Optical cable vibration source external force damage signal identification method based on distributed optical fiber sensing

Abstract

The invention discloses an optical cable vibration source external force damage signal identification method based on distributed optical fiber sensing, relates to the technical field of signal detection, and solves the problem of identification of the optical cable vibration source external force damage signal in the existing urban complex and changeable environment The method comprises the following steps: detecting an optical cable external force damage vibration signal by using a phi-OTDR technology, and constructing a vibration signal data set; preprocessing the vibration signals, then performing multi-feature set extraction on the preprocessed vibration signals, and dividing the preprocessed vibration signals into a training set and a test set; constructing a CatBoost recognition model, and training the recognition model by adopting a training set to obtain a trained CatBoost recognition model; adjusting and optimizing parameters in the trained CatBoost recognition model by adopting an improved whale optimization algorithm so as to obtain an optimized recognition model; and inputting the test set into the optimized recognition model, and performing performance evaluation to realize recognition of the vibration signal. Therefore, various optimization problems can be processed more effectively.

Description

Technical Field

[0001] The present invention relates to the technical field of signal detection, and in particular to a method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing. Background Art

[0002] With the acceleration of urbanization and rapid economic development, optical cables are increasingly used in urban areas. However, in the actual laying environment of urban optical cables, there are many types of external damage sources, such as damage caused by large-scale machine construction such as excavators and road rollers. The construction location and time are uncertain, and they are often affected by strong noise background, multi-source interference and unknown underground conditions in urban areas. Therefore, the accurate identification of various types of external damage sources of urban optical cables is the key point and difficulty of the present invention.

[0003] Traditional methods for identifying external damage to optical cables have problems of low accuracy and poor stability. To solve these problems, researchers have been exploring the field of event identification in distributed optical fiber sensing systems. In 2013, H.Xu et al. introduced the template matching method to the Φ-OTDR system. By establishing a disturbance signal template library, they identified three events: walking, hammering, and touching, achieving an accuracy of 87.5%. However, the template matching method is limited by the coverage of the template and its adaptability to new event types. To solve these problems, in 2017, C.Xu et al. proposed a recognition method based on multi-feature parameters. By extracting features such as short-time energy ratio and short-time level crossing rate, and applying support vector machines for classification, the average recognition accuracy of the four events of tapping, knocking, shaking, and squeezing was increased to 93.8%, but it will be limited by the distinguishing ability of the selected features and the performance of the classifier. By 2023, H. Zhu and others further developed the recognition technology and constructed a two-level support vector machine recognition algorithm for five vibration events: human touch, manual excavation, water flow in the pipe cavity, mechanical excavation and pipe leakage. This algorithm adopts a two-level classification idea through more refined feature processing. After coarse classification and fine classification, it achieves an improvement in the average accuracy rate, further promoting the performance of the Φ-OTDR system in event recognition. However, the algorithm relies too much on the feature processing process, has higher computational complexity, and has unstable recognition performance in the urban background with complex and changeable external damage sources.

[0004] In order to accurately identify the types of external damage sources (such as digging, knocking, etc.) in urban areas with complex and changeable external damage sources, improve the accuracy while taking into account the stability of the model, since SVM is sensitive to kernel function and parameter selection, it is difficult to adjust parameters in urban areas with complex and changeable external damage sources. Therefore, the CatBoost algorithm (CatBoost algorithm is a new machine learning algorithm framework based on Gradient Boosting Decision Tree (GBDT)) that can adapt to different data distributions and noise levels is selected as the core recognition algorithm. The model proposed in the present invention is the optical cable external damage vibration source signal recognition based on Φ-OTDR technology, and the machine learning CatBoost algorithm is applied to it to determine the type of vibration source.

[0005] In summary, the optical cable external break vibration source signal recognition model proposed in the present invention can improve the accuracy of urban optical cable external break detection and provide effective technical support for the safe operation of urban optical cables. Summary of the invention

[0006] In order to solve the problem of identifying external force damage signals of optical cable vibration sources in complex and changeable environments in existing urban areas, the present invention provides an optical cable vibration source external force damage signal identification method based on distributed optical fiber sensing.

[0007] A method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing, the method is implemented by the following steps:

[0008] Step 1: Detect the vibration signal of the optical cable caused by external force damage and construct a vibration signal data set;

[0009] Step 2: preprocessing the vibration signal, then extracting a multi-feature set from the preprocessed vibration signal, merging the data after the multi-feature extraction, and dividing the data into a training set and a test set;

[0010] Step 3: construct a CatBoost recognition model, and train the recognition model using a training set to obtain a trained CatBoost recognition model;

[0011] Step 4: Use the improved whale optimization algorithm to tune the parameters in the trained CatBoost recognition model to obtain an optimized recognition model; input the test set into the optimized recognition model to realize the recognition of vibration signals;

[0012] The improved whale optimization algorithm includes the introduction of nonlinear adaptive weights and sine and cosine operators. The specific process of obtaining the optimized recognition model is as follows:

[0013] Step 41: Introduce and update nonlinear adaptive weights to generate a random number p;

[0014] Step 42: Determine whether the random number p is less than 0.5. If so, execute step 43; otherwise, execute step 44;

[0015] Step 43: Determine the coefficient vector Is it less than 1? If so, use the shrinking and surrounding mechanism to update the position of the search agent; execute steps 4 and 5; otherwise, use the random search mechanism to update the position of the search agent: execute steps 4 and 5;

[0016] Step 44: Use the spiral update mechanism to update the location of the search agent; execute step 45;

[0017] Step 45: Introduce sine and cosine operators to implement secondary update of the position of the search agent, update the optimal solution, and determine whether the iteration is completed. If so, output the optimal solution; otherwise, re-execute steps 41 to 45.

[0018] Beneficial effects of the present invention: The identification method described in the present invention aims to improve the accuracy while taking into account the stability of the model. This method significantly improves the recognition performance of external force damage signals of optical cable vibration sources through a multi-feature set extraction method and an improved model optimization strategy. The innovative multi-feature set extraction method captures the multi-faceted characteristics of the signal, thereby providing rich information for the accurate identification of different types of external force damage signals. Further, through the improved whale optimization algorithm of nonlinear adaptive weights and sine and cosine operators, the combination of the two improves the robustness and global search capability of the algorithm, and uses dynamic adaptability and diversified search to enhance the convergence speed and accuracy, thereby improving the performance of the overall model.

[0019] The method of the present invention uses a machine learning CatBoost algorithm to build a recognition model. For a data set with a large number of category features, it can effectively process and analyze a multi-feature set. The model is optimized using an improved whale optimization algorithm (NSC-WOA) based on nonlinear adaptive weights and sine and cosine operators. The search capability of the algorithm can be adaptively adjusted, the stability of the model in a complex environment is enhanced, and the accurate recognition of the external force damage signal of the optical cable vibration source is achieved.

[0020] In the method of the present invention, an improved whale optimization algorithm based on nonlinear adaptive weights and sine and cosine operators is adopted. The introduction of nonlinear adaptive weights can make the algorithm parameters change adaptively with the iteration process, guide the search agent to explore the search space, balance the development and search stages, and find a suitable transition between the local and the global. The introduction of sine and cosine operators, due to the combination of the dynamic changes of the sine function and the smoothness of the cosine function, realizes richer search behaviors and more diverse update steps and directions, further enhances the global search capability, and effectively balances the development and search stages through the alternating use of sine and cosine functions. By combining nonlinear adaptive weights and sine and cosine operators, the improved whale optimization algorithm enhances its adaptability and flexibility in complex search spaces while maintaining convergence accuracy, thereby more effectively handling a variety of optimization problems.

[0021] The method of the invention has the advantages of high recognition accuracy and wide adaptability. The method is combined with distributed optical fiber sensing technology, can adapt to the vibration signal characteristics of different construction machinery, and realize the recognition of different types of construction machinery. By optimizing CatBoost with nonlinear adaptive weights and sine and cosine operator improved algorithms, the method can effectively improve the recognition accuracy of external force damage signals of optical cable vibration sources and reduce misjudgment and missed judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solution implemented in the present invention, the drawings required for use in the description will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0023] Figure 1 Schematic diagram of the overall steps of the NSC-WOA-CatBoost method based on multi-feature set extraction provided by the present invention.

[0024] Figure 2 The present invention provides a flowchart of the NSC-WOA optimization method.

[0025] Figure 3 A specific schematic diagram of the NSC-WOA-CatBoost method based on multi-feature set extraction provided by the present invention.

[0026] Figure 4 A confusion matrix diagram of the NSC-WOA-CatBoost method based on multi-feature set extraction provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] Specific implementation method 1. Combination Figures 1 to 4This embodiment describes a method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing, that is, an improved whale optimization algorithm based on nonlinear adaptive weights and sine and cosine operators based on CatBoost optimization recognition model (NSC-WOA-CatBoost) for optical cable vibration source external force damage signals. The overall steps are as follows: Figure 1 Shown include:

[0028] S1: Vibration signal data collection: Detect the external vibration signal of the optical cable by using Φ-OTDR technology.

[0029] In this implementation, the Φ-OTDR event data set from Beijing Jiaotong University is used for research. The data set contains six typical events: background noise, digging, knocking, watering, shaking, and walking. The spatiotemporal signal sample of each event consists of 10,000 points in the time domain and 12 adjacent spatial points in the air domain. Five common events related to urban outer destruction (background noise, digging, knocking, shaking, and walking) are selected for experiments.

[0030] S2: Data preprocessing: Preprocess the collected vibration signals, including operations such as data differentiation, to improve signal quality and accuracy.

[0031] In this embodiment, the vibration signal data collected by the Φ-OTDR is preprocessed, and the preprocessing specifically includes data differentiation, which can remove the connection between time and data, reduce the impact of the data change trend caused by time changes on the data itself, and make the data basically stable. The data differentiation includes first-order differentiation, etc., and the formula is as follows:

[0032] X(i)=x(i+1)-x(i)

[0033] Among them, i is the i-th data, x(i) represents the value of the i-th data point, x(i+1) represents the value of the next adjacent i-th data point, and X(i) is the data after difference.

[0034] S3: Feature extraction: Extract multiple feature sets from the preprocessed vibration signal. Merge the many new data samples after feature extraction and then divide them into training set and test set.

[0035] In this embodiment, a multi-feature set extraction is performed on the pre-processed vibration signal, including extracting the statistical features of the differential data. The feature extraction method includes an innovative multi-feature set extraction method, which obtains the spectrum through Fourier transform and calculates the statistical features of the signal. These features can be used to describe the distribution, volatility and spectral characteristics of the differential data. By extracting these features, multiple aspects of differential data information can be obtained, providing a more comprehensive feature representation for subsequent data analysis and modeling. The maximum value X of the statistical feature max , minimum value X min , peak-to-peak value X pk , mean X M , Variance X Var , standard deviation X Std , Energy X E , rectified average value X RMV , RMS X RMS , peak factor X PF , margin factor X MF , waveform factor X WF , Pulse Factor X IF , Kurtosis factor X KF The specific formulas of information entropy H(X) are as follows:

[0036] X max =max(x)

[0037] X min =min(x)

[0038]

[0039]

[0040] Among them, max(x) represents the maximum value of the differential data, min(x) represents the minimum value of the differential data, n is the total number of n data, i is the i-th data, log 10 is the logarithm with base 10, X PV is the signal peak value. The numerators of the peak factor and the pulse factor both represent the peak value of the data, but they are compared with different reference values ​​to describe different characteristics of the data. P(X=i) represents the probability that the data X takes the value i, and log represents the logarithmic operation with base 2. Before signal recognition, extracting multiple features can capture different information of the signal, improve the robustness, generalization ability and classification performance of the model, and significantly improve the effect of signal recognition.

[0041] In this implementation, the above 15 features are used to obtain the multi-directional information of the entire signal in more detail, and then these features are combined using the set principle to facilitate subsequent recognition training, that is, the multi-feature set extraction method. The specific process is as follows:

[0042] Feature set construction: define different feature sets, each set contains a statistical feature, see the above 15 statistical feature formulas for details, such as the maximum value X max ={x max1 ,x max2 ,x max3 ,...,x maxn}, where x maxn is the maximum value feature of the nth signal.

[0043] Feature standardization: Since different features may have different dimensions and numerical ranges, some features in the feature set need to be standardized to ensure that they are comparable when fused. For example, features usually associated with normal distribution are standardized using the Z-score method:

[0044]

[0045] Where x is the original data point, μ is the mean of the original data set, and σ is the standard deviation of the original data set.

[0046] Feature fusion: Use set operations to integrate features, merge features from different sets, and use union operations to form a new feature set S:

[0047] S=X max ∪X min ∪...∪H(X)

[0048] The above formula can be simplified as:

[0049] S={X max ,X min ,...,H(X)}

[0050] Each feature set has its own features, such as the maximum value X max ={x max1 ,x max2 ,x max3 ,...,x maxn}, minimum value X min ={x min1 ,x min2 ,x min3 ,...,x minn}, information entropy H(X) = {x H1 ,x H2 ,x H3 ,...,x Hn}, etc., the merged set is:

[0051] S={x max1 ,...,x maxn ,xmin1 ,...,x minn ,...,x H1 ,...,x Hn}

[0052] Among them, x maxn is the maximum value feature of the nth signal, x minn is the minimum feature of the nth signal, x Hn is the information entropy feature of the nth signal.

[0053] In this embodiment, the multi-feature set extraction method can be regarded as an expansion process of a feature space, and each feature set is regarded as a feature space and merged into a higher-dimensional space. In this new space, each feature is a dimension in the original feature space, and the model needs to learn the distribution of data in this expanded space. The above 15 feature sets provide a comprehensive statistical description of the vibration signal, which helps to capture the global characteristics of the signal. Compared with single feature recognition and signal decomposition, this method is more efficient and applicable, and is particularly suitable for complex and changeable urban environments. It can effectively deal with noise interference and signal variation, especially when the signal characteristics are not obvious or the signal types are diverse. It can provide more information to assist classification. Compared with single feature recognition and signal decomposition, this method is more efficient and applicable.

[0054] S4: Signal recognition: Use the CatBoost algorithm to perform signal recognition. Take the data after feature extraction and merging as input, build a CatBoost recognition model, perform training and prediction, and recognize the vibration signal.

[0055] In this implementation, feature and label data are established based on preprocessed and feature extracted data, and the data set is divided into a training set and a test set in a ratio of 8:2. The CatBoost classifier is trained on the training set, and the classification accuracy, recall rate, precision rate, and F1 score after training are verified by the test set. An improved whale optimization algorithm (NSC-WOA, an improved whale optimization algorithm based on nonlinear adaptive weights and sine and cosine operators) is introduced, and the parameters in the CatBoost model are tuned using the improved whale optimization algorithm. The F1 score evaluated on the verification data set is used as the fitness function, and the number of iterations, learning rate, tree depth, and L2 regularization coefficient are optimized. Finally, the optimized parameter combination is input into a CatBoost classifier model created using a CatBoost classifier (CatBoostClassifier) ​​to obtain the optimal model.

[0056] In this embodiment, the CatBoost algorithm is an ensemble learning algorithm based on the improved gradient boosting decision tree GBDT. The improved CatBoost algorithm has high accuracy and diagnostic ability in the case of small samples, and has some optimization strategies to better deal with the overfitting problem. The CatBoost algorithm has the following three outstanding advantages over the traditional algorithm GBDT:

[0057] 1. CatBoost introduces ordered boosting technology to solve the prediction shift problem of GBDT.

[0058] 2. CatBoost borrows the idea of ​​online learning to introduce the ordered target statistics (Ordered TS) method to directly convert categorical features into numerical statistics, adding direct support for categorical features. This method randomly arranges the data set and calculates the average label value of examples with the same category value, which is located before the given category value in the arrangement. The conversion calculation is based on the following formula

[0059]

[0060] in, Expressed as σ p The k-th dimension feature of the training sample, σ=(σ1,...,σ n ) is a permutation, j from 1 to p-1 represents the partial sample set before the pth sample in the permutation, Expressed as σ j The k-th dimension feature of the training samples, Expressed as σ j The labels of training samples, if the sample and samples belong to the same category in the k-th dimension feature, then If the sample and samples belong to different categories in the k-th dimension feature, then P is the added prior value, and a is its weight coefficient, which can reduce the noise problem of categories with low frequency of occurrence.

[0061] 3. CatBoost uses a symmetric binary tree as the base learner and performs the same splitting criteria on each layer, so it is not easy to overfit, and the stability and prediction speed of the model are also significantly improved. In order to pursue accuracy, more features are selected for training. In the embodiment, the feature extracts multiple feature data and merges them into a comprehensive feature vector as the input of the CatBoost symmetric decision tree.

[0062] S5: Parameter tuning: Use the improved whale optimization algorithm to tune the parameters in the CatBoost recognition model to obtain an optimized model and further improve the performance and accuracy of the model.

[0063] In this implementation, the NSC-WOA algorithm is an improved whale optimization algorithm based on nonlinear adaptive weights and sine and cosine operators, which introduces more randomness and diversity, and has the advantages of high global convergence and avoiding falling into local optimal values. The NSC-WOA algorithm (i.e., NSC-WOA) is used to search for optimization by adding two major changes to the WOA algorithm:

[0064] 1. Introduce nonlinear adaptive weights. Parameters can be adjusted adaptively and This allows the position update mechanism to find a transition between the development and exploration phases. The shrinking encirclement mechanism equation and the random search equation of the whale group are updated as follows:

[0065]

[0066] Each "whale" is a search agent. represents the position vector of the search agent, t is the current iteration number, is a randomly selected position from the search agent, C1 is a nonlinear adaptive weight, and a random number p is generated in [0, 1]. represents the position of the current optimal value (prey), and is a coefficient vector used to control the effect of the search agent's motion, is a random number in [-1, 1], represents the convergence factor, which decreases from 2 to 0 as the iteration progresses. is a random number in [0, 1], with the convergence factor The encirclement mechanism that shrinks as the size decreases can realize the continuous shrinking process of the whale.

[0067] After the introduction of the nonlinear adaptive weight C1, the equation of the spiral update mechanism is the following equation:

[0068]

[0069] Where b is a constant used to constrain the shape of the logarithmic spiral, l is a random number in [-1, 1] used to control the indentation effect of the search agent, Indicates the distance from the current optimal value.

[0070] 2. Introduce the sine and cosine operators. This operator combines the dynamic changes of the sine function and the smoothness of the cosine function to achieve richer search behaviors, produce more diverse update steps and directions, effectively avoid the algorithm from falling into local optimality, enhance the global search capability, and is controlled by only one variable, with low computational complexity, which is suitable for practical application environments with limited resources. The formula for updating the position of the search agent using the sine and cosine operators is as follows:

[0071]

[0072] Among them, r2, r3 and r4 are random numbers. represents the position vector of the search agent, t is the current iteration number, represents the position of the current optimal value (prey), C is a constant, and T is the maximum number of iterations. r1 controls the moving distance of the search agent; r2 controls the phase angle of the sine and cosine wave changes and controls the direction of the search agent; r3 is a random number on [0, 2]; r4 is a random number on [0, 1], which determines whether the position update method of the t+1 generation is the sine method or the cosine method.

[0073] like Figure 2 As shown in the figure, the process of improving the whale optimization algorithm is:

[0074] Step A1, initializing the location and parameters of the search agent;

[0075] Step A2: determine whether the termination condition is met. If yes, execute step A7; otherwise, update the nonlinear adaptive weight and generate a random number p;

[0076] Step A3, determine whether the random number p is less than 0.5, if yes, execute step A4; otherwise, execute the spiral update mechanism; execute step A5;

[0077] Step A4: Determine coefficient vector Is it less than 1? If so, adopt the shrinking and surrounding mechanism to update the position of the search agent; execute step A5; otherwise, adopt the random search mechanism to update the position of the search agent: execute step A5;

[0078] Step A5: introduce sine and cosine operators to search for the position of the agent and update the optimal solution;

[0079] Step A6: Determine whether the iteration is finished. If so, execute step A7; otherwise, return to step A2;

[0080] Step A7: Output the optimal solution and return to step A2.

[0081] In this implementation, the CatBoost recognition model is optimized using the improved whale optimization algorithm. For the purpose of multi-class classification, the model's precision and recall are considered, and the relationship between the two is attempted to be balanced. Therefore, the F1 score is used as the fitness function. The F1 score is the harmonic mean of the precision and recall, and a balance is achieved between the two. The calculation formula is:

[0082]

[0083] Among them, Precision refers to the proportion of samples correctly predicted as positive by the model to all predicted positive samples, and Recall refers to the proportion of samples correctly predicted as positive by the model to all actual positive samples. The average = 'macro' parameter of the f1_score function specifies the calculation of the macro average F1 score, so that the F1 score of each category has the same weight in the final result. Based on CatBoost's built-in mechanism for handling unbalanced data sets, the model is more suitable for situations where the number of class samples is unbalanced, and the overall performance of the model can be more comprehensively evaluated.

[0084] In this implementation, in order to ensure that the NSC-WOA algorithm can effectively search for the best CatBoost recognition model, the maximum number of iterations of the NSC-WOA algorithm is set to 200, the upper and lower bounds of the search space are 2 and 0 respectively, the moving distance of the search agent is a random number generated in [0, 2], and the moving direction of the search agent is a random number generated in [-2, 2]; the number of iterations of the CatBoost algorithm during each training is set to [100, 3000], the decision tree depth range is [1, 16], the learning rate search range is [0.001, 1], and the L2 regularization parameter search range is [1, 10]. Ensure that the best parameter combination can be searched, thereby achieving the CatBoost recognition model with the best recognition effect. At this time, the performance such as accuracy is optimal, but the iteration time becomes longer, and the parameter range can be appropriately narrowed according to the specific situation.

[0085] S6: Model performance evaluation: Based on the comparison between the predicted data and the real data, the performance of the proposed method is evaluated using confusion matrix diagrams and evaluation criteria such as accuracy, precision, recall, and F1 score.

[0086] The accuracy, precision, recall, and F1 score used in the model performance evaluation are verified by the following formula:

[0087]

[0088] Among them, TP is positive class judged as positive class, FP is negative class judged as positive class, FN is positive class judged as negative class, and TN is negative class judged as negative class. In order to take into account the precision (Precision) and recall (Recall) of the model and try to balance the relationship between the two, the F1 score is used as the fitness function. The F1 score is the harmonic mean of the precision rate and the recall rate, achieving a balance between the two.

[0089] like Figure 3 As shown in the figure, the overall steps of the NSC-WOA-CatBoost recognition model based on multi-feature set extraction are:

[0090] Firstly, the vibration signals collected by the Φ-OTDR in the dataset are labeled, and data differential preprocessing and multi-feature set extraction are performed, and the processed feature data are saved for subsequent training; secondly, the NSC-WOA improved whale optimization algorithm is introduced, and the F1 score evaluated on the validation dataset is used as the fitness function. The number of iterations, learning rate, tree depth, and L2 regularization coefficient are optimized, and the parameters are dynamically adjusted to update the best solution and perform iterative optimization; then, the dataset is divided into a training set and a test set in a ratio of 8:2, and the optimal parameter combination obtained by calling the feature data is input into the CatBoost classifier model to train the training set; finally, the classification accuracy, recall rate, precision rate, and F1 score of the trained model are verified by the test set to determine the optimal model performance.

[0091] Specific implementation method 2: Combination Figure 4 This embodiment is an example of the method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing described in the first embodiment. Experiments are conducted using the combination scheme given in the above embodiment to illustrate that the NSC-WOA-CatBoost recognition model based on multi-feature set extraction can improve the accuracy of signal recognition compared with the traditional machine learning model. The specific experimental results are shown in Figure 4 and Table 1, according to Figure 4 From the performance indicators detailed in Table 1, we can see that:

[0092] like Figure 4 As shown in the figure, the accuracy of each category is given in the next line of the predicted label (Predicted Label) for each label. The accuracy of each of the five categories of Background, Digging, Knocking, Shaking, and Walking are 99.8%, 98.0%, 99.4%, 100.0%, and 97.3%, respectively. The fluctuation range of the accuracy is between [97.3, 100.0], and the range is 2.7%. The results are relatively concentrated, indicating that the model is relatively stable.

[0093] Table 1 is a performance measurement table of the NSC-WOA-CatBoost recognition model based on multi-feature set extraction. As shown in Table 1, the precision, recall and F1 score of each category are relatively high, and the overall average accuracy of the model reaches 98.974554%, the average precision reaches 98.938042%, the average recall reaches 98.918448%, and the average F1 score reaches 98.927754%. The model is compared with historical research and found that the average accuracy rate is increased to 98.97%, and it can accurately identify five events, indicating that the performance of the model is very good and can handle multi-classification problems well.

[0094] Table 1

[0095]

[0096]

[0097] This specification describes different embodiments in a progressive manner, and emphasizes the differences between each embodiment and other embodiments. The present invention provides a method for identifying external force damage signals of optical cable vibration sources based on machine learning, namely, the NSC-WOA-CatBoost recognition model based on multi-feature set extraction. The model can accurately identify five events at the same time and has the following advantages in the complex background of urban areas:

[0098] Multi-feature set extraction: By integrating 15 features, it provides multi-faceted information of the signal and improves recognition accuracy. Compared with single feature recognition and signal decomposition, this method is more efficient and applicable. It helps to select features that are effective for all categories and reduce imbalance between categories. In complex urban environments, multi-feature set extraction enhances the adaptability and stability of the system.

[0099] Improved whale optimization algorithm NSC-WOA: Nonlinear adaptive weights allow the algorithm to dynamically adjust parameters during the search process to avoid falling into local optimality and find the global optimal solution, which helps to better adapt to the complexity and variability of urban environments. The introduction of sine and cosine operators can increase the diversity of the search space, and the implementation is simple and the computational cost is low, which helps to adjust the solution in complex urban environments.

[0100] CatBoost can process categorical features, support multi-task learning, and has relatively simple parameter adjustment. After being combined with the improved whale optimization algorithm, its accuracy is improved and its robustness is enhanced. It is more suitable for scenarios such as urban areas with a lot of non-numerical data, complex external sources, and changeable interference.

[0101] The NSC-WOA-CatBoost recognition model based on multiple feature extraction is applied to the urban context, and five common events in urban areas are identified with an average accuracy of 98.97%. It can also achieve the purpose of accurate identification of multiple categories, is more adaptable and stable, and has more practical value in urban areas with complex external sources.

[0102] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing, characterized by: The method is implemented by the following steps: Step 1: Detect the vibration signal of the optical cable caused by external force damage and construct a vibration signal data set; Step 2: preprocessing the vibration signal, then extracting a multi-feature set from the preprocessed vibration signal, merging the data after the multi-feature extraction, and dividing the data into a training set and a test set; Step 3: construct a CatBoost recognition model, and train the recognition model using a training set to obtain a trained CatBoost recognition model; Step 4: Use the improved whale optimization algorithm to tune the parameters in the trained CatBoost recognition model to obtain an optimized recognition model; input the test set into the optimized recognition model to realize the recognition of vibration signals; The improved whale optimization algorithm includes the introduction of nonlinear adaptive weights and sine and cosine operators. The specific process of obtaining the optimized recognition model is as follows: Step 41: Introduce and update nonlinear adaptive weights to generate a random number p; Step 42: Determine whether the random number p is less than 0.

5. If so, execute step 43; Otherwise, execute step 44; Step 43: Determine the coefficient vector Is it less than 1? If so, adopt the shrinking and surrounding mechanism to update the position of the search agent; execute steps 4 and 5; Otherwise, the position of the search agent is updated using a random search mechanism; and steps 4 and 5 are executed; Step 44: Use the spiral update mechanism to update the location of the search agent; execute step 45; Step 4 and 5: Introduce the sine and cosine operators to implement a secondary update of the position of the search agent, update the optimal solution, and determine whether the iteration is completed. If so, output the optimal solution; Otherwise, execute step 41 to step 45 again.

2. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 1 is characterized in that: After step 4, the performance of the optimized recognition model is also evaluated. According to the comparison between the predicted data and the real data, the performance of the recognition model is evaluated using the confusion matrix diagram and evaluation criteria such as accuracy, precision, recall and F1 score.

3. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 1, characterized in that: The specific process of extracting multiple feature sets from the vibration signal in step 2 is as follows: Step 21: Construct a multi-feature set; Set different feature sets, each feature set includes a statistical feature; Step 22: Standardize the features; Step 2: Perform feature fusion; Merge the features in different feature sets to obtain a fused multi-feature set.

4. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 3 is characterized in that: In step 2, multiple feature sets are extracted from the preprocessed vibration signal, including extracting statistical features of differential data; The statistical features include: maximum value X max , minimum value X min , peak-to-peak value X pk , mean X M , variance X Var , standard deviation X Std , Energy X E , rectified average value X RMV , RMS X RMS , peak factor X PF , margin factor X MF , form factor X WF , pulse factor X IF , kurtosis factor X KF and information entropy H(X).

5. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 1, characterized in that: In step 3, a CatBoost recognition model is constructed, which is used to convert categorical features into numerical statistics. In the formula, is the σth p The k-th dimension feature of the training sample, σ=(σ1,...,σ n ) is a permutation, j from 1 to p-1 represents the partial sample set before the pth sample in the permutation, is the σth j The k-th dimension feature of the training samples, is the σth j The labels of training samples; if the sample and samples belong to the same category in the k-th dimension feature, then If the sample and samples belong to different categories in the k-th dimension feature, then P is the added prior value, and a is its weight coefficient.

6. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 1, characterized in that: In step 43, the shrinking and surrounding mechanism is used to update the position of the search agent; the formula is as follows: In the formula, each whale acts as a search agent. is the position vector of the search agent, t is the current iteration number, C1 is the nonlinear adaptive weight, and a random number p is generated in [0, 1]. is the position of the current optimal value; and A vector of coefficients used to control the effect of the search agent's motion, is a random number in [-1, 1]; it can be expressed as follows: in, is the convergence factor, is a random number in [0, 1].

7. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 1, characterized in that: In step 43, a random search mechanism is used to update the position of the search agent. The formula is as follows: In the formula, is a randomly selected location from the search agent.

8. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 1, characterized in that: In step 44, the spiral update mechanism is used to update the position of the search agent, which is expressed as follows: Where b is a constant used to limit the shape of the logarithmic spiral, l is a random number in [-1, 1] used to control the indentation effect of the search agent, Indicates the distance from the current optimal value.

9. The method for identifying external force damage signals of optical cable vibration sources based on distributed optical fiber sensing according to claim 1, characterized in that: In steps 4 and 5, the formula for updating the position of the search agent using the sine and cosine operators is as follows: Where r2, r3 and r4 are random numbers, C is a constant, T is the maximum number of iterations, r1 controls the moving distance of the search agent; r2 controls the phase angle of the sine and cosine waves, which controls the direction of the search agent; r3 is a random number in [0, 2]; r4 is a random number in [0, 1], which determines whether the position update method of the t+1 generation is the sine method or the cosine method.

Images