Security protection method of industrial control system, electronic device, storage medium and program product

Abstract

This application provides a security protection method, electronic device, storage medium, and program product for an industrial control system, relating to the field of security technology. The method dynamically perceives production load and attack trends through a deep reinforcement learning model, adaptively selecting the current combination of execution entities from a heterogeneous execution entity cluster, achieving a dynamic balance between protection strength and production continuity. Simultaneously, based on the fusion and adjudication of execution results and real-time operating status data, it can accurately identify output deviations caused by sensor noise or normal process fluctuations, reducing the false judgment rate and improving the inherent security protection capability of the industrial control system under complex operating conditions.

Description

Technical Field

[0001] This application relates to the field of security technology, and more specifically, to a security protection method for an industrial control system, electronic equipment, storage medium, and program product. Background Technology

[0002] In the existing field of industrial control system security, traditional defense solutions are mainly based on static, clearly defined security architectures. Typical technologies include industrial firewalls, isolation gateways, and feature-matching intrusion detection systems. The core logic of these solutions lies in matching known attack characteristics using a predefined rule base, or blocking unauthorized access through fixed network boundaries. However, this static protection model has obvious structural flaws: its protection rules are fixed and predictable, and attackers can exploit rule vulnerabilities through long-term reconnaissance. Summary of the Invention

[0003] The purpose of this application is to provide a security protection method, electronic device, storage medium, and program product for industrial control systems, so as to improve the problem of poor protection strength of static protection mode in the prior art.

[0004] In a first aspect, embodiments of this application provide a security protection method for an industrial control system, the method comprising: In response to received industrial control commands, a current set of execution entities is determined. This current set of execution entities is selected from a heterogeneous execution entity cluster by a pre-trained deep reinforcement learning model based on the most recently collected production load characteristics and network attack situation characteristics. The heterogeneous execution entity cluster includes multiple execution entities. The industrial control instructions are distributed in parallel to each executor in the current executor group; Receive the execution results returned by each executor and the real-time running status data of each executor; The attack determination result is determined based on the execution result and the real-time running status data.

[0005] In the above implementation process, a deep reinforcement learning model dynamically perceives production load and attack status, adaptively selecting the current combination of execution entities from a heterogeneous execution entity cluster, thus achieving a dynamic balance between protection strength and production continuity. Simultaneously, based on the fusion and adjudication of execution results and real-time operating status data, it can accurately identify output deviations caused by sensor noise or normal process fluctuations, reducing the false judgment rate and improving the inherent security protection capability of the industrial control system under complex operating conditions.

[0006] Optionally, the deep reinforcement learning model is further used to determine a switching frequency based on the most recently collected production load characteristics and network attack situation characteristics, wherein the switching frequency is used to determine the feature collection cycle, and the determination of the current executor combination in response to received industrial control commands includes: In response to the received industrial control command, the current combination of execution entities determined by the deep reinforcement learning model in the most recent acquisition period is obtained.

[0007] In the above implementation process, a deep reinforcement learning model simultaneously outputs the current execution entity combination and the switching frequency. The switching frequency is used to dynamically set the acquisition cycle of production load characteristics and attack situation characteristics, enabling the model to adaptively adjust the rhythm of state updates according to real-time operating conditions. For example, under high load or critical operating conditions, the model outputs a lower switching frequency, extending the acquisition cycle, thereby reducing unnecessary model inference and execution entity replacement, ensuring the continuity and low latency of the instruction processing flow; under low load or high threat scenarios, the model outputs a higher switching frequency, shortening the acquisition cycle, enabling the system to quickly respond to environmental changes and update the execution entity combination in a timely manner, enhancing the dynamic change capability of the attack surface.

[0008] Optionally, the deep reinforcement learning model is a DQN model, and its reward function during the training phase is: , This represents the weighting coefficient, adjusted based on production load characteristics. Indicator of production smoothness This indicates the strength index of safety protection.

[0009] Optionally, the The adjustment method is as follows: when the production load characteristics indicate that the industrial control system is in a high-load state or a critical operating condition, the adjustment is increased. When the production load characteristics indicate that the industrial control system is in a low-load state, the load should be reduced. .

[0010] Optionally, determining the attack determination result based on the execution result and the real-time running status data includes: The health score of each executor is determined based on the real-time operating status data. Based on the health score, the execution results are weighted and voted to obtain candidate results; When the voting consensus is not met, the validity of the candidate results is verified based on industrial process constraints or physical logic laws to determine the attack judgment result.

[0011] In the aforementioned implementation process, by introducing a dynamic evaluation and weighted voting mechanism based on the health score of the executor, executors with better health status have greater say in the adjudication, effectively improving the robustness of the adjudication results to abnormal executors or potentially damaged units. When the weighted voting cannot reach a consensus, it is not directly judged as an attack. Instead, the validity of the candidate results is further verified by utilizing industrial process constraints or physical logic laws, thereby distinguishing between normal process fluctuations caused by sensor noise, environmental disturbances, or differences in the resolution of heterogeneous executors and genuine malicious attacks.

[0012] Optionally, the step of obtaining candidate results by weighted voting on the execution results based on the health score includes: The voting weight of each executor is determined based on the health score, and the voting weight is positively correlated with the health score; The execution results are weighted according to the voting weights to obtain a consistency score, and the candidate result with the highest consistency score and its vote weight percentage are determined.

[0013] In the above implementation process, by establishing a voting weight mechanism positively correlated with the health score of the executor, executors with healthier status and more reliable operation have greater say in the adjudication process. This effectively suppresses the interference of output deviations caused by individual executor performance degradation, resource leakage, or minor attacks on the overall adjudication result. By selecting candidate results based on the weighted consistency score and calculating their vote weight percentage, the consensus strength among the outputs of each executor can be quantitatively characterized, providing a clear and quantifiable basis for determining whether to proceed to process verification.

[0014] Optionally, when the voting consensus is not met, the candidate results are validated based on industrial process constraints or physical logic laws to determine the attack determination result, including: If the weight of the votes is lower than the set percentage, then the consistency of voting is not satisfied. Obtain the process attribute parameters corresponding to the candidate results; If the process attribute parameters meet the industrial process constraints or physical logic laws, then the attack determination result is determined to be that no attack has occurred. If the process attribute parameters do not conform to industrial process constraints or physical logic laws, the attack determination result is that the system has been attacked.

[0015] In the above implementation process, process verification is triggered when the weighted voting percentage falls below a set percentage, rather than directly identifying it as an attack. By extracting the process attribute parameters of candidate results and comparing them with industrial process constraints or physical logic laws, the system can effectively distinguish reasonable output deviations caused by sensor noise, differences in the analysis of heterogeneous actuators, or normal process fluctuations from genuine malicious attacks. When the process attribute parameters conform to physical laws, even if there are discrepancies among actuators, the system still determines that it has not been attacked and adopts the candidate result, thereby significantly reducing the false alarm rate caused by normal process fluctuations in traditional mimicry adjudication methods and avoiding unnecessary actuator reconstruction and alarm interference. Conversely, when a candidate result violates process constraints, the system accurately determines that it has been attacked, even if all actuator outputs are completely consistent.

[0016] Optionally, the real-time operating status data includes: CPU load rate, memory utilization rate, and instruction execution latency, and the health score of each executor is determined by the following calculation formula:

[0017] Where H represents the health score, This represents the rate of change of CPU load rate compared to the standard load rate. This indicates the rate of change in memory usage compared to the standard memory usage. This represents the rate of change of instruction execution latency compared to the standard metric execution latency. , , This represents the weight coefficient corresponding to each item.

[0018] Optionally, after determining the health score of each execution entity based on the real-time operating status data, the method further includes: If the health score of an executor is lower than a set score, the executor is removed from the heterogeneous executor pool. The deep reinforcement learning model is used to select a combination of executors from the heterogeneous executor pool. Call an executor with the same heterogeneous attributes as the executor from the standby executor cluster and add it to the heterogeneous executor pool.

[0019] In the above implementation process, through the dynamic replacement and replenishment mechanism, the system can automatically isolate unhealthy executors and replace them with healthy executors of the same type in a timely manner. This not only avoids the interference of damaged executors with the adjudication results, but also maintains the integrity and diversity of the heterogeneous executor pool, providing reliable candidate resources for deep reinforcement learning models.

[0020] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps of the method provided in the first aspect above are performed.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method provided in the first aspect above.

[0022] Fifthly, embodiments of this application provide a computer program product, including computer program instructions, which, when read and executed by a processor, perform the steps of the method provided in the first aspect above.

[0023] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This application provides a schematic diagram of the hierarchical structure of a security protection system. Figure 2 A schematic diagram of the scheduling process of a dynamic mimicry scheduling layer provided in this application embodiment; Figure 3 This application provides a schematic diagram of the execution flow of an industrial heterogeneous execution layer. Figure 4 A flowchart illustrating a security protection method for an industrial control system provided in this application embodiment; Figure 5 A schematic diagram of the adjudication process of the process fusion adjudication and optimization layer provided in this application embodiment; Figure 6 A structural block diagram of a safety protection device for an industrial control system provided in this application embodiment; Figure 7 This is a schematic diagram of the structure of an electronic device for performing a security protection method for an industrial control system, as provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0027] It should be noted that the terms "system" and "network" in the embodiments of this invention can be used interchangeably. "Multiple" refers to two or more; therefore, in the embodiments of this invention, "multiple" can also be understood as "at least two". "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0028] It should also be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0029] This application provides a security protection method for an industrial control system. This method dynamically perceives production load and attack trends using a deep reinforcement learning model, adaptively selecting the current combination of actuators from a heterogeneous actuator cluster, thus achieving a dynamic balance between protection strength and production continuity. Simultaneously, based on the fusion and adjudication of execution results and real-time operating status data, it can accurately identify output deviations caused by sensor noise or normal process fluctuations, reducing the false judgment rate and improving the inherent security protection capability of the industrial control system under complex operating conditions.

[0030] To facilitate understanding of the method embodiments, the safety protection system will be introduced first (the system mentioned in subsequent embodiments refers to this safety protection system; if the industrial control system is mentioned, its full name will be used). Figure 1 As shown, the security protection system includes an instruction adaptation layer, a dynamic mimicry scheduling layer, an industrial heterogeneous execution layer, and a process fusion adjudication and optimization layer.

[0031] The instruction adaptation layer serves as the entry point for the security protection system. It can directly connect to the control bus of the industrial control system, receive raw industrial control instructions, and unify messages from different protocols into an internally processable format through protocol parsing and standardization conversion. It also performs integrity verification and encryption on the instructions to prevent them from being tampered with during transmission within the system.

[0032] Specifically, the instruction adaptation layer, serving as the system's external interface, connects to the control bus of the industrial control system via a high-performance network processor. When a raw industrial control instruction (such as "set the wire drying temperature to 80℃") arrives in Modbus RTU message form, the instruction adaptation layer first performs deep packet inspection on the message using its built-in industrial protocol adaptive parsing algorithm. This algorithm, based on a state machine model, gradually reconstructs the application layer function code, address field, data field, and cyclic redundancy check code from the physical layer bitstream, thereby identifying the protocol type as Modbus RTU. Subsequently, the instruction adaptation layer converts the parsed information into a standardized format unified within the system (e.g., JSON structure: {"protocol":"modbus","command":"set_temp","value":80,"unit":"℃"}). To ensure the security of instructions during internal system transmission, the instruction adaptation layer also integrates an SM4 encryption chip to perform integrity verification (such as hash value calculation) and encryption processing on standardized tasks. Finally, the instruction adaptation layer sends the processed instructions to the dynamic mimicry scheduling layer via an asynchronous message queue, avoiding front-end blocking due to a surge in instruction concurrency.

[0033] The dynamic mimicry scheduling layer is the decision-making center of the security protection system. It integrates a pre-trained deep reinforcement learning model. This model takes collected production load characteristics (such as throughput and concurrent connections) and attack situation characteristics (such as abnormal message frequency and threat level) as input, and outputs the optimal combination of execution entities. It can also output a switching frequency, which controls the update time of the execution entity combination; it can also refer to the frequency of feature collection, achieving an adaptive balance between protection strength and production continuity. The dynamic mimicry scheduling layer and the instruction adaptation layer are decoupled through an asynchronous message queue. This asynchronous architecture ensures that when industrial instruction concurrency surges, the scheduling layer will not cause front-end instruction blocking due to limited processing capacity.

[0034] Specifically, such as Figure 2As shown, the dynamic mimicry scheduling layer retrieves tasks (instructions) from the asynchronous message queue, but its core function is not to make immediate decisions for a single instruction. Instead, it periodically (e.g., every 5 seconds) runs its internal deep reinforcement learning model. At the beginning of each cycle, the scheduling layer collects two sets of features from sensors and network monitoring probes in the industrial field: one is production load features, including the real-time throughput of the current production line (e.g., the number of pieces processed per hour) and the number of concurrent controller connections, mapped to the [0,1] interval after Max-Min normalization; the other is network attack situation features, such as the frequency of abnormal packets and the number of unauthorized access attempts in the past few seconds, quantized into 1 to 5 levels (level 1 is no threat, level 5 is high risk) according to a preset security knowledge base. These two sets of features are concatenated into a state vector, which is then normalized and quantized before being input into the deep reinforcement learning model that has been trained offline. After forward propagation calculation, the model outputs two decisions: (a) the selected executor combination, for example, selecting three executors from a heterogeneous executor pool: executor A (x86 architecture + Linux real-time system), executor B (ARM architecture + RTOS), and executor C (PowerPC architecture + domestic industrial system); (b) the switching frequency, for example, if the current production load is high and the attack situation is weak, the model outputs a frequency of 0.2Hz (i.e., changing the combination every 5 seconds). This decision is written into the configuration table of the scheduling layer, becoming the effective scheduling strategy for the current period.

[0035] The industrial heterogeneous execution layer is the task execution unit of the security protection system. It contains multiple functionally equivalent but internally different executors (with heterogeneous hardware, operating systems, and firmware versions). Each executor can independently receive and process control commands, and return execution results and its own operating status (CPU, memory, latency, etc.).

[0036] Specifically, the industrial heterogeneous execution layer comprises a cluster of multiple heterogeneous executors. Each executor is a physically isolated computing unit (with its own CPU and memory) and pre-installed with different operating systems and protocol stacks. After the instruction adaptation layer completes the standardization process, standardized instructions can be directly distributed in parallel to all executors in the currently active executor combination via industrial real-time Ethernet (e.g., within a 5-second cycle, the combination is fixed as {A, B, C}). Each executor independently executes the same instructions: for example, executor A uses its built-in Modbus protocol stack to convert the temperature setpoint 80℃ into the corresponding register write operation; executor B, running an RTOS, may have slightly different protocol parsing logic, but ultimately also outputs a control signal. After execution, each executor packages the execution result (such as the output temperature target value) and real-time operating status data (such as the current CPU load rate, memory utilization rate, and response latency of this instruction) and sends them to the process fusion adjudication and optimization layer through a dedicated feedback channel. Meanwhile, the autonomous management module built into the industrial heterogeneous execution layer continuously monitors the health of each execution entity: when the health score of an execution entity (calculated based on the rate of change of CPU, memory, and latency) is lower than a preset threshold (such as 80 points), it is automatically marked as sub-healthy, and a new execution entity with the same heterogeneous attributes is called from the standby pool to take over its work, ensuring the continuous reliability of the cluster.

[0037] In some implementations, such as Figure 3 As shown, in the industrial heterogeneous execution layer, each executor integrates an industrial protocol adaptive parsing module. This module uses a predefined message structure description language (e.g., XML or ASN.1 format to describe the offset, length, and type of each protocol field) to automatically deconstruct and extract features from the received raw industrial control messages (i.e., instructions). Extracted features include, but are not limited to, function codes (indicating operation type, such as read / write), address fields (specifying the target register or data point), data fields (carrying specific control parameters or setpoints), and cyclic redundancy check codes (used for integrity verification). Based on these features, the module constructs a unified message feature vector that uniquely identifies control instructions from different protocols (such as OPC UA, Modbus, S7, DNP3, etc.). By matching the feature vector with a built-in protocol feature library, the module automatically identifies multiple industrial protocols and converts heterogeneous messages into a standardized intermediate representation format (such as JSON or Protocol Buffers) within the system. This standardized conversion process eliminates semantic ambiguity caused by differences in the internal processing logic of different executors (such as different byte order, data structure definitions, or parsing algorithms), ensuring that all executors have a completely consistent understanding and execution of the same instruction, thereby providing a reliable input basis for subsequent weighted voting and process verification.

[0038] The process integration adjudication and optimization layer is the final judgment and feedback unit of the security protection system. It collects the output results and operating status of each actuator via a high-performance backplane bus. Through a multi-dimensional adjudication mechanism combining weighted voting and industrial process constraint verification, it distinguishes between normal output, normal process fluctuations, and malicious attacks, and triggers actuator reconstruction and alarms when an attack is detected. For example, if the adjudication detects an anomaly, the reconstruction module will use a golden mirror to quickly restore the damaged actuator, while the alarm module will push threat intelligence to the security management platform through the network gateway.

[0039] To ensure reliable system operation in harsh industrial environments, this solution standardizes data exchange between different layers, defining clear API interfaces and communication protocols. Specifically, the instruction adaptation layer and the dynamic mimicry scheduling layer use a dedicated task distribution primitive for decoupling interaction. This primitive encapsulates instruction priority, timestamp, and integrity verification information, ensuring efficient delivery and traceability of scheduled tasks. The dynamic mimicry scheduling layer and the industrial heterogeneous execution layer employ an isochronous synchronous transmission protocol. Through precise time synchronization mechanisms and reserved bandwidth, this ensures deterministic end-to-end delays in control instruction transmission, preventing execution response timeouts or instruction out-of-order issues caused by network jitter. At the physical transmission medium level, critical system links (such as the instruction distribution link from the scheduling layer to the execution layer and the status feedback link from the execution layer to the adjudication layer) use Category 6 shielded twisted-pair cable or multimode fiber to reduce signal attenuation and external crosstalk. Meanwhile, all core equipment (including the scheduling server, execution cluster, and adjudication hardware) is deployed in an electromagnetically shielded enclosure, combined with grounding and filtering design, effectively resisting strong electromagnetic interference generated by high-power frequency converters and motor start-ups and shutdowns in industrial settings. These measures comprehensively ensure the signal-to-noise ratio of signal transmission, guarantee data integrity during transmission, and meet the stringent requirements of industrial control systems for real-time performance, determinism, and reliability.

[0040] Through the collaborative work of the above four levels, the security protection system realizes the integration of command access, intelligent scheduling, heterogeneous execution and fusion adjudication, effectively improving the inherent security capabilities of industrial control systems in the face of unknown threats and dynamic operating conditions.

[0041] Before implementing the security protection methods of this solution, a deep engineering deconstruction of the target industrial scenario is necessary. Specifically, the core production processes of the scenario need to be clearly defined. For example, in the tobacco production embodiment, the drying temperature control logic requires an accuracy of ±0.5℃ and a sampling frequency of 10Hz; in the power system embodiment, the grid voltage regulation logic requires millisecond-level response to load fluctuations. Then, the industrial control protocols used in the field are systematically reviewed, including but not limited to Modbus RTU / TCP, OPC UA binary encoding, S7 series proprietary protocols, DNP3, and IEC 61850 communication specifications. Based on this, a real-time response threshold (usually set to no more than 50ms) is determined based on bus contention rate and controller processing cycle to ensure that the intervention of security protection links does not cause timeout errors in the industrial control loop. At the same time, by deeply mining historical attack data, the detection behavior in advanced persistent threats and tampering attacks targeting process parameters (such as illegal write instruction injection) are analyzed to determine the protection priority, setting tasks involving safety instrumented systems and critical logic control as the highest priority. The final output is an industrial scenario requirements specification, which serves as a blueprint for subsequent architecture design. It clearly records the required types and quantities of heterogeneous actuators, the millisecond threshold for scheduling response, and the process constraint parameters that must be followed during the adjudication process (such as the upper limit of the partial derivative of temperature change, the legal range of voltage fluctuation, etc.).

[0042] Based on the above requirements specifications, the next step is to construct an industrial security heterogeneous execution unit cluster. The heterogeneity of this cluster is reflected in three dimensions: In terms of hardware architecture, three distinct architectures are selected: x86 (Intel Core series), ARM (Cortex-A72), and PowerPC (QorIQ series), eliminating common vulnerabilities at the instruction set level; in terms of operating systems, a high-efficiency system based on the Linux kernel, a real-time operating system based on a microkernel architecture, and an industrial control operating system based on a domestically developed kernel are adopted; and in terms of firmware versions, the driver programs and protocol stack source code for each execution unit are logically distinct. At the software logic architecture level, an instruction adaptation layer, a dynamic mimicry scheduling layer, an industrial heterogeneous execution layer, and a process fusion adjudication and optimization layer are designed from bottom to top. Among them, the instruction adaptation layer integrates an industrial protocol adaptive parsing algorithm, which performs deep packet inspection on messages through a state machine model to achieve accurate reconstruction from physical layer bit stream to application layer function code; the core engine of the dynamic mimicry scheduling layer is a scheduling algorithm based on deep reinforcement learning, which can dynamically adjust the combination of execution bodies according to environmental feedback; the process fusion adjudication and optimization layer embeds a multi-dimensional feature fusion adjudication algorithm, which upgrades the traditional bit-level comparison to logic-level verification.

[0043] After the architecture design was completed, each module was physically located in the industrial field according to the design plan. The heterogeneous execution unit cluster was connected to the control network through a high-performance industrial switch, the scheduling server adopted a redundant dual-machine hot standby configuration, and the adjudication module was hardware-deployed in a high-performance FPGA or dedicated computing card. During the execution instruction adaptation and debugging phase, the protocol parsing algorithm performed full-text parsing of the captured industrial control instructions and verified them against the preset standard message template. Subsequently, multi-scenario verification was carried out to simulate different production loads (such as full-load production and downtime maintenance) and attack scenarios of different intensities (such as DoS attacks, replay attacks, and man-in-the-middle attacks), and the scheduling latency, the stuttering rate during task switching, and the adjudication accuracy were obtained in real time. Based on the quantitative data from the test feedback, the neural network weight parameters in the scheduling strategy and the process constraint thresholds in the adjudication logic were finely adjusted until all indicators met the requirements for industrial-grade deployment.

[0044] Please refer to Figure 4 , Figure 4 A flowchart of a security protection method for an industrial control system provided in this application embodiment is included, the method comprising the following steps: Step S110: In response to the received industrial control command, determine the current combination of actuators.

[0045] The current set of executors is selected from a heterogeneous cluster of executors by a pre-trained deep reinforcement learning model based on the most recently collected production load characteristics and network attack situation characteristics. It can be understood as a group of executors currently selected to process industrial control commands, typically consisting of 2-3 executors. This group is dynamically generated by the deep reinforcement learning model and remains fixed within a certain time window.

[0046] Deep reinforcement learning models can employ agent models pre-trained using the DQN (Deep Q-Network) algorithm, or other types of reinforcement learning models, such as agent models pre-trained using the Proximal Policy Optimization (PPO) algorithm. Their inputs are production load characteristics and network attack situation characteristics, and their output is the optimal combination of executors.

[0047] Production load characteristics are parameters used to reflect the real-time load on the industrial site, including but not limited to production line throughput and the number of concurrent controller connections. These parameters can be normalized and used as inputs to the model state.

[0048] Network attack situation characteristics are parameters used to reflect the current network threat level. Based on the security knowledge base, abnormal packet frequency, unauthorized access attempts, etc. are quantified into 1-5 levels and used as input for the model status.

[0049] Before the system goes live, a deep reinforcement learning model can be trained offline. This model takes production load characteristics and network attack situation characteristics as input and outputs the optimal combination of execution entities under the current conditions. After training, the model is deployed in a dynamic mimicry scheduling layer. In some implementations, the model runs periodically, but it can also run in real time. At the beginning of each period, the latest production load characteristics and network attack situation characteristics are collected and input into the model for prediction. The model outputs the combination of execution entities to be used in that period, which is marked as the current combination and remains in effect until the next period.

[0050] The heterogeneous execution cluster consists of multiple execution units, each of which may differ in hardware architecture (such as x86, ARM, PowerPC), operating system (such as real-time Linux, RTOS, and domestic operating system) and firmware version. The aim is to eliminate common vulnerabilities and improve the inherent security capabilities of the system.

[0051] When the system receives an industrial control command from the industrial control bus, it determines the current execution entity combination. For example, the system can directly read the current execution entity combination output from the most recent model inference (e.g., the combination contains three execution entities: one running a Linux real-time system based on x86 architecture, one running an RTOS based on ARM architecture, and one running a domestic industrial operating system based on PowerPC architecture). This combination remains unchanged within a fixed time window after model inference, and the model calculation is not retried due to a single command.

[0052] Of course, if the feature collection is periodic, the deep reinforcement learning model can be input for inference after each collection of production load features and network attack situation features, and the execution body combination can be output. During this period, if an industrial control command is received, the current execution body combination is the most recently output execution body combination of the model, and the latest execution body combination will be output again based on the model when the next feature collection cycle arrives.

[0053] For example, in Example 1, the current execution entity combination is output based on the model's runtime cycle. At time t1, the model's runtime cycle ends, and production load characteristics and network attack situation characteristics are collected and input into the model for inference, resulting in execution entity combination 1. At time t2, the model's runtime cycle ends again, and production load characteristics and network attack situation characteristics are collected again and input into the model for inference, resulting in execution entity combination 2. Therefore, after receiving industrial control commands during the period from time t1 to time t2, the current execution entity combination is execution entity combination 1, and the current execution entity combination at time t2 becomes execution entity combination 2.

[0054] Example 2: The current execution entity combination is output based on the data acquisition cycle. At time t3, the data acquisition cycle arrives, and the production load characteristics and network attack situation characteristics are collected and input into the model for inference, resulting in execution entity combination 3. At time t4, the data acquisition cycle arrives again, and the production load characteristics and network attack situation characteristics are collected again and input into the model for inference, resulting in execution entity combination 4. Therefore, after receiving industrial control instructions during the period from time t3 to time t4, the current execution entity combination is execution entity combination 3, and the current execution entity combination at time t4 becomes execution entity combination 4.

[0055] Regardless of the method, the system can maintain an execution entity combination cache state to store the latest execution entity combination. After the deep reinforcement learning model outputs the optimal execution entity combination each time, it can store the identifiers of each execution entity in the combination into the cache state in a completely overwritten manner. In this way, the cache state always stores the latest execution entity combination. Each time an industrial control command is received, the identifiers of the execution entities can be directly read from the cache state to determine the current execution entity combination.

[0056] Step S120: Distribute industrial control instructions in parallel to each executor in the current executor group.

[0057] Before distributing instructions, the system can first perform protocol adaptation and standardization conversion on the received raw instructions. For example, it can convert Modbus RTU messages into an internally unified JSON format through deep packet inspection. Then, the industrial control instructions are sent to each executor in the current executor combination. Each executor receives and processes the instruction independently. Since these executors differ in hardware, operating system and firmware version, their internal processing paths and potential vulnerabilities are different.

[0058] Step S130: Receive the execution results returned by each executor and the real-time running status data of each executor.

[0059] The execution result of each actuator refers to the control value or instruction response output by each actuator after independently processing the same industrial control instruction, such as temperature setpoint, voltage adjustment step size, etc.

[0060] Real-time running status data refers to dynamic information about each executor during runtime, including CPU load rate, memory occupancy, instruction execution latency, etc., which can be used to calculate the health score of the executor.

[0061] After each executor completes instruction processing, it returns its execution results and real-time operating status data. This data can be collected and transmitted to the process fusion adjudication and optimization layer.

[0062] Step S140: Determine the attack judgment result based on the execution result and real-time running status data.

[0063] The attack determination result is a conclusion drawn from the analysis of the execution result of the executor based on a multi-dimensional feature fusion adjudication mechanism. It can be divided into two categories: not attacked and attacked, and also into three categories: normal output, normal process fluctuation and malicious attack.

[0064] For example, deviations between execution results can be determined, and real-time operational status data can be used to assess the reliability of the executor (e.g., using a health score to characterize reliability in subsequent embodiments). If the deviation is within a set range and the reliability is greater than a set threshold, it is considered that no attack has occurred; otherwise, it is considered that an attack has occurred. If an attack is determined, the system can trigger executor reconstruction (e.g., calling a new heterogeneous executor from the backup pool to replace the contaminated executor) and send an alarm to the security management platform; if no attack is determined, i.e., normal output or normal process fluctuations, then legitimate instructions are output to field equipment for execution.

[0065] In the above implementation process, a deep reinforcement learning model dynamically perceives production load and attack status, adaptively selecting the current combination of execution entities from a heterogeneous execution entity cluster, thus achieving a dynamic balance between protection strength and production continuity. Simultaneously, based on the fusion and adjudication of execution results and real-time operating status data, it can accurately identify output deviations caused by sensor noise or normal process fluctuations, reducing the false judgment rate and improving the inherent security protection capability of the industrial control system under complex operating conditions.

[0066] Based on the above embodiments, the deep reinforcement learning model is also used to determine the switching frequency according to the most recently collected production load characteristics and network attack situation characteristics. The switching frequency is used to determine the feature collection cycle. When determining the current execution body combination, the current execution body combination determined by the deep reinforcement learning model in the most recent collection cycle can be obtained.

[0067] In this implementation, the deep reinforcement learning model is used not only to select the current combination of executors but also to determine a switching frequency. This switching frequency is used to dynamically set the feature acquisition cycle, meaning the system periodically acquires the latest production load characteristics and network attack situation characteristics from the industrial site at this frequency. For example, when the model output switching frequency is 0.2Hz, the corresponding acquisition cycle is 5 seconds, and the system performs feature acquisition and model inference every 5 seconds; when the attack situation intensifies and the model increases the switching frequency to 1Hz, the acquisition cycle is shortened to 1 second, and the system updates the status information at a higher frequency.

[0068] At the start of each acquisition cycle, the system performs the following operations: First, it acquires production load characteristics and network attack situation characteristics within the current cycle, according to the current cycle length (determined by the switching frequency output by the model in the previous cycle). Then, it inputs these characteristics into a deep reinforcement learning model, which outputs two sets of decisions after forward propagation: one is the current execution entity combination (e.g., selecting execution entities A, B, and C from a heterogeneous execution entity pool), and the other is the switching frequency for the next cycle (e.g., maintaining 0.2Hz or adjusting to 0.5Hz). Subsequently, within the current cycle, the system marks this execution entity combination as the current execution entity combination, i.e., the effective combination, and all arriving industrial control commands are distributed in parallel to the respective execution entities in this combination for processing. When the system needs to respond to a command and execute step S110 (determining the current execution entity combination), it directly reads the effective combination set in the most recent model inference, without needing to rerun the model. In this way, the switching frequency determines the rhythm of feature acquisition and model update, and implicitly controls the replacement cycle of the execution body combination. Whenever a new acquisition cycle arrives and model inference is completed, the effective combination is updated to the latest output combination, thereby achieving the effect of dynamic change of the execution body combination with the working conditions and adaptive change frequency to production load and attack situation.

[0069] For example, during peak electricity consumption periods in the power system (high production load, low attack status), the switching frequency of the deep reinforcement learning model output is relatively low, such as 0.1Hz (10-second cycle). The system collects features and re-decides every 10 seconds, keeping the current agent combination unchanged during these 10 seconds, ensuring the stability of instruction distribution under high load. When frequent anomaly scans are detected (attack status escalates to level 4) and production load drops to a low point, the model increases the switching frequency to 0.5Hz (2-second cycle). The system updates features and reselects the agent combination every 2 seconds, causing the attack surface to change rapidly and effectively suppressing the attacker's reconnaissance behavior.

[0070] In the above implementation process, the deep reinforcement learning model simultaneously outputs the current execution body combination and switching frequency, and uses the switching frequency to dynamically set the collection cycle of production load characteristics and attack situation characteristics, so that the model can adaptively adjust the rhythm of state updates according to real-time working conditions.

[0071] Based on the above embodiments, the deep reinforcement learning model is the DQN model, and its reward function during the training phase is: , This represents the weighting coefficient, adjusted based on production load characteristics. Indicator of production smoothness This indicates the strength index of safety protection.

[0072] During the training phase of the deep reinforcement learning model, a training platform simulating an industrial control environment can be constructed. This platform can simulate different production loads and diverse attack scenarios. The goal of training is to enable the DQN model to learn to select the optimal scheduling action (executor combination and switching frequency) based on the current state (production load characteristics and network attack scenario characteristics) to maximize cumulative rewards.

[0073] The input state of the DQN model is composed of production load characteristics and network attack situation characteristics. Its output scheduling actions include selecting a set of executors from a heterogeneous execution pool and setting the switching frequency. Its reward function is used to evaluate the quality of the actions.

[0074] It is a production smoothness index, which can be calculated by monitoring the delay variance of the command flow and the dispersion of process parameters (such as temperature and pressure). It reflects the impact of scheduling actions on production stability. The higher the value, the smaller the disturbance caused by switching.

[0075] Specifically, the latency variance of the instruction stream can be measured by monitoring the uniformity of instruction arrival and processing times over a period of time (such as within a switching cycle). This can be quantified by calculating the variance or standard deviation of the time interval between adjacent instructions. The smaller the variance, the smoother the instruction stream is, and the less noticeable the jitter caused by execution switching.

[0076] The dispersion of process parameters can be assessed by monitoring the fluctuation range of key process variables (such as temperature, pressure, and voltage) during the switching cycle. Specifically, the deviation between the actual and ideal rates of change of process parameters can be calculated, or the variance of the parameter series can be directly calculated.

[0077] The two dimensions can then be weighted and fused to obtain the production smoothness index.

[0078] This indicator represents the strength of security protection and can be evaluated based on factors such as the nonlinearity of the switching frequency (whether it is difficult to predict), attack interception rate, and unknown anomaly detection rate. The higher the value, the better the defense effect.

[0079] The nonlinearity of the switching frequency refers to the unpredictability of the time interval between execution unit switching. If an attacker can predict the timing of the switching, they may be able to precisely select the switching gap to launch an attack. Therefore, the higher the nonlinearity of the switching frequency (i.e., the more random or difficult to model the switching interval changes), the harder it is for an attacker to grasp the attack window. This can be quantified by calculating the information entropy of the actual switching interval sequence or the degree of deviation from a fixed periodic pattern.

[0080] Regarding the interception rate of simulated attacks, during the training phase, known attack samples (such as DoS attacks, replay attacks, man-in-the-middle attacks, and abnormal protocol messages) can be injected, and the proportion that the system can successfully identify and intercept can be statistically analyzed. A higher interception rate indicates that the current scheduling strategy, combined with the adjudication mechanism, can effectively block attacks.

[0081] The detection rate of unknown anomalous messages refers to the proportion of messages that the system can identify that do not conform to any known attack signature but still deviate from normal industrial protocol behavior. This metric reflects the system's ability to detect zero-day exploits and new types of attacks. Specifically, it can be evaluated using unsupervised anomaly detection models (such as anomaly scoring based on protocol state machines or autoencoders): unknown anomalous samples are mixed into the test dataset, and the system is counted on the percentage of those it marks as anomalous.

[0082] The security protection strength index can then be obtained by weighting and fusing the three dimensions mentioned above.

[0083] During the training phase of the DQN model, before training begins, the main network and target network of DQN are initialized, and the experience replay pool capacity and exploration rate are set. (Initially set to 1.0, gradually decreasing to 0.01 during training), discount factor (e.g., 0.95) and batch size (e.g., 32). Meanwhile, multiple operating condition combinations are preset in the simulation environment: production load is divided into low load (throughput <30%), medium load (30%-70%), and high load (>70%); attack scenarios are divided into five levels: no attack, light scanning, medium DoS, replay attack, and APT penetration.

[0084] In each training episode, the environment is first randomly initialized with a working condition (e.g., high load + medium DoS attack), and an initial state S0 is given (containing normalized production load characteristics and attack level values). For each step, the agent... Strategy selection action a t :by A combination of executors and a switching frequency are randomly selected with probability, in order to... The action that produces the highest Q-value in the current network output is selected probabilistically. After the action is executed, the environment updates its state for the next cycle according to the set switching frequency and calculates the immediate reward R. t The reward calculation process is as follows: Based on the switching frequency in the action and the instruction flow within that cycle, simulation calculations are performed. For example, if the switching frequency is too high, it will increase the variance of instruction latency. Decrease; increase if the switching frequency is moderate and the process parameters are stable.

[0085] Based on the feedback from the attack detection module during the execution of the action, calculate If the switching frequency is high and the randomness is strong, the difficulty of attack detection increases. Increase; if the attack is successfully intercepted after switching, an additional bonus is applied.

[0086] Determined based on current production load Values ​​(e.g., under high load) =0.7, under low load =0.4), substitute into the reward function calculation formula. Note that the weight of the security protection strength index in the reward function calculation formula is 1+ This means that the basic weight of safety intensity is always higher than that of production smoothness, reflecting the principle of prioritizing safety while taking stability into account in industrial scenarios.

[0087] Specifically, the weighting coefficients The adjustment method is as follows: when the production load characteristics indicate that the industrial control system is in a high-load state or a critical operating condition, the adjustment is increased. When the production load characteristics indicate that the industrial control system is in a low-load state, the load should be reduced. .

[0088] Dynamic adjustments are made based on production load characteristics to achieve an adaptive balance between production continuity and security defense strength. Production load characteristics include, but are not limited to, the real-time throughput (number of instructions processed per unit time) and concurrent connections (number of controllers or terminals connected simultaneously) of the industrial control system. These characteristics are normalized and mapped to the [0,1] interval, with values ​​closer to 1 indicating higher loads. Critical operating phases refer to special periods in the production process where interruptions or fluctuations are intolerable, such as peak electricity consumption periods in the power system, continuous reaction periods in petrochemical production, or the constant temperature maintenance phase in tobacco drying.

[0089] When the production load characteristic value exceeds the preset high load threshold (e.g., 0.7) or the system explicitly marks it as a critical operating condition, it will... Adjust upwards to the higher value range, typically set to 0.6–0.8 (e.g., take...). =0.7). At this time Increasing the weight of the production smoothness index raises its overall weight, and consequently, the coefficient of the safety protection strength index also increases. When the production load characteristic value is below the preset low load threshold (e.g., 0.3) and the system is in a non-critical operating condition, [the following will occur]. Adjust downwards to a lower value range, typically set to 0.3–0.5 (e.g., take...). =0.4). At this point, the reward function gives a more prominent relative advantage to the safety term, encouraging the model to adopt a more frequent or more random switching strategy to enhance its defense capabilities.

[0090] Before the start of each training round, the model reads the current production load characteristic L and operating condition flag K from the environmental state. If L > 0.7 or K = True, then the parameters are set. =0.7; if L<0.3 and K=False, then set =0.4; For intermediate loads (0.3≤L≤0.7), linear interpolation or maintaining the previous time step value can be used. Values ​​to avoid frequent jumps. The value is used to calculate the immediate reward R obtained after the current action is performed and stored in the experience tuple for network training. Through this dynamic adjustment, the model can perceive during training that choosing conservative, low-frequency switching actions under high load will yield higher rewards (because...). (Weight increase), choosing high-frequency switching actions under low load will yield higher rewards (because) (It has a greater relative advantage), thus autonomously learning scheduling strategies that match the production load.

[0091] Through this dynamic adjustment mechanism, the reward function can guide the model to learn the optimal scheduling strategy based on actual working conditions, achieving an adaptive balance between production stability and safety protection.

[0092] Receive reward R t Then, the environment enters the next state S. t+1 (Updated based on new operating conditions). The experience tuple (S) t ,a t ,R t ,S t+1 The samples are stored in the experience replay pool. Once the number of samples in the replay pool reaches the batch size, a small batch is randomly sampled, the target Q value is calculated, and the parameters of the main network are updated using gradient descent with the mean squared error loss function. The parameters of the main network are copied to the target network every fixed number of steps (e.g., C=100 steps).

[0093] Training continues for multiple rounds until the reward values ​​converge (i.e., the agent consistently obtains high rewards under different operating conditions). For example, under high-load conditions, the agent learns to choose a lower switching frequency (e.g., 0.2Hz) to avoid lag; under low-load and high-threat conditions, the agent chooses a higher switching frequency (e.g., 1Hz). By comparing the cumulative rewards of different actions, the agent eventually learns to dynamically adjust the switching frequency and the combination of executors to achieve an adaptive balance between production continuity and protection strength.

[0094] Therefore, during training, based on the real-time feedback of the reward function R, the dynamic mimicry scheduling layer can output refined decision instructions: First, the executor combination selection decision: the system selects 2 to 3 executors from the candidate heterogeneous executor pool using a pseudo-random selection algorithm or a directional algorithm based on execution history evaluation to form the current executor combination; then, the switching frequency adjustment decision: when facing high-intensity threats, the system sets the switching frequency to a high-frequency mode of 0.5-1Hz to quickly disable the attacker's vulnerability detection; under low-threat and high real-time requirements, the frequency is slowed down to 0.1-0.3Hz to ensure absolute smoothness of task processing.

[0095] After training, the parameters of the main network are fixed and deployed to the dynamic mimicry scheduling layer for the inference phase. During inference, the optimal action is calculated forward based on the current state.

[0096] In the above method of determining the attack judgment result based on the execution result and real-time running status data, the health score of each executor can be determined first based on the real-time running status data. Then, based on the health score, the execution result is weighted and voted to obtain candidate results. If the voting consistency is not satisfied, the validity of the candidate results is verified according to industrial process constraints or physical logic laws to determine the attack judgment result.

[0097] Real-time operational status data may include CPU load rate, memory usage rate, and instruction execution latency, which are monitored and reported by the executor itself.

[0098] The health score is a comprehensive indicator used to quantify the current health status of the executor. The higher the score, the more normal the executor's status, and the higher the reliability of its output results.

[0099] Weighted voting assigns different voting weights to each executor based on their health score. The higher the health score, the greater the voting weight. The execution results returned by all executors are weighted and statistically analyzed. The result with the most votes is called the candidate result, and its voting weight percentage reflects the degree of consensus among the executors.

[0100] Voting consistency refers to whether the weighted percentage of votes for candidate results reaches a preset threshold. If it does, it is considered that there is a high degree of consistency among the implementing bodies, and the candidate results can be directly adopted. If it does not reach the threshold, it indicates that there is a significant disagreement, and further verification is required.

[0101] Industrial process constraints can be predefined physical or logical rules that production industries must follow, such as the temperature change rate must not exceed 2℃ / s, voltage fluctuations must be within ±5% of the rated range, and pressure changes must conform to thermodynamic equations.

[0102] Physical logic laws refer to the laws based on the essence of industrial physical processes, such as the law of inertia, the law of conservation of energy, and the causal relationship of time sequence (the valve must be opened before the material can be injected), which are used to judge whether the execution result violates basic physical common sense.

[0103] After receiving the execution results returned by each executor (for example, the three executors output temperature values ​​respectively, executor A: 80.1℃, executor B: 80.1℃, executor C: 79.9℃) and their real-time running status data (such as CPU load rate, memory usage rate, instruction execution latency), the system first calculates the health score of each executor.

[0104] Specifically, the health score of each executor is determined using the following formula:

[0105] Where H represents the health score, This represents the rate of change of CPU load rate compared to the standard load rate. This indicates the rate of change in memory usage compared to the standard memory usage. This represents the rate of change of instruction execution latency compared to the standard metric execution latency. , , This represents the weight coefficient for each item. The three weight coefficients can be preset according to the role and priority of the executor; for example, for computationally intensive tasks, set... .

[0106] For executor i, its health score can be obtained by substituting it into the above health score calculation formula. Then the calculation result is normalized. For example, executor A has H_A=93 points, executor B has H_B=88 points, and executor C has H_C=76 points.

[0107] Then, the system determines the voting weight of each executor based on its health score (usually allocated proportionally so that the sum of the weights is 1), and performs a weighted vote on the execution results returned by all executors. Figure 5 The adjudication process of the process integration adjudication and optimization layer is as follows: the voting weight is positively correlated with the health score, and then the execution result is weighted according to the voting weight to obtain the consistency score, and the candidate result with the highest consistency score and its voting percentage are determined.

[0108] The consistency score is calculated by summing the voting weights of all executors that output the result for each possible execution result. The higher the score, the stronger the weighted support for the result.

[0109] A simple way to assign weights is: weights , This represents the health score of execution entity i. This represents the sum of the health scores of all executors. In the example above, the sum of their health scores is 93 + 88 + 76 = 257. Based on this formula, executor A has a voting weight of 0.362, executor B has a voting weight of 0.342, and executor C has a voting weight of 0.296. Voting weight can also be negatively correlated with instruction execution latency; that is, the greater the latency, the smaller the weight. One implementation method is... , This indicates the instruction execution delay of executor i. This represents the ratio of the sum of the health scores of all executors to the sum of the instruction execution delays of all executors.

[0110] As in the example above, if the execution result is 80.1℃, and both executors A and B output the result, their consistency score is 0.362 + 0.342 = 0.704. For the execution result 79.9℃, only executor C outputs the result, resulting in a consistency score of 0.296. Then, candidate results are determined. In the example above, the execution result 80.1℃ has the highest consistency score, so it is selected as a candidate result. Its voting percentage is the percentage obtained by dividing the consistency score of the candidate result by the sum of the voting weights of all executors (usually 1), used to measure the degree of consensus among the executors regarding the candidate result. This candidate result has a voting percentage of 70.4%.

[0111] If multiple results tie for the highest score in certain edge cases (e.g., two different results both receive a score of 0.5), the system can pre-set a breaking rule, such as selecting the result with the larger sum of the corresponding executor's health as a candidate result, or randomly selecting one.

[0112] Through the weighted voting process described above, the system comprehensively considers the health status of each executor, giving executors with higher credibility a greater say in the decision-making process, thereby improving the accuracy and resistance to interference of the decision-making results.

[0113] Next, the system determines voting consistency by comparing the weighted percentage of votes for each candidate result with a preset percentage (e.g., 70%). If the weighted percentage is not lower than the preset percentage (e.g., reaching 72%), the candidate result is directly adopted as a valid output and deemed normal, requiring no further verification. If the weighted percentage is lower than the preset percentage (e.g., 63.8% < 70%), it is considered that there is a significant disagreement among the implementing entities, i.e., voting consistency is not satisfied, and the system proceeds to the validity verification step.

[0114] In the validity verification, the system extracts the process attribute parameters corresponding to the candidate results (such as the temperature change rate and voltage adjustment step size under this temperature control command), and calls the preset industrial process constraints and physical logic laws to compare them item by item. If the process attribute parameters meet the industrial process constraints or physical logic laws, the attack judgment result is determined to be that no attack has been made. If the process attribute parameters do not meet the industrial process constraints or physical logic laws, the attack judgment result is determined to be that an attack has been made.

[0115] For example, in a tobacco drying scenario, process constraints require that the temperature change rate not exceed 2℃ / s. If the temperature change rate corresponding to a candidate result is 1.8℃ / s, and its absolute value is within the equipment's rated range, while also conforming to timing logic (e.g., the heater is on), then the candidate result is determined to be a normal process fluctuation. That is, although there are reasonable differences between the actuators (which may stem from heterogeneous hardware resolution accuracy or sensor noise), its output still conforms to physical laws, and the system adopts this result and outputs it to the field equipment. Conversely, if the temperature change rate reaches 3.5℃ / s, far exceeding the process constraints, or if a situation violates physical logic (e.g., the temperature rises abnormally when the heater is not on), then regardless of whether the outputs between the actuators are consistent, it is determined to be a malicious attack.

[0116] In some implementations, when making a decision, the consistency deviation value Δ (the difference between the execution results) can also be obtained to quantify the degree of output difference between different executors.

[0117] During the adjudication process, a consistency deviation value can be introduced for comprehensive judgment. For example, if the vote weight percentage is higher than or equal to a set percentage and the consistency deviation value is less than a set value, the candidate result is directly adopted. If the vote weight percentage is higher than or equal to a set percentage and the consistency deviation value exceeds a set value, it is determined that the voting consistency is not met, and the process proceeds to the validity verification stage. Alternatively, the consistency deviation value can be used in the process verification step to detect whether it conforms to physical laws.

[0118] The adjudication engine can incorporate a physical information-based neural network model or preset process logic operators to perform in-depth validity checks on candidate results. Specifically, if the output data of the executor has low consistency (i.e., a large Δ), but its numerical change curve strictly conforms to preset industrial process thresholds and physical timing logic—for example, reactor pressure changes must follow thermodynamic equations and cannot exhibit abrupt jumps; or the temperature change rate does not exceed 2℃ / s; or the voltage adjustment step size meets inertial constraints—then the system classifies it as a normal process fluctuation. This fluctuation usually originates from differences in the resolution performance of underlying heterogeneous hardware, sensor noise, or normal environmental disturbances, and does not trigger alarms; the system still outputs the candidate result.

[0119] Conversely, even if the output data is completely consistent across all executors (i.e., Δ=0, belonging to a common vulnerability attack scenario), but its value violates the process safety margin (e.g., temperature exceeding equipment limits) or timing logic (e.g., flow is detected before the valve is opened), the adjudication engine will ultimately determine that the system has suffered a coordinated attack. At this time, the system immediately triggers the executor reconstruction module to format and restore the contaminated executors (e.g., reloading firmware from the gold image), and simultaneously sends a high-risk alarm message to the security management platform, including the attack type, timestamp, and identifier of the affected executor.

[0120] To further enhance the adaptability of the ruling, this embodiment differentiates the consistency deviation value Δ in feature extraction for different industrial application scenarios. In the tobacco drying production scenario, considering the inherent hysteresis characteristics of temperature and humidity sensors and the natural non-uniformity of tobacco moisture content distribution, Δ is set to no more than 2% to avoid misjudging normal process fluctuations as attacks. However, in the relay protection scenario of power substations, due to the instantaneous and highly deterministic requirements of electrical signals, any slight deviation may cause the protection device to malfunction. Therefore, Δ is strictly set to no more than 1% to ensure high sensitivity to malicious tampering.

[0121] Through the above process, the system implements an attack judgment mechanism that combines health-weighted voting decisions with validity verification based on process physical constraints. This mechanism can quickly process highly consistent normal outputs and accurately distinguish between harmless process fluctuations and malicious parameter tampering attacks.

[0122] Based on the above embodiments, after determining the health score of each executor, if the health score of an executor is lower than the set score, the executor is removed from the heterogeneous executor pool. The deep reinforcement learning model is used to select a combination of executors from the heterogeneous executor pool, and then call executors with the same heterogeneous attributes as the executor from the standby executor cluster to add them to the heterogeneous executor pool.

[0123] The heterogeneous executor pool is a set of effective executors currently available for deep reinforcement learning models to choose from. When selecting executor combinations, the model selects only from this pool.

[0124] The standby executor cluster is a group of standby executors that share the same heterogeneous attributes (i.e., the same hardware architecture, operating system type, and firmware version characteristics) as the executors in the heterogeneous executor pool. When an executor in the heterogeneous executor pool is removed, the system calls an executor with exactly the same heterogeneous attributes from the standby executor cluster to add to the heterogeneous executor pool, in order to maintain the balance of the number of various heterogeneous executors in the heterogeneous executor pool.

[0125] The system can perform a health assessment on each executor in the heterogeneous executor pool at fixed time intervals, or perform a health assessment after receiving real-time running status data returned by each executor.

[0126] After obtaining the health score of each executor according to the health score calculation method in the above embodiment, its score can be compared with a set score (threshold), which is a predefined lower limit of health, such as 80 points (out of 100). When the health score of an executor is lower than this threshold, the executor is considered to be in a sub-healthy or damaged state and should not continue to participate in scheduling. The system removes it from the heterogeneous executor pool. After removal, the executor will no longer be selected by the deep reinforcement learning model into any executor combination, and the system triggers redundancy replacement logic. If the health score is greater than or equal to the set score, the executor remains in the heterogeneous executor pool and continues to participate in subsequent scheduling selections.

[0127] To maintain the diversity of the number and heterogeneous attributes of executors in the heterogeneous executor pool, the system maintains a backup executor cluster. This cluster pre-deploys a large number of backup executors with the same combination of heterogeneous attributes as those in the heterogeneous executor pool. When an executor is removed, the system searches for and calls a backup executor with identical heterogeneous attributes from the backup executor cluster based on the removed executor's heterogeneous attribute tags (e.g., "ARM architecture, RTOS, protocol stack version V2"). The calling process includes: initializing the backup executor's network configuration and health status, and loading it with the same functional logic as the original executor (e.g., industrial protocol parsing module, control algorithm). Then, the backup executor is added to the heterogeneous executor pool, making it a new candidate for selection by deep reinforcement learning models.

[0128] When a deep reinforcement learning model makes a scheduling decision (i.e., periodically selects the current combination of executors), it selects only from a pool of heterogeneous executors that has undergone health screening. For example, if the pool originally has 6 executors (two each of x86, ARM, and PowerPC), after removing an ARM architecture executor with insufficient health, 5 executors remain in the pool. The system then adds a backup executor with the same ARM architecture from the backup executor cluster to the heterogeneous executor pool, restoring the pool to 6 executors. In this way, the deep reinforcement learning model always faces a complete and healthy set of heterogeneous executors when selecting combinations, thereby ensuring the effectiveness of the scheduling strategy and the inherent security of the system.

[0129] Through the aforementioned dynamic replacement and replenishment mechanism, the system can automatically isolate unhealthy executors and replace them with healthy executors of the same type in a timely manner. This avoids interference from damaged executors with the adjudication results and maintains the integrity and diversity of the heterogeneous executor pool, providing reliable candidate resources for deep reinforcement learning models.

[0130] Please refer to the above method embodiments. Figure 6 , Figure 6 This is a structural block diagram of a safety protection device 200 for an industrial control system provided in an embodiment of this application. The device 200 may be a module, program segment, or code on an electronic device. It should be understood that the device 200 corresponds to the above method embodiment and is capable of performing the various steps involved in the method embodiment. The specific functions of the device 200 can be found in the description above, and detailed descriptions are omitted here to avoid repetition.

[0131] Optionally, the device 200 includes: The combination determination module 210 is used to determine the current execution entity combination in response to the received industrial control command. The current execution entity combination is selected from the heterogeneous execution entity cluster by a pre-trained deep reinforcement learning model based on the most recently collected production load characteristics and network attack situation characteristics. The heterogeneous execution entity cluster includes multiple execution entities. The instruction distribution module 220 is used to distribute the industrial control instructions in parallel to each executor in the current executor combination; The data receiving module 230 is used to receive the execution results returned by each executor and the real-time running status data of each executor; The attack determination module 240 is used to determine the attack determination result based on the execution result and the real-time running status data.

[0132] Optionally, the deep reinforcement learning model is further configured to determine a switching frequency based on the most recently collected production load characteristics and network attack situation characteristics. The switching frequency is used to determine the feature collection cycle. The combination determination module 210 is configured to obtain the current execution combination determined by the deep reinforcement learning model in the most recently collected cycle in response to the received industrial control command.

[0133] Optionally, the deep reinforcement learning model is a DQN model, and its reward function during the training phase is: , This represents the weighting coefficient, adjusted based on production load characteristics. Indicator of production smoothness This indicates the strength index of safety protection.

[0134] Optionally, the The adjustment method is as follows: when the production load characteristics indicate that the industrial control system is in a high-load state or a critical operating condition, the adjustment is increased. When the production load characteristics indicate that the industrial control system is in a low-load state, the load should be reduced. .

[0135] Optionally, the attack determination module 240 is used to determine the health score of each executor based on the real-time running status data; to obtain candidate results by weighted voting on the execution results based on the health score; and to verify the validity of the candidate results based on industrial process constraints or physical logic laws when the voting consistency is not satisfied, so as to determine the attack determination result.

[0136] Optionally, the attack determination module 240 is used to determine the voting weight of each executor based on the health score, wherein the voting weight is positively correlated with the health score; to weight the execution result according to the voting weight to obtain a consistency score, and to determine the candidate result with the highest consistency score and its vote weight percentage.

[0137] Optionally, the attack determination module 240 is configured to determine that the voting consistency is not satisfied if the weight of the vote is lower than a set percentage; obtain the process attribute parameters corresponding to the candidate results; determine that the attack determination result is not attacked if the process attribute parameters meet the industrial process constraints or physical logic laws; and determine that the attack determination result is attacked if the process attribute parameters do not meet the industrial process constraints or physical logic laws.

[0138] Optionally, the real-time operating status data includes: CPU load rate, memory utilization rate, and instruction execution latency, and the health score of each executor is determined by the following calculation formula:

[0139] Where H represents the health score, This represents the rate of change of CPU load rate compared to the standard load rate. This indicates the rate of change in memory usage compared to the standard memory usage. This represents the rate of change of instruction execution latency compared to the standard metric execution latency. , , This represents the weight coefficient corresponding to each item.

[0140] Optionally, the attack determination module 240 is further configured to remove the execution entity from the heterogeneous execution entity pool if the health score of the execution entity is lower than a set score, and the deep reinforcement learning model is configured to select a combination of execution entities from the heterogeneous execution entity pool; and to call execution entities with the same heterogeneous attributes as the execution entity from the standby execution entity cluster to add to the heterogeneous execution entity pool.

[0141] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0142] Please refer to Figure 7 , Figure 7 This application provides a schematic diagram of the structure of an electronic device for executing a security protection method for an industrial control system. The electronic device may include: at least one processor 310, such as a CPU; at least one communication interface 320; at least one memory 330; and at least one communication bus 340. The communication bus 340 is used to establish communication between these components. In this embodiment, the communication interface 320 is used for signaling or data communication with other node devices. The memory 330 may be a high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 330 may also be at least one storage device located remotely from the processor. The memory 330 stores computer-readable instructions; when these instructions are executed by the processor 310, the electronic device performs the aforementioned method process.

[0143] Understandable. Figure 7 The structure shown is for illustrative purposes only; the electronic device may also include components that are more advanced than those shown. Figure 7 The more or fewer components shown, or having the same Figure 7 The different configurations shown. Figure 7 The components shown can be implemented using hardware, software, or a combination thereof.

[0144] This application provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it performs the method process executed by the electronic device in the above method embodiments.

[0145] This embodiment discloses a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can perform the methods provided in the above-described method embodiments, such as including: In response to received industrial control commands, a current set of execution entities is determined. This current set of execution entities is selected from a heterogeneous execution entity cluster by a pre-trained deep reinforcement learning model based on the most recently collected production load characteristics and network attack situation characteristics. The heterogeneous execution entity cluster includes multiple execution entities. The industrial control instructions are distributed in parallel to each executor in the current executor group; Receive the execution results returned by each executor and the real-time running status data of each executor; The attack determination result is determined based on the execution result and the real-time running status data.

[0146] In summary, the embodiments of this application provide a security protection method, electronic device, storage medium, and program product for an industrial control system. This method dynamically perceives production load and attack trends through a deep reinforcement learning model, adaptively selecting the current combination of execution entities from a heterogeneous execution entity cluster, thus achieving a dynamic balance between protection strength and production continuity. Simultaneously, based on the fusion and adjudication of execution results and real-time operating status data, it can accurately identify output deviations caused by sensor noise or normal process fluctuations, reducing the false judgment rate and improving the inherent security protection capability of the industrial control system under complex operating conditions.

[0147] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0148] Furthermore, 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0149] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0150] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0151] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A security protection method for an industrial control system, characterized in that, The method includes: In response to received industrial control commands, a current set of execution entities is determined. This current set of execution entities is selected from a heterogeneous execution entity cluster by a pre-trained deep reinforcement learning model based on the most recently collected production load characteristics and network attack situation characteristics. The heterogeneous execution entity cluster includes multiple execution entities. The industrial control instructions are distributed in parallel to each executor in the current executor group; Receive the execution results returned by each executor and the real-time running status data of each executor; The attack determination result is determined based on the execution result and the real-time running status data.

2. The method according to claim 1, characterized in that, The deep reinforcement learning model is also used to determine the switching frequency based on the most recently collected production load characteristics and network attack situation characteristics. The switching frequency is used to determine the feature collection cycle. The determination of the current executor combination in response to received industrial control commands includes: In response to the received industrial control command, the current combination of execution entities determined by the deep reinforcement learning model in the most recent acquisition period is obtained.

3. The method according to claim 1, characterized in that, The deep reinforcement learning model is a DQN model, and its reward function during the training phase is: , This represents the weighting coefficient, adjusted based on production load characteristics. Indicator of production smoothness This indicates the strength index of safety protection.

4. The method according to claim 3, characterized in that, The The adjustment method is as follows: when the production load characteristics indicate that the industrial control system is in a high-load state or a critical operating condition, the adjustment is increased. When the production load characteristics indicate that the industrial control system is in a low-load state, the load should be reduced. .

5. The method according to claim 1, characterized in that, The step of determining the attack determination result based on the execution result and the real-time running status data includes: The health score of each executor is determined based on the real-time operating status data. Based on the health score, the execution results are weighted and voted to obtain candidate results; When the voting consensus is not met, the validity of the candidate results is verified based on industrial process constraints or physical logic laws to determine the attack judgment result.

6. The method according to claim 5, characterized in that, The step of obtaining candidate results by weighted voting on the execution results based on the health score includes: The voting weight of each executor is determined based on the health score, and the voting weight is positively correlated with the health score; The execution results are weighted according to the voting weights to obtain a consistency score, and the candidate result with the highest consistency score and its vote weight percentage are determined.

7. The method according to claim 6, characterized in that, When the voting consensus is not met, the candidate results are validated based on industrial process constraints or physical logic laws to determine the attack determination result, including: If the weight of the votes is lower than the set percentage, then the consistency of voting is not satisfied. Obtain the process attribute parameters corresponding to the candidate results; If the process attribute parameters meet the industrial process constraints or physical logic laws, then the attack determination result is determined to be that no attack has occurred. If the process attribute parameters do not conform to industrial process constraints or physical logic laws, the attack determination result is that the system has been attacked.

8. The method according to claim 5, characterized in that, The real-time operational status data includes: CPU load rate, memory utilization rate, and instruction execution latency. The health score of each executor is determined using the following formula: Where H represents the health score, This represents the rate of change of CPU load rate compared to the standard load rate. This indicates the rate of change in memory usage compared to the standard memory usage. This represents the rate of change of instruction execution latency compared to the standard metric execution latency. , , This represents the weight coefficient corresponding to each item.

9. The method according to claim 5, characterized in that, After determining the health score of each execution entity based on the real-time operating status data, the method further includes: If the health score of an executor is lower than a set score, the executor is removed from the heterogeneous executor pool. The deep reinforcement learning model is used to select a combination of executors from the heterogeneous executor pool. Call an executor with the same heterogeneous attributes as the executor from the standby executor cluster and add it to the heterogeneous executor pool.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer-readable instructions that, when executed by the processor, perform the method as described in any one of claims 1-9.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it performs the method as described in any one of claims 1-9.

12. A computer program product, characterized in that, It includes computer program instructions, which, when read and executed by a processor, perform the method as described in any one of claims 1-9.

Images