DAS signal cross-scene multi-class classification method, system and device and medium

Abstract

The invention discloses a DAS signal cross-scene multi-category classification method, system and device and a medium, belongs to signal identification and classification in the technical field of optical fiber sensing, and aims to solve the technical problems that in the prior art, only DAS signals of single scene events can be identified, and cross-scene multi-category DAS signals cannot be identified and classified in complex scenes. The method comprises: obtaining a sample and a label; constructing a signal identification and classification model comprising a feature extraction network and an identification and classification network; the identification classification network comprises a tree classifier, each non-leaf node of the tree classifier comprises a node classification sub-network and an output layer, and the node classification sub-network comprises a one-dimensional convolution layer, a batch normalization layer, an ReLU layer, a one-dimensional maximum pooling layer, a one-dimensional convolution layer, a batch normalization layer, an ReLU layer and a one-dimensional maximum pooling layer. The output layer comprises a conversion layer, a full connection layer, a ReLU layer, a full connection layer, a ReLU layer and a Softmax layer which are arranged in sequence; training a signal identification classification model; and classifying the signals in real time.

Description

Technical Field

[0001] The present invention belongs to the field of optical fiber sensing technology, and relates to a method for identifying and classifying optical fiber distributed acoustic sensor (DAS) signals, and in particular to a method, system, device and medium for cross-scenario multi-category classification of DAS signals. Background Art

[0002] Distributed optical fiber sensor (DAS) based on linear phase demodulation phase-sensitive optical time domain reflectometry (Φ-OTDR) technology is a new sensing technology that can realize continuous distributed detection of vibration and sound fields. It is also a sound wave or vibration sensing method with high sensitivity, high cost performance, long sensing distance, wide sensing range and strong anti-electromagnetic interference ability. Due to these unique advantages, various seismic sources can be sensed and located with high precision using widely laid underground reminder cables, making it widely used in seismic exploration, oil / gas pipeline monitoring, natural disaster prediction, communication / power / railway cables and perimeter security.

[0003] In order to improve the sensing capability of DAS, on the one hand, many researchers have improved the hardware, data processing and other parts to ensure the high fidelity of the sensing signal, improve the sensitivity, detection distance, frequency response bandwidth, etc. On the other hand, the post-processing methods of DAS single-source detection and identification have developed rapidly and have been widely studied, including traditional machine learning methods based on artificial feature extraction and new deep learning methods.

[0004] The invention patent application with application number 202410159856.7 discloses a DAS signal classification method based on improved wavelet threshold denoising, which includes the following steps: performing wavelet decomposition on the DAS signal Q to obtain detail coefficients cD{cD1, cD2, ..., cDM} and approximate coefficients cA{cA1, cA2, ..., cAM}; combining the Akaike Information Criterion (AkaikeInformation Criterion, AIC) determines the threshold, uses a new threshold function to perform threshold quantization on the detail coefficient cD after wavelet decomposition, and obtains the processed detail coefficient cD'{cD1', cD2', ..., cDM'}; cDm' represents the mth layer detail coefficient after threshold quantization; uses the processed detail coefficient CD' and the approximate coefficient cAM of the last layer of wavelet decomposition to perform inverse wavelet transform to complete signal reconstruction and obtain the signal Q' after wavelet denoising; extracts the time domain features of the signal Q' after wavelet denoising, uses the extracted time domain features to construct a data set, and divides the data set into a training set and a test set; uses the WKNN algorithm to perform signal classification on the data in the test set. This classification method uses wavelet denoising preprocessing, then extracts signal features to construct a data set, and uses WKNN to classify DAS signals, which improves the classification accuracy.

[0005] The invention patent application with application number 202410135557.X discloses an intelligent DAS and its aliased signal directional target separation method, which first connects the buried optical fiber to the optical fiber vibration sensing system to collect various types of vibration source signals; constructs a single-source signal database, and then constructs a mixed signal data set through the single-source data set; based on the collected signal, constructs a target signal directional separation network; the network consists of two parts: an embedded vector (d-vector) extraction network and a target signal separation network; the embedded vector extraction network generates an embedded vector for the reference signal, and the target signal separation network uses the embedded vector to estimate the target signal in the mixed signal; the embedded vector extraction network includes three LSTM layers and one fully connected layer; the target signal separation network includes three LSTM layers and two fully connected layers, a linear activation unit ReLU function is added before each fully connected layer, and a sigmoid activation function is added after the last fully connected layer. This separation method can achieve the separation of the target signal using a single-channel aliased signal, which solves the problem that the optical fiber sensing system is difficult to detect and identify the target signal due to multi-source interference in a complex buried environment, and has great potential application value.

[0006] In recent years, with the rapid development of artificial intelligence technology, more and more deep learning technologies have been applied to DAS. From basic one-dimensional and two-dimensional convolutional neural networks (1 / 2D-CNN) to more advanced attention-based long short-term memory (ALSTM), dual path network (DPN), multi-scale CNN (MSCNN), Faster recurrent convolutional neural network (R-CNN), efficient YOLO, 1D-CNN-Bi-LSTM and other combination models that simultaneously extract DAS hidden spatiotemporal information, as well as transfer learning, semi-supervision, and unsupervised methods, the detection and recognition capabilities of DAS for events have been effectively improved. However, most of the current DAS signal recognition and classification technologies are concentrated on working effectively in a single scene. For cross-scene multi-category classification tasks in different scenes, their generalization ability is poor, and the recognition and classification effects of multiple signal categories are poor; and when facing different scenes, events in multiple scenes cannot be effectively classified in one network. Summary of the invention

[0007] The purpose of the present invention is to solve the technical problem that the prior art can only identify DAS signals of single scene events and cannot identify and classify DAS signals of multiple categories across scenes in complex scenes. A method, system, device and medium for cross-scene multi-category classification of DAS signals are provided. By adding a constructed tree classifier to the constructed recognition and classification model, effective recognition and classification of multi-category events in multiple scenes can be achieved in a single network.

[0008] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions: A DAS signal cross-scenario multi-category classification method comprises the following steps: Step S1, obtaining signal sample data; Collect various types of vibration source signal samples in various scenarios, and process the vibration source signal samples to obtain the corresponding time-frequency matrix; label the vibration source signal samples to obtain label data; Step S2, constructing a signal recognition and classification model; Construct a signal recognition and classification model, which includes a feature extraction network and a recognition and classification network; the feature extraction network extracts features from the vibration source signal and outputs a feature vector; the recognition and classification network includes a constructed tree classifier, each non-leaf node of the tree classifier includes a node classification subnetwork and an output layer, the node classification subnetwork includes a first one-dimensional convolution layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional maximum pooling layer, a second one-dimensional convolution layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional maximum pooling layer, and the output layer includes a conversion layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a Softmax layer, which are arranged in sequence; Step S3, training a signal recognition classification model; The signal recognition and classification model constructed in step S2 is trained using the time-frequency matrix and label data obtained in step S1 to obtain a final signal recognition and classification model; Step S4, real-time classification of signals; The vibration source signal to be classified is obtained and processed into a time-frequency matrix, and is input into the signal recognition and classification model obtained in step S3, and the signal recognition and classification model outputs the signal category.

[0009] Furthermore, in step S1, when using the vibration source signal sample, the buried optical fiber is connected to the optical fiber vibration sensing system to collect various types of vibration source signals in various scenarios; The optical fiber vibration sensing system includes a narrow linewidth laser, an acousto-optic modulator, an erbium-doped fiber amplifier, a circulator, a phase demodulation module, an interferometer, a coupler, a photodetector and an analog-to-digital converter; the continuous light emitted by the narrow linewidth laser is modulated into pulse light by the acousto-optic modulator, the erbium-doped fiber amplifier amplifies the pulse light and transmits it into the optical fiber through the circulator, the pulse light generates backward Rayleigh scattered light when propagating in the optical fiber, the phase demodulation module receives the Rayleigh scattered light and filters it, and then linearly demodulates the backward Rayleigh scattered light through the interferometer and the coupler in turn to obtain a demodulated signal, the demodulated signal is received by the photodetector and converted into a digital electrical signal through the analog-to-digital converter.

[0010] Furthermore, in step S2, the feature extraction network includes three LSTM layers and one fully connected layer; the time-frequency matrix serves as the input of the feature extraction network, the feature extraction network outputs a feature vector, and the feature vector serves as the input of the recognition and classification network.

[0011] Furthermore, in step S3, when training the signal recognition classification model, the total loss function It is expressed as:

[0012]

[0013]

[0014]

[0015]

[0016]

[0017] in, represents the total loss of feature extraction, represents the total loss of tree classification, Indicates The first The loss of data, Indicates The first samples, Representation sample With The similarity of the centroids of class events, Representation sample With The similarity of the centroids of class events, Indicates the total number of events, Indicates The loss weight of the node classifier is Indicates The loss of the node classifier is Represents the features extracted by the feature extraction network, represents the total number of node classifiers, Represents the one-hot encoding of the labels corresponding to each atomic object class at the current node, represents the predicted probability value, Indicates The number of child nodes of non-leaf nodes, represents the zth one-hot encoding value, Represents the probability value of belonging to the zth prediction.

[0018] Furthermore, the similarity The calculation method is:

[0019]

[0020]

[0021]

[0022] parameter , Use the standard gradient descent algorithm for iteration and update. The specific algorithm is:

[0023]

[0024]

[0025] in, , are all trainable parameters. represents the event centroid; Indicates that samples are removed when calculating the event centroid , making training more stable; represents the number of samples, represents the qth data of the kth type of event, is the learning rate. During training, the Adam optimization algorithm is used to iteratively update the parameters of each node classification subnetwork .

[0026] Furthermore, in step S2, the steps of constructing the tree classifier are: Step S2-1, calculating the average feature representation of each event type; Step S2-2, adding the average feature representation of all event types to the root node; Step S2-3, using spectral clustering to group the time types in the current root node, and treating each group as a child node of the current node; Step S2-4, recursively execute step S2-3 for each child node until the node contains only one event type, that is, a leaf node.

[0027] Furthermore, when a new event category needs to be added, the specific steps are as follows: Step 01: Input the vibration source signal samples corresponding to the newly added event categories into the trained signal recognition and classification model. The signal recognition and classification model assigns the categories of the newly added events to the best matching lowest level non-leaf nodes. ;

[0028] Step 02: To the lowest non-leaf node of the best match of the tree classifier Insert a new leaf node, and update the tree classifier and the signal recognition classification model to obtain a new tree classifier and a new signal recognition classification model;

[0029] The newly inserted leaf node is the newly added event type, and the best matching lowest level non-leaf node in the original tree classifier is Corresponding to the non-leaf nodes in the new tree classifier ;

[0030] Step 03, training a new signal recognition and classification model; during training, first freeze the parameters of the feature extraction network in the new signal recognition and classification model, as well as the leaf nodes in the recognition and classification network. Then, the vibration source signal samples corresponding to the newly added event categories and the best matching lowest-level non-leaf nodes are used to classify the parameters of the node classification subnetwork in the non-leaf nodes other than The samples that originally corresponded to the samples in the new training set are used to train the non-leaf nodes in the new tree classifier. The node classification sub-network is trained.

[0031] A DAS signal cross-scenario multi-category classification system, comprising: The signal sample data acquisition module is used to collect various types of vibration source signal samples in various scenarios, and process the vibration source signal samples to obtain the corresponding time-frequency matrix; the vibration source signal samples are labeled to obtain label data; A signal recognition and classification model construction module is used to construct a signal recognition and classification model, which includes a feature extraction network and a recognition and classification network; the feature extraction network extracts features from the vibration source signal and outputs a feature vector; the recognition and classification network includes a constructed tree classifier, each non-leaf node of the tree classifier includes a node classification subnetwork and an output layer, the node classification subnetwork includes a first one-dimensional convolution layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional maximum pooling layer, a second one-dimensional convolution layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional maximum pooling layer, and the output layer includes a conversion layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a Softmax layer, which are arranged in sequence; The signal recognition and classification model training module is used to train the signal recognition and classification model constructed by the signal recognition and classification model construction module using the time-frequency matrix and label data obtained by the signal sample data acquisition module to obtain the final signal recognition and classification model; The signal real-time classification module is used to obtain the vibration source signal to be classified and process it into a time-frequency matrix, and input the signal recognition and classification model obtained by the signal recognition and classification model training module, and the signal recognition and classification model outputs the signal category.

[0032] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0033] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor is caused to execute the steps of the above method.

[0034] The beneficial effects of the present invention are as follows: 1. In the present invention, a deep learning-based fiber distributed acoustic sensor (DAS) cross-scene multi-category recognition method is proposed for the first time, which realizes the accurate recognition of multiple (more than 10) categories of events in multiple (three or more) different scenes in a single network model, solves the problem that the traditional recognition network can only recognize the DAS signal of a single scene event and cannot recognize and classify cross-scene multi-category DAS signals in complex scenes, and improves the recognition accuracy.

[0035] 2. In the present invention, according to the differences in time-frequency characteristics of different types of vibration sources, the feature extraction of the target signal is realized through encoder-LSTM (feature extraction network); compared with the traditional convolutional neural network, the model parameters of this method are small, the convergence speed is fast, and the timing information in the signal can be effectively captured, so that each feature vector is as close to its center point as possible and away from all other center points, thereby obtaining an effective feature vector, which is convenient for distinguishing different event types.

[0036] 3. In the present invention, a cross-scene multi-category recognition network (recognition classification network) based on a tree classifier is proposed. Compared with the traditional fan-out network structure, it can effectively complete the recognition of event signals in different scenarios and improve the recognition accuracy. Especially when facing event recognition tasks with high similarity, the recognition accuracy is greatly improved, and the network structure is easy to adjust and has strong stability.

[0037] 4. In the present invention, the incremental learning method based on the tree classifier structure can effectively identify new events without retraining the entire network, and its training time is exponentially reduced compared to the traditional method of retraining on a new data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a schematic diagram of the method of the present invention; Figure 2 It is a structural schematic diagram of the optical fiber vibration sensing system of the present invention; Figure 3 It is a schematic diagram of the process of feature extraction in the present invention; Figure 4It is a schematic diagram of the structure of the feature extraction network in the present invention; Figure 5 It is a schematic diagram of a feature tree constructed based on nine types of events in three scenarios in the present invention; Figure 6 It is a schematic diagram of the structure of the identification and classification network in the present invention; Figure 7 It is the scene map, time signal and spectrum map of ten types of events in three scenarios collected in the present invention; Figure 8 This is the visual comparison result of the feature extraction effect of the encoder-LSTM of the present invention and the traditional method; Among them, a-encoder-LSTM, b-one-dimensional convolutional neural network, c-two-dimensional convolutional neural network; Fig. 9 It is the confusion matrix of the test results of the three methods of encoder-LSTM + 1-D CNN Tree-Classifier, encoder-LSTM + SVM Tree-Classifier, and encoder-LSTM + Softmax in the field test of the present invention; Among them, a-encoder-LSTM + 1-D CNN Tree-Classifier, b-encoder-LSTM + SVMTree-Classifier, c-encoder-LSTM + Softmax; Fig.10 It is the five-fold cross-validation result of the three methods of encoder-LSTM + 1-D CNN Tree-Classifier, encoder-LSTM + SVM Tree-Classifier, and encoder-LSTM + Softmax in the field test of the present invention; Fig.11 It is a comparison chart of the test results of the incremental learning method proposed by the present invention and the traditional retraining of the entire data set in the field test of the present invention, including training time, test time and recognition accuracy. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0040] Therefore, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0041] Example 1 In smart city monitoring applications, the vibration signals detected by the optical fiber sensing system usually include multiple typical event types in multiple scenarios, such as manual excavation, mechanical excavation, traffic interference, etc. in urban security scenarios, pipeline leakage signals in pipeline monitoring scenarios, and train passing signals in track monitoring scenarios. Therefore, this embodiment proposes a cross-scenario multi-category classification method for DAS signals, which is based on a tree classifier to achieve accurate recognition of multi-category events in the above multiple scenarios in a single network, and adds an incremental learning function to the classifier, such as Figure 1 As shown, it mainly includes the following steps: Step S1, obtaining signal sample data; Various types of vibration source signal samples are collected in various scenarios, and the vibration source signal samples are processed to obtain corresponding time-frequency matrices; the categories of the vibration source signal samples are marked to obtain label data.

[0042] The buried optical fiber is connected to the optical fiber vibration sensing system to collect various types of vibration source signals in various scenarios and build a signal database for different types of events.

[0043] Step S2, constructing a signal recognition and classification model; A signal recognition and classification model is constructed, which includes a feature extraction network and a recognition and classification network; the feature extraction network extracts features from the vibration source signal and outputs a feature vector; the recognition and classification network includes a constructed tree classifier, and each non-leaf node of the tree classifier includes a node classification subnetwork and an output layer, and the node classification subnetwork includes a first one-dimensional convolutional layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional maximum pooling layer, a second one-dimensional convolutional layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional maximum pooling layer, and the output layer includes a conversion layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a Softmax layer, which are arranged in sequence.

[0044] Step S3, training a signal recognition classification model; The time-frequency matrix and label data obtained in step S1 are used to train the signal recognition and classification model constructed in step S2 to obtain the final signal recognition and classification model.

[0045] Construct a feature extraction network, input the time-frequency matrix of the event signal, calculate the loss function by calculating the similarity matrix of the output feature vector, iteratively update the network parameters, and complete the network training. Complete the construction of the tree classifier based on spectral clustering, add node classifiers, and then complete the training of the tree classifier based on the multi-task learning method, and use the optimal model to identify the event to be tested.

[0046] Step S4, real-time classification of signals; The vibration source signal to be classified is obtained and processed into a time-frequency matrix, and is input into the signal recognition and classification model obtained in step S3, and the signal recognition and classification model outputs the signal category.

[0047] This embodiment proposes for the first time a deep learning-based cross-scene multi-category recognition method for fiber optic distributed acoustic sensors (DAS), which achieves accurate recognition of multiple (more than 10) categories of events in multiple (three or more) different scenes in a single network model. It solves the problem that traditional recognition networks can only recognize DAS signals of single-scene events and cannot recognize and classify cross-scene multi-category DAS signals in complex scenes, and improves the recognition accuracy.

[0048] Example 2 This embodiment further illustrates the optical fiber vibration sensing system in step 1 of the first embodiment.

[0049] In Example 1, a distributed fiber acoustic sensor (DAS) system based on phase-sensitive optical time-domain reflectometry technology is used to collect vibration signals of different types of events. It uses backward Rayleigh scattering to monitor the state change of the optical fiber. Rayleigh scattering is a kind of elastic light scattering caused by medium density fluctuations. When a vibration event occurs at a certain position on the optical fiber, the relative position of the center of Rayleigh scattering will change, thereby changing the phase and realizing the perception of the vibration event.

[0050] The structure of the optical fiber vibration sensing system (DAS system based on the principle of phase-sensitive optical time domain reflectometry (Φ-OTDR)) is as follows: Figure 2 As shown, the system mainly includes a narrow linewidth laser, an acousto-optic modulator, an erbium-doped fiber amplifier, a circulator, a phase demodulation module, an interferometer, a coupler, a photodetector and an analog-to-digital converter; the specific process is: the continuous light emitted by the narrow linewidth laser is modulated into a pulse light with a specific period and pulse width by the acousto-optic modulator, the erbium-doped fiber amplifier amplifies the pulse light and transmits it into the optical fiber through the circulator, the pulse light generates backward Rayleigh scattered light when propagating in the optical fiber, the phase demodulation module receives the Rayleigh scattered light and filters it, and then linearly demodulates the backward Rayleigh scattered light through the interferometer and the coupler in turn to obtain a demodulated signal, the demodulated signal is received by the photodetector and converted into a digital electrical signal by the analog-to-digital converter.

[0051] At each pulse transmission time, the DAS forms an OTDR trace that changes with the spatial position of the optical fiber. By accumulating the OTDR traces received over time, a two-dimensional time-frequency matrix is ​​formed for subsequent analysis and processing.

[0052] Example 3 This embodiment further optimizes the sample data preparation and acquisition process based on the second embodiment.

[0053] In order to verify the performance of the cross-scenario multi-category classification method for DAS signals proposed in this embodiment, this embodiment collected ten types of events in three different scenarios for field testing.

[0054] The five types of events in the following scenario are collected by the DAS system of underground communication cables in Tongren and Wuhan. Figure 7 As shown in (a), a total of five types of events were collected, including 1: mechanical excavation; 2: mechanical tamping; 3: manual excavation; 4: background noise; 5: traffic interference.

[0055] The data in scenario 2 are the actual leakage and pipeline vibration acoustic signals collected in Chongqing. Figure 7 As shown in (b), two types of events are collected, including 6: pipeline leakage; 7: pipeline no leakage.

[0056] The data in scenario 3 is collected by the rail disease monitoring DAS system in Zhangjiakou City, Hebei Province. Figure 7 As shown in (c), a total of three events were collected, including 8: no train passing; 9: train passing; and a new category of events: 10: train passing (with track damage).

[0057] In order to enrich the data, this embodiment performs simple data enhancement on all ten types of data, that is, the original data is sliced, and the length of each segment is 2000 sampling points. The constructed data set is shown in Table 1. The training data, verification data and test data are randomly divided in a ratio of 6:2:2, and there is no duplication among the three.

[0058] Table 1 Typical event databases in three scenarios

[0059]

[0060] Example 4 This embodiment further illustrates the feature extraction network in step 2 of embodiment 1.

[0061] LSTM is more effective for mining and expressing long time series signal features. Therefore, the feature extraction network in this embodiment uses the encoder-LSTM network to extract the feature vector of the DAS signal, making each feature vector as close to its center point as possible and away from all other center points, so as to obtain an effective feature vector and facilitate the distinction of different event types. The basic structure of the feature extraction network is as follows: Figure 3 As shown, it includes three LSTM layers and one fully connected layer, and its structural parameters are shown in Table 2; the time-frequency matrix is ​​used as the input of the feature extraction network, and the feature extraction network outputs the feature vector, which is used as the input of the recognition and classification network.

[0062] Table 2 Structural parameters of feature extraction network (encoder-LSTM network)

[0063] Example 5 This embodiment further illustrates the training of the signal recognition and classification model in step 3 of embodiment 1.

[0064] When training a signal recognition classification model, the total loss function is It is expressed as:

[0065]

[0066] When training the feature extraction network, the total feature extraction loss of the feature extraction network for:

[0067]

[0068] The loss function for each sample signal in the training batch for:

[0069]

[0070] in, Indicates The first samples, Representation sample With The similarity of the centroid of class events, The calculation method is:

[0071]

[0072] The centroid of an event is calculated as follows:

[0073]

[0074] in, Indicates that samples are removed when calculating the event centroid , making training more stable.

[0075] Then, the standard gradient descent algorithm is used to iterate and update the network parameters. , , the specific algorithm is:

[0076]

[0077]

[0078] When training the recognition classification network, the total tree classification loss of the recognition classification network for:

[0079]

[0080] For all node classifiers, this embodiment uses the same cross entropy loss function for learning:

[0081] In addition, this embodiment also uses the Adam optimization algorithm to iteratively update the parameters of each node classification subnetwork , the specific algorithm is:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] The Adam algorithm combines the Momentum algorithm and the RMSProp algorithm to solve the problem that the stochastic gradient descent algorithm may not be able to reach the global optimal point, while making the network converge faster.

[0088] in, represents the total loss of feature extraction, represents the total loss of tree classification, Indicates The first The loss of data, Indicates The first samples, Representation sample With The similarity of the centroids of class events, Representation sample With The similarity of the centroids of class events, Indicates the total number of events, Indicates The loss weight of the node classifier is Indicates The loss of the node classifier is Represents the features extracted by the feature extraction network, represents the total number of node classifiers; Represents the one-hot encoding of the labels corresponding to each atomic object class at the current node, ; represents the predicted probability value, ; Indicates The number of child nodes of a non-leaf node, the maximum value of which is B; represents the zth one-hot encoding value, represents the probability value belonging to the zth prediction, , are all trainable parameters. represents the event centroid; Indicates that samples are removed when calculating the event centroid , making training more stable; represents the number of samples, represents the qth data of the kth type of event, is the learning rate, represents the gradient momentum The correction value of Represents the squared gradient The correction value of , to avoid the moving exponential average having a large difference with the starting value at the beginning of the iteration; represents the smoothing term, set to ; Represents the gradient momentum accumulated by the loss function during the first t-1 rounds of iterations, with an initial value of 0; Represents the square gradient accumulated by the loss function during the first t-1 iterations, with an initial value of 0; Represents the exponential decay rate of the corresponding Momentum algorithm at the tth iteration, Represents the exponential decay rate of the RMSProp algorithm corresponding to the tth iteration; Indicates the exponential decay rate of the corresponding Momentum algorithm, which is set to 0.9; Indicates the exponential decay rate corresponding to the RMSProp algorithm, which is set to 0.999; Indicates The loss weight of the node classifier is Indicates The loss of the node classifier is Represents the features extracted by the feature extraction network [A1].

[0089] Example 6 This embodiment further illustrates the construction and training of the tree classifier in step 2 of embodiment 1.

[0090] First, customize the tree classifier structure according to application requirements. The specific process is as follows: for the features of N types of events extracted by the feature extraction network, this embodiment needs to learn a feature tree , which contains a set of nodes V, a set of feature sets F, and a set of edges E. Each non-leaf node , which contains a set of atomic object classes L(c) and a set of feature subsets corresponding to them For a node, it contains only a subset of the atomic object classes associated with its parent node, that is, if ,but . Each leaf node contains a specific event category, so there are a total of N leaf nodes in the constructed feature tree. Before starting to build the feature tree, this embodiment needs to give a parameter B, which represents its branch number, that is, the number of child nodes under the same parent node. When the number of atomic object classes in a given non-leaf node is less than B, B is set at the node to the number of atomic object classes contained in the node.

[0091] The process of building a tree classifier is as follows: 1) Calculate the average feature representation of N event types; For a given event category, this embodiment chooses to use the average feature of the event to represent its main features. For the The set of event features, in this embodiment, is calculated by The sum of the features extracted by encoder-LSTM for all such events in , and then the average is taken as the overall feature representation of this type of event (i.e., the average feature representation). The overall feature representation (i.e., the average feature representation) of a class of events is defined as :

[0092]

[0093] in, Represents the number of events in the collection, Representing a collection The Class samples.

[0094] 2) Add feature representations of all event types to the root node; 3) Use spectral clustering to group the event types in the current node and treat each group as a child node of the current node; For a given non-leaf nodes, this embodiment obtains the corresponding feature subset, and then uses the Gaussian kernel function to calculate the feature similarity between the two types of events:

[0095]

[0096] in, is the bandwidth, which controls the radial extent of the action.

[0097] By calculating the similarity between the M atomic object classes in the node, we can get the M*M similarity matrix at the node. .

[0098] The Laplacian matrix is ​​then calculated and eigen-decomposed to obtain a low-dimensional representation of the data, thereby increasing the cluster spacing and reducing computation time:

[0099] in, is the degree matrix;

[0100]

[0101] right Perform eigendecomposition and arrange the eigenvalues ​​from small to large, taking the first The eigenvectors corresponding to the eigenvalues ​​form the matrix U, let For U Row vector, then .

[0102] Finally, clustering is performed using the K-means algorithm to obtain the child node set of the current node.

[0103] 4) Recursively perform the third operation on each child node until the node contains only one event type, i.e., a leaf node.

[0104] Figure 5 The results of nine types of events in three different scenarios.

[0105] After the tree classifier is constructed, this embodiment defines a node classifier at each non-leaf node to identify the next-level child node to which the input event belongs, and proposes a multi-task structure learning method to jointly train each node classifier to reduce the computational cost and improve the discrimination ability of the node classifier. Before starting training, the data labels need to be reinitialized to adapt to the distinction criteria of the same event when it arrives at different nodes, such as Figure 5 As shown, under the condition that B is set to 2, in the root node, the labels of events 1, 2, and 3 can be defined as 1, and the labels of events 4, 5, 6, 7, 8, and 9 can be defined as 2. In the first non-leaf node of the next layer, this embodiment needs to define the labels of events 1 and 3 as 1, and the label of event 2 as 2. By analogy, the labels of a class of events at different nodes are obtained to facilitate the training of node classifiers.

[0106] In this embodiment, the encoder-LSTM network is regarded as a shared part, and each node classifier has its own independent network structure to extract the required features. The node classifier is a relatively simple 1-D CNN structure, which speeds up the calculation efficiency. The node classifier network structure is as follows: Figure 6 As shown in the structure, this embodiment first uses 32 one-dimensional convolution kernels of size 4 to extract the initial features after the encoder-LSTM network processing. In the second convolution layer, the one-dimensional convolution kernel of the same size is used to extract additional features, and the number of channels is expanded to 64 to extract deeper features. Finally, the features are input to the output layer, which is composed of two fully connected layers. Finally, the softmax layer is used to convert the obtained feature vector into the corresponding probability vector, and according to the maximum probability in the probability vector, it is determined which node of the next layer this type of event should go to.

[0107] Example 7 This embodiment further illustrates the feature extraction network trained in step 3 of embodiment 1.

[0108] In the process of training the feature extraction network, this embodiment sets the batch size to the number of event types N, the number of samples to 50, and the learning rate to 0.01. In order to verify the effectiveness of the feature extraction network (encoder-LSTM network) for DAS signal feature extraction, this embodiment compares the encoder-LSTM network model proposed by the present invention with the 1-D CNN and 2-D CNN models currently widely used in the field of signal recognition.

[0109] After the three network trainings were completed, the three methods were used to extract and classify the features of nine types of core events to verify the effectiveness of the algorithms. This embodiment uses principal component analysis (PCA) to convert the feature dimension into three dimensions, calculate the average Euclidean distance between the class centers of each type of event features, and visualize the features to observe the feature extraction effect. Figure 8 As shown in Figure 8, compared with 1-D CNN and 1-D CNN, the features of different events extracted by encoder-LSTM can be well separated, and the features of the same type of events can be well integrated, while the features extracted by one-dimensional CNN and two-dimensional CNN are relatively scattered, and the features of different events are easy to overlap. Therefore, encoder-LSTM can more effectively extract the features of different types of events, which provides a good guarantee for the subsequent learning of the tree classifier.

[0110] Example 8 This embodiment further illustrates the recognition and classification network trained in step 3 of embodiment 1.

[0111] The encoder-LSTM + 1-D CNN Tree-Classifier model proposed in the present invention is compared with the encoder-LSTM + SVM Tree-Classifier model using a tree-shaped SVM classifier structure in the classification part, and the traditional encoder-LSTM + Softmax. The original data sets of the three models are all the field experimental data in Table 2. The experimental parameters are set as batch-size = 4, BATCH_SIZE = 4, LR (learning rate) = 0.001, and epoch = 200.

[0112] In this embodiment, the above three models are trained respectively and tested using the test set data to obtain the confusion matrix as follows: Fig. 9 As shown, the results show that the encoder-LSTM + 1-D CNN Tree-Classifier proposed in the present invention has the best recognition performance, followed by encoder-LSTM + Softmax and encoder-LSTM + SVM Tree-Classifier.

[0113] In order to verify the effectiveness and generalization ability of the algorithm proposed in the present invention, this embodiment uses a five-fold cross-validation method to verify the three models. The verification results are as follows: Fig.10The results show that encoder-LSTM + 1-D CNNTree-Classifier has the highest recognition accuracy and the most stable performance. Its average accuracy reaches 96.41%, which is 1.22% and 1.06% higher than the accuracy of the other two models respectively, indicating that the generalization ability of the method proposed in this invention is better than that of the other two models.

[0114] Example 9 This embodiment further illustrates that the tree classifier in Embodiment 1 needs to add a new event category.

[0115] When you need to add a new event category, the specific steps are as follows: Step 01: Input the vibration source signal samples corresponding to the newly added event categories into the trained signal recognition and classification model. The signal recognition and classification model assigns the categories of the newly added events to the best matching lowest level non-leaf nodes. ;

[0116] Step 02: To the lowest non-leaf node of the best match of the tree classifier Insert a new leaf node, and update the tree classifier and the signal recognition classification model to obtain a new tree classifier and a new signal recognition classification model;

[0117] The newly inserted leaf node is the newly added event type, and the best matching lowest level non-leaf node in the original tree classifier is Corresponding to the non-leaf nodes in the new tree classifier ;

[0118] Step 03, training a new signal recognition and classification model; during training, first freeze the parameters of the feature extraction network in the new signal recognition and classification model, as well as the leaf nodes in the recognition and classification network. Then, the vibration source signal samples corresponding to the newly added event categories and the best matching lowest-level non-leaf nodes are used to classify the parameters of the node classification subnetwork in the non-leaf nodes other than The samples that originally corresponded to the samples in the new training set are used to train the non-leaf nodes in the new tree classifier. The node classification sub-network is trained.

[0119] Specifically: For a set of newly added event data, this embodiment designs an incremental learning method to solve its classification problem. Before the training starts, all the data of the newly added events are sequentially passed through the trained encoder-LSTM and tree classifier networks, and reach a best matching non-leaf node c in the penultimate layer of the tree classifier, where the atomic object class originally existing in the node has the greatest similarity with the new event type.

[0120]

[0121] in and It is the parameter of all parent nodes above node c.

[0122] Therefore, the tree classifier of this embodiment can identify a non-leaf node that best matches the new event type. Then, a new leaf node is inserted into the non-leaf node to represent the newly added event type, and the node classifier network of the non-leaf node c is reinitialized, and the output vector length of the output layer of the node classifier is changed from the original B to B+1.

[0123]

[0124] in, and are the new parameters of the classifier at node c. is the predicted probability vector.

[0125] The parameters of the encoder-LSTM network and other node classifiers at non-leaf nodes will be frozen. The node classifier at node c is trained using the dataset of the newly added event type and the dataset of the event type originally contained in the non-leaf node c. The label of the newly added event type is M+1, and the label at the node classifier is redefined as B+1 in this embodiment. The loss function of the node classifier is defined as :

[0126]

[0127] in, represents the loss function during incremental training, The samples representing the newly added event types and the samples of the event types originally included in the non-leaf node c, Indicates the one-hot label of all events including the newly added event type at the non-leaf node. The Adam optimization algorithm is also used to iteratively update the parameters of the node classifier network.

[0128] This example uses event 10 as an incremental event for training and testing, and compares two methods: (1) the incremental learning method proposed in the present invention; and (2) the traditional retraining joint training set (the original training set plus the new event data set). Fig.11 The comparison of the accuracy, training time, and test time between the incremental learning and joint training methods is shown. It can be seen that the incremental learning method proposed in this embodiment can significantly reduce the computing time cost while ensuring the accuracy.

[0129] In summary, the encoder-LSTM + 1-D CNN Tree-Classifier model proposed in this embodiment has the advantages of low computational cost, high accuracy, short training and testing time, and high scalability in identifying new events. It has obvious online real-time processing advantages and is suitable for practical applications.

[0130] Example 10 This embodiment provides a DAS signal cross-scenario multi-category classification system, including: The signal sample data acquisition module is used to collect various types of vibration source signal samples in various scenarios, and process the vibration source signal samples to obtain the corresponding time-frequency matrix; the vibration source signal samples are labeled to obtain label data; A signal recognition and classification model construction module is used to construct a signal recognition and classification model, which includes a feature extraction network and a recognition and classification network; the feature extraction network extracts features from the vibration source signal and outputs a feature vector; the recognition and classification network includes a constructed tree classifier, each non-leaf node of the tree classifier includes a node classification subnetwork and an output layer, the node classification subnetwork includes a first one-dimensional convolution layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional maximum pooling layer, a second one-dimensional convolution layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional maximum pooling layer, and the output layer includes a conversion layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a Softmax layer, which are arranged in sequence; The signal recognition and classification model training module is used to train the signal recognition and classification model constructed by the signal recognition and classification model construction module using the time-frequency matrix and label data obtained by the signal sample data acquisition module to obtain the final signal recognition and classification model; The signal real-time classification module is used to obtain the vibration source signal to be classified and process it into a time-frequency matrix, and input the signal recognition and classification model obtained by the signal recognition and classification model training module, and the signal recognition and classification model outputs the signal category.

[0131] Embodiment 11 A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of a DAS signal cross-scene multi-category classification method.

[0132] The computer device may be a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with the user through a keyboard, a mouse, a remote control, a touch pad, or a voice control device.

[0133] The memory includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or D interface display memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the computer device. Of course, the memory may also include both the internal storage unit of the computer device and its external storage device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the cross-scenario multi-category classification method of the DAS signal. In addition, the memory may also be used to temporarily store various types of data that have been output or are to be output.

[0134] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is generally used to control the overall operation of the computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data, such as running the program code of the DAS signal cross-scenario multi-category classification method.

[0135] Example 12 A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of a DAS signal cross-scene multi-category classification method.

[0136] The computer-readable storage medium stores an interface display program, and the interface display program can be executed by at least one processor so that the at least one processor performs the steps of the above-mentioned DAS signal cross-scene multi-category classification method.

[0137] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server or network device, etc.) to execute the DAS signal cross-scenario multi-category classification method described in the embodiment of the present application.

Claims

1. A DAS signal cross-scenario multi-category classification method, characterized in that: The following steps are involved: Step S1, obtaining signal sample data; Collect various types of vibration source signal samples in various scenarios, and process the vibration source signal samples to obtain the corresponding time-frequency matrix; label the vibration source signal samples to obtain label data; Step S2, constructing a signal recognition and classification model; Construct a signal recognition and classification model, which includes a feature extraction network and a recognition and classification network; the feature extraction network extracts features from the vibration source signal and outputs a feature vector; the recognition and classification network includes a constructed tree classifier, each non-leaf node of the tree classifier includes a node classification subnetwork and an output layer, the node classification subnetwork includes a first one-dimensional convolution layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional maximum pooling layer, a second one-dimensional convolution layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional maximum pooling layer, and the output layer includes a conversion layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a Softmax layer, which are arranged in sequence; Step S3, training a signal recognition classification model; The signal recognition and classification model constructed in step S2 is trained using the time-frequency matrix and label data obtained in step S1 to obtain a final signal recognition and classification model; Step S4, real-time classification of signals; The vibration source signal to be classified is obtained and processed into a time-frequency matrix, and is input into the signal recognition and classification model obtained in step S3, and the signal recognition and classification model outputs the signal category.

2. A DAS signal cross-scenario multi-category classification method as claimed in claim 1, characterized in that: In step S1, when using vibration source signal samples, the buried optical fiber is connected to the optical fiber vibration sensing system to collect various types of vibration source signals in various scenarios; The optical fiber vibration sensing system includes a narrow linewidth laser, an acousto-optic modulator, an erbium-doped fiber amplifier, a circulator, a phase demodulation module, an interferometer, a coupler, a photodetector and an analog-to-digital converter; the continuous light emitted by the narrow linewidth laser is modulated into pulse light by the acousto-optic modulator, the erbium-doped fiber amplifier amplifies the pulse light and transmits it into the optical fiber through the circulator, the pulse light generates backward Rayleigh scattered light when propagating in the optical fiber, the phase demodulation module receives the Rayleigh scattered light and filters it, and then linearly demodulates the backward Rayleigh scattered light through the interferometer and the coupler in turn to obtain a demodulated signal, the demodulated signal is received by the photodetector and converted into a digital electrical signal through the analog-to-digital converter.

3. A DAS signal cross-scenario multi-category classification method as claimed in claim 1, characterized in that: In step S2, the feature extraction network includes three LSTM layers and one fully connected layer; the time-frequency matrix is ​​used as the input of the feature extraction network, and the feature extraction network outputs a feature vector, which is used as the input of the recognition and classification network.

4. A DAS signal cross-scenario multi-category classification method as claimed in claim 1, characterized in that: In step S3, when training the signal recognition classification model, the total loss function is It is expressed as: in, represents the total loss of feature extraction, represents the total loss of tree classification, Indicates The first The loss of data, Indicates The first samples, Representation sample With The similarity of the centroids of class events, Representation sample With The similarity of the centroids of class events, Indicates the total number of events. Indicates The loss weight of the node classifier is Indicates The loss of the node classifier is Represents the features extracted by the feature extraction network, represents the total number of node classifiers, Represents the one-hot encoding of the labels corresponding to each atomic object class at the current node, represents the predicted probability value, Indicates The number of child nodes of non-leaf nodes, represents the zth one-hot encoding value, Represents the probability value of belonging to the zth prediction.

5. A DAS signal cross-scenario multi-category classification method as claimed in claim 4, characterized in that: Similarity The calculation method is: parameter , Use the standard gradient descent algorithm for iteration and update. The specific algorithm is: in, , are all trainable parameters. represents the event centroid; Indicates that samples are removed when calculating the event centroid , making training more stable; represents the number of samples, represents the qth data of the kth type of event, is the learning rate.

6. A DAS signal cross-scenario multi-category classification method as claimed in claim 1, characterized in that: In step S2, the steps of constructing the tree classifier are: Step S2-1, calculating the average feature representation of each event type; Step S2-2, adding the average feature representation of all event types to the root node; Step S2-3, using spectral clustering to group the time types in the current root node, and treating each group as a child node of the current node; Step S2-4, recursively execute step S2-3 for each child node until the node contains only one event type, that is, a leaf node.

7. A DAS signal cross-scenario multi-category classification method as claimed in claim 1, characterized in that: When you need to add a new event category, the specific steps are as follows: Step 01: Input the vibration source signal samples corresponding to the newly added event categories into the trained signal recognition and classification model. The signal recognition and classification model assigns the categories of the newly added events to the best matching lowest level non-leaf nodes. ; Step 02: To the lowest non-leaf node of the best match of the tree classifier Insert a new leaf node, and update the tree classifier and the signal recognition classification model to obtain a new tree classifier and a new signal recognition classification model; The newly inserted leaf node is the newly added event type, and the best matching lowest level non-leaf node in the original tree classifier is Corresponding to the non-leaf nodes in the new tree classifier ; Step 03, training a new signal recognition and classification model; during training, first freeze the parameters of the feature extraction network in the new signal recognition and classification model, as well as the leaf nodes in the recognition and classification network. Then, the vibration source signal samples corresponding to the newly added event categories and the best matching lowest-level non-leaf nodes are used to classify the parameters of the node classification subnetwork in the non-leaf nodes other than The samples that originally corresponded to the samples in the new training set are used to train the non-leaf nodes in the new tree classifier. The node classification sub-network is trained.

8. A DAS signal cross-scenario multi-category classification system, characterized in that: include: The signal sample data acquisition module is used to collect various types of vibration source signal samples in various scenarios, and process the vibration source signal samples to obtain the corresponding time-frequency matrix; the vibration source signal samples are labeled to obtain label data; A signal recognition and classification model construction module is used to construct a signal recognition and classification model, which includes a feature extraction network and a recognition and classification network; the feature extraction network extracts features from the vibration source signal and outputs a feature vector; the recognition and classification network includes a constructed tree classifier, each non-leaf node of the tree classifier includes a node classification subnetwork and an output layer, the node classification subnetwork includes a first one-dimensional convolution layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional maximum pooling layer, a second one-dimensional convolution layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional maximum pooling layer, and the output layer includes a conversion layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a Softmax layer, which are arranged in sequence; The signal recognition and classification model training module is used to train the signal recognition and classification model constructed by the signal recognition and classification model construction module using the time-frequency matrix and label data obtained by the signal sample data acquisition module to obtain the final signal recognition and classification model; The signal real-time classification module is used to obtain the vibration source signal to be classified and process it into a time-frequency matrix, and input the signal recognition and classification model obtained by the signal recognition and classification model training module, and the signal recognition and classification model outputs the signal category.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images