A method and system for detecting and protecting abnormal behavior of power terminals

By collecting multi-dimensional data and using dual-model fusion detection, combined with attack chain pattern matching and power physical consistency verification, a closed-loop self-optimization mechanism is constructed. This solves the problems of detection blind spots and high false alarm rates in power terminal security protection, and achieves efficient identification and protection against abnormal behavior of power terminals.

CN123122915APending Publication Date: 2026-11-17HEXING ELECTRICAL CO LTD +4
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611296781.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing power terminal security protection solutions suffer from several problems, including a single detection mechanism leading to high rates of missed detections and false alarms, a lack of attack chain identification capabilities, the absence of a protection feedback loop preventing performance self-iteration, and insufficient coverage of power terminal-specific issues such as electricity theft and physical attacks.

Method used

By employing multi-dimensional data collection, dual-model fusion detection combining the isolated forest model and the autoencoder model, and combining attack chain pattern matching and power physics consistency verification, a closed-loop self-optimization mechanism is constructed to dynamically adjust the threat level and implement corresponding protection measures.

Benefits of technology

It effectively identifies known and unknown threats, reduces false alarm rate, improves the ability to detect complex attacks, and achieves adaptive real-time security protection at the edge, with an accuracy rate of over 95% and a false alarm rate controlled within 5%, adapting to diverse threat scenarios in power terminals.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122915A_ABST
    Figure CN123122915A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for detecting and protecting abnormal behavior of power terminals. The method includes the following steps: collecting multi-dimensional behavioral data of power terminals and constructing multi-dimensional behavioral feature vectors; inputting the multi-dimensional behavioral feature vectors into an anomaly detection model, performing dual-model fusion detection through an isolated forest model and an autoencoder model to obtain global distribution anomaly detection results and local reconstruction anomaly detection results, fusing and judging the two types of detection results, and outputting anomaly behavior judgment results; performing attack chain pattern matching on continuous abnormal behavior sequences within a preset time window and dynamically adjusting the threat level; and executing corresponding level of security protection measures according to the current threat level. Through the implementation of this invention, known and unknown threats can be effectively identified, the false alarm rate can be reduced, the ability to detect composite attacks can be improved, and adaptive real-time security protection at the edge can be achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power terminal security technology, and in particular to a method and system for detecting and protecting abnormal behavior of power terminals. Background Technology

[0002] With the continuous advancement of the distribution Internet of Things (IoT) and power digitalization, smart meters, concentrators, energy controllers, and other power terminals have been widely deployed at the end of the power distribution network, undertaking core functions such as power metering, data acquisition, equipment control, and two-way interaction. Due to the dispersed deployment environment, open access networks, and limited computing resources of these power terminals, they face diverse security threats such as network intrusion, protocol injection, power tampering, and strong magnetic interference. Electricity theft attacks and terminal takeover incidents occur frequently, posing a severe challenge to the stable operation and data security of the power system. Therefore, intelligent security protection technologies adapted to edge computing scenarios are urgently needed.

[0003] In existing power terminal security protection solutions, some solutions achieve active defense by collecting multi-dimensional state vectors of terminals and calculating behavioral deviation distances. However, they only use a single distance metric mechanism and cannot simultaneously capture global distribution anomalies and local temporal reconstruction anomalies, resulting in blind spots in detection. Some solutions achieve intrusion detection and electricity theft identification based on predefined rules or fixed thresholds, which can only cover known attack types, have a high false alarm rate, and cannot adapt to new attack methods. Most solutions only independently determine single-point anomalies, lack the ability to identify multi-step attack chains, and lack effect evaluation and parameter feedback mechanisms after the protection strategy is implemented, so the detection model cannot continuously self-optimize.

[0004] In summary, there is a need for a method and system for detecting and protecting abnormal behavior of power terminals to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for detecting and protecting abnormal behavior of power terminals. It aims to solve the problems of high false alarm and missed detection rates due to a single detection mechanism, easy omission of compound attacks due to lack of attack chain identification capability, inability to achieve performance self-iteration due to lack of protection feedback closed loop, and insufficient coverage of power terminal-specific electricity theft and physical attacks.

[0006] On the one hand, this application provides a method for detecting and protecting abnormal behavior of power terminals, including the following steps:

[0007] Step 1: Collect multi-dimensional behavioral data from power terminals and construct multi-dimensional behavioral feature vectors;

[0008] Step 2: Input the multi-dimensional behavioral feature vector into the anomaly detection model, perform dual-model fusion detection through the isolated forest model and the autoencoder model, and obtain the global distribution anomaly detection result and the local reconstruction anomaly detection result respectively. Perform fusion judgment on the two types of detection results and output the abnormal behavior judgment result.

[0009] Step 3: Perform attack chain pattern matching on the continuous abnormal behavior sequence within the preset time window, and dynamically adjust the threat level according to the attack chain matching stage, the confidence level of the abnormal behavior and the number of consecutive occurrences.

[0010] Step 4: Implement security measures corresponding to the current threat level;

[0011] Step 5: Evaluate the effectiveness of the security measures and adaptively adjust the detection parameters of the anomaly detection model based on the evaluation results to form a closed-loop operation mechanism consisting of anomaly detection, hierarchical protection, and adaptive parameter optimization.

[0012] By comprehensively covering various threat scenarios of power terminals through multi-dimensional data collection, and adopting dual-model fusion detection to take into account both global distribution and local temporal anomaly identification, combined with dynamic attack chain adjustment and closed-loop self-optimization mechanism, it can effectively identify known and unknown threats, reduce false alarm rate, improve the ability to detect composite attacks, and achieve adaptive real-time security protection at the edge.

[0013] Furthermore, in step 1, the multi-dimensional behavioral data includes at least communication behavioral data, operational behavioral data, power parameter data, and environmental physical data; after time alignment, data cleaning, and normalization preprocessing of the collected multi-dimensional behavioral data, the multi-dimensional behavioral feature vector is constructed by splicing them together.

[0014] Furthermore, in step 2, the anomaly detection model is deployed in the local edge AI computing unit of the power terminal;

[0015] The input to the isolated forest model is a subset of global statistical features, which includes communication traffic statistics, steady-state deviation of power parameters, discrete statistical features of operational behavior, and storage access frequency features.

[0016] The input to the autoencoder model is a subset of time-series features, which includes operation command encoding sequences, protocol message byte sequences, power timing sampling sequences, and current waveform sampling sequences.

[0017] Furthermore, in step 2, the specific rules for fusion determination are as follows:

[0018] When the global anomaly score output by the isolated forest model is greater than or equal to the first anomaly threshold, and the reconstruction error score output by the autoencoder model is greater than or equal to the second anomaly threshold, it is determined to be an confirmed abnormal behavior.

[0019] When either the global anomaly score or the reconstruction error score is greater than or equal to the corresponding anomaly threshold, it is determined to be an anomaly behavior to be confirmed.

[0020] When both scores are less than the corresponding abnormal threshold, it is judged as normal behavior.

[0021] Furthermore, after determining that the behavior is an abnormal behavior to be confirmed, a power physical consistency check is triggered; the power physical consistency check is based on the physical laws of the power system and performs cross-verification of physical laws through at least power balance constraints, voltage-current phase relationship constraints, and energy-power integral consistency constraints.

[0022] If all verification results meet the preset tolerance threshold of the corresponding verification item, the abnormal behavior to be confirmed will be downgraded to an observation state, the sampling frequency of the corresponding data dimension will be increased and continuous monitoring will be carried out.

[0023] If any of the verification results fails to meet the preset tolerance threshold of the corresponding verification item, the abnormal behavior to be confirmed will be upgraded to confirmed abnormal behavior, and the threat level corresponding to the abnormal behavior will be increased.

[0024] Furthermore, in step 3, the attack chain pattern matching is performed using a finite state machine; the built-in attack chain patterns include at least three types of power terminal-specific attack patterns: electricity theft attack chain, terminal control attack chain, and data poisoning attack chain.

[0025] The confidence level of each attack phase is accumulated through exponential smoothing. If no matching event is detected in the corresponding phase within the preset decay period, the confidence level of the corresponding phase is reduced according to the preset decay rule.

[0026] Furthermore, the threat level is divided into three levels from low to high: low, medium, and high; the dynamic adjustment rules are as follows:

[0027] A single confirmed abnormal behavior initially corresponds to a low threat level;

[0028] When the number of consecutive occurrences of abnormal behavior is greater than or equal to the preset number, or when the attack chain matches and enters the probing intrusion phase, the threat level is upgraded to medium threat level.

[0029] When the attack chain matches and enters the stage of gaining privileges or performing sabotage, or when the duration of abnormal behavior is greater than or equal to the preset duration and the overall confidence level of the abnormal behavior is greater than or equal to the preset confidence threshold, the threat level is upgraded to high threat level.

[0030] If no new confirmed abnormal behavior is detected within the preset downgrade period, the threat level will be automatically reduced step by step.

[0031] Furthermore, in step 5, the evaluation dimensions for the effectiveness of security protection measures include at least the termination of abnormal behavior, the recurrence of similar abnormalities, and the degree to which normal business is affected.

[0032] The adaptively adjusted detection parameters include at least a first anomaly threshold, a second anomaly threshold, and dual-model fusion weights;

[0033] The normal behavior baseline of the anomaly detection model is adaptively updated using a sliding window method. The window length is a parameter that can be configured according to the terminal's operating scenario. New samples are incorporated into the original baseline with exponentially decaying weights.

[0034] On the other hand, this application provides a power terminal abnormal behavior detection and security protection system, including:

[0035] The multi-dimensional data acquisition module is used to collect multi-dimensional behavioral data of power terminals and construct multi-dimensional behavioral feature vectors.

[0036] The AI ​​anomaly detection module incorporates an isolated forest model and an autoencoder model to perform dual-model fusion detection on multi-dimensional behavioral feature vectors. It outputs detection results for global distribution anomalies and local reconstruction anomalies respectively, and then merges them to output the anomaly behavior judgment result.

[0037] The attack chain detection module is used to perform attack chain pattern matching on a series of consecutive abnormal behaviors within a preset time window, and dynamically adjust the threat level based on the attack chain matching stage, the confidence level of the abnormal behavior, and the number of consecutive occurrences.

[0038] The protection execution module is used to execute security protection measures of the corresponding level according to the current threat level;

[0039] The strategy optimization module is used to evaluate the effectiveness of security protection measures and adaptively adjust the detection parameters of the AI ​​anomaly detection module based on the evaluation results, forming a closed-loop operation mechanism consisting of anomaly detection, hierarchical protection, and adaptive parameter optimization.

[0040] Furthermore, it also includes a power physical consistency verification module, which has built-in three types of physical verification rules: power balance constraints, voltage-current phase relationship constraints, and energy-power integral consistency constraints. These rules are used to cross-verify the physical laws of the abnormal behavior to be confirmed, and to adjust the comprehensive confidence and threat level of the abnormal behavior based on the verification results.

[0041] In the AI ​​anomaly detection module, the isolated forest model and the autoencoder model adopt a differentiated feature input mechanism, that is, the isolated forest model inputs a subset of global statistical features, and the autoencoder model inputs a subset of time sequence features.

[0042] The substantial effects of this invention:

[0043] 1. In this invention, by using the differentiated division of labor and complementary detection of the isolated forest and autoencoder dual models, combined with cross-validation of the laws of electrical physics, it can simultaneously cover global distribution anomalies and local temporal anomalies, effectively avoiding the detection blind spots of a single model. The anomaly detection accuracy can reach over 95%, and the false alarm rate can be controlled within 5%, greatly improving the ability to detect anti-sample attacks.

[0044] 2. In this invention, a dedicated attack chain pattern library for power terminals is constructed using a finite state machine. This library can identify multi-step continuous attack behaviors and comprehensively determine multiple low-confidence single-point anomalies as high-confidence composite attacks, thus solving the problem that traditional single-point detection is prone to missing advanced persistent attacks.

[0045] 3. In this invention, a closed-loop mechanism of "detection-protection-evaluation-optimization" is constructed, which can automatically adjust the detection threshold and model weight according to the protection performance. It does not require frequent manual maintenance of rules and can adapt to the dynamic changes of terminal operating environment and attack patterns.

[0046] 4. This invention covers four dimensions: communication, operation, power parameters, and environmental physics. It specifically addresses power terminal-specific threat scenarios such as electricity theft and tampering, and physical interference. Compared with general network security solutions, it is more suitable for the business characteristics of power terminals. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of Example 1.

[0049] Figure 2 This is a flowchart of Example 2.

[0050] Figure 3 This is a system principle block diagram of Example 3. Detailed Implementation

[0051] To more clearly illustrate the technical solutions in the embodiments of the 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 only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Example 1:

[0053] Reference Figure 1 As shown, a method for detecting and protecting abnormal behavior of power terminals includes the following steps:

[0054] Collect multi-dimensional behavioral data from power terminals and construct multi-dimensional behavioral feature vectors;

[0055] The multi-dimensional behavioral feature vector is input into the anomaly detection model. The isolated forest model and the autoencoder model are used to perform dual-model fusion detection to obtain global distribution anomaly detection results and local reconstruction anomaly detection results. The two types of detection results are fused and judged to output the abnormal behavior judgment result.

[0056] Within a preset time window, attack chain pattern matching is performed on continuous abnormal behavior sequences, and the threat level is dynamically adjusted based on the attack chain matching stage, the confidence level of abnormal behavior, and the number of consecutive occurrences.

[0057] Implement security measures corresponding to the current threat level;

[0058] The effectiveness of security measures is evaluated, and the detection parameters of the anomaly detection model are adaptively adjusted based on the evaluation results, forming a closed-loop operation mechanism consisting of anomaly detection, hierarchical protection, and adaptive parameter optimization.

[0059] As one implementation method, the multi-dimensional behavioral data includes at least communication behavioral data, operational behavioral data, power parameter data, and environmental physical data; after time alignment, data cleaning, and normalization preprocessing of the collected multi-dimensional behavioral data, the data is spliced ​​together to construct a multi-dimensional behavioral feature vector.

[0060] As one implementation method, the anomaly detection model is deployed in the local edge AI computing unit of the power terminal;

[0061] The input to the isolated forest model is a subset of global statistical features, which includes communication traffic statistics, steady-state deviation of power parameters, discrete statistical features of operational behavior, and storage access frequency features.

[0062] The input to the autoencoder model is a subset of time-series features, which includes operation command encoding sequences, protocol message byte sequences, power timing sampling sequences, and current waveform sampling sequences.

[0063] As one implementation method, the specific rules for fusion determination are as follows:

[0064] When the global anomaly score output by the isolated forest model is greater than or equal to the first anomaly threshold, and the reconstruction error score output by the autoencoder model is greater than or equal to the second anomaly threshold, it is determined to be an confirmed abnormal behavior.

[0065] When either the global anomaly score or the reconstruction error score is greater than or equal to the corresponding anomaly threshold, it is determined to be an anomaly behavior to be confirmed.

[0066] When both scores are less than the corresponding abnormal threshold, it is judged as normal behavior.

[0067] As one implementation method, after determining that an abnormal behavior is to be confirmed, a power physical consistency check is triggered. The power physical consistency check is based on the physical laws of the power system and performs cross-verification of physical laws through at least power balance constraints, voltage-current phase relationship constraints, and energy-power integral consistency constraints.

[0068] If all verification results meet the preset tolerance threshold of the corresponding verification item, the abnormal behavior to be confirmed will be downgraded to an observation state, the sampling frequency of the corresponding data dimension will be increased and continuous monitoring will be carried out.

[0069] If any of the verification results fails to meet the preset tolerance threshold of the corresponding verification item, the abnormal behavior to be confirmed will be upgraded to confirmed abnormal behavior, and the threat level corresponding to the abnormal behavior will be increased.

[0070] As one implementation method, attack chain pattern matching is performed using a finite state machine; the built-in attack chain patterns include at least three types of power terminal-specific attack patterns: electricity theft attack chain, terminal control attack chain, and data poisoning attack chain.

[0071] The confidence level of each attack phase is accumulated through exponential smoothing. If no matching event is detected in the corresponding phase within the preset decay period, the confidence level of the corresponding phase is reduced according to the preset decay rule.

[0072] As one implementation method, the threat level is divided into three levels from low to high: low, medium, and high; the dynamic adjustment rules are as follows:

[0073] A single confirmed abnormal behavior initially corresponds to a low threat level;

[0074] When the number of consecutive occurrences of abnormal behavior is greater than or equal to the preset number, or when the attack chain matches and enters the probing intrusion phase, the threat level is upgraded to medium threat level.

[0075] When the attack chain matches and enters the stage of gaining privileges or performing sabotage, or when the duration of abnormal behavior is greater than or equal to the preset duration and the overall confidence level of the abnormal behavior is greater than or equal to the preset confidence threshold, the threat level is upgraded to high threat level.

[0076] If no new confirmed abnormal behavior is detected within the preset downgrade period, the threat level will be automatically reduced step by step.

[0077] As one implementation method, the evaluation dimensions for the effectiveness of security protection measures should include at least the termination of abnormal behavior, the recurrence of similar abnormalities, and the degree of impact on normal business operations.

[0078] The adaptively adjusted detection parameters include at least a first anomaly threshold, a second anomaly threshold, and dual-model fusion weights;

[0079] The baseline of normal behavior of the anomaly detection model is adaptively updated using a sliding window method. The window length is a parameter that can be configured according to the terminal's operating scenario. New samples are incorporated into the original baseline with exponentially decaying weights.

[0080] Example 2:

[0081] Reference Figure 2 As shown, this embodiment is basically the same as Embodiment 1, except that it provides a method for detecting and protecting abnormal behavior of power terminals. This method is applied to smart meters deployed in user distribution boxes. The meter uses an RK3568 processor as the main controller, integrates an NPU unit with a computing power of 1 TOPS, runs an ARM Linux operating system, and has external temperature and humidity sensors, vibration sensors, and magnetic field detection sensors. The specific implementation process is as follows:

[0082] Step 1: Multi-dimensional behavioral data collection and feature construction

[0083] The system collects four types of behavioral data from the terminal simultaneously, with a detection cycle of 5 seconds:

[0084] Communication behavior data: Collect message interaction records of DL / T645 and DL / T698 protocols, and extract information such as port connection frequency, traffic rate, protocol distribution ratio, and message length distribution;

[0085] Operational behavior data: Collects command execution sequences, parameter modification records, login operation logs, and configuration change records from the terminal, recording the operation object, execution time, and operation subject;

[0086] Power parameter data: Real-time metering data such as three-phase voltage, three-phase current, active power, reactive power, power factor, and cumulative power consumption are collected through metering chips, as well as event alarm records;

[0087] Environmental physical data: Physical state data such as internal temperature, ambient humidity, vibration amplitude, and surrounding magnetic field strength are collected through external sensors.

[0088] The four types of data are aligned according to timestamps, and after preprocessing such as missing value filling, outlier removal, and numerical normalization, they are spliced ​​together to construct a 120-dimensional multi-dimensional behavioral feature vector.

[0089] Step 2: Dual-model fusion anomaly detection

[0090] Multi-dimensional behavioral feature vectors are input into an anomaly detection model deployed on the terminal NPU, and a differentiated feature input mechanism is used to feed them into two detection models respectively:

[0091] A 14-dimensional global statistical feature class is extracted from the feature vector, including traffic cycle distribution characteristics, port access ratio, three-phase voltage imbalance, power factor fluctuation coefficient, instruction execution interval variance, configuration change frequency, and sensitive path access ratio, and input into the isolation forest model. The isolation forest model generates 100 isolation trees based on normal samples pre-training, calculates the average path length of the current sample in all isolation trees, and outputs a global anomaly score in the range of 0-1.

[0092] A 12-dimensional time-series sequence class feature is extracted from the feature vector, including the encoded sequence of the most recent 50 operation commands, the byte sequence of the most recent 100 protocol messages, the active power sampling sequence of the most recent 60 seconds, and the instantaneous current value sequence of the most recent 5 power frequency cycles, which are then input into the autoencoder model. The autoencoder is trained based on normal time-series samples to obtain the encoder-decoder structure. The mean square error between the input features and the reconstructed output is calculated, and after normalization, a reconstruction error score in the 0-1 interval is output.

[0093] A dual-threshold fusion judgment rule is adopted: when the global anomaly score is greater than or equal to 0.65 and the reconstruction error score is greater than or equal to 0.70, it is judged as confirmed abnormal behavior; when only one of the two scores meets the corresponding threshold, it is judged as unconfirmed abnormal behavior; when both scores are less than the corresponding threshold, it is judged as normal behavior.

[0094] For any suspected abnormal behavior, a power physical consistency check is triggered: This check verifies the power balance (input / output power difference less than 0.5% of rated power), voltage / current phase relationship (power factor within the normal inductive range of 0.85-1.0), and energy integral consistency (the difference between energy increment and average power integral less than 1% of rated energy) within the current sampling period. If all three checks meet the tolerance requirements, the anomaly is downgraded to an observation state, the sampling frequency for the corresponding data dimension is increased to 1 second / time, and continuous monitoring is performed. If any check exceeds the tolerance range, the anomaly is upgraded to confirmed, and the threat level is increased by one level.

[0095] Step 3: Attack chain detection and dynamic adjustment of threat level

[0096] A 30-second sliding window is set up to perform attack chain pattern matching on continuous abnormal behavior sequences within the window. Finite state machines are used to model three types of power terminal-specific attack chains, including electricity theft attack chain (electricity collection tampering → protocol data injection → storage tampering → detection evasion), terminal control attack chain (firmware backdoor implantation → parameter tampering → remote takeover → destructive operation), and data poisoning attack chain (training data pollution → model decision misleading → protection strategy failure).

[0097] The confidence level of each attack phase is accumulated using an exponential smoothing method with a smoothing factor of 0.85. If no matching event for the corresponding phase is detected within 5 minutes, the confidence level decays exponentially at a rate of 10% per minute.

[0098] Threat levels are divided into three levels: low, medium, and high. The dynamic adjustment rules are as follows: a single confirmed abnormal behavior is initially judged as a low threat level; when three or more consecutive abnormal behaviors occur, or when the finite state machine transitions to the probing intrusion phase, it is upgraded to a medium threat level; when the finite state machine transitions to the privilege acquisition or sabotage execution phase, or when the abnormal behavior lasts for more than 2 minutes and the overall confidence level of the abnormal behavior is greater than or equal to 0.8, it is upgraded to a high threat level; if no new confirmed abnormal behavior occurs within 5 minutes, the threat level will automatically downgrade step by step.

[0099] Step 4: Implement graded security protection measures

[0100] Implement corresponding protective measures based on the current threat level:

[0101] Low threat level: Logs abnormal events and uploads them to the main operations and maintenance station, without restricting the normal operation of terminal services;

[0102] Medium threat level: Send real-time alerts to operations and maintenance personnel, dynamically restrict access permissions for non-core communication ports, and prohibit parameter modification operations during unauthorized periods;

[0103] High threat level: Send emergency alerts to operations and maintenance personnel, encrypt and protect sensitive metering data, trigger terminal network isolation mechanisms, prohibit all untrusted external connections, and retain core metering functions.

[0104] Step 5: Evaluation of protection effectiveness and closed-loop optimization

[0105] After the protective measures are implemented, continuous monitoring is conducted for 2 minutes, and the protection effectiveness score is calculated from three dimensions: whether the abnormal behavior has stopped, whether the same type of abnormality has been reproduced, and whether normal metering business has been affected. Finally, a quantitative score in the range of 0-1 is obtained.

[0106] The protection effectiveness score is fed back to the anomaly detection module: if the protection effectiveness score is greater than or equal to 0.8, it indicates that the detection sensitivity is insufficient, and the first and second anomaly thresholds are reduced in increments of 0.02; if the protection effectiveness score is less than or equal to 0.3, it indicates that false alarms have affected normal business operations, and the two anomaly thresholds are increased in increments of 0.02.

[0107] Meanwhile, the normal behavior baseline of the anomaly detection model is adaptively updated using a sliding window with a window length of 5000 samples. New samples are incorporated into the original baseline with a decay coefficient of 0.95, so that the baseline adapts to the slow changes in the normal operating state of the terminal.

[0108] In this embodiment, the AI ​​model is lightweightly deployed on the local NPU of the terminal, without relying on cloud computing. The latency of a single detection is less than 100ms, the protection response time is less than 1 second, and it can still run independently when the network is interrupted, making it suitable for edge deployment scenarios of power terminals.

[0109] Example 3:

[0110] Reference Figure 3 As shown, this embodiment is basically the same as embodiment 1, except that it provides a power terminal abnormal behavior detection and security protection system, deployed in a smart energy meter as in embodiment 2, to implement the protection method of embodiment 2. The system hardware relies on the terminal's RK3568 main control chip, NPU computing unit, metering chip, and peripheral sensors. The software includes the following functional modules:

[0111] The multi-dimensional data acquisition module connects to the terminal's communication interface, metering chip, operation log system, and peripheral sensors to synchronously collect four types of operational data: communication behavior, operation behavior, power parameters, and environmental physics. After completing time alignment, data cleaning, and normalization preprocessing, the data is spliced ​​to construct a multi-dimensional behavioral feature vector and output to the AI ​​anomaly detection module.

[0112] The AI ​​anomaly detection module is deployed in the terminal NPU computing unit and includes an isolated forest detection unit, an autoencoder detection unit, and a fusion judgment unit. The isolated forest detection unit receives a subset of global statistical features and outputs a global anomaly score. The autoencoder detection unit receives a subset of time-series features and outputs a reconstruction error score. The fusion judgment unit outputs three types of judgment results based on a dual threshold rule: confirmed anomaly, unconfirmed anomaly, and normal.

[0113] The power physics consistency verification module has three types of verification rules built in: power balance constraints, voltage-current phase relationship constraints, and energy-power integral consistency constraints. It receives the unconfirmed anomaly results output by the AI ​​anomaly detection module, retrieves the power operation data of the corresponding time period for cross-verification of physical laws, and adjusts the anomaly confidence and corresponding threat level according to the verification results.

[0114] The attack chain detection module has a built-in finite state machine unit and a power-specific attack chain pattern library. The pattern library stores the stage transition rules for three types of attacks: electricity theft attack chain, terminal control attack chain, and data poisoning attack chain. The module performs stage matching on continuous abnormal behavior sequences in a preset time window, calculates the matching confidence of each stage through exponential smoothing, and outputs the attack chain matching stage and situation assessment results.

[0115] The protection execution module has a built-in hierarchical protection policy library. After receiving threat level instructions, it executes corresponding protection actions. Under low threat, it performs log recording and reporting; under medium threat, it performs alarms and dynamic access control; under high threat, it performs emergency alarms, data encryption and network isolation. It also supports real-time synchronization of protection policies when the threat level is dynamically adjusted.

[0116] The strategy optimization module has a built-in effect evaluation unit and parameter adaptation unit. After the protection is implemented, the effect evaluation unit calculates a quantitative score of the protection effect from three dimensions: abnormal termination, abnormal reproduction, and business impact. The parameter adaptation unit adaptively adjusts the first abnormal threshold, the second abnormal threshold, and the dual-model fusion weight of the AI ​​anomaly detection module according to the score results, while driving the sliding window update of the normal behavior baseline to form a complete self-optimization closed loop.

[0117] The security monitoring application module connects to the operations and maintenance main station to display the terminal security status, threat level, attack type, current stage of the attack chain, and protection execution status in real time. It also supports operations and maintenance personnel to remotely view alarm details and adjust protection configurations.

[0118] It should be noted that while the preferred embodiments of the present invention are provided in the specification and accompanying drawings, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are not intended to impose additional limitations on the content of the present invention; their purpose is to provide a more thorough and comprehensive understanding of the disclosure of the present invention. Furthermore, the above-described technical features can be combined with each other to form various embodiments not listed above, all of which are considered to be within the scope of the present invention specification. Moreover, those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for detecting and protecting the abnormal behavior of power terminals, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional behavioral data from power terminals and construct multi-dimensional behavioral feature vectors; Step 2: Input the multi-dimensional behavioral feature vector into the anomaly detection model, perform dual-model fusion detection through the isolated forest model and the autoencoder model, and obtain the global distribution anomaly detection result and the local reconstruction anomaly detection result respectively. Perform fusion judgment on the two types of detection results and output the abnormal behavior judgment result. Step 3: Perform attack chain pattern matching on the continuous abnormal behavior sequence within the preset time window, and dynamically adjust the threat level according to the attack chain matching stage, the confidence level of the abnormal behavior and the number of consecutive occurrences. Step 4: Implement security measures corresponding to the current threat level; Step 5: Evaluate the effectiveness of the security measures and adaptively adjust the detection parameters of the anomaly detection model based on the evaluation results to form a closed-loop operation mechanism consisting of anomaly detection, hierarchical protection, and adaptive parameter optimization.

2. The method for detecting and protecting abnormal behavior of power terminals according to claim 1, characterized in that, In step 1, the multi-dimensional behavioral data includes at least communication behavioral data, operational behavioral data, power parameter data, and environmental physical data. After time alignment, data cleaning, and normalization preprocessing of the collected multi-dimensional behavioral data, the multi-dimensional behavioral feature vector is constructed by splicing the data together.

3. The method for detecting and protecting abnormal behavior of power terminals according to claim 1, characterized in that, In step 2, the anomaly detection model is deployed in the local edge AI computing unit of the power terminal. The input to the isolated forest model is a subset of global statistical features, which includes communication traffic statistics, steady-state deviation of power parameters, discrete statistical features of operational behavior, and storage access frequency features. The input to the autoencoder model is a subset of time-series features, which includes operation command encoding sequences, protocol message byte sequences, power timing sampling sequences, and current waveform sampling sequences.

4. The method for detecting and protecting abnormal behavior of power terminals according to claim 1, characterized in that, In step 2, the specific rules for fusion determination are as follows: When the global anomaly score output by the isolated forest model is greater than or equal to the first anomaly threshold, and the reconstruction error score output by the autoencoder model is greater than or equal to the second anomaly threshold, it is determined to be an confirmed abnormal behavior. When either the global anomaly score or the reconstruction error score is greater than or equal to the corresponding anomaly threshold, it is determined to be an anomaly behavior to be confirmed. When both scores are less than the corresponding abnormal threshold, it is judged as normal behavior.

5. The method for detecting and protecting abnormal behavior of power terminals according to claim 4, characterized in that, Once an abnormal behavior is identified as pending confirmation, a power physical consistency check is triggered. The power physical consistency check is based on the physical laws of the power system and performs cross-verification of physical laws through at least power balance constraints, voltage-current phase relationship constraints, and energy-power integral consistency constraints. If all verification results meet the preset tolerance threshold of the corresponding verification item, the abnormal behavior to be confirmed will be downgraded to an observation state, the sampling frequency of the corresponding data dimension will be increased and continuous monitoring will be carried out. If any of the verification results fails to meet the preset tolerance threshold of the corresponding verification item, the abnormal behavior to be confirmed will be upgraded to confirmed abnormal behavior, and the threat level corresponding to the abnormal behavior will be increased.

6. The method for detecting and protecting abnormal behavior of power terminals according to claim 1, characterized in that, In step 3, the attack chain pattern matching is performed using a finite state machine; the built-in attack chain patterns include at least three types of power terminal-specific attack patterns: electricity theft attack chain, terminal control attack chain, and data poisoning attack chain. The confidence level of each attack phase is accumulated through exponential smoothing. If no matching event is detected in the corresponding phase within the preset decay period, the confidence level of the corresponding phase is reduced according to the preset decay rule.

7. The method for detecting and protecting abnormal behavior of power terminals according to claim 1, characterized in that, The threat levels are divided into three levels from low to high: low, medium, and high; the dynamic adjustment rules are as follows: A single confirmed abnormal behavior initially corresponds to a low threat level; When the number of consecutive occurrences of abnormal behavior is greater than or equal to the preset number, or when the attack chain matches and enters the probing intrusion phase, the threat level is upgraded to medium threat level. When the attack chain matches and enters the stage of gaining privileges or performing sabotage, or when the duration of abnormal behavior is greater than or equal to the preset duration and the overall confidence level of the abnormal behavior is greater than or equal to the preset confidence threshold, the threat level is upgraded to high threat level. If no new confirmed abnormal behavior is detected within the preset downgrade period, the threat level will be automatically reduced step by step.

8. The method for detecting and protecting abnormal behavior of power terminals according to claim 1, characterized in that, In step 5, the evaluation dimensions for the effectiveness of security protection measures include at least the termination of abnormal behavior, the recurrence of similar abnormalities, and the degree to which normal business is affected. The adaptively adjusted detection parameters include at least a first anomaly threshold, a second anomaly threshold, and dual-model fusion weights; The normal behavior baseline of the anomaly detection model is adaptively updated using a sliding window method. The window length is a parameter that can be configured according to the terminal's operating scenario. New samples are incorporated into the original baseline with exponentially decaying weights.

9. A power terminal abnormal behavior detection and security protection system, used to implement the protection method as described in any one of claims 1-8, characterized in that, include: The multi-dimensional data acquisition module is used to collect multi-dimensional behavioral data of power terminals and construct multi-dimensional behavioral feature vectors. The AI ​​anomaly detection module incorporates an isolated forest model and an autoencoder model to perform dual-model fusion detection on multi-dimensional behavioral feature vectors. It outputs detection results for global distribution anomalies and local reconstruction anomalies respectively, and then merges them to output the anomaly behavior judgment result. The attack chain detection module is used to perform attack chain pattern matching on a series of consecutive abnormal behaviors within a preset time window, and dynamically adjust the threat level based on the attack chain matching stage, the confidence level of the abnormal behavior, and the number of consecutive occurrences. The protection execution module is used to execute security protection measures of the corresponding level according to the current threat level; The strategy optimization module is used to evaluate the effectiveness of security protection measures and adaptively adjust the detection parameters of the AI ​​anomaly detection module based on the evaluation results, forming a closed-loop operation mechanism consisting of anomaly detection, hierarchical protection, and adaptive parameter optimization.

10. The power terminal abnormal behavior detection and security protection system according to claim 9, characterized in that, It also includes a power physical consistency verification module, which has built-in three types of physical verification rules: power balance constraints, voltage-current phase relationship constraints, and energy-power integral consistency constraints. These rules are used to cross-verify the physical laws of the abnormal behavior to be confirmed, and to adjust the comprehensive confidence and threat level of the abnormal behavior based on the verification results. In the AI ​​anomaly detection module, the isolated forest model and the autoencoder model adopt a differentiated feature input mechanism, that is, the isolated forest model inputs a subset of global statistical features, and the autoencoder model inputs a subset of time sequence features.