Perimeter vibration signal classification method and system based on vibration fingerprint recognition, and medium

Abstract

This invention discloses a method, system, and storage medium for classifying perimeter vibration signals based on vibration fingerprint recognition. The method employs a high-sampling-rate multi-axis vibration sensor network to collect time-series data of complete perimeter vibration events. It constructs complete vibration event samples based on five-element event identifiers, avoiding the feature fragmentation problem of traditional window truncation. By independently extracting multi-dimensional time-frequency features from the sensor's physical axes, and combining this with a median-interquartile range robust normalization algorithm, feature optimization and fusion are completed to construct a highly recognizable vibration fingerprint feature vector. During model training, a hybrid sampling strategy combining synthetic minority class oversampling and TomekLinks boundary cleanup is used to address sample imbalance. This is further enhanced by hierarchical cross-validation, hyperparameter global search, early stopping mechanism, and class weight optimization strategies to achieve optimal training of the gradient-boosted multi-classification model. Finally, the model is solidified to achieve accurate classification of vibration events, outputting vibration type labels and confidence probabilities.

Description

Technical Field

[0001] This invention relates to the field of perimeter security monitoring technology, specifically to a method, system, and medium for classifying perimeter vibration signals based on vibration fingerprint recognition. Background Technology

[0002] Perimeter security is a core component of security protection for key areas such as airports, factories, industrial parks, borders, and warehouses. Perimeter monitoring systems based on vibration sensor networks have become mainstream security equipment due to their advantages of convenient deployment, strong anti-interference capabilities, and all-weather operation. These systems determine the presence of abnormal intrusion behavior within the area by collecting vibration signals from perimeter protection facilities. However, existing technologies suffer from several core flaws in practical applications, severely limiting the accuracy and practicality of security monitoring. Specific problems are as follows: 1. Vibration identification is limited in scope and cannot differentiate event types. Existing traditional vibration monitoring solutions only have a binary classification capability of "abnormal / normal," which can only identify whether vibration has occurred, but cannot differentiate the specific type of vibration event. Security personnel cannot distinguish whether the vibration originates from intrusive acts such as climbing, knocking, or cutting, or from non-threatening disturbances such as wind and rain, vehicle traffic, or environmental construction, resulting in a lack of effective data support for security decisions.

[0003] 2. Frequent false alarms due to non-threat events lead to low security efficiency. Non-intrusive vibration events such as vibrations from natural environments like wind and rain, vibrations from passing vehicles on roads, and vibrations from routine construction in the surrounding area continuously trigger alarms in traditional security systems. However, current technology lacks the ability to automatically distinguish and filter between threatening and non-threatening vibrations. A large number of invalid alarms consume security manpower and resources, significantly reducing the effectiveness and reliability of security early warning systems.

[0004] 3. Imbalanced distribution of multi-class samples, resulting in extremely poor accuracy in minority class identification. In perimeter vibration scenarios, there are a massive number of common vibration samples such as wind and rain, and vehicle traffic, while the number of rare vibration samples such as intrusions such as climbing, cutting, and intentional knocking is extremely small. The training dataset suffers from a severe class imbalance problem. Traditional classification models tend to over-bias the majority class samples during training, directly leading to extremely low accuracy in identifying rare intrusion events and posing a serious risk of missed security alerts.

[0005] 4. Defects in sampling processing methods, resulting in loss of global vibration features. Existing technologies generally use fixed-size sliding windows to truncate sampling data to construct training samples, which fragments the same continuous vibration event, destroying the temporal continuity and global pattern features of the vibration signal. The model cannot learn the complete vibration event pattern, further reducing the accuracy of classification and recognition.

[0006] 5. Low sampling frequency leads to inherent limitations in frequency domain recognition. Different vibration events exhibit significant frequency characteristics: climbing vibrations are below 5Hz, impact vibrations are 20-40Hz, and cutting vibrations can reach over 50Hz. Existing equipment generally operates with a low sampling rate below 100Hz. According to the Nyquist sampling theorem, its effective recognition frequency limit is less than 50Hz, failing to fully cover various high-frequency intrusion vibration signals. This results in insufficient frequency domain resolution, difficulty in distinguishing between high and low frequency vibration events, and extremely low recognition error tolerance.

[0007] In summary, there is an urgent need for a perimeter vibration signal classification scheme that can take into account global feature extraction, sample balance optimization, accurate identification of multiple types, and adaptability to vibration signals across the entire frequency band, thereby addressing the various shortcomings of existing technologies. Summary of the Invention

[0008] This invention aims to solve the technical problems of existing perimeter vibration identification technology, such as weak vibration type differentiation ability, high false alarm rate of non-threat events, missed detection of rare events due to sample imbalance, loss of features in sliding window sampling, and limited frequency domain recognition due to low sampling rate. It provides a perimeter vibration signal classification method, system and storage medium based on vibration fingerprint recognition, which can realize high-precision and fully automatic classification and identification of multiple types of perimeter vibration events.

[0009] To solve the above technical problems, the present invention provides a method for classifying perimeter vibration signals based on vibration fingerprint recognition, the method comprising the following steps: S1, through a multi-axis vibration sensor network with a preset sampling frequency, collect time-series sampling data corresponding to a complete vibration event within the perimeter protection area. The time-series sampling data includes all sampling points of multiple sensors within the complete event time period. S2, perform zero-bias correction on the vibration event time series data, group the data according to the combination of sensor physical axis and feature type, independently extract the time domain statistical features and frequency domain features of each physical axis, and perform robust standardization on the extracted features according to the grouping, and concatenate the standardized features of all physical axes into a vibration fingerprint feature vector. S3, input the vibration fingerprint feature vector into the pre-trained gradient boosting multi-classification model for inference, output the classification label of the vibration type to which the vibration event belongs and its confidence probability, and complete the classification and identification of the perimeter vibration event; The training process of the gradient boosting multi-class classification model includes: Step A: Obtain a set of vibration event samples with labeled vibration types, and extract the vibration fingerprint feature vector for each sample; Step B involves applying a hybrid sampling strategy of oversampling and boundary cleanup to process the vibration fingerprint feature vector set to address its imbalance. Step C: Use hierarchical cross-validation combined with parameter search to fine-tune the hyperparameters of the gradient boosting multi-classification model, and determine the optimal hyperparameter combination based on the cross-validation accuracy. Step D: Retrain the gradient boosting multi-classification model on the complete balanced feature set using the optimal hyperparameter combination.

[0010] Furthermore, in step A, a complete event time period is used as a single sample. Specifically, all sampling points collected by the same sensor under the event identifier information composed of device identifier, communication interface identifier, sensor number, event start time and end time are grouped into one sample. The five pieces of information, namely device identifier, communication interface identifier, sensor number, event start time and end time, form a five-element identifier, which is a uniquely determined vibration sample. After extracting features independently for each physical axis of the sensor, the feature vectors of all physical axes are concatenated to form the final vibration fingerprint feature vector.

[0011] Furthermore, the independent extraction of time-domain statistical features and frequency-domain features for each physical axis specifically involves: for each sensor physical axis, independently extracting multiple time-domain statistics and frequency-domain spectral descriptors from the complete sampling data of the vibration event; The time-domain statistics and frequency-domain spectral descriptions of each physical axis are concatenated into a unified vibration fingerprint feature vector after the grouping robust normalization process is completed.

[0012] Furthermore, the robust standardization process by grouping specifically involves using the median of the feature values ​​within each group as the location parameter and the interquartile range of the feature values ​​within each group as the scale parameter, and then using the median and the interquartile range for robust standardization. When the absolute value of the interquartile range of a certain group is less than the preset protection threshold, the interquartile range of that group is forcibly set to the preset protection threshold.

[0013] Furthermore, the hybrid sampling strategy of oversampling and boundary cleanup in step B is as follows: First, synthetic minority class oversampling technology is used to interpolate and generate synthetic samples in the feature space between minority class samples and their nearest neighbors to increase the number of minority class samples; Then, the TomekLinks algorithm is used to detect nearest neighbor pairs of samples in the sample space that belong to different categories, and these pairs are removed from the sample set to clean up the category boundaries.

[0014] Furthermore, the search space of hyperparameters in step C specifically includes: number of leaf nodes, maximum tree depth, learning rate, number of estimators, minimum number of child node samples, minimum split gain value, L1 regularization coefficient, L2 regularization coefficient, subsampling ratio, column sampling ratio, and path smoothing coefficient.

[0015] Furthermore, step C further includes: during the parameter search process, for each group of candidate hyperparameters, a validation subset is used as an early stopping validation set, and training is terminated early when the multi-class classification loss function of the validation set no longer decreases for a specified number of consecutive rounds.

[0016] Furthermore, the gradient boosting multi-class classification model training process also introduces a class weight mechanism, specifically: calculating a balanced weight value based on the distribution of each class in the training samples, so that the class with fewer samples receives a higher loss weight than the class with more samples.

[0017] This invention discloses a perimeter vibration signal classification system based on vibration fingerprint recognition. The system is used to implement the aforementioned classification method, and the system includes: The data acquisition module is used to acquire time-series data of vibration events within a complete event time period collected by the vibration sensor network; The feature extraction module is used to perform zero-bias correction on the vibration event time series data, independently extract time-domain statistical features and frequency-domain features according to the sensor physical axis, and perform robust standardization by grouping according to physical axis-feature type. After standardization, the features are concatenated into a vibration fingerprint feature vector. The model training module is used to execute the training process of the gradient boosting multi-class model described in steps A to D. The model training module has a built-in hybrid sampling equalization algorithm, hierarchical cross-validation and hyperparameter optimization program, which is used to complete the optimization of the labeled dataset, iterative training of model parameters and fixation of the optimal model. The fingerprint classification module includes a pre-trained gradient boosting multi-classification model, which is used to infer the vibration fingerprint feature vector, complete the vibration event category inference, and output the vibration type label and confidence probability.

[0018] The present invention provides a computer-readable storage medium storing a computer-executable program, which, when executed by a processor, implements all the steps of the perimeter vibration signal classification method based on vibration fingerprint recognition.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves precise subdivision of multiple vibration types, solving the problem of single identification dimensions. Based on the full coverage of high and low frequency vibration signals collected by the sampling rate, this invention constructs a unique vibration fingerprint through the fusion of multi-dimensional features in the time and frequency domains. Relying on a gradient-enhanced multi-classification model, it can accurately distinguish intrusion vibrations such as human climbing, knocking, and cutting, as well as environmental interference vibrations such as wind, rain, vehicles, and construction. It breaks through the limitations of traditional binary classification and provides accurate event type basis for security decision-making.

[0020] 2. This invention reduces the false alarm rate for non-threatening events and improves security efficiency. By autonomously learning the fingerprint characteristics of different vibration events through a model, this invention can automatically distinguish between threatening and non-threatening vibration events, accurately filter out invalid interference alarms, solve the problem of frequent false alarms in traditional systems, and greatly reduce the manpower required for security.

[0021] 3. This invention optimizes the sample imbalance problem and prevents the underreporting of rare events. This invention employs a dual optimization strategy of "oversampling supplementation + boundary cleanup," which solves the problem of insufficient minority class samples and eliminates interference from class boundaries. Combined with an adaptive class weighting mechanism and hierarchical cross-validation, it significantly improves the accuracy of identifying rare intrusion vibration events and addresses the technical pain point of the model being biased towards the majority class.

[0022] 4. This invention preserves global vibration features and avoids feature fragmentation and loss. This invention abandons the traditional sliding window truncation sampling method, using complete vibration events as sample units and dividing independent samples based on a five-element unique identifier. This fully preserves the temporal continuity and global pattern features of vibration events, significantly improving the stability and accuracy of model classification and recognition.

[0023] 5. Full-band signal coverage, adaptable to various vibration scenarios. Adopting a sampling rate of no less than 100Hz, it meets the requirements of the Nyquist sampling theorem, effectively identifying all vibration types covering the frequency domain from 0-50Hz. This overcomes the shortcoming of traditional low-sampling-rate equipment in being unable to identify high-frequency intrusive vibrations, resulting in stronger environmental adaptability. Attached Figure Description

[0024] Figure 1 This is an overall flowchart of the vibration fingerprint recognition method of the present invention.

[0025] Figure 2-4 This is a schematic diagram of the hybrid sampling strategy of oversampling and boundary cleanup of the present invention.

[0026] Figure 5 This is a flowchart of the gradient boosting multi-classification model training process of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the scope of protection of the present invention. Obviously, the embodiments described in this invention are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] Example 1: The specific structure of the present invention is as follows: Please refer to the appendix. Figure 1 A perimeter vibration signal classification method based on vibration fingerprint recognition: This embodiment discloses the core classification method of the present invention, which fully realizes the entire process from data collection, feature optimization, model training to intelligent classification, and accurately solves various defects of the prior art. The specific steps are as follows: The first step involves comprehensive high-fidelity vibration data acquisition, establishing a multi-axis vibration sensor network. All sensors are uniformly configured with a sampling frequency of no less than 100Hz, ensuring a Nyquist frequency of 50Hz, fully covering the frequency range of all types of vibration signals, including low-frequency climbing vibration, mid-frequency impact vibration, and high-frequency cutting vibration. The sensor network is deployed throughout the perimeter protection area, continuously acquiring vibration time-series data in real time.

[0030] Simultaneously, a five-element unique identifier rule is adopted to divide independent vibration samples. Using a combined identifier consisting of the device's unique code, communication interface number, sensor number, vibration event start time, and vibration event end time, all consecutive sampling points collected by the same sensor under the same five-element identifier are defined as a complete vibration event sample. This abandons the traditional sliding window truncation method, ensuring that each sample corresponds to a complete and continuous vibration event, completely avoiding the fragmentation of temporal features and fully preserving the global vibration pattern. The five-element identifier, composed of the unique code, communication interface number, sensor number, vibration event start time, and vibration event end time, uniquely identifies the vibration sample. Step 2: Data Preprocessing and Multidimensional Feature Extraction First, all collected vibration time-series data are zero-biased to eliminate baseline drift and zero-point offset errors caused by long-term operation of the sensor hardware, thus ensuring the accuracy of the original data.

[0031] Subsequently, the data is grouped according to the combination of sensor physical axes (X1 / X2 / X3 / Y1 / Y2 / Y3 / Z1 / Z2 / Z3) and feature types (time-domain features, frequency-domain features). Feature extraction is performed independently for each physical axis to avoid feature loss caused by multi-axis data mixing. Among them, time-domain statistical features extract 12 core statistical quantities such as mean, root mean square, peak value, kurtosis, skewness, waveform factor, and impulse factor to comprehensively characterize the time-domain fluctuation characteristics of vibration signals. Frequency-domain features extract 10 spectral descriptive quantities such as total spectral energy, fundamental frequency, fundamental frequency amplitude, high-frequency energy proportion, and spectral entropy to accurately distinguish the frequency domain distribution differences of different vibrations.

[0032] Step 3: Robust standardization of groups (to address issues such as data scale differences and outlier interference): For each group of feature data (single physical axis + single feature type), a robust normalization algorithm is used to optimize the data, avoiding the shortcomings of traditional mean-variance normalization which is easily affected by outliers. Specifically, the calculation method is as follows: the median of the grouped feature data is used as the position correction parameter to eliminate the influence of extreme outlier offsets; the interquartile range of the grouped data is used as the scaling parameter to complete feature normalization.

[0033] Simultaneously, a data protection threshold is set, with a preset threshold of 0.1. When the absolute value of the interquartile range of a certain group is less than this threshold, the interquartile range of that group is forcibly assigned a value of 0.1, completely avoiding the problem of standardized numerical divergence caused by the denominator approaching 0, and ensuring the stability and validity of all feature data. After standardization, all standardized feature vectors of all physical axes are sequentially concatenated according to their dimensions to form a unique vibration fingerprint feature vector with uniform dimensions and high recognizability.

[0034] Step 4: Imbalanced Model Dataset Optimization (Addressing the issues of scarce minority class samples and mixed boundary conditions): In the perimeter data collected in this embodiment, conventional samples such as wind and rain, and vehicle traffic account for over 80%, while rare intrusion samples such as climbing and cutting account for less than 20%, indicating a severe imbalance in the dataset. This invention employs a hybrid sampling strategy for step-by-step optimization: 1. Minority oversampling supplementation: The SMOTE algorithm is used to traverse all minority class intrusion samples, select the three nearest neighbor samples of each minority class sample, and perform linear interpolation in the feature space to automatically generate synthetic rare samples, unifying the number of samples of each class to the same order of magnitude and making up for the data shortage of rare samples. 2. Category Boundary Cleaning: The Tomek Links algorithm is used to traverse and detect the supplemented dataset, and to filter out mixed sample pairs that are nearest neighbors in the feature space but belong to different categories. All sample pairs with blurred class boundaries are removed, and the interference of class overlap is thoroughly cleaned up, so that the decision boundaries of each category sample are clearly distinguishable, which greatly improves the classification accuracy of the model.

[0035] Specifically, such as Figure 2-4 As shown, the hybrid sampling process consists of three steps: (a) the majority class in the original data is much larger than the minority class, which leads to model skewness when trained directly; (b) linear interpolation is performed between minority class samples to generate synthetic samples so that the number of samples in each class tends to be balanced; (c) the nearest neighbor samples at the class boundaries that belong to different classes are detected and removed to clean up the decision boundaries and reduce inter-class overlap.

[0036] Step 5: Gradient boosting for optimal parameter training of the multi-class classification model (addressing issues such as inefficiency of manual tuning, model overfitting, and insufficient minority class weights). This embodiment uses the LightGBM gradient boosting multi-class classification model, combined with multiple optimization strategies to complete model training: 1. Hyperparameter global search: The preset hyperparameter search space covers 11 core parameters, including maximum tree depth (3-10), learning rate (0.01-0.03), number of estimators (500, 800, 1000), L1 / L2 regularization coefficient, and subsampling ratio. It uses a combination of grid search and random search to traverse all parameter combinations. 2. Hierarchical cross-validation: Five-fold hierarchical cross-validation is adopted to ensure that the class distribution of each fold of the training set and validation set is completely consistent with the original dataset, avoiding training bias caused by uneven data partitioning. The average cross-validation accuracy is used as the criterion for judging the quality of parameters. 3. Early stopping mechanism: For each group of candidate hyperparameters, the number of early stopping rounds is set to 50 rounds. If the multi-class loss value on the validation set does not decrease for 50 consecutive rounds, the current training is terminated immediately to effectively prevent model overfitting. 4. Adaptive Class Weights: The loss weights are calculated inversely to the number of samples in each class. Higher loss weights are set for rare invasion classes such as climbing and cutting, forcing the model to focus on learning the features of minority class samples and solving the problem of the model being biased towards the majority class. 5. Model solidification: Select the hyperparameter combination with the highest cross-validation accuracy as the optimal parameters, retrain on a complete balanced dataset using these parameters, and finally solidify to obtain the optimal gradient boosting multi-classification model.

[0037] Step 6: Real-time vibration event classification and reasoning: In the actual deployment and inference phase, following the same process as the training phase—collection, correction, feature extraction, standardization, and feature splicing—vibration fingerprint feature vectors are constructed from the real-time collected vibration event data. These vectors are then input into the solidified optimal model for inference. The model automatically outputs the specific type of the current vibration event (intrusion, wind and rain interference, vehicle passage, construction disturbance, etc.) and its corresponding confidence probability, achieving accurate classification and identification, enabling precise early warning of threat events and automatic filtering of invalid interference.

[0038] Example 2: Perimeter vibration signal classification system based on vibration fingerprint recognition: This embodiment is a dedicated system for implementing the classification method described in Embodiment 1. It features a modular design, independent operation, and can be directly embedded into various perimeter security platforms. The specific module functions are as follows: 1. Data Acquisition Module: Connects to an external multi-axis vibration sensor network to receive 100Hz high sampling rate vibration time-series data in real time. It automatically divides complete vibration event samples according to the five-element identifier, filters invalid empty data, and ensures the integrity and standardization of input data.

[0039] 2. Feature Extraction Module: Built-in zero-bias correction algorithm, multi-physical axis time-frequency feature extraction program, and grouped robust normalization program, which can automatically complete raw data preprocessing, multi-dimensional feature extraction, feature optimization and fingerprint vector concatenation without manual intervention.

[0040] 3. Model Training Module: Integrates a hybrid sampling equalization algorithm, a 5-fold hierarchical cross-validation program, a hyperparameter global search program, an early stopping control program, and a class weight calculation program. It supports automatic dataset optimization, iterative model training, optimal parameter selection, and model solidification, and can autonomously iterate and optimize vibration data in different scenarios.

[0041] 4. Fingerprint Classification Module: Equipped with a solidified optimal gradient boosting multi-classification model, it receives vibration fingerprint feature vectors in real time, completes inference calculations at high speed, and outputs standardized classification results and confidence parameters. It supports integration with security early warning platforms to achieve automatic alarms and data uploads.

[0042] Specifically, such as Figure 5 As shown, the training process includes: hierarchical division of the training set and test set, mixed sampling of oversampling and boundary cleaning to balance the training set, hierarchical K-fold cross-validation with early stopping to prevent overfitting, random search of the hyperparameter space to determine the optimal combination with the highest average accuracy, full retraining and evaluation of model performance.

[0043] Example 3: Computer-readable storage medium: This embodiment provides a computer-readable storage medium containing a computer program that can be called and executed by a processor, server, or embedded device. When the program runs, it can completely execute all the steps of the perimeter vibration signal classification method based on vibration fingerprint recognition in Embodiment 1. It can be deployed on various hardware terminals such as security hosts, edge computing devices, and cloud servers, and has extremely strong adaptability.

[0044] In summary, this invention addresses the five core shortcomings of existing perimeter vibration recognition technologies by innovatively proposing a multi-class intelligent recognition scheme based on multi-axis vibration fingerprint recognition. It upgrades the entire process from data acquisition, sample construction, feature optimization, data balancing, model training, to inference applications. This invention ensures full coverage of vibration signals across the entire frequency band through 100Hz high sampling rate acquisition, and solves the feature fragmentation problem of traditional window truncation by relying on a complete event sample construction mode. It constructs a highly recognizable vibration fingerprint through multi-dimensional time-frequency feature fusion and robust standardization algorithms, achieving accurate differentiation of multiple types of vibration events. Through multiple optimizations including hybrid sampling strategies, hierarchical cross-validation, early stopping mechanisms, and adaptive class weights, it thoroughly solves problems such as low accuracy in rare event recognition, model bias, and overfitting caused by imbalanced multi-class samples, significantly improving the recognition accuracy of rare intrusion vibration events and effectively filtering invalid alarms caused by environmental interference. Meanwhile, this invention adopts a modular and automated design, requiring no manual intervention throughout the entire process. It can be flexibly deployed in various perimeter security systems, exhibiting strong compatibility and practicality. It effectively compensates for the shortcomings of traditional perimeter vibration monitoring technology, significantly improves the intelligent and precise monitoring capabilities of perimeter security systems, and possesses extremely high engineering application value and promotion prospects.

[0045] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for classifying perimeter vibration signals based on vibration fingerprint recognition, characterized in that, The method includes the following steps: S1, through a multi-axis vibration sensor network with a preset sampling frequency, collect time-series sampling data corresponding to a complete vibration event within the perimeter protection area. The time-series sampling data includes all sampling points of multiple sensors within the complete event time period. S2, perform zero-bias correction on the vibration event time series data, group the data according to the combination of sensor physical axis and feature type, independently extract the time domain statistical features and frequency domain features of each physical axis, and perform robust standardization on the extracted features according to the grouping, and concatenate the standardized features of all physical axes into a vibration fingerprint feature vector. S3, input the vibration fingerprint feature vector into the pre-trained gradient boosting multi-classification model for inference, output the classification label of the vibration type to which the vibration event belongs and its confidence probability, and complete the classification and identification of the perimeter vibration event; The training process of the gradient boosting multi-class classification model includes: Step A: Obtain a set of vibration event samples with labeled vibration types, and extract the vibration fingerprint feature vector for each sample; Step B involves applying a hybrid sampling strategy of oversampling and boundary cleanup to process the vibration fingerprint feature vector set to address its imbalance. Step C: Use hierarchical cross-validation combined with parameter search to fine-tune the hyperparameters of the gradient boosting multi-classification model, and determine the optimal hyperparameter combination based on the cross-validation accuracy. Step D: Retrain the gradient boosting multi-classification model on the complete balanced feature set using the optimal hyperparameter combination.

2. The method for classifying perimeter vibration signals based on vibration fingerprint recognition according to claim 1, characterized in that, In step A, a complete event time period is used as a single sample. Specifically, all sampling points collected by the same sensor under the event identifier information composed of device identifier, communication interface identifier, sensor number, event start time and end time are grouped into one sample. The five pieces of information, namely device identifier, communication interface identifier, sensor number, event start time and end time, form a five-element identifier, which is a uniquely determined vibration sample. After extracting features independently for each sensor physical axis, the feature vectors of all physical axes are concatenated to form the final vibration fingerprint feature vector.

3. The method for classifying perimeter vibration signals based on vibration fingerprint recognition according to claim 1, characterized in that, The independent extraction of time-domain statistical features and frequency-domain features for each physical axis specifically involves: for each sensor physical axis, independently extracting multiple time-domain statistics and frequency-domain spectral descriptors from the complete sampling data of the vibration event; The time-domain statistics and frequency-domain spectral descriptions of each physical axis are concatenated into a unified vibration fingerprint feature vector after the grouping robust normalization process is completed.

4. The method for classifying perimeter vibration signals based on vibration fingerprint recognition according to claim 1, characterized in that, The robust standardization process by grouping is specifically as follows: using the median of the feature values ​​within each group as the location parameter and the interquartile range of the feature values ​​within each group as the scale parameter, robust standardization is performed using the median and the interquartile range. When the absolute value of the interquartile range of a certain group is less than the preset protection threshold, the interquartile range of that group is forcibly set to the preset protection threshold.

5. The method for classifying perimeter vibration signals based on vibration fingerprint recognition according to claim 1, characterized in that, The hybrid sampling strategy of oversampling and boundary cleanup in step B is as follows: First, synthetic minority class oversampling technique is used to interpolate and generate synthetic samples in the feature space between minority class samples and their nearest neighbors to increase the number of minority class samples. Then, the TomekLinks algorithm is used to detect nearest neighbor pairs of samples in the sample space that belong to different categories, and these pairs are removed from the sample set to clean up the category boundaries.

6. The method for classifying perimeter vibration signals based on vibration fingerprint recognition according to claim 1, characterized in that, The search space of hyperparameters in step C specifically includes: number of leaf nodes, maximum tree depth, learning rate, number of estimators, minimum number of child node samples, minimum split gain, L1 regularization coefficient, L2 regularization coefficient, subsampling ratio, column sampling ratio, and path smoothing coefficient.

7. The method for classifying perimeter vibration signals based on vibration fingerprint recognition according to claim 1, characterized in that, Step C further includes: during the parameter search process, for each group of candidate hyperparameters, a validation subset is used as an early stopping validation set, and training is terminated early when the multi-class classification loss function of the validation set no longer decreases for a specified number of consecutive rounds.

8. The method for classifying perimeter vibration signals based on vibration fingerprint recognition according to claim 1, characterized in that, The gradient boosting multi-class classification model training process also introduces a class weight mechanism, which specifically calculates a balanced weight value based on the distribution of each class in the training samples, so that the class with fewer samples receives a higher loss weight than the class with more samples.

9. A perimeter vibration signal classification system based on vibration fingerprint recognition, characterized in that, This system is used to implement the classification method according to any one of claims 1-8, and the system comprises: The data acquisition module is used to acquire time-series data of vibration events within a complete event time period collected by the vibration sensor network; The feature extraction module is used to perform zero-bias correction on the vibration event time series data, independently extract time-domain statistical features and frequency-domain features according to the sensor physical axis, and perform robust standardization by grouping according to physical axis-feature type. After standardization, the features are concatenated into a vibration fingerprint feature vector. The model training module is used to execute the training process of the gradient boosting multi-classification model described in steps A to D of claim 1. The model training module has a built-in hybrid sampling equalization algorithm, hierarchical cross-validation and hyperparameter optimization program, which is used to complete the optimization of the labeled dataset, iterative training of model parameters and fixation of the optimal model. The fingerprint classification module includes a pre-trained gradient boosting multi-classification model, which is used to infer the vibration fingerprint feature vector, complete the vibration event category inference, and output the vibration type label and confidence probability.

10. A computer-readable storage medium storing a computer-executable program, characterized in that, When the computer executable program is called and executed by the processor, it implements all the steps of the perimeter vibration signal classification method based on vibration fingerprint recognition as described in any one of claims 1 to 8.

Images