A leak detection method based on multi-sensor time-lapse correction and deep learning

Abstract

The present application 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 vibrations are arranged at different positions of a monitored pipeline or container for collecting leakage-related signals; the time series signals collected by the plurality of sensors are time-aligned according to the position differences thereof in space; the plurality of sensor signals subjected to time alignment are compared and subjected to differential processing; the environmental noise is modeled and updated in real time, the detection parameters are dynamically adjusted according to the noise level, the statistical characteristics of the background signals are continuously monitored, and the noise model is established; the pre-trained deep learning model is used to perform pattern recognition and classification judgment on the signals subjected to noise reduction processing; and an alarm and display module: when a suspected leakage is detected, an alarm is triggered and leakage information is recorded. The present application can reliably detect weak leakage signals under strong background noise conditions.

Description

TECHNICAL FIELD

[0001] The application 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. BACKGROUND

[0002] In the fields of nuclear power plants, petrochemical industry, etc., leakage monitoring is crucial for ensuring safety and reducing losses. However, the high-noise field environment poses a great challenge to leakage detection: the background noise generated by normally operating equipment often overwhelms the leakage sound, making it difficult for traditional detection methods to detect small leaks in a timely manner. Existing acoustic leakage detection techniques mainly include the following categories:

[0003] Manual listening and simple sensor monitoring: such as using a listening rod, a simple microphone, or a liquid listener to listen to the pipeline leakage sound at positions such as valves and fire hydrants. This method has been used since the 19th century, but it is not sensitive to small leaks, and the weak leakage sound signal can be easily masked by environmental noise, resulting in low reliability. The operator often needs to review multiple times to avoid misjudging environmental noise as leakage sound.

[0004] Noise recording and statistical analysis: a noise recorder can record pipeline noise for a long time, and determine the existence of leakage by statistically analyzing the changes in noise intensity. However, in strong interference situations, this method is easily affected by non-leakage factors, has a high false alarm rate, and requires a long monitoring period to extract signal change characteristics.

[0005] Double-sensor correlation detection: a leakage noise correlation instrument uses sensors deployed on both sides of the pipeline to calculate the correlation based on the time difference of the propagation of the leakage sound at the two measurement points to determine the leakage location. This method is effective for locating medium to large scale leaks, but in complex high-noise environments, the correlation peak value may be overwhelmed or shifted by noise, leading to missed detection or misjudgment. In addition, the correlation method usually requires manual threshold setting and is difficult to adapt to non-stationary noise interference.

[0006] Single-sensor signal processing and machine learning: with the development of computing technology, research has emerged that uses signal processing and pattern recognition to automatically detect leaks. For example, literature reports that empirical mode decomposition is used to extract pipeline vibration signal features, and then a one-dimensional convolutional neural network is used to realize water pipe leakage detection. For another example, a deep learning model is used to input time-frequency features obtained by short-time Fourier transform or wavelet transform to improve the recognition ability of complex leakage patterns. These methods have improved detection accuracy to some extent. However, traditional signal processing methods are still sensitive to noise and are difficult to adapt to non-stationary signals, and have high computational overhead, making real-time monitoring difficult. Purely relying on single-sensor data also lacks spatial redundancy information, and the reliability will decrease once the sensor signal is disturbed by occasional noise.

[0007] In summary, the prior art is prone to false negatives / false positives or requires complex signal processing and is not adaptive when detecting leaks in a high-noise environment. Therefore, it is necessary to provide a new technical solution that combines multi-sensor data fusion and advanced signal processing and deep learning algorithms to accurately and robustly detect leaks in a high-noise environment. SUMMARY

[0008] The present application aims to provide a leak detection method based on multi-sensor time shift correction and deep learning, which combines 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 leak signals in a strong background noise environment and achieve high-precision detection of weak leak signals in industrial pipelines and other facilities.

[0009] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0010] A leak detection method based on multi-sensor time shift correction and deep learning, comprising a multi-sensor data acquisition module: arranging multiple sensors sensitive to leak acoustic waves or vibrations at different positions of the monitored pipeline or container for collecting leak-related signals; a time shift correction module: time aligning the time series signals collected by the multiple sensors according to their spatial position differences; a differential noise reduction module: comparing and differentially processing the time-aligned multiple sensor signals; an adaptive noise modeling module: modeling and updating the environmental noise in real time, dynamically adjusting the detection parameters according to the noise level, continuously monitoring the statistical properties of the background signal, and establishing a noise model; a deep learning analysis module: using a pre-trained deep learning model to identify patterns and classify the signals after noise reduction; an alarm and display module: triggering an alarm and recording leak information when a suspected leak is detected.

[0011] The multi-sensor data acquisition module: the multiple sensors include accelerometers, acoustic emission sensors, or hydrophones.

[0012] The differential noise reduction module: for the data of the multiple sensors, the average reduction method and principal component analysis are used to extract the noise subspace and eliminate the main noise components from the signal.

[0013] The adaptive noise modeling module: the statistical properties of the background signal include mean, variance, and spectral distribution; when the environmental noise changes, the weight of the differential noise reduction or the threshold setting is adjusted; at night when the environmental noise decreases, the detection sensitivity is automatically increased, and during the day when the noise increases, the threshold is increased to reduce false positives.

[0014] Deep learning analysis module: adopt deep neural network, including convolutional neural network, long short-term memory network or their combination as leakage discrimination model, network input is multiple sensor signals processed time domain waveform, frequency domain features or time-frequency image; through the automatic feature extraction ability of deep learning model, the characteristic mode of leakage is refined from the data, and the judgment result is output, including outputting the probability of leakage occurrence, leakage level or no leakage indication.

[0015] Time shift correction module: known multiple sensor positions and sound speed: pre-acquire the position of each sensor relative to the pipeline coordinate, when suspected to occur leakage, select the signal of a reference sensor as reference, correct the time offset of other sensor signals according to the distance difference between them and the reference sensor, offset time According to the formula , , The distance from the sensor And the reference sensor To the reference point, The propagation speed of leakage sound in medium, align all sensor signals to the same time reference by applying corresponding time offset; correlation analysis estimates time delay: without the need for accurate position parameters, cross-correlation method is used to automatically estimate the time delay between sensor pairs, for two sensor signals And , calculate the cross-correlation function , find the time The most likely delay difference between the signals, align the signal of sensor With respect to the signal of sensor Offset , select one sensor as reference in turn, pair with the remaining sensors to get the correlation, get the time delay estimation value of all sensors relative to the reference, if necessary, combine known distance method or use multiple sensor data redundancy to improve robustness.

[0016] The multiple sensor signals after time alignment are recorded as , The number of sensors, calculate the difference between multiple sensor signals, for adjacent sensor pairs , calculate , difference signal Eliminate And Commonly existing part; fuse the multiple difference signals calculated again: one method is to add up or average all difference signals; another method is to use principal component analysis or independent component analysis method to extract multiple channel signals The main noise components and signal components of interest are extracted, a covariance matrix of the multiple sensor signals is constructed, eigenvalue decomposition is performed on the covariance matrix, 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; and the average of the differences of adjacent sensors is taken as the comprehensive difference signal :

[0017]

[0018] Difference and accumulation are performed on each pair ; the signal after difference fusion is filtered by a frequency band filter to further improve the signal-to-noise ratio; if the leakage sound is known to be concentrated in a specific frequency band, a band-pass filter is designed to extract the energy of the frequency band, and other irrelevant frequency noises are filtered out; the short-time energy or envelope is calculated to smooth the instantaneous peak interference; when the number of sensors is large or the distribution range is large, a beam forming technique is used to increase the gain for a specific position; if the approximate propagation direction or speed range of the leakage wave is known, a beam former is designed to have high gain for the leakage signal and to attenuate noise in other directions, so as to achieve the purpose of noise reduction.

[0019] Background signals in a no-leakage state are collected at the start or in the early stage of running as reference samples, an initial noise model is established by analyzing the samples, the mean , standard deviation , and power spectrum distribution of the background noise are calculated, if a pure background sample cannot be obtained in advance, the noise model is updated by detecting a signal segment with no suspicious event for a long time during the running process; when continuously running, the noise model is constantly updated according to new data, and the method is sliding window statistics or exponential weighted average, and the estimation of the background noise energy adopts:

[0020] is the total signal energy observed at present, is a smoothing coefficient between 0 and 1, is the updated noise energy estimation value, and through this recursive noise model, the average noise level of the current environment can be gradually reflected, and the changes in the spectrum are tracked and maintained respectively a background noise power spectrum updated over time is used to identify the appearance of new noise components or the disappearance of original noise components.

[0021] After the noise model is updated, the parameter settings of each module are affected, and adaptive control is realized:

[0022] Differential noise reduction parameter: if the noise model shows that the current global noise level is rising, increase the strength of noise subspace weakening in the differential fusion process, increase the weight in the differential signal summation average, so that more common components are canceled out, if the noise spectrum changes, adjust the passband frequency of the filter to match the new leakage signal prominent frequency band;

[0023] Detection decision threshold: the neural network output probability is greater than the threshold value to determine whether there is leakage, and the threshold value is dynamically adjusted according to the current noise level, the threshold value is increased to reduce false positives when the noise is high, and the threshold value is reduced to capture weak leakage when the noise is low;

[0024] Deep learning model input adjustment: when the noise spectrum changes significantly, trigger the adjustment of the model input features, increase the feature quantity for the new noise frequency band, or enable the adaptive batch normalization mechanism in the model, so that the model remains robust to changes in mean / variance;

[0025] When the construction noise temporarily increases, automatically adjust the sensitivity to avoid false positives; when the construction noise decreases, restore the high sensitivity to capture potential leaks.

[0026] Convert the preprocessed signal into appropriate feature input: one method is to use the differential noise-reduced time series 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 input them into the network: calculate the short-time Fourier transform of to get the time-varying frequency spectrum , take its amplitude or power to get the time-frequency image, input it into the convolutional neural network as the input image, extract statistical features, including energy, peak frequency, spectral barycenter in multiple frequency bands, or calculate the mel frequency cepstral coefficient to get a fixed-length feature vector, input it into the fully connected network or support vector machine classifier, select the normalized frequency spectrum of the differential signal as the input of the deep convolutional neural network model: take Short-time Fourier transform analysis is performed with a window length of 1 second to obtain a frequency spectrum diagram per second, and the amplitude of the frequency spectrum diagram is normalized, and then input into a two-dimensional convolutional neural network after being subtracted by the mean value and divided by the standard deviation; a structure combining a convolutional neural network and a fully connected layer is used for binary classification of whether there is leakage, and the model includes: a convolutional layer + a pooling layer: a plurality of convolutional layers extract local feature patterns in the input frequency spectrum diagram, including a sudden increase in energy of a specific frequency combination, and a pooling layer is connected after each convolutional layer to reduce the data dimension and increase the invariance: the first convolutional kernel size is 5*5 to extract primary features; the second convolutional kernel size is 3*3 to extract higher-level features; the size is reduced through 2*2 maximum pooling; a fully connected layer: the feature map output by the last convolutional layer is flattened, and one or two layers of fully connected neurons are connected to combine the features extracted by convolution and perform classification and discrimination, and a layer of 128 neurons is used as an implicit layer with a ReLU activation function and a Sigmoid output layer activation function to output a leakage probability; in the training, the fully connected layer uses the Dropout strategy to randomly discard part of the neurons to prevent overfitting and improve the generalization ability of the model to new data; in the training process, a cross-entropy loss function and an Adam optimizer are used, or a deep model with other architectures is used: two groups of sensor signals are used as inputs or time-based RNN / LSTM networks to capture time dynamics; before being put into practical application, a large amount of sample data is used to train the deep learning model offline, and the training data includes signals under normal working conditions and signals under various simulated leakage working conditions, covering various noise scenes, and the samples are obtained by combining field collection and simulation: background noise samples are recorded when there is no leakage, and leakage signal samples are obtained by manufacturing controllable small hole leaks on the experimental pipeline; or use an existing leakage detection public data set, combine a pre-processing method to perform data enhancement, artificially superimpose different levels of noise on the leakage signal during training, or mix the noise component generated by the adaptive noise modeling module with the leakage signal to expand the training sample distribution; after training is completed, the model performance is verified on independent test data, including accuracy, recall rate and false alarm rate indicators; the deep learning model outputs a probability value or a classification score for the existence of leakage for each analysis window, and a decision threshold is set for the final decision: if , it is determined that there is leakage at present, and an alarm is triggered; otherwise, it is determined that there is no leakage; under the action of the adaptive noise modeling module, the noise level is adjusted, or a dynamic threshold that changes over time is used to further reduce false alarms; if a plurality of windows in succession determine that there is leakage, the alarm level is raised and corresponding safety measures are performed, and the time and possible location area of the leakage event are recorded for reference by maintenance personnel.

[0027] The beneficial effects obtained by the present application are:

[0028] ​The present application combines the five technical means of multi-sensor fusion, time shift correction, differential noise reduction, adaptive modeling and deep learning together for the first time to form a complete solution. In existing solutions, although there are applications of multi-sensor correlation method or machine learning detection method, there is no public report of integrating all the above modules for leak detection. The combined innovation of the present application effectively avoids the limitations of a single method and is significantly different from the prior art.

[0029] The multi-sensor of the present application provides spatial information redundancy, ensures signal synchronization superposition with time shift correction, improves signal-to-noise ratio from the physical layer, further suppresses background noise interference with differential noise reduction, and ensures automatic optimization of algorithm parameters with adaptive noise model as the environment changes. Finally, the deep learning model extracts high-dimensional features and makes intelligent discrimination. Each link gains layer by layer, so that the system can reliably distinguish between leakage signals and noise in a strong noise background. Laboratory simulation and field testing show that the leak detection accuracy of the present method in a high noise environment is greatly improved, several percentage points higher than the baseline scheme without using the combined strategy, and the false alarm rate is significantly reduced.

[0030] The adaptive modeling of the present application enables the system to automatically adjust to different noise levels without frequent manual intervention for parameter adjustment, ensuring the stability of long-time operation. In addition, the algorithms of each step of the present application are optimized to realize real-time processing: time shift correction and differential operation have low computational complexity, adaptive noise model uses incremental update strategy with small overhead, and deep learning part can efficiently run on embedded GPU or special AI chip. Therefore, the whole system is suitable for online real-time monitoring and does not miss the leak warning due to calculation delay.

[0031] The present application ingeniously combines traditional signal processing methods with emerging deep learning technology and achieves better results than existing technologies in high-noise leak detection scenarios. This cross-domain combination has non-obviousness: for example, the idea of multi-sensor differential noise reduction combined with deep neural networks is not a simple superposition of existing technologies, but a new idea for leak detection problems. Although existing research has reported the combination of multi-sensor data fusion or signal decomposition with neural networks, it has not been foreseen to integrate time alignment, differential noise reduction and adaptive optimization into one. The overall scheme of the present application significantly exceeds the expectations of existing technologies and therefore has outstanding creativity.

[0032] The present application is not only suitable for long-distance pipeline leak detection, but also can be extended to storage tank leak detection, building water supply and drainage pipe network leakage and other scenarios that require leak detection in a noisy environment. By adjusting the sensor type and deep learning model structure, the system can be customized for different media (gas, liquid) and different sizes of leaks, and has strong practical value.

[0033] In summary, the present application has significant innovation compared with the prior art, and can effectively improve the accuracy and reliability of leakage detection in a high-noise environment. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 A flow chart of a leakage detection method based on multi-sensor time shift correction and deep learning;

[0035] Figure 2 A system architecture diagram of a leakage detection system based on multi-sensor time shift correction and deep learning;

[0036] In the figure: 1, the device to be tested; 2, the sensor. DETAILED DESCRIPTION

[0037] The present application will be described in detail below in conjunction with the drawings and specific embodiments.

[0038] As shown in Figure 1 , 2 The leakage detection method based on multi-sensor time shift correction and deep learning includes:

[0039] Multi-sensor data acquisition module: a plurality of sensors are arranged at different positions of the monitored pipeline or container for collecting leakage-related signals. These sensors are acoustic sensors sensitive to leakage sound waves or vibrations, such as accelerometers, acoustic emission sensors or hydrophones, etc., distributed at key points in the leakage monitoring area. The multi-sensor arrangement provides spatial information redundancy, which can capture the propagation signals of the leakage sound waves at different positions and provide basic data for subsequent processing.

[0040] Time shift correction module: the time series signals collected by each sensor are time-aligned according to their position differences in space. Since the leakage sound waves transmitted to different sensors have different propagation delays, the time delay needs to be estimated according to the distance difference between each sensor and the leakage source or through signal correlation, and the signal is corrected and synchronized. This time shift correction aligns the same leakage event in the multi-sensor signals, thereby facilitating subsequent fusion analysis.

[0041] Difference denoising module: compare and difference process the time-aligned multi-sensor signals to weaken the influence of background noise. Specifically, the difference signals between each sensor signal can be calculated to eliminate the noise components common to them. Since environmental noise, such as fluid noise in the pipeline or external mechanical noise, is usually spatially coherent or uniformly distributed, each sensor will receive similar noise components; while the leakage signal is spatially characterized by local features, with obvious changes in intensity with sensor position. Therefore, difference processing of multi-sensor signals can highlight the differences in the leakage signal, achieving the effect of difference denoising. For example, for two adjacent sensors A and B, the A signal can be subtracted from the B signal to cancel out the common part of the background noise, thereby highlighting the leakage component in the difference between the two. For data from multiple sensors, noise subspace can also be extracted using methods such as average reduction and principal component analysis to eliminate major noise components from the signal.

[0042] Adaptive noise modeling module: real-time modeling and updating of environmental noise, dynamically adjusting detection parameters according to noise level. This module continuously monitors the statistical properties of the background signal, such as mean, variance, spectral distribution, etc., during system operation, establishing a noise model. When the environmental noise changes, the adaptive model can adjust the parameters of the previous modules accordingly, such as adjusting the weight of difference denoising or updating the threshold setting, to adapt to the new noise level. This adaptive mechanism ensures that the algorithm maintains stable performance under different working conditions, avoiding the use of fixed parameters that may fail in certain situations. For example, when the environmental noise decreases at night, the model can automatically increase the detection sensitivity, while in the daytime when the noise increases, the threshold is increased to reduce false positives.

[0043] Deep learning analysis module: uses a pre-trained deep learning model to perform pattern recognition and classification on the noise-reduced signals. This module can use deep neural networks such as convolutional neural networks (CNN), long short-term memory networks (LSTM), or their combination as a leakage discrimination model. The network input can be the time-domain waveform, frequency-domain features such as power spectrum, mel-frequency cepstral coefficients (MFCC), or time-frequency images such as spectrograms of short-time Fourier transform, wavelet transform coefficients, etc. After processing the multi-sensor signals, the deep learning model can extract the characteristic patterns of the leakage from the data and output the judgment results, such as the probability of leakage occurrence, the level of leakage, or a simple indication of the presence or absence of leakage. Deep learning models can recognize complex signal patterns and subtle feature changes, greatly improving the reliability and accuracy of weak signal detection.

[0044] Alarm and display module: triggers an alarm and records leakage information when a suspected leak is detected.

[0045] This example takes monitoring of high pressure fluid pipeline leak as an example, each sensor is an acoustic sensor attached to the pipe wall, such as a piezoelectric accelerometer, evenly distributed at different positions of the pipeline. The signals of these sensors are synchronously sampled, and a suitable sampling rate is selected according to 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.

[0046] Time shift correction: the propagation of leak sound in the medium has a limited speed, about 340 m / s in air, and up to thousands of m / s in steel pipeline. The distances of multiple sensors from the leak source are different, resulting in time delay differences in the collected leak signals. If not corrected, the same leak event recorded by different sensors will be misaligned on the time axis, and direct superposition will cause the signal characteristics to be blurred. Therefore, time shift correction, also known as time alignment or time synchronization, is first needed for the multi-sensor signals.

[0047] Method one: known sensor position and sound speed: the positions of each sensor relative to the pipeline coordinates are obtained in advance. When a leak is suspected, the signal of a reference sensor is selected as a reference, and the signals of other sensors are time-shifted and corrected according to the distance difference between them and the reference sensor. The offset time can be calculated according to the formula , where , is the distance from the sensor and the reference sensor to a reference point, such as the start of the pipeline, is the propagation speed of leak sound in the medium, which can be estimated according to the pipeline material and fluid properties. By applying the corresponding time offset, the signals of all sensors are aligned to the same time reference.

[0048] Method two: correlation analysis to estimate time delay: without the need for accurate position parameters, cross-correlation method can also be used to automatically estimate the time delay between sensor pairs. For two sensor signals and , the cross-correlation function is calculated, and the time at which the correlation function peak occurs is found, which is the most likely delay difference between the signals. The signal of sensor is shifted by relative to the signal of sensor to align. By selecting one sensor as a reference and pairing it with the remaining sensors to calculate the correlation, the time delay estimates of all sensors relative to the reference can be obtained. The correlation method can still accurately align the main related signals even in the presence of noise, but when the noise is very strong, the correlation peak may not be obvious, so the known distance method or the use of multi-sensor data redundancy can be used to improve robustness if necessary.

[0049] After time alignment, the data streams of each sensor are aligned on the time axis. If these synchronized signals are superimposed, such as averaging or summing, the leakage signal will be enhanced due to the consistent phase on each channel, while the irrelevant noise will be partially cancelled out due to the random phase on each channel, thereby improving the signal-to-noise ratio of the overall signal. This lays the foundation for the next step of differential noise reduction.

[0050] Differential noise reduction and signal fusion:

[0051] The time-synchronized multi-sensor signals are denoted as , , where N is the number of sensors. These signals contain leakage signal components and background noise components. To further extract the leakage characteristics, the present application introduces differential noise reduction processing based on synchronized signals, that is, using the correlation and difference information between multi-sensor signals to suppress noise and highlight leakage.

[0052] Signal difference calculation: First, calculate the difference between each pair of multi-sensor signals. For example, for adjacent sensor pairs , calculate . The difference signal will eliminate the common part of and . If the background noise is relatively uniform in space, the similar noise components in and are cancelled out by subtraction, while the leakage signal, due to the difference in amplitude or phase at two locations, still remains in . This difference operation is equivalent to a spatial high-pass filter, filtering out signals that are uniformly distributed in space and highlighting signals that change locally.

[0053] Multi-signal fusion: The calculated multiple difference signals are again subjected to fusion processing. One method is to superimpose and sum or average all difference signals. Another method is to use principal component analysis (PCA) or independent component analysis (ICA) to extract the main noise components and signal components of interest from the multi-channel signals . Specifically, the covariance matrix of the multi-sensor signals can be constructed, and its eigenvalues are decomposed; the eigenvector corresponding to the largest eigenvalue represents the main common component, which is mostly noise, and the eigenvector corresponding to the smallest eigenvalue represents the difference component, which may contain the leakage signal. By discarding the main characteristic components and retaining the secondary characteristic components, noise reduction is achieved. For real-time implementation, it can be simplified to summing and averaging the difference signals to reduce the computational load. In this embodiment, the average of the difference between adjacent sensors is taken as the comprehensive difference signal , and the formula is:

[0054]

[0055] where each pair is differenced and summed. In ideal case, the ambient noise is mostly cancelled out, leaving only the local anomaly component caused by the leak.

[0056] Filtering and feature enhancement: after the fusion of the differential signals , frequency band filtering is applied as needed to further improve the signal-to-noise ratio. If it is known that the leak sound is mainly concentrated in a certain frequency band, such as a few hundred Hz to a few kHz, a band-pass filter can be designed to extract the energy of this frequency band and filter out other irrelevant frequency noise. The short-time energy or envelope can also be calculated to smooth out the instantaneous peak interference. After this series of differential and filtering processes, the leak signal features will be significantly prominent, laying the foundation for subsequent pattern recognition.

[0057] In some cases, such as a large number of sensors or a large distribution range, it is also possible not to be limited to simple pairwise differencing, but to use more complex spatial filtering algorithms, such as Beamforming technology, to gain in a specific location. Beamforming achieves signal enhancement from a certain spatial direction or propagation speed by summing multiple sensor signals with appropriate weights, which is equivalent to a generalized time-shift alignment and summation process. If the approximate propagation direction or speed range of the leak wave is known, a beamformer can be designed to have high gain for the leak signal and attenuation for noise in other directions, thereby achieving the purpose of noise reduction. These are extended implementations of the differential noise reduction module.

[0058] Adaptive noise modeling and parameter adjustment: the background noise in industrial sites can change significantly over time, such as equipment start-stop, human interference, environmental wind noise, etc., which can cause noise level fluctuations. To ensure the robustness of the leak detection algorithm to noise changes, the present invention designs an adaptive noise modeling module to estimate the noise online and dynamically adjust the related parameters.

[0059] Noise model initialization: the system can collect a segment of background signal without leak at the start or early stage of operation as a reference sample. By analyzing this sample, an initial noise model is established. For example, the mean , standard deviation , power spectrum distribution , etc. These statistical characteristics constitute the initial noise model, which is used to guide subsequent processing. If it is not possible to obtain a pure background sample in advance, the noise model can also be updated during operation by detecting signal segments that have not appeared for a long time without suspicious events.

[0060] Real-time updating strategy: when the system is continuously running, the noise model will be constantly updated according to new data. Common methods are sliding window statistics or exponential weighted average. For example, for the estimation of background noise energy, the following can be used:​

[0061] where is the current observed total signal energy, is a smoothing factor between 0 and 1, e.g. 0.9, is the updated noise energy estimate. With this recursion, the noise model can gradually reflect the average noise level of the current environment. At the same time, it can also track the changes in the spectrum separately - for example, maintain a background noise power spectrum that updates over time to identify the emergence of new noise components or the disappearance of existing noise components.

[0062] Parameter adaptive adjustment: After the noise model is updated, it will feedback the parameter settings of the previous processing modules to achieve adaptive control. Specifically, it includes:

[0063] Differential noise reduction parameters: If the noise model shows that the current global noise level is rising, the system can increase the weakening strength of the noise subspace in the differential fusion process. For example, increase the weight in the aforementioned sum average of the differential signal, so that more common components are canceled out. If the noise spectrum changes, the passband frequency of the filter can also be adjusted to match the new leakage signal prominent frequency band.

[0064] Detection decision threshold: Before or after deep learning analysis, it is often necessary to set a decision threshold, such as a neural network output probability greater than a certain value to determine a leak. The adaptive noise modeling module can dynamically adjust the threshold size according to the current noise level. When the noise is high, increase the threshold to reduce false positives, and when the noise is low, reduce the threshold to capture weak leaks.

[0065] Deep learning model input adjustment: If the noise spectrum changes significantly, it may cause the performance of the original trained model to decline. The adaptive noise modeling module can trigger adjustments to the model input features, such as increasing the feature quantity for new noise frequency bands, or enabling adaptive batch normalization mechanisms in the model to make the model robust to changes in mean / variance.

[0066] Through the above adaptive mechanisms, the environment noise characteristics can be continuously learned to ensure stable operation under various working conditions. For example, when construction noise temporarily increases, the sensitivity is automatically adjusted to avoid false positives; when the noise decreases after the construction is completed, the high sensitivity is restored to capture potential leaks. Compared with a system with fixed parameters, the adaptive system significantly reduces manual intervention and false positives / negatives.

[0067] Deep learning analysis and leak discrimination:

[0068] After the foregoing pretreatment, the leakage characteristics in the signal have been extracted and enhanced as much as possible. Next, the core deep learning analysis module is entered, and the signal is subjected to pattern recognition by a trained neural network model to determine whether there is leakage.

[0069] Feature extraction and input construction: In order to enable the deep learning model to more efficiently learn the leakage pattern, it is usually necessary to convert the pre-processed signal into a suitable feature input. One method is to use the differential denoised time series as input, and let the neural network extract the time domain features itself. Another more common method is to perform certain transformations on to extract frequency domain or time-frequency domain features, which are then used as network inputs. For example:

[0070] Calculate the short-time Fourier transform STFT of to obtain the time-varying frequency spectrum , and take the amplitude or power to obtain a time-frequency image. This is equivalent to converting a segment of signal into matrix-form data time x frequency, which can be used as an input image for a convolutional neural network. Extract several statistical features, such as energy in multiple frequency bands, peak frequency, spectral centroid, or calculate MFCC mel frequency cepstral coefficients, etc., to obtain a fixed-length feature vector, which is input into a fully connected network or a support vector machine classifier. If a sequence model such as an LSTM network is used, the feature sequence segmented by time frame can be directly input, for example, several dimensions of features calculated every 0.1 seconds. In this way, the LSTM can capture the time evolution pattern.

[0071] In this embodiment, the normalized frequency spectrum of the differential signal is selected as the input of the deep CNN model. The specific method is as follows: perform STFT analysis on with a window length of 1 second to obtain a frequency spectrum per second, and normalize the frequency spectrum amplitude - subtract the mean and divide by the standard deviation, and then input a two-dimensional convolutional neural network.

[0072] 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 convolutional neural network CNN and fully connected layers for binary classification of whether there is leakage. The model mainly includes:

[0073] Convolutional layer + pooling layer: several convolutional layers extract local feature patterns in the input frequency spectrum, such as sudden increases in energy of specific frequency combinations, etc. After each convolution, a pooling layer is connected to reduce the data dimension and increase the invariance. For example, the first layer of convolution kernel size 5x5 extracts primary features; the second layer of convolution kernel size 3x3 extracts higher-level features; and the size is reduced by 2x2 max pooling.

[0074] Fully connected layer: flatten the feature map output from the last convolutional layer, connect one or two layers of fully connected neurons to combine the features extracted by convolution and make classification decision. Use one hidden layer with 128 neurons, activation function ReLU and output layer activation function Sigmoid to output the probability of leakage.

[0075] Dropout regularization: use Dropout strategy to randomly discard part of the neurons in the fully connected layer during training to prevent overfitting and improve the model's generalization ability to new data.

[0076] The entire network contains about several hundred thousand to one million trainable parameters. Cross-entropy loss function and Adam optimizer are used in the training process. Other deep models with different architectures can also be used, such as dual-channel input, which takes two sets of sensor signals as input or time-based RNN / LSTM network to capture temporal dynamics. However, experiments have found that for the differential frequency spectrum features extracted by the present application, CNN has already learned the spatial frequency spectrum features and pattern differences of the leakage very well.

[0077] Model training and validation: before putting the system 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, and should cover multiple noise scenarios. Samples can be obtained by a combination of field collection and simulation: for example, record background noise samples when there is no leakage, and obtain leakage signal samples by making controllable small hole leakage on the experimental pipeline; or use existing public leakage detection data sets and combine them with the data enhancement method of the present application. 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 components generated by the adaptive noise modeling module can be mixed with the leakage signal to expand the training sample distribution.

[0078] After training, the model performance is verified on independent test data, including accuracy, recall rate, false alarm rate, etc. In the test of the embodiment, the deep learning model has achieved excellent detection effect under various noise conditions. For example, the accuracy of small leakage detection in a high noise background in the laboratory is more than 95%, which is much higher than the accuracy of about 80% of the traditional correlation method under the same noise ratio. The model confusion matrix shows that almost all the leakage samples are correctly identified, and only a few false alarms are made on the non-leakage samples.

[0079] Leakage decision and output: the deep learning model outputs a probability value or a classification score for each analysis window indicating the existence of leakage. Set the decision threshold such as 0.5 corresponding to a probability of 50% for the final decision: if If a leak is detected, an alarm is triggered; otherwise, no leak is detected. Under the action of the adaptive noise modeling module, The alarm level can be adjusted based on noise levels, or a dynamic threshold that changes over time can be used to further reduce false alarms. If multiple consecutive windows detect a leak, the alarm level is escalated and corresponding safety measures are implemented, such as automatically shutting off valves and notifying inspection personnel. Simultaneously, the system records the time and possible location of the leak—which sensor group detected it first can provide a rough estimate of which section of the pipeline the leak occurred in—for maintenance personnel's reference.

[0080] Deep learning models can also output richer information, such as multi-class classification to determine the severity level of a leak. For example, the model can be trained to classify leaks as small, medium, and large, outputting different level labels to allow operations personnel to determine appropriate response measures. Within the framework of this invention, by simply adding samples of different leak flow sizes to the training data and assigning corresponding labels, the model can learn to distinguish the scale of a leak based on signal amplitude characteristics. In our implementation, we have added a regression module for leak intensity, enabling the model to estimate the energy ratio of leak noise to background noise, thereby indirectly inferring the size of the leak.

[0081] The workflow of a leak detection method based on multi-sensor time-shift correction and deep learning is as follows:

[0082] Deployment and Initialization: Multiple sensors are deployed on the monitored pipeline and connected to the signal processing unit. Initialization involves collecting background noise samples under leak-free conditions to establish an initial noise model and set initial detection thresholds and other parameters.

[0083] Data Acquisition and Synchronization: Continuously acquire signals from each sensor, timestamp and buffer the data. Whenever a detection cycle arrives—for example, every second—retrieve the digital signal sequence from each sensor within that cycle, calculate the time offset using a preset distance or correlation method, and synchronize the signal sequences.

[0084] Differential analysis and preprocessing: The difference between the synchronization signals is calculated and fused to obtain a signal with enhanced leakage characteristics. .right Perform necessary filtering or feature calculations to extract feature vectors or time-frequency graphs representing the signal characteristics of the current time period.

[0085] Noise Model Update: Update the background noise model parameters based on the latest signal over a period of time. If there are no obvious signs of leakage, it is mainly used to learn the noise; if anomalies are detected, it is decided whether to include them in the model update. Adjust parameters such as the difference algorithm and detection threshold to adapt to the new noise level.

[0086] Deep learning discrimination: input the extracted features into a deep learning model to calculate the leakage probability or classification result. Apply adaptive threshold to make a decision to obtain whether there is leakage and possible confidence, level, etc.

[0087] Result output and alarm: if it is determined that leakage occurs, record the event and trigger an alarm - audible and visual alarm or remote notification, while saving the detailed data of the period for further analysis. If there is no leakage, continue to monitor the next cycle. The above process is continuously cycled to achieve the purpose of real-time online monitoring of leakage.

[0088] The present application can realize early automatic detection of pipeline leakage in a noisy environment. When leakage occurs, the multi-sensor network timely captures weak anomalies and discriminates the leakage signal through intelligent algorithm, and will not miss the alarm due to background noise interference. Those skilled in the art can build an actual system for use or make a mobile leakage monitoring tool kit according to the needs of different equipment monitoring and detection.

Claims

1. A multi-sensor time-lapse correction and deep learning based leak detection method, characterized in that: The method comprises the following steps: a multi-sensor data acquisition module: arranging multiple sensors sensitive to leakage sound waves or vibrations at different positions of the monitored pipeline or container to collect leakage-related signals; a time shift correction module: time aligning the time series signals collected by multiple sensors according to their position differences in space; a differential noise reduction module: comparing and differentially processing the time-aligned multiple sensor signals; An adaptive noise modeling module: modeling and updating the environmental noise in real time, dynamically adjusting the detection parameters according to the noise level, continuously monitoring the statistical characteristics of the background signal, and establishing a noise model; A deep learning analysis module: using a pre-trained deep learning model to perform pattern recognition and classification judgment on the noise-reduced signals; an alarm and display module: triggering an alarm and recording leakage information when a suspected leakage is detected.

2. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 1, wherein: The multi-sensor data acquisition module comprises at least one of the following sensors: an accelerometer, an acoustic emission sensor, or a hydrophone.

3. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 1, wherein: The differential noise reduction module: for the data of multiple sensors, uses the average reduction method or principal component analysis to extract the noise subspace and eliminate the main noise components from the signal.

4. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 1, wherein: The adaptive noise modeling module: the statistical characteristics of the background signal include mean, variance, and spectral distribution; when the environmental noise changes, the weight parameters of the differential noise reduction module are dynamically adjusted or the detection threshold is updated to adapt to the change in noise level.

5. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 1, wherein: The deep learning analysis module: uses a deep neural network, including a convolutional neural network, a long short-term memory network, or a combination thereof as a leakage discrimination model, the network input is the time domain waveform, frequency domain feature or time frequency image of the processed multiple sensor signals; 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 result is output, the judgment result includes at least one of the following: outputting the probability of leakage occurrence, the leakage level or the indication of whether there is leakage.

6. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 1, wherein: The time shift correction module: given the positions of multiple sensors and the sound speed, select the signal of a reference sensor as the reference, correct the time offset of other sensor signals according to their distance difference with the reference sensor, the offset time Δt is calculated by the formula Δt = (dj - di) / v, where dj and di are the distances from sensor j and the reference sensor to the reference point, v is the propagation speed of the leakage sound in the medium, all sensor signals are aligned to the same time reference by applying the corresponding time offset; estimate the time delay by correlation analysis: for two sensor signals xi(t) and xj(t), calculate their cross-correlation function R(τ), find the time τ_max where the correlation function peak value is located, as the most likely time delay difference between the two signals, offset the signal of sensor j relative to the signal of sensor i by τ_max to achieve alignment, select each sensor as a reference in turn, calculate the correlation function with the remaining sensors to obtain the time delay estimation value of all sensors relative to the reference sensor; combine the results of the known distance calculation method or use the redundancy information of multiple sensor data to improve the robustness of the time shift correction result.

7. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 3, wherein: Differential noise reduction module: the time-aligned multiple sensor signals are differentially processed to obtain multiple differential signals; then the differential signals are fused and reduced in noise by the following two ways: (1) all differential signals are superimposed and summed or averaged; (2) principal component analysis or independent component analysis is used to extract the main common noise components and the differential signal components of interest from the multi-channel differential signals, and the noise reduction is realized by discarding the main noise components and retaining the secondary signal components to obtain a comprehensive differential signal D(t), and the comprehensive differential signal D(t) is further processed for signal enhancement, including: applying frequency band filtering to improve the signal-to-noise ratio, adjusting the passband of the filter to extract the energy of the frequency band where the leakage signal is located and filter out irrelevant noise frequencies; calculating the short-time energy or envelope of D(t) to smooth the instantaneous peak interference; and in the case of a large number of sensors or a wide distribution range, the beamforming technique is used to enhance the spatial gain of the leakage signal and attenuate the noise in other directions, thereby further improving the signal-to-noise ratio.

8. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 4, wherein: The adaptive noise modeling module includes initialization and dynamic updating of the noise model: at system startup or initial operation, a background signal without leakage is collected as a reference sample, and the initial statistical characteristics of the background noise, including mean, standard deviation and power spectrum distribution, are calculated by analyzing the sample to establish an initial noise model; if a pure background sample cannot be obtained in advance, a signal segment with no suspicious events for a long time is detected during system operation to establish or update the noise model, and the noise model is continuously updated based on newly entered data, and a sliding window statistical method or an exponential weighted average method is used to recursively calculate the noise parameters; for the estimation of the background noise energy, the exponential weighted average formula is used for updating, so that the noise model gradually reflects the changes in the background noise level in the current environment, and the changes in the noise signal spectrum characteristics are tracked, and a background noise power spectrum P(f, t) that is updated over time is maintained, which is used to identify the appearance of new noise frequency components or the disappearance of original noise components.

9. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 8, wherein: The updated noise model is used to adaptively adjust the parameters of each detection module to improve the adaptability of the system to changes in the noise environment, including: when the current background noise level increases, the suppression strength of the differential noise reduction module on the noise subspace is enhanced, and the noise weight coefficient in the summation and averaging of the differential signals is increased to offset more common noise components; when the main frequency band of the leakage signal changes, the passband frequency of the filter is adjusted accordingly to match the new leakage signal frequency band, thereby ensuring the effective extraction of the leakage characteristics by the differential noise reduction; the threshold of the deep learning discrimination module is dynamically adjusted based on the noise model, the decision threshold is increased to reduce false positives when the background noise is high, and the threshold is reduced to capture weak leaks when the background noise is low; when the spectral characteristics of the environmental noise change significantly, the update of the input features of the deep learning model is triggered, the feature quantities for the newly appeared noise frequency band are increased, or the adaptive batch normalization mechanism in the model is enabled, so that the model remains robust to changes in the mean and variance of the input data.

10. The multi-sensor time-lapse correction and deep learning based leak detection method of claim 5, wherein: The preprocessed signal is subjected to feature conversion to generate features suitable for the input of a deep learning model: firstly, the differential denoised time series D(t) is used as the input of the deep learning model, and the time domain feature pattern of the leakage signal is automatically extracted by the model; secondly, the sequence D(t) is subjected to frequency domain or time-frequency domain transformation, and the frequency domain parameters or time-frequency images representing the features of the leakage signal are extracted, and the extracted features are used as the input of the deep learning model, the deep learning model including a convolutional neural network for extracting features in the input signal, and a layer of fully connected neurons for combining the extracted features and outputting a classification result, so as to determine whether there is leakage.

Images