A method for detecting pipeline leaks by integrating multi-dimensional time-frequency domain features of vibration signals

Abstract

This invention provides a pipeline leak detection method that integrates multi-dimensional time-frequency domain features of vibration signals, relating to pipeline leak detection in complex pipeline network monitoring. This method utilizes wavelet packet-based EEMD (Extremely Employable Epipolar Dynamics) fast spectral kurtosis joint analysis for environmental interference separation to remove noise from the signal. Further data filtering is achieved by combining wavelet packet and fast spectral kurtosis joint analysis to separate and denoise the vibration signal from environmental interference. Feature extraction is then performed from the time domain, frequency domain, and time-frequency domain to obtain feature sequences. Finally, a model based on stacked sparse autoencoders is used. After pre-training, the reconstruction error is used as a monitoring statistic, and the reconstruction of each sensor is obtained based on the variational structure. The perceived health status is determined using a kernel density estimation method. This method effectively solves noise interference, promptly captures the dynamic changes in pipeline status, reduces false alarms and missed alarms when pipeline leaks occur, and improves the leak detection performance of the pipeline health monitoring system.

Description

Technical Field

[0001] This invention relates to the field of pipeline leakage detection during the monitoring of complex pipeline networks, and in particular to a pipeline leakage detection method that integrates multi-dimensional time-frequency domain features of vibration signals. Background Technology

[0002] Pipeline transportation is widely used to transport essential energy sources such as natural gas, oil, and water. Over time and with the increasing complexity of pipeline usage scenarios, pipelines naturally corrode or develop fatigue cracks, leading to leaks. This results in negative social impacts such as energy loss, environmental pollution, and economic losses, and may even endanger human lives. To ensure the efficient and stable operation of pipeline networks, real-time pipeline monitoring systems are crucial. Pipeline health status monitoring systems can monitor pipeline operating status in real time and detect leaks promptly. However, sensors are susceptible to external noise interference during pipeline health monitoring, making it difficult to capture the dynamic changes in pipeline conditions and distinguish between interference and leaks. Therefore, eliminating noise interference, improving leak detection performance, and minimizing false alarms and missed alarms are important research issues in pipeline leak detection methods.

[0003] Currently, pipeline leak detection can be mainly divided into three steps: signal denoising, feature extraction, and anomaly detection. (1) Signal denoising is an important part of data preprocessing in the leak detection framework. The quality of the signal seriously affects the feature extraction part. Since environmental noise is ubiquitous, the signal data collected by the sensor will be affected by noise, which will reduce the signal-to-noise ratio and make the extracted features less discriminative and difficult to distinguish the pipeline operating status. (2) Signal feature extraction is an indispensable step in the leak detection framework. It provides valuable information support for the subsequent combination of detection models to distinguish between normal and leaking states, thereby significantly improving the leak detection performance. At present, most signal feature extraction methods are based on the time domain and frequency domain. Time domain and frequency features are most suitable for reflecting static change characteristics, but for dynamic changes, these two features are often difficult to capture dynamic changes. (3) Pipeline leak detection can be regarded as detecting an anomaly. In actual pipeline health status monitoring, leaks are rare, and it is difficult to obtain data on the leaking state. In most cases, the pipeline is operating in a healthy state. At present, anomaly detection methods based on clustering and classification are more common. Such detection methods require a large amount of data with anomaly labels and are difficult to apply to such scenarios.

[0004] Therefore, the vibration signals contained noise during the monitoring of complex pipeline networks, and the difficulty in capturing the dynamic changes in pipeline conditions, are technical problems that urgently need to be solved. Summary of the Invention

[0005] To address the challenges of noise-laden vibration signals and the difficulty in capturing dynamic changes in pipeline conditions during complex pipeline monitoring, this invention provides a pipeline leak detection method that integrates multi-dimensional time-frequency domain features of vibration signals, particularly relevant in industrial pipelines with leaks and signal interference. The method primarily includes:

[0006] S1: Use a filter to remove power frequency interference from the original signal S, and obtain the signal matrix after removing power frequency interference.

[0007] S2: Using the EEMD decomposition method to decompose the signal matrix The IMF components are decomposed to obtain IMF components, and then filtered to obtain effective IMF components.

[0008] S3: The selected IMF components are decomposed using wavelet packets. The energy distribution of each component is analyzed, and the corresponding optimal frequency bands in the spectral kurtosis map are used for component reconstruction. The denoised signal is obtained by combining the reconstructed components.

[0009] S4: Based on the noise-reduced signal matrix obtained in step S3 The sliding window method is used to dynamically capture characteristic changes. For the signal within the window, feature sequences corresponding to time-domain features, frequency-domain features, information entropy features, and time-frequency domain features are extracted to obtain time-domain feature statistics. and frequency domain characteristic statistics And the vibration sensor channel signal Extracting the information entropy feature statistics yields the following results. and time-frequency domain features

[0010] S5: A pipeline leak detection model is built using a stacked sparse autoencoder, and the model is trained. The reconstruction error is used as a detection statistic, and the reconstruction error J is calculated by comparing the predicted and tested values ​​obtained from each sensor. SSAE Then, the confidence level α is designed using the Gaussian kernel density estimation method, and the set threshold J is obtained. α By comparing the reconstruction error J SSAE With the set threshold J α Anomaly detection is performed on the size of the interval; if the reconstruction error J is... SSAE Greater than the set threshold J α If so, a pipeline leak will be detected.

[0011] A storage device that stores instructions and data for implementing a pipeline leak detection method that integrates multi-dimensional time-frequency domain features of vibration signals.

[0012] A pipeline leak detection device that integrates multi-dimensional time-frequency domain features of vibration signals includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a pipeline leak detection method that integrates multi-dimensional time-frequency domain features of vibration signals.

[0013] The beneficial effects of the technical solution provided by this invention are as follows: For raw data collected from actual industrial processes, this invention uses feature extraction methods to extract trend features of the raw signals; a deep learning algorithm model is used to reconstruct the extracted signal data, and anomalies are judged based on the reconstructed data. To solve the problem of signal interference from external factors, this invention utilizes a wavelet packet-based EEMD fast spectral kurtosis joint analysis environmental interference separation method to separate noise in the signal. By combining wavelet packet and fast spectral kurtosis joint analysis for further data filtering, environmental interference separation and denoising of vibration signals are achieved. A multivariate video feature extraction method is proposed, extracting features from the time domain, frequency domain, and time-frequency domain to obtain feature sequences. Finally, a model based on stacked sparse autoencoders is used. After pre-training, the reconstruction error is used as a monitoring statistic, and the reconstruction of each sensor is obtained according to the variational structure. The health status is judged and anomalies are detected by kernel density estimation method. This effectively solves noise interference, can timely capture the dynamic changes in pipeline status processes, and reduces the false alarm and missed alarm rates when pipeline leaks occur, thus improving the leak detection performance of the pipeline health monitoring system. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0015] Figure 1 This is a flowchart of a pipeline leakage detection method that integrates multi-dimensional time-frequency domain features of vibration signals in an embodiment of the present invention.

[0016] Figure 2 This is a framework diagram of the pipeline leakage detection method that integrates multi-dimensional time-frequency domain features of vibration signals in an embodiment of the present invention.

[0017] Figure 3 This is a framework diagram of a pipeline leak detection and alarm based on a sparse stacked autoencoder in an embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram of the time domain, frequency domain, entropy value, and time-frequency feature matrix image in an embodiment of the present invention.

[0019] Figure 5 This is a schematic diagram of an image used in an embodiment of the present invention to calculate the threshold of detection statistical indicators based on kernel density estimation.

[0020] Figure 6 These are the experimental results of three groups of leak detection experiments: AE, StAE, SpAE, and SSAE, as shown in the embodiments of this invention.

[0021] Figure 7 This is a schematic diagram of the hardware device working in an embodiment of the present invention. Detailed Implementation

[0022] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] Please refer to Figure 1-2 , Figure 1-2 These are a flowchart and a framework diagram of a pipeline leakage detection method that integrates multi-dimensional time-frequency domain features of vibration signals, as described in an embodiment of the present invention. Specifically, it includes:

[0024] S1: Use a filter to remove power frequency interference from the original signal S, and obtain the signal matrix after removing power frequency interference.

[0025] S2: Using the EEMD decomposition method to decompose the signal matrix The IMF components are decomposed to obtain IMF components, and then filtered to obtain effective IMF components.

[0026] S3: The selected IMF components are decomposed using wavelet packets. The energy distribution of each component is analyzed, and the corresponding optimal frequency bands in the spectral kurtosis map are used for component reconstruction. The denoised signal is obtained by combining the reconstructed components.

[0027] The specific steps for the above three steps are as follows:

[0028] To address the issue of environmental noise interference in the collected vibration signal, it is necessary to first process the power frequency interference; otherwise, it will affect subsequent environmental interference separation. Because the Butterworth filter has the smoothest passband for the same filter order, it is used as a tool for removing power frequency interference. The original signal S is then filtered for power frequency interference to obtain... Then, the EEMD decomposition method is selected to add white noise to the original signal. Decomposition yields multiple intrinsic mode function (IMF) components:

[0029]

[0030] in, It is the k-th IMF component after the i-th addition of white noise. The residual of the signal is used to represent the average trend of the signal. Signal matrix The j-th sensor signal, For signal The i-th disturbance signal.

[0031] To better separate environmental interference and obtain effective components, the IMF components need to be filtered after obtaining them. Wavelet packet decomposition and fast spectral kurtosis analysis are then used to reconstruct new components, achieving signal denoising. The fast spectral kurtosis analysis algorithm used in this invention is based on multi-level spectrum division after STFT transformation, then using the spectral kurtosis formula to calculate the spectral kurtosis of the signal. It requires first normalizing the IMF1 and the original signal to 0-1 before calculating the corresponding spectral kurtosis and correlation coefficient.

[0032] The formula for calculating kurtosis is as follows:

[0033]

[0034] The formula for calculating the correlation coefficient is as follows:

[0035]

[0036] in, For the k-th IMF component and the original signal The correlation coefficient between them Defined as a signal The j-th sensor signal, and These are signals The mean of the k-th IMF component and L is the total number of signal sampling points.

[0037] threshold r th for:

[0038]

[0039] in, is the mean of the correlation coefficients, and n represents the number of levels.

[0040] According to the conditions Furthermore, valid IMF components were selected based on kurtosis and spectral kurtosis values ​​greater than 3. Subsequently, the obtained IMF components were decomposed using wavelet packet decomposition, and nodes for reconstructing the IMF components were selected based on the energy distribution of each frequency band and the optimal filtering parameters. After reconstructing and merging the IMF components, the denoised signal was obtained.

[0041] S4: Based on the noise-reduced signal matrix obtained in step S3 The sliding window method is used to dynamically capture characteristic changes. For the signal within the window, feature sequences corresponding to time-domain features, frequency-domain features, information entropy features, and time-frequency domain features are extracted to obtain time-domain feature statistics. and frequency domain characteristic statistics And the vibration sensor channel signal Extract to obtain the information entropy feature statistic and time-frequency domain features The specific implementation process is as follows:

[0042] (1) Use a sliding window with a set window size of τ and a step size of l to perform feature extraction on the denoised signal Calculate the samples within [k, k - 1 + τ] starting from time k, and judge k + l < K - τ. Define the signal length as L. If it is satisfied, slide the window by l step sizes and calculate the time-domain features. After that, construct the time-domain feature matrix It represents the matrix constructed by all time-domain feature vectors of the signal of the j-th sensor channel

[0043] (2) First, construct each feature as a feature column vector where (·) is the transpose of the row vector, represents the feature vector is the total length of and represents rounding down

[0044] According to the obtained vibration sensor channel signal Perform a one-sided FFT spectrum transformation within [k, k - 1 + τ] and define the frequency-domain signal as Its frequency sampling f s is the sampling frequency. According to the spectral kurtosis analysis of the frequency signal after STFT transformation obtain the spectral kurtosis S j (ω) calculated for the signal of the j-th sensor channel, where <·>​​​​​​​​​​​​​​​​​​​​​​​​​​​

[0050] Singular spectral entropy is an analytical method that combines singular spectrum and information entropy. Firstly, it... Transform into n vectors Y1,…,Y of length m∈[2,τ-1]. n , where vector And n = k + τ - m, combine the n sub-vectors to construct the trajectory matrix Y = (Y1, ..., Y... n )∈R m×n

[0051]

[0052] Next, singular value decomposition is performed on the covariance matrix of matrix Y to obtain the singular value spectrum λ = (λ1,...,λ2). m The probability density of the singular value spectrum (i = 1, ..., m) can be obtained by calculating the proportion of each singular value:

[0053]

[0054] Then, by combining the information entropy formula, the singular spectral entropy H can be obtained. Y (k), calculate the singular spectral entropy eigenvector. Similar to the power spectral entropy process:

[0055]

[0056] Similar to Transform into n matrices Z = (Z1, ..., Zn) of length m ∈ [2, τ-1]. n ), where vector in This represents the mean of the subvectors. Next, Z... i With Z c Calculate the maximum value of the corresponding point difference and record d. ic =max(|Z i -Z c |), traverse all combinations and calculate similarity using fuzzy membership:

[0057]

[0058] Where σ represents the gradient of the similarity tolerance boundary determined by the weight, r is the width of the fuzzy function boundary, and then the similarity statistic is calculated:

[0059]

[0060] Let m = m + 1, and repeat the steps to obtain the fuzzy entropy:

[0061]

[0062] For parameter selection, this invention uses m, σ = 2, and r = 0.15, which ultimately yields the fuzzy entropy feature vector.

[0063] Regarding the selection of permutation entropy features, firstly... Transform into n' matrices Q = (Q1, ..., Qn') of length m ∈ [2, τ-1]. n ),in This is the time delay parameter. A sequential encoding method is selected, and the probability of each type occurring is defined as follows:

[0064]

[0065] Where N(·) represents the number of items, the permutation entropy is calculated according to the information entropy formula:

[0066]

[0067] Where m = 6, Sorting Entropy After calculating the above four entropy features, construct the entropy feature matrix. This represents a matrix constructed from all the entropy eigenvectors of the signal from the j-th sensor channel.

[0068] (4) This invention extracts time-frequency features from the time-frequency graph based on the entropy value definition method. First, for the signal... Starting from time k, calculate the signal function by performing STFT transformation on τ signal points within the range [k, k-1+τ]. Simultaneously, wavelet basis functions are selected as window functions, and then CWT transformation is performed to obtain continuous wavelet transform functions:

[0069]

[0070] The CWT transform process is essentially a convolution of the signal with a wavelet function, where ψ a,b Wavelet function (a>0):

[0071]

[0072] Here, 'a' and 'b' are both time scale parameters. Although the frequency ω is not explicitly used as a wavelet function parameter, it is implicitly included in parameter 'a'. Next, time-frequency feature extraction is performed. The time-frequency map and scale map obtained after the two transformations are divided into N equal-area time-frequency map patches, and their total area is calculated. and Using the ratio of sub-block area to total area as the probability density function:

[0073]

[0074] The STFT entropy and CWT entropy are calculated using the information entropy formula:

[0075]

[0076] If the time-frequency characteristic conditions are the same, the calculation stops. Construct the time-frequency characteristic matrix. Feature column vector Similar to the time domain, the two time-frequency characteristic matrices, STFT entropy and CWT entropy, are finally combined to obtain... Finally, the four feature extraction methods are combined to obtain the feature matrix. like Figure 4 As shown.

[0077] S5: A pipeline leak detection model is built using a stacked sparse autoencoder, and the model is trained. The reconstruction error is used as a detection statistic, and the reconstruction error J is calculated by comparing the predicted and tested values ​​obtained from each sensor. SSAE Then, the confidence level α is designed using the Gaussian kernel density estimation method, and the set threshold J is obtained. α ,like Figure 5 As shown, CDF represents the cumulative distribution function, and PDF represents the probability density function. The region is determined on the CDF based on the confidence level α, and the boundary points correspond to... Figure 5 The values ​​(0.0234, 0.99) on the horizontal axis represent the set threshold J corresponding to the confidence level α. α By comparing the reconstruction error J SSAE With the set threshold J α Anomaly detection is performed on the size of the interval; if the reconstruction error J is... SSAE Greater than the set threshold J α If the data falls outside the confidence interval, a pipeline leak will be detected. The confidence level is used to determine the confidence interval, and data points outside this interval are identified as outliers. Therefore, it is used to determine the threshold (if 95% is set, it means that data points falling outside the 95% interval are outliers).

[0078] This invention proposes a pipeline leak detection method based on stacked sparse autoencoders. After denoising and feature extraction, to address the problem that detection methods require large amounts of data containing anomaly labels, making them unsuitable for this invention, an unsupervised model is used for leak detection. To ensure that the output sequence can be converted into an alarm signal, an anomaly detection model based on a stacked autoencoder structure is considered. The model is then pre-trained, and the reconstruction error is used as a detection statistic. The reconstruction error of each sensor is obtained through a variational structure, and the detection threshold is obtained using a kernel density estimation method. Finally, the pipeline health status is displayed in the form of a leak alarm signal. The pipeline leak detection alarm framework is as follows: Figure 3 As shown.

[0079] An anomaly detection model is based on a stacked autoencoder (SAE) structure, which appears to consist of multiple SAE modules stacked layer by layer. An SAE module is a three-layer symmetric neural network containing: an input layer x∈R n Hidden layer h∈R p Output layer Its dimensionality change is n→p→n, the purpose of which is to achieve this through the encoder f. θ and decoder g θ The process makes the output The model aims to make the input x as similar as possible. To improve the model's anomaly detection accuracy, a sparse penalty constraint structure is introduced for the stacked autoencoder. The training of the resulting stacked sparse autoencoder (SSAE) is not a one-step process, but rather a layer-by-layer training. The entire model training process is as follows:

[0080] (1) Model training: First, normalize the data, and then initialize all parameters X∈R in the model. M×L Where M is the number of variables and L is the length of the time series. The data preparation phase involves dividing the data into training and validation sets, and reorganizing the data dimensions to facilitate model training. For example, where X i This represents the training input for the i-th batch of training, n. bs This indicates that each batch size enters the first Action Block (AE). During training, a complete AE block is used for training, and during the encoding phase, the encoder f... θ The encoded y1 is the hidden layer output.

[0081] y1=f θ (X i =ReLU(WX) i +b)

[0082] Where θ = {W, b} are the encoder parameters, the input layer dimension n1 = N × M, the hidden layer dimension p1, and the output layer dimension n1, this invention selects the ReLU function as the activation function. In the decoding stage, the encoded y1 is processed by the decoder to obtain the reconstructed matrix.

[0083] y1=g θ' (y1)=ReLU(W′X i +b)

[0084] Where θ' = {W', b'} are the decoder parameters, which are then compared with the original X. i Constructing the loss function The loss function used in this invention is the MSE (mean squared error) loss function. The network parameters are iteratively updated through gradient optimization to complete the training of the first AE block; that is, the network learns n1→p1→n1. Then, the output layer is removed to obtain the change n1→p1, and the output of the hidden layer h1 is passed to the next AE block. This process continues until all AE blocks are trained.

[0085] (2) Sparse Penalty Constraint: Sparsity reduces computational complexity and storage requirements by decreasing the number of parameters in a neural network. This accelerates training and inference while maintaining or even improving model accuracy. Adding a sparsity constraint essentially deactivates certain neurons; a neuron's output closer to 1 indicates activation, while a output closer to 0 indicates deactivation. Sparse Penalty Calculation: Before reconstructing the loss function, the sparsity constraint needs to be converted into a sparsity penalty term, based on the KL divergence and the activation level of the neural units. Construct a sparsity penalty term with the sparsity parameter ρ.

[0086]

[0087] in The sum of the KL divergences of the p units within the hidden layer represents the sparsity penalty term. Combining this with the aforementioned loss function, a loss function incorporating the sparsity penalty is constructed.

[0088]

[0089] Where β is the weight that controls the sparsity penalty, and n is the input dimension of each AE block.

[0090] (3) Calculation of detection statistics and threshold setting: The reconstruction error is calculated using the input test data and the output prediction data, and the abnormal threshold of the reconstruction data is determined as the detection statistics.

[0091] After each batch of training, the stacked sparse autoencoder model described above constructs a reconstruction error by combining the input and output data.

[0092]

[0093] in The floor sign indicates that the reconstructing errors of all batches are combined to obtain the reconstructing error sequence J. SSAE

[0094]

[0095] Since the model is trained using normal samples, the resulting reconstruction error should be relatively small. However, when anomalies occur, the model's reconstruction error will be relatively large, making the threshold calculation crucial. This paper employs a kernel density estimation method to solve for the threshold, with the specific steps as follows:

[0096] (1) Calculation of probability density function: First, select the kernel function and the kernel density estimation bandwidth, and then calculate the probability density function based on the reconstruction error sequence J. SSAE The kernel density function is obtained by fitting and solving the kernel density function, which is the probability density function.

[0097] (2) The cumulative distribution function is obtained by integrating the probability density function.

[0098] (3) Threshold J α Solution: Design a confidence level α, and find the corresponding threshold J using the cumulative distribution function. α Since normal data is used to train the model, the confidence level α is generally selected between 0.95 and 0.99. The threshold is calculated during the training phase and does not need to be calculated during the validation phase.

[0099] Finally, by comparing J SSAE and J α By analyzing the magnitude relationships, abnormal data and their corresponding positions in the sequence can be identified one by one. Table 3 compares the pipeline leak detection results of the method proposed in this invention with those of different models.

[0100] Table 1. Statistical confidence intervals for each model, α = 0.99 threshold.

[0101]

[0102] Verification indicators and comprehensive

[0103] Evaluation metrics are crucial for determining model performance. Common evaluation metrics in anomaly detection include accuracy, recall, and precision. However, for the statistical metrics output by the model proposed in this invention, accuracy, false alarm rate (FAR), false alarm rate (MAR), and detection delay (DD) will be combined to comprehensively evaluate the model.

[0104] The formula for calculating the accuracy of the detection model SSAE proposed in this invention is as follows, where R ACC ∈[0,1], N TP N represents the number of sample points where the leak was successfully detected as a leak. TN N represents the number of sample points that were successfully detected as normal. FN N represents the number of sample points where the leak was detected as a normal state.FP This indicates the number of sample points that were detected as leaking under normal conditions.

[0105]

[0106] False alarm rate R FAR ∈[0,1] represents the evaluation index for when the detection model will misjudge a leak, and its calculation formula is:

[0107]

[0108] The false negative rate R MAR ∈[0,1] represents the evaluation index by which the detection model classifies a leak as normal, and its calculation formula is:

[0109]

[0110] When the k-th sample point is detected as leaking, all subsequent τ' samples are also leaking, meaning all samples within the interval [k, k+τ') are leaking. Therefore, the k-th sample point can be considered the leak detection starting point. If this condition is not met, let k = k+1, and repeat the above steps until the condition is satisfied. If the actual leak starting point is k', then the detection delay represents the time difference between the leak detection starting point and the actual leak point: T. DD =k-k'.

[0111] Components were selected for the five parts: signal denoising and four types of feature values. Different training sets were used for training, and normal and abnormal sample points were used as validation sets. The impact of each component on the detection results was observed by calculating the accuracy, false alarm rate, and false negative rate. The ablation experiment results are shown in Table 2, with the symbol (√) indicating that the component was selected.

[0112] Table 2 Ablation experiments on signal denoising and four types of eigenvalues.

[0113]

[0114] To ensure comparability of the comparative experiments, the encoding and decoding dimensionality variations of the AE, SpAE, and StAE models are identical to those of the proposed model. The comparison results are shown in Table 3. Due to a step size l = 100, the detection delay T... DD The unit is 100 milliseconds (ms). For example... Figure 6 The figure shows the statistical results of different methods for different leak detection experiments.

[0115] Table 3 Comparison of pipeline leak detection results using different models

[0116]

[0117] Please see Figure 7 , Figure 7 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a pipeline leakage detection device 401 that integrates multi-time-frequency domain features of vibration signals, a processor 402, and a storage device 403.

[0118] A pipeline leak detection device 401 that integrates multi-time-frequency domain features of vibration signals: The pipeline leak detection device 401 that integrates multi-time-frequency domain features of vibration signals implements the pipeline leak detection method that integrates multi-time-frequency domain features of vibration signals.

[0119] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the pipeline leakage detection method that integrates multi-dimensional time-frequency domain features of vibration signals.

[0120] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the pipeline leakage detection method that integrates multi-dimensional time-frequency domain features of vibration signals.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting pipeline leaks by integrating multi-dimensional time-frequency domain features of vibration signals, characterized in that: include: S1: Use a filter to process the original signal Power frequency interference is filtered out to obtain the signal matrix after power frequency interference removal. ; S2: Using the EEMD decomposition method to decompose the signal matrix The IMF components are decomposed to obtain IMF components, and then filtered to obtain effective IMF components. S3: Decompose the selected IMF components using wavelet packets, analyze the energy distribution of each component after wavelet packet decomposition, and reconstruct the components by combining the corresponding optimal frequency bands in the spectral kurtosis map. Then, combine the reconstructed components to obtain the denoised signal matrix. ; S4: Based on the noise-reduced signal matrix obtained in step S3 The sliding window method is used to dynamically capture characteristic changes. For the signal within the window, feature sequences corresponding to time-domain features, frequency-domain features, information entropy features, and time-frequency domain features are extracted to obtain the time-domain features. and frequency domain features and the vibration sensor channel signal Extracting the information entropy feature statistics yields the following results. and time-frequency domain features ; S5: A pipeline leak detection model is built using a stacked sparse autoencoder, and the model is trained. The reconstruction error is used as a detection statistic, and the reconstruction error J is calculated by comparing the predicted and tested values ​​obtained from each vibration sensor. SSAE Then, the confidence level is designed using the Gaussian kernel density estimation method. Thus, the set threshold is obtained. By comparing the reconstruction error J SSAE With set threshold Anomaly detection is performed on the size of the interval; if the reconstruction error J is... SSAE Greater than the set threshold If so, a pipeline leak will be detected.

2. The pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain features of vibration signals as described in claim 1, characterized in that: In S2, rapid spectral kurtosis analysis and autocorrelation coefficients are used to screen IMF components.

3. The pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain features of vibration signals as described in claim 1, characterized in that: In S2, each decomposition yields J IMF components and a sum of residuals. The formula for calculating the IMF components is: in, It is the first After the first addition of white noise decomposition, the [number]th One IMF component, For the residual of the signal, Signal matrix The original signal from the j-th vibration sensor. For signal No. A disturbance signal.

4. The pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain features of vibration signals as described in claim 1, characterized in that: In S2, according to the conditions Furthermore, if the kurtosis value and spectral kurtosis are greater than 3, valid IMF components are selected. Here, kurtosis is a signal... The fourth-order distance, and It is a signal The mean and standard deviation, For the first Each IMF component and the original signal The correlation coefficient between them For threshold; The formula for calculating kurtosis is as follows: The formula for calculating the threshold is: in, For the threshold, The mean of the correlation coefficients. This indicates the number of levels.

5. The pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain features of vibration signals as described in claim 4, characterized in that: In S3, based on the STFT change, a fast spectral kurtosis analysis algorithm is used to divide the spectrum into multiple levels. Then, the kurtosis value and correlation coefficient of the signal are calculated using the spectral kurtosis formula. The formula for calculating the correlation coefficient is as follows: in, For the first Each IMF component and the original signal The correlation coefficient between them Signal matrix The original signal from the j-th vibration sensor. and These are the original signals. and the The mean of each IMF component, This represents the total number of signal sampling points. The obtained IMF components are decomposed using wavelet packet decomposition, and nodes for reconstructing the IMF components are selected based on the energy proportion distribution of each frequency band and the optimal filtering parameters. After reconstructing and merging the IMF components, the denoised signal is obtained. .

6. The pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain features of vibration signals as described in claim 1, characterized in that: S4 specifically refers to: S41: Utilize the set window size Step size is The sliding window for the denoised signal Perform feature extraction, from Start calculating time Within the samples, determine The signal length is defined as If so, slide the window. The time-domain features are calculated by taking a step size and then calculating the time-domain features, thereby obtaining the time-domain feature statistics constructed from all time-domain feature vectors of the j-th vibration sensor channel signal. ; S42: Based on the obtained vibration sensor channel signal ,exist Perform FFT one-sided spectral transformation within the domain and define the frequency domain signal as... Its frequency sampling It is the sampling frequency; based on the frequency signal after STFT change. The spectral kurtosis of the j-th vibration sensor channel signal is obtained through spectral kurtosis analysis. : in, , It is a time averaging operation, and Finally, the frequency domain feature statistics constructed from all frequency domain feature vectors of the j-th vibration sensor channel signal are obtained. ; S43: Obtain the power spectral entropy by combining the information entropy formula. : in, Represents the power spectrum. Power spectrum The probability density function is: As the window slides, the next calculation is performed. Until the power spectral entropy eigenvector is obtained. ; Singular spectral entropy is an analytical method based on the combination of singular spectrum and information entropy. It first involves... Transform into A length of subvectors , where subvector and ,Will The trajectory matrix is ​​obtained by merging the subvectors. : Next, the matrix Singular value spectrum is obtained by performing singular value decomposition on the covariance matrix. The probability density of the singular value spectrum can be obtained by calculating the proportion of each singular value: ， Then, by combining the information entropy formula, the singular spectral entropy can be obtained. Calculate the eigenvectors of the singular spectrum entropy. : Will Transform into A length of matrix , where vector , express The mean of the subvectors, Z c Indicates the first vectors ; Z i With Z c The maximum corresponding point difference value is recorded as Iterate through all combinations and calculate Z using fuzzy membership. i and Z c Similarity: in, The weights indicate that the gradient of the similarity tolerance boundary is determined. The width of the fuzzy function boundary; Then calculate the similarity statistic: Again replace Repeat the steps to obtain the fuzzy entropy: Finally, the fuzzy entropy feature vector is obtained. ; The permutation entropy is calculated using the information entropy formula: in, This represents the probability of each sequential type encoding arrangement occurring. Indicates the number of vibration sensors, and: in, This represents a function to count items. This is the time delay parameter; Finally, the sorting entropy is obtained. ; After calculating the above four entropy features, we obtain the information entropy feature matrix statistics constructed from all the entropy feature vectors of the j-th vibration sensor channel signal. ; S44: Extracting time-frequency features from the time-frequency graph based on the entropy definition method. First, for the signal... No. Calculated from time Inside The signal function is obtained by performing STFT transformation on each signal point. Simultaneously, wavelet basis functions are selected as window functions, and then CWT transformation is performed to obtain continuous wavelet transform functions: in, For wavelet functions, and All are time-scale parameters, and ; The time-frequency diagram and scale diagram obtained after the two transformations are divided into equal parts. Time-frequency patches of equal area and Calculate the total area of ​​each of the two graphs above. and The ratio of sub-block area to total area is used as the probability density function: in, This represents the total area of ​​the time-frequency plot. The total area of ​​the scale map is represented, and the STFT entropy and CWT entropy are calculated using the information entropy formula: If the time-frequency threshold feature conditions are the same, the calculation stops, and the time-frequency threshold feature matrix is ​​constructed. Feature column vector Similar to the time domain, the final result is obtained by merging the two time-frequency threshold feature column vectors: STFT entropy and CWT entropy. ; Finally, the four feature extraction methods are combined to obtain the feature matrix. .

7. The pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain features of vibration signals as described in claim 1, characterized in that: In S5, the model training process is as follows: S51: Model Training: First, normalize the data, then initialize all parameters within the model. ,in It is the number of variables. The time series length is specified; the data is divided into training and validation sets during the data preparation phase, and the data dimensions are reorganized to facilitate model training. Indicates the first Batch training input, This indicates that each batch of samples enters the first AE block, and during training, it is trained as a complete AE block. During the encoding phase, it passes through the encoder. Get the code This is the output of the hidden layer: in For encoder parameters, input layer dimensions n1 represents the output layer dimension, and ReLU is used as the activation function; additionally, Represents the hidden layer dimension; during the decoding phase, encoding... The reconstructed data is obtained through the decoder. : in These are the decoder parameters, which are then compared with the original input. Constructing the loss function The mean squared error (MSE) is chosen as the loss function; then, the network parameters are iteratively updated through gradient optimization to complete the training of the first AE block, meaning the network has learned... Then remove the output layer to obtain The changes will affect the hidden layer. The output is passed to the next AE block; Then begin training the next AE block, assuming it has already been trained to the [number]th [block]. The first AE block, first freeze the first... The parameters of each AE block are set, and training them layer by layer is done separately to prevent the deep network from getting stuck in local minima. During training, the loss function for each AE is different; here, a general formula is defined as follows: , of which The input of each AE block is and output as The training was completed according to the above steps. The change satisfies the current input layer dimension. Equal to the previous hidden layer dimension Then train the next AE block one by one until all AE blocks are trained. S52: Sparsity penalty constraint: Sparsity reduces computational complexity and required storage space by reducing the number of parameters in a neural network; Adding sparsity constraints essentially deactivates certain neurons; that is, the closer a neuron's output is to 1, the more activated it is, and vice versa. Sparse penalty calculation: Before reconstructing the loss function, the sparsity constraints need to be transformed into sparsity penalty terms, based on the KL divergence and the activation level of the neural units. and sparsity parameters Construct a sparsity penalty term: in Within the hidden layer The sum of the KL divergences of each unit is the sparsity penalty term; combined with the loss function, a loss function with sparsity penalty is constructed: in, To control the weight of the sparsity penalty, The input dimension for each AE block; S53: Reconstruct the data using the input test data and the output prediction data, and calculate the anomaly threshold of the reconstructed data as a detection statistical indicator.

8. A storage device, characterized in that: The storage device stores instructions and data to implement the pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain characteristics of vibration signals as described in any one of claims 1 to 7.

9. A pipeline leakage detection device that integrates multi-dimensional time-frequency domain characteristics of vibration signals, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the pipeline leakage detection method based on the fusion of multi-dimensional time-frequency domain characteristics of vibration signals as described in any one of claims 1 to 7.

Images