DAS external damage prevention method and system based on deep learning

Abstract

The invention provides a DAS external damage prevention method and system based on deep learning, and the method comprises the steps: collecting an environment vibration signal through a DAS, carrying out the preprocessing of the environment vibration signal, and building a time period feature database; based on the time period feature database, training a special identification model of each monitoring time period by adopting a CNN-LSTM hybrid network to obtain a multi-time period initial sub-model; arranging the multi-period initial sub-models into a preset scene model through model similarity analysis, establishing a dynamic updating mechanism, and optimizing model parameters by adopting updated data to obtain an optimized multi-period model; and a current time period model in the optimized multi-time period model is monitored and matched in real time, a monitoring alarm result is output by utilizing multi-dimensional judgment, and a model library is continuously updated based on the monitoring alarm result.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to a deep learning-based DAS (Disruption and Protection System) method and system for preventing external damage. Background Technology

[0002] Distributed Acoustic Sensing (DAS) technology uses optical fibers as continuous sensors to monitor environmental vibrations in real time by detecting phase changes in optical signals. It is widely used in pipeline safety, perimeter protection, and subway damage prevention. Traditional DAS systems rely on human experience or simple threshold judgments, making it difficult to cope with complex environmental noise (such as vehicle traffic and weather interference), resulting in high false alarm rates and delayed response.

[0003] In recent years, artificial intelligence (AI) technology has been introduced into DAS data analysis, using deep learning models (such as CNN and LSTM) to improve event recognition capabilities. However, existing AI solutions still face two major challenges: first, dynamic environmental changes (such as day-night differences and traffic peaks) lead to insufficient model generalization ability; second, the high real-time computing requirements make it difficult to run efficiently on resource-constrained edge devices.

[0004] Existing DAS (Disruption-Avoidance System) technologies for preventing external damage mainly employ two types of solutions: (1) Fixed threshold detection: Based on the signal amplitude or frequency, a static threshold is set. It is simple but has poor adaptability and cannot distinguish between real threats (such as mechanical excavation) and normal interference (such as subway vibration). The false alarm rate exceeds 30%.

[0005] (2) Single AI model: Using CNN or LSTM to train a global model, although it improves the recognition rate to a certain extent, it does not take into account the differences in time periods (such as the different signal characteristics during the day / night), which causes the accuracy of the model to drop sharply in dynamic scenarios (such as misidentifying animal activities as intrusion at night).

[0006] It can be seen that the following problems exist: (1) Poor environmental adaptability: Static models cannot dynamically adapt to changing factors such as time periods and weather; (2) Low computational efficiency: Parallel processing of data throughout the day leads to resource waste and restricts edge deployment; (3) The model is seriously redundant: similar scenarios (such as adjacent time periods) are repeatedly modeled, and the common features are not used to optimize storage and computation.

[0007] Therefore, there is an urgent need for an intelligent method to prevent external damage that combines high precision, low latency, and adaptability to environmental changes. Summary of the Invention

[0008] This invention provides a deep learning-based DAS (Disruption-Oriented Assist System) method and system for preventing external damage, in order to address the shortcomings of existing technologies.

[0009] In a first aspect, the present invention provides a deep learning-based DAS (Data Access System) anti-tampering method, comprising: Environmental vibration signals are acquired through DAS, and the environmental vibration signals are preprocessed to establish a time-period feature database. Based on the time period feature database, a dedicated recognition model for each monitoring time period is trained using a CNN-LSTM hybrid network to obtain multiple initial sub-models for different time periods; The initial sub-models for multiple time periods are organized into a preset scenario model through model similarity analysis, a dynamic update mechanism is established, and the model parameters are optimized using updated data to obtain an optimized multi-time period model. By real-time monitoring and matching of the current time period model in the optimized multi-time period model, multi-dimensional judgment is used to output monitoring and alarm results, and the model library is continuously updated based on the monitoring and alarm results.

[0010] According to the present invention, a deep learning-based DAS (Disruptive Amplifier) ​​method for preventing external damage involves acquiring environmental vibration signals via DAS, preprocessing the environmental vibration signals, and establishing a time-period feature database, including: Multiple environmental vibration signals were collected via DAS within a preset time period; Multiple environmental vibration signals are sequentially subjected to wavelet transform for noise filtering, signal normalization, and short-time Fourier transform to convert the time-domain signals into time-frequency feature matrices. The time-frequency feature matrix is ​​divided into time periods by using equal-interval durations as individual monitoring units to form multiple continuous monitoring time periods. Multidimensional feature vectors are extracted from these multiple continuous monitoring time periods to construct the time period feature database.

[0011] According to the deep learning-based DAS (Disruptive Amplifier) ​​method for preventing external damage provided by the present invention, after acquiring environmental vibration signals through DAS, preprocessing the environmental vibration signals, and establishing a time-period feature database, the method further includes: An automatic quality detection mechanism is adopted to evaluate preset signal indicators in real time, remove abnormal data segments from the time period feature database, and trigger an automatic data repair process or an automatic data re-collection process.

[0012] According to the present invention, a deep learning-based DAS (Disruption-Oriented Assault System) method for preventing external attacks is provided. Based on the time-period feature database, a dedicated identification model for each monitoring time period is trained using a CNN-LSTM hybrid network to obtain multiple initial sub-models for different time periods, including: Input feature processing is performed on the time period feature database to extract the number of sampling points of a single monitoring unit and the number of frequency bands divided by short-time Fourier transform in the corresponding time-frequency feature matrix; A five-layer two-dimensional convolutional module is used to extract convolutional features from the input features, and a multi-dimensional feature vector is output. The time-period feature database is processed using a bidirectional LSTM network to obtain continuous time-period features; An attention mechanism is introduced into the continuous features of the time period to enhance key features and obtain significant vibration features; The softmax function is used in the output layer for decision classification to obtain the multi-time initial sub-model.

[0013] According to the present invention, a deep learning-based DAS (Data Access System) anti-external attack method is provided, which organizes the multi-time-period initial sub-models into a preset scenario model through model similarity analysis, establishes a dynamic update mechanism, and optimizes the model parameters using updated data to obtain an optimized multi-time-period model, including: For each initial sub-model, extract the weight matrix of the last hidden feature transformation to construct the model feature space. The dimensions of the weight matrix include the dimension of the LSTM hidden units and the dimension of the number of frequency bands divided by the short-time Fourier transform. Based on the consistency of responses of each model to the same time-frequency features, the similarity between models is measured by projection distance. A bottom-up aggregated hierarchical clustering and control clustering method is adopted to construct a hierarchical clustering tree based on the similarity matrix; For each cluster in the hierarchical clustering tree, a central model is selected; Obtain the original model for any time period, and establish a mapping function from the original time period to the representative model based on the central model; The parameters of all models in each representative model cluster are fused; Based on the new time-period data of the fused parameters and environmental vibration signals, the representative model is incrementally updated to obtain the optimized multi-time-period model.

[0014] According to the present invention, a deep learning-based DAS (Data Access System) anti-external attack method is provided, which monitors and matches the current time period model in the optimized multi-time period model in real time, uses multi-dimensional judgment, outputs monitoring and alarm results, and continuously updates the model library based on the monitoring and alarm results, including: Match the optimal model corresponding to the current time period, perform real-time analysis on the environmental vibration signal, and output the probability of current data anomalies; By combining historical time period data and spatial location information for multi-dimensional cross-validation, if a potential threat is identified, a graded early warning mechanism is triggered based on event type, confidence level, and spatial distribution characteristics. Establish a continuous learning mechanism to automatically feed manually confirmed event samples back to the model optimization queue, thereby enabling the dynamic evolution of detection capabilities.

[0015] According to the present invention, a deep learning-based DAS (Data Access System) anti-external damage method includes a graded early warning mechanism comprising: Automatically record and continuously track the characteristics of preset low-risk events; Activate audible and visual alarms and push location information for pre-set medium- and high-risk events; The system will be activated in real time to respond to confirmed high-risk events.

[0016] Secondly, the present invention also provides a deep learning-based DAS (Disruption-Avoidance System) for preventing external damage, comprising: The preprocessing module is used to collect environmental vibration signals through DAS, preprocess the environmental vibration signals, and establish a time-period feature database; The training module is used to train a dedicated recognition model for each monitoring period based on the time period feature database using a CNN-LSTM hybrid network, thereby obtaining multiple initial sub-models for each time period. The optimization module is used to organize the initial sub-models of the multiple time periods into a preset scene model through model similarity analysis, establish a dynamic update mechanism, and optimize the model parameters using updated data to obtain an optimized multi-time period model. The output module is used to monitor and match the current time period model in the optimized multi-time period model in real time, use multi-dimensional judgment to output monitoring and alarm results, and continuously update the model library based on the monitoring and alarm results.

[0017] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the deep learning-based DAS anti-tampering method as described above.

[0018] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the deep learning-based DAS anti-external attack method as described above.

[0019] The deep learning-based DAS (Disruption Assurance System) method and system for preventing external damage provided by this invention continuously receives pre-processed real-time vibration data streams, automatically matches the optimal representative model corresponding to the current time period for real-time analysis through an intelligent scheduling engine, uses a CNN-LSTM hybrid network to train a dedicated recognition model for each time period, establishes a dynamic update mechanism, periodically optimizes model parameters with new data, intelligently matches the model for the current time period during real-time monitoring, supports high-precision full-database matching and efficient single-model matching dual modes, immediately triggers multi-level early warning upon detecting a threat, and achieves sub-second response and meter-level positioning accuracy. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts of the deep learning-based DAS anti-external damage method provided by the present invention; Figure 2 This is the second flowchart of the deep learning-based DAS anti-external attack method provided by the present invention; Figure 3 This is a schematic diagram of the structure of the DAS anti-external damage system based on deep learning provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] To address the limitations of existing technologies, this invention proposes a deep learning-based DAS (Data Access System) anti-external breaching method, such as... Figure 1 As shown, it includes: Step 100: Collect environmental vibration signals through DAS, preprocess the environmental vibration signals, and establish a time period feature database; Step 200: Based on the time period feature database, a dedicated recognition model for each monitoring time period is trained using a CNN-LSTM hybrid network to obtain multiple initial sub-models for each time period; Step 300: Organize the initial sub-models of the multiple time periods into a preset scene model through model similarity analysis, establish a dynamic update mechanism, optimize the model parameters using updated data, and obtain an optimized multi-time period model; Step 400: By real-time monitoring and matching the current time period model in the optimized multi-time period model, using multi-dimensional judgment, output monitoring and alarm results, and continuously update the model library based on the monitoring and alarm results.

[0024] This invention's system collects environmental vibration signals in real time via distributed optical fibers, dividing the entire day into 5-minute monitoring periods. The signals for each period undergo preprocessing such as noise reduction and time-frequency conversion to establish a period feature database. Next, a CNN-LSTM hybrid network is used to train a dedicated recognition model for each period, constructing a total of 288 initial sub-models to optimize the recognition capability of external damage events in different scenarios (such as subway operation and quiet nighttime). Then, through model similarity analysis, the initial models are restructured into 7 typical scenario models, establishing a dynamic update mechanism to periodically optimize model parameters with new data, ensuring accuracy while reducing computational burden. Finally, during real-time monitoring, the system intelligently matches the model for the current period, supporting both high-precision full-database matching and efficient single-model matching. Upon detecting a threat, it immediately triggers multi-level early warnings, achieving sub-second response and meter-level positioning accuracy.

[0025] It should be noted that the system reorganized the initial 288 5-minute time period models (corresponding to the entire 24 hours) into 7 typical high-speed rail scenario models through cluster analysis, and their correspondence is as follows: (1) M0 (normal high-speed rail operation period): corresponds to the off-peak period when trains run at a constant speed in the initial model (such as the sub-model of the non-entry and exit period between 9:00 and 16:00). (2) M1 (high-speed rail entry vibration period): a sub-model that integrates all deceleration entry periods within 5km of all stations (characterized by enhanced low-frequency vibration). (3) M2 (high-speed rail departure vibration period): a sub-model integrating the train acceleration departure period (characterized by high-frequency vibration dominance); (4) M3 (morning and evening peak hours): merge the high-frequency train tracking interval sub-models for 7:00-9:00 AM and 5:00-7:00 PM; (5) M4 (Extreme Weather Interference Period): A sub-model that aggregates all affected periods under meteorological conditions such as strong winds / heavy rain; (6) M5 (Nighttime sunroof maintenance period): A special vibration mode sub-model for the maintenance work period from 0:00 to 4:00 in the early morning; (7) M6 (Quiet Period): Covers all sub-models of time periods when no trains pass through (such as non-running periods in remote sections). Each representative model is formed by weighted fusion of 20-500 initial sub-model parameters. For example, M3 integrates 42 sub-models for morning and evening peak hours, while M6 integrates more than 400 models for nighttime off-peak hours.

[0026] Specifically, the logical flow of the specific method in this embodiment of the invention is as follows: Figure 2 As shown, it includes: First, the system collects environmental vibration signals in real time through distributed optical fibers, divides the monitoring period of the day into 5-minute units, performs preprocessing such as noise reduction and time-frequency conversion on the signals of each period, and establishes a period feature database.

[0027] The system collects environmental vibration signal data in real time through a distributed fiber optic sensor network deployed around critical infrastructure. This network employs phase-sensitive optical time-domain reflectometry (Φ-OTDR) technology, enabling ultra-high sensitivity detection of acoustic vibration signals propagating along optical fibers, with a spatial resolution of up to 10 meters and a sampling frequency of up to 10 kHz. The system first performs multi-stage preprocessing on the raw signal: the first stage uses wavelet transform for noise filtering, effectively separating environmental background noise from the effective vibration signal; the second stage performs signal normalization to eliminate signal strength differences caused by distance attenuation; and the third stage uses short-time Fourier transform to convert the time-domain signal into a time-frequency feature matrix. The preprocessed data is divided into 5-minute monitoring units, forming 288 continuous monitoring periods throughout the day. Data from each period is stored separately, and multi-dimensional feature vectors, including signal energy, frequency domain features, and time-domain statistics, are extracted to construct a complete period feature database. To ensure data quality, the system also includes an automatic quality detection mechanism that evaluates indicators such as signal-to-noise ratio and continuity in real time, removing abnormal data segments and triggering a re-collection process. Meanwhile, the system will establish an environmental baseline model for each monitoring period and record the typical vibration characteristic patterns of that period under normal conditions, providing a comparison benchmark for subsequent anomaly detection.

[0028] The system automatically identifies and removes abnormal data segments using multi-dimensional rules, mainly including five types of anomalies: low signal-to-noise ratio (e.g., fiber optic cable breakage), signal interruption or saturation (e.g., equipment failure), abnormal frequency band energy distribution (exceeding the typical vibration range of high-speed rail), spatial continuity failure (signals from adjacent monitoring points are uncorrelated), and deviation from historical baseline (abnormal and continuous vibration intensity). These judgment criteria are optimized based on the operating characteristics of high-speed rail and employ a cross-validation mechanism between adjacent sensors to reduce the false positive rate and ensure the reliability of the data input for subsequent analysis. For short-term anomalies, the system automatically repairs them; for long-term anomalies, it switches to a backup channel to collect data.

[0029] Secondly, a CNN-LSTM hybrid network was used to train a dedicated recognition model for each time period, and a total of 288 initial sub-models were constructed to optimize the external damage event recognition capabilities for different scenarios (such as subway operation, nighttime silence, etc.).

[0030] Based on the time-segment feature database established in step one, the system employs a deep convolutional-recurrent neural network (CNN-LSTM) hybrid architecture to train a dedicated recognition model for each 5-minute time segment. This model design fully considers the time-frequency features extracted during the preprocessing stage. The specific implementation process is as follows: (1) Input feature processing: The model input directly uses the time-frequency feature matrix generated in the preprocessing stage. ,in: T=3000 represents the number of sampling points within a 5-minute time span (determined by the 10kHz sampling rate in step one), and F=128 represents the number of frequency bands divided by the short-time Fourier transform in step one. This input feature maintains the complete time-frequency characteristics of the signal processing in step one.

[0031] (2) Convolutional feature extraction: Based on the time-frequency feature matrix obtained in step one, a five-layer two-dimensional convolutional module is designed:

[0032] in, Using 3×3 convolution kernels, local patterns of time-frequency features are extracted step by step. The ReLU activation function is used; after each convolutional layer, 2×2 max pooling is used to gradually abstract the feature representation, and finally output a 256-dimensional feature vector.

[0033] (3) Temporal dependency modeling: To capture the complete event evolution within a defined 5-minute timeframe, a bidirectional LSTM network is employed.

[0034] The input at the current time step (from the CNN output features) Hidden state (dimension 128); This layer is specifically designed to process the continuous time-series characteristics established in step one, and to learn the temporal evolution patterns of vibration events.

[0035] (4) Attention Enhancement: Introducing attention mechanisms to strengthen key features:

[0036] in, The mean of the time period characteristics (from the time period baseline in step one). The baseline mean for the time period; , Projection matrix; q Attention vector.

[0037] This mechanism can highlight the significant vibration features discovered in the preprocessing step one.

[0038] (5) Classification decision: The final output layer uses softmax classification.

[0039] in This indicates that combining dynamic features with time-based baseline features improves classification robustness. It is a feature concatenation operation. Represents the dynamic feature vector at the current moment, which comes from the output of the LSTM layer (dimension 128) and contains the temporal evolution features of the real-time vibration signal (such as the vibration pattern when a high-speed train passes by). The baseline feature vector (128 dimensions) for the time period is derived from the historical mean of the time period preprocessed in step one, representing the typical vibration characteristics (such as background environmental noise) of this 5-minute time period under normal, event-free conditions. [;] indicates vertical concatenation of vectors, ultimately generating a 256-dimensional hybrid feature vector. Its function is to fuse dynamic and static features, integrating transient features detected in real time. Compared with historical baseline In contrast, this enhances the model's sensitivity to anomalous events.

[0040] Then, the initial model was reorganized into seven typical scenario models through model similarity analysis, and a dynamic update mechanism was established to optimize the model parameters with new data on a regular basis, so as to reduce the computational burden while ensuring accuracy.

[0041] The dynamic time-segmented modeling and intelligent reorganization mechanism and the time-dimensional dynamic modeling method adopted in this invention break through the limitations of traditional single models. They not only preserve the feature differences of different time periods, but also greatly reduce the computational complexity through model fusion, thereby achieving synergistic optimization of recognition accuracy and system performance.

[0042] Based on the 288 time-segment sub-models trained in step two, this step achieves intelligent compression and optimization of the model library through feature space analysis and model similarity calculation. The specific implementation process is as follows: (1) Model feature space construction: For each sub-model M_i (i∈[1,288]), extract its last hidden feature transformation matrix: in: This represents the weight matrix of the Lth layer (the layer before the classification layer) of model M_i. =256 is the dimension of the LSTM hidden unit in step two. =128 is consistent with the frequency band dimension F in step one.

[0043] (2) Model similarity calculation: Projected distance is used to measure the similarity between models: This metric takes into account the consistency of the responses of each model in step two to the same time-frequency features (output of step one).

[0044] (3) Hierarchical clustering analysis: Constructing a hierarchical clustering tree based on the similarity matrix:

[0045] in: This represents bottom-up hierarchical clustering, with a threshold value. Controlling the clustering granularity ensures consistent classification behavior of intra-class models in step two. (4) Representative model selection: For each cluster Choose the central model:

[0046] The representative model to be retained must have an average accuracy of no less than 95% on the test set in step two.

[0047] (5) Feature space mapping: Establish a mapping function from the original time period to the representative model:

[0048] Where Mt represents the original model corresponding to time period t.

[0049] (6) Model fusion optimization: For each representative model, fuse the parameters of all models within its cluster:

[0050] in It includes all trainable parameters of the CNN-LSTM from step two.

[0051] (7) Dynamic update mechanism: Based on the newly collected data in step one, the representative model is incrementally updated:

[0052] in: This is data for the new time period (from the preprocessing workflow in step one). =0.01 is the learning rate. This refers to the Focal Loss defined in step two.

[0053] Finally, during real-time monitoring, the system intelligently matches the model for the current time period, supporting both high-precision full-database matching and efficient single-model matching modes. Upon detecting a threat, it immediately triggers multi-level warnings, achieving sub-second response and meter-level positioning accuracy. The system provides two intelligent matching modes: the high-precision full-database matching mode simultaneously compares all seven representative models, suitable for critical sections such as high-speed rail stations, ensuring high accuracy in complex scenarios; the efficient single-model matching mode only calls the single optimized model corresponding to the current time period, suitable for monitoring regular sections, significantly improving real-time performance. The two modes can automatically switch according to the risk level. When the confidence level of single-model matching is insufficient, full-database verification is automatically activated, ensuring accurate capture of high-risk events while achieving efficient resource utilization, forming a flexible monitoring system.

[0054] Building upon the optimized model library established in the first three stages, this step enables 24 / 7 real-time monitoring and intelligent early warning. The system continuously receives pre-processed real-time vibration data streams and automatically matches the optimal representative model for the current time period for real-time analysis via an intelligent scheduling engine. During monitoring, the system not only assesses the anomaly probability of the current data but also performs multi-dimensional cross-validation using historical time period data and spatial positioning information to ensure the reliability of the detection results. When a potential threat is identified, the system intelligently triggers a tiered early warning mechanism based on event type, confidence level, and spatial distribution characteristics: for low-risk events, it automatically records characteristics and continuously tracks them; for medium- and high-risk events, it immediately activates audible and visual alarms and pushes location information; and for confirmed high-risk events, it links with the emergency response system in real time. Simultaneously, the system establishes a continuous learning mechanism, automatically feeding back manually confirmed event samples to the model optimization queue, enabling dynamic evolution of detection capabilities. This complete "perception-decision-response-optimization" closed loop ensures that the system maintains high detection accuracy and rapid response capabilities amidst complex environmental changes. This invention employs multimodal real-time response and closed-loop evolution to construct a full-link intelligent system encompassing "data-model-decision." By real-time matching of time-period models with vibration characteristics, combined with spatiotemporal consistency verification, it achieves sub-second event detection. Its unique three-level early warning mechanism comprehensively assesses risk based on event probability, spatial distribution, and historical trajectories, supporting tiered responses from recording and tracking to emergency linkage. The system also innovatively establishes a dynamic feedback loop, automatically feeding confirmed event data back to the model library. Through incremental learning, it continuously optimizes representative models, enabling the system to autonomously evolve with environmental changes and maintain a recognition accuracy exceeding 95% over the long term.

[0055] The deep learning-based DAS anti-external damage system provided by this invention will be described below. The deep learning-based DAS anti-external damage system described below can be referred to in correspondence with the deep learning-based DAS anti-external damage method described above.

[0056] Figure 3This is a schematic diagram of the structure of the DAS anti-external damage system based on deep learning provided in an embodiment of the present invention, as shown below. Figure 3 As shown, it includes: a preprocessing module 31, a training module 32, an optimization module 33, and an output module 34, wherein: The preprocessing module 31 is used to collect environmental vibration signals through DAS, preprocess the environmental vibration signals, and establish a time period feature database; the training module 32 is used to train a dedicated recognition model for each monitoring time period based on the time period feature database using a CNN-LSTM hybrid network to obtain a multi-time period initial sub-model; the optimization module 33 is used to organize the multi-time period initial sub-models into a preset scene model through model similarity analysis, establish a dynamic update mechanism, and optimize the model parameters using updated data to obtain an optimized multi-time period model; the output module 34 is used to match the current time period model in the optimized multi-time period model through real-time monitoring, use multi-dimensional judgment, output monitoring alarm results, and continuously update the model library based on the monitoring alarm results.

[0057] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communication interface 420, a memory 430, and a communication bus 440. The processor 410, communication interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a deep learning-based DAS (Disruption-Oriented Assault System) anti-tampering method. This method includes: acquiring environmental vibration signals via DAS; preprocessing the environmental vibration signals to establish a time-period feature database; training a dedicated recognition model for each monitoring time period using a CNN-LSTM hybrid network based on the time-period feature database to obtain multiple initial sub-models for each time period; organizing the multiple initial sub-models into a preset scenario model through model similarity analysis, establishing a dynamic update mechanism, and optimizing model parameters using updated data to obtain an optimized multi-time-period model; matching the current time-period model in the optimized multi-time-period model through real-time monitoring, using multi-dimensional judgment, outputting monitoring and alarm results, and continuously updating the model library based on the monitoring and alarm results.

[0058] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the deep learning-based DAS anti-external damage method provided by the above methods. The method includes: acquiring environmental vibration signals through DAS; preprocessing the environmental vibration signals to establish a time-period feature database; training a dedicated recognition model for each monitoring time period based on the time-period feature database using a CNN-LSTM hybrid network to obtain multiple time-period initial sub-models; organizing the multiple time-period initial sub-models into preset scene models through model similarity analysis, establishing a dynamic update mechanism, optimizing model parameters using updated data, and obtaining optimized multi-time-period models; matching the current time-period model in the optimized multi-time-period models through real-time monitoring, using multi-dimensional judgment, outputting monitoring alarm results, and continuously updating the model library based on the monitoring alarm results.

[0060] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for preventing external attacks on DAS (Data Access System), characterized in that, include: Environmental vibration signals are collected by distributed optical fiber acoustic sensing (DAS), and the environmental vibration signals are preprocessed to establish a time-period feature database. Based on the time period feature database, a dedicated recognition model for each monitoring time period is trained using a CNN-LSTM hybrid network to obtain multiple initial sub-models for different time periods; The initial sub-models for multiple time periods are organized into a preset scenario model through model similarity analysis, a dynamic update mechanism is established, and the model parameters are optimized using updated data to obtain an optimized multi-time period model. By real-time monitoring and matching of the current time period model in the optimized multi-time period model, multi-dimensional judgment is used to output monitoring and alarm results, and the model library is continuously updated based on the monitoring and alarm results.

2. The deep learning-based DAS anti-external attack method according to claim 1, characterized in that, Environmental vibration signals are acquired via DAS, preprocessed, and a time-period feature database is established, including: Multiple environmental vibration signals were collected via DAS within a preset time period; Multiple environmental vibration signals are sequentially subjected to wavelet transform for noise filtering, signal normalization, and short-time Fourier transform to convert the time-domain signals into time-frequency feature matrices. The time-frequency feature matrix is ​​divided into time periods by using equal-interval durations as individual monitoring units to form multiple continuous monitoring time periods. Multidimensional feature vectors are extracted from these multiple continuous monitoring time periods to construct the time period feature database.

3. The deep learning-based DAS anti-external attack method according to claim 2, characterized in that, After acquiring environmental vibration signals via DAS, preprocessing the environmental vibration signals, and establishing a time-period feature database, the process also includes: An automatic quality detection mechanism is adopted to evaluate preset signal indicators in real time, remove abnormal data segments from the time period feature database, and trigger an automatic data repair process or an automatic data re-collection process.

4. The deep learning-based DAS anti-external attack method according to claim 1, characterized in that, Based on the aforementioned time-period feature database, a dedicated recognition model for each monitoring time period is trained using a CNN-LSTM hybrid network, resulting in multiple initial sub-models for different time periods, including: Input feature processing is performed on the time period feature database to extract the number of sampling points of a single monitoring unit and the number of frequency bands divided by short-time Fourier transform in the corresponding time-frequency feature matrix; A five-layer two-dimensional convolutional module is used to extract convolutional features from the input features, and a multi-dimensional feature vector is output. The time-period feature database is processed using a bidirectional LSTM network to obtain continuous time-period features; An attention mechanism is introduced into the continuous features of the time period to enhance key features and obtain significant vibration features; The softmax function is used in the output layer for decision classification to obtain the multi-time initial sub-model.

5. The deep learning-based DAS anti-external attack method according to claim 1, characterized in that, The initial sub-models across multiple time periods are organized into a preset scenario model through model similarity analysis. A dynamic update mechanism is established, and the model parameters are optimized using updated data to obtain an optimized multi-time period model, including: For each initial sub-model, extract the weight matrix of the last hidden feature transformation to construct the model feature space. The dimensions of the weight matrix include the dimension of the LSTM hidden units and the dimension of the number of frequency bands divided by the short-time Fourier transform. Based on the consistency of responses of each model to the same time-frequency features, the similarity between models is measured by projection distance. A bottom-up aggregated hierarchical clustering and control clustering method is adopted to construct a hierarchical clustering tree based on the similarity matrix; For each cluster in the hierarchical clustering tree, a central model is selected; Obtain the original model for any time period, and establish a mapping function from the original time period to the representative model based on the central model; The parameters of all models in each representative model cluster are fused; Based on the new time-period data of the fused parameters and environmental vibration signals, the representative model is incrementally updated to obtain the optimized multi-time-period model.

6. The deep learning-based DAS anti-external attack method according to claim 1, characterized in that, By real-time monitoring and matching the current time period model in the optimized multi-time period model, and using multi-dimensional judgment, monitoring and alarm results are output. Based on the monitoring and alarm results, the model library is continuously updated, including: Match the optimal model corresponding to the current time period, perform real-time analysis on the environmental vibration signal, and output the probability of current data anomalies; By combining historical time period data and spatial location information for multi-dimensional cross-validation, if a potential threat is identified, a graded early warning mechanism is triggered based on event type, confidence level, and spatial distribution characteristics. Establish a continuous learning mechanism to automatically feed manually confirmed event samples back to the model optimization queue, thereby enabling the dynamic evolution of detection capabilities.

7. The deep learning-based DAS anti-external attack method according to claim 6, characterized in that, The tiered early warning mechanism includes: Automatically record and continuously track the characteristics of preset low-risk events; Activate audible and visual alarms and push location information for pre-set medium- and high-risk events; The system will be activated in real time to respond to confirmed high-risk events.

8. A deep learning-based DAS (Disruption-Avoidance System) for preventing external damage, characterized in that, include: The preprocessing module is used to collect environmental vibration signals through DAS, preprocess the environmental vibration signals, and establish a time-period feature database. The training module is used to train a dedicated recognition model for each monitoring period based on the time period feature database using a CNN-LSTM hybrid network, thereby obtaining multiple initial sub-models for each time period. The optimization module is used to organize the initial sub-models of the multiple time periods into a preset scene model through model similarity analysis, establish a dynamic update mechanism, and optimize the model parameters using updated data to obtain an optimized multi-time period model. The output module is used to monitor and match the current time period model in the optimized multi-time period model in real time, use multi-dimensional judgment to output monitoring and alarm results, and continuously update the model library based on the monitoring and alarm results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the deep learning-based DAS anti-external attack method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based DAS anti-external attack method as described in any one of claims 1 to 7.

Images