Leakage detection method based on multi-sensor time shift correction and deep learning

Abstract

The invention belongs to the technical field of leakage detection, and particularly relates to a leakage detection method based on multi-sensor time shift correction and deep learning. A plurality of sensors sensitive to leakage sound waves or vibration are arranged at different positions of a monitored pipeline or container and used for collecting leakage related signals; time alignment is carried out on the time sequence signals collected by the multiple sensors according to the position difference of the time sequence signals in the space; comparing and differentiating the plurality of sensor signals subjected to time alignment; modeling and updating environmental noise in real time, dynamically adjusting detection parameters according to noise level, continuously monitoring statistical characteristics of background signals, and establishing a noise model; a pre-trained deep learning model is used to carry out mode identification and classification determination on the signal after noise reduction processing; and the alarm and display module is used for triggering alarm and recording leakage information when suspected leakage is detected. According to the invention, weak leakage signals can be reliably detected under the condition of strong background noise.

Description

Technical Field

[0001] The present invention belongs to the technical field of leakage detection, and in particular relates to a leakage detection method based on multi-sensor time-shift correction and deep learning. Background Art

[0002] In nuclear power plants, petrochemicals, and other fields, leak detection is crucial for ensuring safety and minimizing losses. However, high-noise on-site environments pose significant challenges to leak detection: background noise generated by normally operating equipment often drowns out leak sounds, making it difficult for traditional detection methods to detect even small leaks in a timely manner. Existing acoustic leak detection technologies primarily fall into the following categories: Manual listening and simple sensor monitoring: For example, listening rods, simple microphones, or liquid microphones are used to listen for leaks at locations such as valves and fire hydrants. This method has been used since the 19th century, but it is not sensitive to small leaks, and weak leak signals are easily masked by ambient noise, making it unreliable. Operators often need to double-check to avoid misinterpreting ambient noise as leaks.

[0003] Noise recording and statistical analysis: Noise recorders can record pipeline noise over long periods of time and use statistical analysis of changes in noise intensity to determine the presence of leaks. However, in environments with strong interference, this method is susceptible to non-leakage factors, resulting in a high false alarm rate and requiring a long monitoring period to extract signal change characteristics.

[0004] Dual-sensor correlation detection: Leak noise correlators deploy sensors on both sides of the pipeline and use the time difference between the propagation time of the leak sound at the two measurement points to perform correlation calculations to determine the leak's location. This method is effective for locating medium-sized and larger leaks, but in complex, high-noise environments, the correlation peak can be overwhelmed or offset by the noise, leading to missed detections or misjudgments. Furthermore, correlation methods typically require manual threshold setting and are difficult to adapt to non-stationary noise interference.

[0005] Single-sensor signal processing and machine learning: With the advancement of computing technology, research has emerged on the use of signal processing and pattern recognition to automatically detect leaks. For example, literature reports on the use of empirical mode decomposition to extract pipeline vibration signal features, combined with a one-dimensional convolutional neural network to detect water pipe leaks. Another example is the use of time-frequency features obtained through short-time Fourier transforms or wavelet transforms to input deep learning models to improve the ability to recognize complex leak patterns. These methods have improved detection accuracy to a certain extent. However, traditional signal processing methods are often still sensitive to noise, difficult to adapt to non-stationary signals, and computationally expensive, making them difficult to implement in real-time monitoring. Relying solely on data from a single sensor also lacks spatial redundancy, and reliability decreases if the sensor signal is interfered with by occasional noise.

[0006] In summary, existing leak detection technologies in high-noise environments are either prone to missed or false positives or require complex signal processing and lack adaptability. Therefore, it is necessary to provide a new technical solution that combines multi-sensor data fusion with advanced signal processing and deep learning algorithms to achieve accurate and robust leak detection in high-noise environments. Summary of the Invention

[0007] The purpose of the present invention is to provide a leak detection method based on multi-sensor time-shift correction and deep learning. It integrates multi-sensor signal acquisition, time-shift correction (time delay alignment), differential noise reduction, adaptive noise modeling and deep learning analysis. It can reliably detect weak leakage signals under strong background noise conditions and achieve high-precision detection of weak leakage signals in facilities such as industrial pipelines.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is: A leak detection method based on multi-sensor time-shift correction and deep learning includes a multi-sensor data acquisition module: multiple sensors sensitive to leakage sound waves or vibrations are arranged at different positions of the monitored pipeline or container to collect leakage-related signals; a time-shift correction module: time-aligns the time series signals collected by multiple sensors according to their position differences in space; a differential noise reduction module: compares and performs differential processing on the multiple sensor signals that have been time-aligned; an adaptive noise modeling module: models and updates the environmental noise in real time, dynamically adjusts the detection parameters according to the noise level, continuously monitors the statistical characteristics of the background signal, and establishes a noise model; a deep learning analysis module: uses a pre-trained deep learning model to perform pattern recognition and classification judgment on the signal after noise reduction processing; and an alarm and display module: triggers an alarm and records the leakage information when a suspected leak is detected.

[0009] Multi-sensor data acquisition module: Multiple sensors include accelerometers, acoustic emission sensors or hydrophones.

[0010] Differential noise reduction module: For data from multiple sensors, the average subtraction method and principal component analysis are used to extract the noise subspace and remove the main noise components from the signal.

[0011] Adaptive noise modeling module: The statistical characteristics of the background signal include mean, variance, and spectral distribution; when the ambient noise changes, the differential noise reduction weight is adjusted or the threshold setting is updated; when the ambient noise decreases at night, the detection sensitivity is automatically increased; when the noise increases during the day, the threshold is increased to reduce false alarms.

[0012] Deep learning analysis module: uses deep neural networks, including convolutional neural networks, long short-term memory networks or their combinations as leakage discrimination models. The network input is the time domain waveform, frequency domain features or time-frequency images of multiple sensor signals after processing. Through the automatic feature extraction capability of the deep learning model, the characteristic pattern of the leakage is extracted from the data, and the judgment results are output, including the probability of leakage occurrence, leakage level or the presence or absence of leakage indication.

[0013] Time shift correction module: Knowing the positions and sound speeds of multiple sensors: Get the position of each sensor relative to the pipeline coordinates in advance. When a leak is suspected, select the signal of a reference sensor as a reference and perform time shift correction on the signals of other sensors based on the distance difference between them and the reference sensor. The time shift correction is done. According to the formula calculate, 、 For sensors and reference sensor The distance to the reference point, The propagation speed of the leakage sound in the medium is calculated by applying the corresponding time offset to align the signals of all sensors to the same time base; Correlation analysis estimates the time delay: Without the need for precise position parameters, the cross-correlation method is used to automatically estimate the time delay between the sensor pairs. and , calculate their cross-correlation function , find the moment when the correlation function peaks The most likely delay difference between the signals will be The signal relative to the sensor Signal offset Alignment: Select one sensor as the reference in turn, pair it with the remaining sensors and calculate the correlation, and obtain the estimated value of the time delay of all sensors relative to the reference. If necessary, combine the known distance method or use the redundancy of multiple sensor data to improve robustness.

[0014] The time-aligned multiple sensor signals are recorded as , For the number of sensors, calculate the difference between the two pairs of multiple sensor signals. ,calculate , differential signal eliminate and The common parts in the multi-channel signal are fused again: one method is to superimpose and sum or average all the differential signals; another method is to use principal component analysis or independent component analysis to obtain the multi-channel signal. Extract the main noise components and signal components of interest, construct the covariance matrix of multiple sensor signals, and perform eigendecomposition on them. The eigenvector corresponding to the maximum eigenvalue represents the main common component, and the eigenvector corresponding to the minimum eigenvalue represents the difference component. Noise reduction is achieved by discarding the main characteristic components and retaining the secondary characteristic components. The average of the differences of adjacent sensors is taken as the comprehensive differential signal : For each pair Perform differential and accumulation; the signal after differential fusion , apply frequency band filtering as needed to further improve the signal-to-noise ratio; if it is known that the leakage sound is concentrated in a specific frequency band, design a bandpass filter to extract the energy of this frequency band and filter out other irrelevant frequency noise; Calculate short-term energy or envelope to smooth out instantaneous spike interference. When there are a large number of sensors or they are distributed over a large area, use beamforming technology to target specific locations for gain. If the approximate propagation direction or speed range of the leakage wave is known, design a beamformer to provide high gain for the leakage signal and attenuate noise in other directions, thereby achieving noise reduction.

[0015] At the start-up or initial operation, collect the background signal in the no-leak state as the baseline sample, establish the initial noise model by analyzing this sample, and calculate the mean value of the background noise , standard deviation , power spectrum distribution If pure background samples cannot be obtained in advance, the noise model is updated during operation by detecting signal segments where no suspicious events occur for a long time. During continuous operation, the noise model is continuously updated based on new data using sliding window statistics or exponentially weighted averages. The background noise energy is estimated using: is the total signal energy currently observed, is a smoothing coefficient between 0 and 1, This recursive noise model can gradually reflect the average noise level of the current environment and track the changes in the spectrum separately - maintaining a background noise power spectrum updated over time. , which is used to identify the appearance of new noise components or the disappearance of original noise components.

[0016] After the noise model is updated, it affects the parameter settings of each module to achieve adaptive control: Differential noise reduction parameters: If the noise model indicates that the current global noise level is increasing, the noise subspace attenuation during differential fusion is increased, and the weight in the summation and averaging of the differential signals is increased to offset more common components. If the noise spectrum changes, the filter passband frequency is adjusted to match the new prominent frequency band of the leakage signal. Detection decision threshold: When the probability of the neural network output is greater than the threshold, it is judged as a leak. The threshold is dynamically adjusted according to the current noise level. When the noise is high, the threshold is increased to reduce false positives, and when the noise is low, the threshold is lowered to capture weak leaks. Deep learning model input adjustment: Significant changes in the noise spectrum trigger adjustments to the model input features, such as adding features specific to the new noise frequency band or enabling the adaptive batch normalization mechanism in the model to make the model robust to changes in mean / variance. When construction noise increases temporarily, the sensitivity is automatically lowered to avoid false alarms; when the noise decreases after construction is completed, the high sensitivity is restored to capture potential leaks.

[0017] Convert the preprocessed signal into a suitable feature input: One method is to use the time series after difference denoising As input, let the neural network extract time domain features by itself; another method is to Transform to extract frequency domain or time-frequency domain features, and then use them as network input: Calculate The short-time Fourier transform of , take its amplitude or power to get the time-frequency image, as the input image of the convolutional neural network, extract statistical features, including energy, peak frequency, spectral center of gravity in multiple frequency bands or calculate Mel frequency cepstral coefficients to obtain a fixed-length feature vector, input to the fully connected network or support vector machine classifier, select the differential signal The normalized spectrum map of is used as the input of the deep convolutional neural network model: Perform short-time Fourier transform analysis with a window length of 1 second to obtain a spectrum graph per second, and normalize the amplitude of the spectrum graph - subtract the mean and divide it by the standard deviation before inputting it into a two-dimensional convolutional neural network; use a structure combining a convolutional neural network and a fully connected layer for binary classification leakage judgment. The model includes: convolution layer + pooling layer: several convolution layers extract local feature patterns in the input spectrum graph, including energy surges of specific frequency combinations. Each convolution layer is followed by a pooling layer to reduce the data dimension and increase invariance: the first layer has a convolution kernel size of 5×5 to extract primary features; the second layer has a convolution kernel size of 3×3 to extract higher-level features; the size is reduced by 2×2 maximum pooling; fully connected layer: flatten the feature map output by the last convolution layer, connect one or two layers of fully connected neurons, and use them to combine the features extracted by convolution and perform classification judgment. Use a hidden layer with 128 neurons as the activation function ReLU and an output layer activation function Sigmoid to output the leakage probability; during training, the fully connected layer uses the Dropout strategy to randomly discard some neurons to prevent overfitting and improve the model Ability to generalize to new data; use cross entropy loss function and Adam optimizer during training, or use deep models with other architectures: use two sets of sensor signals as input or time-series-based RNN / LSTM network to capture temporal dynamics; before putting into practical application, use a large amount of sample data to train the deep learning model offline. The training data includes signals under normal working conditions and signals under various simulated leakage conditions, covering a variety of noise scenarios, and obtain samples through a combination of on-site collection and simulation: record background noise samples when there is no leakage, and create controllable small hole leakage on the experimental pipeline to obtain leakage signal samples; or use existing leak detection public data sets, combined with preprocessing methods for data enhancement, artificially superimpose different levels of noise on the leakage signal during training, or use the noise components generated by the adaptive noise modeling module to mix with the leakage signal to expand the distribution of training samples; after training, verify the model performance on independent test data, including accuracy, recall rate and false alarm rate indicators; the deep learning model outputs a probability value of the existence of a leak for each analysis window Or classification score, set decision threshold For final judgment: If , it is determined that there is a leak and an alarm is triggered; otherwise it is determined that there is no leak; under the action of the adaptive noise modeling module, Adjust according to the noise level, or use dynamic thresholds that change over time to further reduce false alarms; if multiple consecutive windows are determined to have leaks, the alarm level will be raised and corresponding safety measures will be implemented. The time and possible location of the leak incident will be recorded for reference by maintenance personnel.

[0018] The beneficial effects achieved by the present invention are: This invention combines five technical approaches—multi-sensor fusion, time-shift correction, differential noise reduction, adaptive modeling, and deep learning—for the first time, into a comprehensive solution. While existing solutions employ multi-sensor correlation or machine learning detection methods, there are no public reports of integrating all of these modules for leak detection. This innovative combination effectively avoids the limitations of individual approaches, significantly differentiating it from existing technologies.

[0019] The multi-sensor approach of this invention provides spatial information redundancy, which, combined with time-shift correction, ensures synchronous signal superposition, improving the signal-to-noise ratio from a physical perspective. Differential noise reduction further suppresses background noise interference. An adaptive noise model ensures that algorithm parameters are automatically optimized as the environment changes. Finally, a deep learning model extracts high-dimensional features and makes intelligent judgments. The layered gains in each link enable the system to reliably distinguish leakage signals from noise in a strong noise background. Laboratory simulations and field tests have shown that this method significantly improves the accuracy of leak detection in high-noise environments, increasing by several percentage points compared to a baseline solution that does not adopt this combined strategy, and significantly reducing missed detection and false alarm rates.

[0020] The adaptive modeling of this invention enables the system to automatically adjust to varying noise levels, eliminating the need for frequent manual parameter adjustments and ensuring long-term operational stability. Furthermore, the algorithms in each step of the invention have been optimized for real-time processing: time-shift correction and differential operations require minimal computational effort, the adaptive noise model utilizes an incremental update strategy with minimal overhead, and the deep learning component can run efficiently on embedded GPUs or dedicated AI chips. This makes the entire system suitable for online, real-time monitoring, ensuring that leak warnings are not missed due to computational delays.

[0021] The present invention cleverly combines traditional signal processing methods with emerging deep learning technologies, achieving superior results to existing technologies in high-noise leak detection scenarios. This cross-disciplinary combination is non-obvious: for example, the idea of multi-sensor differential denoising combined with deep neural networks is not a simple superposition of existing technologies, but a completely new approach to addressing the challenges of leak detection. While existing research has reported on combining multi-sensor data fusion or signal decomposition with neural networks, the integration of time alignment, differential denoising, and adaptive tuning was not foreseen. The overall solution of the present invention significantly exceeds the expectations of the prior art and therefore possesses outstanding creativity.

[0022] This invention is not only suitable for long-distance pipeline leak detection, but can also be extended to scenarios requiring leak detection in noisy environments, such as storage tank leaks and building water supply and drainage network leaks. By adjusting the sensor type and deep learning model structure, the system can be customized for different media (gas, liquid) and leaks of varying sizes, demonstrating its high practical value.

[0023] In summary, the present invention has significant innovation compared to the existing technology and can effectively improve the accuracy and reliability of leakage detection in high-noise environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a leak detection method based on multi-sensor time-shift correction and deep learning; Figure 2 This is an architecture diagram of a leak detection system based on multi-sensor time-shift correction and deep learning; In the figure: 1. Device under test; 2. Sensor. DETAILED DESCRIPTION

[0025] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] like Figure 1 、 2 As shown in the figure, the leak detection method based on multi-sensor time shift correction and deep learning includes: Multi-sensor data acquisition module: Multiple sensors are deployed at different locations along the monitored pipeline or vessel to collect leak-related signals. These sensors are acoustic sensors sensitive to leak sound waves or vibrations, such as accelerometers, acoustic emission sensors, or hydrophones. They are distributed at key points within the leak monitoring area. This multi-sensor deployment provides spatial information redundancy, capturing the propagation of leak sound waves at different locations and providing foundational data for subsequent processing.

[0027] The time-shift correction module aligns the time-series signals collected by each sensor based on their spatial position differences. Since leak acoustic waves experience different propagation delays when reaching different sensors, the signals need to be synchronized and corrected based on the distance difference between each sensor and the leak source or by estimating the time delay through signal correlation. This time-shift correction ensures that the signals from multiple sensors are aligned for the same leak event, facilitating subsequent fusion analysis.

[0028] Differential Noise Reduction Module: This module compares and performs differential processing on time-aligned multi-sensor signals to mitigate the impact of background noise. Specifically, it calculates the differential signal between each sensor signal to eliminate common noise components. Because environmental noise, such as fluid noise in a pipeline or external mechanical noise, is typically spatially coherent or evenly distributed, each sensor receives similar noise components. Leakage signals, on the other hand, exhibit localized characteristics, with intensity varying significantly with sensor position. Therefore, performing differential noise reduction on multi-sensor signals can highlight differences in leakage signals, achieving a differential noise reduction effect. For example, for two adjacent sensors A and B, signal B can be subtracted from signal A to cancel out the common background noise components, thereby highlighting the leakage component within the difference between the two. For data from multiple sensors, methods such as average subtraction and principal component analysis can also be used to extract the noise subspace and remove the main noise components from the signal.

[0029] Adaptive Noise Modeling Module: This module models and updates ambient noise in real time, dynamically adjusting detection parameters based on noise levels. During system operation, this module continuously monitors the statistical characteristics of background signals, such as mean, variance, and spectral distribution, to establish a noise model. When ambient noise changes, the adaptive model can adjust the parameters of previous modules accordingly, such as adjusting the weights of differential noise reduction or updating threshold settings to adapt to the new noise level. This adaptive mechanism ensures that the algorithm maintains stable performance under different operating conditions, avoiding the failure of fixed parameters in certain situations. For example, when ambient noise decreases at night, the model can automatically increase detection sensitivity, while when noise increases during the day, the threshold is raised to reduce false alarms.

[0030] Deep Learning Analysis Module: This module utilizes a pre-trained deep learning model to perform pattern recognition and classification on noise-reduced signals. This module can employ deep neural networks, such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), or a combination thereof, as leakage detection models. Network inputs can include processed multi-sensor signals in the time domain, frequency domain features (power spectra, Mel-frequency cepstral coefficients (MFCCs), or time-frequency images (short-time Fourier transform spectrograms, wavelet transform coefficients), etc. Through the deep learning model's automatic feature extraction capabilities, characteristic patterns of leakage are extracted from the data and a judgment result is output, such as the probability of leakage, leakage level, or a simple leak presence / absence indication. Deep learning models are capable of recognizing complex signal patterns and subtle feature changes, significantly improving the reliability and accuracy of weak signal detection.

[0031] Alarm and display module: triggers an alarm and records leakage information when a suspected leak is detected.

[0032] This example uses monitoring high-pressure fluid pipeline leaks as an example. Each sensor is an acoustic sensor attached to the pipe wall, such as a piezoelectric accelerometer, and is evenly distributed at different locations along the pipeline. These sensor signals are synchronously sampled, and an appropriate sampling rate is selected based on the spectral characteristics of the leak sound to cover the main frequency band of the leak sound. For example, the sampling rate is set to 50 kHz.

[0033] Time-shift correction: Leakage sound propagates at a finite speed in a medium—approximately 340 m / s in air and up to several thousand m / s in steel pipes. Multiple sensors are located at varying distances from the leak source, resulting in arrival delay differences in the collected leak signals. Without correction, the same leak event recorded by different sensors will be misaligned on the timeline, and direct superposition will blur the signal characteristics. Therefore, it is first necessary to perform time-shift correction on the multi-sensor signals, also known as time alignment or time synchronization. Time-shift correction is achieved using the following two methods: Method 1: Known sensor position and sound velocity: The position of each sensor relative to the pipeline coordinates is obtained in advance. When a leak is suspected, the signal of a reference sensor is selected as a reference, and the other sensor signals are time-shifted according to the distance difference between them and the reference sensor. According to the formula Calculate, where 、 For sensors and reference sensor The distance to a reference point, such as the starting point of a pipeline, The propagation speed of the leakage sound in the medium can be estimated based on the pipe material and fluid properties. By applying the corresponding time offset, the signals of all sensors are aligned to the same time base.

[0034] Method 2: Correlation analysis to estimate time delay: In the absence of precise position parameters, the cross-correlation method can also be used to automatically estimate the time delay between sensor pairs. and , calculate their cross-correlation function , find the moment when the correlation function peaks The most likely delay difference between the signals will be The signal relative to the sensor Signal offset Alignment is now achieved. By selecting one sensor as the reference and correlating it with the remaining sensors, we can obtain the estimated time delays of all sensors relative to the reference. The correlation method can accurately align the main correlation signals in the presence of noise, but the correlation peak may not be obvious when the noise is strong. Therefore, if necessary, it can be combined with the known distance method or exploit multi-sensor data redundancy to improve robustness.

[0035] After time-shift correction is complete, the data streams from each sensor are aligned on the time axis. If these synchronized signals are then superimposed, such as by averaging or summing, the leakage signal, due to its consistent phase across channels, will be enhanced. Unrelated noise, due to its random phase across channels, will partially cancel each other out, thereby improving the overall signal-to-noise ratio. This lays the foundation for differential noise reduction in the next step.

[0036] Differential noise reduction and signal fusion: The time-synchronized multi-channel sensor signals are recorded as , is the number of sensors. These signals contain both leakage signal components and background noise components. To further extract leakage features, the present invention introduces differential noise reduction based on the synchronized signals. This process leverages the correlation and difference information between multi-sensor signals to suppress noise and highlight leakage.

[0037] Signal difference calculation: First calculate the difference between the two multi-sensor signals. For example, for adjacent sensor pairs ,calculate . Differential signaling will eliminate and If the background noise is relatively consistent in space, then and The similar noise components in the image are cancelled out by subtraction, but the leakage signal still remains in the image due to the difference in amplitude or phase between the two locations. This differential operation is equivalent to a spatial high-pass filter, which filters out spatially uniformly distributed signals and highlights locally changing signals.

[0038] Multi-signal fusion: Fusion is performed on multiple differential signals obtained by calculation. One method is to sum or average all differential signals. Another method is to use principal component analysis (PCA) or independent component analysis (ICA) to obtain the multi-channel signal. The main noise components and the signal components of interest are extracted from the covariance matrix of the multi-sensor signal. Specifically, the covariance matrix of the multi-sensor signal can be constructed and the eigenvalues are decomposed; the eigenvector corresponding to the maximum eigenvalue represents the main common component - most of which is noise, and the eigenvector corresponding to the minimum eigenvalue represents the difference component - which may contain leakage signals. Noise reduction is achieved by discarding the main characteristic components and retaining the secondary characteristic components. For real-time implementation, it can be simplified to sum and average the differential signals to reduce the amount of calculation. In this embodiment, the average of the differences of adjacent sensors is taken as the comprehensive differential signal , the formula is: For each pair Differences were taken and summed. Ideally, the environmental noise is basically canceled out, leaving only the local abnormal components caused by leakage.

[0039] Filtering and feature enhancement: signal after differential fusion , apply frequency band filtering as needed to further improve the signal-to-noise ratio. If it is known that the leakage sound is mainly concentrated in a specific frequency band, such as a few hundred Hz to a few kHz, a bandpass filter can be designed to extract the energy of this frequency band and filter out other irrelevant frequency noise. Calculate the short-term energy or envelope to smooth out instantaneous spike interference. After this series of differential and filtering processes, the leakage signal characteristics will be significantly highlighted, laying the foundation for subsequent pattern recognition.

[0040] In some cases, such as when there are a large number of sensors or they are distributed over a wide area, it's possible to go beyond simple pairwise differentiation and employ more complex spatial filtering algorithms, such as beamforming, to target specific locations for gain. Beamforming enhances signals originating from a specific spatial direction or propagation velocity by summing signals from multiple sensors with appropriate weights. This is equivalent to a generalized time-shift alignment and summation process. If the approximate propagation direction or velocity range of the leakage wave is known, a beamformer can be designed to provide high gain for the leakage signal and attenuate noise from other directions, thereby achieving noise reduction. These are all extended implementations of the differential noise reduction module.

[0041] Adaptive Noise Modeling and Parameter Adjustment: Background noise levels in industrial sites can vary significantly over time. For example, equipment startup and shutdown, human interference, and ambient wind noise can all cause noise level fluctuations. To ensure the robustness of the leak detection algorithm to noise variations, this paper designs an adaptive noise modeling module that estimates noise online and dynamically adjusts relevant parameters.

[0042] Noise model initialization: When the system is started or in the early stages of operation, a background signal without leakage can be collected as a baseline sample. By analyzing this sample, an initial noise model is established. For example, the mean value of the background noise is calculated. , standard deviation , power spectrum distribution These statistical features form the initial noise model, which is used to guide subsequent processing. If pure background samples cannot be obtained in advance, the noise model can be updated during operation by detecting signal segments without suspicious events for a long time.

[0043] Real-time update strategy: When the system is running continuously, the noise model will be continuously updated based on new data. Common methods are sliding window statistics or exponentially weighted average. For example, to estimate the background noise energy, you can use: in is the total signal energy currently observed, is a smoothing coefficient between 0 and 1, such as 0.9, is the updated noise energy estimate. Through this recursion, the noise model can gradually reflect the average noise level of the current environment. At the same time, it is also possible to track the changes in the spectrum separately - for example, maintaining a background noise power spectrum that is updated over time , which is used to identify the appearance of new noise components or the disappearance of original noise components.

[0044] Parameter adaptive adjustment: After the noise model is updated, it will feedback and affect the parameter settings of the previous processing modules to achieve adaptive control. Specifically including: Differential noise reduction parameters: If the noise model indicates an increase in the current global noise level, the system can increase the intensity of the noise subspace reduction during the differential fusion process. For example, the weight in the aforementioned summation and averaging of the differential signals can be increased to cancel out more common components. If the noise spectrum changes, the filter passband frequency can also be adjusted to match the new prominent frequency band of the leakage signal.

[0045] Detection decision threshold: Before or after deep learning analysis, it's often necessary to set a decision threshold. For example, a neural network output probability greater than a certain value indicates a leak. The Adaptive Noise Modeling module dynamically adjusts this threshold based on the current noise level. Raising the threshold when noise is high reduces false positives, while lowering it when noise is low helps detect subtle leaks.

[0046] Deep learning model input adjustments: Significant changes in the noise spectrum can cause previously trained models to degrade in performance. The Adaptive Noise Modeling module can trigger adjustments to model input features, such as adding features specific to the new noise frequency band or enabling adaptive batch normalization in the model to make it robust to changes in mean and variance.

[0047] This adaptive mechanism continuously learns the characteristics of ambient noise, ensuring stable operation under various working conditions. For example, when construction noise temporarily increases, the system automatically lowers sensitivity to avoid false alarms. Once construction is complete and noise subsides, it returns to high sensitivity to detect potential leaks. Compared to fixed-parameter systems, adaptive systems significantly reduce manual intervention and false alarms.

[0048] Deep learning analysis and leakage identification: After the aforementioned preprocessing, the leakage characteristics in the signal have been extracted and enhanced as much as possible. Next, the core deep learning analysis module enters, where a trained neural network model performs pattern recognition on the signal to determine whether there is a leak.

[0049] Feature extraction and input construction: In order for deep learning models to learn leakage patterns more efficiently, it is usually necessary to convert the preprocessed signal into a suitable feature input. One method is to use the time series after differential denoising. As input, let the neural network extract time domain features by itself. Another more common method is to Perform certain transformations to extract frequency domain or time-frequency domain features, which are then used as network input. For example: calculate The short-time Fourier transform STFT of the time-varying spectrum is obtained , taking its amplitude or power to obtain a time-frequency image. This is equivalent to converting a signal into a matrix of data in the form of time × frequency, which can be used as the input image for a convolutional neural network. Several statistical features are extracted, such as energy, peak frequency, and spectral center of gravity in multiple frequency bands, or MFCC Mel-frequency cepstral coefficients are calculated to obtain a fixed-length feature vector, which is then input into a classifier such as a fully connected network or support vector machine. If a sequence model such as an LSTM network is used, a feature sequence segmented by time frame can be directly input, for example, a multi-dimensional feature calculated every 0.1 second. This allows the LSTM to capture temporal evolution patterns.

[0050] In this embodiment, the differential signal is selected The normalized spectrum graph of is used as the input of the deep CNN model. The specific approach is: STFT analysis is performed with a window length of 1 second to obtain a spectrum graph per second. The spectrum graph amplitude is normalized by subtracting the mean and dividing by the standard deviation before inputting into a two-dimensional convolutional neural network.

[0051] Deep learning model structure: The structure of the neural network model can be designed as needed. To balance complexity and performance, this embodiment uses a structure combining a convolutional neural network (CNN) with a fully connected layer for binary classification of leak detection. The model mainly includes: Convolutional layer + pooling layer: Several convolutional layers extract local characteristic patterns in the input spectrogram, such as energy spikes at specific frequency combinations. Each convolution layer is followed by a pooling layer to reduce data dimensionality and increase invariance. For example, the first convolution kernel size is 5×5 to extract primary features, while the second convolution kernel size is 3×3 to extract higher-level features. 2×2 max pooling is used to reduce the size.

[0052] Fully connected layer: Flattens the feature map output by the last convolutional layer and connects one or two layers of fully connected neurons to combine the features extracted by convolution and perform classification. A hidden layer with 128 neurons uses a ReLU activation function and a Sigmoid output layer activation function to output leakage probabilities.

[0053] Dropout regularization: During training, the Dropout strategy is used to randomly discard some neurons in the fully connected layer to prevent overfitting and improve the model's ability to generalize to new data.

[0054] The entire network contains approximately hundreds of thousands to one million trainable parameters. The training process uses the cross-entropy loss function and the Adam optimizer. Deep models with other architectures can also be used, such as dual-channel input—using two sets of sensor signals as separate inputs—or time-series-based RNN / LSTM networks to capture temporal dynamics. However, experiments have shown that for the differential spectral features extracted by this method, CNNs are already able to effectively learn the spatial spectral characteristics and pattern differences of leakage.

[0055] Model training and verification: Before the system is put into practical application, a large amount of sample data is needed to train the deep learning model offline. The training data includes signals under normal working conditions and signals under various simulated leakage conditions, preferably covering a variety of noise scenarios. Samples can be obtained by combining on-site collection and simulation: for example, recording background noise samples when there is no leakage, creating a controllable small hole leak on the experimental pipeline to obtain leakage signal samples; or using existing leak detection public data sets, combined with the preprocessing method of the present invention to perform data enhancement. In order to improve the robustness of the model to high noise, different levels of noise can be artificially superimposed on the leakage signal during training, or the noise component generated by the adaptive noise modeling module can be mixed with the leakage signal to expand the distribution of training samples.

[0056] After training, the model's performance was verified on independent test data, including metrics such as accuracy, recall (a complement to missed detection rate), and false alarm rate. In the tests of the embodiment, the deep learning model achieved excellent detection results under various noise conditions. For example, the accuracy of detecting small leaks simulated in a high-noise laboratory background exceeded 95%, far exceeding the approximately 80% accuracy of the traditional correlation method under the same signal-to-noise ratio. The model confusion matrix showed that almost all leak samples were correctly identified, while only a very small number of no-leak samples were falsely reported.

[0057] Leakage judgment and output: The deep learning model outputs a probability value of leakage for each analysis window Or classification score. Set decision threshold ——For example, 0.5 corresponds to a probability of 50% for the final decision: if , then it is determined that there is a leak and an alarm is triggered; otherwise it is determined that there is no leak. Under the action of the adaptive noise modeling module, Thresholds can be adjusted based on noise levels or employ dynamic, time-varying thresholds to further reduce false alarms. If a leak is detected across multiple consecutive windows, the alarm level is raised and appropriate safety measures are implemented, such as automatically closing valves and notifying inspection personnel. The system also records the time and possible location of the leak—information that can be used by maintenance personnel to roughly determine the pipeline segment where the leak occurred, based on which sensor group detected it first.

[0058] Deep learning models can also output richer information, such as multi-classification to determine the severity of a leak. For example, the model can be trained to categorize leaks as small, medium, or large, outputting labels of varying severity levels, allowing operations personnel to determine appropriate response measures. Under the framework of this invention, simply by adding samples of varying leak flow rates to the training data and applying corresponding labels, the model can learn to distinguish leak size based on signal amplitude characteristics. Our implementation incorporates a regression module for leak intensity, enabling the model to estimate the energy ratio of the leak noise relative to the background noise, thereby indirectly inferring leak size.

[0059] The workflow of a leak detection method based on multi-sensor time-shift correction and deep learning is as follows: Deployment and initialization: Multiple sensors are placed on the monitored pipeline and connected to a signal processing unit. Initially, background noise samples are collected when there is no leakage, an initial noise model is established, and parameters such as the initial detection threshold are set.

[0060] Data acquisition and synchronization: Continuously collect sensor signals, timestamp, and cache the data. At each detection cycle—for example, every second—the digital signal sequence from each sensor within that cycle is retrieved and time offsets are calculated using a preset distance or correlation method to align and synchronize the signal sequences.

[0061] Difference and preprocessing: Calculate the difference between the synchronized signals and fuse them to obtain a signal with enhanced leakage characteristics .right Perform necessary filtering or feature calculations to extract the feature vector or time-frequency diagram that represents the signal characteristics of the current period.

[0062] Noise Model Update: Updates the background noise model parameters based on the most recent signal. If there are no obvious leaks, this is primarily used to learn the noise. If an anomaly is detected, it is considered for model update. Parameters such as the differential algorithm and detection threshold are adjusted to accommodate the new noise level.

[0063] Deep Learning Discrimination: The extracted features are fed into a deep learning model to calculate the leakage probability or classification result. An adaptive threshold is applied to determine whether a leak has occurred, along with the possible confidence level and classification level.

[0064] Result Output and Alarm: If a leak is detected, the event is recorded and an alarm is triggered—audio, visual, or remote notification. Detailed data for that period is also saved for further analysis. If no leak is detected, the monitoring cycle continues. This process is repeated continuously, achieving real-time online leak monitoring.

[0065] This invention enables early, automated detection of pipeline leaks in noisy environments. When a leak occurs, the multi-sensor network promptly captures subtle anomalies and uses intelligent algorithms to identify leak signals, eliminating potential alerts missed due to background noise. Those skilled in the art can build a practical system based on this technology or create a portable leak monitoring tool kit to monitor and detect different devices as needed.

Claims

1. A leak detection method based on multi-sensor time-shift correction and deep learning, characterized by: It includes a multi-sensor data acquisition module: multiple sensors sensitive to leakage sound waves or vibrations are arranged at different locations of the monitored pipeline or container to collect leakage-related signals; a time shift correction module: time-aligns the time series signals collected by multiple sensors according to their position differences in space; and a differential noise reduction module: compares and performs differential processing on the time-aligned multiple sensor signals. Adaptive noise modeling module: models and updates environmental noise in real time, dynamically adjusts detection parameters based on noise levels, continuously monitors the statistical characteristics of background signals, and establishes a noise model; Deep learning analysis module: uses a pre-trained deep learning model to perform pattern recognition and classification judgment on the signal after noise reduction processing; alarm and display module: triggers an alarm and records the leakage information when a suspected leak is detected.

2. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 1, characterized in that: Multi-sensor data acquisition module: Multiple sensors include accelerometers, acoustic emission sensors or hydrophones.

3. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 1, characterized in that: Differential noise reduction module: For data from multiple sensors, the average subtraction method and principal component analysis are used to extract the noise subspace and remove the main noise components from the signal.

4. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 1, characterized in that: Adaptive noise modeling module: The statistical characteristics of the background signal include mean, variance, and spectral distribution; when the ambient noise changes, the differential noise reduction weight is adjusted or the threshold setting is updated; when the ambient noise decreases at night, the detection sensitivity is automatically increased; when the noise increases during the day, the threshold is increased to reduce false alarms.

5. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 1, characterized in that: Deep learning analysis module: uses deep neural networks, including convolutional neural networks, long short-term memory networks or their combinations as leakage discrimination models. The network input is the time domain waveform, frequency domain features or time-frequency images of multiple sensor signals after processing. Through the automatic feature extraction capability of the deep learning model, the characteristic pattern of the leakage is extracted from the data, and the judgment results are output, including the probability of leakage occurrence, leakage level or the presence or absence of leakage indication.

6. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 1, characterized in that: Time shift correction module: Knowing the positions and sound speeds of multiple sensors: Get the position of each sensor relative to the pipeline coordinates in advance. When a leak is suspected, select the signal of a reference sensor as a reference and perform time shift correction on the signals of other sensors based on the distance difference between them and the reference sensor. The time shift correction is done. According to the formula calculate, 、 For sensors and reference sensor The distance to the reference point, The propagation speed of the leakage sound in the medium is calculated by applying the corresponding time offset to align the signals of all sensors to the same time base; Correlation analysis estimates the time delay: Without the need for precise position parameters, the cross-correlation method is used to automatically estimate the time delay between the sensor pairs. and , calculate their cross-correlation function , find the moment when the correlation function peaks The most likely delay difference between the signals will be The signal relative to the sensor Signal offset Alignment: Select one sensor as the reference in turn, pair it with the remaining sensors and calculate the correlation, and obtain the estimated value of the time delay of all sensors relative to the reference. If necessary, combine the known distance method or use the redundancy of multiple sensor data to improve robustness.

7. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 3, characterized in that: The time-aligned multiple sensor signals are recorded as , For the number of sensors, calculate the difference between the two pairs of multiple sensor signals. ,calculate , differential signal eliminate and The common parts in the multi-channel signal are fused again: one method is to superimpose and sum or average all the differential signals; another method is to use principal component analysis or independent component analysis to obtain the multi-channel signal. Extract the main noise components and signal components of interest, construct the covariance matrix of multiple sensor signals, and perform eigendecomposition on them. The eigenvector corresponding to the maximum eigenvalue represents the main common component, and the eigenvector corresponding to the minimum eigenvalue represents the difference component. Noise reduction is achieved by discarding the main characteristic components and retaining the secondary characteristic components. The average of the differences of adjacent sensors is taken as the comprehensive differential signal : For each pair Perform differential and accumulation; the signal after differential fusion , apply frequency band filtering as needed to further improve the signal-to-noise ratio; if it is known that the leakage sound is concentrated in a specific frequency band, design a bandpass filter to extract the energy of this frequency band and filter out other irrelevant frequency noise; Calculate short-term energy or envelope to smooth out instantaneous spike interference. When there are a large number of sensors or they are distributed over a large area, use beamforming technology to target specific locations for gain. If the approximate propagation direction or speed range of the leakage wave is known, design a beamformer to provide high gain for the leakage signal and attenuate noise in other directions, thereby achieving noise reduction.

8. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 4, characterized in that: At the start-up or initial operation, collect the background signal in the no-leak state as the baseline sample, establish the initial noise model by analyzing this sample, and calculate the mean value of the background noise , standard deviation , power spectrum distribution If pure background samples cannot be obtained in advance, the noise model is updated during operation by detecting signal segments where no suspicious events occur for a long time. During continuous operation, the noise model is continuously updated based on new data using sliding window statistics or exponentially weighted averages. The background noise energy is estimated using: is the total signal energy currently observed, is a smoothing coefficient between 0 and 1, This recursive noise model can gradually reflect the average noise level of the current environment and track the changes in the spectrum separately - maintaining a background noise power spectrum updated over time. , which is used to identify the appearance of new noise components or the disappearance of original noise components.

9. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 8, characterized in that: After the noise model is updated, it affects the parameter settings of each module to achieve adaptive control: Differential noise reduction parameters: If the noise model indicates that the current global noise level is increasing, the noise subspace attenuation during differential fusion is increased, and the weight in the summation and averaging of the differential signals is increased to offset more common components. If the noise spectrum changes, the filter passband frequency is adjusted to match the new prominent frequency band of the leakage signal. Detection decision threshold: When the probability of the neural network output is greater than the threshold, it is judged as a leak. The threshold is dynamically adjusted according to the current noise level. When the noise is high, the threshold is increased to reduce false positives, and when the noise is low, the threshold is lowered to capture weak leaks. Deep learning model input adjustment: Significant changes in the noise spectrum trigger adjustments to the model input features, such as adding features specific to the new noise frequency band or enabling the adaptive batch normalization mechanism in the model to make the model robust to changes in mean / variance. When construction noise increases temporarily, the sensitivity is automatically lowered to avoid false alarms; when the noise decreases after construction is completed, the high sensitivity is restored to capture potential leaks.

10. The leak detection method based on multi-sensor time shift correction and deep learning according to claim 5, characterized in that: Convert the preprocessed signal into a suitable feature input: One method is to use the time series after difference denoising As input, let the neural network extract time domain features by itself; Another method is to Transform to extract frequency domain or time-frequency domain features, and then use them as network input: Calculate The short-time Fourier transform of , take its amplitude or power to get the time-frequency image, as the input image of the convolutional neural network, extract statistical features, including energy, peak frequency, spectral center of gravity in multiple frequency bands or calculate Mel frequency cepstral coefficients to obtain a fixed-length feature vector, input to the fully connected network or support vector machine classifier, select the differential signal The normalized spectrum map of is used as the input of the deep convolutional neural network model: Perform short-time Fourier transform analysis with a window length of 1 second to obtain a spectrum graph per second, and normalize the amplitude of the spectrum graph - subtract the mean and divide it by the standard deviation before inputting it into a two-dimensional convolutional neural network; use a structure combining a convolutional neural network and a fully connected layer for binary classification leakage judgment. The model includes: convolution layer + pooling layer: several convolution layers extract local feature patterns in the input spectrum graph, including energy surges of specific frequency combinations. Each convolution layer is followed by a pooling layer to reduce the data dimension and increase invariance: the first layer has a convolution kernel size of 5×5 to extract primary features; the second layer has a convolution kernel size of 3×3 to extract higher-level features; the size is reduced by 2×2 maximum pooling; fully connected layer: flatten the feature map output by the last convolution layer, connect one or two layers of fully connected neurons, and use them to combine the features extracted by convolution and perform classification judgment. Use a hidden layer with 128 neurons as the activation function ReLU and an output layer activation function Sigmoid to output the leakage probability; during training, the fully connected layer uses the Dropout strategy to randomly discard some neurons to prevent overfitting and improve the model Ability to generalize to new data; use cross entropy loss function and Adam optimizer during training, or use deep models with other architectures: use two sets of sensor signals as input or time-series-based RNN / LSTM network to capture temporal dynamics; before putting into practical application, use a large amount of sample data to train the deep learning model offline. The training data includes signals under normal working conditions and signals under various simulated leakage conditions, covering a variety of noise scenarios, and obtain samples through a combination of on-site collection and simulation: record background noise samples when there is no leakage, and create controllable small hole leakage on the experimental pipeline to obtain leakage signal samples; or use existing leak detection public data sets, combined with preprocessing methods for data enhancement, artificially superimpose different levels of noise on the leakage signal during training, or use the noise components generated by the adaptive noise modeling module to mix with the leakage signal to expand the distribution of training samples; after training, verify the model performance on independent test data, including accuracy, recall rate and false alarm rate indicators; the deep learning model outputs a probability value of the existence of a leak for each analysis window Or classification score, set decision threshold For final judgment: If , it is determined that there is a leak and an alarm is triggered; otherwise it is determined that there is no leak; under the action of the adaptive noise modeling module, Adjust according to the noise level, or use a dynamic threshold that changes over time to further reduce false alarms; if multiple consecutive windows are determined to have a leak, the alarm level will be raised and corresponding safety measures will be implemented. The time and possible location of the leak incident will be recorded for reference by maintenance personnel.

Images