A DOFS signal classification method based on DWT-RMG-TabPFN

Abstract

The application provides a signal classification method based on the following: collecting an original signal for a single event, calculating a differential signal of a current time slot; and segmenting, calculating a segmented differential signal corresponding to each segment; then, respectively decomposing the segmented differential signal, extracting corresponding low-frequency approximate coefficients and high-frequency detail coefficients, and further establishing a matrix corresponding to each segment; then, traversing the matrix, extracting the features of the current time slot; finally, inputting the features of various events, using a network to encode, obtaining an input sequence, and combining a mask matrix to input into a reasoning mechanism, obtaining a reasoning result, and further classifying and predicting a test event. The application uses a neural network combined with features to obtain a high classification accuracy.

Description

Technical Field

[0001] This invention relates to the fields of signal processing and machine learning, specifically to a DOFS (Distributed Fiber Optic Sensing) signal classification method based on DWT-RMG-TabPFN (Discrete Wavelet Transformation-Riemannian manifold geometry-TabularPrior-data Fitted Network). Background Technology

[0002] In existing technologies, DOFS signal classification is mainly achieved using traditional machine learning algorithms. The process is as follows: first, the signal is denoised and decomposed; then, features are extracted from the signal components; and finally, the extracted feature vectors are input into the machine learning model to complete the event classification.

[0003] Common methods for decomposing DOFS signals include Fourier decomposition and variational mode decomposition. However, Fourier decomposition is difficult to capture the local abrupt changes in the signal and has a fixed frequency resolution; variational mode decomposition has low adaptability to the signal.

[0004] Furthermore, as the monitoring environment becomes increasingly complex, signals are highly susceptible to noise interference, making it difficult to extract effective features. Feature extraction is often limited to Euclidean space and cannot characterize implicit features in complex environments, thus seriously affecting the accuracy of event classification.

[0005] For DOFS signal classification, traditional methods are mostly based on explicit feature extraction and fixed mapping functions to build classifiers. Their mathematical foundation usually relies on the assumptions of feature independence and statistical stationarity, which makes it difficult to reflect the dynamic changes in signal distribution. This makes it impossible for the fixed mapping function of traditional models to adaptively adjust to distribution drift, and also impossible to capture the contextual dependencies and task condition differences between different events, resulting in a significant decrease in adaptability.

[0006] DOFS signal classification has been widely used in many fields. For example, algorithms such as CNN and KNN have achieved good signal classification results in structural health monitoring, intrusion detection, and pipeline monitoring.

[0007] With the development of smart cities and the Internet of Things, distributed fiber optic sensing technology is widely used in many fields such as geological exploration, automobile manufacturing and aerospace, which puts forward higher requirements for signal classification accuracy and model generalization ability.

[0008] Therefore, to achieve highly accurate DOFS signal classification in complex scenarios and meet the requirements of signal classification, it is necessary to establish an efficient and accurate DOFS signal classification method to improve event identification efficiency and DOFS signal classification accuracy. This will provide efficient and accurate anomaly identification capabilities for many application areas of distributed fiber optic sensing technology, enabling practitioners to make better decisions. Summary of the Invention

[0009] To address the aforementioned problems, this invention provides a DOFS signal classification method based on DWT-RMG-TabPFN, which can improve the accuracy of DOFS signal classification.

[0010] The specific steps are as follows:

[0011] Step 1: For a single event M, acquire the raw DOFS signals of adjacent time slots in the distributed fiber optic sensor and calculate the current time slot. The differential signal;

[0012] Time slot differential signal The formula is:

[0013]

[0014] In the formula, For time slots The original signal;

[0015] Step 2: Divide the differential signal into segments using a segmented window function and calculate the segmented differential signal corresponding to each segment;

[0016] Time slot The Middle Segmented window function for signal segmentation :

[0017]

[0018] In the formula, This is the scaling factor; The signal length; For time slots The total number of segments in the differential signal.

[0019] Time slot differential signal Divided into Section, No. Segmented differential signal The calculation is as follows:

[0020]

[0021] Step 3: Perform DWT decomposition on each segment of the differential signal and extract the corresponding low-frequency approximation coefficients and high-frequency detail coefficients.

[0022] For the first Segmented differential signal The formula for extracting the coefficients is as follows:

[0023]

[0024] In the formula, For the first Low-frequency approximation coefficients extracted from the segmented differential signal of the segment; For the first High-frequency detail coefficients extracted from segmented differential signals; Indicates the first Channel 1 The approximation coefficients of the layer capture the low-frequency characteristics of the signal; among which the channels , This represents the total number of channels in the DWT decomposition. , The total number of layers for DWT decomposition; the number of channels and layers is set manually according to the actual situation. These are the parameters of the low-pass filter; Indicates the first Wavelet coefficients; For detail coefficients, representing the first... aisle High-frequency components at the scale; These are the parameters for the high-pass filter.

[0025] Step 4: Using the low-frequency approximation coefficients and high-frequency detail coefficients corresponding to each segment of the differential signal, establish the SPD matrix corresponding to each segment;

[0026] Specifically:

[0027] First, regarding the first The segmented differential signal is used to construct the high and low frequency component coefficient matrix of the signal using its corresponding low-frequency approximation coefficients and high-frequency detail coefficients. :

[0028]

[0029] Then, the coefficient matrix of the high and low frequency components of the signal. Calculate the covariance to obtain the first... SPD matrix of segmented differential signals :

[0030]

[0031] In the formula, Representation matrix The row mean matrix, matrix Size is .

[0032] Step 5: Traverse the SPD matrix corresponding to each segment and extract the current time slot. The RMG features of event M are the RMG features of event M.

[0033] Specifically:

[0034] First, regarding time slots For each segment of the SPD matrix, randomly perform pairwise traversal and calculate the SPD matrix. and Riemannian distance :

[0035]

[0036] Indicates the first The and the first The geodesic distance between two points on the manifold represented by an SPD matrix. ; This represents the Frobenius norm.

[0037] Then, using Riemannian manifold distance Calculating the Riemann center of mass :

[0038]

[0039] SPD matrix Any SPD matrix in the dataset.

[0040] Next, using the Riemann center of mass Calculate the relative angles of Riemannian manifolds :

[0041]

[0042] In the formula, This indicates that the Riemannian manifold Mapped to the Riemannian center of mass The tangent space with base point, i.e.

[0043] .

[0044] Finally, the above features are spliced ​​together to obtain the time slot. RMG features:

[0045]

[0046] In the formula, Represents RMG features; Represents vectorization operators; This represents the feature splicing function.

[0047] Step Six: Input The RMG features of the event type are encoded using a TabPFN network to obtain the input sequence. ;

[0048] First, RMG feature set of class events Divide the dataset into training and testing sets; and manually label each type of event with event tags.

[0049] Then, feature vectors are selected from the training set. , and the event label Encode the code to obtain the encoded result. :

[0050]

[0051] In the formula, Indicates the first RMG characteristics of similar events; Indicates the first Event tag; , These are the parameters of the weight matrix. For paranoia vector parameters;

[0052] Then, the feature vectors of all events in the test set are... Encode the code to obtain the encoded result. :

[0053]

[0054] In the formula, Encode the obtained test set; For the paranoia vector parameters.

[0055] Finally, all the encoding results from the training set and the encoding results from the test set are combined to form the input sequence. :

[0056] .

[0057] Step 7: Input sequence Combined with mask matrix The common input is processed through the inference mechanism of the Transformer to obtain the inference result. :

[0058]

[0059] In the formula, the mask matrix .

[0060] Step 8: Utilize the results of reasoning Classify and predict test events:

[0061]

[0062] In the formula, The predicted output is the event classification result; This is a custom weight matrix; A custom paranoia matrix; Regarding the reasoning results The set of all categories obtained by performing softmax probability normalization.

[0063] The advantages of this invention are:

[0064] (1) This invention overcomes the problems of poor ability to capture local mutation features of Fourier decomposition signals and low physical adaptability of variational mode decomposition. The signal decomposition is more comprehensive and can effectively improve the accuracy of signal classification.

[0065] (2) This invention overcomes the problem that traditional signal feature extraction methods are limited to Euclidean space, and the extracted features are more complete, which helps the neural network to accurately classify signals.

[0066] (3) This invention approximates the TabPFN neural network through a Bayesian inference integral process and combines it with RMG to achieve a high classification accuracy. This shows that this invention can achieve a good classification effect when classifying DOFS signals. Attached Figure Description

[0067] Figure 1 This is a flowchart of a DOFS signal classification method based on DWT-RMG-TabPFN according to the present invention;

[0068] Figure 2 This is a schematic diagram of the TabPFN neural network described in this invention;

[0069] Figure 3 This is the result of classifying four groups of DOFS signals using this invention; Detailed Implementation

[0070] The invention will be further described in detail below through implementation examples.

[0071] This invention provides a DOFS signal classification method based on DWT-RMG-TabPFN, which can improve the accuracy of DOFS signal classification, such as... Figure 1 As shown, the specific steps are as follows:

[0072] Step 1: For a single event M, acquire the raw DOFS signals of adjacent time slots in the distributed fiber optic sensor and calculate the current time slot. The differential signal;

[0073]

[0074] In the formula, For time slots The original signal; Indicates a differential signal; Indicates time.

[0075] Step 2: Divide the differential signal into segments using a segmented window function and calculate the segmented differential signal corresponding to each segment;

[0076] Time slot The Middle Segmented window function for signal segmentation :

[0077]

[0078] In the formula, This is the scaling factor; The signal length; For time slots The total number of segments in the differential signal.

[0079] Time slot differential signal Divided into Section, No. Segmented differential signal The calculation is as follows:

[0080]

[0081] Step 3: Perform DWT decomposition on each segment of the differential signal and extract the corresponding low-frequency approximation coefficients and high-frequency detail coefficients.

[0082] For the first Segmented differential signal The formula for extracting the coefficients is as follows:

[0083]

[0084] In the formula, For the first Low-frequency approximation coefficients extracted from the segmented differential signal of the segment; For the first High-frequency detail coefficients extracted from segmented differential signals; Indicates the first Channel 1 The approximation coefficients of the layer capture the low-frequency characteristics of the signal; among which the channels , This represents the total number of channels in the DWT decomposition. , The total number of layers for DWT decomposition; the number of channels and layers is set manually according to the actual situation. These are the parameters of the low-pass filter; Indicates the first Wavelet coefficients; For detail coefficients, representing the first... aisle High-frequency components at the scale; These are the parameters for the high-pass filter.

[0085] Step 4: Using the low-frequency approximation coefficients and high-frequency detail coefficients corresponding to each segment of the differential signal, establish the SPD matrix corresponding to each segment;

[0086] Specifically:

[0087] First, regarding the first The segmented differential signal is used to construct the high and low frequency component coefficient matrix of the signal using its corresponding low-frequency approximation coefficients and high-frequency detail coefficients. :

[0088]

[0089] Then, the coefficient matrix of the high and low frequency components of the signal. Calculate the covariance to obtain the first... SPD matrix of segmented differential signals :

[0090]

[0091] In the formula, Representation matrix The row mean matrix, matrix Size is .

[0092] Step 5: Traverse the SPD matrix corresponding to each segment and extract the current time slot. The RMG features of event M are the RMG features of event M.

[0093] Specifically:

[0094] First, regarding time slots For each segment of the SPD matrix, randomly perform pairwise traversal and calculate the SPD matrix. and Riemannian distance :

[0095]

[0096] Indicates the first The and the first The geodesic distance between two points on the manifold represented by an SPD matrix. ; This represents the Frobenius norm.

[0097] Then, using Riemannian manifold distance Calculating the Riemann center of mass :

[0098]

[0099] SPD matrix Any SPD matrix in the given form, where the expression satisfying the above formula is represented as follows: This is the Riemann center of mass.

[0100] Next, using the Riemann center of mass Calculate the relative angles of Riemannian manifolds :

[0101]

[0102] In the formula, This indicates that the Riemannian manifold Mapped to the Riemannian center of mass The tangent space with base point, i.e.

[0103] .

[0104] Finally, the above features are spliced ​​together to obtain the time slot. RMG features:

[0105]

[0106] In the formula, Represents RMG features; Represents vectorization operators; This represents the feature splicing function.

[0107] Step Six: Input The RMG features of the event type are encoded using a TabPFN network to obtain the input sequence. ;

[0108] First, RMG feature set of class events Divide the dataset into training and testing sets; and manually label each type of event with event tags.

[0109] Then, feature vectors are selected from the training set. , and the event label Encode the code to obtain the encoded result. :

[0110]

[0111] In the formula, Indicates the first RMG characteristics of similar events; Indicates the first Event tag; , These are the parameters of the weight matrix. For paranoia vector parameters;

[0112] Then, the feature vectors of all events in the test set are... Encode the code to obtain the encoded result. :

[0113]

[0114] In the formula, Encode the obtained test set; For the purpose of several events The test set consists of; For the paranoia vector parameters.

[0115] Finally, all the encoding results from the training set and the encoding results from the test set are combined to form the input sequence. :

[0116] .

[0117] Step 7: Input sequence Combined with mask matrix The common input is processed through the inference mechanism of the Transformer to obtain the inference result. :

[0118]

[0119] In the formula, the mask matrix .

[0120] Step 8: Utilize the results of reasoning Classify and predict test events:

[0121]

[0122] In the formula, The predicted output is the event classification result; This is a custom weight matrix; This is a custom paranoia matrix. Regarding the reasoning results The set of all categories obtained by performing softmax probability normalization;

[0123] From category set The option with the highest probability is selected as the predicted category for the test event.

[0124] Example:

[0125] The selected dataset contains 1000 samples, each with a signal length of 1000. There are 200 samples each of DOFS signals from car horns, rain, welding, pedestrians passing by, and normal conditions. Using random sampling, 150 samples were selected from each of the five datasets as the training set, and the remaining 50 samples were used as the test set. Ultimately, a total of 750 samples were used for training, and 250 samples were used for testing.

[0126] The specific steps are as follows:

[0127] (1) Determine the differential signal of the original signal :

[0128] By selecting one set of impact signal data and performing calculations, the original vibration signal can be obtained. The resulting differential signal is (-0.11, 0.04, 0.02, … 0.04). The values ​​are (0.16, -0.03, -0.08, ..., -0.07).

[0129] (2) Determine the signal segmentation window function :

[0130] The signal length is 1000, evenly divided into 10 segments, each segment being [length missing]. If it is 100, then Determine the segmented window function for the signal. For example, in paragraph 3:

[0131]

[0132] (3) Determine the differential segmented signal :

[0133] Calculate the segmented signal using a window function. ,get The inequality is (0.16, -0.04, -0.21, ..., 0.03).

[0134] (4) Perform DWT decomposition to determine the low-frequency approximation coefficients for each signal segment. and high frequency detail coefficient :

[0135] Perform DWT decomposition on the third segment of the signal to obtain the low-frequency approximation coefficients. (-0.03, -0.02, -0.02,…, -0.03) represents the high-frequency detail coefficients. The range is (-0.01, -0.03, 0.01,…, 0.05; -0.03,0.02, 0.03,…, -0.03;……).

[0136] (5) Establish the SPD matrix

[0137] First, for the third segment of the signal obtained above, establish the high and low frequency component coefficient matrix of the signal:

[0138]

[0139] Then, to Calculate the covariance to obtain the SPD matrix:

[0140]

[0141] (6) Establish RMG feature representation

[0142] First, calculate the Riemannian distance, and obtain... The inequality is (0.3, 0.4, …, 0.9).

[0143] Then calculate the relative angles of the Riemannian manifold. The range is (1.34, 2.32, ..., 1.31).

[0144] Then, the RMG feature representation is established, and the result is... The inequality is (0.3, 0.4, …, 0.9, 1.34, 2.32, …, 1.31, -0.01, -0.03, 0.01, …, 0.05, 0.02, 0.32, …, 0.03).

[0145] (6) Construct the TabPFN network:

[0146] Therefore, in this embodiment, the input feature vector has a dimension of 21 and the sample categories are 5. The TabPFN network is constructed as follows: Figure 2 As shown;

[0147] (7) Determine the discriminant function of the output variable. :

[0148] Set output layer parameters The inequality is (0.2, 0.3, 1, 1.2, 0.3, 2.1, 3.2).

[0149] Paranoia parameter The expression is (0.03, -0.02, 0.02, 0.05, 0.07, 0.09, -0.02).

[0150] The trained TabPFN neural network is tested using test set samples to achieve DOFS signal classification, and the classification results are obtained, thus completing the DOFS signal classification method based on DWT-RMG-TabPFN.

[0151] To verify the accuracy of the present invention in classifying DOFS signals, four sets of vibration signal classification experiments were conducted. The experimental results are as follows: Figure 3 As shown. By Figure 3 As can be seen, the DOFS signal identification method established in this invention maintains an accuracy rate of over 99% in classifying DOFS signals, achieving high accuracy while ensuring stability, and demonstrating good classification results. This indicates that the DOFS signal classification method established in this invention is effective, providing a better method for establishing accurate vibration signal classification models, and has certain practical value.

Claims

1. A DWT-RMG-TabPFN-based DOFS signal classification method, characterized in that, The specific steps are as follows: Step one, for a single event M, collect the original DOFS signals of adjacent time slots in the distributed optical fiber sensor, and calculate the difference signal of the current time slot ; Step 2: Divide the differential signal into segments using a segmented window function and calculate the segmented differential signal corresponding to each segment; Time slot differential signal Divided into Section, No. Segmented differential signal The calculation is as follows: For time slots The Middle Segmentation window function for the signal segment; Step 3: Perform DWT decomposition on each segment of the differential signal and extract the corresponding low-frequency approximation coefficients and high-frequency detail coefficients. Step 4: Using the low-frequency approximation coefficients and high-frequency detail coefficients corresponding to each segment of the differential signal, establish the SPD matrix corresponding to each segment; Step 5: Traverse the SPD matrix corresponding to each segment and extract the current time slot. The RMG features of event M are the RMG features of event M. Step Six: Input The RMG features of the event type are encoded using a TabPFN network to obtain the input sequence. ; First, RMG feature set of class events Divide the dataset into training and testing sets; and manually label each type of event with event tags. Then, feature vectors are selected from the training set. , and the event label Encode the code to obtain the encoded result. : In the formula, Indicates the first RMG characteristics of similar events; Indicates the first Event tag; , These are the parameters of the weight matrix. For paranoia vector parameters; Then, the feature vectors of all events in the test set are... Encode the code to obtain the encoded result. : In the formula, Encode the obtained test set; For paranoia vector parameters; Finally, all the encoding results from the training set and the encoding results from the test set are combined to form the input sequence. : Step 7: Input sequence Combined with mask matrix The common input is used in the inference mechanism of the Transformer to obtain the inference result. : In the formula, the mask matrix ; Step 8: Utilize the results of reasoning Classify and predict test events: In the formula, The predicted output is the event classification result; This is a custom weight matrix; A custom paranoia matrix; Regarding the reasoning results The set of all categories obtained by performing softmax probability normalization.

2. The method as described in claim 1, characterized in that, The time slot of step one differential signal The formula is: In the formula, For time slots The original signal.

3. The method as described in claim 1, characterized in that, In step two, the time slot The Middle Segmented window function for signal segmentation : In the formula, This is the scaling factor; The signal length; For time slots The total number of segments in the differential signal.

4. The method as described in claim 1, characterized in that, In step three, the first... Segmented differential signal The formula for extracting the coefficients is as follows: In the formula, For the first Low-frequency approximation coefficients extracted from the segmented differential signal of the segment; For the first High-frequency detail coefficients extracted from segmented differential signals; Indicates the first Channel 1 The approximation coefficients of the layer capture the low-frequency characteristics of the signal; Among them, the channel , This represents the total number of channels in the DWT decomposition. , The total number of layers for DWT decomposition; the number of channels and layers is set manually according to the actual situation. These are the parameters of the low-pass filter; Indicates the first Wavelet coefficients; For detail coefficients, representing the first... aisle High-frequency components at the scale; These are the parameters for the high-pass filter.

5. The method as described in claim 4, characterized in that, Step four specifically involves: First, regarding the first The segmented differential signal is used to construct the high and low frequency component coefficient matrix of the signal using its corresponding low-frequency approximation coefficients and high-frequency detail coefficients. : Then, the coefficient matrix of the high and low frequency components of the signal. Calculate the covariance to obtain the first... SPD matrix of segmented differential signals : In the formula, Representation matrix The row mean matrix, matrix Size is .

6. The method as described in claim 4 or 5, characterized in that, Step five specifically involves: First, regarding time slots For each segment of the SPD matrix, randomly perform pairwise traversal and calculate the SPD matrix. and Riemannian distance : Indicates the first The and the first The geodesic distance between two points on the manifold represented by an SPD matrix. ; Denotes the Frobenius norm; Then, using Riemannian manifold distance Calculating the Riemann center of mass : SPD matrix Any SPD matrix in the dataset; Next, using the Riemann center of mass Calculate the relative angles of Riemannian manifolds : In the formula, This indicates that the Riemannian manifold Mapped to the Riemannian center of mass The tangent space with base point, i.e. Finally, the above features are spliced ​​together to obtain the time slot. RMG features: In the formula, Represents RMG features; Represents vectorization operators; This represents the feature splicing function.

Images