Leakage on-line identification technology under complex noise background condition
By employing multi-stage denoising and feature extraction, combined with KL divergence and mean teacher models, the accuracy and robustness issues of pipeline leak detection under complex noise backgrounds are addressed, achieving efficient leak identification.
Patent Information
- Application Number
- CN202511519685.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-23
- Publication Date
- 2026-02-17
AI Technical Summary
Existing pipeline leak detection methods are inefficient and inaccurate in complex noise environments, struggle to effectively utilize time-domain, frequency-domain, and time-frequency characteristics, and rely on human experience, resulting in poor adaptability.
A multi-stage denoising strategy is adopted, combining low-pass filtering, wavelet denoising, and singular value decomposition to construct a feature matrix and extract time-domain, frequency-domain, and time-frequency-domain features. Features are selected using KL divergence and semi-supervised learning is performed using the mean teacher model to achieve leakage identification.
It significantly improves the signal-to-noise ratio of pipeline acoustic signals, ensures the accuracy and stability of feature extraction, enhances the accuracy and robustness of leakage detection, and has good prospects for engineering applications.
Smart Images
Figure CN121545547A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of pipeline leakage monitoring and intelligent signal processing, and particularly relates to a leakage online identification process under complex noise background conditions. BACKGROUND
[0002] The water supply pipe network system is a large and complex network system, mainly including pipes, water pumps, water towers, water storage tanks and the like. Leakage of the water supply pipe will cause waste of water resources, and at the same time, will cause some secondary disasters, such as damage to roadbeds and building foundations, infiltration of pollutants and minerals and the like, so it is of great significance to timely and accurately detect the water leakage point of the water supply pipe.
[0003] The current pipe materials include steel pipes, cast iron, ductile cast iron, concrete and plastic, among which the use of environmentally friendly and sanitary plastic pipes is increasingly widespread. However, since the sound wave signal in the plastic pipe will be rapidly weakened, a more efficient and accurate method is needed for pipe leakage detection by sound signals.
[0004] The existing pipe network leakage detection devices include the following: sound stick: divided into metal stick and wooden stick, the biggest advantage is easy to carry, and has relatively sensitive leakage detection performance, and is not restricted by the test environment. During detection, the sound stick is placed at one end on the ground, and the tester determines whether the water pipe is leaking through the sense of hearing; electronic amplification leak detector: through analog signal processing or digital signal processing, the digital frequency analysis, minimum value and instantaneous value analysis are completed; tracer leak detector: a certain concentration of gas is injected into the pipe. If there is leakage in the pipe, the pressure difference between the inside and outside of the pipe will occur. The position of the ground where the gas concentration increases is the position of the leakage; related detection instrument: a detection device is installed at both ends of the pipe to collect the water leakage signals in the pipe, and then the signals are transmitted to the related calculation equipment to determine whether the pipe is leaking; hydrophone sensor: detecting the friction sound of the leakage in water, and converting the sound into an electric signal.
[0005] Although the principle of the traditional sound listening method is relatively simple and widely applied, it completely depends on the experience of the listener and the degree of water leakage, and the method steps are complicated, the manpower scheduling is large, and the efficiency is low. Moreover, the method of detecting pipe network leakage according to the physical way is not ideal for noise processing in the sound signal. At the same time, the importance of time domain feature selection, frequency domain feature selection and time-frequency feature selection is ignored, resulting in great defects in the traditional method. SUMMARY
[0006] The purpose of the present application is to solve the problems existing in the prior art, and to provide a leakage online identification process under complex noise background conditions.
[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:
[0008] A leakage online identification process under complex noise background conditions, comprising:
[0009] S1, multi-stage denoising; in order to solve the problems of large noise interference, low feature extraction quality and weak adaptability in the traditional pipe network sound signal detection method, a multi-stage denoising preprocessing strategy is adopted as the basic link of the whole sound signal recognition system, which significantly improves the accuracy and robustness of subsequent feature expression. Specifically, a three-stage denoising process of low-pass filtering, wavelet denoising and singular value decomposition denoising is adopted;
[0010] S2, feature matrix construction:
[0011] The abnormal sound signal generally lasts for 2-3 seconds, L=3 seconds, S=1 second, the total length of the whole signal is T seconds, and the number of sub-sections N is
[0012]
[0013] [·] represents rounding down; the continuous time sound signal x svd (t) after denoising is divided into N frame sections: Feature matrix:
[0014]
[0015] Wherein represents the feature vector of the i-th frame section;
[0016] S3, feature selection, from different time frames t n and frequency points f k the following features are extracted:
[0017] (1) time domain feature: amplitude a∈R 1×N
[0018] (2) frequency domain feature:
[0019] Fourier transform of signal
[0020] Power spectral density p PSD ∈R 1×N
[0021]
[0022] Marginal spectrum m s ∈R 1×N
[0023] Hilbert-Huang transform is performed on to obtain Hilbert spectrum H i (fk t), compute marginal spectrum
[0024] where f k is the frequency sample point, t j is the time sample point;
[0025] spectrum centroid c s ∈ R 1×N
[0026]
[0027] spectrum centroid bandwidth b s ∈ R 1×N
[0028]
[0029] higher-order spectrum h s ∈ R 1×N
[0030]
[0031] (3) time-frequency domain feature: time-frequency spectrum s tf ∈ R 1×N is the energy distribution at different time frames and frequency points;
[0032] Mel-frequency cepstral coefficient m ∈ R 1×N
[0033]
[0034] F -1 [·] is the inverse Fourier transform, and the final required Mel-frequency cepstral coefficient m i is obtained.
[0035] Through the formulas of time domain features, frequency domain features, and Mel-frequency cepstral coefficients, the features are extracted to form a feature vector:
[0036] F(j,:), integer j ∈ [1, N];
[0037] where
[0038]
[0039] Feature selection based on KL divergence
[0040] To quantitatively evaluate the separability of the leakage and loss sample distribution p(F[:,m]) and the non-leakage and loss sample distribution q(F[:,m]) on the mth feature, the KL divergence is used to calculate the distribution difference:
[0041]
[0042] where F[:,m] represents all sample feature values corresponding to the mth feature dimension in the feature matrix p(F[:,m]) and q(F[:,m]) represent the probability density distributions of the loss samples and non-loss samples in this feature dimension, respectively;
[0043] The KL divergence values D KL (p||q) of the loss samples and non-loss samples obtained by the time domain features, frequency domain features, and mel frequency cepstrum coefficients are sorted as follows:
[0044]
[0045] The final feature matrix F is constructed * = D KL [1,2,3,4,5]∈R N×5 , where
[0046]
[0047] S4, loss judgment:
[0048] After obtaining the feature matrix F * , a semi-supervised classification learning of loss and non-loss acoustic signals is performed using a mean teacher model, which is represented as a M(·) function composed of a student model f θ (·) and a teacher model f θ′ (·), where the parameters θ' of the teacher model are updated by an exponential moving average of the student model parameters θ without artificial assignment:
[0049]
[0050] During the training process, the prediction results of the feature by the teacher model F * are used as pseudo labels to guide the student model learning, and the consistency of the outputs of the two is constrained by a consistency loss:
[0051] L cons =||f θ (F*)-f θ′ (F*)|| 2
[0052] The teacher model is used for prediction, and the output results are binarized to obtain the final loss judgment value:
[0053]
[0054] where the indicator function
[0055]
[0056] Therefore respectively represent non-leakage and leakage states, the following steps are normal if the result is 0, and leakage if the result is 1.
[0057] Preferably, low-pass filter noise reduction:
[0058] In order to effectively retain the main spectral information of the pipeline leakage sound signal and suppress the middle and high frequency background noise, the cut-off frequency of the low-pass filter is set to f c = 300 Hz; the reason for selecting 300 is that a large number of sample spectrum analysis shows that the leakage signal energy is mainly concentrated in 300 Hz, which is the best compromise point of energy concentration degree and noise suppression ability, and experiments also show that the signal-to-noise ratio is improved the most at this cut-off frequency, and the subsequent classification accuracy is higher, which has good practicability and adaptability;
[0059] The continuous original sound signal x(t)
[0060] is low-pass filtered to filter out high-frequency noise and obtain a preliminary noise reduction signal
[0061] x LF (t) = LPF(x(t)).
[0062] Preferably, wavelet denoising:
[0063] The preliminary noise reduction signal x LF (t) is decomposed by wavelet transform with a total of m layers
[0064]
[0065] A m (t) represents the mth layer approximation coefficient of the low-frequency part, and D m (t) represents the mth layer detail coefficient of the high-frequency part, and for each layer of detail coefficient D n (t), an integer n ∈ [1, j] applies a threshold function Γ(·)
[0066]
[0067] After that, the approximation coefficient A m (t) and the threshold-processed detail coefficient are inverse wavelet transformed to obtain a second-stage noise reduction signal
[0068]
[0069] Preferably, hetero-value decomposition denoising:
[0070] The wavelet denoised signal x WF(t) further performing singular value decomposition denoising, by processing singular values of the signal, that is, retaining larger singular values and suppressing smaller singular values, the noise part in the signal can be removed, and a final denoised signal is obtained
[0071] x svd (t) = SVD(x WF (t))
[0072] Wherein SVD(·) represents the singular value decomposition denoising process.
[0073] Compared with the prior art, the beneficial effects of the present application are: by constructing a multi-stage denoising mechanism, the signal-to-noise ratio of the pipeline sound signal in a complex environment is significantly improved, ensuring the accuracy and stability of subsequent feature extraction; on this basis, the KL divergence evaluation mechanism is used to select modeling features with strong discrimination ability among time domain, frequency domain and time-frequency domain features, to realize accurate expression of leakage and loss features; finally, the mean teacher model is introduced after the commonly used classification model to enhance semi-supervised learning, which effectively utilizes unlabeled data while enhancing the generalization ability and environmental adaptability of the model;
[0074] The present application has significant technical effects in improving the accuracy, robustness and intelligent level of the pipeline network leakage detection system, and has good engineering application prospect and popularization value. BRIEF DESCRIPTION OF DRAWINGS
[0075] Figure 1 A pipeline network leak detection system flow chart of a leakage online identification process under complex noise background conditions is proposed. DETAILED DESCRIPTION
[0076] Referring to Figure 1 A leakage online identification process under complex noise background conditions comprises the following steps:
[0077] S1, multi-stage denoising; in order to solve the problems of large noise interference, low feature extraction quality and weak adaptability in the traditional pipeline sound signal detection method, a multi-stage denoising preprocessing strategy is adopted as the basic link of the entire sound signal recognition system, which significantly improves the accuracy and robustness of subsequent feature expression, specifically, a three-stage denoising process of low-pass filtering, wavelet denoising and singular value decomposition denoising is adopted.
[0078] A multi-stage denoising mode including low-pass filtering, wavelet denoising and singular value decomposition is constructed, which gradually eliminates different frequency band and different type of noise interference for the original pipeline sound signal, effectively improves the signal quality, and provides higher signal-to-noise ratio of basic features for subsequent feature extraction, and the multi-stage denoising mode can adapt to non-stationary, weak signal, high noise background and other complex environments, and is the key guarantee of system robustness
[0079] S2, Feature matrix construction:
[0080] The abnormal acoustic signal usually lasts for 2-3 seconds, taking L=3 seconds and S=1 second, the total length of the whole signal is T seconds, and the number of sub-sections N is
[0081]
[0082] [·] represents rounding down; the continuous time acoustic signal x svd (t) after denoising is divided into N frame sections: Feature matrix:
[0083]
[0084] wherein represents the feature vector of the i-th frame section;
[0085] S3, Feature selection, from different time frames t n and frequency points f k the following features are extracted:
[0086] (1) Time domain feature: amplitude a∈R 1×N
[0087] (2) Frequency domain feature:
[0088] Fourier transform of the signal
[0089] Power spectral density p PSD ∈R 1×N
[0090]
[0091] Marginal spectrum m s ∈R 1×N
[0092] Hilbert-Huang transform is performed on to obtain the Hilbert spectrum H i (f k ,t), and the marginal spectrum is calculated
[0093] wherein f k is the frequency sampling point, and t j is the time sampling point;
[0094] Spectral centroid c s ∈R 1×N
[0095]
[0096] Spectral centroid bandwidth b s ∈R 1×N
[0097]
[0098] Higher-order spectrum h s ∈R 1×N
[0099]
[0100] (3) Time-frequency domain characteristics: time spectrum s tf ∈R 1×N It refers to the energy distribution at different time frames and frequency points;
[0101] Mel frequency cepstral coefficients m∈R 1×N
[0102]
[0103] F -1 [·] represents the inverse Fourier transform, yielding the final required Mel-frequency cepstral coefficients m. i ;
[0104] Extracting the time-domain features, frequency-domain features, and formulas for Mel-frequency cepstral coefficients. Features, forming a feature vector:
[0105] F(j,:), where j∈[1,N];
[0106] in
[0107]
[0108] This paper proposes a construction approach that integrates time-domain, frequency-domain, and time-frequency-domain features. Combining statistical standards such as peak value, smoothness, average energy, and energy concentration, KL divergence analysis is performed on each type of feature between leaked and non-leaked samples to quantitatively evaluate the distinguishing ability. The paper also selects highly separable feature combinations required to construct the classification model. This feature selection can effectively enhance the accuracy of the intelligent leak identification model for official website leaks.
[0109] Feature selection based on KL divergence
[0110] To quantitatively evaluate the difference between the distribution of leaky samples p(F[:,m]) and the distribution of non-leaky samples q(F[:,m]) on the m-th feature. Separability is determined by calculating the distributional differences using KL divergence:
[0111]
[0112] Where F [:, m] represents all sample eigenvalues corresponding to the mth eigen dimension in the feature matrix p(F [:, m]) and q(F [:, m]) represent the probability density distribution of the loss sample and the non-loss sample on the eigen dimension, respectively;
[0113] The KL divergence values D of the loss sample and the non-loss sample obtained by the formula of the time domain feature, the frequency domain feature and the mel frequency cepstrum coefficient KL (p||q) are sorted as follows:
[0114]
[0115] The final feature matrix F is constructed * = D KL [1, 2, 3, 4, 5]∈R N×5 , wherein
[0116]
[0117] An end-to-end identification system composed of a multi-stage denoising module, a feature extraction and selection module, and a classification module (a preliminary identification model and a mean teacher enhancement model) is constructed, the overall process is closely linked, realizes accurate and rapid judgment of the pipe network loss signal under the condition of limited labeled samples, and has good engineering practicability and intelligent characteristics
[0118] S4, loss judgment:
[0119] The feature matrix F is obtained * After that, the mean teacher model is used for semi-supervised classification learning of the loss and non-loss sound signals, and the mean teacher model is represented as the M(·) function, which is composed of the student model f θ (·) and the teacher model f θ′ (·), wherein the parameters θ' of the teacher model are updated by the exponential moving average method of the student model parameters θ, without artificial assignment:
[0120]
[0121] In the training process, the prediction result of the feature by the teacher model F * is used as a pseudo label to guide the student model learning, and the consistency of the outputs of the two is constrained by the consistency loss:
[0122] L cons =||f θ (F*)-f θ′ (F*)|| 2
[0123] The teacher model is used for prediction, and the output result is binarized to obtain the final loss determination value:
[0124]
[0125] where the indicator function
[0126]
[0127] Thus respectively represent non-leakage and leakage states, the following steps are normal if the result is 0, if the result is 1, leakage;
[0128] Low-pass filter noise reduction:
[0129] In order to effectively retain the main spectral information of the pipeline leakage sound signal and suppress the background noise of the middle and high frequencies, the cut-off frequency of the low-pass filter is set to f c = 300 Hz; The reason for selecting 300 is that a large number of sample spectrum analysis shows that the leakage signal energy is mainly concentrated in 300 Hz, which is the best compromise point of energy concentration degree and noise suppression ability. Experiments also show that the signal-to-noise ratio is improved the most at this cut-off frequency, the subsequent classification accuracy is higher, and it has good practicability and adaptability.
[0130] The continuous original sound signal x(t)
[0131] is low-pass filtered to filter out high-frequency noise and obtain a preliminary noise reduction signal
[0132] x LF (t) = LPF(x(t));
[0133] Low-pass filter noise reduction:
[0134] In order to effectively retain the main spectral information of the pipeline leakage sound signal and suppress the background noise of the middle and high frequencies, the cut-off frequency of the low-pass filter is set to f c = 300 Hz; The reason for selecting 300 is that a large number of sample spectrum analysis shows that the leakage signal energy is mainly concentrated in 300 Hz, which is the best compromise point of energy concentration degree and noise suppression ability. Experiments also show that the signal-to-noise ratio is improved the most at this cut-off frequency, the subsequent classification accuracy is higher, and it has good practicability and adaptability.
[0135] The continuous original sound signal x(t)
[0136] is low-pass filtered to filter out high-frequency noise and obtain a preliminary noise reduction signal
[0137] x LF (t) = LPF(x(t));
[0138] Odd value decomposition noise reduction:
[0139] The wavelet denoised signal xWF (t) Further singular value decomposition denoising is performed, noise in the signal is removed by processing singular values of the signal, that is, larger singular values are retained and smaller singular values are suppressed, and a final denoised signal is obtained
[0140] x svd (t) = SVD(x WF (t))
[0141] Wherein SVD(·) represents a singular value decomposition denoising process.
[0142] In the application, a multi-stage denoising mode including low-pass filtering, wavelet denoising and singular value decomposition is constructed, different frequency bands and different types of noise interference in the original pipeline sound signal are gradually eliminated, the signal quality is effectively improved, and higher signal-to-noise ratio basic features are provided for subsequent feature extraction, the multi-stage denoising mode can adapt to non-stationary, weak signal, high noise background and other complex environments, and is a key guarantee for system robustness.
[0143] A construction idea of fusing time domain, frequency domain and time-frequency domain features is proposed, KL divergence analysis is performed on each type of feature between leakage and non-leakage samples in combination with statistical standards such as peak value, smoothness, average energy and energy concentration, the discrimination ability is quantitatively evaluated and distinguished, and a high separability feature combination required by a classification model is selected and constructed, the feature selection can effectively enhance the accuracy of the intelligent leakage identification model on online leakage;
[0144] An end-to-end identification system composed of a multi-stage denoising module, a feature extraction and selection module and a classification module (a preliminary identification model and a mean teacher enhanced model) is constructed, the overall process is closely connected, precise and rapid judgment of pipeline leakage signals under the condition of limited labeled samples is realized, and the system has good engineering practicability and intelligent characteristics.
Claims
1. A process for on-line leak identification in complex noise background conditions, characterized in that, Comprising the following steps: S1, multi-stage denoising; in view of the problems of large noise interference, low feature extraction quality and weak adaptability in the traditional pipe network sound signal detection method, a "multi-stage denoising" preprocessing strategy is adopted as the basic link of the whole sound signal recognition system, which significantly improves the accuracy and robustness of the subsequent feature expression. Specifically, a three-stage denoising process of low-pass filtering, wavelet denoising and singular value decomposition denoising is adopted; S2, feature matrix construction: The abnormal sound signal generally lasts for 2-3 seconds, L=3 seconds, S=1 second, the total length of the whole signal is T seconds, and the number of sub-sections N is [·] represents rounding down; the continuous-time acoustic signal x svd (t) is divided into N frame segments: Feature matrix: wherein represents the feature vector of the i-th frame segment; S3, feature selection, from different time frames t n and frequency points f k the following features are extracted respectively: (1) Time domain feature: amplitude a e R 1×N (2) Frequency domain features: Fourier transform of the signal Power spectral density p PSD ∈ R 1×N marginal spectrum m s ∈ R 1×N right Perform the Hilbert-Huang transform to obtain the Hilbert spectrum H. i (f k ,t), calculate the marginal spectrum where f k is the frequency sample point, and t j is the time sample point. Spectral centroid c s ∈ R 1×N Spectral centroid bandwidth b s ∈ R 1×N Higher order spectrum h s ∈ R 1×N (3) Time-frequency domain features: time-frequency spectrum s tf ∈R 1×N is the energy distribution at different time frames and frequency points; Mel-frequency cepstral coefficients m e R 1×N F -1 [·] is the inverse Fourier transform, resulting in the final Mel-frequency cepstral coefficients m i ; Features are extracted by time domain features, frequency domain features and mel-frequency cepstral coefficients formula features, The feature vector is composed of: F(j,:), integer j∈[1,N]; Where Feature selection based on KL divergence To quantitatively evaluate the distribution of leakage samples p(F[ :, m]) and the distribution of non-leakage samples q(F[ :, m]) on the mth feature The separability is calculated by the KL divergence of the distribution difference: where F[ :, m] represents all sample eigenvalues corresponding to the mth eigen dimension in the feature matrix p(F[ :, m]) and q(F[ :, m]) represent the probability density distribution of the loss sample and the non-loss sample respectively on the eigen dimension. The KL divergence values D of the leaky and non-leaky samples obtained from the formula of the time domain features, the frequency domain features, and the mel-frequency cepstral coefficients KL The (p||q) ordering is as follows: Constructing the final feature matrix F * = D KL [1, 2, 3, 4, 5] e R N×5 wherein S4, leakage and loss judgment: obtaining a feature matrix F * After that, the semi-supervised classification learning of the missing and non-missing acoustic signals is performed by using the mean teacher model, and the mean teacher model is represented as a M(·) function composed of a student model f θ (·) and a teacher model f θ′ (·). The parameters θ' of the teacher model are updated by the exponential moving average of the parameters θ of the student model, without artificial assignment. During training, the teacher model F * The prediction results of the features are used as pseudo labels to guide the student model to learn, and the consistency of the outputs of the two is constrained through a consistency loss: L cons =||f θ (F*)-f θ′ (F*)|| 2 The output result is binarized using the teacher model to obtain the final leakage and loss judgment value: Where the indicator function Thus 0 = no loss, 1 = loss.
2. The process of claim 1, wherein, Low-pass filter denoising: In order to effectively retain the main spectral information of the pipeline leakage sound signal and suppress the middle and high frequency background noise, the cut-off frequency of the low-pass filter is set to f c = 300 Hz; the reason for selecting 300 is that a large number of sample spectrum analyses show that the leakage signal energy is mainly concentrated in 300 Hz, which is the best compromise point of energy concentration degree and noise suppression ability, experiments also show that the signal-to-noise ratio is improved most at this cut-off frequency, the subsequent classification accuracy is higher, and good practicability and adaptability are possessed. The continuous original sound signal x(t) is subjected to low-pass filtering to filter out high-frequency noise and obtain the preliminary denoised signal x LF (t) = LPF(x(t)).
3. The process of claim 1, wherein, Wavelet denoising: to the preliminary noise-reduced signal x LF (t) performing a wavelet transform with m layers of decomposition A m (t) represents the m-th layer approximation coefficient of the low frequency part, D m (t) represents the m-th layer detail coefficient of the high frequency part, D n (t) for each layer, an integer n e [1, j] applies a threshold function Γ(·) After this, the approximation coefficients A m (t) and the thresholded detail coefficients performing an inverse wavelet transform to obtain a second stage denoised signal 4. The process of claim 1, wherein, Singular value decomposition denoising: x is the signal after wavelet denoising WF (t) further singular value decomposition denoising, by processing the singular values of the signal, that is, retaining larger singular values and suppressing smaller singular values, the noise part in the signal can be removed, and the final denoised signal is obtained x svd (t) = SVD(x WF (t)) Where SVD(·) represents the singular value decomposition denoising process.