DAS signal classification method based on improved wavelet threshold denoising
Patent Information
- Application Number
- CN202410159856.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-04
- Publication Date
- 2026-09-04
- Estimated Expiration
- 2044-02-04
AI Technical Summary
[0006]DAS采集的信号多为非线性非平稳的信号，高斯滤波和均值滤波等方法，去噪效果较差；低通滤波和带通滤波等经典滤波算法，可能会导致信号中重要的非线性特征丢失或模糊；经验模态分解类算法去噪会有模态混叠的现象产生，处理起来较为麻烦
[0060] 1) In the wavelet denoising preprocessing of this invention, the threshold determination is combined with the AIC algorithm, which makes the threshold more accurate, improves the effect of threshold quantization of detail coefficients, and makes the nearest neighbor determined by the KNN algorithm more accurate.
Smart Images
Figure CN118194154B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of DAS signal processing technology, and specifically to a DAS signal classification method based on improved wavelet threshold denoising. Background Technology
[0002] Distributed Optical Fiber Acoustic Sensing (DAS) is a novel sensing technology capable of continuous distributed detection of vibration and sound fields, and it has been widely applied in fields such as seismic exploration, pipeline monitoring, power cables, and perimeter security. However, research on the application of DAS technology in event classification, especially the classification of various similar events, has become a bottleneck in its application development.
[0003] The signals acquired by the DAS system reflect the characteristics of different vibration sources, requiring appropriate methods to classify different events. The K-Nearest Neighbors (KNN) classification algorithm is a classic machine learning algorithm, belonging to supervised learning. Traditional KNN uses majority voting to assign the most frequent category among the K nearest neighbors to the sample to be classified. Each neighbor contributes equally to the final classification, without considering the similarity differences between samples. Weighted KNN (WKNN), on the other hand, assigns a weight to each nearest neighbor, typically calculated based on distance; closer samples have higher weights. During final classification, the category of the sample is determined by weighted voting, taking into account the similarity differences between neighbors, resulting in more accurate classification results.
[0004] KNN classifies samples based on distance metrics. When the dataset contains noise, the classification accuracy is affected, impacting the classification results. Noise may bias the selection of nearest neighbors, thus affecting classification accuracy. Therefore, denoising the raw signal acquired by DAS to improve the signal-to-noise ratio and enhance signal features is crucial for subsequent signal classification. Wavelet transform methods are widely used in signal denoising, and wavelet thresholding denoising has become the most common signal denoising method in practical engineering applications due to its ease of implementation and preservation of original signal features.
[0005] Defects and shortcomings of existing technology:
[0006] Signals acquired by DAS are mostly nonlinear and non-stationary. Methods such as Gaussian filtering and mean filtering have poor denoising effects. Classical filtering algorithms like low-pass and band-pass filtering may cause the loss or blurring of important nonlinear features in the signal. Empirical Mode Decomposition (EMD) algorithms can produce mode aliasing, which is difficult to handle. Wavelet thresholding is well-suited for such signals, but the low accuracy of traditional threshold estimation methods affects subsequent threshold quantization. Hard thresholding functions have good denoising effects but reduce signal smoothness, while soft thresholding functions maintain signal smoothness but have poor denoising effects. Both methods zero out high-frequency coefficients below the wavelet threshold, resulting in the loss of some useful information. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention proposes a DAS signal classification method based on improved wavelet threshold denoising. First, the DAS signal is preprocessed using an improved wavelet threshold denoising method. The threshold is then determined using the AIC algorithm, improving the accuracy of the wavelet threshold. A new threshold function is then used for threshold quantization. Compared to traditional soft and hard threshold functions, this threshold function retains some detail coefficients below the threshold, preserving more signal features and achieving superior denoising performance. Finally, the WKNN algorithm is used for signal classification, significantly improving classification accuracy.
[0008] The technical solution adopted in this invention is as follows:
[0009] The DAS signal classification method based on improved wavelet threshold denoising includes the following steps:
[0010] Step 1: Perform wavelet decomposition on the DAS signal Q to obtain the detail coefficients cD{cD1, cD2, ..., cD...} M} and approximation coefficients cA{cA1, cA2, ..., cA M}, where: M is the wavelet decomposition level, cD m cA m These are the detail coefficients and approximation coefficients for the m-th layer, respectively, where m = 1, 2, ..., M;
[0011] Step 2: Determine the threshold using the Akaike Information Criterion (AIC), and apply a new threshold function to perform threshold quantization on the detail coefficients cD after wavelet decomposition, obtaining the processed detail coefficients cD'{cD1', cD2', ..., cD...} M '};cD m ' represents the detail coefficient of the m-th layer after threshold quantization;
[0012] Step 3: Use the processed detail coefficients CD' and the approximation coefficients cA of the last layer of wavelet decomposition. M Perform inverse wavelet transform to reconstruct the signal and obtain the wavelet-denoised signal Q'.
[0013] Step 4: Extract the time-domain features of the wavelet-denoised signal Q', construct a dataset using the extracted time-domain features, and divide the dataset into a training set and a test set;
[0014] Step 5: Use the WKNN algorithm to classify the signals in the test set.
[0015] In step 2, the detail coefficient cD of the m-th layer m Threshold quantization processing specifically includes the following steps:
[0016] Step 2.1: Calculate the detail coefficients cD m The AIC value of the t-th sampling point, obtain the flag:
[0017] AIC(t)=t*lg(var(cD m (1:t)))+(Nt-1)*lg(var(cD m (t+1:N)))t=2, 3..., N-1
[0018] Where var is the variance calculation function; N is the cD. m The length of the detail coefficient cDm; t represents the t-th sampling point of the detail coefficient cDm; AIC is the name of a sequence, and AIC(t) represents the detail coefficient cDm. m The AIC value of the t-th sampling point.
[0019] Store the signal sampling point number corresponding to the minimum AIC value in the flag.
[0020] Step 2.2: Determine the detail coefficients cD m Threshold λ:
[0021] Let the detail factor cD m Let T be the sampling point number corresponding to the maximum value. Construct a new sequence AIC_cD using the following formula. m :
[0022]
[0023] Use the new sequence AIC_cD m The threshold λ is determined by combining heuristic thresholding, as follows:
[0024] The heuristic threshold calculation method is as follows:
[0025] First, calculate β and γ using the following formula:
[0026]
[0027]
[0028] Where LL is the sequence AIC_cD m The length of β < γ,
[0029]
[0030] When β≥γ
[0031] AIC_cD m Take the absolute value, sort in ascending order, and then square it to get a new sequence f(k): f(k) = (sort(|AIC_cD)). m (k)|)) 2 k = 1, 2, ..., L;
[0032] Here, sort is the ascending order sorting function.
[0033] If the square root of the kth element of the threshold f(k) is taken as the threshold λ, then the risk generated by this threshold is risk(k).
[0034]
[0035] Find the value of k corresponding to the minimum value of risk(k) to obtain λ′;
[0036]
[0037] when hour,
[0038] λ=λ′
[0039] when hour,
[0040]
[0041] Step 2.3: Adjust the detail coefficients (cD) m Threshold quantization is performed to obtain cD m ';
[0042] The new threshold function is as follows:
[0043]
[0044] Where ω is cD m The coefficient values corresponding to each sampling point For each sampling point after threshold quantization, the coefficient values are as follows: when cD m After threshold quantization of all coefficients, cD is obtained.m Pi represents π (pi).
[0045] The detail coefficients of each layer are processed according to steps 2.1 to 2.3 to obtain the processed detail coefficients cD'{cD1', cD2', ..., cD M '}.
[0046] In step 3, the processed detail coefficients CD' are used along with the approximation coefficients cA of the last layer of wavelet decomposition. M Perform inverse wavelet transform to reconstruct the signal, obtaining the wavelet-denoised signal Q'; the details are as follows:
[0047] The processed detail coefficients CD' are compared with the approximation coefficients cA of the last layer of wavelet decomposition. M Arranged as follows:
[0048] wav = [cA M cD M 'cD M-1 ′……cD1′]
[0049] Implement inverse wavelet transform using the waverec function in MATLAB:
[0050] Q′=waverec(wav,len,wave)
[0051] Where len is cA M , cD M ′,cD M-1 '……cD1', the matrix formed by the lengths of Q; wave is the wavelet basis used for wavelet decomposition.
[0052] In step 5, the samples S in the test set are classified using the WKNN algorithm, which specifically includes the following steps:
[0053] 5.1 Calculate the distance: Calculate the Euclidean distance between sample S and all samples in the training set. The Euclidean distance D between sample X (X1, X2, ..., XN) and sample Y (Y1, Y2, ..., YN) is calculated as follows:
[0054]
[0055] Where X1, X2, ..., XN are the features of sample X, and Y1, Y2, ..., YN are the features of sample Y;
[0056] 5.2 Determine the nearest neighbors: Select the K training samples with the shortest Euclidean distance to sample S, where K is a constant.
[0057] 5.3. Weighting: The reciprocal of the distance between a neighboring sample and sample S (1 / D) is used as the weight of the neighboring sample;
[0058] 5.4 Voting: Calculate the sum of the weights of neighboring samples of the same class, and take the class of the sample S as the class of the sample with the maximum weight. Perform operations 5.1 to 5.4 on all samples in the test set to obtain the classification results of all samples in the test set.
[0059] This invention discloses a DAS signal classification method based on improved wavelet threshold denoising, with the following technical advantages:
[0060] 1) In the wavelet denoising preprocessing of this invention, the threshold determination is combined with the AIC algorithm, which makes the threshold more accurate, improves the effect of threshold quantization of detail coefficients, and makes the nearest neighbor determined by the KNN algorithm more accurate.
[0061] 2) In the wavelet denoising preprocessing of this invention, a new threshold function is proposed to process the detail coefficients below the threshold, thereby preserving more signal features and improving the accuracy of signal classification.
[0062] 3) This invention employs wavelet denoising preprocessing, then extracts signal features to construct a dataset, and uses WKNN to classify DAS signals, thereby improving classification accuracy. Attached Figure Description
[0063] Figure 1 This is a flowchart of the present invention.
[0064] Figure 2 This is the time-domain plot of the original DAS signal.
[0065] Figure 3 A comparison chart showing the noise reduction effects of traditional filtering methods on simulated signals.
[0066] Figure 4 The image shows a comparison of the wavelet threshold denoising effects on simulated signals.
[0067] Figure 5 Comparison of noise reduction effects for welding signals acquired by DAS.
[0068] Figure 6 The image shows the KNN classification results for the DAS signal.
[0069] Figure 7 The image shows the WKNN classification results for the DAS signal. Detailed Implementation
[0070] The DAS signal classification method based on improved wavelet threshold denoising has the following process: Figure 1 As shown:
[0071] Step 1: Perform wavelet decomposition on the DAS signal Q to obtain the detail coefficients cD{cD1, cD2, ..., cD...} M} and approximation coefficients cA{cA1, cA2, ..., cA M}, where M is the wavelet decomposition level, cD m cA m These are the detail coefficients and approximation coefficients for the m-th layer, respectively.
[0072] Step 2: Determine the threshold using AIC, and apply a new threshold function to perform threshold quantization on the detail coefficients cD after wavelet decomposition, obtaining the processed detail coefficients cD'{cD1', cD2', ..., cD...} M '}, where the detail coefficients cD of the m-th layer are... m The specific steps of threshold quantization are as follows:
[0073] 2.1: Calculate the detail coefficients cD m Get the AIC value of the t-th sampling point and obtain the flag;
[0074]
[0075] Where var is the variance calculation function, and N is the cD. m The length of the signal sample point corresponding to the minimum AIC value is stored in the flag.
[0076] 2.2: Determine the detail coefficients cD m Threshold λ;
[0077] Let the detail factor cD m Let T be the sampling point number corresponding to the maximum value. Construct a new sequence AIC_cD using the following formula. m :
[0078]
[0079] Use the new sequence AIC_cD m The threshold λ is determined by combining heuristic thresholding.
[0080] 2.3: Detail coefficient cD m Threshold quantization is performed to obtain cD m ';
[0081] The new threshold function is as follows:
[0082]
[0083] Where ω is cD m The coefficient values corresponding to each sampling point For each sampling point after threshold quantization, the coefficient values are as follows: when cD m After threshold quantization of all coefficients, cD is obtained. m '.
[0084] The detail coefficients of each layer are processed according to steps 2.1 to 2.3 to obtain the processed detail coefficients cD'{cD1', cD2', ..., cD M '}.
[0085] Step 3: Use the processed detail coefficients CD' and the approximation coefficients cA of the last layer of wavelet decomposition. M Perform inverse wavelet transform to reconstruct the signal and obtain the wavelet-denoised signal Q'.
[0086] Step 4: Extract time-domain features of signal Q' such as the number of peaks, waveform factor, margin factor, peak factor, kurtosis and skewness. Use the extracted time-domain features to construct a dataset and divide the dataset into training set and test set.
[0087] The calculation formula is as follows:
[0088] Waveform factor:
[0089] Margin factor:
[0090] Peak factor:
[0091] data: Input data, i.e., Q';
[0092] rms: Root Mean Square function, calculates the root mean square of the input data;
[0093] mean: The average function, which calculates the average of the input data;
[0094] max: The maximum value function, which finds the maximum value of the input data;
[0095] min: Minimum value function, finds the minimum value of the input data;
[0096] The number of peaks, kurtosis, and skewness were obtained using the built-in MATLAB functions findpeaks, kurtosis, and skewness, respectively.
[0097] Step 5: Classify the signals in the test set using the WKNN algorithm. The specific steps for classifying samples S in the test set using the WKNN algorithm are as follows:
[0098] 5.1 Calculate the distance: Calculate the Euclidean distance between sample S and all samples in the training set. The Euclidean distance D between sample X (X1, X2, ..., XN) and sample Y (Y1, Y2, ..., YN) is calculated as follows:
[0099]
[0100] Where X1, X2, ..., XN are the features of sample X, and Y1, Y2, ..., YN are the features of sample Y.
[0101] 5.2 Determine the nearest neighbors: Select the K training samples with the shortest Euclidean distance to sample S, where K is a constant.
[0102] 5.3. Assign weights: The reciprocal of the distance between a neighboring sample and sample S (1 / D) is used as the weight of the neighboring sample.
[0103] 5.4 Voting: Calculate the sum of the weights of neighboring samples of the same category, and take the category of the maximum value as the category of sample S.
[0104] Perform the operations from 5.1 to 5.4 on all samples in the test set to obtain the classification results for all samples in the test set.
[0105] Example 1:
[0106] The present invention classifies real DAS signals in the following manner according to the following steps:
[0107] The DAS public dataset from Hong Kong Polytechnic University is referenced in [1]: Zhu K, Huan WU, Chao S, et al. Phase recognition in distributed fiber-optic acoustic sensor using intensity and phase stacked convolutional neural network with data augmentation [J]. Optics Express, 2021, 29(3). Experimental verification was performed. The dataset contains 10 DAS vibration signals, and 5 of them, namely welding, thunderstorm, shoveling, rain, and jackhammer, were selected for verification. First, the data of the above 5 signals in the dataset were spliced together. The spliced time domain signal is as follows. Figure 2 As shown, the spliced signal was then divided into groups of 10,000 points each, with 177 groups for welding, 187 groups for thunderstorm, 176 groups for shoveling, 186 groups for rain, and 156 groups for jackhammer, for a total of 882 samples.
[0108] Wavelet denoising was performed on 882 samples using the db4 wavelet basis and a decomposition level M of 3. The wavelet in the wavelet reconstruction is 'db4'.
[0109] First, the performance of the denoising algorithm was evaluated using simulated signals, specifically the 'blocks' signal provided by MATLAB. The denoising algorithms compared included three traditional filtering methods: median filtering, mean filtering, and Gaussian filtering; and three wavelet denoising methods: soft thresholding, hard thresholding, and improved thresholding. In wavelet denoising, the threshold was determined solely using the heursure method. The signal-to-noise ratio (SNR) and root mean square error (RSE) are shown in Table 1. The denoising results using the heursure method with AIC to determine the threshold are shown in Table 2. The original ideal signal, the noisy signal, the median-filtered signal, the mean-filtered signal, and the Gaussian-filtered signal are shown in Table 2. Figure 3 The wavelet denoising signals using the heursure method to determine the threshold, and the wavelet denoising signals using the soft thresholding function, hard thresholding function, and improved thresholding function combined with AIC, are shown in [link to wavelet denoising signal analysis]. Figure 4 .
[0110] Table 1 Evaluation of Noise Reduction Effect of Simulated Signals
[0111]
[0112] In Table 1, the signal-to-noise ratios (SNRs) after Gaussian filtering and mean filtering are 16.1966 dB and 15.4373 dB, respectively, indicating poor denoising performance. Median filtering, however, achieves an SNR of 17.1564 dB, which is better than the 16.6306 dB achieved with the soft thresholding function. However, it still lags behind the 17.4808 dB SNR achieved with the hard thresholding function. The thresholding function used in this invention achieves the best denoising effect, reaching 17.5922 dB, and the root mean square error between the denoised signal and the original signal is 0.4617, meaning it has the smallest deviation from the original signal, thus verifying the practicality of the thresholding function of this invention.
[0113] Table 2 Evaluation of Wavelet Denoising Effect Based on AIC Threshold Determination
[0114]
[0115] Table 2 shows that threshold quantization, performed after determining the threshold using AIC, improves the effectiveness of threshold quantization. The SNR of the soft threshold function increased from 16.6306 dB to 16.8094 dB, the SNR of the hard threshold function increased from 17.4808 dB to 17.6252 dB, and the SNR of the improved threshold function increased from 17.5922 dB to 17.6254 dB. This demonstrates that determining the threshold using AIC can improve the wavelet denoising effect and verifies the practicality of the method.
[0116] Based on Tables 1 and 2, the method of combining AIC to determine the threshold and using the improved threshold function for denoising, as used in this invention, achieves the best results.
[0117] 882 sets of temporal features were extracted from the denoised signal, including peak count, waveform factor, margin factor, peak factor, kurtosis, and skewness. A dataset was constructed using these extracted features and WKNN classification was performed. In the WKNN classification algorithm, K was set to 3, 4, or 5. The algorithm was tested 50 times, with the training and test sets split equally in each iteration. 735 sets were randomly selected as the training set, and 147 sets as the test set. The average accuracy of DAS signal classification was calculated and compared with the KNN classification algorithm. In soft and hard thresholding, the threshold determination only used the heursure method. The improved thresholding method described in this invention incorporates AIC and uses an improved threshold function. The original welding signal samples and denoised signals acquired by DAS are shown below. Figure 5 The classification results of KNN and WKNN are as follows: Figure 6 and Figure 7 As shown.
[0118] exist Figure 6 In the data analysis, when K=3, the classification accuracy of KNN using the original DAS data is 75.32%, while the classification accuracies using the traditional soft thresholding function, hard thresholding function, and the improved threshold denoising method of this invention are 85.25%, 85.63%, and 87.01%, respectively. Compared with the original data, the classification accuracy has increased by 9.93%, 10.31%, and 11.69%, respectively. When K=4, the classification accuracy of the original DAS data is 74.08%, while the classification accuracies using the traditional soft thresholding function, hard thresholding function, and the improved threshold denoising method of this invention are 83.56%, 84.20%, and 85.59%, respectively. Compared with the original data, the classification accuracy has increased by 9.48%, 10.12%, and 11.51%, respectively. When K=5, the classification accuracy of the original DAS data is 74.50%, while the classification accuracies after using the traditional soft thresholding function, hard thresholding function, and the improved threshold denoising method of this invention are 83.89%, 84.26%, and 85.67%, respectively. Compared with the original data, the classification accuracy is improved by 9.39%, 9.76%, and 11.17%, respectively. When K is 3, 4, and 5, the average classification accuracy of the original data is 74.63%, while the average classification accuracy after using the traditional soft thresholding function, hard thresholding function, and the improved threshold denoising method of this invention are 84.23%, 84.69%, and 86.09%, respectively, improving by 9.60%, 10.06%, and 11.46%, respectively. The average accuracy of the method of this invention is improved by 1.86% and 1.40% compared with the traditional soft thresholding function and hard thresholding function, respectively.
[0119] exist Figure 7In the data analysis, when K=3, the classification accuracy of the original DAS data using WKNN was 77.06%, while the classification accuracies after using the traditional soft thresholding function, hard thresholding function, and the improved thresholding denoising method of this invention were 86.81%, 86.88%, and 87.76%, respectively. Compared with the original data, the classification accuracy improved by 9.75%, 9.82%, and 10.70%, respectively. When K=4, the classification accuracy of the original DAS data was 77.25%, while the classification accuracies after using the traditional soft thresholding function, hard thresholding function, and the improved thresholding denoising method of this invention were 87.38%, 87.51%, and 88.40%, respectively. Compared with the original data, the classification accuracy improved by 10.13%, 10.26%, and 11.15%, respectively. When K=5, the classification accuracy of the original DAS data is 76.93%, while the classification accuracies after using the traditional soft thresholding function, hard thresholding function, and the improved threshold denoising method of this invention are 86.40%, 86.46%, and 87.39%, respectively. Compared with the original data, the classification accuracy has been improved by 9.47%, 9.53%, and 10.46%, respectively. When K is 3, 4, and 5, the average classification accuracy of the original data is 77.08%, while the average classification accuracy after using the traditional soft thresholding function, hard thresholding function, and the improved threshold denoising method of this invention are 86.86%, 86.95%, and 87.85%, respectively, representing improvements of 9.78%, 9.87%, and 10.77%.
[0120] Combination Figure 6 , Figure 7 When K=3, 4, 5, the WKNN algorithm improves the average accuracy of classification of original data, traditional soft threshold function, hard threshold function and the improved threshold denoising of the present invention by 2.45%, 2.63%, 2.26% and 1.76% respectively compared with the KNN algorithm.
Claims
1. A DAS signal classification method based on improved wavelet threshold denoising, characterized in that... Includes the following steps: Step 1: Perform wavelet decomposition on the DAS signal Q to obtain the detail coefficients cD{cD1, cD2, ..., cD...} M } and approximation coefficients cA{cA1, cA2, ..., cA M }, where: M is the wavelet decomposition level, cD m cA m These are the detail coefficients and approximation coefficients for the m-th layer, respectively, where m = 1, 2, ..., M; Step 2: Determine the threshold based on the Akaike Information Content Criterion (AIC), and use a new threshold function to perform threshold quantization on the detail coefficients cD after wavelet decomposition, obtaining the processed detail coefficients cD'{cD1', cD2', ..., cD...} M '};cD m ' represents the detail coefficient of the m-th layer after threshold quantization; Step 3: Use the processed detail coefficients CD' and the approximation coefficients cA of the last layer of wavelet decomposition. M Perform inverse wavelet transform to reconstruct the signal and obtain the wavelet-denoised signal Q'. Step 4: Extract the time-domain features of the wavelet-denoised signal Q', construct a dataset using the extracted time-domain features, and divide the dataset into a training set and a test set; Step 5: Use the WKNN algorithm to classify signals in the test set; In step 2, the detail coefficient cD of the m-th layer m Threshold quantization processing specifically includes the following steps: Step 2.1: Calculate the detail coefficients cD m The AIC value of the t-th sampling point, obtain the flag: ； Where var is the variance calculation function; N is the cD. m The length of the detail coefficient cDm; t represents the t-th sampling point of the detail coefficient cDm, and AIC(t) represents the detail coefficient cDm. m The AIC value of the t-th sampling point; store the sequence number of the signal sampling point corresponding to the minimum AIC value into flag; Step 2.2: Determine the detail coefficients cD m threshold : Let the detail factor cD m Let T be the sampling point number corresponding to the maximum value. Construct a new sequence using the following formula. : ； Use the new sequence AIC_cD m Determining the threshold by combining heuristic thresholding The details are as follows: The heuristic threshold calculation method is as follows: First, calculate β and γ using the following formula: ； ； Where LL is the sequence AIC_cD m Length; When β < γ, When β≥γ AIC_cD m Take the absolute value, sort in ascending order, and then square it to obtain a new sequence f(k): ； Where sort is the ascending order sorting function; Take the square root of the k-th element of the threshold f(k) as the threshold. If the threshold is such that the risk generated is risk(k); ； Take the k value corresponding to the minimum risk(k) to obtain ; ； when hour, ； when hour, ； Step 2.3: Adjust the detail coefficients (cD) m Threshold quantization is performed to obtain cD m '; The new threshold function is as follows: ； in, For cD m The coefficient values corresponding to each sampling point For each sampling point after threshold quantization, the coefficient values are as follows: when cD m After threshold quantization of all coefficients, cD is obtained. m '; Pi represents Π; The detail coefficients of each layer are processed according to steps 2.1 to 2.3 to obtain the processed detail coefficients cD'{cD1', cD2', ..., cD M '}.
2. The DAS signal classification method based on improved wavelet threshold denoising according to claim 1, characterized in that: In step 3, the processed detail coefficients CD' are used along with the approximation coefficients cA of the last layer of wavelet decomposition. M Perform inverse wavelet transform to reconstruct the signal, obtaining the wavelet-denoised signal Q'; the details are as follows: The processed detail coefficients CD' are compared with the approximation coefficients cA of the last layer of wavelet decomposition. M Arranged as follows: ； Implement inverse wavelet transform using the waverec function in MATLAB: ； Where len is , , , A matrix consisting of lengths; The wavelet basis used for wavelet decomposition.
3. The DAS signal classification method based on improved wavelet threshold denoising according to claim 1, characterized in that: In step 5, the samples S in the test set are classified using the WKNN algorithm, which specifically includes the following steps: 5.1 Calculate the distance: Calculate the Euclidean distance between sample S and all samples in the training set. The Euclidean distance D between sample X (X1, X2, ..., XN) and sample Y (Y1, Y2, ..., YN) is calculated as follows: ； Where X1, X2, ..., XN are the features of sample X, and Y1, Y2, ..., YN are the features of sample Y; 5.2 Determine the nearest neighbors: Select the K training samples with the shortest Euclidean distance to sample S, where K is a constant; 5.
3. Weighting: The reciprocal of the distance between a neighboring sample and sample S is used as the weight of the neighboring sample; 5.4 Voting: Calculate the sum of the weights of neighboring samples of the same category, and take the category of the maximum value as the category of sample S; Perform the operations from 5.1 to 5.4 on all samples in the test set to obtain the classification results for all samples in the test set.