A secure interaction method and system

Abstract

This application discloses a secure interaction method and system, belonging to the field of industrial intelligent control technology. The method includes: responding to a data interaction request, providing acquired industrial field data to a second inference side via a first control side; based on the industrial field data, performing asynchronous inference via the second inference side to generate control suggestion values ​​and feeding them back to the first control side; according to preset process safety constraint rules, performing compliance judgment on the control suggestion values ​​via the first control side to obtain a judgment result; generating a secure interaction result based on the judgment result; when the judgment result is compliant, confirming the control suggestion value as valid interaction data; when the judgment results are non-compliant for a preset number of consecutive times, triggering an anomaly isolation mechanism, blocking the control suggestion values, outputting an isolation status, or triggering a degradation response. By performing compliance judgment and anomaly isolation on the AI ​​inference suggestion values, non-compliant commands are prevented from acting on the field, ensuring the safe and stable operation of the industrial control system.

Description

Technical Field

[0001] This application relates to the field of industrial intelligent control technology, and in particular to a safe interaction method and system. Background Technology

[0002] With the deep application of artificial intelligence in complex industrial control fields such as waste incineration, the current common intelligent control architecture is "DCS system + back-end server". This means the DCS is responsible for hard real-time logic control, while the back-end server runs AI algorithms to generate decision commands, which are then fed back to the DCS for execution. However, this architecture is affected by network transmission and server load fluctuations. Data interaction and command response cycles typically exceed 300ms, making it unable to cope with instantaneous fluctuations in operating conditions such as heterogeneous waste composition. This can easily lead to coking in the furnace or excessive pollutants. Furthermore, back-end server failures or network interruptions can directly paralyze intelligent control. Long-distance communication under harsh operating conditions is also prone to interference, resulting in high packet loss rates, and the system reliability is severely insufficient.

[0003] Currently, some solutions attempt to push AI inference down to the edge using "soft PLC + additional AI module", but there are still significant drawbacks: on the one hand, the millisecond-level load fluctuations of AI inference tasks compete for computing resources with microsecond-level hard real-time control tasks, causing spatiotemporal coupling and resulting in control task jitter or delay; on the other hand, AI models are prone to outputting erroneous instructions that deviate from the process limits under complex working conditions, and existing solutions lack mandatory multi-dimensional safety constraints and anomaly isolation mechanisms. If non-compliant abnormal instructions are directly adopted and executed, it is very easy to cause serious industrial accidents, making it difficult to guarantee the safety and stability of the control system. Summary of the Invention

[0004] This application provides a secure interaction method and system to at least solve the problems in the prior art, such as high data interaction latency and susceptibility to network interference, spatiotemporal coupling between AI inference and real-time control, and the lack of security constraints and anomaly isolation mechanisms for the uncertainty of AI inference, which can easily lead to system loss of control or industrial accidents.

[0005] In a first aspect, this application provides a secure interaction method, the method comprising: In response to a data interaction request, the acquired industrial field data is provided to the second inference side via the first control side; Based on the industrial field data, asynchronous inference is performed by the second inference side to generate control suggestion values ​​and feed them back to the first control side, thus completing the initial interactive response. Based on the preset process safety constraint rules, the first control side performs a compliance determination on the control suggestion value and obtains the determination result. Based on the judgment result, a safe interaction result is generated. When the judgment result is compliant, the control suggestion value is confirmed as valid interaction data. When the judgment result is non-compliant for a preset number of consecutive times, an abnormal isolation mechanism is triggered to block the control suggestion value and output an isolation status or trigger a degradation response as a safe interaction result.

[0006] The above technical solution generates control suggestion values ​​through asynchronous interaction between the first control side and the second inference side. The first control side then performs a mandatory compliance judgment on the suggestion values ​​according to preset process safety constraints. When non-compliance occurs, anomaly isolation and degradation responses are triggered, achieving spatiotemporal decoupling of control tasks and inference tasks, as well as safe interception of uncertain AI outputs. Its beneficial effects are: asynchronous inference avoids interference from AI computation time and load fluctuations on the real-time operation cycle of the first control side, while limiting AI output to suggestion values ​​that need to be verified. Abnormal inference results that do not comply with process safety constraints are intercepted through compliance judgment, and direct shielding and degradation are performed when non-compliance is judged, preventing abnormal instructions from being executed erroneously, and ensuring the control continuity and operational safety of the industrial control system when AI inference is abnormal or erroneous.

[0007] Secondly, this application provides a secure interaction system applied to a control system including a first control side and a second inference side, the system comprising: The data interaction module is used to respond to data interaction requests and provide the acquired industrial field data to the second inference side through the first control side; An asynchronous inference module, located on the second inference side, is used to perform asynchronous inference based on the industrial field data, generate control suggestion values ​​and feed them back to the first control side to complete the initial interactive response; The safety constraint module, located on the first control side, is used to determine the compliance of the control suggestion value according to the preset process safety constraint rules and obtain the determination result. The result execution module, located on the first control side, is used to generate a secure interaction result based on the judgment result. When the judgment result is compliant, the control suggestion value is confirmed as valid interaction data. When the judgment result is non-compliant for a preset number of consecutive times, an abnormal isolation mechanism is triggered to block the control suggestion value and output an isolation status or trigger a degradation response as a secure interaction result.

[0008] Thirdly, this application provides an electronic device including one or more processors and one or more memories, wherein at least one piece of program code is stored in the one or more memories, the program code being loaded and executed by the one or more processors to implement the operations performed by the secure interaction method.

[0009] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the operations performed by the secure interaction method.

[0010] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described secure interaction methods. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0012] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating a secure interaction method provided in this application embodiment. Figure 1 ; Figure 2 A flowchart illustrating a secure interaction method provided in this application embodiment. Figure 2 ; Figure 3 A flowchart illustrating a secure interaction method provided in this application embodiment. Figure 3 ; Figure 4 A flowchart illustrating a secure interaction method provided in this application embodiment. Figure 4 ; Figure 5 A flowchart illustrating a secure interaction method provided in this application embodiment. Figure 5 ; Figure 6 A flowchart illustrating a secure interaction method provided in this application embodiment. Figure 6 ; Figure 7 A flowchart illustrating a secure interaction method provided in this application embodiment. Figure 7 ; Figure 8 This is a schematic diagram of the structure of a secure interaction system provided in an embodiment of this application; Figure 9 A schematic diagram of system hardware architecture and memory partitioning based on an asymmetric multiprocessing architecture is provided for embodiments of this application; Figure 10This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 11 A timing diagram illustrating the asynchronous interaction between the real-time control domain and the AI ​​computing domain, provided for embodiments of this application; Figure 12 Flowchart of the logic verification and circuit breaker mechanism of the security constraint module provided in the embodiments of this application; Figure 13 This is a schematic diagram of the mixed programming interface of AI function blocks and ladder diagrams in the integrated development environment provided in the embodiments of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0015] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, the said secure interaction method, article, or device that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, the said secure interaction method, article, or device. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0016] In related technologies, existing basic control hardware such as PLCs, while possessing microsecond-level hard real-time performance, have limited computing power and cannot run complex AI algorithms. Industrial control computers, while having strong computing power, lack real-time performance, and the uncertainty of AI inference output poses potential security risks. In complex industrial scenarios such as waste incineration, the currently widely adopted "DCS + server" intelligent control architecture suffers from high data transmission latency, with interaction response cycles typically exceeding 300ms or even 500ms. It cannot cope with instantaneous fluctuations in operating conditions, easily leading to coking in the furnace or excessive pollutants. Furthermore, backend server failures or network interruptions can easily paralyze intelligent functions. In harsh environments, long-distance communication suffers from high packet loss rates (10%-15%), resulting in insufficient system reliability. In addition, it also faces problems such as high deployment and maintenance costs, long debugging cycles for protocol adaptation of equipment from different manufacturers, uneven utilization of server resources, and poor adaptability, including the inability to dynamically adjust algorithm parameters in real time according to on-site operating conditions.

[0017] Existing edge-side integration improvement solutions, such as adding edge gateways or using soft PLCs and additional AI modules, still rely on backend server decisions. They also face the risk of spatiotemporal coupling between AI inference load fluctuations and competition for computing resources between hard real-time control tasks, leading to control task jitter or delays. Furthermore, the lack of a mandatory safety constraint mechanism for the uncertainty of AI inference makes it easy for erroneous instructions to be executed directly, causing serious industrial accidents. In addition, the development environments for logic control and AI inference are disconnected, making system integration and debugging difficult.

[0018] To address the aforementioned technical challenges, this application presents a secure interaction scheme that features asynchronous decoupling of control and inference, along with mandatory compliance determination. The scheme involves a first control side providing industrial field data to a second inference side for asynchronous inference, avoiding interference from inference time and load fluctuations on control real-time performance. Simultaneously, the first control side performs mandatory compliance determination on the control suggestion values ​​generated by the second inference side based on preset process safety constraints. Only when compliance is determined is the suggestion value recognized as valid interactive data; otherwise, an anomaly isolation mechanism is triggered to block the suggestion value and output an isolation status or trigger a degradation response. This achieves spatiotemporal decoupling of control and inference tasks, as well as mandatory security interception and fallback for uncertain AI outputs, ensuring the real-time performance and operational safety of the industrial control system under complex operating conditions.

[0019] The application scenarios of the technical solutions provided in the embodiments of this application are described below.

[0020] The technical solutions provided in this application are mainly applied in the field of industrial automation control, which has extremely high requirements for real-time performance, security, and intelligent collaboration, especially in complex working conditions that require the integration of hard real-time logic control and artificial intelligence reasoning. Specifically, they mainly include the following application scenarios: 1. Intelligent control scenario for the entire waste incineration process In the field of waste incineration, due to the severe heterogeneity of waste composition, large fluctuations in calorific value, drastic fluctuations in operating conditions, and stringent environmental emission requirements, the traditional "DCS + server" architecture, which relies on backend servers, suffers from instruction latency exceeding 300ms, easily leading to furnace coking, dioxin surges, or pollutant exceedances. The technical solution proposed in this application can be widely applied to the following key stages of waste incineration: Waste feeding and grate speed control: The edge-end intelligent controller uses AI algorithms to identify changes in the calorific value of waste in real time, dynamically adjusts the feeding amount and grate speed, and ensures millisecond-level response through hard real-time control to avoid unstable combustion.

[0021] Furnace temperature control: For instantaneous fluctuations such as sudden rises / falls in furnace temperature, AI inference directly generates control suggestion values ​​at the edge side, and through safety constraint modules, such as threshold verification and change amplitude verification, ensures that the temperature adjustment command does not exceed the process safety envelope, preventing overheating damage or flameout.

[0022] Flue gas purification, such as deacidification and denitrification: real-time prediction of pollutant concentrations such as NOx and SO2, precise control of ammonia injection and other reagent dosage, and prevention of AI outputting erroneous instructions that violate the reaction mechanism through process mechanism verification, ensuring that emissions meet standards and do not waste reagents.

[0023] Leachate treatment: In harsh on-site environments with high temperature and high corrosion, control logic and AI inference are deployed directly at the edge to avoid packet loss and delay caused by long-distance communication, ensuring that the heavy metals in the leachate are discharged in compliance with standards.

[0024] 2. Smart Manufacturing and Automated Production Line Scenarios In intelligent manufacturing and automated production lines, the production process requires a high degree of flexibility and precise dynamic adjustment, while ensuring the safety of personnel and equipment. Industrial robot control: AI inference can dynamically adjust the robot's operation path and obstacle avoidance strategy based on environmental perception data such as vision. The heterogeneous architecture of this solution ensures that fluctuations in AI computing load will not interfere with the microsecond-level hard real-time response of the robot's underlying servo control. The safety constraint module can prevent AI from outputting commands that exceed the physical limits of the robotic arm or cause acceleration impacts, thus ensuring equipment safety.

[0025] Production line status monitoring and parameter adjustment: The AI ​​model monitors the production line status in real time and predicts the probability of failure, dynamically adjusting production parameters. Through a cross-domain shared memory mechanism, the control domain and the computing domain collaborate efficiently. When the AI ​​model outputs an incorrect speed adjustment command due to sensor anomalies, compliance judgment can immediately intercept and downgrade to safe shutdown logic.

[0026] 3. Intelligent Traffic Control Scenarios In intelligent transportation systems, roadside edge control units need to simultaneously process massive amounts of data and execute real-time signal control. Traffic light timing optimization: AI inference processes traffic flow data in real time to optimize traffic light timing to improve traffic efficiency. This solution decouples the traffic flow inference task from the hard real-time control task of traffic lights. Even if the AI ​​inference is delayed due to the large amount of data, the basic control cycle of the traffic lights can still operate independently and stably. At the same time, the timestamp alignment mechanism can discard outdated AI timing suggestions to avoid causing traffic chaos.

[0027] 4. Control of other complex industrial processes with high reliability requirements This includes, but is not limited to, chemical reaction process control, metallurgical blast furnace control, and new energy battery electrode coating control. These scenarios typically exhibit strong coupling of multiple variables and nonlinear dynamic characteristics, and control errors often lead to serious safety accidents. The edge-side hard real-time control, AI inference, and mandatory safety verification architecture provided in this application can effectively mitigate the impact of AI uncertainty on the underlying industrial processes through hardware-level resource isolation and safety envelope verification, while simultaneously introducing AI to optimize processes.

[0028] After introducing the implementation environment and application scenarios of the embodiments of this application, the technical solutions provided by the embodiments of this application are described below. (See also...) Figure 1 Taking the waste incineration edge intelligent controller as the executing entity as an example, the safe interaction method includes the following steps.

[0029] Step S101: In response to the data interaction request, the acquired industrial field data is provided to the second inference side through the first control side.

[0030] A data interaction request is used to request cross-domain data transmission between the first control side and the second inference side to initiate the AI ​​inference process. The data interaction request can be a control signal or event within the system used to trigger cross-domain data transmission. For example, the aforementioned data interaction request may include, but is not limited to: a timed trigger signal generated when the control cycle of the first control side arrives; or a data ready notification generated by the first control side after completing on-site data acquisition and preprocessing; or a data pull instruction issued by the second inference side to the shared data area when computing power is idle.

[0031] The sender of the data interaction request can be the first control side or a system timer. The receiver of the aforementioned data interaction request can be the first control side.

[0032] Understandably, upon receiving the data interaction request, the first control side can respond to the request by writing industrial field data into the shared data area, thereby enabling the second inference side to obtain the required data from the shared data area.

[0033] The first control side can be the real-time control domain in the system, and the second inference side can be the AI ​​computing domain in the system.

[0034] For example, the first control side can be a hardware core running a real-time operating system (RTOS) and responsible for performing hard real-time logic control tasks; the second inference side can be a hardware core running a general-purpose operating system and integrating an AI acceleration processor and responsible for performing AI learning and inference tasks.

[0035] Industrial field data may include, but is not limited to, multi-source sensor data such as furnace temperature, flue gas NOx / SO2 concentration, waste feed rate, and grate speed during the waste incineration process.

[0036] Specifically, within a preset control cycle, the first control side is responsible for collecting sensor data from the industrial site as industrial site data. After acquiring this industrial site data, the first control side performs preliminary preprocessing, such as removing obviously invalid and abnormal data, and provides the multi-source site data required for AI inference to the second inference side through methods such as cross-domain shared memory. This data provision process adopts a splitting and asynchronous mechanism to ensure that data interaction does not affect the independent execution and response of the first control side's own hard real-time logic control tasks.

[0037] Step S102: Based on industrial field data, asynchronous inference is performed through the second inference side to generate control suggestion values ​​and feed them back to the first control side, completing the initial interactive response.

[0038] Asynchronous inference refers to the process by which the second inference side performs inference calculations independently of the first control side's runtime cycle. In other words, the second inference side can perform AI model inference calculations on a time scale different from the first control side's hard real-time control cycle, based on its own computing power and scheduling.

[0039] The control recommendation value can refer to the preliminary suggested instructions for adjusting industrial process parameters, calculated by the second inference side using an AI model. It is understood that this control recommendation value is not the final executable control instruction, but rather data to be verified by the first control side for safety checks.

[0040] The initial interactive response can refer to the process of initial data and inference result transmission and reception between the first control side and the second inference side, the result of which is a control suggestion value to be verified.

[0041] For example, the asynchronous inference described above may include the process of the second inference side reading industrial field data from the shared data area and running AI algorithms such as combustion optimization models and pollutant emission prediction models to generate control recommendation values ​​such as grate speed adjustment values ​​and ammonia injection quantity recommendation values.

[0042] Understandably, due to the use of asynchronous inference, the AI ​​computation process on the second inference side will not consume the computing resources of the first control side, thereby avoiding the interference of AI inference load fluctuations on hard real-time control tasks.

[0043] Specifically, preprocessed industrial field data provided by the first control side is read from shared memory via the second inference side. Asynchronous inference calculations are then performed using locally deployed AI models, such as combustion optimization models and pollutant emission prediction models. Because asynchronous inference is used, the AI ​​calculation process on the second inference side does not consume computing resources on the first control side, thus avoiding interference from AI inference load fluctuations on the hard real-time control task. After inference is complete, preliminary control suggestion values ​​are generated by the second inference side, and these preliminary control key values ​​are fed back to the first control side via shared memory, completing the initial data interaction response.

[0044] Step S103: Based on the preset process safety constraint rules, the first control side performs a compliance judgment on the control recommendation value and obtains the judgment result.

[0045] Pre-defined process safety constraints can refer to a set of rules used to define the safe and feasible domain, which are formed by combining industry standards such as environmental emission limits, equipment safety parameters such as rated operating range, and process mechanism logic.

[0046] Compliance determination refers to the process by which the first control side compares and verifies the recommended control value with the safe and feasible domain to determine whether the recommended value is within a safe range and whether it may cause risks. The safe and feasible domain can be defined by pre-set process safety constraints, such as industry standards, equipment safety parameters, and process mechanism logic, allowing the control parameters output by AI inference to operate safely.

[0047] The determination result can be either compliant or non-compliant. Compliance indicates that the control recommendation value is within the safe and feasible domain and can be used as valid interaction data; non-compliance indicates that the control recommendation value exceeds the safety boundary or violates the process logic and needs to be intercepted or processed.

[0048] For example, compliance determination can include multi-dimensional verification, such as: determining whether the parameters in the recommended values ​​exceed the safety threshold range, whether the change range of the parameters is too large, and whether the recommended values ​​conform to the process mechanism, etc.

[0049] Specifically, after receiving the control suggestion value from the second inference side, the first control side inputs the control suggestion value into the built-in safety constraint module for verification. This compliance determination is based on preset process safety constraint rules. These rules combine industry standards for the industrial site, such as environmental emission limits, equipment safety parameters such as rated operating range, and process mechanism logic, forming a clearly defined safe and feasible domain. The first control side performs multi-dimensional verification, such as determining whether the parameters in the control suggestion value exceed safety thresholds, whether the parameter changes are excessive, and whether the suggestion value conforms to the process mechanism, to determine whether the control suggestion value is within the safe and feasible domain, thereby obtaining a compliance or non-compliance determination.

[0050] Step S104: Based on the judgment result, generate a secure interaction result.

[0051] Specifically, when the judgment result is compliant, the control suggestion value is confirmed as valid interactive data; when the judgment result is non-compliant for a preset number of consecutive times, the abnormal isolation mechanism is triggered, the control suggestion value is blocked, and the isolation status or a downgrade response is output as a safe interactive result.

[0052] Specifically, when the judgment result is compliant, it indicates that the control suggestion value is safe and feasible. The first control side confirms the control suggestion value as valid interactive data, which can then be converted into standardized control commands and sent to the actuator for execution, realizing intelligent optimization control. When the judgment result is non-compliant, it indicates that the control suggestion value may have safety risks due to abnormal input or algorithm uncertainty. At this time, the abnormal isolation mechanism is immediately triggered to block the control suggestion value, cut off its output channel to the actuator, and output an isolation status or trigger a degradation response, such as seamlessly switching to traditional deterministic logic algorithms for control or triggering a safe shutdown as a safe interaction result. This effectively avoids the impact of AI inference uncertainty on the underlying industrial control and ensures the safe and stable operation of the system.

[0053] To prevent the system from frequently entering a degraded state due to occasional data fluctuations, an anomaly isolation mechanism is triggered when the judgment results are non-compliant for a preset number of consecutive times. Specifically, when the judgment result is non-compliant, the first control side does not immediately trigger anomaly isolation, but instead starts an anomaly observation period and accumulates the number of consecutive non-compliant judgment results. During this observation period, the first control side uses the valid interaction data from the previous control cycle to maintain stable operation. When the cumulative number of consecutive non-compliant results reaches a preset number, such as 3 consecutive times, the first control side confirms that the second inference side has fallen into a continuous anomaly. At this time, the anomaly isolation mechanism is officially triggered, the control suggestion value is blocked and the isolation status is output or a degraded response is triggered.

[0054] This embodiment generates control suggestion values ​​through asynchronous interaction between the first control side and the second inference side. The first control side then performs a mandatory compliance judgment on the suggestion values ​​according to preset process safety constraints. If non-compliance is detected, anomaly isolation and degradation responses are triggered. This achieves asynchronous decoupling of control tasks and inference tasks, as well as safe interception of uncertain AI outputs. Its beneficial effects are: asynchronous inference avoids interference from AI computation time and load fluctuations on the real-time operation cycle of the first control side, while limiting AI output to suggestion values ​​that need to be verified. Abnormal inference results that do not comply with process safety constraints are intercepted through compliance judgment, and direct shielding and degradation are performed when non-compliance is detected, preventing abnormal instructions from being executed erroneously and ensuring the control continuity and operational safety of the industrial control system when AI optimization is introduced.

[0055] It should be noted that the above steps S101-S104 are a simplified description of the embodiments provided in this application.

[0056] To more clearly illustrate the secure interaction method provided by the above embodiments of this application, the following detailed description of this application is provided in conjunction with the accompanying drawings and specific embodiments.

[0057] The methods provided in the embodiments of this application will be described in more detail below with reference to some examples. See also Figure 2 Based on the industrial field data, the second inference side performs asynchronous inference to generate control suggestion values ​​and feeds them back to the first control side, completing the initial interactive response. This process specifically includes the following steps: Step S201: Write the industrial field data and the corresponding collection timestamp into the shared data area through the first control side.

[0058] The shared data area can refer to a memory region accessible to both the first control side and the second inference side, used to enable cross-domain data exchange. The shared data area can be implemented using a lock-free circular buffer, which is used to avoid mutex lock contention between the first control side and the second inference side during concurrent read / write operations.

[0059] Specifically, the first control side operates continuously with a hard real-time control cycle at the microsecond level. For example, in the real-time control domain, the control cycle is ≤1ms. When collecting sensor data from the waste incineration site, such as furnace temperature and flue gas concentration, a unique acquisition timestamp is added to each frame of data. After preliminary preprocessing, this industrial site data is written to the shared data area. The shared data area is implemented through a lock-free circular buffer. Since the write operation of the first control side and the read operation of the second inference side are executed concurrently, the lock-free circular buffer, through a hardware-level memory mapping mechanism, avoids the deadlock risk and performance bottleneck caused by the traditional mutex lock mechanism during concurrent read and write operations between the first control side and the second inference side, ensuring the real-time performance and zero jitter of the data writing on the first control side.

[0060] Step S202: Based on the shared data area, asynchronous inference is performed by asynchronously reading data from the second inference side to generate preliminary control recommendation values.

[0061] Asynchronous reading can refer to the second inference side actively pulling data from the shared data area according to its own computing cycle, without synchronizing with the write action of the first control side.

[0062] Asynchronous inference refers to the process by which the second inference side performs inference calculations independently of the first control side's runtime cycle. In other words, the second inference side can perform AI model inference calculations on a time scale different from the first control side's hard real-time control cycle, based on its own computing power and scheduling.

[0063] The preliminary control recommendation value can refer to the preliminary recommended instruction for adjusting industrial process parameters, calculated by the second inference side using an AI model. It is understood that this preliminary control recommendation value is not the final executable control instruction, but rather data to be verified by the first control side for safety checks.

[0064] Specifically, the second inference side operates independently of the first control side on a microsecond-level cycle. This could be a periodically running AI computing domain, which, based on its own computing power, asynchronously reads the required field data from the shared data area on a millisecond-level cycle, such as 10ms, and runs complex AI models, such as combustion optimization models, for inference calculations. Because the read and write operations are decoupled in time, the first control side does not need to wait for the inference results from the second inference side to continue executing the control logic for the next cycle, thus avoiding interference from load fluctuations or computational time consumption of the AI ​​inference task on the hard real-time control cycle.

[0065] Step S203: Associate the initial control recommendation value with the corresponding timestamp to obtain the initial control recommendation value with timestamp, and feed the initial control recommendation value with timestamp back to the shared data area.

[0066] The associated timestamp can refer to binding the inference result with the timestamp of the original input data on which it is based, so as to ensure that the control command and the field industrial data that triggered the command are consistent in logical timing.

[0067] Specifically, after the second inference side completes the inference to generate preliminary control suggestion values, such as grate speed adjustment values ​​and ammonia injection quantity suggestion values, the generated preliminary control suggestion values ​​are bound and associated with the acquisition timestamps of the original input data on which the inference is based, and written into the output buffer of the shared data area. Through the synchronous association of timestamps, the consistency of the control command and the field industrial data that triggers the command in logical timing is ensured, providing a basis for subsequent timeliness verification.

[0068] Step S204: Use the first control side to periodically poll the shared data area, match and verify the timestamp carried by the preliminary control suggestion value with the current control cycle, and determine the timestamp difference.

[0069] Periodic polling can refer to the operation of the first control side actively checking whether there is any new data written to the shared data area within each control cycle of the first control side.

[0070] Matching verification can refer to the process of comparing the time stamp carried by the preliminary control recommendation value with the current time to confirm the timeliness of the recommendation value.

[0071] The timestamp difference can be used to represent the time difference between the operating condition time on which the initial control recommendation value was based and the execution time of the current first control side.

[0072] Specifically, see Figure 11 The asynchronous interaction timing diagram shown illustrates that within each hard real-time control cycle, the first control side actively polls the output buffer of the shared data area to check whether the second inference side has written new inference results. If a new preliminary control suggestion value is detected, the first control side extracts the acquisition timestamp carried by the new preliminary control suggestion value and compares it with the current control cycle time to calculate the timestamp difference, thereby determining whether the AI ​​suggestion value corresponds to the current real-time operating condition.

[0073] Step S205: When the timestamp difference does not exceed the preset timeliness threshold and the inference result is ready, the preliminary control recommendation value is determined as the control recommendation value.

[0074] The preset time threshold can refer to the maximum time difference that allows control commands to lag behind actual operating conditions. If the time threshold is exceeded, the command is considered to have expired.

[0075] The inference result is ready, which means that the second inference side has successfully written the preliminary control recommendation value into the shared data area and the data status is marked as readable.

[0076] Specifically, complex operating conditions such as waste incineration have extremely high requirements for control response latency, typically ≤100ms. If the preliminary control suggestion value is ready and its timestamp difference does not exceed the preset timeliness threshold, such as ≤50ms, it indicates that the AI ​​inference result can still reflect the current instantaneous operating condition changes, and the data timeliness meets the standard. At this time, the first control side confirms that the preliminary control suggestion value is the time-aligned data to be verified, reads it as the control suggestion value, and prepares to enter the subsequent safety verification process.

[0077] Step S206: When the timestamp difference exceeds the preset timeliness threshold or the inference result is not ready, the preliminary control suggestion value is determined to be invalid, and the interaction data valid in the previous period is used as the control suggestion value.

[0078] "Inference results not ready" can refer to a situation where, within the current control cycle, the second inference side has not yet completed the inference calculation, or although it has completed it, it has not yet written the preliminary control recommendation value into the shared data area.

[0079] Valid interaction data in the previous cycle can refer to control data that has been verified as safe and executable by the first control side during the previous control cycle.

[0080] Specifically, if excessive computing power load causes the second inference side to take too long to infer, resulting in a timestamp difference exceeding a preset timeliness threshold, or if the inference result has not yet been written in the current cycle (i.e., not ready), it indicates that the AI ​​output is lagging behind drastic changes in actual operating conditions, such as a sudden rise in furnace temperature. Forcing execution could lead to serious accidents such as coking in the furnace. Therefore, the first control side determines that the incomplete or expired preliminary control suggestion value is invalid and discards it. The valid interaction data that has been verified and successfully executed in the previous control cycle is automatically used as the control suggestion value to be verified this week. This ensures the continuity of waste incineration control commands, avoids control interruptions or drastic fluctuations in operating conditions due to AI inference delays, and thus completes the initial interaction response.

[0081] This embodiment achieves concurrent read / write operations between the first control side and the second inference side by employing a lock-free circular buffer, and by attaching timestamps to the interactive data and performing periodic polling and timeliness alignment checks. This realizes asynchronous spatiotemporal decoupling of control and inference tasks, as well as timeliness filtering of data. Its beneficial effects are as follows: on the one hand, it avoids the deadlock risk and performance bottleneck caused by mutex lock contention, and protects the hard real-time control task from interference from AI inference time consumption and load fluctuations, ensuring the determinism and absolute real-time performance of the control cycle. On the other hand, it promptly eliminates lagging AI inference results through timeliness threshold judgment and automatically uses the valid data from the previous cycle as the control suggestion value to be verified, effectively preventing expired instructions from erroneously intervening in transient operating conditions, ensuring the continuity and safety stability of the industrial control process, thereby completing the initial interactive response.

[0082] In some embodiments, see Figure 3 The preset process safety constraints include threshold constraints and variation range constraints; Based on preset process safety constraints, the first control side performs a compliance determination on the control recommendation value to obtain a determination result, including: Step S301: Based on the threshold constraints in the preset process safety constraint rules, compare the control parameters in the control recommendation values ​​with the preset safety threshold range to obtain the first comparison result.

[0083] Specifically, after the first control side, such as the safety constraint module in the real-time control domain, receives the control suggestion value generated by the second inference side, it first extracts various control parameters, such as the furnace temperature setpoint, grate speed, and ammonia injection rate during the waste incineration process. These parameters are then compared in real time with the safety threshold ranges pre-set based on industrial process requirements and equipment safety parameters. For example, the furnace temperature safety threshold is set to 850℃-1050℃ and the grate speed is set to 0.5-2.0m / min. This is to determine whether the absolute parameter values ​​output by the AI ​​exceed the limits, thereby obtaining the first comparison result.

[0084] Step S302: When the control parameters exceed the safety threshold range, the first comparison result triggers a non-compliance condition.

[0085] Specifically, if any control parameter exceeds the corresponding safety threshold range, such as the furnace temperature setting value being higher than 1050℃ or lower than 850℃, it indicates that the AI ​​suggested instruction has exceeded the physical or process safety boundary, which may lead to serious consequences such as equipment damage or excessive pollutant emissions. At this time, the first comparison result immediately triggers non-compliance conditions and intercepts the over-limit instruction.

[0086] Step S303: Obtain the parameter difference based on the variation range constraint in the preset process safety constraint rules.

[0087] The parameter difference is used to represent the magnitude of change between the control recommendation value and the effective interaction data of the previous control cycle. Specifically, industrial field conditions, such as waste incineration, are sensitive to sudden changes in control parameters. Even if the absolute value of the AI ​​output is within the threshold, excessively large adjustment steps can cause drastic system fluctuations. Therefore, the parameter difference between the control recommendation value of the current cycle and the effective interaction data that has been verified and executed in the previous control cycle is calculated. This parameter difference reflects the magnitude and rate of the control adjustment suggested by the AI.

[0088] Step S304: Compare the parameter difference with the preset maximum allowable variation range to obtain the second comparison result.

[0089] Specifically, the calculated parameter difference is compared with the maximum allowable change range preset based on the dynamic characteristics of the operating conditions, such as a single change in furnace temperature not exceeding 50°C and a single change in ammonia injection not exceeding 10%, in order to assess whether the adjustment action suggested by the AI ​​is too aggressive, thus obtaining the second comparison result.

[0090] Step S305: When the parameter difference exceeds the maximum allowable variation range, the second comparison result triggers a non-compliance condition.

[0091] Specifically, if the parameter difference exceeds the maximum allowable range of change, it means that if the control recommendation is executed, it will cause a sudden change in the parameters, which is very likely to cause drastic fluctuations in the system operating conditions, such as causing dangerous situations like furnace shutdown or a sudden increase in flue gas concentration. In this case, the second comparison result triggers the non-compliance condition to prevent excessive adjustment from disrupting the stability of the control process.

[0092] Step S306: If neither the first comparison result nor the second comparison result triggers the non-compliance condition, the obtained judgment result is compliant.

[0093] Specifically, the first comparison result is compliant only if the control recommendation value does not exceed the absolute safety boundary of the process and equipment, and does not cause a dangerous risk of sudden change in operating conditions. In other words, when the instruction output by the AI ​​is within the safe and feasible domain in both the absolute value and the rate of change, the first control side determines that the control recommendation value is compliant and allows it to enter the subsequent execution process as safe interactive data.

[0094] This embodiment achieves dual logical verification of AI inference output in both absolute value and rate of change by using threshold constraints that compare the control parameters of the control recommendation value with the preset safety threshold range, and by using variation constraints that calculate the parameter difference with the effective data of the previous cycle and compare it with the maximum allowable variation range. Its beneficial effect is that it can not only effectively intercept out-of-limit commands that exceed the physical boundaries of the process or equipment to avoid equipment damage or environmental violations, but also prevent aggressive commands with excessive adjustment steps from causing furnace shutdown or sudden parameter changes and other operating condition fluctuations. It limits the uncertainty of AI inference to a safe and feasible domain, thereby effectively ensuring the stability and operational safety of the industrial control system while introducing AI optimization.

[0095] In some embodiments, see Figure 4 The preset process safety constraint rules include process mechanism constraints and redundancy comparison constraints; The step of determining the compliance of the control suggestion value based on preset process safety constraints, through the first control side, and obtaining the determination result includes: Step S401: Based on the process mechanism constraints in the preset process safety constraint rules, substitute the control recommendation value into the preset process mechanism algorithm to obtain the theoretically reasonable value.

[0096] Specifically, the first control side embeds the process mechanism algorithm of the industrial site, such as the combustion dynamics model or flue gas purification reaction mechanism model in the waste incineration process. When the first control side receives the control suggestion value output by the second inference side, it substitutes the current operating parameters into the solidified mechanism algorithm to calculate the theoretically reasonable grate speed or ammonia injection rate under the current waste feed rate and furnace temperature, and uses this as the theoretically reasonable value.

[0097] Step S402: Compare the deviation between the control recommendation value and the theoretical reasonable value to obtain the third comparison result.

[0098] Specifically, the difference between the control recommendation value output by the AI ​​algorithm and the theoretically reasonable value calculated based on objective physical and chemical laws is calculated to clarify the degree to which the AI ​​reasoning result deviates from the logic of the process mechanism, thereby assessing whether the recommendation value violates the actual production law.

[0099] Step S403: When the deviation exceeds the preset allowable deviation range, the third comparison result triggers non-compliance conditions.

[0100] Specifically, if the deviation between the recommended control value and the theoretically reasonable value exceeds the normal fluctuation range allowed by the process, it indicates that the instructions output by the AI ​​do not conform to the actual process logic. For example, the correspondence between the ammonia injection amount and the nitrogen oxide concentration violates the reaction mechanism. In this case, non-compliant conditions are triggered, and erroneous operation instructions that do not conform to the process rules are intercepted from the root.

[0101] Step S404: Obtain two sets of control recommendation values ​​based on the redundancy comparison constraints in the preset process safety constraint rules.

[0102] The two sets of control suggestion values ​​represent the inference results generated by the primary inference algorithm and the backup inference algorithm, respectively. Specifically, to further prevent a single AI algorithm from producing erroneous outputs due to input disturbances or model degradation, a dual-mode AI inference algorithm (primary and backup) is deployed on the second inference side. When acquiring control suggestion values, the first control side simultaneously retrieves the inference results independently generated by the primary AI algorithm and the backup AI algorithm, respectively, as the two sets of control suggestion values ​​to be verified.

[0103] Step S405: Compare the deviations between the two sets of control recommendation values ​​to obtain the fourth comparison result.

[0104] Specifically, the safety constraint module on the first control side calculates the parameter differences between the two sets of control suggestion values ​​output by the main inference algorithm and the backup inference algorithm, thereby assessing the consistency of different models' judgments on the same working condition and determining whether the AI ​​inference process is stable and reliable.

[0105] Step S406: When the deviation exceeds the preset consistency threshold, the fourth comparison result triggers non-compliance conditions.

[0106] Specifically, if the deviation between the instructions output by the two sets of algorithms exceeds the preset consistency judgment range, it indicates that the current AI inference result is highly uncertain and cannot reach a consensus on its own. At this time, non-compliance conditions are triggered, and the disputed instruction is refused to be executed.

[0107] Step S407: If neither the third comparison result nor the fourth comparison result triggers the non-compliance condition, the obtained judgment result is compliant.

[0108] Specifically, only when the control recommendation value conforms to the physicochemical laws of the objective process mechanism and the results of dual AI inference are consistent and do not have high uncertainty, will the first control side determine that the recommendation value is logically correct and the inference is reliable, judge it as compliant, and allow it to enter the subsequent execution stage as the final safe interaction data.

[0109] This embodiment achieves verification of AI inference output in two dimensions: the consistency of objective process logic and the stability of the algorithm itself. This is achieved by using process mechanism constraints—substituting control suggestion values ​​into preset process mechanism algorithms to calculate theoretically reasonable values ​​and comparing deviations—and by using redundancy comparison constraints—obtaining control suggestion values ​​generated by two sets of inference algorithms (primary and backup) and comparing deviations. Its beneficial effects are that it can intercept erroneous instructions that violate actual physical and chemical laws, ensuring that control actions meet real process requirements, and effectively identify and prevent output anomalies caused by excessive uncertainty of a single AI model. Through dual verification of objective mechanism and algorithm redundancy, the influence space of AI black box characteristics is further compressed, thereby providing a deeper level of security for industrial control systems under complex working conditions.

[0110] In some embodiments, see Figure 5 The preset process safety constraint rules include input data validity constraints; The step of determining the compliance of the control suggestion value based on preset process safety constraints, through the first control side, and obtaining the determination result includes: Step S501: Obtain the original input data corresponding to the control suggestion value according to the input data validity constraints in the preset process safety constraint rules.

[0111] Specifically, through the first control side, such as the safety constraint module within the real-time control domain, after receiving the control suggestion value output from the second inference side, the conventional logic of only performing surface-level verification of the AI ​​output result is broken. Instead, the safety defense line is moved forward to the data source. Based on the acquisition timestamp or data association identifier carried by the control suggestion value, the data is traced back to the shared data area to extract the original input data that triggered the AI ​​inference, such as the flue gas analyzer readings, furnace thermocouple temperatures, and material layer thickness sensor data at the waste incineration site. This ensures that the correspondence between the source and the result is traceable, providing a basis for subsequent data source verification.

[0112] Step S502: Perform validity verification on the original input data and obtain the verification result.

[0113] Specifically, the harsh industrial environment makes sensors prone to malfunction or measurement errors. Allowing abnormal data to be input into the AI ​​model will inevitably lead to "garbage in, garbage out" (GIGO). Therefore, the first control side performs a comprehensive multi-dimensional verification of the original input data according to preset input data validity constraints. This includes, but is not limited to, checking whether the data exceeds the range (e.g., a thermocouple open circuit causing a full-scale temperature signal deviation, exceeding the sensor's physical measurement limits), whether there is a dead value stagnation (e.g., a sensor jamming causing data to remain completely stagnant for multiple consecutive control cycles), whether there are signal abrupt changes (e.g., electromagnetic interference causing non-physical spikes in the data), and whether the underlying communication messages are marked as packet loss or verification errors. This comprehensive assessment determines whether the raw materials input into the AI ​​model truly and reliably reflect the actual working conditions on site.

[0114] Step S503: When the verification result is abnormal, the verification result triggers non-compliance conditions, and the resulting judgment result is non-compliant.

[0115] Specifically, if the verification result of the original input data exhibits any of the aforementioned abnormal characteristics, it indicates that the preconditions upon which the AI ​​inference relies have failed. For example, if the oxygen sensor data is stuck, the AI ​​model may misjudge the combustion state and incorrectly adjust the air supply volume drastically, leading to overheating or flameout in the furnace. In this case, even if the absolute value and range of change of the control recommendation value itself seem reasonable, it is essentially an erroneous inference based on false preconditions. Therefore, the verification result triggers non-compliant conditions, and the first control side determines that the control recommendation value is non-compliant and intercepts it, cutting off the transmission path of abnormal input to the underlying actuators and preventing misleading instructions generated based on distorted data from causing miscontrol and safety accidents.

[0116] Step S504: If the verification result does not trigger non-compliance conditions, the judgment result obtained is compliant.

[0117] Specifically, only when all the original input data passes validity verification, confirming that the operating condition data collected by the field sensors is authentic, complete, and in a healthy working state, does the inference result derived by the AI ​​model based on this reliable data have physical meaning and reference value. At this point, the input data does not trigger non-compliant conditions, and the first control side determines that the control recommendation value is compliant, allowing it to proceed to further verification in dimensions such as thresholds, rates of change, or mechanisms, or to enter the execution stage as safe interactive data, thereby constructing a fundamental defense line to ensure control security from the data source.

[0118] This embodiment achieves source verification of AI inference results from the output end to the input end by acquiring the original input data corresponding to the control suggestion value and verifying its validity according to the validity constraints of the input data. Its beneficial effect is that it can identify and intercept erroneous control commands derived from invalid inputs caused by sensor failures or abnormal acquisition from the data source, avoid the AI ​​model from generating misleading outputs based on distorted data and causing miscontrol, thereby effectively blocking the transmission of abnormal inputs to the underlying actuators, and further improving the robustness and operational safety of the industrial control system under complex and harsh working conditions.

[0119] In some embodiments, see Figure 6 The triggering of the degradation response includes: Step S601: Stop the inference task on the second inference side and disconnect the output channel of the control suggestion value.

[0120] Specifically, when the compliance assessment result of the first control side is serious non-compliance, or when an untrusted state such as model collapse or continuous abnormal output is detected in the AI ​​control domain of the second inference side, a circuit breaker mechanism is executed. At this time, the first control side sends a forced suspension or termination signal to the second inference side to stop the AI ​​inference computing tasks within the second inference side and release computing resources. At the same time, the output channel of the control suggestion value from the second inference side to the first control side is logically blocked or physically cut off at the data interaction layer, such as stopping the writing of data to the unlocked circular buffer or suspending the read operation of the first control side. This prevents any unverified or potentially dangerous AI instructions from continuing to penetrate into the underlying execution mechanism, thereby achieving rapid physical isolation of the AI ​​system.

[0121] Step S602: Generate safety control instructions based on industrial field data.

[0122] Specifically, after cutting off the AI ​​system's intervention capabilities, to prevent the industrial site from spiraling out of control, the first control side, namely the real-time control domain, seamlessly takes over the system's leadership. At this point, the first control side acquires raw industrial site data—unprocessed by the AI ​​model, such as real-time temperature, pressure, and flow rate collected by sensors—and reverts to a preset safety fallback control strategy. This preset safety fallback control strategy, based on built-in conventional automation logic, such as classic PID control algorithms or pre-set safety hold / degradation process curves, calculates safety control commands—for example, maintaining the current safe output or slowly decreasing it to the safe operating range at a fixed slope. This ensures that even if the AI ​​intelligent function fails, the industrial site can still maintain basic safe and stable operation or safely shut down.

[0123] This embodiment achieves physical and logical isolation between the AI ​​system and the real-time control system, as well as seamless demotion and takeover of control, by stopping the AI ​​inference task and cutting off its output channel, while independently generating safety control commands based on industrial field data. Its beneficial effect is that when the AI ​​system experiences serious anomalies or continuously outputs untrusted commands, it can block the execution path of erroneous commands to avoid catastrophic consequences, and ensure that the industrial field can still retreat to the basic automation control mode after losing AI assistance, maintaining the safety baseline and stable operation of the production process, thereby providing a reliable backup and emergency self-healing mechanism for the entire intelligent control system.

[0124] In some embodiments, see Figure 7 After triggering a degradation response, the security interaction methods also include: Step S701: When the preset abnormal recovery conditions are met, asynchronous inference is performed through the second inference side to generate a new control suggestion value.

[0125] Specifically, during degraded operation—meaning that while the first control side independently executes safety control commands—AI-optimized control is not abandoned, but the operational status is continuously monitored. When preset anomaly recovery conditions are met, such as: input data validity verification continuously returning to normal, fault codes that triggered AI inference anomalies being eliminated, or degraded safe operation reaching the set observation period, the second inference side is reactivated. In the background, based on real-time industrial field data, the asynchronous inference task is restarted via the second inference side to generate new control recommendation values. This process is asynchronous and does not interfere with the safety fallback control currently being executed by the first control side, ensuring stable operating conditions during the recovery period.

[0126] Step S702: Based on the preset process safety constraint rules, determine the compliance of the new control recommendation value and obtain the determination result.

[0127] Specifically, the control recommendations newly generated by the second inference side during the recovery period cannot directly take over control; they need to undergo a security review by the first control side again. The first control side invokes preset process safety constraints that are completely consistent with normal operation, including threshold constraints, variation range constraints, process mechanism constraints, redundancy comparison constraints, and input data validity constraints, to conduct a comprehensive compliance assessment of the new control recommendations. This verifies whether the AI ​​model's inference logic has escaped its previous abnormal state and whether its output has regained physical rationality and engineering safety.

[0128] Step S703: When the judgment results of a preset number of consecutive times are all compliant, switch back to the first control side and replace the currently executed safety control command with a new control suggestion value for interactive response.

[0129] Specifically, to prevent the AI ​​system from blindly regaining control based on occasional, single compliant outputs, thus causing frequent switching between basic automated control and AI control—that is, to prevent control oscillations or the ping-pong effect—this step sets a confidence recovery threshold. Only when the new control recommendation value consistently passes all compliance checks for multiple consecutive control cycles (i.e., a preset number of consecutive cycles, such as 3 or 5), fully demonstrating that the AI ​​inference function has returned to normal and its output is consistently reliable, will the control rollback mechanism be triggered. At this point, the first control side switches back to AI interaction mode, smoothly replaces the currently executed safety control command with the verified new control recommendation value, and sends it to the actuator, thereby safely and seamlessly restoring the system to the AI-optimized control state.

[0130] This embodiment achieves tentative recovery and rigorous confidence accumulation verification of the AI ​​system in the background after degradation by triggering asynchronous inference on the second inference side when the abnormal recovery conditions are met, determining the compliance of the new control suggestion value, and switching back control to replace the safety control command only after a preset number of consecutive compliance determinations. Its beneficial effect is that it avoids system control oscillations caused by blindly switching back due to occasional compliance outputs of the AI, and ensures that the AI ​​smoothly takes over control only after it has truly restored its continuous and stable inference capabilities. This realizes a safe closed loop of system degradation and intelligent recovery, and maximizes the recovery of the benefits of AI-optimized control while adhering to the bottom line of industrial safety.

[0131] Figure 8 This is a schematic diagram of the structure of a secure interaction system provided in an embodiment of this application. See also... Figure 8 The secure interaction system 8 includes: a data interaction module 801, an asynchronous reasoning module 802, a security constraint module 803, and a result execution module 804.

[0132] The data interaction module 801 is used to respond to a data interaction request and provide the acquired industrial field data to the second inference side through the first control side; The asynchronous inference module 802, located on the second inference side, is used to perform asynchronous inference based on the industrial field data, generate control suggestion values ​​and feed them back to the first control side to complete the initial interactive response; The safety constraint module 803, located on the first control side, is used to determine the compliance of the control suggestion value according to the preset process safety constraint rules and obtain the determination result. The result execution module 804, located on the first control side, is used to generate a secure interaction result based on the judgment result. When the judgment result is compliant, the control suggestion value is confirmed as valid interaction data. When the judgment result is non-compliant for a preset number of consecutive times, an abnormal isolation mechanism is triggered to block the control suggestion value and output an isolation status or trigger a degradation response as a secure interaction result.

[0133] It should be noted that the secure interaction system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the secure interaction system and the secure interaction method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the secure interaction method embodiments, which will not be repeated here.

[0134] To support the efficient operation and spatiotemporal decoupling of the aforementioned system modules, this application provides a hardware and underlying system embodiment based on an asymmetric multiprocessing (AMP) architecture. See also Figure 9 The diagram illustrates the asymmetric multiprocessing hardware architecture and memory partitioning of the system. This underlying system serves as the hardware foundation for the aforementioned secure interaction system and specifically includes: The heterogeneous computing unit includes a multi-core central processing unit (CPU) and a dedicated AI acceleration coprocessor (such as an NPU or FPGA / PLD). The FPGA / PLD coprocessor is dedicated to hard real-time logic control tasks, responsible for millisecond-level response control in key stages of waste incineration; the NPU coprocessor is dedicated to AI computing tasks such as large-scale mathematical matrix operations and neural network inference. Hardware-level resource isolation between the two types of tasks is achieved to avoid resource contention issues at the hardware level.

[0135] The asymmetric dual-kernel architecture specifically includes: The real-time control domain corresponds to the first control side. It occupies one physical core of the CPU (such as Core 0), runs a real-time operating system (RTOS), and is responsible for executing the logic control and I / O data refresh of the IEC 61131-3 standard. It directly interacts with the DCS system in real time to ensure microsecond-level response in critical aspects such as furnace temperature control. The AI ​​computing domain corresponds to the second inference side. It utilizes the remaining CPU cores (such as Core 1-3) and the NPU accelerator, runs a general-purpose operating system (such as Linux), and is responsible for executing AI learning and inference tasks, running AI algorithms related to waste incineration to generate initial control instructions.

[0136] Cross-domain shared memory bus: Data exchange between the real-time control domain and the AI ​​computing domain is achieved through shared memory, such as a lockless circular buffer, avoiding the performance bottlenecks and deadlock risks associated with traditional mutex locks. Data transmission employs a priority sorting mechanism, setting real-time control commands and urgent on-site data as the highest priority to ensure that hard real-time control tasks are not interfered with by AI inference data transmission.

[0137] The hardware implementation of the safety constraint module is embedded within the real-time control domain and implemented using FPGA hardware circuitry. It is responsible for comprehensive safety verification of the inference results transmitted from the AI ​​computing domain. The safety feasible domain is defined based on: waste incineration industry standards such as GB 18485, on-site equipment safety parameters (i.e., the equipment manufacturer's rated operating range), and the waste incineration process mechanism. Ultimately, these parameters are pre-set and solidified in the FPGA hardware circuitry in the form of parameter thresholds, variation thresholds, and process mechanism matching ranges. Leveraging the high-speed parallel processing capabilities of the FPGA, millisecond-level forced verification of the AI ​​inference results is achieved.

[0138] This embodiment isolates the physical core through an asymmetric multiprocessing architecture, ensuring that the control cycle jitter of the real-time control domain remains minimal regardless of fluctuations in AI computing load, thus guaranteeing absolute real-time performance at the device's underlying level. Simultaneously, it achieves asynchronous collaboration between microsecond-level control and millisecond-level inference through a lock-free circular buffer, avoiding the deadlock risk caused by mutex locks and providing a robust hardware-level spatiotemporal decoupling foundation for the aforementioned secure interaction method.

[0139] In one specific implementation, in conjunction with the above system architecture, see [link to relevant documentation]. Figure 12 The flowchart shown illustrates the logic verification and circuit breaker mechanism of the security constraint module. The security constraint module 803 employs FPGA hardware-level verification combined with multi-dimensional collaborative verification, leveraging the high-speed parallel processing capabilities of the FPGA to achieve millisecond-level forced verification of AI inference results. The specific internal verification and circuit breaker logic of this module, in the context of waste incineration processes, is as follows: Step 1: Data Reception. The security constraint module receives preliminary control commands, such as grate speed and ammonia injection rate, transmitted by the AI ​​computing domain asynchronous inference module 802 via the cross-domain shared memory bus. At the same time, it retrieves the original input data corresponding to the AI ​​inference, such as real-time furnace temperature and flue gas NOx concentration.

[0140] Step 2: Input Data Backtracking Verification. The original input data undergoes a secondary validity check using the FPGA's built-in data processing module. If serious anomalies are found in the input data, such as missing data or a thermocouple open circuit causing the temperature deviation to exceed the range, the AI ​​inference result is directly deemed invalid, triggering the circuit breaker mechanism and skipping subsequent verifications.

[0141] Step 3: Threshold Verification. The FPGA threshold comparison circuit compares each parameter in the initial AI control command with the preset safety threshold range, such as comparing the furnace temperature setpoint to 850℃-1050℃ and the grate speed to 0.5-2.0m / min. If any parameter exceeds the threshold range, an anomaly is recorded, and the process proceeds to Step 6; if all parameters are within the threshold range, the process proceeds to Step 4.

[0142] Step 4: Change Amplitude Verification. The FPGA difference calculation module calculates the parameter change amplitude of the current AI command and the valid control command of the previous cycle, and compares it with the preset maximum allowable change amplitude, such as a single change in furnace temperature ≤ 50℃ and a single change in ammonia injection ≤ 10%. If the change amplitude exceeds the range, an anomaly is recorded, and the process proceeds to Step 6; if it is within the allowable range, the process proceeds to Step 5.

[0143] Step 5: Process Mechanism and Redundancy Comparison and Verification. On one hand, the FPGA inputs the AI ​​instructions into the fixed waste incineration process mechanism algorithm, calculates the theoretically reasonable value, and compares the deviation between the AI ​​instructions and the theoretical value. The allowable deviation is ≤3%. If it exceeds this, an anomaly is recorded. On the other hand, the inference results of the primary and backup AI algorithms in the AI ​​calculation domain are compared. If the deviation between the two sets of instructions is >5%, the inference result is determined to be uncertain and an anomaly is recorded. If neither of the above two items is abnormal, the AI ​​instruction is deemed qualified and transmitted to the real-time control domain for execution. If either item is abnormal, proceed to Step 6.

[0144] Step 6: Anomaly Detection and Circuit Breaker Trigger. Based on the verification results of Steps 2-5, if any anomaly is found, the AI ​​inference result is determined to be invalid or the control command is unsafe. The circuit breaker mechanism is immediately triggered, the inference task in the AI ​​computing domain is stopped, the output channel of the AI ​​control command is cut off, and a circuit breaker signal is sent to the result execution module 804 to start the backup safety control algorithm.

[0145] In one specific implementation, the degradation response and DCS control handover process of the result execution module, after receiving the fuse signal from the safety constraint module 803, the core of the degradation response triggered by the result execution module 804 is: automatically exiting the edge AI control mode and handing over control back to the automatic combustion control system (ACC) of the DCS system. The specific processing flow is as follows: Switching Triggering and Execution: When the safety constraint module triggers the circuit breaker mechanism, the real-time control domain immediately sends a control transfer request and abnormal information, including the abnormality type and parameters, to the DCS system via a preset communication protocol (such as Modbus TCP or OPC UA). Upon receiving the request, the DCS system automatically activates the ACC module to take over full-process control authority. The edge-end intelligent controller switches to data acquisition and feedback mode, only responsible for collecting and preprocessing on-site data and transmitting it to the DCS system, without generating any control commands. The switching process response time is ≤50ms, ensuring that the waste incineration operation is not interrupted or experiences drastic fluctuations.

[0146] Post-switch safety measures: After the ACC module takes over, the following actions will be performed: For stable operation control, prioritize adjusting all control parameters to safe baseline values, such as adjusting the furnace temperature to 950℃±20℃ and the grate speed to 1.0m / min, to avoid excessive parameter fluctuations and prevent furnace shutdown or coking. Abnormal adaptation and adjustment; if the cause of the abnormality is abnormal input data, such as instrument failure, the ACC module will automatically switch to backup instrument data for control; if it is due to AI inference deviation, the current safe operating conditions will be maintained and no aggressive adjustments will be performed. Early warning and coordination: Issues audible and visual warning signals to on-site maintenance personnel, and simultaneously uploads abnormal information to the maintenance management platform to remind them to investigate the fault.

[0147] Control rollback mechanism: After maintenance personnel troubleshoot and resolve AI control anomalies, they send an AI control recovery command to the edge intelligent controller through the maintenance management platform, and the AI ​​computing domain restarts asynchronous inference. The safety constraint module performs a comprehensive verification of new AI commands. After a preset number of consecutive verifications, such as 3 or 5 consecutive verifications, a control rollback request is sent to the DCS system. Once the DCS system confirms stable operating conditions, it automatically relinquishes control and restores the edge AI intelligent optimization control mode.

[0148] In one specific implementation, the model lightweight deployment unit and the hybrid compilation environment are described in [reference needed]. Figure 13 The diagram shown illustrates a hybrid programming interface combining AI function blocks and ladder diagrams within the integrated development environment. To address the issues of fragmented logic control and AI inference development, and the resulting integration difficulties, this application's system also includes a lightweight model deployment unit, enabling seamless integration from the host computer to the edge controller. Knowledge distillation and quantization: The trained high-performance deep learning model is subjected to knowledge distillation and INT8 / INT16 quantization on the host computer to generate a lightweight model adapted to edge computing power, reducing the computing resources and memory usage required for inference, and improving inference efficiency and response speed.

[0149] Automated Deployment and Hybrid Programming: The host computer software provides a programming interface compliant with the IEC 61131-3 standard, allowing users to directly drag and drop "AI inference function blocks" into the ladder diagram programming environment. Lightweight models, such as those in .onnx format, are automatically deployed to the system's AI computing domain via the compilation tool, and a shared memory variable mapping table between the real-time control domain and the AI ​​computing domain is automatically generated. This eliminates the need for manual writing of cross-domain low-level communication code, achieving collaborative work and seamless integration between hard real-time logic control and AI inference.

[0150] This application also provides an electronic device, which may specifically be an edge intelligent controller based on an asymmetric multiprocessing architecture, such as a waste incineration edge intelligent controller. Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0151] Typically, electronic device 10 includes one or more processors 1001 and one or more memories 1002.

[0152] Processor 1001 may include one or more processing cores, such as a multi-core ARM Cortex-A series processor. To achieve spatiotemporal decoupling of control and inference, processor 1001 adopts an asymmetric multiprocessing (AMP) architecture, physically divided into a real-time control domain and an AI computing domain. The real-time control domain occupies at least one physical core (e.g., Core 0) and integrates hardware logic circuits such as FPGA (Field-Programmable Gate Array) or PLA (Programmable Logic Array) to run a real-time operating system (RTOS), performing microsecond-level hard real-time logic control, I / O data refresh, and hardware-based security constraint verification. The AI ​​computing domain occupies the remaining processor cores (e.g., Core 1-3) and integrates a dedicated AI (Artificial Intelligence) accelerator processor (e.g., NPU) to run a general-purpose operating system (e.g., Linux) and perform deep learning inference tasks such as waste incineration optimization and pollutant emission prediction. This hardware-level resource isolation of heterogeneous computing units fundamentally avoids the interference of AI inference load fluctuations with the execution of logic control.

[0153] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory and non-volatile memory, such as one or more flash memory devices.

[0154] In one specific implementation, the memory is divided into two types: a memory for storing the first control-side program code and a memory for storing the second inference-side program code. These two types of memory are physically isolated. To ensure the security and reliability of the system, the memory for the first control-side program code must be physically isolated.

[0155] In some embodiments, the high-speed random access memory is logically divided into a dedicated area for the real-time control domain, a dedicated area for the AI ​​computing domain, and a cross-domain shared memory area. The cross-domain shared memory area is implemented using a lock-free circular buffer, used for efficient asynchronous data exchange between the real-time control domain and the AI ​​computing domain, avoiding mutex lock contention during concurrent read / write operations. The non-transitory computer-readable storage medium in memory 1002 stores at least one computer program, which is executed by processor 1001 to implement the secure interaction method provided in the embodiments of the secure interaction method described in this application.

[0156] Those skilled in the art will understand that Figure 10 The structure shown does not constitute a limitation on the electronic device 10, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0157] In addition, the device provided in the embodiments of this application may specifically be a chip, component or module. The chip may include a connected processor and a memory. The memory is used to store instructions. When the processor calls and executes the instructions, the chip can execute the secure interaction method provided in the above embodiments.

[0158] This embodiment also provides a computer-readable storage medium storing computer program code, which, when run on a computer, causes the computer to execute the secure interaction method provided in the above embodiment.

[0159] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to implement the secure interaction method provided in the above embodiment.

[0160] In this embodiment, the device, computer-readable storage medium, computer program product, or chip are all used to execute the secure interaction method provided above. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects of the secure interaction method provided above, and will not be repeated here.

[0161] Through the above description of the embodiments, those skilled in the art can understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0162] In the embodiments provided in this application, it should be understood that the disclosed apparatus and the described secure interaction method can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0163] The above description is only a specific implementation of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A secure interaction method, characterized in that, The method includes: In response to a data interaction request, the acquired industrial field data is provided to the second inference side via the first control side; Based on the industrial field data, asynchronous inference is performed by the second inference side to generate control suggestion values ​​and feed them back to the first control side, thus completing the initial interactive response. Based on the preset process safety constraint rules, the first control side performs a compliance determination on the control suggestion value and obtains the determination result. Based on the judgment result, a safe interaction result is generated. When the judgment result is compliant, the control suggestion value is confirmed as valid interaction data. When the judgment result is non-compliant for a preset number of consecutive times, an anomaly isolation mechanism is triggered to block the control suggestion value and output an isolation status or trigger a degradation response as a safe interaction result.

2. The secure interaction method according to claim 1, characterized in that, Based on the industrial site data, asynchronous inference is performed by the second inference side to generate control suggestion values ​​and feed them back to the first control side, completing the initial interactive response, including: The industrial field data and the corresponding collection timestamps are written into the shared data area through the first control side; Based on the shared data area, asynchronous inference is performed by asynchronously reading data from the second inference side to generate preliminary control suggestion values; Associating the initial control suggestion value with a corresponding timestamp yields an initial control suggestion value carrying a timestamp, and feeding the initial control suggestion value carrying a timestamp back to the shared data area; The first control side periodically polls the shared data area, matches and verifies the timestamp carried by the preliminary control suggestion value with the current control cycle, and determines the timestamp difference; When the timestamp difference does not exceed the preset timeliness threshold and the inference result is ready, the preliminary control suggestion value is determined as the control suggestion value; When the timestamp difference exceeds a preset timeliness threshold or the inference result is not ready, the preliminary control suggestion value is determined to be invalid, and the interaction data valid in the previous cycle is used as the control suggestion value to complete the preliminary interaction response.

3. The secure interaction method according to claim 2, characterized in that, The shared data area is implemented using a lock-free circular buffer, which is used to avoid mutex lock contention between the first control side and the second inference side during concurrent read and write operations.

4. The secure interaction method according to claim 1, characterized in that, The preset process safety constraint rules include threshold constraints and variation range constraints; The step of determining the compliance of the control suggestion value based on preset process safety constraints, through the first control side, and obtaining the determination result includes: Based on the threshold constraints in the preset process safety constraint rules, the control parameters in the control suggestion value are compared with the preset safety threshold range to obtain the first comparison result; When the control parameter exceeds the safety threshold range, the first comparison result triggers a non-compliance condition. Based on the variation range constraints in the preset process safety constraint rules, the parameter difference is obtained. The parameter difference is used to represent the variation range between the control suggestion value and the effective interaction data of the previous control cycle. The parameter difference is compared with the preset maximum allowable variation range to obtain a second comparison result; When the parameter difference exceeds the maximum allowable change range, the second comparison result triggers a non-compliance condition. If neither the first comparison result nor the second comparison result triggers a non-compliance condition, the resulting judgment is compliant.

5. The secure interaction method according to claim 1, characterized in that, The preset process safety constraint rules include process mechanism constraints and redundancy comparison constraints; The step of determining the compliance of the control suggestion value based on preset process safety constraints, through the first control side, and obtaining the determination result includes: Based on the process mechanism constraints in the preset process safety constraint rules, the control suggestion value is substituted into the preset process mechanism algorithm to obtain the theoretically reasonable value; The deviation between the recommended control value and the theoretically reasonable value is compared to obtain the third comparison result; When the deviation exceeds the preset allowable deviation range, the third comparison result triggers a non-compliance condition; Based on the redundancy comparison constraints in the preset process safety constraint rules, two sets of control suggestion values ​​are obtained. The two sets of control suggestion values ​​are used to represent the inference results generated by the main inference algorithm and the backup inference algorithm, respectively. The deviations between the two sets of control recommendation values ​​are compared to obtain the fourth comparison result; When the deviation exceeds the preset consistency threshold, the fourth comparison result triggers a non-compliance condition; If neither the third nor the fourth comparison result triggers a non-compliance condition, the resulting judgment is compliant.

6. The secure interaction method according to claim 1, characterized in that, The preset process safety constraint rules include input data validity constraints; The step of determining the compliance of the control suggestion value based on preset process safety constraints, through the first control side, and obtaining the determination result includes: Based on the input data validity constraints in the preset process safety constraint rules, obtain the original input data corresponding to the control suggestion value; The original input data is validated to obtain the validation result; If the verification result is abnormal, the verification result triggers a non-compliance condition, and the resulting judgment result is non-compliant. If the verification result does not trigger any non-compliance conditions, the resulting judgment is compliant.

7. The secure interaction method according to claim 1, characterized in that, When the judgment results are all non-compliant for a preset number of consecutive times, the exception isolation mechanism is triggered, including: When the judgment result is non-compliant, an abnormal observation period is initiated. During the abnormal observation period, the number of consecutive non-compliance judgments is accumulated, and the valid interaction data from the previous control cycle is used during the accumulation period. If the judgment results are all non-compliant for a preset number of consecutive times within the preset observation period, the abnormal isolation mechanism is triggered. The triggering of the degradation response includes: Stop the inference task on the second inference side and disconnect the output channel of the control suggestion value; Based on the industrial site data, safety control commands are generated.

8. The secure interaction method according to claim 7, characterized in that, After triggering the degradation response, the method further includes: When the preset anomaly recovery conditions are met, asynchronous reasoning is performed through the second reasoning side to generate new control suggestion values; Based on the preset process safety constraint rules, the new control recommendation value is judged for compliance, and the judgment result is obtained; When the judgment results are all compliant after a preset number of consecutive times, switch back to the first control side and replace the currently executed security control command with the new control suggestion value for interactive response.

9. A secure interactive system, characterized in that, A control system comprising a first control side and a second inference side, the system comprising: The data interaction module is used to respond to data interaction requests and provide the acquired industrial field data to the second inference side through the first control side; An asynchronous inference module, located on the second inference side, is used to perform asynchronous inference based on the industrial field data, generate control suggestion values ​​and feed them back to the first control side to complete the initial interactive response; The safety constraint module, located on the first control side, is used to determine the compliance of the control suggestion value according to the preset process safety constraint rules and obtain the determination result. The result execution module, located on the first control side, is used to generate a secure interaction result based on the judgment result. When the judgment result is compliant, the control suggestion value is confirmed as valid interaction data. When the judgment result is non-compliant for a preset number of consecutive times, an abnormal isolation mechanism is triggered to block the control suggestion value and output an isolation status or trigger a degradation response as a secure interaction result.

10. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable program code; A processor is configured to call and run the executable program code from the memory, causing the electronic device to perform the secure interaction method as described in any one of claims 1 to 8.

Images