External breaking vibration detection method and system based on wavelet transform and time convolution network

Abstract

The invention discloses an external breaking vibration detection method and system based on wavelet transform and a time convolution network. The method comprises the following steps: S1, collecting three-component (X / Y / Z) vibration data of underground pipeline monitoring points; s2, selecting a db4 wavelet basis function to carry out wavelet transformation on each data segment, and extracting X, Y and Z three-channel multi-scale time-frequency features; s3, configuring a multi-channel time convolution network model; s4, collecting set type mechanical vibration data, constructing a data set, and training a multi-channel time convolution network model to obtain a model file; and S5, inputting the three-channel multi-scale time-frequency features into the multi-channel time convolutional network model loaded with the model file to obtain an identified vibration type result and confidence, and taking the vibration type corresponding to the confidence greater than a preset threshold as an external breaking vibration detection result. According to the method, multiple external breaking types can be accurately identified, and interpretability analysis based on physical characteristics such as dominant frequency, energy distribution and periodicity can be provided.

Description

Technical Field

[0001] This invention relates to a method and system for detecting external vibration damage to urban underground infrastructure, meeting the vibration monitoring needs of the surrounding environment of key facilities such as urban subway tunnels, underground pipelines (e.g., gas, water supply, and communication), long-distance oil and gas pipelines, and power cable channels. Background Technology

[0002] In the process of urbanization, underground space development is becoming increasingly frequent. The mechanical vibrations generated by various construction machinery (such as excavators, pile drivers, road rollers, and heavy vehicles) during construction are a major external source of damage, leading to underground structural damage, pipeline rupture, loosening of joints, and even secondary disasters (such as gas leaks and ground subsidence). Therefore, constructing a high-precision, low-false-alarm, and highly interpretable intelligent identification system for external vibration is of great significance for achieving closed-loop management of "proactive early warning—precise positioning—type identification—emergency response".

[0003] With the advancement of smart city and urban lifeline engineering safety monitoring systems, vibration monitoring technology based on sensors such as distributed optical fibers, accelerometers, and seismic detectors is widely used in the protection of underground infrastructure. However, existing systems still face the following core challenges in practical applications:

[0004] The signals are highly complex: multiple vibration sources coexist in urban environments, including traffic vibrations, industrial equipment operation, natural earthquakes, and various construction machinery operations. These signals exhibit non-stationary and nonlinear characteristics in the time domain, and their frequency bands overlap significantly, making it difficult for traditional threshold alarms or spectrum analysis methods to effectively distinguish between different types of external sources of vibration.

[0005] Insufficient classification accuracy: While existing deep learning methods such as 1D-CNN and LSTM can process time series data, their ability to capture multi-scale frequency features is limited. For example, excavators exhibit periodic impact characteristics (dominant frequency 5–20Hz), pile drivers produce transient strong impacts (10–40Hz), heavy vehicles exhibit continuous rolling vibrations (2–15Hz), and road rollers exhibit regular low-frequency vibrations (10–30Hz), while natural earthquakes are characterized by a combination of wide-bandwidth, long-duration body waves and surface waves. If only the original time-domain input is used, the model cannot fully extract these physical differences.

[0006] The methods are highly dependent on the sensor network topology: some methods require densely deployed sensor arrays for direction-of-arrival (DOA) estimation or wave velocity inversion, which is costly and complex to deploy. On the other hand, sparse deployments result in insufficient spatial information, affecting localization and classification performance.

[0007] Therefore, there is an urgent need for a new integrated method that can accurately identify multiple types of external damage and provide decision-making basis based on physical feature matching, thereby improving the robustness and reliability of the system. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for detecting external vibration based on wavelet transform and time convolutional network, so as to achieve high-precision classification of five typical vibration events: excavators, pile drivers, road rollers, heavy vehicles, and background noise.

[0009] To this end, the present invention provides a method for detecting external vibration based on wavelet transform and temporal convolutional network, comprising the following steps: S1, collecting three-component (X / Y / Z) vibration data from monitoring points of underground pipelines; S2, splitting the three-component vibration data into T data segments, selecting the db4 wavelet basis function to perform wavelet transform on each data segment to obtain four layers of high-frequency details and one layer of low-frequency trend, thereby extracting multi-scale time-frequency features of the X, Y, and Z channels; S3, configuring a multi-channel temporal convolutional network model, wherein the input tensor dimension of the model is [B,C,T], where B is the batch size, C is the multi-scale time-frequency features of the X, Y, and Z channels, and the model output is [P(1), P(2),---,P(CL)], where P(1) to P(CL) are the confidence levels corresponding to the set vibration type; S4. Collect mechanical vibration data of the specified types, construct a dataset, and obtain the X, Y, and Z three-channel multi-scale time-frequency features after processing in step S2. Train a multi-channel temporal convolutional network based on these features to obtain a model file. S5. Input the X, Y, and Z three-channel multi-scale time-frequency features obtained in step S2 into the multi-channel temporal convolutional network model with the model file loaded, and obtain the identified vibration type results and confidence scores. Vibration types with confidence scores greater than a preset threshold are taken as the external failure vibration detection results.

[0010] According to another aspect of the present invention, an external vibration detection system based on wavelet transform and time convolutional network is provided, including a cloud management platform and a distributed terminal, wherein the distributed terminal stores a computer program one, which, when executed, implements step S1 of the external vibration detection method described above, and the cloud management platform stores a computer program two, which, when executed, implements steps S2-S5 of the external vibration detection method described above.

[0011] The present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the above-described external vibration detection method.

[0012] This method extracts multi-scale time-frequency features from three-component seismic signals using wavelet transform to enhance the separability of different types of vibrations. It also constructs a deep temporal convolutional network (TCN) model to efficiently learn the temporal dynamic evolution of wavelet domain features, ultimately achieving high-precision classification of five typical vibration events: excavators, pile drivers, road rollers, heavy vehicles, and background noise. This method is wavefield data-driven and can provide interpretable analysis based on physical characteristics such as dominant frequency, energy distribution, and periodicity.

[0013] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0015] Figure 1 This is a flowchart of the external vibration detection method based on wavelet transform and time convolutional network of the present invention;

[0016] Figure 2 This is a structural block diagram of an external vibration detection system according to an embodiment of the present invention;

[0017] Figure 3 This is a structural block diagram of a cloud management platform according to an embodiment of the present invention;

[0018] Figure 4 This is a structural block diagram of a distributed terminal according to an embodiment of the present invention;

[0019] Figure 5 The frequency signals are obtained by decomposing the Y-axis vibration signal caused by a high-speed pile driver into db4 wavelets.

[0020] Figure 6 The backbone network structure of the multi-channel temporal convolutional network model used in this invention is shown. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] like Figure 1 As shown, the external vibration detection method based on wavelet transform and time convolutional network of the present invention includes the following steps S1 to S5.

[0023] S1. Collect three-component (X / Y / Z) vibration data from each monitoring point of the underground pipeline.

[0024] First, establish an external vibration detection system, such as... Figure 2 As shown, the system consists of a cloud management platform and a distributed terminal network. Figure 3 As shown, the cloud management platform integrates modules such as a data receiving unit, a data decryption unit, an artificial intelligence computing unit, and an application unit. Figure 4 As shown, the distributed terminal integrates a three-component vibration sensor acquisition unit, a temperature and humidity sensor acquisition unit, a BeiDou positioning and timing unit, a data encryption unit, a data transmission unit, and a communication unit.

[0025] Distributed terminals are deployed at 100-150 meter intervals along the underground pipeline to form a sensor network. Using wired or wireless communication units, they periodically transmit collected data blocks to a cloud management platform. Each data block contains three-component (X / Y / Z) vibration data with a sampling duration of 60 seconds, along with information such as time, device number, longitude, latitude, altitude, temperature, and humidity. The data blocks are compressed and encrypted using the LZ4 compression algorithm before being sent to the cloud management platform.

[0026] S2. Vibration feature extraction based on wavelet decomposition

[0027] The cloud management platform decrypts and decompresses each received terminal data to obtain a data block. The vibration data within the data block is then divided into T data segments using a sliding window method, with a sliding step size of [missing value]. Window size is This invention takes =1 second, =2 seconds, then T is 59. Let the three-component waveform segment of each data segment T be... Assuming sampling frequency =500Hz, which is the number of data points collected per second. One window is collected at a time. If the data is taken as a data segment, then the number of data points in each data segment is N=1000.

[0028] Considering that the energy of excavators (vibration frequency 5–20Hz), pile drivers (vibration frequency 10–40Hz), road rollers (vibration frequency 10–30Hz), and heavy vehicles (vibration frequency 2–15Hz) is mainly concentrated in the low-frequency band, this invention selects the db4 wavelet basis function to perform wavelet transform on each data segment in order to capture such non-stationary transient signals. db4 has fourth-order vanishing moments, orthogonality, and good time-frequency localization capabilities, which can effectively suppress polynomial trend terms and retain impact characteristics. A four-level decomposition is adopted, as shown in Table 1, which can control computational complexity while retaining key features and avoid information redundancy caused by excessive decomposition. Figure 5 The following is a set of frequency signals obtained by decomposing the Y-axis vibration signal caused by a high-speed pile driver using db4 wavelet decomposition. The vertical axis in the figure represents the vibration amplitude, which is dimensionless.

[0029] Table 1. Frequency range covered by db4 wavelet decomposition

[0030]

[0031] The wavelet basis decomposition of 4 layers actually yields 5 layers, including 4 layers of high-frequency details. , , , ) and Layer 1 low-frequency trend (A4).

[0032] For each layer ( , , , , Calculate the statistics for the wavelet coefficients of ) respectively:

[0033] Mean:

[0034] Standard deviation:

[0035] energy:

[0036] kurtosis:

[0037] in, For sample size, For the first One observation value, Sample mean, s is the sample standard deviation.

[0038] Based on this, two more global features are calculated:

[0039] Energy proportion of each layer :

[0040]

[0041] It reflects the ratio of the j-th energy to the total energy.

[0042] Wavelet entropy H characterizes the uncertainty of signal energy distribution. For example, the vibration energy caused by pile drivers and road rollers is usually concentrated, while random noise energy is uniformly distributed. Its formula is defined as follows:

[0043]

[0044] Therefore, each channel generates 4 (mean, standard deviation, energy, kurtosis) * 5 layers + 5 (energy percentage per layer) + 1 (wavelet entropy) = 26 dimensions. The X, Y, and Z channels together generate 78-dimensional feature vectors.

[0045] S3, Multichannel Temporal Convolutional Network Model

[0046] Temporal convolutional networks (TCNs) outperform RNNs / LSTMs in long sequence modeling thanks to their causal convolution, dilated convolution, and residual connection mechanisms, and support parallel computation, making them suitable for real-time inference.

[0047] The model input tensor has dimensions [B,C,T]=[B,78,59], where B is the batch size, C is a 78-dimensional feature vector, and T is 59. The model output is [P(1), P(2),---,P(CL)]. P(CL) is the confidence level corresponding to the vibration type CL.

[0048] The TCN backbone network structure is as follows: Figure 1 As shown. RC is a dilated convolution module, and the dilation factor is... The receptive field grows exponentially; when d=4, the receptive field is 3*(1+2+4)=21 steps. Maxpool is the maximum pooling operation; FC is a fully connected layer; ReLU is a non-linear activation function; Dropout is a random dropout layer used to suppress overfitting; FC(CL) is a fully connected layer with output CL; Softmax is a normalized exponential function layer, where CL is the total number of vibration types. Dilated convolution is defined as:

[0049]

[0050] Where d is the dilation factor; k is the kernel size; Use the ReLU activation function; T=59 (according to sampling step size) =1 second, sampling window =2 seconds, T=59 corresponds to 60 seconds of data). Indicates the index of the current time step; Indicates the inner inductance of the convolution kernel; represents the kernel weights; b represents the bias term.

[0051] Loss function: Weighted cross-entropy loss is used to mitigate class imbalance.

[0052]

[0053] Where N represents the total number of samples in the batch, i.e., the number of samples processed in one forward propagation (e.g., N=32); j represents the category index. This invention mainly focuses on five types of vibration: excavators, pile drivers, road rollers, heavy vehicles, and background noise. This represents the weight coefficient of the j-th class; Let represent the true label. If the i-th sample belongs to the j-th class, then... It is 1 if it is true, otherwise it is 0. The label represents the model's prediction.

[0054] S4, Model Training

[0055] Vibration data from various types of machinery were collected, covering different distances and soil conditions as much as possible. A total of 3630 sets of three-component acceleration data were collected, each set lasting 60 seconds. These included 742 sets from excavators, 768 sets from pile drivers, 635 sets from road rollers, 616 sets from heavy vehicles, and 869 sets from background noise. The data were then split into training / validation and test sets at a 9:1 ratio.

[0056] After extracting vibration features from wavelet decomposition using the training and validation sets, a temporal convolutional network model is trained to obtain the final model file.

[0057] S5. In practical applications, the cloud-based artificial intelligence computing unit processes the received real-time data and runs a temporal convolutional network model loaded with the optimal parameter model file to obtain the identified vibration type and confidence level. When the confidence level is greater than a preset threshold, it is determined that the vibration of that type exists, and an early warning message is sent to the application platform.

[0058] Compared with the prior art, the method of the present invention has the following technical features / advantages:

[0059] (1) The first "wavelet feature + TCN" external damage detection architecture is created: the physical interpretability of wavelet transform is deeply integrated with the deep temporal modeling capability of TCN, which significantly improves the classification performance. Table 2 shows the comparative experiments based on the constructed dataset and related methods.

[0060] Table 2. Comparison of this method and related methods in experimental results

[0061]

[0062] In the table, 1D-CNN is a one-dimensional convolutional neural network, LSTM is a long short-term memory network, and SVM is a support vector machine.

[0063] (2) In view of the non-stationarity and environmental noise interference in underground vibration signals, this invention analyzes the collected vibration data of different types, and effectively separates the components of different frequency bands based on the db4 wavelet basis 4-layer decomposition, extracts feature components with discriminative characteristics, enhances the ability to capture micro-vibration signals, and provides high-quality input features for time convolutional network models.

[0064] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting external vibration damage based on wavelet transform and time convolutional network, characterized in that, Includes the following steps: S1. Collect three-component (X / Y / Z) vibration data from underground pipeline monitoring points; S2. Divide the three-component vibration data into T data segments, select the db4 wavelet basis function to perform wavelet transform on each data segment, and obtain four layers of high-frequency details and one layer of low-frequency trend, thereby extracting the multi-scale time-frequency features of the X, Y and Z channels. S3. Configure a multi-channel temporal convolutional network model, where the input tensor dimension is [B,C,T], where B is the batch size, C is the X, Y, Z three-channel multi-scale time-frequency features, and the model output is [P(1), P(2),---,P(CL)], where P(1) to P(CL) are the confidence levels corresponding to the set vibration types; S4. Collect mechanical vibration data of a specified type, construct a dataset, and obtain X, Y, and Z three-channel multi-scale time-frequency features after processing in step S2. Based on this, train a multi-channel temporal convolutional network to obtain the model file. S5. Input the X, Y, and Z three-channel multi-scale time-frequency features obtained in step S2 into the multi-channel temporal convolutional network model with the model file loaded, and obtain the identified vibration type results and confidence level. The vibration type corresponding to the confidence level greater than the preset threshold is taken as the external failure vibration detection result.

2. The external vibration detection method based on wavelet transform and time convolutional network according to claim 1, characterized in that, Extracting multi-scale time-frequency features of the X, Y, and Z channels includes: calculating the mean, standard deviation, energy, kurtosis, energy proportion, and wavelet entropy of each layer of wavelet coefficients obtained from wavelet decomposition, as well as the multi-scale time-frequency features of the X, Y, and Z channels.

3. The external vibration detection method based on wavelet transform and time convolutional network according to claim 1, characterized in that, Distributed terminals are deployed at the underground pipeline monitoring points. Each distributed terminal is equipped with a three-component acceleration sensor to collect three-component (X / Y / Z) vibration signals.

4. The external vibration detection method based on wavelet transform and time convolutional network according to claim 3, characterized in that, The distributed terminals are deployed at intervals of 100-150 meters along the underground pipeline to form a sensor network. They periodically transmit the collected data blocks to the cloud management platform. Each data block contains three-component (X / Y / Z) vibration data with a sampling duration of 60 seconds.

5. The external vibration detection method based on wavelet transform and time convolutional network according to claim 4, characterized in that, The data block is divided into T data segments using the sliding window method, with a sliding step size of . =1 second, window size is =2 seconds.

6. The external vibration detection method based on wavelet transform and time convolutional network according to claim 1, characterized in that, The vibration types are set as five categories: excavator, pile driver, road roller, heavy vehicle, and background noise.

7. The external vibration detection method based on wavelet transform and time convolutional network according to claim 1, characterized in that, The backbone of the multichannel temporal convolutional network model includes three dilated convolutional modules with dilation factors of d=1, d=2, and d=4, respectively.

8. The external vibration detection method based on wavelet transform and time convolutional network according to claim 1, characterized in that, The weighted cross-entropy loss function was chosen as the loss function for training the multi-channel temporal convolutional network model to alleviate class imbalance.

9. An external vibration detection system based on wavelet transform and time convolutional network, characterized in that, The method includes a cloud management platform and a distributed terminal, wherein the distributed terminal stores a computer program one, which, when executed, implements step S1 of the method according to claim 1; and the cloud management platform stores a computer program two, which, when executed, implements steps S2-S5 of the method according to claim 1.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images