An intelligent early warning method for preventing intrusion in urban tunnels based on distributed acoustic sensing

Abstract

The present invention discloses an intelligent early warning method for urban tunnel intrusion prevention based on distributed acoustic sensing, which relates to the fields of underground monitoring and rail safety early warning. Specifically, it includes: inputting data records of distributed optical fibers; windowing the records in standard sizes, calculating spectral amplitudes, intercepting ranges, truncating data, and normalizing; training a convolutional autoencoder to reconstruct data, and using the encoder part to encode and reduce the dimension of the data; training a random forest classifier using the feature vectors after preprocessing and dimension reduction of real-time data; preprocessing continuous data, reducing the dimension of the data with a convolutional encoder, and then detecting and classifying it with a random forest; performing post-processing on the detection results based on the screening of signal spatio-temporal continuity to obtain accurate detection and classification results. The present invention uses a convolutional autoencoder and a random forest algorithm to monitor construction signals in distributed acoustic sensing data in real time, detect and accurately classify them with high spatio-temporal resolution, and achieve intelligent early warning for rail construction based on distributed optical fibers.

Description

Technical Field

[0001] The present invention relates to the field of underground monitoring and track safety warning, and in particular to an intelligent warning method for preventing intrusion into urban tunnels based on distributed acoustic sensing. Background Art

[0002] Intrusion refers to the situation where an object exceeds the allowable contour size line within the boundary of the safe operation of the track traffic. Intrusion accidents will affect the operation of trains and threaten the safety of passengers. In severe cases, it may cause major safety accidents endangering people's lives and property. Construction machinery operating above the subway tunnel is the main potential threat of intrusion into the tunnel. Therefore, timely detection and understanding of the construction dynamics above the tunnel will be an important measure to protect track safety.

[0003] Traditional track inspection is carried out manually. High-frequency manual inspection can effectively prevent foreign object intrusion caused by construction, but the labor cost is too high. Therefore, there is an urgent need to develop a new technology that can overcome the limitations of manual inspection and achieve real-time tunnel construction monitoring with low labor and material costs. Summary of the Invention

[0004] To solve the above problems, the present invention provides an intelligent warning method for preventing intrusion into urban tunnels based on distributed acoustic sensing, which uses a convolutional autoencoder and a random forest algorithm to real-time monitor the construction signals in the distributed acoustic sensing data, and can detect and accurately classify them with high spatio-temporal resolution, realizing intelligent warning of track construction based on distributed optical fiber.

[0005] To achieve the above object, the present invention provides an intelligent warning method for preventing intrusion into urban tunnels based on distributed acoustic sensing, specifically including the following steps:

[0006] Step S1: Input the vibration data recorded by the distributed optical fiber acoustic wave sensor, and the vibration data includes the waveforms detected by all channels;

[0007] Step S2: Implement sliding windowing and data preprocessing on the continuous distributed acoustic sensing monitoring real-time data in a standardized process;

[0008] Step S3: Use the continuous data to train the convolutional autoencoder to reconstruct the spectrum;

[0009] Step S4: Use the encoder to preprocess and compress and reduce the dimension of the real-time data, and train the random forest classifier;

[0010] Step S5: Preprocess the real-time data stream and use the encoder to reduce the dimension, and then use the random forest for classification to obtain the detection result.

[0011] Preferably, determine the standard window size according to the propagation distance, duration, and detection accuracy required for the engineering machinery signal to be detected.

[0012] Preferably, in step S2, the data preprocessing includes performing a fast Fourier transform on the time axis of the single window data to the frequency domain, intercepting the required frequency range, taking the amplitude, and truncating and normalizing the data with percentiles.

[0013] Preferably, in step S3, the convolutional autoencoder includes an encoder part and a decoder part. The encoder part and the decoder part are composed of multiple convolutional layers with the same total number of layers and symmetry. The residual is set as the mean square error loss between the input and the output, and the network is trained with a dynamic learning rate.

[0014] Preferably, in step S4, the manually picked noise, subway, and construction signals with sliding windows and the preprocessed real-time data are input into the encoder part, and an integrated detection and classification random forest classifier is trained using the encoded feature vectors.

[0015] Preferably, in step S5, the convolutional autoencoder and the random forest classifier are connected. The continuous data in the real-time data stream with sliding windows and the two-dimensional batch data obtained by preprocessing are flattened and then input into the network to obtain the prediction result, and the prediction result is projected back to the two-dimensional spatio-temporal domain.

[0016] Preferably, it includes a distributed acoustic sensing device accessing the track optical fiber and a supporting mobile workstation for data storage and data processing.

[0017] Preferably, according to the duration and propagation distance of the detected construction signal, a continuity constraint parameter is set, and the detection results with spatio-temporal continuity weaker than the threshold are identified as discrete false detections and removed.

[0018] Therefore, the present invention adopts the above-mentioned intelligent early warning method for preventing intrusion in urban tunnels based on distributed acoustic sensing, and has the following beneficial effects:

[0019] (1) In the present invention, the distributed acoustic sensing technology uses an optical fiber as a sensor and measures the axial strain rate using the phase of the backscattered Rayleigh light. The distributed acoustic sensing technology can turn an optical cable dozens of kilometers long into dense sensors arranged at meter intervals and can sample at a sampling rate far exceeding that of traditional seismological instruments.

[0020] (2) The present invention uses spatio-temporal continuity constraints to screen out misidentifications, improving the accuracy; and outputs the category, location, and time of the detected construction signal.

[0021] (3) The present invention uses a convolutional autoencoder and a random forest algorithm to monitor the construction signals in the distributed acoustic sensing data in real time, can detect and accurately classify them with high spatio-temporal resolution, and realizes intelligent early warning for track construction based on distributed optical fiber.

[0022] (4) In the present invention, the random forest is an ensemble learning method for classification and regression, which is a voter composed of multiple decision trees. After reducing the dimensionality of the two-dimensional frequency-domain distributed acoustic sensing data through a convolutional autoencoder, a relatively stable classification result is obtained using the random forest algorithm, thereby realizing the detection and classification of construction signals, and having high computational efficiency, enabling real-time monitoring.

[0023] (5) In the present invention, the convolutional autoencoder is a symmetric neural network composed of multiple convolutional layers, consisting of an encoder and a decoder, mainly used in fields such as data denoising and data dimensionality reduction. Performing a fast Fourier transform on the distributed acoustic sensing data to the frequency domain can effectively erase time information and solve the translational invariance problem of convolutional neural networks.

[0024] The technical solution of the present invention will be further described in detail below through the accompanying drawings and embodiments. Description of the Drawings

[0025] Figure 1 It is the method flow chart of an intelligent early warning method for urban tunnel anti-intrusion based on distributed acoustic sensing according to the present invention;

[0026] Figure 2 It is the waveform domain and frequency domain characteristics of the day and night noise signals recorded by distributed optical fibers in the present invention;

[0027] Figure 3 It is the waveform domain and frequency domain characteristics of the day and night subway signals recorded by distributed optical fibers in the present invention;

[0028] Figure 4 It is the waveform domain and frequency domain characteristics of various construction signals recorded by distributed optical fibers in the present invention;

[0029] Figure 5 It is the structure and flow chart of the convolutional autoencoder combined with the random forest in the present invention;

[0030] Figure 6 It is the reconstruction effect diagram of the spectrum by the convolutional autoencoder in the present invention;

[0031] Figure 7 It is the effect diagram of constraining the detection result using continuity in the present invention. Specific Embodiments

[0032] The technical solution of the present invention will be further described below through the accompanying drawings and embodiments.

[0033] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those of ordinary skill in the field to which the present invention belongs.

[0034] In the present invention, words such as "including" or "comprising" mean that the elements before this word cover the elements listed after this word, and do not exclude the possibility of also covering other elements. The orientation or positional relationship indicated by terms such as "inside", "outside", "above", "below", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation to the present invention. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In the present invention, unless otherwise clearly specified and limited, terms such as "attachment" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be directly connected, or indirectly connected through an intermediate medium. It can be the communication inside two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0035] As Figure 1 shown, an intelligent early warning method for preventing intrusion in urban tunnels based on distributed acoustic sensing includes a distributed acoustic sensing device for accessing track optical fibers and a mobile workstation for data storage and data processing.

[0036] Use the trained convolutional encoder network and random forest network to achieve data dimensionality reduction and feature vector detection and classification.

[0037] Use spatio-temporal continuity constraints to screen out misidentifications and improve the accuracy rate; output the types, locations, and times of the detected construction signals.

[0038] Determine the standard cut window size according to the propagation distance, duration, and detection accuracy required for the engineering equipment signals to be detected.

[0039] Set the continuity constraint parameters according to the duration and propagation distance of the detected construction signals, and identify the detection results with spatio-temporal continuity weaker than the threshold as discrete false detections and remove them to reduce the system false alarm rate.

[0040] Specifically, it includes the following steps:

[0041] Step S1: Input the vibration data recorded by the distributed fiber optic acoustic wave sensing. The vibration data includes the waveforms detected by all channels.

[0042] Step S2: Implement sliding window cutting and data preprocessing on the continuous real-time data of distributed acoustic sensing monitoring in a standardized process.

[0043] In step S2, data preprocessing includes performing a fast Fourier transform on the single intercepted window data to the frequency domain along the time axis, intercepting the required frequency range, taking the amplitude, and truncating and normalizing the data using percentiles.

[0044] Step S3: Use continuous data to train a convolutional autoencoder to reconstruct the spectrum;

[0045] In step S3, the convolutional autoencoder includes an encoder part and a decoder part. The encoder part and the decoder part are composed of multiple convolutional layers with the same total number of layers and symmetry. The residual is set as the mean squared error loss between the input and the output, and the network is trained with a dynamic learning rate.

[0046] Step S4: Use the encoder to preprocess and compress and reduce the dimension of the real-time data, and train a random forest classifier;

[0047] In step S4, the manually picked noise, subway, and construction signals with sliding windows and the preprocessed real-time data are input into the encoder part, and an integrated detection and classification random forest classifier is trained using the encoded feature vectors.

[0048] Step S5: Preprocess the real-time data stream and use the encoder to reduce the dimension, and then use a random forest for classification to obtain the detection result. In step S5, the convolutional autoencoder and the random forest classifier are connected. The continuous data in the real-time data stream with sliding windows and the two-dimensional batch data obtained by preprocessing are flattened and input into the network to obtain the prediction result, and the prediction result is projected back to the two-dimensional spatio-temporal domain.

[0049] Embodiment

[0050] The method for detecting construction signals around a tunnel is applied, and specifically includes the following steps:

[0051] Slide and intercept the seismic waveforms recorded by the distributed optical fiber, perform a fast Fourier transform to the frequency domain, and intercept the response frequency band where the construction signals are located;

[0052] Use the spectra of a large amount of continuous data to train a convolutional autoencoder to obtain a reliable encoder as a data feature extraction tool;

[0053] Use the encoder to reduce the dimension of the spectrogram to be detected to obtain the compressed feature vector;

[0054] Manually identify construction events to make labels, perform data augmentation and then divide them into a training set and a validation set, using 0~n to represent noise, subway signals, and different construction categories;

[0055] Compress the spectra of the labeled events into feature vectors and train a random forest model;

[0056] Connect the continuous data to the convolutional encoder and the random forest network to obtain the prediction result.

[0057] According to this embodiment, a tunnel perimeter construction signal detection system based on distributed fiber optic acoustic sensing frequency domain coding dimensionality reduction and random forest algorithm is proposed, including: performing windowing and data preprocessing on continuous distributed acoustic sensing monitoring data according to a standardized process; connecting the trained convolutional encoder and random forest network, flattening the two-dimensional batch data obtained by sliding window and inputting it into the network to obtain a prediction result, and projecting the prediction result back to the two-dimensional spatio-temporal domain; using the spatio-temporal continuity of the construction signal to constrain the prediction result, removing discrete false detections, and outputting the detected construction signal category, location and time.

[0058] Distributed acoustic sensing technology uses optical fiber as a sensor and measures the axial strain rate using the phase of the backscattered Rayleigh light. Distributed acoustic sensing technology can turn a fiber optic cable dozens of kilometers long into a dense array of sensors spaced at meter intervals and can sample at a sampling rate far exceeding that of traditional seismological instruments. In an operating subway tunnel, communication optical cables and reserved optical cables are often laid, usually hanging on the iron racks on the tunnel wall along the tunnel. These optical cables can all be connected to a distributed acoustic sensing interrogator at one end, thus becoming dense vibration sensors.

[0059] Field experiments in the Hefei Subway show that even in a relatively poor optical cable - tunnel wall coupling mode such as the wall-mounted type, distributed acoustic sensing can still demodulate high-fidelity tunnel wall vibration signals. As Figure 2 shown, the background noise signals recorded by the distributed optical fiber in the experiment, the upper one is the waveform diagram, the lower one is the spectrogram, and the left and right are the noise characteristics during the day and at night respectively. The background noise level during the day is significantly higher than that at night, and the "common mode noise" generated by the vibration of the device interrogator itself is significantly visible in the night noise. As Figure 3 shown, the subway operation signals recorded by the distributed optical fiber in the experiment, the left and right are the signals generated by the subway trains running in the up and down directions respectively. Since the optical cable is hung on the subway tunnel wall in the down direction, the wind pressure signal generated by the subway operation can be recorded. As Figure 4 shown, the four construction machinery signals recorded by the distributed optical fiber in the experiment, the upper left, upper right, lower left, and lower right are from an excavator, a roller, a crusher, and a drill respectively. It can be seen that the distributed acoustic sensing record can accurately reflect the characteristics such as the frequency range and spatial distribution of various construction signals.

[0060] In the data preprocessing stage, the continuous data of the distributed optical fiber is downsampled to 200 Hz in time and multi-channel stacked to a 10-meter trace interval in space. The data is windowed in 10 traces and 20 seconds, and data augmentation is achieved by adjusting the step size. The fast Fourier transform is performed on the intercepted continuous data on the time axis to obtain the spectrum, and only the amplitude part in the range of 0 - 80 Hz is retained, resulting in two-dimensional data of 10×1600. The data is truncated and normalized with the fifth percentile of the spectrum as the lower bound and the ninety-fifth percentile as the upper bound.

[0061] As Figure 5 shown, the upper part shows the structure of the convolutional autoencoder neural network, which consists of 7 convolutional layers and 7 transposed convolutional layers symmetric to the former. The first half is the encoder, and the second half is the decoder. The encoder compresses and encodes the input data of 1×10×1600 to obtain a feature vector with a size of 8×2×13 = 208, while the decoder restores the feature vector to the original-sized data. The training objective is that the output of the decoder has as small a residual as possible compared to the input of the encoder. At this time, the encoded feature vector can represent the main features of the input data. As Figure 6 shown, the reconstruction effect of the convolutional autoencoder on the intercepted spectrum. The good consistency between the input and output demonstrates the reliability of the encoding and compression of the autoencoder.

[0062] All preprocessed manually annotated data is input into the autoencoder to obtain their corresponding feature vectors. The feature vectors and corresponding labels are used to train Figure 5 the random forest model shown in the lower half of , obtaining a feature vector classifier to determine whether the feature vector belongs to noise or a certain type of construction signal. The subway signal can be output as a separate class or merged with noise as one class. Connecting the preprocessing program, convolutional encoder, and random forest classifier together results in a detector that can directly detect and classify construction signals by sliding a window on the original two-dimensional data. The detection precision rate and recall rate of the first model on the test set are shown in Table 1. The test results show that at least 89.23% of the construction alarms made by this artificial intelligence model are real alarms, and 70.62% of the events detected manually are also detected by the model. The model sets different weight balancing strategies considering the different abilities of four types of construction to damage the ground. For example, the signal of the excavator is not significant and its invasiveness is not strong, so the model tries to improve its precision rate and reduce the requirement for its recall rate; the drilling machine has the greatest invasion risk and requires ensuring as high a recall rate as possible. The precision rates of the four types of construction equipment, namely excavator, roller, crusher, and drilling machine, are 76.65%, 92.71%, 87.39%, and 76.63% respectively, and the recall rates are 25.55%, 73.85%, 71.39%, and 81.84% respectively. This shows that the model can effectively monitor and warn of ground construction conditions.

[0063] Table 1

[0064] ;

[0065] In practical applications, the early warning system requires a lower false alarm rate. Constraints on spatio-temporal continuity can be added in the post-processing stage of the model to reduce false alarms. Considering the duration and propagation distance of construction signals, the continuity parameter requires that the results of sliding window detection be spatio-temporally continuous; otherwise, it is judged as a false detection. As Figure 7 shown, the effect of continuity constraints. For a certain segment of distributed acoustic sensing recorded data, this algorithm is used for detection after sliding window. The horizontal axis corresponds to the distance, and the vertical axis corresponds to the time. The colored blocks of different colors represent various detected construction signals. The upper figure is the original output, and the lower figure is the output after screening using spatio-temporal continuity constraints. Table 2 shows the test results for a single event after adding continuity constraints. At this time, the precision rate of the model for construction signals is increased to 99.34%, significantly reducing the false detection rate and improving the practicability. After testing, this algorithm can access the distributed acoustic sensing data stream for real-time detection.

[0066] Table 2

[0067] ;

[0068] Therefore, the present invention adopts the above-mentioned intelligent early warning method for preventing intrusion in urban tunnels based on distributed acoustic sensing, uses a convolutional autoencoder and a random forest algorithm to monitor construction signals in distributed acoustic sensing data in real time, performs detection and accurate classification with high spatio-temporal resolution, and realizes intelligent early warning for track construction based on distributed optical fibers.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements do not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An urban tunnel anti-intrusion intelligent early warning method based on distributed acoustic sensing, characterized in that: The specific steps include: Step S1: inputting vibration data recorded by distributed optical fiber acoustic wave sensing, the vibration data including waveforms detected by all channels; Step S2: Sliding windowing and data preprocessing are performed on the continuous distributed acoustic sensor monitoring real-time data in a standardized process; in step S2, the data preprocessing includes performing a time axis fast Fourier transform on the single windowed data to the frequency domain, intercepting the required frequency domain range, taking the amplitude, and truncating and normalizing the data by percentile; Step S3: Reconstruct the spectrum by training the convolutional autoencoder using continuous data; Step S4: using the encoder to preprocess and compress the real-time data and reduce its dimension, and train the random forest classifier; Step S5: preprocess the real-time data stream and use the encoder to reduce the dimension and then use the random forest to classify to obtain the detection result; The method for detecting construction signals around a tunnel specifically includes the following steps: The seismic waveform recorded by the distributed optical fiber is windowed by sliding, and fast Fourier transformed to the frequency domain to intercept the response frequency band where the construction signal is located; In the data preprocessing stage, the continuous data of the distributed optical fiber is downsampled to 200 Hz in time, and multiple channels are superimposed in space to a channel spacing of 10 meters. The data is windowed with 10 channels and 20 seconds, and data augmentation is achieved by adjusting the step size. The continuous data is fast Fourier transformed on the time axis to obtain the spectrum, and only the The amplitude of the Hertz part was obtained to obtain 10 × 1600 two-dimensional data, which were truncated and normalized with the fifth percentile of the spectrum as the lower bound and the ninety-fifth percentile as the upper bound; Use the spectrum of a large amount of continuous data to train the convolutional autoencoder to obtain a reliable encoder as a data feature extraction tool; Use an encoder to reduce the dimension of the spectrum graph to be detected to obtain a compressed feature vector; The convolutional autoencoder consists of 7 convolutional layers and 7 transposed convolutional layers symmetrical to the former. The first half is the encoder and the second half is the decoder. The encoder compresses and encodes the input data of 1×10×1600 to obtain a feature vector of size 8×2×13=208, and the decoder restores the feature vector to the original size of data. Manually identify construction events and create labels, and then divide the data into training and validation sets after data augmentation. representing noise, subway signals, and different construction categories; Compress the labeled event spectrum into feature vectors and train the random forest model; Connect the continuous data to the convolution encoder and random forest network to get the prediction results; By connecting the preprocessor, convolutional encoder and random forest classifier, we get a detector that directly detects and classifies construction signals by sliding a window on the original two-dimensional data.

2. According to the method of claim 1, the method is characterized by: The standard window size is determined based on the propagation distance, duration and detection accuracy required of the engineering equipment signal to be detected.

3. According to the method of claim 2, the method is characterized by: In step S3, the convolutional autoencoder includes an encoder part and a decoder part. The encoder part and the decoder part are composed of multiple convolutional layers with the same total number of layers and symmetry. The residual is set to the mean square error loss between the input and output, and the network is trained with a dynamic learning rate.

4. According to the method of claim 3, the method is characterized by: In step S4, the manually picked noise, subway and construction signal sliding windows and preprocessed real-time data are input into the encoder part, and the encoded feature vector is used to train the integrated detection and classification random forest classifier.

5. According to the method of claim 4, the method is characterized by: In step S5, the convolutional autoencoder and the random forest classifier are connected, the continuous data sliding window in the real-time data stream and the two-dimensional batch data obtained by preprocessing are flattened and input into the network to obtain the prediction result, and the prediction result is projected back to the two-dimensional spatiotemporal domain.

6. An urban tunnel anti-intrusion intelligent early warning method based on distributed acoustic sensing according to any one of claims 1 to 5, characterized in that: It includes distributed acoustic sensing equipment connected to the track optical fiber and a matching mobile workstation for data storage and data processing.

7. The method for preventing intrusion in urban tunnels based on distributed acoustic sensing according to claim 6 is characterized in that: The continuity constraint parameters are set according to the duration and propagation distance of the detected construction signal, and the detection results with spatiotemporal continuity weaker than the threshold are identified as discrete false detections and removed.

Images