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

Abstract

The application discloses a DAS signal cross-scene multi-category classification method, system, device and medium, belongs to signal recognition classification in the field of optical fiber sensing technology, and aims to solve the technical problems that the DAS signal can only recognize single scene events in the prior art, and the DAS signal cannot be recognized and classified in a complex scene. The application comprises the following steps: acquiring samples and labels; constructing a signal recognition classification model comprising a feature extraction network and an identification 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, the node classification sub-network comprises a one-dimensional convolution layer, a batch normalization layer, a ReLU layer, a one-dimensional maximum pooling layer, a one-dimensional convolution layer, a batch normalization layer, a ReLU layer and a one-dimensional maximum pooling layer, and the output layer comprises a transformation layer, a full connection layer, a ReLU layer, a full connection layer, a ReLU layer and a Softmax layer arranged in sequence; training the signal recognition classification model; and classifying signals in real time.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical fiber sensing, and relates to a method for identifying and classifying signals of a distributed acoustic sensor (DAS) of an optical fiber, in particular to a DAS signal cross-scene multi-class classification method, system, device and medium. BACKGROUND

[0002] A distributed acoustic sensor (DAS) based on linear phase demodulation phase-sensitive optical time domain reflection (Φ-OTDR) technology is a new type of sensing technology that can realize continuous distributed detection of vibration and acoustic field, and is also a high-sensitivity, high-performance, long-sensing-distance, wide-sensing-range and strong anti-electromagnetic interference acoustic or vibration sensing method. Due to these unique advantages, it can be widely used in seismic exploration, oil / gas pipeline monitoring, natural disaster prediction, communication / power / railway cable and perimeter security, etc. by using the widely laid underground alarm cables to achieve high-precision sensing and positioning of various seismic sources.

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

[0004] The application patent application with the application number 202410159856.7 discloses a DAS signal classification method based on improved wavelet threshold denoising, which includes the following steps: wavelet decomposition is performed on the DAS signal Q to obtain the detail coefficients cD{cD1, cD2, …, cDM} and the approximation coefficients cA{cA1, cA2, …, cAM}; the threshold value is determined in combination with the Akaike Information Criterion (AIC), a new threshold function is used to perform threshold quantization processing on the detail coefficients cD after wavelet decomposition to obtain the processed detail coefficients cD'{cD1', cD2', …, cDM'}; cDm' represents the mth layer of detail coefficients after threshold quantization processing; the processed detail coefficients CD' and the approximation coefficients cAM of the last layer of wavelet decomposition are used to perform wavelet inverse transform to complete signal reconstruction, and the wavelet denoised signal Q' is obtained; the time domain features of the wavelet denoised signal Q' are extracted, the extracted time domain features are used to construct a data set, and the data set is divided into a training set and a test set; the WKNN algorithm is used to classify the data in the test set. The classification method uses wavelet denoising preprocessing, then extracts signal features to construct a data set, and uses WKNN to classify the DAS signal, thereby improving the classification accuracy.

[0005] The application number 202410135557.X discloses a kind of intelligent DAS and its aliasing signal directional target separation method, it first connects buried optical fiber access optical fiber vibration sensing system and collects multiple types of vibration source signals;Single source signal database is constructed, and then mixed signal data set is constructed by single source data set;Based on the collected signal, target signal directional separation network is constructed;Network is composed of two parts: embedding vector (d-vector) extraction network and target signal separation network;Embedding vector extraction network generates embedding vector for reference signal, and target signal separation network uses embedding vector to estimate target signal in mixed signal;Embedding vector extraction network includes three layers of LSTM layer and a layer of fully connected layer;Target signal separation network includes three layers of LSTM and two layers of fully connected layer, 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.The separation method can realize the separation of target signal using single aliasing signal, solve the problem that target signal detection and identification are difficult due to multi-source interference of optical fiber sensing system in complex buried environment, and have great potential application value.

[0006] In recent years, with the rapid development of artificial intelligence technology, more and more deep learning technologies are applied to DAS. From basic one-dimensional and two-dimensional convolutional neural network (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 combined model for simultaneously extracting DAS hidden space-time information, and transfer learning, semi-supervised and unsupervised methods, the detection and recognition ability of DAS for events is effectively improved. However, most of the current DAS signal recognition and classification technologies work effectively in a single scene, and the generalization ability is poor for cross-scene multi-class classification tasks in different scenes, and the recognition and classification effect of multiple signal categories is poor. Moreover, when facing different scenes, the events in multiple scenes cannot be effectively classified in one network. SUMMARY

[0007] The present application aims to solve the technical problems that the existing DAS signal can only recognize single scene events, and cannot recognize and classify cross-scene multi-class DAS signals in complex scenes. A DAS signal cross-scene multi-class classification method, system, device and medium are provided, which adds a constructed tree classifier to the constructed recognition and classification model to realize effective recognition and classification of multi-class events in multiple scenes in a single network.

[0008] The application specifically adopts the following technical scheme to achieve the above-mentioned purpose.

[0009] A DAS signal cross-scene multi-class classification method, comprising the following steps:

[0010] Step S1, acquiring signal sample data;

[0011] Collecting vibration source signal samples of multiple types in multiple scenes, and processing the vibration source signal samples to obtain corresponding time-frequency matrices; labeling the vibration source signal samples to obtain label data;

[0012] Step S2, constructing a signal recognition classification model;

[0013] The signal recognition classification model includes a feature extraction network and a recognition classification network; the feature extraction network extracts features from the vibration source signal and outputs a feature vector; the recognition classification network includes a tree-shaped classifier, and each non-leaf node of the tree-shaped classifier includes a node classification sub-network and an output layer; the node classification sub-network 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 arranged in sequence; and the output layer includes a transformation layer, a first full connection layer, a third ReLU layer, a second full connection layer, a fourth ReLU layer, and a Softmax layer arranged in sequence;

[0014] Step S3, training the signal recognition classification model;

[0015] The time-frequency matrix and the label data obtained in step S1 are used to train the signal recognition classification model constructed in step S2, so as to obtain a final signal recognition classification model;

[0016] Step S4, real-time signal classification;

[0017] The vibration source signal to be classified is processed into a time-frequency matrix and input into the signal recognition classification model obtained in step S3, and the signal recognition classification model outputs the signal class.

[0018] Further, in step S1, the buried optical fiber is connected to an optical fiber vibration sensing system to collect vibration source signal samples of multiple types in multiple scenes.

[0019] The fiber optic 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. Continuous light emitted by the narrow-linewidth laser is modulated into pulsed light by the acousto-optic modulator. The erbium-doped fiber amplifier amplifies the pulsed light and transmits it into the optical fiber through the circulator. When the pulsed light propagates in the optical fiber, it generates backscattered Rayleigh light. The phase demodulation module receives and filters the Rayleigh scattered light, and then linearly demodulates the backscattered light through the interferometer and coupler 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.

[0020] 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 to the feature extraction network, the feature extraction network outputs a feature vector, and the feature vector serves as the input to the recognition and classification network.

[0021] Furthermore, in step S3, when training the signal recognition and classification model, the overall loss function... Represented as:

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] in, This represents the total loss from feature extraction. This represents the total loss for tree-based classification. Indicates the first The first type of event The loss of data points Indicates the first The first type of event One sample, Indicates sample With the Similarity of centroids of similar events Indicates sample With the Similarity of centroids of similar events Indicates the total number of event categories. Indicates the first The loss weights of each node classifier Indicates the first The loss of a classifier with individual nodes This represents the features extracted by the feature extraction network. This represents the total number of node classifiers. This represents the one-hot encoding of the labels corresponding to the atomic object classes at the current node. This represents the predicted probability value. Indicates the first The number of child nodes of a non-leaf node. This represents the z-th one-hot encoded value. This represents the probability value of belonging to the z-th prediction.

[0028] Furthermore, similarity The calculation method is as follows:

[0029]

[0030]

[0031]

[0032] parameter , The standard gradient descent algorithm is used for iteration and updating. The specific algorithm is as follows:

[0033]

[0034]

[0035] in, , All of these are trainable parameters. Indicates the center of gravity of the event; This indicates removing samples when calculating the event centroid. This makes training more stable; Indicates the number of samples. This represents the q-th data point of the k-th type of event. is the learning rate. During training, the Adam optimization algorithm is used to iteratively update the parameters of the classification subnetwork at each node. .

[0036] Furthermore, in step S2, the steps for constructing the tree classifier are as follows:

[0037] Step S2-1: Calculate the average feature representation for each event type;

[0038] Step S2-2: Add the average feature representation of all event types to the root node;

[0039] Step S2-3: Use spectral clustering to group the time types in the current root node, and treat each group as a child node of the current node;

[0040] Step S2-4: Recursively execute step S2-3 for each child node until the node contains only one event type, i.e., a leaf node.

[0041] Furthermore, when it is necessary to add a new event category, the specific steps are as follows:

[0042] Step 01: Input the vibration source signal samples corresponding to the newly added event category into the trained signal recognition and classification model. The signal recognition and classification model will assign the category of the newly added event to the lowest-level non-leaf node with the best match. ;

[0043] Step 02, feed the best-matched non-leaf node of the tree classifier. Insert a new leaf node and update the tree classifier and signal recognition classification model to obtain a new tree classifier and a new signal recognition classification model;

[0044] The newly inserted leaf node represents the new event type, and is the lowest-level non-leaf node that represents the best match in the original tree classifier. Corresponding to the non-leaf nodes in the new tree classifier ;

[0045] Step 03: Train the new signal recognition and classification model. During training, first freeze the parameters of the feature extraction network and the non-leaf nodes in the recognition and classification network of the new signal recognition and classification model. The parameters of the node classification subnetwork in the non-leaf nodes other than the newly added event categories are then used to classify the vibration source signal samples corresponding to the newly added event categories and the best-matched lowest-level non-leaf nodes. The original samples were used to form a new training set, and finally, the new training set was used to classify the non-leaf nodes in the new tree classifier. The node classification subnetwork is trained.

[0046] A multi-class classification system for DAS signals across scenarios includes:

[0047] The signal sample data acquisition module is used to collect various types of vibration source signal samples in various scenarios, process the vibration source signal samples to obtain the corresponding time-frequency matrix, and label the vibration source signal samples to obtain label data.

[0048] The signal recognition and classification model construction module is used to build 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 pre-constructed tree classifier. Each non-leaf node of the tree classifier includes a node classification sub-network and an output layer. The node classification sub-network includes, in sequence, a first one-dimensional convolutional layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional max pooling layer, a second one-dimensional convolutional layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional max pooling layer. The output layer includes, in sequence, a transformation layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a softmax layer.

[0049] 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, so as to obtain the final signal recognition and classification model.

[0050] The real-time signal classification module is used to acquire the vibration source signal to be classified, process it into a time-frequency matrix, and input it into the signal recognition and classification model obtained by the signal recognition and classification model training module. The signal recognition and classification model outputs the signal category.

[0051] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method described above.

[0052] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method.

[0053] The beneficial effects of this invention are as follows:

[0054] 1. This invention proposes for the first time a cross-scene multi-category recognition method for fiber optic distributed acoustic sensors (DAS) based on deep learning. It achieves accurate recognition of multiple (more than 10) categories of events in multiple (three or more) different scenarios in a single network model. This solves the problem that traditional recognition networks can only recognize DAS signals of single scene events and cannot recognize and classify DAS signals of multiple categories across scenarios in complex scenarios, and improves the accuracy of recognition.

[0055] 2. In this invention, based on the differences in time-frequency characteristics of different types of vibration sources, the feature extraction of the target signal is achieved through encoder-LSTM (feature extraction network). Compared with traditional convolutional neural networks, this method has smaller model parameters, faster convergence speed, and can effectively capture the temporal information in the signal, making each feature vector as close as possible to its center point and as far away from all other center points, thereby obtaining effective feature vectors, which facilitates the differentiation of different event types.

[0056] 3. In this 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, improve the recognition accuracy, especially when facing event recognition tasks with high similarity, it greatly improves the recognition accuracy, and the network structure is easy to adjust and has strong stability.

[0057] 4. In this 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 reduced exponentially compared with the traditional method of retraining on a new dataset. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the method of the present invention;

[0059] Figure 2 This is a schematic diagram of the fiber optic vibration sensing system in this invention;

[0060] Figure 3 This is a flowchart illustrating the feature extraction process in this invention.

[0061] Figure 4 This is a schematic diagram of the feature extraction network structure in this invention;

[0062] Figure 5 This is a schematic diagram of the feature tree constructed based on nine types of events in three scenarios in this invention;

[0063] Figure 6 This is a schematic diagram of the structure of the identification and classification network in this invention;

[0064] Figure 7 These are the scene diagrams, time signals, and spectrum diagrams of ten types of events in three scenarios collected in this invention;

[0065] Figure 8 This is a visual comparison of the feature extraction performance of the encoder-LSTM of this invention with traditional methods;

[0066] Among them, a-encoder-LSTM, b-one-dimensional convolutional neural network, and c-two-dimensional convolutional neural network;

[0067] Figure 9 This is a confusion matrix of the test results of three methods in the field test of this invention: encoder-LSTM + 1-D CNN Tree-Classifier, encoder-LSTM + SVM Tree-Classifier, and encoder-LSTM + Softmax.

[0068] Among them, a-encoder-LSTM + 1-D CNN Tree-Classifier, b-encoder-LSTM + SVMTree-Classifier, c-encoder-LSTM + Softmax;

[0069] Figure 10 The results of five-fold cross-validation of three methods—encoder-LSTM + 1-D CNN Tree-Classifier, encoder-LSTM + SVM Tree-Classifier, and encoder-LSTM + Softmax—during field testing of this invention.

[0070] Figure 11 This is a comparison chart of the test results of the incremental learning method proposed in this invention and the traditional method of retraining the entire dataset in the field test of this invention, including training time, testing time and recognition accuracy. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0072] Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0073] Example 1

[0074] In smart city monitoring applications, vibration signals detected by fiber optic sensing systems typically encompass multiple typical event types across various scenarios. Examples include manual excavation, mechanical excavation, and traffic interference in urban security scenarios; pipeline leak 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. Based on a tree classifier, it achieves accurate identification of multiple categories of events across these scenarios within a single network, adding incremental learning functionality to the classifier. Figure 1 As shown, it mainly includes the following steps:

[0075] Step S1: Acquire signal sample data;

[0076] Multiple types of vibration source signal samples were collected from various scenarios, and the vibration source signal samples were processed to obtain the corresponding time-frequency matrix; the categories of the vibration source signal samples were labeled to obtain label data.

[0077] By connecting buried optical fibers to an optical fiber vibration sensing system, various types of vibration source signals in multiple scenarios can be collected, and a signal database for different types of events can be constructed.

[0078] Step S2: Construct a signal recognition and classification model;

[0079] 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 pre-constructed tree classifier. Each non-leaf node of the tree classifier includes a node classification sub-network and an output layer. The node classification sub-network includes, in sequence, a first one-dimensional convolutional layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional max pooling layer, a second one-dimensional convolutional layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional max pooling layer. The output layer includes, in sequence, a transformation layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a softmax layer.

[0080] Step S3: Train the signal recognition and classification model;

[0081] Using the time-frequency matrix and label data obtained in step S1, the signal recognition and classification model constructed in step S2 is trained to obtain the final signal recognition and classification model.

[0082] A feature extraction network is constructed, taking the time-frequency matrix of the event signal as input and calculating the loss function by calculating the similarity matrix of the output feature vectors. The network parameters are iteratively updated to complete the network training. A tree classifier is constructed based on spectral clustering, and node classifiers are added. Then, the tree classifier is trained using a multi-task learning method, and the optimal model is used to identify the test event.

[0083] Step S4: Real-time signal classification;

[0084] The vibration source signal to be classified is acquired and processed into a time-frequency matrix, and then input into the signal recognition and classification model obtained in step S3. The signal recognition and classification model outputs the signal category.

[0085] This embodiment proposes for the first time a cross-scene multi-category recognition method for fiber optic distributed acoustic sensors (DAS) based on deep learning. It achieves accurate recognition of multiple (more than 10) categories of events in multiple (three or more) different scenarios 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 DAS signals of multiple categories across scenarios in complex scenarios, and improves the recognition accuracy.

[0086] Example 2

[0087] This embodiment further illustrates the fiber optic vibration sensing system in step 1 of Embodiment 1.

[0088] In Example 1, a distributed optical fiber acoustic sensing (DAS) system based on phase-sensitive optical time-domain reflectometry is used to collect vibration signals from different types of events. It utilizes backscattering Rayleigh scattering to monitor changes in the state of the optical fiber. Rayleigh scattering is an elastic light scattering originating from fluctuations in the medium density. When a vibration event occurs at a certain location on the optical fiber, the relative position of the Rayleigh scattering center changes, thereby altering the phase and enabling the sensing of the vibration event.

[0089] The structure of the fiber optic vibration sensing system (a 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 as follows: continuous light emitted by the narrow-linewidth laser is modulated into pulsed light with a specific period and pulse width by the acousto-optic modulator. The erbium-doped fiber amplifier amplifies the pulsed light and transmits it into the optical fiber through the circulator. Backscattered Rayleigh light is generated as the pulsed light propagates in the fiber. The phase demodulation module receives and filters the Rayleigh scattered light, and then linearly demodulates it sequentially through the interferometer and coupler 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.

[0090] During each pulse transmission time, the DAS forms an OTDR trace that varies with the spatial position of the fiber. By accumulating the OTDR traces received over time, a two-dimensional time-frequency matrix is ​​formed for subsequent analysis and processing.

[0091] Example 3

[0092] This embodiment is based on Embodiment 2, and further optimizes the process of preparing and obtaining sample data.

[0093] To verify the performance of the DAS signal cross-scenario multi-category classification method proposed in this embodiment, this embodiment collected ten types of events in three different scenarios for field testing.

[0094] The five types of events described below are collected by the DAS system of underground communication cables in the two cities of Tongren and Wuhan. For example... Figure 7 As shown in (a), five types of events were collected, including 1: mechanical excavation; 2: mechanical tamping; 3: manual excavation; 4: background noise; and 5: traffic interference.

[0095] The data in Scenario 2 consists of actual pipeline vibration acoustic signals collected in Chongqing during leaks and leak repairs. For example... Figure 7 As shown in (b), two types of events were collected, including 6: pipeline leak; 7: pipeline not leaking.

[0096] The data in Scenario 3 was collected by the DAS system for monitoring rail defects in Zhangjiakou City, Hebei Province. For example... Figure 7 As shown in (c), a total of three events were collected, including 8: No train passed; 9: Train passed; and a new category of event: 10: Train passed (with track defects).

[0097] To enrich the data, this embodiment performed simple data augmentation on all ten data categories, i.e., slicing the original data into segments of 2000 sampling points each. The constructed dataset is shown in Table 1. The training, validation, and test data were randomly divided in a 6:2:2 ratio, with no overlap among the three.

[0098] Table 1 Typical event databases in three scenarios

[0099]

[0100]

[0101] Example 4

[0102] This embodiment further illustrates the feature extraction network in step 2 of embodiment 1.

[0103] LSTM is more effective for mining and representing the features of long-term sequence signals. Therefore, in this embodiment, the feature extraction network uses an encoder-LSTM network to extract feature vectors from the DAS signal, making each feature vector as close as possible to its center point and as far away from all other center points as possible, thereby obtaining effective feature vectors that facilitate the differentiation of different event types. The basic structure of the feature extraction network is as follows: Figure 3 As shown, it consists of 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, the feature extraction network outputs a feature vector, and the feature vector is used as the input of the recognition and classification network.

[0104] Table 2. Structural parameters of the feature extraction network (encoder-LSTM network)

[0105]

[0106] Example 5

[0107] This embodiment further illustrates the training of the signal recognition and classification model in step 3 of embodiment 1.

[0108] When training a signal recognition and classification model, the overall loss function is... Represented as:

[0109]

[0110] When training a feature extraction network, the total feature extraction loss of the feature extraction network is... for:

[0111]

[0112] Loss function for each sample signal in the training batch for:

[0113]

[0114] in, Indicates the first The first type of event One sample, Indicates sample With the The similarity of the centroids of similar events. The calculation method is as follows:

[0115]

[0116] The centroid of an event is calculated using the following formula:

[0117]

[0118]

[0119] in, This indicates removing samples when calculating the event centroid. This makes training more stable.

[0120] Then, the standard gradient descent algorithm is used to iterate and update the network parameters. , The specific algorithm is as follows:

[0121]

[0122]

[0123] When training a recognition and classification network, the total tree-like classification loss of the network is... for:

[0124]

[0125] This embodiment uses the same cross-entropy loss function to learn all node classifiers:

[0126]

[0127] Furthermore, this embodiment also uses the Adam optimization algorithm to iteratively update the parameters of the classification subnetwork at each node. The specific algorithm is as follows:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133] The Adam algorithm combines the Momentum algorithm and the RMSProp algorithm, which solves the problem that stochastic gradient descent may not be able to reach the global optimum, and at the same time makes the network converge faster.

[0134] in, This represents the total loss from feature extraction. This represents the total loss for tree-based classification. Indicates the first The first type of event The loss of data points Indicates the first The first type of event One sample, Indicates sample With the Similarity of centroids of similar events Indicates sample With the Similarity of centroids of similar events Indicates the total number of event categories. Indicates the first The loss weights of each node classifier Indicates the first The loss of a classifier with individual nodes This represents the features extracted by the feature extraction network. This represents the total number of node classifiers; This represents the one-hot encoding of the labels corresponding to the atomic object classes at the current node. ; This represents the predicted probability value. ; Indicates the first The maximum value of the number of child nodes of a non-leaf node is B; This represents the z-th one-hot encoded value. This represents the probability value of belonging to the z-th prediction. , All of these are trainable parameters. Indicates the center of gravity of the event; This indicates removing samples when calculating the event centroid. This makes training more stable; Indicates the number of samples. This represents the q-th data point of the k-th type of event. For learning rate, Represents gradient momentum The correction value, Represents the squared gradient The correction value is used to avoid a large difference between the moving exponential average and the initial value at the beginning of the iteration. Indicates the smoothing term, set to ; This represents the gradient momentum accumulated by the loss function during the first t-1 iterations, with an initial value of 0; This represents the squared gradient accumulated by the loss function during the first t-1 iterations, with an initial value of 0; This represents the exponential decay rate of the Momentum algorithm at iteration t. This represents the exponential decay rate of the RMSProp algorithm at the t-th iteration. This represents the exponential decay rate of the corresponding Momentum algorithm, set to 0.9; This represents the exponential decay rate of the corresponding RMSProp algorithm, set to 0.999; Indicates the first The loss weights of each node classifier Indicates the first The loss of a classifier with individual nodes This represents the features extracted by the feature extraction network [A1].

[0135] Example 6

[0136] This embodiment further illustrates the construction and training of the tree classifier in step 2 of embodiment 1.

[0137] First, the tree classifier structure is customized 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. It contains a set of nodes V, a set of features F, and a set of edges E. Each non-leaf node... It contains a set of atomic object classes L(c) and a corresponding set of feature subsets. For a given 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 completed feature tree. Before starting to build the feature tree, this embodiment needs to provide a parameter B, which represents the number of branches, 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, then B is set to the number of atomic object classes contained in that node.

[0138] The process of building a tree classifier is as follows:

[0139] 1) Calculate the average feature representation of N event types;

[0140] For a given event category, this embodiment chooses to use the average characteristic of the event to represent its main characteristics. Let... For the first The set of event characteristics, in this embodiment, is calculated. The sum of features extracted from all events of this type using encoder-LSTM is accumulated, and then the average is calculated as the overall feature representation (i.e., average feature representation) of this type of event. This embodiment will use the first... The overall characteristic representation (i.e., the average characteristic representation) of a class of events is defined as follows: :

[0141]

[0142] in, Indicates the number of events in the set. Represents a set The first in A sample of class.

[0143] 2) Add the characteristic representations of all event types to the root node;

[0144] 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;

[0145] For a given first For each non-leaf node, this embodiment obtains its corresponding feature subset, and then uses the Gaussian kernel function to calculate the feature similarity between the two types of events:

[0146]

[0147] in, It refers to bandwidth, the radial range used for control.

[0148] By calculating the pairwise similarity between the M atomic object classes in this node, we obtain the M*M similarity matrix at this node. .

[0149] Then, the Laplacian matrix is ​​calculated and its eigenvalues ​​are decomposed to obtain a low-dimensional representation of the data, thereby increasing the cluster spacing and reducing computation time.

[0150]

[0151] in, It is a degree matrix;

[0152]

[0153] right Perform eigenvalue decomposition, sort the eigenvalues ​​from smallest to largest, and take the top [values]. The eigenvectors corresponding to the eigenvalues ​​form a matrix U, let For U's first row vectors, then .

[0154] Finally, the K-means algorithm is used for clustering to obtain the set of child nodes of the current node.

[0155] 4) Recursively perform the third operation on each child node until the node contains only one event type, i.e., a leaf node.

[0156] Figure 5 The results show the tree structure for nine types of events in three different scenarios.

[0157] After the tree-structured 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. A multi-task structure learning method is proposed to jointly train the various node classifiers, reducing computational cost while improving their discriminative ability. Before training begins, the data labels need to be reinitialized to adapt to the discrimination criteria for the same event reaching different nodes, such as... Figure 5 As shown, with B set to 2, in the root node, the labels for events 1, 2, and 3 can be defined as 1 in this embodiment, and the labels for events 4, 5, 6, 7, 8, and 9 can be defined as 2 in this embodiment. In the first non-leaf node of the next layer, the labels for events 1 and 3 need to be defined as 1, and the label for event 2 needs to be defined as 2. This process continues to obtain the labels for a class of events at different nodes, which is necessary for training the node classifier.

[0158] In this embodiment, the encoder-LSTM network is considered a shared part, and each node classifier has its own independent network structure to extract the features it needs. The node classifier uses a relatively simple 1-D CNN structure to accelerate computation. The node classifier network structure is as follows: Figure 6 As shown, in this structure, this embodiment first uses 32 one-dimensional convolutional kernels of size 4 to extract the initial features processed by the encoder-LSTM network. In the second convolutional layer, one-dimensional convolutional kernels of the same size are used to extract additional features, and the number of channels is increased to 64 to extract deeper features. Finally, the features are input to the output layer, which consists of two fully connected layers. Finally, a softmax layer is used to convert the obtained feature vector into a corresponding probability vector. Based on the maximum probability in the probability vector, it is determined which node in the next layer should be used for this type of event.

[0159] Example 7

[0160] This embodiment further illustrates the feature extraction network trained in step 3 of embodiment 1.

[0161] In 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. To verify the effectiveness of the feature extraction network (encoder-LSTM network) in extracting features from DAS signals, this embodiment compares the proposed encoder-LSTM network model with the 1-D CNN and 2-D CNN models widely used in the field of signal recognition.

[0162] After training the three networks, features of nine core events were extracted and classified using these three methods to verify the effectiveness of the algorithms. This embodiment utilizes Principal Component Analysis (PCA) to reduce the feature dimension to three dimensions, calculates the class centers of each type of event feature and the average Euclidean distance between class centers, and visualizes the features to observe the feature extraction effect. Figure 8 As shown in Figure 8, compared with 1-D CNN and 2-D CNN, the features extracted by encoder-LSTM can be well separated for different events, and the features of the same type of events can be well integrated. In contrast, the features extracted by 1-D CNN and 2-D CNN are relatively scattered, and the features of different events are prone to overlap. Therefore, encoder-LSTM can extract features of different types of events more effectively, providing a good guarantee for the learning of the subsequent tree classifier.

[0163] Example 8

[0164] This embodiment further illustrates the recognition and classification network trained in step 3 of embodiment 1.

[0165] The proposed encoder-LSTM + 1-D CNN Tree-Classifier model is compared with the encoder-LSTM + SVM Tree-Classifier model, which uses a tree-structured SVM classifier in the classification part, and the traditional encoder-LSTM + Softmax model. The original datasets for all three models are the field experimental data shown in Table 2. Experimental parameters were set as follows: batch-size = 4, BATCH_SIZE = 4, LR (learning rate) = 0.001, and andepoch = 200.

[0166] In this embodiment, the three models described above are trained and tested using test set data to obtain the confusion matrix, as shown below. Figure 9 As shown, the results indicate that the encoder-LSTM + 1-D CNN Tree-Classifier proposed in this invention has the best recognition performance, followed by encoder-LSTM + Softmax and encoder-LSTM + SVM Tree-Classifier.

[0167] To verify the effectiveness and generalization ability of the algorithm proposed in this invention, this embodiment employs a five-fold cross-validation method to validate three models. The validation results are as follows: Figure 10As shown in the figure. The results show that the encoder-LSTM + 1-D CNNTree-Classifier has the highest recognition accuracy and the most stable performance, with an average accuracy of 96.41%, which is 1.22% and 1.06% higher than the accuracy of the other two models, respectively. This indicates that the generalization ability of the method proposed in this invention is better than the other two models.

[0168] Example 9

[0169] This embodiment further explains the need to add event categories to the tree classifier in Embodiment 1.

[0170] When you need to add a new event category, the specific steps are as follows:

[0171] Step 01: Input the vibration source signal samples corresponding to the newly added event category into the trained signal recognition and classification model. The signal recognition and classification model will assign the category of the newly added event to the lowest-level non-leaf node with the best match. ;

[0172] Step 02, feed the best-matched non-leaf node of the tree classifier. Insert a new leaf node and update the tree classifier and signal recognition classification model to obtain a new tree classifier and a new signal recognition classification model;

[0173] The newly inserted leaf node represents the new event type, and is the lowest-level non-leaf node that represents the best match in the original tree classifier. Corresponding to the non-leaf nodes in the new tree classifier ;

[0174] Step 03: Train the new signal recognition and classification model. During training, first freeze the parameters of the feature extraction network and the non-leaf nodes in the recognition and classification network of the new signal recognition and classification model. The parameters of the node classification subnetwork in the non-leaf nodes other than the newly added event categories are then used to classify the vibration source signal samples corresponding to the newly added event categories and the best-matched lowest-level non-leaf nodes. The original samples were used to form a new training set, and finally, the new training set was used to classify the non-leaf nodes in the new tree classifier. The node classification subnetwork is trained.

[0175] Specifically:

[0176] For a new set of event data, this embodiment designs an incremental learning method to solve its classification problem. Before training begins, all data of the new events are sequentially passed through the pre-trained encoder-LSTM and tree classifier network, reaching the best-matching non-leaf node c in the penultimate layer of the tree classifier. The atomic object class that originally existed in this node has the greatest similarity to the new event type.

[0177]

[0178] in and It is the parameter of all parent nodes above node c.

[0179] Therefore, the tree classifier in this embodiment can identify the best-matching non-leaf node for a new event type. Then, a new leaf node is inserted into the non-leaf node to represent the new event type, and the node classifier network of the non-leaf node c is reinitialized, with the output vector length of the output layer of the node classifier changed from the original B to B+1.

[0180]

[0181] in, and These are the new parameters for the classifier at node c. It is the predicted probability vector.

[0182] The parameters of the encoder-LSTM network and the classifiers at other non-leaf nodes are frozen. The classifier at node c is trained using the dataset of the newly added event types and the dataset of the event types originally contained in the non-leaf node c. The label for the newly added event types is M+1, while the label at this node classifier is redefined as B+1 in this embodiment. The loss function of this node classifier is defined as... :

[0183]

[0184] in, This represents the loss function during incremental training. This represents samples of newly added event types and samples of event types originally contained in non-leaf node c. This represents the one-hot label of all events at the non-leaf node, including the newly added event type. The parameters of the node classifier network are also iteratively updated using the Adam optimization algorithm.

[0185] In this embodiment, event 10 is used as an incremental event for training and testing, and two methods are compared: (1) the incremental learning method proposed in this invention; (2) the traditional retraining of the joint training set (the original training set plus the new event dataset). Figure 11 The comparison of accuracy, training time, and testing time between incremental learning and joint training methods is shown. It can be seen that the incremental learning method proposed in this embodiment can significantly reduce computation time costs while ensuring accuracy.

[0186] 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 recognizing new events. It has obvious advantages in online real-time processing and is suitable for practical applications.

[0187] Example 10

[0188] This embodiment provides a DAS signal cross-scene multi-category classification system, including:

[0189] The signal sample data acquisition module is used to collect various types of vibration source signal samples in various scenarios, process the vibration source signal samples to obtain the corresponding time-frequency matrix, and label the vibration source signal samples to obtain label data.

[0190] The signal recognition and classification model construction module is used to build 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 pre-constructed tree classifier. Each non-leaf node of the tree classifier includes a node classification sub-network and an output layer. The node classification sub-network includes, in sequence, a first one-dimensional convolutional layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional max pooling layer, a second one-dimensional convolutional layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional max pooling layer. The output layer includes, in sequence, a transformation layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a softmax layer.

[0191] 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, so as to obtain the final signal recognition and classification model.

[0192] The real-time signal classification module is used to acquire the vibration source signal to be classified, process it into a time-frequency matrix, and input it into the signal recognition and classification model obtained by the signal recognition and classification model training module. The signal recognition and classification model outputs the signal category.

[0193] Example 11

[0194] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform steps of a DAS signal cross-scene multi-class classification method.

[0195] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0196] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), 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 the 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, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device. Of course, the memory may include both internal storage units and external storage devices of the computer 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 DAS signal cross-scene multi-category classification method. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0197] 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 typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, for example, to run program code for the DAS signal cross-scene multi-class classification method.

[0198] Example 12

[0199] A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform steps of a DAS signal cross-scene multi-class classification method.

[0200] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to cause the at least one processor to perform the steps of the DAS signal cross-scene multi-category classification method described above.

[0201] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they 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 this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the DAS signal cross-scenario multi-category classification method described in the embodiments of this application.

Claims

1. A method for cross-scene multi-category classification of DAS signals, characterized in that, Includes the following steps: Step S1: Acquire signal sample data; Collect various types of vibration source signal samples from multiple scenarios, process the vibration source signal samples to obtain the corresponding time-frequency matrix, and label the vibration source signal samples to obtain label data; Step S2: Construct 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 pre-constructed tree classifier. Each non-leaf node of the tree classifier includes a node classification sub-network and an output layer. The node classification sub-network includes, in sequence, a first one-dimensional convolutional layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional max pooling layer, a second one-dimensional convolutional layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional max pooling layer. The output layer includes, in sequence, a transformation layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a softmax layer. Step S3: Train the signal recognition and classification model; Using the time-frequency matrix and label data obtained in step S1, the signal recognition and classification model constructed in step S2 is trained to obtain the final signal recognition and classification model. Step S4: Real-time signal classification; The vibration source signal to be classified is acquired and processed into a time-frequency matrix, and then input into the signal recognition and classification model obtained in step S3. The signal recognition and classification model outputs the signal category. In step S3, when training the signal recognition and classification model, the total loss function is... Represented as: in, This represents the total loss from feature extraction. This represents the total loss for tree-based classification. Indicates the first The first type of event The loss of data points Indicates the first The first type of event One sample, Indicates sample With the Similarity of centroids of similar events Indicates sample With the Similarity of centroids of similar events Indicates the total number of event categories. Indicates the first The loss weights of each node classifier Indicates the first The loss of a classifier with individual nodes This represents the features extracted by the feature extraction network. This represents the total number of node classifiers. This represents the one-hot encoding of the labels corresponding to the atomic object classes at the current node. This represents the predicted probability value. Indicates the first The number of child nodes of a non-leaf node. This represents the z-th one-hot encoded value. This represents the probability value of belonging to the z-th prediction.

2. The method for cross-scene multi-category classification of DAS signals as described 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 under various scenarios. The fiber optic 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. Continuous light emitted by the narrow-linewidth laser is modulated into pulsed light by the acousto-optic modulator. The erbium-doped fiber amplifier amplifies the pulsed light and transmits it into the optical fiber through the circulator. When the pulsed light propagates in the optical fiber, it generates backscattered Rayleigh light. The phase demodulation module receives and filters the Rayleigh scattered light, and then linearly demodulates the backscattered light through the interferometer and coupler 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.

3. The method for cross-scene multi-category classification of DAS signals as described 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 serves as the input to the feature extraction network, the feature extraction network outputs a feature vector, and the feature vector serves as the input to the recognition and classification network.

4. The method for cross-scene multi-category classification of DAS signals as described in claim 1, characterized in that: Similarity The calculation method is as follows: parameter , The standard gradient descent algorithm is used for iteration and updating. The specific algorithm is as follows: in, , All of these are trainable parameters. Indicates the center of gravity of the event; This indicates removing samples when calculating the event centroid. This makes training more stable; Indicates the number of samples. This represents the q-th data point of the k-th type of event. This is the learning rate.

5. The method for cross-scene multi-category classification of DAS signals as described in claim 1, characterized in that: In step S2, the steps for constructing the tree classifier are as follows: Step S2-1: Calculate the average feature representation for each event type; Step S2-2: Add the average feature representation of all event types to the root node; Step S2-3: Use spectral clustering to group the time types in the current root node, and treat 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, i.e., a leaf node.

6. The method for cross-scene multi-category classification of DAS signals as described 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 category into the trained signal recognition and classification model. The signal recognition and classification model will assign the category of the newly added event to the lowest-level non-leaf node with the best match. ; Step 02: Find the lowest-level non-leaf nodes that best match the tree classifier. Insert a new leaf node and update the tree classifier and signal recognition classification model to obtain a new tree classifier and a new signal recognition classification model; The newly inserted leaf nodes represent the new event types, and the lowest-level non-leaf nodes that represent the best match in the original tree classifier. Corresponding to the non-leaf nodes in the new tree classifier ; Step 03: Train the new signal recognition and classification model. During training, first freeze the parameters of the feature extraction network and the non-leaf nodes in the recognition and classification network of the new signal recognition and classification model. The parameters of the node classification subnetwork in the non-leaf nodes other than the newly added event categories are then used to classify the vibration source signal samples corresponding to the newly added event categories and the best-matched lowest-level non-leaf nodes. The original samples were used to form a new training set, and finally, the new training set was used to classify the non-leaf nodes in the new tree classifier. The node classification subnetwork is trained.

7. A DAS signal cross-scene 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, process the vibration source signal samples to obtain the corresponding time-frequency matrix, and label the vibration source signal samples to obtain label data. The signal recognition and classification model construction module is used to build 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 pre-constructed tree classifier. Each non-leaf node of the tree classifier includes a node classification sub-network and an output layer. The node classification sub-network includes, in sequence, a first one-dimensional convolutional layer, a first batch normalization layer, a first ReLU layer, a first one-dimensional max pooling layer, a second one-dimensional convolutional layer, a second batch normalization layer, a second ReLU layer, and a second one-dimensional max pooling layer. The output layer includes, in sequence, a transformation layer, a first fully connected layer, a third ReLU layer, a second fully connected layer, a fourth ReLU layer, and a softmax layer. 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, so as to obtain the final signal recognition and classification model. The real-time signal classification module is used to acquire the vibration source signal to be classified, process it into a time-frequency matrix, and input it into the signal recognition and classification model obtained by the signal recognition and classification model training module. The signal recognition and classification model outputs the signal category. In the signal recognition and classification model training module, the total loss function during signal recognition and classification model training is... Represented as: in, This represents the total loss from feature extraction. This represents the total loss for tree-based classification. Indicates the first The first type of event The loss of data points Indicates the first The first type of event One sample, Indicates sample With the Similarity of centroids of similar events Indicates sample With the Similarity of centroids of similar events Indicates the total number of event categories. Indicates the first The loss weights of each node classifier Indicates the first The loss of a classifier with individual nodes This represents the features extracted by the feature extraction network. This represents the total number of node classifiers. This represents the one-hot encoding of the labels corresponding to the atomic object classes at the current node. This represents the predicted probability value. Indicates the first The number of child nodes of a non-leaf node. This represents the z-th one-hot encoded value. This represents the probability value of belonging to the z-th prediction.

8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The system stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

Images