a dual-branch neural network model, and the noise matching degree is obtained by comparing it with a pre-constructed noise feature library; at the same time, hydraulic measurement anomaly indicators are calculated based on the flow rate and pressure. The leakage confidence level, the hydraulic measurement anomaly index, and the inverse value of the noise matching degree are weighted and fused according to a set weight to obtain a comprehensive leakage judgment value. When the comprehensive leakage judgment value exceeds a set threshold, it is judged as a leakage event.

2. The method according to claim 1, characterized in that, The temperature normalization compensation includes: pre-constructing a temperature-feature mapping table, which records the offset of each acoustic feature relative to 5℃ at different temperatures; obtaining a compensation vector through linear interpolation based on the currently acquired temperature; and compensating the extracted acoustic features to the equivalent features at a reference temperature, which is set to 20℃.

3. The method according to claim 1, characterized in that, In the calculation of the sub-band energy ratio, the effective frequency range is divided into 4 sub-bands according to the frequency boundaries of 25Hz, 200Hz, 850Hz, 1500Hz and 2500Hz, and the ratio of energy in each sub-band to the total energy is calculated.

4. The method according to claim 1, characterized in that, This also includes stress compensation for audio frame energy: calculating the compensated audio frame energy based on the acquired stress. in, For the energy of the original audio frame, The median design pressure for the pipeline network. This is the pressure currently being collected.

5. The method according to claim 1, characterized in that, The method for calculating the hydraulic pressure measurement anomaly index is as follows: extract the minimum flow rate within a set time period in the early morning of each day as the minimum flow rate for that night, and record the corresponding pressure value; calculate the flow-pressure ratio. The flow rate pressure ratio This is the ratio of the minimum flow rate at night to the square root of the corresponding pressure, and is expressed as... The set number of days moving average is used as one of the judgment criteria; if the minimum nighttime flow exceeds the upper limit of the historical flow baseline and the corresponding pressure is lower than the lower limit of the historical pressure baseline for several consecutive days, the value of the hydraulic measurement anomaly index is increased by 0.5; if the moving average exceeds a certain multiple of its historical baseline for the same period, the accumulated value of the hydraulic measurement anomaly index is increased by 0.5 again, and the upper limit of the value of the hydraulic measurement anomaly index is 1.

0.

6. The method according to claim 1, characterized in that, It also includes dynamic longitudinal comparison and leakage expansion trend criteria: For each monitoring node, daily feature records for a set number of past days are maintained. Using a weekly cycle, different historical baselines for audio features are constructed for different weekday types (weekdays and weekends). The deviation vector of the current feature relative to the historical baseline of the same weekday type is calculated, and the maximum value of each element of the deviation vector is recorded as . ; A leak is considered to be in an expanding state when any one of the following conditions is met: Multiple consecutive time windows Strictly incremental; The slope of the linear regression of the sensitive subband energy ratio within a set time period exceeds a set first threshold. The year-on-year difference in leakage confidence over multiple consecutive periods is positive and the cumulative absolute value exceeds the set second threshold; Linear regression was performed on the time series of minimum nighttime flow and the corresponding pressure. The slope of the time series of minimum nighttime flow was positive and exceeded the set third threshold, while the slope of the time series of corresponding pressure was negative and both exceeded the set fourth threshold.

7. The method according to claim 6, characterized in that, This also includes distinguishing between incremental leakage and the expansion of existing leakage: When the When a leak jumps from below the fifth threshold to above the sixth threshold and remains above the seventh threshold for multiple time windows, while the average year-on-year difference in leakage confidence over the most recent periods is less than a set value, it is considered an incremental leak. When the leakage confidence year-on-year difference is positive for multiple consecutive periods and accumulates to exceed a set value, and the slope of the sensitive sub-band energy ratio exceeds a set first threshold, but the When fluctuations are within the set range, it is determined that the stock leakage is expanding.

8. The method according to claim 1, characterized in that, The rules for determining the frequency centroid and frequency dispersion are as follows: When the center of gravity of the frequency is greater than 950Hz and the frequency dispersion is greater than 250Hz, it is determined to be a metal pipe leak; when the center of gravity of the frequency is less than 620Hz and the frequency dispersion is greater than 280Hz, it is determined to be a plastic pipe leak; when the frequency dispersion is less than 150Hz, it is determined to be narrowband mechanical noise and is not considered a leak.

9. The method according to claim 1, characterized in that, It also includes the model's closed-loop self-learning step: After obtaining maintenance feedback data, the feature vector of the false alarm frame is added to the corresponding category of the noise feature library and the mean value of the category is recalculated; The feature vectors of the missed frames are added to the training dataset, and the dual-branch neural network model is incrementally trained to update the model parameters. The temperature-feature mapping table was recalibrated based on the newly accumulated cross-seasonal data; Refit the pressure compensation coefficient.

10. A water supply network leakage detection system integrating audio and electromagnetic water meter multi-parameter data, characterized in that, include: Multiple multi-parameter intelligent monitoring terminals, each terminal integrating a sound sensor, an electromagnetic water meter, a pressure sensor, and a temperature sensor. Each sensor is connected to a communication module via a bus. The reporting cycle of the communication module is configurable and is used to collect pipeline radiation noise audio, flow rate, pipeline pressure, and temperature of the monitoring nodes and upload them according to the reporting cycle. A cloud platform or edge computing gateway is used to receive data uploaded by each monitoring terminal and to execute the method as described in any one of claims 1 to 9.

Images