Method and system for identifying working sound signals of water supply pipeline

Abstract

The present application discloses a method and system for identifying working condition sound signals of a water supply pipeline, the method comprising: obtaining working condition sound signals of the inner wall of the water supply pipeline; inputting the working condition sound signals into a leakage identification model: inputting the working condition sound signals into a first processing branch line, obtaining a first result feature related to the characteristics of the working condition sound signals in the time domain; inputting the working condition sound signals into a second processing branch line, obtaining a second result feature related to the characteristics of the working condition sound signals in the frequency domain; inputting the working condition sound signals into a third processing branch line, obtaining a third result feature related to the characteristics of the working condition sound signals in the modal domain; splicing the three result features, and obtaining an output result of a water supply pipeline leak or a water supply pipeline no leak. The present application realizes water supply pipeline leakage detection through working condition sound signals propagated through the inner wall of the water supply pipeline, thereby improving the accuracy and robustness of water supply pipeline leakage detection.

Description

Technical Field

[0001] The present application relates to the technical field of water supply pipeline leakage detection, and in particular to a method and system for identifying working condition sound signals of a water supply pipeline. Background Art

[0002] Leakage in water network pipelines will not only cause a huge waste of water resources, but also bring serious economic losses and even disasters to related companies and families. In related technologies, according to different pipeline leakage detection methods, it can be divided into direct leak detection method and indirect leak detection method. The direct leak detection method relies on manual regular inspection of the pipeline status. The indirect leak detection method relies on sensors to monitor physical variables such as flow and magnetic flux in real time. The direct leak detection method can only detect shallow buried water supply pipelines. The indirect leak detection method requires the purchase of professional tracer equipment or has problems such as high noise and poor anti-interference ability, resulting in low detection accuracy. Summary of the invention

[0003] In order to address the deficiencies of the prior art, the purpose of the present application is to provide a method and system for identifying operating sound signals of a water supply pipe, which can realize water supply pipe leakage detection through the operating sound signals propagated through the inner wall of the water supply pipe, thereby improving the accuracy and robustness of water supply pipe leakage detection.

[0004] To achieve the above objectives, this application adopts the following technical solutions:

[0005] In a first aspect, the present application provides a method for identifying an operating sound signal of a water supply pipe, the method comprising: obtaining an operating sound signal of an inner wall of the water supply pipe; inputting the operating sound signal into a leakage identification model to obtain an output result, the output result being that the water supply pipe is leaking or that the water supply pipe is not leaking; wherein the leakage identification model comprises a first processing branch, a second processing branch and a third processing branch in parallel; inputting the operating sound signal into the leakage identification model to obtain an output result comprises: inputting the operating sound signal into the first processing branch to obtain a first result feature, the first result feature being related to a feature of the operating sound signal in the time domain; inputting the operating sound signal into the second processing branch to obtain a second result feature, the second result feature being related to a feature of the operating sound signal in the frequency domain; inputting the operating sound signal into the third processing branch to obtain a third result feature, the third result feature being related to a feature of the operating sound signal in the modal; splicing the first result feature, the second result feature and the third result feature to obtain an output result.

[0006] In one embodiment, the first processing branch and the second processing branch each include at least one sub-branch, and the third processing branch includes a plurality of sub-branches.

[0007] In one embodiment, an operating sound signal is input into a third processing branch to obtain a third result feature, including: performing modal decomposition on the operating sound signal using an adaptive FIR filter in the third processing branch to obtain multiple modal components; inputting the multiple modal components into multiple sub-branches in the third processing branch, wherein one modal component corresponds to one sub-branch; for each modal component in a sub-branch, determining a modal learning network corresponding to the modal component from a modal learning network set, and using the corresponding modal learning network to process the modal component on the sub-branch to obtain a modal transition feature corresponding to the modal component; wherein the modal learning network set includes multiple modal learning networks, parameters of the modal learning network are related to the modal components, and one modal component corresponds to at least one modal learning network; performing modal fusion based on the multiple modal transition features obtained in the multiple branches to obtain the third result feature.

[0008] In one embodiment, the third processing branch includes four sub-branches; the adaptive FIR filter in the third processing branch is used to perform modal decomposition on the operating sound signal to obtain four modal components; four corresponding modal learning networks are used to process the modal components in the four sub-branches to obtain four corresponding modal transition features; wherein the features of the operating sound signal represented by the four modal components are different, and the time domain diagrams and spectrum diagrams corresponding to the four modal components are different.

[0009] In one embodiment, modal fusion is performed based on multiple modal transition features obtained from multiple sub-branches to obtain a third result feature, including: each sub-branch in the third processing branch includes a multi-layer perceptron, four modal transition features are used as four-way vector signals, and each vector signal is input into the multi-layer perceptron in the corresponding sub-branch to obtain one output vector; the four output vectors are added together to obtain the third result feature of the operating sound signal.

[0010] In one embodiment, the modal learning network is a one-dimensional network, and the modal learning network includes at least a convolution block group, a residual block group and an output layer group; wherein the convolution block group includes a front convolution block and a rear convolution block, and the residual block group includes a front residual block and a rear residual block; for the modal component in each sub-branch, the corresponding modal learning network is used to process the modal component to obtain a modal transition feature corresponding to the modal component, including: for the modal component in each sub-branch, the modal component is input into the front convolution block, the output of the front convolution module is used as the input of the front residual block, the output of the front residual block is used as the input of the rear convolution block, the output of the rear convolution block is used as the input of the rear residual block, the output of the rear residual block is used as the input of the output layer group, and the output layer group outputs the modal transition feature corresponding to the modal component.

[0011] In one embodiment, an operating condition sound signal is input into a first processing branch to obtain a first result feature, including: obtaining a target signal segment corresponding to the operating condition sound signal, and processing the operating condition sound signal to obtain a target signal that matches the target signal segment; using a time domain learning network to process the target signal to obtain a time domain feature of the operating condition sound signal; wherein the time domain learning network includes multiple convolution blocks and multiple residual blocks, and the convolution kernel size of the convolution block is different from the convolution kernel size of the residual block; processing the operating condition sound signal to obtain a target signal that matches the target signal segment, including: using a filter to remove non-target signals in the operating condition sound signal to obtain a target signal, and the passband cutoff frequency of the filter is less than the stopband cutoff frequency; inputting the operating condition sound signal into a second processing branch to obtain a second result feature, including: performing a fast Fourier transform on the operating condition sound signal input into the second processing branch to obtain a frequency domain transformed signal; using the frequency domain learning network on the branch to process the frequency domain transformed signal to obtain a second result feature.

[0012] In one embodiment, the first result feature, the second result feature and the third result feature are spliced ​​to obtain the output result, including: when the dimensions and lengths of the first result feature, the second result feature and the third result feature are consistent, using a set function to vector splice the first result feature, the second result feature and the third result feature along a set direction to obtain the output result.

[0013] In one embodiment, a first data set is obtained, the first data set includes positive samples and negative samples, the positive samples are operating sound signals when a first preset number of water supply pipes leak, and the negative samples are operating sound signals when a second preset number of water supply pipes are not leaking; the first data set is used to pre-train a leakage identification model; a second data set is obtained, the second data set includes a first factor and a second factor, the first factor is the ratio of positive samples predicted to be positive samples to all predicted positive samples, and the second factor is the ratio of positive samples predicted to be positive samples to actual positive samples; based on the first factor and the second factor, the leakage identification model is corrected.

[0014] In a second aspect, the present application provides an identification system for operating sound signals of a water supply pipe, the identification system comprising a signal acquisition device and an electronic device; the signal acquisition device is configured to collect operating sound signals from the inner wall of the water supply pipe; the electronic device is configured to obtain the operating sound signals collected by the signal acquisition device, and identify the operating sound signals according to any method of the first aspect.

[0015] The above-mentioned method and system for identifying the operating sound signal of the water supply pipeline obtains the operating sound signal of the inner wall of the water supply pipeline, and inputs the operating sound signal into a leakage identification model including multiple processing branches, so that the operating sound signal passes through the first processing branch to obtain a first result feature related to the characteristics of the operating sound signal in the time domain, passes through the second processing branch to obtain a second result feature related to the characteristics of the operating sound signal in the frequency domain, passes through the third processing branch to obtain a third result feature related to the characteristics of the operating sound signal in the modal domain, and then splices the first result feature, the second result feature and the third result feature of the operating sound signal to obtain an output result of water supply pipeline leakage or water supply pipeline non-leakage. This method can achieve more accurate and comprehensive feature recognition of complex operating sound signals through a leakage identification model that includes multiple processing branches. It also improves the robustness of identifying interference and noise in the operating sound signals while being sensitive to identifying abnormal features of the operating sound signals, so as to avoid the problem of poor anti-interference ability of pipeline leakage detection due to the weakening of complex operating sound signals due to medium absorption, thereby improving the accuracy and robustness of water supply pipeline leakage detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the structure of a leakage identification model provided by an embodiment;

[0017] Figure 2 A schematic diagram of an environment when a water supply pipeline leaks provided by an embodiment;

[0018] Figure 3 A first flow chart of a method for identifying working condition sound signals of a water supply pipeline provided by an embodiment;

[0019] Figure 4 A schematic diagram of performing modal decomposition based on an adaptive FIR filter provided by an embodiment;

[0020] Figure 5 A schematic diagram of modal fusion based on a multi-layer perceptron provided by an embodiment;

[0021] Figure 6 A schematic diagram of the structure of a one-dimensional network model provided by an embodiment;

[0022] Figure 7 A schematic diagram of a filter frequency response provided for one embodiment;

[0023] Figure 8 A schematic diagram of the structure of a residual block provided by an embodiment;

[0024] Fig. 9 A schematic diagram of the structure of a leakage identification model provided by another embodiment;

[0025] Fig.10 A second flow chart of a method for identifying working condition sound signals of a water supply pipeline provided by an embodiment;

[0026] Fig.11 A structural block diagram of a system for identifying working condition sound signals of a water supply pipeline provided in one embodiment. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the specific implementation manner of the present application will be clearly and completely described below in conjunction with the drawings in the implementation manner of the present application.

[0028] In the related art, according to the different ways of pipeline leakage detection, it can be divided into direct leak detection method and indirect leak detection method. Among them, the indirect leak detection method includes pipeline leak detection based on acoustic signals. The detection method based on the propagation of acoustic signals in the medium around the water supply pipeline is to use some equipment to monitor the acoustic signals on the ground above the water supply pipeline, and determine the leakage situation and the location of the leak by analyzing the degree of attenuation of the acoustic signals during the propagation process. However, when the leakage point is far away or the pipeline is buried deep, the acoustic signal will become weak due to the absorption of the medium (such as stones, gravel and other medium particles), that is, the detection method based on the propagation of acoustic signals in the medium around the water supply pipeline has poor anti-interference ability. In the related art, the detection method based on the propagation of acoustic signals on the inner wall of the water supply pipeline is prone to problems of high noise and low precision in the face of complex working conditions.

[0029] The method for identifying the operating sound signal of a water supply pipeline provided in the present application determines the leakage of the water supply pipeline by detecting the operating sound signal of the inner wall of the water supply pipeline. This can avoid the problem that the operating sound signal becomes weak due to the absorption of the medium, resulting in poor anti-interference ability of pipeline leakage detection and a narrow range of use. At the same time, the operating sound signal is input into a leakage identification model including multiple processing branches to obtain an output result of water supply pipeline leakage or water supply pipeline no leakage. The complex operating sound signal is identified and processed from multiple angles and directions, thereby improving the robustness and accuracy of water supply pipeline leakage detection.

[0030] The following is a detailed description of the method for identifying working sound signals of a water supply pipeline in the present application in conjunction with the accompanying drawings.

[0031] like Figure 1The figure shows a schematic diagram of the structure of a leakage identification model provided by an embodiment of the present application. Based on the leakage identification model, a method for identifying the working condition sound signal of a water supply pipeline includes: obtaining the working condition sound signal of the inner wall of the water supply pipeline; inputting the working condition sound signal into the leakage identification model to obtain an output result, and the output result is that the water supply pipeline leaks or the water supply pipeline has no leakage. Among them, the leakage identification model includes a first processing branch line, a second processing branch line and a third processing branch line in parallel.

[0032] Specifically, inputting the operating sound signal into the leakage identification model to obtain the output result includes: inputting the operating sound signal into the first processing branch to obtain a first result feature, and the first result feature is related to the feature of the operating sound signal in the time domain; inputting the operating sound signal into the second processing branch to obtain a second result feature, and the second result feature is related to the feature of the operating sound signal in the frequency domain; inputting the operating sound signal into the third processing branch to obtain a third result feature, and the third result feature is related to the feature of the operating sound signal in the modal domain; and splicing the first result feature, the second result feature and the third result feature to obtain the output result.

[0033] It should be noted that if there is a leak in the water supply pipeline, the operating sound signal at the leak point will produce specific signal characteristics, such as Figure 2 The figure shows a schematic diagram of the environment when the water supply pipeline leaks. In the case of a water supply pipeline leak, the operating sound signal includes a pipeline vibration signal generated by fluid turbulence inside the water supply pipeline, an impact vibration signal generated by fluid leakage impacting the medium, and a medium friction vibration signal generated by friction collision caused by the injection of fluid in the water supply pipeline. These sound signals are different from the operating sound signals when the pipeline is not leaking. When the output result contains at least one of the pipeline vibration signal, impact vibration signal and medium friction vibration signal, it can be determined that the water supply pipeline is leaking.

[0034] Since the operating sound signal propagates along the inner wall of the pipeline, the operating sound signal of the inner wall of the pipeline can be captured by sensors installed around the outside of the pipeline (such as magnetic piezoelectric acceleration sensors). For example, if the sensors are installed on the outside of the pipeline at a set distance, if a leak occurs in the pipeline between the two sensors, the operating sound signal can be captured by any sensor. For another example, since pipeline leakage often occurs at the pipeline connection, the sensor can be installed at the connection between the pipeline and another pipeline to ensure that the leaking operating sound signal can be captured.

[0035] Since the water supply pipeline is located underground and the sensor collection environment is usually complex, the collected working sound signal usually contains random interference noise. Therefore, the working sound signal is input into a leakage identification model including multiple processing branches to process the complex working sound signal with high precision and multi-angle, so as to obtain the output result of whether the water supply pipeline has a leak. Among them, in the first processing branch, fast Fourier transform, time domain filtering algorithm, etc. can be used to analyze the characteristics of the working sound signal changing over time, that is, the first result feature. The first result feature is related to the characteristics of the working sound signal in the time domain, such as the waveform, amplitude, phase, period, etc. of the working sound signal, which can reflect the intensity and change trend of the working sound signal.

[0036] In the second processing branch, Fourier transform, wavelet transform, etc. can be used to analyze the distribution characteristics of the working condition sound signal in frequency, that is, the second result characteristics. The second result characteristics are related to the characteristics of the working condition sound signal in the frequency domain, such as the frequency mean of the working condition sound signal, the energy of the frequency component, the frequency width, etc. The second result characteristics can reflect the distribution of the working condition sound signal at different frequencies.

[0037] In the third processing branch, the characteristic mode decomposition (FMD) can be used to perform modal decomposition on the working condition sound signal to extract the third result feature. The third result feature is related to the modal characteristics of the working condition sound signal, such as the energy distribution and spectrum of the working condition sound signal. The characteristic mode decomposition takes into account the impulsiveness and periodicity of the working condition sound signal at the same time. It can decompose the original working condition sound signal into the third result feature with local time-frequency properties, reflecting the vibration mode of the working condition sound signal on different time scales and in different frequency ranges, so as to understand and analyze the structure and characteristics of the working condition sound signal in a targeted manner.

[0038] Furthermore, the first result feature, the second result feature and the third result feature can be combined by using methods such as feature fusion, feature selection or feature weighting. For example, a support vector machine, random forest, neural network or other classification algorithms can be used to combine these features to obtain an output result that the water supply pipeline is leaking or the water supply pipeline is not leaking.

[0039] In summary, the method for identifying the operating sound signal of the water supply pipe in this embodiment obtains the operating sound signal of the inner wall of the water supply pipe, and inputs the operating sound signal into a leakage identification model including multiple processing branches, so that the operating sound signal passes through the first processing branch to obtain a first result feature related to the characteristics of the operating sound signal in the time domain, passes through the second processing branch to obtain a second result feature related to the characteristics of the operating sound signal in the frequency domain, passes through the third processing branch to obtain a third result feature related to the characteristics of the operating sound signal in the modal domain, and then splices the first result feature, the second result feature and the third result feature of the operating sound signal to obtain an output result of water supply pipe leakage or water supply pipe non-leakage. This method can achieve more accurate and comprehensive feature recognition of complex operating sound signals through the leakage identification model that includes multiple processing branches. It also improves the robustness of identifying interference and noise in the operating sound signals while being sensitive to identifying abnormal features of the operating sound signals, so as to avoid the problem of poor anti-interference ability of pipeline leakage detection due to the weak complex operating sound signals due to absorption by the medium, thereby improving the accuracy and robustness of water supply pipeline leakage detection.

[0040] In one embodiment, the first processing branch and the second processing branch each include at least one sub-branch, and the third processing branch includes multiple sub-branches. Each sub-branch in the third processing branch processes and extracts different features, for example, one branch extracts frequency distribution, one branch extracts energy distribution, or one branch extracts a specific spectrum graph feature, and one branch extracts another specific spectrum graph feature, thereby further reflecting the vibration mode of the working condition sound signal on different time scales and in different frequency ranges, making the understanding and analysis of the structure and characteristics of the working condition sound signal more targeted, achieving more accurate and comprehensive feature recognition, and further improving the accuracy and robustness of water supply pipeline leakage detection.

[0041] In order to improve the robustness of water supply pipeline leakage identification, in one embodiment, Figure 3 As shown, the working condition sound signal is input into the third processing branch to obtain the third result feature, including the following steps:

[0042] Step S301: Use the adaptive FIR filter in the third processing branch to perform modal decomposition on the working condition sound signal to obtain multiple modal components.

[0043] Among them, the adaptive FIR (Finite Impulse Response) filter is a digital filter that can adaptively adjust its coefficients according to the input signal. It can optimize its performance according to the characteristics of the working condition sound signal. Modal decomposition can decompose complex working condition sound signals into modal components (Intrinsic Mode Function, IMF) with different frequencies and amplitudes. The modal components represent the vibration mode of the working condition sound signal in the time dimension.

[0044] Step S302: input multiple modal components into multiple sub-branches in the third processing branch, wherein one modal component corresponds to one sub-branch.

[0045] Specifically, different weights may be set for the multiple sub-branches to respectively perform modal decomposition of the multiple modal components to different degrees, and extract multiple features of the operating sound signal.

[0046] Step S303: for each modal component in a sub-branch, determine the modal learning network corresponding to the modal component from the modal learning network set, and use the corresponding modal learning network to process the modal component on the sub-branch to obtain the modal transition characteristics corresponding to the modal component; wherein the modal learning network set includes multiple modal learning networks, the parameters of the modal learning network are related to the modal component, and one modal component corresponds to at least one modal learning network.

[0047] Among them, the modal learning network is a pre-trained deep learning network. Each modal learning network is specifically used to process a specific type of modal component to obtain the modal transition characteristics corresponding to the modal component. The modal transition characteristics represent the vibration mode of the operating sound signal in different frequency ranges.

[0048] Step S304: Perform modal fusion based on multiple modal transition features obtained from multiple branches to obtain a third result feature.

[0049] Among them, modal fusion can adopt a variety of methods, such as attention mechanism, feature weighting or feature splicing, etc. The purpose of modal fusion is to integrate the modal transition features from different modal components to form a comprehensive feature to represent the third result feature. This comprehensive feature can more comprehensively describe the depth characteristics of the working condition sound signal. The final third result feature will be used to analyze whether there is a leak in the water supply pipeline.

[0050] Through the above description, the third result feature reflecting the intrinsic depth characteristics of the operating sound signal can be more effectively extracted from the operating sound signal, so as to understand and analyze the structure and characteristics of the operating sound signal more specifically and improve the robustness of the leakage identification model.

[0051] In one embodiment, Figure 4As shown, the third processing branch includes four sub-branches; the adaptive FIR filter in the third processing branch is used to perform modal decomposition on the operating sound signal to obtain four modal components; four corresponding modal learning networks are used to process the modal components in the four sub-branches to obtain four corresponding modal transition features; wherein, the characteristics of the operating sound signal represented by the four modal components are different, and the time domain diagrams and spectrum diagrams corresponding to the four modal components are different. For example, one modal component mainly contains low-frequency leakage sound, while another modal component mainly contains high-frequency noise or environmental sound. The time domain diagram of each modal component shows the change of sound pressure over time, and the spectrum diagram shows the intensity of different frequency components. In this way, the leakage identification model can understand the operating sound signal more comprehensively, thereby improving the leakage detection performance in complex environments.

[0052] In one embodiment, modal fusion is performed based on multiple modal transition features obtained from multiple sub-branches to obtain a third result feature, including: each sub-branch in the third processing branch includes a multi-layer perceptron, four modal transition features are used as four-way vector signals, and each vector signal is input into the multi-layer perceptron in the corresponding sub-branch to obtain one output vector; the four output vectors are added together to obtain the third result feature of the operating sound signal.

[0053] A multi-layer perceptron is a deep neural network, including an input layer, a hidden layer, and an output layer. In this embodiment, a two-layer multi-layer perceptron is used, which is more helpful to extract and strengthen each third result feature. Each modal transition feature corresponds to a multi-layer perceptron. For example, the best multi-layer perceptron corresponding to each modal transition feature is determined by cross-validation. Figure 5 As shown in FIG. 1 , a schematic diagram of modal fusion is shown. The modal transition feature can be converted into a vector signal, and the vector signal is input into the corresponding multi-layer perceptron. The multi-layer perceptron processes the input vector signal through its internal hidden layer and nonlinear activation layer to obtain an output vector. The four output vectors are added together, which can be a simple element-by-element addition or a more complex weighting or other fusion strategy to obtain the third result feature of the working condition sound signal.

[0054] In this embodiment, through the hierarchy of the multi-layer perceptron, the abstract modal transition features are converted into more advanced, abstract and effective third result features, so that the leakage identification model can identify the operating sound signals from different angles and levels, thereby improving the accuracy and robustness of the identification.

[0055] In order to avoid the problem that the accuracy of the modal learning network of deep learning decreases as it gradually deepens, in one embodiment, the modal learning network is a one-dimensional network model, and the modal learning network at least includes a convolution block group, a residual block group, and an output layer group, wherein the convolution block group includes a front convolution block and a rear convolution block, and the residual block group includes a front residual block and a rear residual block. For the modal components in each sub-branch, the corresponding modal learning network is used to process the modal components to obtain modal transition features corresponding to the modal components, including: for the modal components in each sub-branch, the modal components are input to the front convolution block, the output of the front convolution module is used as the input of the front residual block, the output of the front residual block is used as the input of the rear convolution block, the output of the rear convolution block is used as the input of the rear residual block, the output of the rear residual block is used as the input of the output layer group, and the output layer group outputs the modal transition features corresponding to the modal components.

[0056] like Figure 6 As shown in FIG. , it is a schematic diagram of the network structure of a one-dimensional network model, wherein the one-dimensional network model is a one-dimensional residual network. Specifically, the front convolution block and the rear convolution block respectively include a convolution layer, an activation layer, and a pooling layer. The front convolution block is used to perform preliminary feature extraction on the modal component, and the rear convolution block further extracts features from the output of the front residual block. The convolution kernel size of the rear convolution block can be different from that of the front convolution block to meet the requirements of feature extraction at different levels. The front residual block and the rear residual block respectively include a convolution layer, an activation layer, and a pooling layer. The output result of the convolution layer is added to the input of the front residual block or the rear residual block, and then processed by another activation layer to realize jump connection to solve the gradient disappearance problem and allow the one-dimensional network model to learn more complex features to capture higher-level abstract features. The output layer group includes two fully connected layers, which map the features extracted by the rear residual block to the final modal transition features. The network structure parameters of each layer of the one-dimensional residual network are shown in Table 1.

[0057] Table 1

[0058]

[0059] In this way, highly representative modal transition features can be extracted from the operating sound signal, thereby avoiding the problem of decreased accuracy of the modal learning network in the process of gradual deepening of deep learning and improving the accuracy of the modal learning network.

[0060] In one embodiment, the adaptive FIR filter is used to perform modal decomposition on the working condition sound signal to obtain multiple modal components, which specifically includes the following steps:

[0061] a. Load the operating sound signal and enter the preset mode number n.

[0062] b. Initialize the filter bank, where the filter bank includes k FIR filters, where k is an integer greater than 1.

[0063] c. Start iteration by decomposing the working sound signal with a filter bank to obtain multiple modal components.

[0064] d. Estimate the modal period and update the coefficients of the FIR filter based on the operating sound signal. The modal period can be estimated by analyzing the autocorrelation spectrum of the operating sound signal, and the coefficients of the FIR filter can be updated accordingly. The autocorrelation spectrum describes the view of the similarity between the operating sound signal and itself at different time lags as the frequency changes. The filter coefficients are updated by the point where the autocorrelation spectrum reaches the local maximum after the zero point, which can identify the periodic characteristics of the operating sound signal to more effectively estimate the modal period.

[0065] e. Determine whether the number of iterations reaches the preset number, if not, return to step c. The preset number can be set according to experience.

[0066] f. Calculate the co-correlation coefficient of the modal component and obtain the coefficient matrix, determine the kurtosis value based on the estimated modal period, and remove the modal component with the smallest kurtosis value. The co-correlation coefficient is used to determine the similarity or linear relationship between two modal components. The kurtosis is a characteristic number that characterizes the peak height of the data distribution curve at the mean value. The kurtosis value is sensitive to extreme values. Therefore, the modal component with the smallest kurtosis value is removed to obtain the third result feature of the more impulsive working sound signal.

[0067] g. Determine whether the modal number reaches the preset modal number n. If not, return to step c to continue iteration; if yes, take the retained modal component as the final modal component.

[0068] In this way, the decomposition is not limited by the filter shape, bandwidth and center frequency. Moreover, the impulsiveness and periodicity of the operating sound signal can be considered at the same time. It is targeted at the operating sound signal and robust to other interference and noise signals.

[0069] In order to reduce the impact of noise in the operating sound signal on the recognition accuracy, in one embodiment, the operating sound signal is input into at least one sub-branch in the first processing branch to obtain a first result feature, including: obtaining a target signal segment corresponding to the operating sound signal, and processing the operating sound signal to obtain a target signal that matches the target signal segment; using a time domain learning network to process the target signal to obtain the time domain feature of the operating sound signal; wherein the time domain learning network includes multiple convolution blocks and multiple residual blocks, and the convolution kernel size of the convolution block is different from the convolution kernel size of the residual block.

[0070] Specifically, the target signal segment is extracted from the original working sound signal and is directly related to the water supply pipeline leakage. The time domain learning model is a deep learning network used to process time series data and extract features from it. The structure, parameter setting and feature extraction operation of the time domain learning model refer to the processing process of the above-mentioned one-dimensional network model. The time domain learning model also includes multiple convolution blocks and multiple residual blocks. The convolution kernel size of each convolution block is the same and usually small to capture the local features of the target signal. The residual block helps to solve the gradient vanishing problem in the deep network. The convolution kernel size of each residual block is the same, but the convolution kernel size of the residual block is different from the convolution kernel size of the convolution block to capture a wider range of features in the target signal. By inputting the target signal into the time domain learning network, after alternating stacking of multiple layers of convolution blocks and residual blocks, and the features extracted by each convolution block and residual block are accumulated in the depth direction, rich time domain features are gradually extracted from the target signal, thereby reducing the impact of noise in the working sound signal on the recognition accuracy and improving the accuracy of identifying water supply pipeline leakage.

[0071] In one embodiment, the operating sound signal is processed to obtain a target signal that matches the target signal segment, including: using a filter to remove non-target signals in the operating sound signal to obtain the target signal, and the passband cutoff frequency of the filter is less than the stopband cutoff frequency. Specifically, the passband cutoff frequency refers to the highest effective frequency of the filter within the transmission range, and the stopband cutoff frequency refers to the degree to which the filter blocks certain frequency ranges reaching a certain threshold. In the filter of this embodiment, the passband cutoff frequency is preferably 0.3πHZ, and the stopband cutoff frequency is preferably 0.5πHZ. The filter can also set the stopband attenuation to -40 decibels, and the window function selects the Hanning window. At this time, the filter frequency response is as follows: Figure 7 In this way, high-frequency noise or other interference can be removed, and low-frequency signals related to leakage can be retained to reduce the interference of noise and other irrelevant signals.

[0072] In one embodiment, the operating condition sound signal is input into at least one sub-branch in the second processing branch to obtain the second result feature, including: performing a fast Fourier transform on the operating condition sound signal input into at least one sub-branch in the second processing branch to obtain a frequency domain transformation signal; using the frequency domain learning network on the sub-branch to process the frequency domain transformation signal to obtain the second result feature. Among them, the fast Fourier transform (Fast Fourier Transform, FFT) can convert the signal from the time domain to the frequency domain. For the operating condition sound signal, the fast Fourier transform can reveal the components and intensities of the operating condition sound signal at different frequencies, and obtain the spectrum of the frequency domain transformation signal. The frequency domain learning network is a deep learning network used to process the frequency domain transformation signal and extract the second result feature therefrom. The structure, parameter setting and feature extraction operation of the frequency domain learning network refer to the processing process of the above-mentioned one-dimensional network model. Through the frequency domain learning network, the frequency domain transformation signal is deeply learned, and a highly representative second result feature can be extracted from the frequency domain transformation signal, thereby avoiding the problem of the accuracy of the frequency domain learning network of deep learning decreasing in the process of gradual deepening, and improving the accuracy of the frequency domain learning network.

[0073] In one embodiment, the architecture and parameters of the modality learning network, the time domain learning network, and the frequency domain learning network are the same. The learning network includes a convolution block and a residual block, wherein the convolution block outputs two original features to the residual block, wherein one original feature is processed by convolution, activation, and pooling in the residual block and then outputs the processed intermediate feature, and the intermediate feature is added to the other original feature and processed by activation in the residual block as the output result of the residual block.

[0074] More specifically, the residual block contains a convolution layer, an activation layer, and a pooling layer. The convolution layer is used to implement the convolution operation, the activation layer is used to implement the activation operation through the activation function, and the pooling layer is used to implement the pooling operation. The internal structure of the residual block is as follows Figure 8 As shown in the figure, the two original features output by the convolution block are sent to the residual block, one of which is processed by the convolution layer, activation layer and pooling layer in the residual block to output the intermediate feature, and the other original feature is added to the intermediate feature and activated by the activation layer to achieve jump connection to solve the gradient disappearance problem and obtain the output result of the residual block, thereby capturing higher-level abstract features.

[0075] In one embodiment, the first result feature, the second result feature and the third result feature of the operating sound signal are combined to obtain an output result, including: when the dimensions and lengths of the first result feature, the second result feature and the third result feature are consistent, a set function is used to vector-spin the first result feature, the second result feature and the third result feature along a set direction to obtain an output result.

[0076] Among them, in order to ensure that the lengths of the first result feature, the second result feature, and the third result feature are consistent, the lengths can be adjusted by padding or truncating. The setting function is used to perform vector concatenation, and the setting direction can be the row direction or the column direction. For example, assuming that the dimensions of the first result feature, the second result feature, and the third result feature are all one-dimensional and have the same length, the first result feature is represented as time_features, the second result feature is represented as freq_features, and the third result feature is represented as modal_features, then, using the setting function concatenate(), the output result along the column direction (axis=1) can be expressed as F_output =concatenate ((time_features, freq_features, modal_features), axis=1).

[0077] Through the above method, the first result feature, the second result feature and the third result feature of the operating sound signal can be effectively integrated, and the features can be spliced ​​by setting functions. This method improves the utilization rate and analysis accuracy of complex operating sound signals, and provides more effective information support for leakage identification of water supply pipelines.

[0078] In one embodiment, Fig. 9 As shown, it is a structural diagram of another leakage identification model. The working condition sound signal is input into the leakage identification model, and the working condition sound signal is pre-processed by the low-pass filter of the first processing branch to obtain the target signal, and then input into the time domain learning network to obtain the first result feature of the working condition sound signal; the working condition sound signal is subjected to fast Fourier transform through the second processing branch to obtain the frequency domain transformation signal, and then input into the frequency domain learning network for processing to obtain the second result feature of the working condition sound signal; the working condition sound signal is subjected to modal decomposition by using an adaptive FIR filter through the third processing branch to obtain four modal components, and then input into the modal learning network for processing to obtain four corresponding modal transition features, and then the four modal transition features are used as four-way vector signals, input into the corresponding multi-layer perceptron, and four-way output vectors are obtained, and then the obtained four-way output vectors are added to obtain the third result feature of the working condition sound signal. Finally, the first result feature, the second result feature and the third result feature are spliced ​​to obtain the output result of water supply pipeline leakage or water supply pipeline non-leakage.

[0079] In order to improve the accuracy and reliability of the leakage identification model, in one embodiment, Fig.10 As shown, the method for identifying the working sound signal of the water supply pipeline also includes the following steps:

[0080] Step S1001, obtaining a first data set, the first data set including positive samples and negative samples, the positive samples are operating sound signals when a first preset number of water supply pipes are leaking, and the negative samples are operating sound signals when a second preset number of water supply pipes are not leaking.

[0081] Step S1002: pre-training a leakage recognition model using the first data set.

[0082] Step S1003 , obtaining a second data set, the second data set includes a first factor and a second factor, the first factor is the ratio of positive samples predicted as positive samples to all predicted positive samples, and the second factor is the ratio of positive samples predicted as positive samples to actual positive samples.

[0083] Step S1004: modifying the leakage identification model based on the first factor and the second factor.

[0084] Exemplarily, the first preset number is 2000, the second preset number is 1100, the positive samples are 2000 working sound signals of the water supply pipeline leaking with a duration of 1 minute, and the negative samples are 1100 working sound signals of the water supply pipeline without leakage with a duration of 1 minute. The leakage identification model is pre-trained using the first data set to obtain a data set in which positive samples are predicted to be positive samples and negative samples are predicted to be positive samples, as the second data set. Furthermore, the proportion of positive samples predicted to be positive samples to all predicted positive samples is obtained by calculation, which is the first factor, also known as the precision rate; the proportion of positive samples predicted to be positive samples to actual positive samples is the second factor, also known as the recall rate. Based on the first factor and the second factor, the ratio of twice the product of the first factor and the second factor to the sum of the first factor and the second factor can also be obtained as a reference score, and the first factor, the second factor and the reference score are combined to correct the leakage identification model.

[0085] Table 2 shows a comparison of the correction effects of the leakage identification model of this embodiment and the support vector machine. It can be found that the recognition accuracy of the leakage identification model of this embodiment can reach 99%, which can better realize the recognition of water supply pipeline leakage.

[0086] Table 2

[0087]

[0088] Through the above description, pre-training is performed through the first data set to obtain the first factor and the second factor, and the leakage identification model is corrected, so as to continuously optimize the structure and parameters of the leakage identification model and improve the accuracy and reliability of the leakage identification model.

[0089] Based on the same inventive concept, the present application also provides a system for implementing the above-mentioned method for identifying working condition sound signals of water supply pipes. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in the embodiment of the recognition system for working condition sound signals of water supply pipes provided below can be referred to the limitations of the recognition method for working condition sound signals of water supply pipes in the above text, and will not be repeated here.

[0090] like Fig.11 As shown, the embodiment of the present application provides a system for identifying working condition sound signals of a water supply pipeline, and the identification system includes a signal acquisition device 1101 and an electronic device 1102;

[0091] The signal acquisition device 1101 is configured to collect the working condition sound signal of the inner wall of the water supply pipe;

[0092] The electronic device 1102 is configured to obtain the operating sound signal collected by the signal collection device 1101, and identify the operating sound signal according to the method for identifying the operating sound signal of the water supply pipeline in any of the above embodiments.

[0093] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the claims attached to this application.

Claims

1. A method for identifying working condition sound signals of a water supply pipeline, characterized in that: The method comprises: Acquire the working sound signal of the inner wall of the water supply pipe; Inputting the working condition sound signal into a leakage identification model to obtain an output result, wherein the output result is a water supply pipeline leakage or a water supply pipeline no leakage; wherein the leakage identification model includes a first processing branch line, a second processing branch line and a third processing branch line in parallel; The operating sound signal is input into the leakage identification model to obtain output results including: Inputting the working condition sound signal into a first processing branch to obtain a first result feature, wherein the first result feature is related to a feature of the working condition sound signal in the time domain; Inputting the working condition sound signal into a second processing branch to obtain a second result feature, where the second result feature is related to a feature of the working condition sound signal in the frequency domain; Inputting the working condition sound signal into a third processing branch to obtain a third result feature, wherein the third result feature is related to a modal feature of the working condition sound signal; The first result feature, the second result feature and the third result feature are spliced ​​to obtain the output result; wherein the third processing branch line includes a plurality of sub-branches, and the working condition sound signal is input into the third processing branch line to obtain the third result feature, including: Using an adaptive FIR filter in a third processing branch line to perform modal decomposition on the working condition sound signal to obtain a plurality of modal components; Inputting the plurality of modal components into a plurality of sub-branches in the third processing branch, wherein one modal component corresponds to one sub-branch; For each modal component in a sub-branch, a modal learning network corresponding to the modal component is determined from a modal learning network set, and the modal component on the sub-branch is processed using the corresponding modal learning network to obtain a modal transition feature corresponding to the modal component; wherein the modal learning network set includes a plurality of modal learning networks, parameters of the modal learning networks are related to the modal components, and one modal component corresponds to at least one modal learning network; Modal fusion is performed based on multiple modal transition features obtained from multiple sub-branches to obtain the third result feature.

2. The method for identifying working condition sound signals of a water supply pipeline according to claim 1, characterized in that: The first processing branch and the second processing branch each include at least one sub-branch.

3. The method for identifying working condition sound signals of a water supply pipeline according to claim 1, characterized in that: The third processing branch includes four sub-branches; the adaptive FIR filter in the third processing branch is used to perform modal decomposition on the operating sound signal to obtain four modal components; four corresponding modal learning networks are used to process the modal components in the four sub-branches to obtain four corresponding modal transition features; wherein the characteristics of the operating sound signal represented by the four modal components are different, and the time domain diagrams and spectrum diagrams corresponding to the four modal components are different.

4. The method for identifying working condition sound signals of a water supply pipeline according to claim 1, characterized in that: The third result feature is obtained by performing modal fusion based on multiple modal transition features obtained from multiple branches, including: Each sub-branch in the third processing branch includes a multi-layer perceptron, and the four modal transition features are used as four-way vector signals. For each way of vector signals, the signals are input into the multi-layer perceptron in the corresponding sub-branch to obtain an output vector. The four output vectors are added together to obtain a third result feature of the operating sound signal.

5. The method for identifying working condition sound signals of a water supply pipeline according to claim 1, characterized in that: The modality learning network is a one-dimensional network, and the modality learning network at least includes a convolution block group, a residual block group and an output layer group; wherein the convolution block group includes a front convolution block and a rear convolution block, and the residual block group includes a front residual block and a rear residual block; For each modal component in each sub-branch, the modal component is processed by a corresponding modal learning network to obtain a modal transition feature corresponding to the modal component, including: For the modal component in each sub-branch, the modal component is input into the front convolution block, the output of the front convolution block is used as the input of the front residual block, the output of the front residual block is used as the input of the rear convolution block, the output of the rear convolution block is used as the input of the rear residual block, the output of the rear residual block is used as the input of the output layer group, and the output layer group outputs the modal transition feature corresponding to the modal component.

6. The method for identifying working condition sound signals of a water supply pipeline according to claim 1, characterized in that: Inputting the working condition sound signal into the first processing branch to obtain the first result feature includes: obtaining a target signal segment corresponding to the working condition sound signal, and processing the working condition sound signal to obtain a target signal matching the target signal segment; using a time domain learning network to process the target signal to obtain a time domain feature of the working condition sound signal; wherein the time domain learning network includes a plurality of convolution blocks and a plurality of residual blocks, and the convolution kernel size of the convolution block is different from the convolution kernel size of the residual block; processing the working condition sound signal to obtain a target signal matching the target signal segment includes: using a filter to remove non-target signals in the working condition sound signal to obtain a target signal, and the passband cutoff frequency of the filter is less than the stopband cutoff frequency; Inputting the operating sound signal into the second processing branch to obtain the second result feature includes: performing a fast Fourier transform on the operating sound signal input into the second processing branch to obtain a frequency domain transformation signal; and using the frequency domain learning network of the second processing branch to process the frequency domain transformation signal to obtain the second result feature.

7. The method for identifying working condition sound signals of a water supply pipeline according to claim 1, characterized in that: The first result feature, the second result feature, and the third result feature are concatenated to obtain the output result, including: When the dimensions and lengths of the first result feature, the second result feature and the third result feature are consistent, a set function is used to perform vector concatenation on the first result feature, the second result feature and the third result feature along a set direction to obtain an output result.

8. The method for identifying working condition sound signals of a water supply pipeline according to claim 1, characterized in that: Acquire a first data set, the first data set comprising positive samples and negative samples, the positive samples being operating sound signals of a first preset number of water supply pipes when they are leaking, and the negative samples being operating sound signals of a second preset number of water supply pipes when they are not leaking; Pre-training the leakage identification model using the first data set; Acquire a second data set, the second data set includes a first factor and a second factor, the first factor is the ratio of positive samples predicted as positive samples to all predicted positive samples, and the second factor is the ratio of positive samples predicted as positive samples to actual positive samples; The leakage identification model is modified based on the first factor and the second factor.

9. A system for identifying working condition sound signals of a water supply pipeline, characterized in that: The identification system includes a signal acquisition device and an electronic device; The signal acquisition device is configured to collect the working condition sound signal of the inner wall of the water supply pipe; The electronic device is configured to obtain the operating sound signal collected by the signal acquisition device, and identify the operating sound signal according to the method for identifying the operating sound signal of the water supply pipeline according to any one of claims 1 to 8.

Images