Distributed fiber optic acoustic sensing event hybrid neural network recognition system and method

Abstract

The application belongs to the field of optical fiber sensing, and discloses a hybrid neural network identification system and method for distributed optical fiber acoustic wave sensing events. The system comprises a preliminary feature extraction module, a deep feature extraction module and an event identification module. The preliminary feature extraction module comprises a preliminary feature extraction model based on a quantum neural network, the deep feature extraction module comprises a deep feature extraction model based on a deep learning neural network, and the event identification module comprises an event identification model based on a quantum neural network. The main steps of the identification method include extracting preliminary features of intrusion events, extracting deep features and identifying events according to the deep features. The application combines traditional neural networks and quantum neural networks, can capture complex data patterns that are difficult for traditional neural networks to express, effectively improves the identification accuracy of intrusion events, has strong adaptability in a strong noise complex environment, has good intelligent level, and is worth wide popularization and application.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical fiber sensing, in particular to a hybrid neural network identification system and method for distributed optical fiber acoustic sensing events. BACKGROUND

[0002] Distributed acoustic sensing (DAS) system is a device that uses optical fiber as a sensing element for distributed multi-point simultaneous monitoring. DAS system uses Rayleigh scattering in optical fiber to accurately perceive and analyze acoustic signals. The basic principle is that when there is external vibration along the optical fiber, the phase of the light wave at the corresponding position will change accordingly. The phase change and external vibration are linearly related. Because different types of external vibration have different characteristics, the phase can be demodulated through data processing technology, so as to identify different types of external vibration.

[0003] In perimeter security and oil and gas pipeline transportation monitoring, DAS system has significant advantages such as long-distance transmission and anti-electromagnetic interference compared with traditional monitoring methods. When an intrusion event occurs, different types of acoustic signals are usually generated. By capturing and analyzing the characteristics of these acoustic signals, the type and location of the intrusion event can be identified. Applying DAS system to intrusion event monitoring can achieve distributed and all-weather monitoring and early warning, reducing the loss caused by intrusion events.

[0004] Although DAS system has great application potential in intrusion event monitoring, it still faces some challenges in accurately identifying intrusion events and improving identification efficiency: 1. Currently, optical fibers are mainly buried in underground environments. Due to the complexity of underground environments and the influence of external noise sources, it is easy to interfere with the sensing signal, leading to a decrease in signal quality and increasing the difficulty of data processing; 2. The signal of the intrusion event itself has non-stationary characteristics, with large frequency variation and complex characteristics, making it difficult to extract valuable features from the signal. Traditional time-frequency domain signal analysis methods are difficult to extract valuable features; 3. In practical applications, different types of intrusion events may produce similar acoustic signals, such as excavation construction and some large vehicles passing through, which may produce similar signals in the low-frequency part. The subtle differences between these similar signals make it challenging to accurately identify different event types; 4. DAS system has high requirements for real-time monitoring, and when monitoring over long distances, it will generate a large amount of data. Existing technologies still have certain bottlenecks in quickly and efficiently processing large-scale data, affecting the performance and application effect of the system. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a hybrid neural network identification system and method for distributed optical fiber acoustic sensing events, which can quickly and accurately identify intrusion events in strong background noise.

[0006] To solve the above technical problems, the application first proposes a hybrid neural network identification system for distributed optical fiber acoustic sensing events, comprising a preliminary feature extraction module, a deep feature extraction module and an event identification module;

[0007] The preliminary feature extraction module comprises a preliminary feature extraction model based on a quantum neural network, which is used to input the intrusion event data collected by the distributed optical fiber acoustic sensing system to extract the preliminary features of the intrusion event;

[0008] The deep feature extraction module comprises a deep feature extraction model based on a deep learning neural network, which is used to input the preliminary features to extract the deep features of the intrusion event;

[0009] The event identification module comprises an event identification model based on a quantum neural network, which is used to input the deep features of the intrusion event data to identify the corresponding intrusion event.

[0010] Further, the preliminary feature extraction model, the deep feature extraction model and the event identification model need to be trained before formal use, and the training method comprises:

[0011] A certain amount of training data is obtained as the input of the preliminary feature extraction model, so that the preliminary feature extraction model, the deep feature extraction model and the event identification model are trained according to the respective model characteristics until the models converge.

[0012] The training of the preliminary feature extraction model, the deep feature extraction model and the event identification model adopts cross-entropy loss and cosine annealing algorithm.

[0013] Optimally, the identification system further comprises a data preprocessing module for denoising preprocessing of the intrusion event data before the intrusion event data collected by the distributed optical fiber acoustic sensing system is input into the preliminary feature extraction module.

[0014] The application also provides a hybrid neural network identification method for distributed optical fiber acoustic sensing events, comprising the following steps:

[0015] The feature extraction module uses its preliminary feature extraction model based on a quantum neural network to extract the preliminary features of the intrusion event data collected by the distributed optical fiber acoustic sensing system;

[0016] The deep feature extraction module uses its deep feature extraction model based on a deep learning neural network to input the preliminary features to extract the deep features of the intrusion event data;

[0017] The event recognition model built on the quantum neural network in the event recognition module takes the deep features of the intrusion event data as input and identifies the corresponding intrusion event.

[0018] Preferably, the preliminary feature extraction includes the following steps in sequence:

[0019] First, m (m=3, 5, 7) qubits are used , and RY rotation quantum gate coding is performed on every m data points,

[0020] RZ rotation quantum gate operation is performed on the coded m qubits,

[0021] Controlled non-gate is applied to realize entanglement between qubits,

[0022] RZ rotation quantum gate operation is performed on the m qubits,

[0023] The m qubits are The Pauli expectation observation is taken as the quantum neural network convolution output value of m different channels, i.e., the preliminary feature of the intrusion event data.

[0024] Preferably, the event recognition model takes the deep features of the intrusion event data as input and identifies the corresponding intrusion event, including the following steps:

[0025] n qubits are used to correspond to n kinds of intrusion events, and the deep features are flattened,

[0026] The flattened deep features are output through a fully connected layer to obtain an n-dimensional vector,

[0027] RY rotation quantum gate angle coding is performed on the n qubits,

[0028] Controlled RZ rotation quantum gate operation is performed on the qubit pairs to realize mutual entanglement of the qubit pairs;

[0029] RY rotation quantum gate operation is performed on all qubits;

[0030] Controlled RZ rotation quantum gate operation is performed on the qubit pairs ,

[0031] The Pauli expectation observation value of each qubit is measured,

[0032] The n probability values are compared, and the event corresponding to the largest probability value is selected as the final recognition result of the system.

[0033] Further optimization is to perform denoising preprocessing on the intrusion event data before inputting it into the preliminary feature extraction module.​

[0034] Preferably, the denoising preprocessing comprises a method of spectral subtraction and wavelet packet denoising.

[0035] Compared with the prior art, the present application has the following beneficial effects:

[0036] (1) At present, most DAS intrusion event recognition methods mainly rely on traditional machine learning or deep learning technology. However, in a strong background noise environment, the recognition accuracy is low, and there are difficulties in recognizing events with small feature differences or similar events. The present application includes data preprocessing and noise reduction, hybrid quantum neural network feature extraction, and quantum neural network output. By introducing quantum neural network and using quantum bit entanglement and other quantum characteristics, the key features in the noisy signal can be more accurately captured, the sensitivity to different events is enhanced, and efficient and accurate classification can be made in a complex environment. The present application can be used in important scenes such as perimeter intrusion early warning and monitoring of oil and gas pipelines. By laying monitoring optical cables, real-time collection of vibration signals can accurately identify external intrusion events such as mechanical excavation and manual construction, and provide more reliable protection for public safety.

[0037] (2) The present application combines spectral subtraction and wavelet packet denoising technology to enhance the inhibition of noise. Spectral subtraction can suppress background noise and improve signal-to-noise ratio to a certain extent by processing signals in the frequency domain and estimating noise amplitude; wavelet packet denoising can further decompose signals in the time domain and remove unnecessary noise. When processing signals denoised by spectral subtraction and wavelet packet technology, the quantum neural network can better capture the key features in the signal through the quantum superposition state, reduce the influence of noise on the recognition result, and improve the recognition accuracy and stability in a strong background noise environment.

[0038] (3) The present application combines quantum neural network in feature extraction, and through quantum gate encoding and rotating quantum gate operation of quantum neural network, it can effectively extract the potential rules and complex features of data, process high-dimensional data features that traditional neural networks cannot fully exploit, avoid feature loss problem in high-dimensional data analysis, and has significant feature extraction advantage for non-stationary signals collected by DAS system.

[0039] (4) In the quantum neural network output module of the present application, n quantum bits are used to correspond to n kinds of intrusion events, and efficient entanglement between quantum bits is realized through RY rotating quantum gate and control RZ rotating quantum gate. The final model compares the probability values of the n events, selects the event corresponding to the maximum probability as the recognition result. The entanglement between quantum bits enhances the sensitivity of the system to subtle differences, so that events with similar signals can be distinguished, and the classification accuracy and robustness are improved.

[0040] (5) The present application optimizes the learning rate by introducing the cosine annealing algorithm, so that the training process is more efficient. The cosine annealing algorithm can dynamically adjust the learning rate according to different stages of the training process, avoid overfitting while accelerating the convergence of the model, so that the quantum neural network can be efficiently trained in a short time, and the response speed and computing efficiency of the system in practical application are improved. BRIEF DESCRIPTION OF DRAWINGS

[0041] The technical solutions of the present application will be further specifically described below in combination with the drawings and specific embodiments.

[0042] Figure 1 The hardware arrangement diagram of the distributed optical fiber acoustic sensing system involved in the present application.

[0043] Figure 2 The organization structure diagram of the main function module of the hybrid neural network recognition system of the distributed optical fiber acoustic sensing event of the present application.

[0044] Figure 3 The relationship structure diagram of the preliminary feature extraction model based on quantum neural network, the deep feature extraction model based on deep learning neural network and the event recognition model of quantum neural network of the present application.

[0045] Figure 4 The overall flowchart of the hybrid neural network recognition method of the distributed optical fiber acoustic sensing event provided by the present application.

[0046] Figure 5 The mechanism schematic diagram of the preliminary feature extraction model of intrusion event based on quantum neural network provided by the present application.

[0047] Figure 6 The mechanism schematic diagram of the intrusion event recognition model based on quantum neural network provided by the present application. DETAILED DESCRIPTION

[0048] Generally, the hardware arrangement of the distributed acoustic sensing (DAS) system involved in the present application is shown in Figure 1 The distributed acoustic sensing (DAS) system therein, at different times, different places of the test optical fiber, includes: a narrow line width laser 1, a 10:90 optical fiber coupler 2, an acousto-optic modulator 3, an erbium-doped fiber amplifier 4, an optical fiber Bragg grating 5, a circulator 6, an erbium-doped fiber amplifier 7, an optical fiber Bragg grating 8, a 50:50 optical fiber coupler 9, a balanced photodetector 10, and a data acquisition card 11. Through the DAS system, different types of mechanical construction, human operation and other behaviors that may cause interference or threat to the monitoring area in perimeter security, oil and gas pipeline monitoring and other scenes are collected, and the collected data is divided into a training set and a test set.

[0049] When collecting signals, the light output by the narrow linewidth laser 1 is divided into two light paths through the fiber coupler 2. The light with an energy ratio of 90% is used as probe light, and the light with an energy ratio of 10% is used as local light. The probe light is modulated into pulses by the acousto-optic modulator 3, and the optical pulses are amplified in energy by the erbium-doped fiber amplifier 4. The fiber Bragg grating 5 is used to filter the amplified spontaneous emission of the erbium-doped fiber amplifier 4. Then, the probe light enters the test optical fiber through the optical fiber circulator 6. The reflected backscattering light is amplified by the erbium-doped fiber amplifier 7 after passing through the optical fiber circulator 6, and the fiber Bragg grating 8 is used to filter the amplified spontaneous emission of the erbium-doped fiber amplifier 7. Then, the backscattering light is coupled with the local light in the 50:50 fiber coupler 9, and then passes through the balanced photodetector 10, and then the data acquisition card 11 completes the data acquisition process.

[0050] As shown in Figure 2 , in combination Figure 3 , the distributed optical fiber acoustic wave sensing event hybrid neural network recognition system of the present application comprises a data preprocessing module, a preliminary feature extraction module, a deep feature extraction module and an event recognition module.

[0051] The data preprocessing module is used for denoising preprocessing of the intrusion event data before the intrusion event data collected by the distributed optical fiber acoustic wave sensing system is input into the preliminary feature extraction module.

[0052] The preliminary feature extraction module comprises a preliminary feature extraction model based on a quantum neural network, which is used to extract preliminary features of the intrusion event by taking the intrusion event data collected by the distributed optical fiber acoustic wave sensing system as input;

[0053] The deep feature extraction module comprises a deep feature extraction model based on a deep learning neural network, which is used to extract deep features of the intrusion event by taking the preliminary features as input;

[0054] The event recognition module comprises an event recognition model based on a quantum neural network, which is used to recognize the corresponding intrusion event by taking the deep features of the intrusion event data as input.

[0055] Before formal use, the preliminary feature extraction model, the deep feature extraction model and the event recognition model need to be trained, and the training method comprises: taking a certain amount of training data as input of the preliminary feature extraction model, so that the preliminary feature extraction model, the deep feature extraction model and the event recognition model are trained according to the respective model operation mechanism until the models converge. In the training process, the cross-entropy loss and the cosine annealing algorithm are adopted. Generally, when the models converge, the accuracy rate of event recognition reaches 90%.

[0056] The present application also provides a distributed optical fiber acoustic wave sensing event hybrid neural network recognition method, as shown in Figure 4As shown, comprising the following steps:

[0057] Step 1: Preprocessing of the distributed optical fiber acoustic wave sensing system to collect intrusion event data, including: performing spectral subtraction and wavelet packet denoising to improve the signal-to-noise ratio of the signal. Among them, the spectral subtraction is performed on each frame signal after framing Fast Fourier transform is performed, where is the first frame, and its frequency spectrum is obtained. The amplitude spectrum and phase spectrum of the signal are represented as

[0058] ;

[0059] Then the amplitude spectrum is processed to estimate the noise amplitude, and the denoised amplitude spectrum , represents the noise amplitude estimation value of the first frame signal, is a parameter for controlling the denoising strength, the processed amplitude spectrum and the phase spectrum of the original signal are combined to obtain the denoised frequency spectrum , and finally the denoised frequency spectrum of all frames is inverse transformed to obtain the denoised signal.

[0060] Wavelet packet denoising uses Db2 wavelet to perform 3-layer wavelet packet decomposition on the signal. Each layer of decomposition divides the signal into two parts: a low-frequency part and a high-frequency part. Set the hard threshold value, all wavelet coefficients less than the threshold value are directly set to zero, and the other coefficients remain unchanged. Finally, the remaining signal coefficients are used for signal reconstruction to obtain the signal after wavelet packet denoising.

[0061] Step 2: The preliminary feature extraction module extracts preliminary features from the denoised intrusion event data. As shown in Figure 5 , in the extraction of preliminary features, the quantum neural network first uses m (m = 3, 5, 7) quantum bits to perform RY rotation quantum gate encoding corresponding to every m data points. The matrix form of the RY rotation quantum gate is as follows:

[0062] ;

[0063] Among them, is the angle calculated from the data points , and the operation process is as follows:

[0064] ;

[0065] Then, the encoded m quantum bits are subjected to RZ rotation quantum gate operation, and the matrix form of the RZ rotation quantum gate is as follows:

[0066] ;

[0067] wherein are learnable parameters trained by the quantum neural network. Next, a CNOT (Controlled-Not) gate is applied to achieve entanglement between the qubits. The CNOT gate is a quantum entanglement operation, usually composed of two qubits: one is the control qubit, and the other is the target qubit. If the control qubit is in state , the state of the target qubit will be flipped. The CNOT gate is applied to qubits (control qubit) and (target qubit) in sequence . In this way, entanglement can be created between the qubits.

[0068] Then, RZ rotation quantum gate operations are performed on the m qubits. The Pauli expectation observations of the last m qubits are respectively taken as the quantum neural network convolution output values of m different channels.

[0069] Step 3: The deep feature extraction module uses its deep feature extraction model based on a deep learning neural network to extract deep features of the intrusion event data. That is, the preliminary features processed by the preliminary feature extraction model are input into the deep feature extraction model of the deep learning neural network for further convolution to achieve deep feature extraction.

[0070] The deep learning neural network is composed of multiple residual convolution blocks and attention mechanisms. The residual convolution block is composed of repeated one-dimensional convolution, batch normalization layer, ReLU activation function, and maximum pooling.

[0071] After the stacked residual convolution blocks, a one-dimensional attention mechanism is used. The attention mechanism is composed of a channel attention mechanism and a spatial attention mechanism. The channel attention mechanism uses adaptive average pooling and maximum pooling operations to extract global channel information, processes the pooled results through one-dimensional convolution to further compress the information dimension, uses the ReLU activation function to introduce nonlinearity, and obtains the final channel weight through the Sigmoid function; the spatial attention mechanism uses one-dimensional convolution to weight the input in the spatial dimension to capture spatial features and calculate the attention weight of each position through the Sigmoid activation function.

[0072] Step 4: The extracted deep features are flattened, and then reduced to n dimensions using a fully connected layer, and then input into the event recognition model based on the quantum neural network. The event recognition model performs event recognition processing. For example Figure 6As shown, the process includes: the quantum neural network output module adopts n qubits, corresponding to n kinds of intrusion events in specific scenarios such as oil and gas pipeline monitoring, perimeter security monitoring, etc., to perform RY rotation quantum gate angle encoding on n-dimensional features, and then perform control RZ rotation quantum gate operations on qubit pairs 、 、 、 、 . The first qubit in the qubit pair is the control qubit, and the second qubit is the target qubit. The target qubit performs RZ rotation quantum gate operation according to the state of the control qubit, and the rotation angle is determined by the parameters learnable by the quantum neural network , realizing the mutual entanglement of the qubit pairs.

[0073] Then perform RY rotation quantum gate operations on all qubits; next, perform control RZ rotation quantum gate operations on qubit pairs 、 、 、 、 . Finally, measure the Braun expectation observation value of each qubit, and the expected value of each qubit obtained by measurement represents the probability of the corresponding event. By comparing the n values, the type of event corresponding to the maximum value is selected as the final event recognition result output.

[0074] In summary, the present application combines quantum neural networks and traditional neural networks, and its advantages mainly include: (1) effectively improving the signal-to-noise ratio of the signal, enhancing the recognizability of the intrusion event, and having good adaptability to the complex environment in the real world; (2) being able to capture complex features that classical neural networks cannot obtain, improving the feature representation capability; (3) through the superposition and entanglement characteristics of the quantum neural network, the subtle differences in the signal can be captured more finely, effectively improving the system's ability to recognize similar event signals; (4) through the quantum bit and rotation quantum gate operation of the quantum neural network output module, fast and accurate classification is realized, improving the precision and response speed of the intrusion event recognition.

[0075] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A hybrid neural network identification system for distributed fiber acoustic wave sensing events, characterized in that, The preliminary feature extraction module, the deep feature extraction module and the event recognition module are included. The preliminary feature extraction module includes a preliminary feature extraction model based on a quantum neural network, which is used to extract preliminary features of an intrusion event by taking intrusion event data collected by a distributed optical fiber acoustic wave sensing system as input. The deep feature extraction module includes a deep feature extraction model based on a deep learning neural network, which is used to extract deep features of the intrusion event by taking the preliminary features as input. The event recognition module includes an event recognition model based on a quantum neural network, which is used to identify the corresponding intrusion event by taking the deep features of the intrusion event data as input. The preliminary feature extraction module extracts preliminary features of an intrusion event, The following methods are included: First, m (m=3, 5, 7) qubits are used RY rotation quantum gate encoding is performed corresponding to every m data points, RZ rotation quantum gate operation is performed on the encoded m qubits, A controlled NOT gate is applied to realize entanglement between qubits, RZ rotation quantum gate operation is performed on the m qubits, m qubits of - the Pauli expectation observations are output values of the quantum neural network convolution for m different channels, respectively, i.e. preliminary features of the intrusion event data.

2. The hybrid neural network identification system of distributed acoustic sensing events of optical fiber of claim 1, wherein, The preliminary feature extraction model, the deep feature extraction model and the event recognition model need to be trained before formal use, and the training method includes: A certain amount of training data is obtained as input of the preliminary feature extraction model, so that the preliminary feature extraction model, the deep feature extraction model and the event recognition model are trained according to the respective model characteristics until the models converge.

3. The hybrid neural network distributed optical fiber acoustic event identification system of claim 1, wherein, The training of the preliminary feature extraction model, the deep feature extraction model and the event recognition model adopts cross-entropy loss and cosine annealing algorithm.

4. The hybrid neural network distributed optical fiber acoustic event identification system of claim 1, wherein, The recognition system further includes a data preprocessing module for denoising preprocessing of the intrusion event data collected by the distributed optical fiber acoustic wave sensing system before the intrusion event data is input into the preliminary feature extraction module.

5. A hybrid neural network identification method for distributed acoustic sensing events based on claim 1, characterized in that, The following steps are included: The feature extraction module uses its preliminary feature extraction model based on a quantum neural network to extract preliminary features of the intrusion event data collected by the distributed optical fiber acoustic wave sensing system; The deep feature extraction module uses its deep feature extraction model based on a deep learning neural network to extract deep features of the intrusion event data by taking the preliminary features as input; The event recognition model built on the quantum neural network in the event recognition module takes the deep features of the intrusion event data as input to identify the corresponding intrusion event; The feature extraction module uses its preliminary feature extraction model based on a quantum neural network to extract preliminary features of the intrusion event data collected by the distributed optical fiber acoustic wave sensing system, including the following steps: First, m (m=3, 5, 7) qubits are used RY rotation quantum gate encoding is performed corresponding to every m data points, RZ rotation quantum gate operation is performed on the encoded m qubits, A controlled NOT gate is applied to realize entanglement between qubits, RZ rotation quantum gate operation is performed on the m qubits, m qubits of - the Pauli expectation observations are output values of the quantum neural network convolution for m different channels, respectively, i.e. preliminary features of the intrusion event data.

6. The distributed optical fiber acoustic wave sensing event hybrid neural network identification method of claim 5, wherein, The event recognition model takes the deep features of the intrusion event data as input to identify the corresponding intrusion event, including the following steps: Using n qubits , corresponding to n kinds of intrusion events respectively, the deep features are flattened, The flattened deep features are output as n-dimensional vectors through a fully connected layer, The n qubits are encoded with RY rotation quantum gate angles, Pairs of qubits 、 、 、 、 Control the RZ rotation quantum gate operation to achieve mutual entanglement of quantum bit pairs; RY rotation quantum gate operation is performed on all qubits; Pairs of qubits , , , , Controlled-RZ rotation quantum gate operations, measurements of each qubit - the Pauli expectation observables, The n probability values are compared, and the event corresponding to the maximum probability value is selected as the final recognition result of the system.

7. The distributed hybrid neural network identification method of claim 5, wherein, Before the intrusion event data is input into the initial feature extraction module, the intrusion event data is preprocessed by denoising.

8. The method of claim 7, wherein, The denoising preprocessing includes a spectral subtraction method and a wavelet packet denoising method.

Images