Water supply pipeline leakage detection method based on BD combined convolutional neural network

Abstract

The invention discloses a water supply pipeline leakage detection method based on a BD combined convolutional neural network, relates to a water supply pipeline leakage detection method, and aims to solve the problem of poor recognition accuracy of a recognition model due to the fact that a blind deconvolution filter cannot be directly combined with a deep learning algorithm in existing water supply pipeline leakage detection. The method comprises the following steps: filtering an obtained noisy signal through a BD filter to obtain a frequency domain filtering signal; performing parameter optimization training on the convolutional neural network model by using the frequency domain filtering signal to generate a water supply pipeline leakage identification model; and the current noise-containing signal of the leakage point location is introduced into the input end of the water supply pipeline leakage identification model, and a detection result is output. The method has the beneficial effect that the accuracy of identifying the to-be-detected signal by the water supply pipeline identification model is improved.

Description

Technical Field

[0001] The invention relates to a water supply pipeline leakage detection method. Background Art

[0002] With the continuous development of urban water supply systems, leak detection of water supply networks has gradually become an important part of infrastructure maintenance; in leak detection of water supply networks, feature extraction plays a key role; leaks usually produce irregular changes in pressure, flow, vibration or sound in the pipeline, and these changes often have significant characteristics of leaks. We have collected relevant data through sensors, and the task of feature extraction is to extract effective information that can reflect the characteristics of leaks from these raw data, and identify abnormal patterns related to leaks for subsequent analysis.

[0003] At present, feature extraction methods are divided into two categories: data-driven methods and signal processing methods. Data-driven methods promote automatic feature extraction and pattern recognition, such as convolutional generative adversarial networks, but due to their black box characteristics, these models lack transparency and their internal mechanisms are unclear, which limits their application in real life; while in the field of signal processing methods, a large number of algorithms with strict mathematical foundations have been proposed to extract fault-related features, including wavelet transform, variational mode decomposition, BD (blind deconvolution) method, etc. However, the results of feature extraction using wavelet transform depend on the selected wavelet basis function and have weak adaptability; variational mode decomposition will have problems such as mode aliasing, inability to effectively extract high-frequency components, and uncontrollable number of decomposition layers.

[0004] Compared with other signal feature extraction algorithms, the BD method has the unique advantages of strong adaptability and no restrictions on bandwidth and center frequency. These characteristics are conducive to the blind deconvolution filter to more effectively extract leakage signal features. However, the existing blind deconvolution filters cannot be directly used in conjunction with deep learning algorithms, resulting in poor recognition accuracy of the recognition model. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that in the existing water supply network leakage detection, the blind deconvolution filter cannot be directly used in conjunction with the deep learning algorithm, resulting in poor accuracy of the recognition model. A water supply pipeline leakage detection method based on BD combined with convolutional neural network is proposed.

[0006] The water supply pipeline leakage detection method based on BD combined with convolutional neural network described in the present invention comprises the following steps:

[0007] The following steps are involved:

[0008] The current noisy signal of the leakage point is passed to the input end of the water supply pipeline leakage identification model, and the output end of the water supply pipeline leakage identification model outputs the detection result.

[0009] Furthermore, the method for generating the water supply pipeline leakage identification model is:

[0010] Step 1: Obtain the historical noise signal of the leakage point;

[0011] Step 2: Filter the historical noisy signal obtained in step 1 through a BD filter to obtain a frequency domain filtered signal;

[0012] Step 3: Use the frequency domain filtering signal obtained in step 2 to perform parameter optimization training on the convolutional neural network model to generate a water supply pipeline leakage recognition model.

[0013] Further, the historical noise-containing signal includes a leakage acoustic signal of a leak point and a non-leakage acoustic signal of a leak point;

[0014] Among them, the method for collecting leakage sound signals at the leakage point is: to check the actual water supply pipeline network, and use a vibration sensor to collect leakage signals at the pipeline leakage point; the method for collecting non-leakage sound signals at the leakage point is: after repairing the pipeline leakage point, use the vibration sensor again to collect non-leakage signals at that point.

[0015] Furthermore, the specific process of obtaining the historical noise signal of the leakage point in step 1 is as follows:

[0016] 35 leakage points in four different cities were selected, and the acoustic signals were sampled every 2 minutes at the leakage points. 72 leakage data and 72 non-leakage data were collected from each leakage point during the day and night. 2×2×72=288 data were collected from each leakage point, for a total of 35×288=10080 acoustic emission data. The sampling frequency of each data was 4762 Hz, and each data had a total of 42859 data points.

[0017] Furthermore, in step 2, the BD filter includes a time domain quadratic filter and a frequency domain linear filter;

[0018] The time domain secondary filter and the frequency domain linear filter are connected together in a cascade manner; wherein the time domain secondary filter serves as a first-stage BD filter; and the frequency domain linear filter serves as a second-stage BD filter.

[0019] Furthermore, the method for constructing the BD filter in step 2 is: reconstructing the leakage signal by estimating the transmission path function using the measured signal;

[0020] Given the measured signal x, the leakage source signal d, and the noise n, the signal transfer process is defined as follows:

[0021] x=d*h d +n*h n

[0022] Among them, h d is the transfer function of the time domain quadratic filter; h n is the transfer function of the frequency domain linear filter, * is the convolution operation;

[0023] The approximate leakage source signal y is restored by constructing a BD filter f. The process is defined as follows:

[0024] y=x*f=(d*h d +n*h n )*f≈d

[0025] The expression of the secondary convolutional neural network used by the time domain secondary filter is:

[0026] y=σ((W1*x+b1)⊙(W2*x+b2)+W3*(x⊙x)+b3)

[0027] Among them, σ(~) is the activation function ReLU, and the specific mathematical expression of the activation function ReLU is: σ(β)=max(0,β), β is the input of the activation function; W1 is the weight of the first layer of the quadratic neural network; W2 is the weight of the second layer of the quadratic neural network; W3 is the weight of the third layer of the quadratic neural network; b1 is the bias of the first layer of the quadratic neural network; b2 is the bias of the second layer of the quadratic neural network; b3 is the bias of the third layer of the quadratic neural network; ⊙ represents matrix dot multiplication;

[0028] The loss function of the secondary convolutional neural network is expressed as:

[0029]

[0030] Among them, l t is the loss function; k is the kth layer of the quadratic neural network, N is the total number of layers of the quadratic neural network, and y(k) is the output of the quadratic convolutional neural network;

[0031] The frequency domain is filtered using a fully connected neural network and Fourier transform to construct a frequency domain linear filter;

[0032] The signal output by the time domain quadratic filter is defined as Apply the Fourier transform F(·) to convert the signal into the frequency domain:

[0033]

[0034] in, Output signal of the time domain quadratic filter Frequency domain signal after Fourier transformation;

[0035] According to the convolution theorem, the expression for implementing the frequency domain linear filter using a fully connected neural network is as follows:

[0036]

[0037] Among them, N is the total number of layers of the quadratic neural network, i is the i-th layer of the quadratic neural network, and w i (f) is the weight of the i-th layer of the quadratic neural network at frequency f, b i is the bias of the i-th layer of the quadratic neural network; is the frequency domain signal after filtering;

[0038] Apply the inverse Fourier transform F -1 Convert the signal to the time domain:

[0039]

[0040] in, is the time domain signal after inverse Fourier transform;

[0041] Finally, an optimization function based on envelope spectrum is designed under the frequency domain linear filter;

[0042] The design process of the optimization function is:

[0043] The Hilbert transform is defined as:

[0044]

[0045] Among them, t is the time quantity; τ is the time offset; h(t) is the output of the input signal x(t) after Hilbert transform. According to the convolution theorem, we get:

[0046]

[0047] Where, j is an imaginary number; is the frequency domain signal under Hilbert transform; sgn(f) is the sign function; is the frequency domain signal after filtering;

[0048] The discrete analytical signal z(n) obtained by Hilbert transform is:

[0049]

[0050] Among them, z(n) is the analytical signal in discrete form obtained by Hilbert transform; is a time domain signal after filtering in discrete form; is the time domain signal under the discrete form of Hilbert transform;

[0051] Then the envelope spectrum ES(f) is defined as the Fourier spectrum of the analytical signal z(n) mode, and its mathematical expression is as follows:

[0052]

[0053] Then, the optimization function l f The expression is:

[0054]

[0055] Where f is the frequency of the quadratic neural network.

[0056] Furthermore, the specific definition method of the convolutional neural network model in step three is:

[0057] Step 31: Determine the convolution layer;

[0058] use and denote the weight and bias of the kernel of the i-th convolutional layer in layer l, and use u l (m) represents the mth local area in layer l, so the convolution process is described as follows:

[0059]

[0060] in, is the characteristic value of the mth local area in layer l+1;

[0061] Step 32: Determine the pooling layer;

[0062] The maximum pooling operation is used to sample the convolutional layer output, and its formula is as follows:

[0063]

[0064] Among them, γ is the γth local area to be pooled, is the input feature corresponding to the γth local area to be pooled under the i-th convolution kernel in layer l, W is the pooling depth, P i l+1 (j) is the output value of the ath pooling layer under the i-th convolution kernel in layer l+1;

[0065] Step 33: Determine the cross entropy loss function;

[0066]

[0067] Among them, Loss represents the cross entropy loss function; G is the total number of samples; g is the g-th pseudo-copy; c g is the true label of the g-th sample; d g is the predicted probability of the g-th sample.

[0068] Furthermore, the specific process of parameter optimization training of the convolutional neural network model in step 3 is as follows:

[0069] A training sample data set and a training set label, wherein the training sample data set is a two-dimensional matrix composed of frequency domain filtered signals after passing through a BD filter, and the training set label is the signal type in the training sample data set, including leakage and non-leakage;

[0070] The training sample data set is input into the convolutional neural network model, and after logical output, it is compared with the training set label, and the loss value is obtained through the cross entropy loss function. The convolutional neural network model is automatically updated according to the size of the loss value and the iterative operation is continued until a certain number of iterations is reached to obtain the optimal parameters with the minimum loss, and the historical noisy signal of the input leakage point can be correctly classified;

[0071] The parameters are the weights and biases of each layer of the network.

[0072] Furthermore, the specific method for evaluating the performance of the water supply pipeline leakage identification model is as follows:

[0073] The best performing weight and bias information on the training set is loaded onto the test set for inference. Six indicators are used to evaluate the performance of the model: true positive rate, false positive rate, F1 score, accuracy, ROC curve, and AUC. The expressions are as follows:

[0074]

[0075] In the formula, TP means that the leak in the pipeline is correctly detected; FN means that the leak in the pipeline is not detected; TN means that the leak actually occurred and was successfully detected; FP means that the leak did not actually occur but was mistakenly detected as a leak; TPR is the true positive rate; FPR is the false positive rate; PRE is the accuracy rate;

[0076] AUC is the area under the ROC curve, which is used to quantify the performance of the ROC curve. The value range of AUC is between 0 and 1. The physical meaning of AUC is: when a positive sample and a negative sample are randomly selected, the water supply pipe leakage recognition model determines the probability that the positive sample score is greater than the negative sample score. The closer the AUC is to 1.0, the stronger the classification ability of the water supply pipe leakage recognition model.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] The present invention uses two cascaded BD filters to extract features and perform frequency filtering on signals, trains a large number of processed signals through a convolutional neural network, and establishes a water supply pipeline recognition model; the two cascaded BD filters are directly used in conjunction with a deep learning algorithm to improve the accuracy of the water supply pipeline recognition model in recognizing the signals to be detected. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a schematic diagram of the implementation principle of a water supply pipeline leakage detection method based on BD combined with convolutional neural network described in the first specific implementation mode;

[0080] Figure 2 The training and validation loss graphs of the algorithm provided in the first embodiment;

[0081] Figure 3 The ROC-AUC curve diagram of the water supply pipe leakage identification model evaluation provided in the first specific implementation mode;

[0082] Figure 4 This is the confusion matrix of the algorithm provided in the first specific implementation mode. DETAILED DESCRIPTION

[0083] Specific implementation method 1. Combination Figures 1 to 4 The present embodiment is described. The water supply pipeline leakage detection method based on BD combined with convolutional neural network described in the present embodiment comprises the following steps:

[0084] The current noise-containing signal of the leakage point is passed to the input end of the water supply pipeline leakage identification model, and the output end of the water supply pipeline leakage identification model outputs the detection result. Among them, the current noise-containing signal of the leakage point is the sound signal to be detected;

[0085] In this embodiment, the method for generating the water supply pipeline leakage identification model is:

[0086] Step 1: Obtain the historical noise signal of the leakage point;

[0087] Step 2: Filter the historical noisy signal obtained in step 1 through a BD filter to obtain a frequency domain filtered signal;

[0088] Step 3: Use the frequency domain filtering signal obtained in step 2 to perform parameter optimization training on the convolutional neural network model to generate a water supply pipeline leakage recognition model.

[0089] In this embodiment, the historical noise-containing signal includes a leakage acoustic signal of a leak point and a non-leakage acoustic signal of a leak point;

[0090] Among them, the method for collecting leakage sound signals at the leakage point is: to check the actual water supply pipeline network, and use a vibration sensor to collect leakage signals at the pipeline leakage point; the method for collecting non-leakage sound signals at the leakage point is: after repairing the pipeline leakage point, use the vibration sensor again to collect non-leakage signals at that point.

[0091] In this embodiment, the specific process of obtaining the historical noise signal of the leakage point in step 1 is as follows:

[0092] 35 leakage points in four different cities were selected, and the acoustic signals were sampled every 2 minutes at the leakage points. 72 leakage data and 72 non-leakage data were collected from each leakage point during the day and night. 2×2×72=288 data were collected from each leakage point, for a total of 35×288=10080 acoustic emission data. The sampling frequency of each data was 4762 Hz, and each data had a total of 42859 data points.

[0093] In this embodiment, the BD filter in step 2 includes a time domain quadratic filter and a frequency domain linear filter;

[0094] The time domain secondary filter and the frequency domain linear filter are connected together in a cascade manner; wherein the time domain secondary filter serves as a first-stage BD filter; and the frequency domain linear filter serves as a second-stage BD filter.

[0095] In this embodiment, the method for constructing the BD filter in step 2 is: reconstructing the leakage signal by estimating the transmission path function using the measured signal;

[0096] Given the measured signal x, the leakage source signal d, and the noise n, the signal transfer process is defined as follows:

[0097] x=d*h d +n*h n

[0098] Among them, h d is the transfer function of the time domain quadratic filter; h n is the transfer function of the frequency domain linear filter, * is the convolution operation;

[0099] The goal of the BD filter is to extract leakage-related features from the historical noisy signal of the leakage point. To achieve this goal, the approximate leakage source signal y is restored by constructing a BD filter f. The process is defined as follows:

[0100] y=x*f=(d*h d +n*h n )*f≈d

[0101] However, accurate estimation of the transfer function and its frequency response is often impractical due to the complexity of the system, and detection is further complicated by the presence of unpredictable noise;

[0102] In view of the non-stationarity and periodicity of the leakage feature, this embodiment uses a quadratic convolutional neural network to construct the first-order BD filter, that is, a time-domain quadratic filter; the quadratic convolutional neural network has many advantages. In terms of efficiency, the quadratic convolutional neural network can use polynomial-level neurons to approximate functions, while the existing neural network requires exponential neurons; in terms of feature representation, the quadratic network can achieve polynomial approximation, in contrast, the existing network uses piecewise approximation through nonlinear activation functions; in terms of representing complex functions, polynomial approximation is better than piecewise approximation;

[0103] The expression of the secondary convolutional neural network used by the time domain secondary filter is:

[0104] y=σ((W1*x+b1)⊙(W2*x+b2)+W3*(x⊙x)+b3)

[0105] Among them, σ(·) is the activation function ReLU, and the specific mathematical expression of the activation function ReLU is: σ(β)=max(0,β), β is the input of the activation function; W1 is the weight of the first layer of the quadratic neural network; W2 is the weight of the second layer of the quadratic neural network; W3 is the weight of the third layer of the quadratic neural network; b1 is the bias of the first layer of the quadratic neural network; b2 is the bias of the second layer of the quadratic neural network; b3 is the bias of the third layer of the quadratic neural network; ⊙ represents matrix dot multiplication;

[0106] The activation function ReLU is widely used as an activation unit to accelerate the convergence of convolutional neural networks. The activation function ReLU makes the weights of shallow layers easier to train when adjusting parameters using the back-propagation learning method.

[0107] In this embodiment, two secondary convolutional neural network layers are used to form a symmetrical structure based on the above secondary convolutional neural network neurons, imitating a multi-layer deconvolution filter. The first layer maps the input into 16 channels, and the second layer merges these 16 channels into a single output, and the dimension of the output remains the same as the input.

[0108] The loss function of the secondary convolutional neural network is expressed as:

[0109]

[0110] Among them, l t is the loss function; k is the kth layer of the quadratic neural network, N is the total number of layers of the quadratic neural network, and y(k) is the output of the quadratic convolutional neural network; this method realizes the time domain quadratic filter through the quadratic convolutional neural network and effectively realizes the feature extraction.

[0111] The frequency domain is filtered using a fully connected neural network and Fourier transform to construct a frequency domain linear filter;

[0112] The signal output by the time domain quadratic filter is defined as Apply the Fourier transform F(·) to convert the signal into the frequency domain:

[0113]

[0114] in, Output signal of the time domain quadratic filter Frequency domain signal after Fourier transformation;

[0115] According to the convolution theorem, the expression for implementing the frequency domain linear filter using a fully connected neural network is as follows:

[0116]

[0117] Among them, N is the total number of layers of the quadratic neural network, i is the i-th layer of the quadratic neural network, and w i (f) is the weight of the i-th layer of the quadratic neural network at frequency f, b i is the bias of the i-th layer of the quadratic neural network; is the frequency domain signal after filtering;

[0118] Apply the inverse Fourier transform F -1 Convert the signal to the time domain:

[0119]

[0120] in, is the time domain signal after inverse Fourier transform;

[0121] Finally, an optimization function based on envelope spectrum is designed under the frequency domain linear filter;

[0122] The design process of the optimization function is:

[0123] The Hilbert transform is defined as:

[0124]

[0125] Among them, t is the time quantity; τ is the time offset; h(t) is the output of the input signal x(t) after Hilbert transform. According to the convolution theorem, we get:

[0126]

[0127] Where, j is an imaginary number; is the frequency domain signal under Hilbert transform; sgn(f) is the sign function; is the frequency domain signal after filtering;

[0128] The discrete analytical signal z(n) obtained by Hilbert transform is:

[0129]

[0130] Among them, z(n) is the analytical signal in discrete form obtained by Hilbert transform; is a time domain signal after filtering in discrete form; is the time domain signal under the discrete form of Hilbert transform;

[0131] Then the envelope spectrum ES(f) is defined as the Fourier spectrum of the analytical signal z(n) mode, and its mathematical expression is as follows:

[0132]

[0133] Then, the optimization function l f The expression is:

[0134]

[0135] Where f is the frequency of the quadratic neural network.

[0136] This method implements a frequency domain linear filter through a fully connected neural network and adds an optimization function based on the envelope spectrum, which enhances the signal sparsity in the frequency domain, effectively reduces the influence of noise frequency components, and realizes filtering in the frequency domain.

[0137] In this embodiment, the specific definition method of the convolutional neural network model in step 3 is:

[0138] Convolutional neural network is a deep learning model that has been widely used in image processing, fault detection, speech recognition and other fields. It can effectively extract and learn the features of input data through convolution operation. Compared with the matrix operation of traditional network, its operation speed has been greatly improved. The convolution layer uses a series of convolution kernels to convolve the input data and extract the features of the input data.

[0139] Step 31: Determine the convolution layer;

[0140] use and denote the weight and bias of the kernel of the i-th convolutional layer in layer l, and use u l (m) represents the mth local area in layer l, so the convolution process is described as follows:

[0141]

[0142] Among them, r i l+1(m) is the eigenvalue of the mth local area in layer l+1. Since the convolution operation is a linear processing of data, the output result after convolution is processed through the activation layer to introduce nonlinear factors, thereby improving the learning ability of the model.

[0143] The main function of the pooling layer is to reduce the size of the output features of the convolutional layer, thereby reducing the number of parameters and the amount of calculation of the model;

[0144] Step 32: Determine the pooling layer;

[0145] The maximum pooling operation is used to sample the convolutional layer output, and its formula is as follows:

[0146]

[0147] Among them, γ is the γth local area to be pooled, is the input feature corresponding to the γth local area to be pooled under the i-th convolution kernel in layer l, W is the pooling depth, P i l+1 (j) is the output value of the ath pooling layer under the i-th convolution kernel in layer l+1;

[0148] In order to reduce the complexity and computational complexity of the convolutional neural network model, the convolution layer is used to extract the input features, and then the output features of the convolution layer are reduced in dimension through the pooling layer;

[0149] The cross entropy loss function is usually used for classification problems. It measures the difference between the predicted distribution and the true label distribution. The core idea of ​​cross entropy comes from information theory and is used to describe the distance between two probability distributions.

[0150] Step 33: Determine the cross entropy loss function;

[0151]

[0152] Among them, Loss represents the cross entropy loss function; G is the total number of samples; g is the g-th pseudo-copy; c g is the true label of the g-th sample; d g is the predicted probability of the g-th sample.

[0153] In this embodiment, a convolutional network is used to identify leakage signals, but attention should be paid to the size of the convolution kernel. If the convolution kernel is set too small, the features of the leakage signal may not be fully included in the receptive field, resulting in feature loss.

[0154] In order to prevent this problem from occurring, the convolutional neural network model used in this embodiment adopts a convolution check code with a length of 64 for feature extraction in the first convolution layer. This feature extraction method can reduce feature loss. In subsequent convolution layers, a convolution check code with a length of 3 is used for further refined extraction, and the features contained in the leakage signal are analyzed layer by layer. After the above operation, the features in the leakage signal can be effectively obtained, thereby ensuring the recognition accuracy of the model.

[0155] In this embodiment, the specific process of performing parameter optimization training on the convolutional neural network model in step 3 is:

[0156] A training sample data set and a training set label, wherein the training sample data set is a two-dimensional matrix composed of frequency domain filtered signals after passing through a BD filter, and the training set label is the signal type in the training sample data set, including leakage and non-leakage;

[0157] The training sample data set is input into the convolutional neural network model. After the logical output, it is compared with the training set label. The loss value is obtained through the cross entropy loss function. The convolutional neural network model is automatically updated according to the size of the loss value and the iterative operation is continued until a certain number of iterations is reached to obtain the optimal parameters with the minimum loss, such as Figure 2 As shown, the convolutional neural network model with the most optimized parameters is finally obtained, and the historical noisy signal of the input leakage point can be correctly classified;

[0158] The parameters are the weights and biases of each layer of the network.

[0159] In this embodiment, the specific method for evaluating the performance of the water supply pipeline leakage identification model is:

[0160] The best performing weight and bias information on the training set is loaded onto the test set for inference. Six indicators are used to evaluate the performance of the model: true positive rate, false positive rate, F1 score, accuracy, ROC curve, and AUC. The expressions are as follows:

[0161]

[0162] In the formula, TP means that the leak in the pipeline is correctly detected; FN means that the leak in the pipeline is not detected; TN means that the leak actually occurred and was successfully detected; FP means that the leak did not actually occur but was mistakenly detected as a leak; TPR is the true positive rate; FPR is the false positive rate; PRE is the accuracy rate;

[0163] like Figure 3As shown in the figure, AUC is the area under the ROC curve, which is used to quantify the performance of the ROC curve; the value range of AUC is between 0 and 1. The physical meaning of AUC is: randomly selecting a positive sample and a negative sample, the probability that the water supply pipeline leakage recognition model judges the positive sample score to be greater than the negative sample score; the closer the AUC is to 1.0, the stronger the classification ability of the water supply pipeline leakage recognition model.

[0164] In this embodiment, the first 75% of the data points of the same continuous leakage (non-leakage) signal are used as the training set, the last 25% are used as the test set, and 20% of the data points in the training set are used as the validation set and sent to the model for training. The same continuous leakage (non-leakage) signal processed by BD filtering is used as the test data set and sent to the model for testing.

[0165] In the test phase, every 2048 data points in this continuous signal are taken as a slice, and each slice in the test set is tested. The test results of all slices in the test set of this signal are summarized, and the test result with the most occurrences is taken as the final test result of this signal test set. The specific confusion matrix is ​​as follows Figure 4 The leakage detection performance of the water supply pipeline leakage identification model is shown in Table 1:

[0166] Table 1

[0167] Models F1 TPR FPR PRE AUC BDCNN 0.9971 0.9971 0.0028 0.9972 0.9998

[0168] The water supply pipeline leakage detection method based on BD joint convolutional neural network can realize the detection of water supply pipeline leakage and has high model recognition accuracy.

[0169] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A water supply pipeline leakage detection method based on BD combined with convolutional neural network, characterized in that: The following steps are involved: The current noisy signal of the leakage point is passed to the input end of the water supply pipeline leakage identification model, and the output end of the water supply pipeline leakage identification model outputs the detection result.

2. A water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 1, characterized in that: The generation method of the water supply pipeline leakage identification model is: Step 1: Obtain the historical noise signal of the leakage point; Step 2: Filter the historical noisy signal obtained in step 1 through a BD filter to obtain a frequency domain filtered signal; Step 3: Use the frequency domain filtering signal obtained in step 2 to perform parameter optimization training on the convolutional neural network model to generate a water supply pipeline leakage identification model.

3. A water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 2, characterized in that: The historical noise-containing signals include leakage sound signals of leaking points and non-leakage sound signals of leaking points; Among them, the method for collecting leakage sound signals at the leakage point is: to check the actual water supply pipeline network, and use a vibration sensor to collect leakage signals at the pipeline leakage point; the method for collecting non-leakage sound signals at the leakage point is: after repairing the pipeline leakage point, use the vibration sensor again to collect non-leakage signals at that point.

4. A water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 3, characterized in that: The specific process of obtaining the historical noise signal of the leakage point in step 1 is as follows: 35 leakage points in four different cities were selected, and the acoustic signals were sampled every 2 minutes at the leakage points. 72 leakage data and 72 non-leakage data were collected from each leakage point during the day and night. 2×2×72=288 data were collected from each leakage point, for a total of 35×288=10080 acoustic emission data. The sampling frequency of each data was 4762 Hz, and each data had a total of 42859 data points.

5. The water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 2 is characterized in that: In step 2, the BD filter includes a time domain quadratic filter and a frequency domain linear filter; The time domain secondary filter and the frequency domain linear filter are connected together in a cascade manner; wherein the time domain secondary filter serves as a first-stage BD filter; and the frequency domain linear filter serves as a second-stage BD filter.

6. A water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 5, characterized in that: The construction method of the BD filter in step 2 is: using the measured signal to reconstruct the leakage signal by estimating the transmission path function; Given the measured signal x, the leakage source signal d, and the noise n, the signal transfer process is defined as follows: x=d*h d +n*h n Among them, h d is the transfer function of the time domain quadratic filter; h n is the transfer function of the frequency domain linear filter, * is the convolution operation; The approximate leakage source signal y is restored by constructing a BD filter f. The process is defined as follows: y=x*f=(d*h d +n*h n )*f≈d The expression of the secondary convolutional neural network used by the time domain secondary filter is: y=σ((W1*x+b1)⊙(W2*x+b2)+W3*(x⊙x)+b3) Among them, σ(·) is the activation function ReLU, and the specific mathematical expression of the activation function ReLU is: σ(β)=max(0,β), β is the input of the activation function; W1 is the weight of the first layer of the quadratic neural network; W2 is the weight of the second layer of the quadratic neural network; W3 is the weight of the third layer of the quadratic neural network; b1 is the bias of the first layer of the quadratic neural network; b2 is the bias of the second layer of the quadratic neural network; b3 is the bias of the third layer of the quadratic neural network; ⊙ represents matrix dot multiplication; The loss function of the secondary convolutional neural network is expressed as: Among them, l t is the loss function; k is the kth layer of the quadratic neural network, N is the total number of layers of the quadratic neural network, and y(k) is the output of the quadratic convolutional neural network; The frequency domain is filtered using a fully connected neural network and Fourier transform to construct a frequency domain linear filter; The signal output by the time domain quadratic filter is defined as Apply the Fourier transform F(~) to convert the signal into the frequency domain: in, Output signal of the time domain quadratic filter Frequency domain signal after Fourier transformation; According to the convolution theorem, the expression for implementing the frequency domain linear filter using a fully connected neural network is as follows: Where N is the total number of layers of the quadratic neural network, i is the i-th layer of the quadratic neural network, and w i (f) is the weight of the i-th layer of the quadratic neural network at frequency f, b i is the bias of the i-th layer of the quadratic neural network; is the frequency domain signal after filtering; Apply the inverse Fourier transform F -1 Convert the signal to the time domain: in, is the time domain signal after inverse Fourier transform; Finally, an optimization function based on envelope spectrum is designed under the frequency domain linear filter; The design process of the optimization function is: The Hilbert transform is defined as: Among them, t is the time quantity; τ is the time offset; h(t) is the output of the input signal x(t) after Hilbert transform. According to the convolution theorem, we get: Where, j is an imaginary number; is the frequency domain signal under Hilbert transform; sgn(f) is the sign function; is the frequency domain signal after filtering; The discrete analytical signal z(n) obtained by Hilbert transform is: Among them, z(n) is the analytical signal in discrete form obtained by Hilbert transform; is a time domain signal after filtering in discrete form; is the time domain signal under the discrete form of Hilbert transform; Then the envelope spectrum ES(f) is defined as the Fourier spectrum of the analytical signal z(n) mode, and its mathematical expression is as follows: Then, the optimization function l f The expression is: Where f is the frequency of the quadratic neural network.

7. A water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 2, characterized in that: The specific definition method of the convolutional neural network model in step three is: Step 31: Determine the convolution layer; use and denote the weight and bias of the kernel of the i-th convolutional layer in layer l, and use u l (m) represents the mth local area in layer l, so the convolution process is described as follows: in, is the characteristic value of the mth local area in layer l+1; Step 32: Determine the pooling layer; The maximum pooling operation is used to sample the convolutional layer output, and its formula is as follows: Among them, γ is the γth local area to be pooled, is the input feature corresponding to the γth local area to be pooled under the i-th convolution kernel in layer l, W is the pooling depth, P i l+1 (j) is the output value of the ath pooling layer under the i-th convolution kernel in layer l+1; Step 33: Determine the cross entropy loss function; Among them, Loss represents the cross entropy loss function; G is the total number of samples; g is the g-th pseudo-copy; c g is the true label of the g-th sample; d g is the predicted probability of the g-th sample.

8. A water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 7, characterized in that: The specific process of parameter optimization training of the convolutional neural network model in step 3 is as follows: A training sample data set and a training set label, wherein the training sample data set is a two-dimensional matrix composed of frequency domain filtered signals after passing through a BD filter, and the training set label is the signal type in the training sample data set, including leakage and non-leakage; The training sample data set is input into the convolutional neural network model, and after logical output, it is compared with the training set label, and the loss value is obtained through the cross entropy loss function. The convolutional neural network model is automatically updated according to the size of the loss value and the iterative operation is continued until a certain number of iterations is reached to obtain the optimal parameters with the minimum loss, and the historical noisy signal of the input leakage point can be correctly classified; The parameters are the weights and biases of each layer of the network.

9. A water supply pipeline leakage detection method based on BD combined with convolutional neural network according to claim 1, 2 or 8, characterized in that: The specific method for evaluating the performance of the water supply pipeline leakage identification model is: The best performing weights and bias information on the training set are loaded onto the test set for inference. Six indicators are used to evaluate the performance of the model: true positive rate, false positive rate, F1 score, accuracy, ROC curve, and AUC. Its expression is as follows: In the formula, TP means that the leak in the pipeline is correctly detected; FN means that the leak in the pipeline is not detected; TN means that the leak actually occurred and was successfully detected; FP means that the leak did not actually occur but was mistakenly detected as a leak; TPR is the true positive rate; FPR is the false positive rate; PRE is the accuracy rate; AUC is the area under the ROC curve, which is used to quantify the performance of the ROC curve. The value range of AUC is between 0 and 1. The physical meaning of AUC is: randomly selecting a positive sample and a negative sample, the probability that the water supply pipeline leakage recognition model determines that the positive sample score is greater than the negative sample score; The closer the AUC is to 1.0, the stronger the classification ability of the water supply pipe leakage identification model is.

Images