Industrial control parameter optimization system and method based on neural-symbolic dual-flow architecture

Abstract

This invention provides an industrial control parameter optimization system and method based on a neural symbolic dual-stream architecture, belonging to the field of industrial intelligent control and automation technology. The system includes a slow-flow cognitive subsystem, a fast-flow control subsystem, and an intermediate data interaction and storage module. This invention achieves time-scale decoupling between AI cognitive computing and underlying mechanical control, resolving system resource competition and timing collapse issues caused by their simultaneous operation; it cuts off the direct drive link between the probabilistic AI model and the underlying physical actuators, eliminating the risk of unauthorized control and reshaping the underlying security defenses of industrial control; and it constructs a low-cost, low-risk, non-intrusive intelligent upgrade architecture adapted to existing legacy PLC / DCS systems, avoiding the high implementation risks of modifying the underlying core control logic.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control and automation technology, and in particular to an industrial control parameter optimization system and method based on a neural symbolic dual-stream architecture. Background Technology

[0002] Industry-driven needs: As the manufacturing industry transforms towards intelligence and flexibility, the demand for improved production efficiency, refined energy management, predictive equipment maintenance, and adaptive global optimization of process parameters has significantly increased in industrial settings. Traditional industrial control systems such as PLCs / DCS, designed based on static rule matrices, conventional linear PID algorithms, and fixed-sequence execution processes, struggle to cope with the numerous nonlinear variables arising from raw material fluctuations, environmental parameter changes, and equipment wear in dynamic manufacturing environments. They are unable to perform complex multivariate coupled decision-making and global real-time optimization, and they also cannot directly analyze and utilize unstructured tacit knowledge such as operator experience and maintenance manuals, resulting in a waste of industrial data assets.

[0003] Technological evolution trend: Generative artificial intelligence technology, represented by edge small language models (SLM), has the ability to fuse cross-modal data and make logical reasoning. Integrating its cognitive ability into industrial control loops to assist control decision-making has become the core technological evolution direction in the field of industrial automation.

[0004] Existing core technological bottlenecks: Current solutions for introducing large / small language models into industrial control loops suffer from two unavoidable technical pain points in engineering implementation. First, there's the issue of real-time mismatch and system execution jitter. The hard real-time control requirements of underlying industrial actuators and the soft real-time characteristics of AI model inference have a significant time scale difference. Time delay fluctuations in AI inference directly penetrate to the physical layer, disrupting the stability of the original closed-loop control. Second, there's the risk of uncontrollable security and unauthorized control. Generative AI inherently suffers from the "illusion" problem. Current solutions, where AI directly takes over physical actuators, cannot intercept abnormal out-of-bounds commands, easily leading to equipment damage and safety accidents. Summary of the Invention

[0005] This invention provides an industrial control parameter optimization system and method based on a neural symbolic dual-stream architecture. It decouples the time scale of AI cognitive computing and underlying mechanical control, resolving the system resource competition and timing collapse problems caused by their simultaneous operation on the same time axis. It cuts off the direct drive link between the probabilistic AI model and the underlying physical actuator, eliminating the risk of unauthorized control and reshaping the underlying security defense of industrial control. It also constructs a low-cost, low-risk, non-intrusive intelligent upgrade architecture adapted to existing aging PLC / DCS systems, avoiding the high implementation risk of modifying the underlying core control logic.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An industrial control parameter optimization system based on a neural symbolic dual-flow architecture includes a slow-flow cognitive subsystem, a fast-flow control subsystem, and an intermediate data interaction and storage module.

[0008] The slow-flow cognitive subsystem is deployed in edge computing devices in industrial settings. It has a built-in small language model finely tuned for the industrial vertical domain. It is used to acquire multimodal operating condition data of the industrial site through bypass listening. Combined with built-in historical work order data and process expert experience data, it performs global optimization calculation of process parameters with the goal of minimizing the overall production cost and generates suggested values ​​for process parameters. The slow-flow cognitive subsystem only has the permission to write suggested values ​​for process parameters to the intermediate data interaction register module. It is prohibited from issuing hardware interrupt commands and direct drive commands for physical actuators.

[0009] The intermediate data interaction register module is located in the accessible storage area of ​​the fast flow control subsystem, serving as a one-way communication isolation interface between the slow flow cognition subsystem and the fast flow control subsystem. It is used to receive and store the process parameter suggestion values ​​written by the slow flow cognition subsystem, and the writing action of the slow flow cognition subsystem to the intermediate data interaction register module does not trigger a hardware interrupt of the fast flow control subsystem.

[0010] The fast-flow control subsystem is deployed in industrial controller hardware closely attached to the physical actuators in the industrial field. It runs on an embedded real-time operating system and features hard real-time fixed-cycle cyclic scanning. It is granted unique control over the physical actuators within the system and is prohibited from responding to external hardware interrupt requests or drive commands directly issued by external systems. Within each fixed scan cycle, the fast-flow control subsystem actively retrieves currently stored process parameter suggestions from the intermediate data interaction register module. Within the same scan cycle, it performs a deterministic safety check on the retrieved process parameter suggestions. If the check passes, the process parameter suggestions are processed into the final control target value. This final control target value is then substituted into the built-in closed-loop control algorithm to generate a drive signal, which is sent to the corresponding physical actuator to complete the control action. If the check fails, the safe control target value from the previous scan cycle is maintained, and an anomaly alarm is triggered.

[0011] In this specification, the multimodal operating condition data acquired by the slow flow cognitive subsystem includes time-series operating condition data collected by industrial field sensors, process procedure text data, equipment maintenance record data, and experience record data of on-site operators; when the slow flow cognitive subsystem performs global optimization calculation of process parameters, it simultaneously meets the preset process quality constraints.

[0012] In this specification, the intermediate data interaction register module adopts a non-volatile data storage area inside the industrial controller, or an independent dual-port RAM module mounted on the industrial real-time bus; the intermediate data interaction register module only grants one-way write permission to the specified storage address to the slow flow cognitive subsystem, and only grants read permission to the corresponding storage address to the fast flow control subsystem.

[0013] In this specification, the industrial controller used in the fast flow control subsystem includes a programmable logic controller, a distributed control system controller, or an independent motion controller; the fixed cyclic scanning period of the fast flow control subsystem is 1 millisecond to 20 milliseconds, the time deviation of the scanning period is controlled at the microsecond level, and its internal control logic is hard-coded in accordance with the IEC 61131-3 international standard.

[0014] In this specification, the deterministic safety verification performed by the fast flow control subsystem on the recommended values ​​of process parameters includes static boundary interval verification and dynamic rate of change verification. The static boundary interval verification is used to determine whether the recommended values ​​of process parameters are within a preset safe physical range, and the dynamic rate of change verification is used to determine whether the parameter change rate corresponding to the recommended values ​​of process parameters is within a preset safe rate of change range.

[0015] In this specification, when the fast flow control subsystem performs static boundary interval verification and dynamic rate of change verification, if either verification fails, the currently retrieved process parameter recommendation value is determined to be invalid. The system maintains the safe control target value of the previous scan cycle, clears the invalid process parameter recommendation value in the intermediate data interaction register module, and triggers the corresponding type of abnormal alarm. If both verifications pass, the process parameter recommendation value is determined to be valid. The valid process parameter recommendation value is then subjected to smooth approximation processing to generate the final control target value.

[0016] In this specification, the slow-flow cognitive subsystem cannot directly access the input / output image area of ​​the fast-flow control subsystem, and cannot directly issue any drive commands to the physical actuator; the fast-flow control subsystem only executes the drive signals generated based on the final control target value within its own scan cycle, and does not execute any drive commands directly issued by external systems.

[0017] In this manual, when the system is deployed in an existing industrial control system, there is no need to modify the core closed-loop regulation and control program code built into the fast flow control subsystem. Only by allocating the corresponding storage address to the intermediate data interaction register module and granting the slow flow cognition subsystem one-way write permission to the corresponding storage address, the non-intrusive deployment of the system can be completed.

[0018] In this specification, the operating cycle of the slow-flow cognitive subsystem is not constrained by the fixed scanning cycle of the fast-flow control subsystem. The inference delay and computational load fluctuations of the slow-flow cognitive subsystem are buffered through the intermediate data interaction register module and will not interfere with the fixed scanning cycle of the fast-flow control subsystem.

[0019] The industrial control parameter optimization method based on the neural symbolic two-stream architecture, implemented based on any one of the above-mentioned industrial control parameter optimization systems based on the neural symbolic two-stream architecture, includes the following steps:

[0020] Step 1. The slow flow cognitive subsystem acquires multimodal operating condition data from the industrial site through bypass monitoring, combines it with built-in historical work order data and process expert experience data, performs global optimization calculation of process parameters, and generates suggested values ​​for process parameters.

[0021] Step 2. The slow flow cognitive subsystem silently writes the generated process parameter suggestions into the intermediate data interaction register module. The writing action does not trigger a hardware interrupt in the fast flow control subsystem.

[0022] Step 3. Within each fixed scan cycle, the fast flow control subsystem actively retrieves the currently stored process parameter suggestion values ​​from the intermediate data interaction register module;

[0023] Step 4. Within the same scan cycle, the fast flow control subsystem completes the deterministic safety verification of the pulled process parameter suggestion values. If the verification passes, the process parameter suggestion values ​​are processed into the final control target values, substituted into the built-in closed-loop control algorithm to generate drive signals, and sent to the corresponding physical actuators to complete the control actions. If the verification fails, the safety control target value of the previous scan cycle is maintained and an abnormal alarm is triggered.

[0024] In summary, this invention has at least the following beneficial effects: It constructs a bottom-level security isolation defense line for industrial control, achieving secure convergence of AI model output. Through architectural decoupling and a dual interception algorithm, the AI ​​model is strictly confined to the parameter optimization suggestion functional layer. Even if the large model exhibits unpredictable illusory output, the underlying hard-coded logic defense line can ensure that physical equipment maintains a safe operating state, completing control convergence from the probabilistic domain to the deterministic domain, fundamentally eliminating the security risks of AI overstepping its authority. It achieves dual decoupling in the spatiotemporal dimensions, completely resolving the core contradiction between AI computational latency and the industrial hard real-time control cycle. Slow and fast flows are completely decoupled at the physical and logical levels. Unstable factors such as inference latency, network fluctuations, and computational load changes in the slow flow are buffered and absorbed by static registers, without affecting the hard real-time scanning cycle of the fast flow. This allows the industrial system to fully retain the original core control indicators of hard real-time and low jitter while introducing large-model cognitive intelligence. It achieves low-cost, non-intrusive intelligent upgrades of existing industrial control systems. Without modifying the complex core closed-loop control code inside the old PLC / DCS system, only the read and write permissions of the corresponding static data blocks and specific addresses need to be opened to complete the access of the intelligent parameter optimization system. This greatly reduces the implementation risk and transformation cost of intelligent transformation of old industrial control equipment, and has strong engineering implementation versatility and scenario adaptability. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall workflow of the present invention based on the neural symbolic dual-stream architecture.

[0026] Figure 2 This is a schematic diagram of the internal autonomous fetching and deterministic verification logic of the fast stream involved in this invention. Detailed Implementation

[0027] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] like Figure 1 As shown, the system exhibits physical isolation and logical decoupling; the upper part is a slow flow, the lower part is a fast flow, and the middle part uses a data storage module for unidirectional writing and active retrieval of static values. This embodiment provides an industrial control parameter optimization system based on a neural symbolic dual-flow architecture, including a slow flow cognitive subsystem, a fast flow control subsystem, and an intermediate data interaction storage module;

[0029] The slow-flow cognitive subsystem is deployed in edge computing devices in industrial settings. It has a built-in small language model finely tuned for the industrial vertical domain. It is used to acquire multimodal operating condition data of the industrial site through bypass listening. Combined with built-in historical work order data and process expert experience data, it performs global optimization calculation of process parameters with the goal of minimizing the overall production cost and generates suggested values ​​for process parameters. The slow-flow cognitive subsystem only has the permission to write suggested values ​​for process parameters to the intermediate data interaction register module. It is prohibited from issuing hardware interrupt commands and direct drive commands for physical actuators.

[0030] The intermediate data interaction register module is located in the accessible storage area of ​​the fast flow control subsystem, serving as a one-way communication isolation interface between the slow flow cognition subsystem and the fast flow control subsystem. It is used to receive and store the process parameter suggestion values ​​written by the slow flow cognition subsystem, and the writing action of the slow flow cognition subsystem to the intermediate data interaction register module does not trigger a hardware interrupt of the fast flow control subsystem.

[0031] The fast-flow control subsystem is deployed in industrial controller hardware closely attached to the physical actuators in the industrial field. It runs on an embedded real-time operating system and features hard real-time fixed-cycle cyclic scanning. It is granted unique control over the physical actuators within the system and is prohibited from responding to external hardware interrupt requests or drive commands directly issued by external systems. Within each fixed scan cycle, the fast-flow control subsystem actively retrieves currently stored process parameter suggestions from the intermediate data interaction register module. Within the same scan cycle, it performs a deterministic safety check on the retrieved process parameter suggestions. If the check passes, the process parameter suggestions are processed into the final control target value. This final control target value is then substituted into the built-in closed-loop control algorithm to generate a drive signal, which is sent to the corresponding physical actuator to complete the control action. If the check fails, the safe control target value from the previous scan cycle is maintained, and an anomaly alarm is triggered.

[0032] In some embodiments, the multimodal operating condition data acquired by the slow flow cognitive subsystem includes time-series operating condition data collected by industrial field sensors, process procedure text data, equipment maintenance record data, and experience record data of field operators; when the slow flow cognitive subsystem performs global optimization calculation of process parameters, it simultaneously satisfies preset process quality constraints.

[0033] In some embodiments, the intermediate data interaction register module adopts a non-volatile data storage area inside the industrial controller, or an independent dual-port RAM module mounted on the industrial real-time bus; the intermediate data interaction register module only grants one-way write permission to the specified storage address to the slow flow cognitive subsystem, and only grants read permission to the corresponding storage address to the fast flow control subsystem.

[0034] In some embodiments, the industrial controller used in the fast flow control subsystem includes a programmable logic controller, a distributed control system controller, or an independent motion controller; the fixed cyclic scanning period of the fast flow control subsystem is 1 millisecond to 20 milliseconds, the time deviation of the scanning period is controlled at the microsecond level, and its internal control logic is hard-coded in accordance with the IEC 61131-3 international standard.

[0035] In some embodiments, the deterministic safety verification performed by the fast flow control subsystem on the recommended values ​​of process parameters includes static boundary interval verification and dynamic rate of change verification; the static boundary interval verification is used to determine whether the recommended values ​​of process parameters are within a preset safe physical range, and the dynamic rate of change verification is used to determine whether the parameter change rate corresponding to the recommended values ​​of process parameters is within a preset safe rate of change range.

[0036] In some embodiments, when the fast flow control subsystem performs static boundary interval verification and dynamic rate of change verification, if either verification fails, the currently retrieved process parameter recommendation value is determined to be invalid, the safety control target value of the previous scan cycle is maintained, the invalid process parameter recommendation value in the intermediate data interaction register is cleared, and the corresponding type of abnormal alarm is triggered; if both verifications pass, the process parameter recommendation value is determined to be valid, and a smooth approximation process is performed on the valid process parameter recommendation value to generate the final control target value.

[0037] In some embodiments, the slow-flow cognitive subsystem cannot directly access the input / output image area of ​​the fast-flow control subsystem and cannot directly issue any drive commands to the physical actuator; the fast-flow control subsystem only executes the drive signals generated based on the final control target value within its own scan cycle and does not execute any drive commands directly issued by external systems.

[0038] In some embodiments, when the system is deployed in an existing industrial control system, there is no need to modify the core closed-loop regulation and control program code built into the fast flow control subsystem. Only by allocating the corresponding storage address to the intermediate data interaction register module and granting the slow flow cognition subsystem one-way write permission to the corresponding storage address, the non-intrusive deployment of the system can be completed.

[0039] In some embodiments, the operating cycle of the slow-flow cognitive subsystem is not constrained by the fixed scanning cycle of the fast-flow control subsystem. The inference delay and computational load fluctuations of the slow-flow cognitive subsystem are buffered through the intermediate data interaction register module, and will not interfere with the fixed scanning cycle of the fast-flow control subsystem.

[0040] In some embodiments, the fine-tuning dataset for the small language model (SLM) for industrial vertical domains built into the slow flow cognitive subsystem includes at least two types of industrial scenario structured datasets and unstructured datasets. The structured datasets include historical time-series operating condition datasets, historical work order datasets, process parameter and product quality correlation datasets, and equipment operation lifecycle datasets corresponding to the industrial scenario. The unstructured datasets include process procedure texts, equipment maintenance manuals, on-site operator experience record texts, historical fault handling record texts, and industry safety standard texts corresponding to the industrial scenario.

[0041] In some embodiments, the core input of the small language model is a standardized multimodal feature vector. This feature vector is generated by the slow-flow cognitive subsystem after normalizing and extracting features from the multimodal operating condition data acquired by bypass monitoring. The feature vector dimension is fixed to the preset dimension adapted to the model input layer. The feature vector contains at least three types of core features: statistical features of time-series operating condition data, process constraint features, and equipment operating status features. The core output of the small language model is a suggested value of process parameters in a preset format. The output format is a numerical vector that corresponds one-to-one with the process parameter to be optimized. The precision and dimensions of each value in the vector are consistent with the field-collected data of the corresponding process parameter. There is no additional natural language description content, ensuring that it can be directly written into the intermediate data interaction storage module.

[0042] In some embodiments, the inference triggering rules of the small language model include two types: periodic triggering and event triggering. The inference period of periodic triggering is preset to 10s~300s, which can be adapted and adjusted according to the process inertia and operating condition change rate of the corresponding industrial scenario. The inference period is always greater than the scanning period of the fast flow control subsystem to ensure the decoupling of the time scale between slow flow inference and fast flow control. The triggering conditions of event triggering include sudden operating condition events, product quality deviation events, and equipment status abnormal events. When the slow flow cognition subsystem identifies the corresponding triggering event through bypass monitoring data, it immediately starts an asynchronous inference and updates the recommended values ​​of process parameters.

[0043] In some embodiments, the base model of the small language model is an open-source lightweight large language model with a parameter range of 1B to 10B that supports edge deployment, including but not limited to lightweight versions of the Llama series, lightweight versions of the Qwen series, and Phi series models. The base model needs to complete supervised fine-tuning and alignment training in the industrial vertical field. After fine-tuning, the inference latency of the model does not exceed 500ms, and it can be deployed and inferred locally on the AI ​​acceleration chip of the edge computing node without relying on cloud services.

[0044] In some embodiments, the nonlinear process quality constraint function Based on the process mechanisms, product quality standards, and industry safety regulations of the corresponding industrial scenarios, its general construction rule is: using the predicted future operating state vector The function takes quantifiable process quality indicators as input and outputs them. The function mapping relationship is determined by the process mechanism formula, the nonlinear regression model fitted to historical data, or the hard constraint rules in the expert rule base. The function output value is positively correlated with the process quality, that is, the larger the output value, the better the predicted process quality.

[0045] In some embodiments, the nonlinear process quality constraint function The input is Predicted operating condition state vector at time step This vector is generated by the slow-flow cognitive subsystem based on the current time. The data is obtained by time-series prediction of the operating conditions, the suggested values ​​of the process parameters to be evaluated, and the process context data. The vector contains at least the key process parameters directly related to product quality. The output of the function is a single-value quantified process quality score. The range of the score is pre-normalized to the interval [0,100], which corresponds to the full range of process quality from unqualified to optimal.

[0046] In some embodiments, for continuous production industrial scenarios, the nonlinear process quality constraint function The system employs a mechanism- and data fusion approach. Inputs include predicted continuous process parameters such as temperature, pressure, flow rate, and material ratio. Outputs are quantitative scores of predicted product purity, pass rate, and impurity content. The constraint is that the score must not be lower than a preset minimum pass threshold. For batch production industrial scenarios, the nonlinear process quality constraint function A phased construction approach is adopted, with corresponding sub-constraint functions set for different process stages of batch production. Each sub-function corresponds to the quality constraint requirements of its own process. The entire batch production process must ensure that the output of all sub-constraint functions is not lower than the minimum quality threshold of the corresponding stage.

[0047] In some embodiments, the maximum permissible change per cycle in the smooth approximation algorithm The tuning is determined based on three core dimensions: the mechanical characteristics of the corresponding actuator, the inertial characteristics of the process, and the sensitivity of product quality to parameter fluctuations. The general determination method is as follows: First, based on the rated mechanical parameters of the actuator, determine the maximum single-step adjustment amount for shock-free operation of the actuator, as... The upper limit threshold is determined; then, based on the inertial characteristics of the process, the maximum rate of change that will not cause oscillations in the process parameters is determined, and converted into the maximum change in a single cycle, which is used as... The intermediate threshold is determined; finally, based on the product quality's sensitivity to parameter fluctuations, the intermediate threshold is adjusted to obtain the final result. Values.

[0048] In some embodiments, The general reference principle is: for thermal processes with large inertia and large time lag, The value should not exceed 0.5% / s of the total width of the corresponding process parameter safety range, which is converted into the change within a single scan cycle of the fast flow control subsystem; for fast-response fluid control scenarios, The value should not exceed 2% / s of the total width of the corresponding process parameter safety range, which is converted into the change within a single scan cycle of the fast flow control subsystem; for high-precision mixing scenarios... The value of is no more than 0.1% / s of the corresponding process parameter's rated value, which is converted into the change within a single scan cycle of the fast flow control subsystem.

[0049] In some embodiments, the general judgment rule for the dynamic change rate verification performed by the fast flow control subsystem is as follows: for the recommended value of the process parameter to be verified, firstly, based on the actual operating parameters of the current cycle and the historical N cycles, calculate the actual change rate of the parameter; then compare the calculated actual change rate with the preset safe change rate threshold. If the actual change rate does not exceed the safe threshold, the dynamic change rate verification is determined to pass; if the actual change rate exceeds the safe threshold, the dynamic change rate verification is determined to fail.

[0050] In some embodiments, the entire process of anomaly handling for dynamic rate of change verification is as follows: First, when the dynamic rate of change verification is determined to fail, the fast flow control subsystem immediately triggers a rejection mechanism, refusing to apply the currently retrieved process parameter recommendation value; Second, a degradation clamping algorithm is executed, and the maximum allowable clamping value for the current cycle is calculated based on the difference between the actual rate of change of the current parameter and the safety threshold, clamping the control target value to a safe range; Third, the current process parameter recommendation value in the intermediate data interaction register is cleared, and a dynamic over-limit anomaly alarm is triggered, recording the abnormal event, abnormal parameter value, actual rate of change, and handling action in the system log; Fourth, the control target value is maintained unchanged for at least three subsequent consecutive scan cycles, and the normal parameter adjustment process is resumed after the parameter rate of change falls back to within the safety threshold.

[0051] In some embodiments, the proportional gain of the incremental PID controller Integral gain Differential gain The basic principle of its tuning is as follows: First, the basic parameters are tuned using the critical proportional gain method, the decay curve method, or the Ziegler-Nichols method to obtain the initial PID parameters; then, based on the control requirements of the corresponding industrial scenario, the parameters are adapted and adjusted, among which the proportional gain... The adjustment focuses on ensuring system response speed, with integral gain as the core objective. The adjustment focuses on eliminating the steady-state error of the system, and the differential gain The adjustment focuses on suppressing system overshoot and oscillation; the tuned PID parameters must ensure that when the system changes stepwise from the setpoint, the overshoot does not exceed 10%, the adjustment time does not exceed the maximum response time allowed by the process, and there is no continuous oscillation.

[0052] In some embodiments, the following precautions should be taken when tuning PID parameters to adapt to the dual-flow architecture of the present invention: First, PID parameter tuning must be completed based on the fixed scan cycle of the fast-flow control subsystem. During the tuning process, the scan cycle should be kept consistent with the actual operating cycle to ensure that the parameters are adapted to the hard real-time scan characteristics; Second, since the setpoint of the PID controller in this architecture is output by a smooth approximation algorithm without large step changes, the derivative gain can be appropriately reduced. The value of the PID parameter should be chosen to avoid excessive derivative action that could amplify system noise. Third, for retrofitting existing equipment, the PID parameters can be directly reused from the stable closed-loop control parameters already running in the original PLC / DCS system, without readjustment, ensuring the control stability of the system after non-intrusive retrofitting. Fourth, the integral gain... The tuning needs to take into account both the steady-state accuracy of the system and the ability to resist integral saturation. It can be used in conjunction with the integral limiting algorithm to avoid the integral saturation problem caused by the long-term limitation of parameters.

[0053] In some embodiments, to address the technical problem of semantic misalignment between the neural semantic space of the slow-flow cognitive subsystem and the symbolic rule space of the fast-flow control subsystem, resulting in implicit process conflicts in the parameter suggestion values ​​output by the SLM even after single-dimensional numerical verification, a neural-symbolic bidirectional semantic alignment mechanism is constructed. This bidirectional semantic alignment mechanism consists of two stages: pre-constraint injection and post-consistency verification.

[0054] In the pre-constraint injection stage, the fast-flow control subsystem converts all its hard-coded symbolic rules (including single-parameter safety intervals, multi-parameter interlocking rules, process timing constraints, and equipment interlocking logic) into first-order logic predicate expressions and constraint prompt templates conforming to the SLM input format. These are then unidirectionally transmitted to the slow-flow cognitive subsystem through the dedicated read-only constraint storage area of ​​the intermediate data interaction register module. Before executing global optimization reasoning, the slow-flow cognitive subsystem fixes the constraint prompt template to the prefix of the reasoning prompt and uses the first-order logic predicate expression as a forced stopping condition in the reasoning process. This ensures that the optimization reasoning of the SLM is subject to the hard constraints of the fast-flow symbolic rules throughout the entire process, and parameter optimization is only performed within the process space allowed by the symbolic rules.

[0055] In the post-consistency verification stage, before the static boundary interval verification, the fast flow control subsystem adds a symbolic consistency verification step. The retrieved process parameter recommendations are substituted into all symbolic rules for SAT satisfiability verification. Only when the parameter recommendations satisfy the logical constraints of all symbolic rules will the subsequent numerical verification stage begin. If the verification fails, the parameter recommendations are directly determined to be invalid, and the safe state of the previous cycle is maintained. At the same time, the unsatisfied rule entries are fed back to the slow flow cognitive subsystem, triggering the slow flow's secondary constraint reasoning.

[0056] This embodiment injects the hard-coded symbol rules of the underlying PLC into the inference pre-process of the SLM, constructing a two-way closed loop where "neural inference is constrained by symbol rules throughout the process and symbol verification is based on neural semantic alignment". This reduces the conflict between AI output and underlying control rules from the root, rather than just doing post-hoc remediation, thus improving the safety protection effect.

[0057] In some embodiments, to address the technical problem that single-dimensional parameter boundary verification cannot intercept the implicit risk of "each single parameter is within the safe range, but the combination of multiple parameters exceeds the critical safety boundary of the process", a multi-parameter coupled dynamic safety envelope verification mechanism is constructed. This mechanism, together with the original static boundary verification and dynamic rate of change verification, constitutes a three-level safety verification system for the fast flow control subsystem.

[0058] The dynamic safety envelope verification mechanism is implemented as follows: First, the slow-flow cognitive subsystem, based on the historical full-condition data from bypass monitoring, uses an algorithm that fuses SLM and Gaussian process regression to fit a multi-parameter coupled critical safety surface corresponding to the industrial scenario. The input to the critical safety surface is N core control parameters directly related to process safety, and the output is the process safety margin under the parameter combination. When the safety margin is ≤0, it means that the parameter combination has exceeded the critical safety boundary. The slow-flow cognitive subsystem converts the fitted critical safety surface into a piecewise linearized symbolic equation executable by the fast-flow control subsystem. It then periodically updates the fast-flow control subsystem through the dedicated storage area of ​​the intermediate data interaction register module. The update cycle is consistent with the inference cycle of the slow-flow, and the update action does not trigger a hardware interrupt of the fast-flow.

[0059] After the fast flow control subsystem passes the static boundary interval verification, it immediately performs dynamic safety envelope verification: it substitutes the currently pulled multiple sets of process parameter recommendations into the symbolic critical safety surface equation to calculate the safety margin of the current parameter combination; if the safety margin is greater than the preset safety threshold, the envelope verification is deemed to have passed and the system proceeds to the subsequent dynamic rate of change verification stage; if the safety margin is less than or equal to the preset safety threshold, the system determines that the parameter combination has a hidden risk of exceeding the limit, directly triggers the clamping algorithm, refuses to apply the set of parameter recommendations, and simultaneously feeds back the alarm information of insufficient safety margin and the critical parameter combination to the slow flow cognition subsystem to correct the optimization space of the slow flow.

[0060] This embodiment combines "slow flow high-dimensional fitting of safety surface - conversion to PLC executable symbolic equation - fast flow real-time coupling verification" to utilize the high-dimensional fitting capability of AI while retaining the deterministic execution characteristics of fast flow, thus solving the long-standing but undetected hidden out-of-bounds problem and significantly improving the system's safety protection dimension.

[0061] In some embodiments, to address the technical problem of timestamp misalignment between the operating condition data of the slow-flow asynchronous sensing and the real-time sampling data of the fast-flow, which leads to a mismatch between the optimization parameters and the actual operating conditions during the execution of the fast-flow and causes control deviations, a spatiotemporal matching mechanism of asynchronous sensing-synchronous anchoring is constructed to achieve spatiotemporal alignment between the slow-flow optimization parameters and the actual operating conditions of the fast-flow without destroying the hard real-time scanning characteristics of the fast-flow.

[0062] The spatiotemporal matching mechanism is implemented as follows: First, the fast-flow control subsystem marks the collected operating condition sensor data with a global synchronization timestamp corresponding to the corresponding period within each scanning cycle. The global synchronization timestamp is generated based on the PLC's internal crystal oscillator clock and is strictly synchronized with the scanning cycle, with a timestamp accuracy of no less than 1ms. The fast-flow control subsystem sets up a dedicated operating condition snapshot ring buffer in the intermediate data interaction register module. The full amount of operating condition snapshot data with timestamps is written into the buffer at fixed intervals. The buffer retains the operating condition data of the most recent N cycles. The slow-flow cognition subsystem can only perform read-only operations on the buffer and cannot write to it.

[0063] When the Slow Flow cognitive subsystem performs global optimization, it first reads the latest operating condition snapshot data from the circular buffer. Using the timestamp of this snapshot as the reference anchor point, it performs operating condition prediction and parameter optimization. The generated process parameter recommendation values ​​must include the timestamp of the reference anchor point and mark the effective time window of the parameter recommendation value. The Slow Flow writes the parameter recommendation values ​​with timestamp and effective time window into the intermediate data interaction register module.

[0064] When the fast-flow control subsystem retrieves parameter suggestions, it first checks whether the difference between the reference anchor timestamp of the parameter and the timestamp of the current scan cycle is within the parameter's valid time window. If the difference is within the valid time window, the parameter's spatiotemporal matching is deemed valid, and the system proceeds to the subsequent safety verification stage. If the difference exceeds the valid time window, the parameter is deemed outdated and invalid, and the parameter suggestion is discarded directly. The system maintains the safe operating state of the previous cycle and triggers a re-inference of the slow-flow system. The duration of the valid time window is preset according to the rate of change of the corresponding industrial scenario and is always greater than the maximum inference delay of the slow-flow system but less than the process inertia time constant.

[0065] This embodiment, without compromising the hard real-time performance of the fast stream or altering the core decoupling architecture of the fast and slow streams, solves the inherent spatiotemporal misalignment problem of the asynchronous architecture through a non-obvious design of "timestamp anchoring - snapshot buffer - effective time window verification," significantly improving the accuracy of parameter optimization and enhancing control performance.

[0066] In some embodiments, to address the technical problem that older PLCs with limited computing power cannot support newly added multi-dimensional security verification algorithms and that forcibly implanting them would disrupt the original hard real-time scanning cycle, a lightweight shadow verification and dual-track rollback mechanism is constructed to achieve a non-intrusive security upgrade with zero computing power consumption, without modifying the original core control code and scanning logic of the older PLCs.

[0067] The shadow verification and dual-track fallback mechanism is implemented as follows: the security verification logic of the fast flow control subsystem in the original scheme is split into two levels: basic backup verification and full shadow verification, wherein:

[0068] The basic safety check retains only the core single-parameter static boundary clamping logic, which is written into the end of the original scan cycle of the old PLC in a very simple ladder diagram / structured text code. The code execution time does not exceed 5% of the original scan cycle and does not affect the hard real-time characteristics of the PLC at all. The basic safety check only intercepts extreme out-of-bounds parameter values. After the check passes, the parameters are directly substituted into the original PID closed loop for execution.

[0069] The full shadow verification logic is deployed in the edge computing node where the slow flow cognitive subsystem is located, and is executed in strict synchronization with the scan cycle of the old PLC: the edge node obtains the actual operating parameters of each scan cycle of the PLC, the pulled parameter suggestion values, and the sensor feedback data in real time through bypass listening. The full safety verification logic of symbol consistency verification, multi-parameter coupling envelope verification, and dynamic rate of change verification is executed in parallel in the edge node. The verification process is synchronized with the execution process of the PLC, and the verification delay does not exceed the scan cycle of the PLC.

[0070] The execution logic of the dual-track rollback mechanism is as follows: if all shadow checks pass, the normal execution of the PLC is not interfered with; if the full shadow check determines that the parameters have a security risk, the edge node immediately writes the preset safety rollback parameters and emergency clamping instructions to the PLC through the dedicated emergency control address of the intermediate data interaction register module. The basic backup check logic of the PLC responds to the emergency clamping instruction first, forcibly switching the control target value to the safety rollback parameters, and triggering an emergency alarm at the same time. The whole process does not require modification of the original closed-loop control code of the PLC. Only one line of emergency instruction reading logic needs to be added to the original scan cycle, and the occupation of PLC computing power is negligible.

[0071] This embodiment separates the "edge-side shadow verification + PLC-side baseline verification" process, which does not occupy the core computing power of the old PLC at all, while achieving full-dimensional security protection and further amplifying the core advantages of non-intrusive transformation.

[0072] In some embodiments, to address the technical problems of existing technologies that can only intercept AI hallucination output after the fact and cannot suppress hallucinations at the source, and that SLM models cannot adapt to continuous optimization due to working condition drift, a hallucination suppression and model self-optimization mechanism based on neural-symbolic closed-loop iteration is constructed. The symbolic execution results of fast flow are used to inversely constrain and optimize the slow flow SLM model, thereby achieving source suppression of hallucinations and lifelong self-optimization of the model.

[0073] The closed-loop iteration mechanism is implemented as follows: First, a dedicated operation feedback log area is set up in the intermediate data interaction register module. In each scan cycle, the fast flow control subsystem writes the verification results of the parameter suggestion value, the final executed control parameters, the actual working condition data fed back by the sensor, the process quality detection results, and the abnormal alarm information into the log area after marking the corresponding timestamp. The slow flow cognition subsystem can only perform read-only operations on the log area.

[0074] The slow-flow cognitive subsystem reads the operation feedback logs at fixed intervals to construct a model optimization dataset. The dataset is divided into a positive sample set and a negative sample set. The positive sample set contains the parameter suggestions and corresponding operating condition data for samples that have passed verification, executed successfully, and whose process quality and energy consumption indicators have reached the optimization targets. The negative sample set contains the parameter suggestions and corresponding operating condition data for samples that have failed verification and been intercepted, caused abnormal operating conditions after execution, or whose process quality is not up to standard. The reasons for the interception of the samples are also labeled (such as symbol rule conflicts, coupling envelope out-of-bounds, rate of change exceeding limits, etc.).

[0075] In the hallucination suppression stage, the slow-flow cognition subsystem constructs a hallucination constraint rule base based on the negative sample set. It converts the parameter range, logical conflict rules, and working conditions corresponding to the negative samples into negative prompts and prohibitions for SLM inference. These are loaded before each optimization inference to prevent the SLM from outputting parameter suggestion values ​​that have been verified as invalid or illegal from the source.

[0076] In the model self-optimization phase, the Slow Flow cognitive subsystem performs low-rank adaptation (LoRA) lightweight fine-tuning based on positive and negative sample sets. The objective function of the fine-tuning simultaneously includes two optimization objectives: minimizing overall production costs and maximizing parameter validation pass rate. This ensures that the parameter suggestions output by the fine-tuned model satisfy both the economic optimization objective and conform to the symbolization rules of Fast Flow to the greatest extent, significantly reducing the probability of hallucination output. The fine-tuning process is performed during the idle computing power periods of edge computing nodes, without affecting the normal inference business of Slow Flow. The fine-tuned model needs to undergo offline testing, and only after the validation pass rate reaches a preset threshold will it be switched to the online inference model.

[0077] This embodiment upgrades the "one-way parameter transfer" to "neural-symbolic bidirectional closed-loop iteration", which suppresses AI illusions from the source and solves the industry pain point that SLM models cannot adapt to working condition drift.

[0078] In some embodiments, to address the technical issues that a single-controller dual-stream architecture cannot achieve global collaborative optimization of multiple devices at the production line level, and that cross-device interlocking protection cannot adapt to dynamic optimization parameters, a production line-level global collaborative optimization and interlocking protection mechanism for a distributed dual-stream architecture is constructed. The original single-node dual-stream architecture is extended into a three-level distributed architecture of "global slow flow - local slow flow - local fast flow", which achieves global optimization and dynamic interlocking protection at the production line level while ensuring the hard real-time security of a single device.

[0079] The three-level distributed architecture is implemented as follows:

[0080] The production line-level global slow flow subsystem is deployed in a production line-level industrial server. It has a built-in production line-level global optimization SLM model. It acquires the operating data, material flow data, and quality inspection data of all equipment in the production line through bypass listening. With the goal of minimizing the overall production cost and maximizing the overall capacity of the production line, it performs production line-level global parameter optimization, generates the target parameter constraint range and cross-device collaboration rules for each device, and writes them to the global constraint storage area of ​​the intermediate data interaction register module of the corresponding device through the production line industrial Ethernet. The write operation does not trigger any local fast flow hardware interruption.

[0081] The device-level local slow flow subsystem is deployed in the edge computing node of the corresponding device. It has a built-in single-device optimization SLM model. Within the parameter constraint range issued by the global slow flow, it performs fine-grained parameter optimization at the single device level and generates process parameter suggestion values, which are written to the local intermediate data interaction register module. The optimization process of the local slow flow is restricted by the global constraint range and cross-device collaboration rules throughout the process and must not break the global constraints.

[0082] The device-level local fast flow subsystem is deployed in the PLC of the corresponding device. It maintains the original hard real-time scanning cycle and safety verification logic, while adding a cross-device dynamic interlock verification step: The local fast flow obtains the current operating parameters of the upstream and downstream related devices in the production line through the production line real-time bus, and substitutes the retrieved local parameter suggestion values ​​and the upstream and downstream device parameters into the cross-device collaboration rules issued by the global slow flow to perform interlock verification. The parameter suggestion value is only executed when the interlock verification passes and all local safety verifications pass. If the interlock verification fails, clamping back is directly triggered, and the interlocking abnormality information is reported to the global slow flow subsystem to trigger the re-optimization of global parameters.

[0083] This embodiment extends the dual-stream architecture in a distributed manner, retaining the core advantages of single-node "decoupling of fast and slow streams and hard real-time security protection" while realizing global collaborative optimization and dynamic interlocking protection at the production line level. It breaks through the application boundaries of conventional technologies in the field and has extremely strong industrial application value.

[0084] In some embodiments, for the industrial vertical domain fine-tuning small language model built into the slow-flow cognitive subsystem, the selection range and selection judgment rules of its basic model are clarified. The selection of the basic model must simultaneously meet three core principles: industrial edge deployment adaptability, industrial scenario inference adaptability, and intellectual property compliance. The selection range is limited to open-source lightweight large language models with 1B to 10B parameters that support local deployment on the edge, including but not limited to the lightweight versions of the Llama series, Qwen series, Phi series, and Gemma series open-source models. The selection judgment indicators include: single inference latency not exceeding 500ms, edge memory usage not exceeding 16GB, support for INT4 / INT8 quantization compression, open-source license allowing commercial modification and deployment in industrial scenarios, and having a complete low-rank adaptation (LoRA) lightweight fine-tuning interface to ensure that the selected model can fully adapt to the hardware conditions and operating requirements of industrial edge computing devices.

[0085] In some embodiments, for the training dataset of the small language model fine-tuned in the industrial vertical field, the classification composition, preprocessing rules and annotation specifications of the dataset are clearly defined. The training dataset is divided into two main categories: structured industrial datasets and unstructured industrial datasets, wherein:

[0086] The structured industrial dataset includes historical time-series operating condition datasets corresponding to industrial scenarios, process parameter-product quality correlation datasets, equipment lifecycle operation datasets, and historical energy consumption and cost statistics datasets. The datasets need to cover all operating condition scenarios, including equipment start-up and shutdown, stable operation, operating condition fluctuations, and fault anomalies. The data sample size for a single scenario should not be less than 10,000 records. During preprocessing, outlier removal, missing value interpolation, time-series alignment, and normalization and standardization should be completed, and the datasets should be divided into training set, validation set, and test set in an 8:1:1 ratio.

[0087] Unstructured industrial datasets include process procedure texts, equipment maintenance manuals, industry safety standards, historical fault handling records, on-site operator experience records, and historical process optimization work order data corresponding to industrial scenarios. During preprocessing, text cleaning, deduplication, sentence segmentation, industrial entity recognition, and relation extraction are performed. During annotation, the process constraints, safety boundary conditions, parameter optimization experience, and fault handling logic corresponding to the text content are clearly defined, forming an industrial instruction-response pair annotation dataset with no less than 5,000 annotation samples per scenario.

[0088] In some embodiments, the core objectives and phased training logic of the fine-tuning of the small language model in the industrial vertical field are clearly defined. The core objectives of the fine-tuning are multi-objective joint optimization, including: First, maximizing the parameter verification pass rate, so that the probability of the process parameter suggestion value output by the model passing the full safety verification of the fast flow control subsystem is not less than 99%, thereby reducing the probability of invalid output and phantom output from the source; Second, maximizing the optimization target achievement rate, so that the parameter suggestion value output by the model can achieve the optimization target of reducing overall production costs and improving process quality, with a target achievement rate of not less than 95%; Third, controllable inference latency, ensuring that the single-round inference latency of the model on the target edge hardware is stable within 500ms after fine-tuning; Fourth, phantom output suppression, so that the parameter value output by the model is strictly limited within the process safety feasible domain, with no invalid output exceeding the physical limits of the equipment and process constraints;

[0089] The phased training logic consists of three stages: The first stage is domain pre-training, which involves continuing pre-training based on an unstructured text dataset covering the entire industrial scenario, enabling the model to learn professional terminology, process logic, safety regulations, and industry knowledge within the vertical industrial domain; the second stage is supervised fine-tuning, which involves performing LoRA lightweight fine-tuning on the dataset based on labeled instruction-response pairs, allowing the model to adapt to the core task of "multimodal operating condition data input - process parameter suggested value output," aligning with the optimization requirements of industrial control scenarios; the third stage is alignment training, which involves building a reward model based on the safety verification rules and process constraints of the fast flow control subsystem, and completing model alignment through human feedback reinforcement learning (RLHF) or direct preference optimization (DPO), ensuring that the model output strictly conforms to industrial safety hard constraints and suppressing hallucination output.

[0090] In some embodiments, the input / output interface specifications of the small language model are clearly defined to ensure that the model's input / output is fully compatible with the slow-flow cognitive subsystem and the intermediate data interaction and storage module, wherein:

[0091] The input interface adopts a standardized multimodal feature vector input specification. The fixed input format is a two-dimensional feature matrix that matches the dimension of the model input layer. The matrix rows correspond to the feature types, and the columns correspond to the temporal dimensions of the features. The input interface can automatically receive the temporal condition features, process constraint features, equipment status features, and expert experience semantic features preprocessed by the slow flow cognitive subsystem without additional manual input. The input interface also reserves access ports for trigger signals such as sudden change events and quality deviation events to adapt to the asynchronous inference requirements triggered by events.

[0092] The output interface adopts a fixed-format numerical vector output specification. The output content is a numerical vector that corresponds one-to-one with the process parameter to be optimized. The precision and dimensions of each value in the vector are completely consistent with the field-acquired data of the corresponding process parameter and the parameter operation precision of the fast flow control subsystem, without any additional natural language description. The output vector is accompanied by three types of additional information: parameter reference anchor point timestamp, parameter effective time window, and parameter corresponding operating condition scenario code. The additional information adopts a fixed-length standardized encoding format and is written together with the parameter numerical vector into the designated storage address of the intermediate data interaction register module. The output interface is strictly limited to outputting only process parameter values ​​and prohibits the output of non-parameter content such as control commands, interrupt commands, and hardware driver code, which is fully compatible with the permission control rules of the slow flow cognitive subsystem.

[0093] In some embodiments, for the multimodal operating condition data acquired by the slow-flow cognitive subsystem, its preprocessing logic is defined. This preprocessing logic is executed by the slow-flow cognitive subsystem on a local edge computing device without interfering with the normal operation of the fast-flow control subsystem. Differentiated preprocessing is performed for different modal data.

[0094] For time-series operating condition data collected by sensors in industrial sites, outlier processing, missing value filling, resampling alignment, and normalization are performed sequentially. Outliers are removed using a combination of the 3σ criterion and box plot method, missing values ​​are filled using linear interpolation or Kalman filtering, resampling is aligned to a uniform sampling frequency, and finally, the min-max normalization method is used to map the data to the [0,1] interval to eliminate the influence of differences in the dimensions of different parameters on feature extraction.

[0095] For unstructured text data such as process specifications, equipment maintenance records, and operator experience records, text cleaning, word segmentation, industrial domain entity extraction, and semantic vectorization are performed sequentially. A pre-trained word vector model in the industrial vertical domain is used to complete text vectorization. A pre-trained named entity recognition model is used to extract core entities such as process parameters, equipment names, safety constraints, and optimization experience from the text. Finally, a fixed-dimensional semantic feature vector is generated through average pooling to achieve the numerical transformation of unstructured text data.

[0096] In some embodiments, the fusion method of multimodal operating condition data is specified, and a two-level fusion architecture of "intramodal feature extraction - cross-modal feature fusion" is adopted to achieve deep fusion of time-series operating condition data and unstructured text data, wherein:

[0097] The first stage is intramodal feature extraction. For the preprocessed time-series operating condition data, a one-dimensional convolutional neural network (1D-CNN) and a gated recurrent unit (GRU) are used to jointly extract time-series dynamic features, operating condition change trend features, and equipment operating status features. For the preprocessed text semantic feature vector, a fully connected network is used to extract process constraint features, expert experience features, and safety rule features, and generate fixed-dimensional single-modal feature vectors respectively.

[0098] The second stage is cross-modal feature fusion, which uses a cross-modal attention mechanism to achieve adaptive fusion of different single-modal features. Through attention weight allocation, it automatically focuses on core effective features under different operating conditions. When the operating conditions are stable, the weight ratio of time-series dynamic features is increased; when the operating conditions fluctuate or quality is abnormal, the weight ratio of process constraints and expert experience features is increased. The mathematical expression for the fusion process is: ;in, For the fused cross-modal features, It is a time-series single-modal feature. For textual unimodal features, For cross-modal attention computation function, For feature splicing operations, , The weight matrix and bias terms are learnable, and the resulting fused feature has a fixed dimension, which can be directly input into a small language model to complete subsequent inference.

[0099] In some embodiments, the full-dimensional working condition feature vector is explicitly defined. The construction method of the feature vector As the core input for the slow-flow cognitive subsystem to perform global optimization calculations, a modular fixed-dimensional construction method is adopted. The total dimension of the feature vector is completely matched with the dimension of the input layer of the small language model. Its composition includes four core feature modules: The first module is the time-series operating condition statistical feature module, accounting for 40% of the dimension. It includes the statistical features of the mean, variance, extreme values, rate of change, and fluctuation amplitude of process parameters within the current time and historical sliding window, as well as the real-time numerical features of core operating parameters such as temperature, pressure, flow rate, and material ratio collected by sensors, reflecting the real-time operating status of the current operating condition; The second module is the process constraint rule feature module, accounting for 20% of the dimension. It includes process quality requirements, safety, and other features. The system comprises four modules: a standard module for encoding hard constraints, a module for encoding hard constraints, a module for encoding parameter safety ranges, and a module for encoding multi-parameter interlocking constraints; a module for defining the feasible domain boundaries for parameter optimization; a module for defining equipment operating status features (20% dimensionality), which includes equipment operating time, predicted wear levels, status features corresponding to maintenance records, and equipment failure risk probability features, reflecting the current health status and operational capabilities of the equipment; and a module for defining expert experience semantic features (20% dimensionality), which consists of fused unstructured text semantic features, including semantic features corresponding to historical optimization experience, fault handling experience, and operating condition adaptation experience, realizing the numerical embedding of implicit expert knowledge; and a feature vector. The slow-flow cognitive subsystem completes the update before each inference, and the update cycle is consistent with the slow-flow inference cycle. After the update is completed, a validity check is performed. Only after the check passes can the small language model be input to ensure that the feature vector is completely matched with the current real-time operating conditions.

[0100] In some embodiments, a secure physical zone is defined. The determination method and preset principles are as follows: the safe physical zone adopts a three-level hierarchical setting method, from the outside to the inside, namely the equipment limit zone, the process safety zone, and the optimized operation zone. The determination rules for each level of zone are as follows:

[0101] The limiting range of the equipment is the absolute boundary of the parameter. , Based on the rated parameters on the nameplate of the actuator and production equipment, mechanical and physical limits, and national and industry mandatory safety standards, these parameters are absolute red lines that cannot be crossed, and under no circumstances are they allowed to exceed this range.

[0102] The process safety zone is nested within the equipment limit zone. , Based on the process procedures, product quality standards, and process safety interlock rules for the corresponding industrial scenario, and with reference to the parameter fluctuation range of historical stable operating conditions, it is ensured that when the parameters are running within this range, the basic process production requirements and safety requirements can be met.

[0103] The optimized operating range is nested within the process safety range. , Based on historical process optimization data and expert experience, the core feasible region for parameter optimization of the slow flow cognitive subsystem is determined, ensuring that the optimization calculation is always performed within a safe range with optimization space.

[0104] Once the interval is preset, it needs to be jointly reviewed and confirmed by process experts and equipment engineers. After the review is approved, it is hard-coded into the control logic of the fast flow control subsystem and can only be modified through authorized offline configuration. The slow flow recognition subsystem and any external system have no right to modify the interval parameters.

[0105] In some embodiments, the maximum allowable step size is specified. The entire process tuning method and preset principles, the The tuning process must simultaneously consider the mechanical safety of the actuator, the stability of the process, and the consistency of product quality. A three-level quantitative tuning process is used to determine the final value.

[0106] The first step is to tune the mechanical characteristic boundaries of the actuator. Based on the actuator's rated mechanical parameters, response speed, and shock-free operation requirements, the maximum adjustment that the actuator can withstand in a single cycle is calculated as... The upper limit threshold is set to ensure that parameter adjustments do not cause mechanical shock, increased wear, or malfunction of the actuator;

[0107] The second step involves adapting and tuning the inertial characteristics of the process. Based on the inertial time constant and lag time of the process, a step response test is used to determine the maximum parameter change rate that will not cause oscillations or exceed limits in the process parameters. This rate is then converted into the maximum change within a single scan cycle of the fast flow control subsystem, and used as... The intermediate threshold;

[0108] The third step is product quality sensitivity correction and tuning. Based on the test results of the product quality sensitivity to parameter fluctuations, the intermediate threshold is adjusted. The more sensitive the product quality is to parameter fluctuations, the higher the sensitivity. The smaller the value, the better, ensuring that smooth parameter adjustments will not cause excessive fluctuations in product quality.

[0109] Preset for different industrial scenarios General reference specifications: for thermal processes with large inertia and large time lag, The value should not exceed 0.5% / s of the total width of the corresponding process parameter safety range, and is converted into the change within a single scan cycle of a fast flow; for fast-response fluid control scenarios, The value should not exceed 2% / s of the total width of the safety range for the corresponding process parameters; this applies to high-precision mixing and precision machining scenarios. The value should not exceed 0.1% / s of the corresponding process parameter's rated value.

[0110] In some embodiments, the PID gain parameter is explicitly defined. , , The tuning method and preset principles adapted to this dual-stream architecture are as follows: the parameter tuning is divided into three stages: basic tuning, architecture adaptation optimization, and field verification, to ensure smooth approximation of PID control and complete adaptation of safety verification logic to the dual-stream architecture.

[0111] The first stage is the basic parameter tuning, which uses the critical proportional gain method, attenuation curve method, or Ziegler-Nichols method to complete the initial parameter tuning. The tuning process is performed based on the fixed scan cycle of the fast current control subsystem. During the tuning process, the scan cycle is completely consistent with the actual operating cycle, thus obtaining the initial parameters. , , parameter;

[0112] The second stage involves adaptation and optimization of the dual-stream architecture. Based on the architectural characteristics of this invention, the initial parameters are optimized and adjusted. The optimization principles include: since the setpoint of the PID controller in this architecture is output by a smooth approximation algorithm without significant step changes, the derivative gain can be appropriately reduced. The value of should be chosen to avoid excessive differential action that could amplify system noise; integral gain The tuning needs to balance steady-state accuracy and resistance to integral saturation, and should be used in conjunction with an integral limiting algorithm to avoid integral saturation problems caused by long-term parameter constraints; proportional gain The tuning is centered on ensuring the system's response speed, while also ensuring that there is no oscillation or overshoot during the smooth adjustment of parameters;

[0113] The third stage is on-site verification and solidification. The calibrated parameters need to be verified by continuous on-site stable operation for more than 72 hours. The verification indicators include: the overshoot does not exceed 10% when the set value changes by a step, the adjustment time does not exceed the maximum response time allowed by the process, there is no continuous oscillation, and the steady-state error meets the process accuracy requirements. After the verification is passed, it is hard-coded into the control logic of the fast flow control subsystem.

[0114] For retrofitting existing industrial equipment, the PID parameters already stably running in the original PLC / DCS system can be directly reused without readjustment; only a smooth approximation algorithm is needed. The values ​​are fine-tuned to ensure that the control stability of the system is not affected after non-intrusive modification.

[0115] The industrial control parameter optimization method based on the neural symbolic two-stream architecture, implemented based on any one of the above-mentioned industrial control parameter optimization systems based on the neural symbolic two-stream architecture, includes the following steps:

[0116] Step 1. The slow flow cognitive subsystem acquires multimodal operating condition data from the industrial site through bypass monitoring, combines it with built-in historical work order data and process expert experience data, performs global optimization calculation of process parameters, and generates suggested values ​​for process parameters.

[0117] Step 2. The slow flow cognitive subsystem silently writes the generated process parameter suggestions into the intermediate data interaction register module. The writing action does not trigger a hardware interrupt in the fast flow control subsystem.

[0118] Step 3. Within each fixed scan cycle, the fast flow control subsystem actively retrieves the currently stored process parameter suggestion values ​​from the intermediate data interaction register module;

[0119] Step 4. Within the same scan cycle, the fast flow control subsystem completes the deterministic safety verification of the pulled process parameter suggestion values. If the verification passes, the process parameter suggestion values ​​are processed into the final control target values, substituted into the built-in closed-loop control algorithm to generate drive signals, and sent to the corresponding physical actuators to complete the control actions. If the verification fails, the safety control target value of the previous scan cycle is maintained and an abnormal alarm is triggered.

[0120] The technical concept of this invention is as follows:

[0121] The core of this invention is a neural-symbolic dual-stream architecture that decouples and restricts the probabilistic generalization reasoning (neural branch) based on deep neural networks and the deterministic physical safety control (symbolic branch) based on strict industrial control language at both the physical hardware and system logic levels, thereby constructing an asynchronous and restricted interactive safety control closed loop.

[0122] The core architecture consists of three independent but collaborative modules, with the core operational logic as follows:

[0123] Fast Stream control subsystem: Deployed within the PLC / DCS / independent motion controller hardware closely integrated with the physical equipment, it runs on an embedded real-time operating system, possessing hard real-time millisecond-level cyclic scanning capabilities. Its control logic is hard-coded according to the IEC 61131-3 international standard and is granted the final and sole physical execution control within the system, prohibiting responses to external hardware interrupt requests and externally issued drive commands. Fast Stream's internal autonomous fetching and deterministic verification logic is as follows... Figure 2 As shown, this illustrates the process branches of a PLC that process external input data using hard-coded logic within a single scan cycle of a high-speed loop.

[0124] The Slow Stream cognitive subsystem is deployed at the edge IoT gateway in the industrial field or in an industrial server with an AI acceleration chip. It runs a small language model (SLM) in an asynchronous soft real-time cycle, which is not constrained by the underlying hard real-time clock. It listens to multimodal industrial data in bypass mode, constructs full-dimensional operating condition features, builds an optimization objective function based on embedded expert experience and historical data, completes global asynchronous optimization of process parameters, and outputs suggested values ​​for target parameters.

[0125] Intermediate data interaction register module: As a unidirectional data isolation area for cross-stream communication, it is set in the non-volatile storage area inside the fast flow control subsystem or in the independent dual-port RAM of the real-time bus; it only allows slow flow to write static parameter suggestion values, and the writing action does not trigger hardware interrupts of fast flow, realizing the interception of dynamic control execution flow and unidirectional secure transmission of static data flow.

[0126] The core workflow of the system consists of four strictly decoupled steps: the slow stream completes bypass sensing and asynchronous inference to generate process optimization parameters → the slow stream silently writes the optimization parameters as suggested values ​​into the intermediate register module → the fast stream actively pulls the parameter values ​​from the register module according to its own fixed scan cycle → within a single scan cycle, the fast stream completes deterministic verification such as boundary clamping and smooth approximation of the pulled parameters. If the verification passes, the parameters are sent to the underlying PID controller to complete closed-loop drive; if the verification fails, the safe operating state of the previous cycle is maintained and an abnormal alarm is triggered.

[0127] The overall architecture logic of the neural symbolic dual-stream parallelism: This system constructs three core modules that operate independently and collaboratively. These three modules together form an asynchronous, constrained interactive control loop.

[0128] Slow Stream: Handled by the SLM on edge computing nodes. It is responsible for long-term analysis and strategy formulation of multi-dimensional complex information, and its operation is not constrained by the underlying microsecond-level clock.

[0129] Fast Stream: Handled by the underlying PLC / DCS. It maintains a stable millisecond-level cyclic scan and performs final verification and rejection logic for instructions involving physical safety.

[0130] Intermediate data interaction register module: As a one-way data isolation area for cross-stream communication, it intercepts dynamic control execution flow and only allows static data flow to pass through.

[0131] Fast Stream control subsystem:

[0132] 1. Operating Environment and Hard Real-Time Characteristics: Deployed in PLC, DCS, or standalone motion controller hardware closely attached to the physical equipment. The core is based on an embedded real-time operating system (RTOS). Its operating mechanism is a continuous infinite loop scan cycle (ScanCycle), with a cycle time of [value missing]. (generally This cycle possesses hard real-time performance, with time deviation. .

[0133] 2. Symbolic deterministic logic attributes: Its internal control logic is hard-coded in accordance with IEC 61131-3 international standards (such as ladder diagrams (LD) and structured text (ST)). The logic represents rigorously verified deterministic rules and has the control capability to handle absolute priority events such as emergency stops (E-Stop).

[0134] 3. Control and Access Control Settings: In the system architecture design, FastStream is granted final and sole physical execution control. The FastStream system is prohibited from passively listening to and executing any externally triggered hardware interrupt requests, and is prohibited from receiving drive instructions written directly to the I / O output image area by external networks or SlowStream.

[0135] Slow Stream Cognitive Subsystem:

[0136] 1. Operating Environment and Asynchronous Soft Real-Time Characteristics: Deployed in smart IoT edges near the industrial site or in industrial servers equipped with dedicated AI acceleration chips. Runs a small language model (SLM) containing billions of parameters. Its inference cycle is set to... It exhibits nonlinear fluctuation characteristics, that is Furthermore, there is significant time jitter.

[0137] 2. Multimodal perception and high-dimensional target optimization algorithm:

[0138] Slow Flow monitors multimodal data in bypass mode to construct a full-dimensional feature vector of the current operating condition. SLM utilizes its high-dimensional fitting capabilities to solve constrained optimization problems.

[0139] The objective function for slow flow optimization is set as minimizing the overall production cost. Its mathematical model is expressed as:

[0140] ;

[0141] in, For the minimum value operator, For small language model expectation operators, For cost calculation function, Set a set of parameters for the process to be optimized (such as temperature setpoint and flow rate ratio). The space for exploring processes as perceived by the model. Historical work orders and expert experience context are embedded in the SLM. Nonlinear process quality constraints must also be met. (That is, the predicted future state must meet the minimum quality requirements) ), This is a nonlinear process quality constraint function. for The state vector of operating conditions at any given time. The optimal solution output by the SLM. This refers to the generated target parameters (optimal control parameter vector). Specifically, the slow-flow cognitive subsystem uses real-time operating data within each optimization cycle. Based on the input, combined with production context data Construct a complete decision-making scenario and analyze the control parameter vector using a built-in small language model (SLM). In the feasible region The system performs a global optimization, ultimately returning the result that minimizes the expected overall production cost. Minimize the optimal control parameter vector This information is then output as a process optimization suggestion to the system interaction layer.

[0142] Intermediate data interaction and storage module:

[0143] 1. Module definition and physical location: Located in the non-volatile memory area inside the fast flow control subsystem (such as the DB data block, D register area, or recipe temporary storage area of ​​the PLC), or an independent dual-port RAM module mounted on the real-time bus.

[0144] 2. One-way asynchronous write isolation and instruction dimensionality reduction mechanism:

[0145] The operational intent of slow-flow output is converted into static numerical suggestions in this module. .

[0146] Slow flow writes to this register The operation remains silent and does not trigger the hardware interrupt signal of the fast stream processor.

[0147] A dual-stream asynchronous interactive workflow algorithm for optimizing industrial control parameters:

[0148] The optimization control process consists of four strictly decoupled steps, and its underlying mathematical and logical execution flow is as follows:

[0149] 1. Step 1 (Slow Flow Bypass Awareness and Asynchronous Inference): Without interfering with the underlying runtime, the slow flow solves the above optimization objective function to obtain the latest optimization parameters to be set. .

[0150] 2. Step 2 (Writing Slow Flow Static Values): Slow Flow will... As Write to the target address of the register.

[0151] 3. Step 3 (Active Fast Stream Extraction): In each scan period of the fast stream discrete time series k Inside, the PLC actively performs a read operation to obtain the current value in the register, and sets the retrieved value to... .

[0152] 4. Step 4 (Deterministic Verification and Smooth Execution Algorithm Based on Symbolic Logic): Fast Stream Acquisition Later, in one Complete the following algorithm derivation within the period:

[0153] (1) Boundary clamping algorithm (Clamp Function): Utilizes hard-coded secure physical regions within the fast stream. Intercepting abnormal data:

[0154] ;

[0155] For other cases, This is the target value for safety control in the k-th cycle; this formula ensures that when the AI ​​outputs an out-of-bounds value, the system maintains the safe state of the previous cycle. And trigger an abnormal status alarm.

[0156] (2) Smooth Approximation Algorithm (Ramp Function): To prevent sudden changes in the target value from causing mechanical shock to the servo, the fast flow algorithm performs step limit processing on the safe suggested value, setting the maximum allowable step size to be... :

[0157] ;

[0158] The final control target value for the k-th period, The final control target value for the (k-1)th cycle. It is a symbolic function;

[0159] (3) Closed-loop drive output (Discrete PID): The smoothed final target value Substitute it into the underlying incremental discrete PID controller to output the physical electrical signal that drives the actuator. :

[0160] ;

[0161] in The control increment for the k-th period, For proportional gain, For integral gain, For differential gain, The control error for the k-th cycle is... , This represents the current actual feedback value from the field sensors. The control error for the (k-1)th cycle is... This represents the control error for the (k-2)th cycle. The absolute quantity of the physical electrical signal actually sent to the actuator (such as valve, motor, variable frequency pump, frequency converter, robotic arm, etc.) is... It is obtained by accumulating historical increments: ,in It is the absolute control quantity that was executed in the previous cycle.

[0162] Technical effects:

[0163] 1. Constructing an isolation barrier to achieve safe convergence of AI model output: This invention changes the approach of AI directly taking over underlying control. Through the dual blocking of clamping and smooth approximation algorithms, AI is confined to the functional layer that provides optimization suggestions. Even if the large model produces unpredictable hallucinations, the underlying physical devices still rely on hard logic defenses to maintain safe operation, achieving control convergence from the probabilistic domain to the deterministic domain.

[0164] 2. Spatiotemporal decoupling resolves the conflict between computational delay and control cycle: due to and Physically and logically decoupled, the inference latency, network fluctuations, or changes in model computational load of the slow stream are all absorbed by the static registers, without blocking the millisecond-level scan cycle of the fast stream. This allows industrial systems to introduce cognitive intelligence from models with hundreds of billions of parameters without compromising their hard real-time performance and low jitter.

[0165] 3. Provides a low-cost, non-intrusive architecture for retrofitting legacy equipment: Reconstructing the core logic of legacy DCS systems in actual industrial settings carries extremely high risks. This architecture eliminates the need to modify the complex closed-loop control code within legacy PLCs; it only requires opening static data blocks (DB blocks) and providing specific addresses to the edge gateway to achieve non-intrusive access to the parameter optimization system.

[0166] Example 1. Adaptive Optimization Control of Cement Rotary Kiln Temperature in the Building Materials Industry:

[0167] 1. Scenario Description and System Constraints: In the rotary kiln calcination process of a cement plant, the computing power of the existing DCS system is limited. The process objective is to reduce gas energy consumption by having edge AI calculate the target calcination temperature based on the real-time raw material quality.

[0168] 2. Dual-stream system deployment and algorithm execution:

[0169] Slow Stream Asynchronous Inference: Edge SLM Utilizes Its Objective Function The target calcination temperature for energy consumption optimization under the current operating conditions is calculated. And silently write the kiln temperature recipe temporary register exposed by the DCS.

[0170] Fast stream active pull and defense line verification: Old DCS system in its 15ms ( During the cyclic scan, the value of the register is retrieved. Then, the hard-coded interval determination formula is substituted to determine... Whether it is valid or not.

[0171] Smooth approximation: The formula holds true, and the DCS triggers the Ramp function logic. The maximum temperature change step size is set to... The system calculates approximate values ​​step by step and sends them to the PID control block to adjust the gas valve, thereby avoiding combustion oscillations caused by setting large range setpoints all at once.

[0172] 3. Abnormal data testing mechanism: If the edge computing device malfunctions or the AI ​​model outputs the target set temperature... DCS pulls Then, it is determined that the system is outside the range, and the system triggers the defense algorithm. Maintain the control target at the safe value of the previous cycle (e.g.) ), to prevent equipment from operating beyond its limits.

[0173] Example 2. Dynamic optimization of batch process rate in high-risk reactors in the fine chemical industry:

[0174] 1. Scenario Description and System Constraints: Catalytic polymerization is exothermic. Excessive catalyst feed rate can lead to heat accumulation and potentially cause safety accidents. The process objective is to dynamically optimize the catalyst acceleration feed curve to fit the critical safety boundary using an SLM model.

[0175] 2. Differential safety interception verification of dual-flow system:

[0176] Fast Flow Differential Operator Configuration: Besides setting static parameters, the underlying PLC... In addition to the absolute temperature limit, dynamic differential protection logic has been added to its periodic scan to calculate the reaction temperature rise rate using a differential formula:

[0177] ; For continuous temperature change rate, This represents the temperature sampling value for the kth period. The temperature sample value for the k-Nth period is... This is the length of the sliding window; if the calculated temperature rise rate exceeds the safety limit... If so, the system is determined to be in a high-risk state.

[0178] Slow Flow Calculation Output: The SLM model outputs the target feed rate setpoint for increasing production capacity. And write it to the register.

[0179] Calculation of rigid constraints for fast flow: The PLC acquires this value at a frequency of 5ms. Although the actual reactor temperature is at this time... Below The upper limit is set, but the PLC calculates using the aforementioned differential formula that the recent temperature rise rate has exceeded the threshold.

[0180] Degradation clamping action: The PLC's symbolic execution engine triggers the clamping algorithm (Clamp): It refuses to apply the given value 15.0, forces the execution of the minimum guarantee calculation, and outputs... And send the clamping value to the variable frequency pump; This is the critical safety threshold (9.5 L / min in this example).

[0181] Results evaluation: This scenario demonstrates that the system, based on symbolic logic, not only has the ability to intercept static numerical values, but also can mitigate the potential lag risk of AI output through a dynamic differential model.

[0182] Key technical points of this invention:

[0183] 1. Decoupling of the physical and mathematical aspects of the time axis: dividing the industrial control system into sections that are not subject to strict periodic constraints and can handle nonlinear optimization calculations. The slow-flow cognitive engine, driven by a fixed crystal oscillator beat It also employs a fast-flow control engine that executes discrete PID closed-loop equations.

[0184] 2. Register-based static dimensionality reduction data transfer mechanism: This mechanism uses a low-level non-volatile memory area as an airlock bridge. External intervention instructions to the physical world are converted into a set of static constants that cannot trigger hardware interrupts. .

[0185] 3. Discrete safety clamping operation under control transfer: Fastflow hardware actively polls and reads data based on its own cyclic rhythm, and forcibly embeds a mathematical clamping algorithm based on absolute envelope and a rate of change differential blocking model to build a formal computational defense line at the bottom layer.

[0186] Two-stream time-scale decoupled architecture: includes runtime domain as Furthermore, the slow-flow cognitive subsystem that executes the target nonlinear optimization model has a runtime in the hard real-time microsecond range. Furthermore, the fast-flow control subsystem executes the discrete physical regulation equations, and the intermediate data register module serves as the communication isolation interface; its system characteristics are as follows: the slow-flow cognitive subsystem is prohibited from issuing action interruption instructions, and only has the authority to write the expected value of static process parameters to the specified address of the register module.

[0187] Active fetching and mathematical clamping interception method: The fast flow control subsystem does not respond to external interrupt signals and autonomously fetches registered data according to the hardware cyclic scan cycle; it generates physical drive output. Previously, the system was forced to perform boundary clamping operations and differential rate of change verification based on symbolic logic within the current scan cycle. If the operation result exceeded the system's preset safe mathematical envelope, the system would execute a historical data retention or clamping rollback procedure to ensure that the drive device remained in the historical safe state domain.

[0188] Edge optimization algorithm application: The slow flow system listens to the waveforms of heterogeneous time-series sensors and the natural language process specification text in a bypass manner; based on the fused feature vector, an optimization objective function containing process constraint constants is constructed, and the continuous operating parameters (such as the follow-up temperature control curve) that minimize the objective function value are asynchronously reconstructed and derived in the computational domain, thereby isolating the influence of complex calculations on the core control loop.

[0189] Non-intrusive control architecture topology: Without modifying the core closed-loop regulation control program code of the underlying controller, the industrial control communication protocol is used to map or allocate specific data blocks of its internal shared memory to the external computing gateway as unidirectional data write addresses. Through the scanning and reading mechanism of the original controller, the external computing parameters are smoothly integrated into the traditional execution process.

Claims

1. An industrial control parameter optimization system based on a neural-symbolic dual-flow architecture, characterized in that, It includes a slow-flow cognitive subsystem, a fast-flow control subsystem, and an intermediate data interaction and storage module; The slow-flow cognitive subsystem is deployed in edge computing devices in industrial settings. It has a built-in small language model finely tuned for the industrial vertical domain. It is used to acquire multimodal operating condition data of the industrial site through bypass listening. Combined with built-in historical work order data and process expert experience data, it performs global optimization calculation of process parameters with the goal of minimizing the overall production cost and generates suggested values ​​for process parameters. The slow-flow cognitive subsystem only has the permission to write suggested values ​​for process parameters to the intermediate data interaction register module. It is prohibited from issuing hardware interrupt commands and direct drive commands for physical actuators. The intermediate data interaction register module is located in the accessible storage area of ​​the fast flow control subsystem, serving as a one-way communication isolation interface between the slow flow cognition subsystem and the fast flow control subsystem. It is used to receive and store the process parameter suggestion values ​​written by the slow flow cognition subsystem, and the writing action of the slow flow cognition subsystem to the intermediate data interaction register module does not trigger a hardware interrupt of the fast flow control subsystem. The fast-flow control subsystem is deployed in industrial controller hardware closely attached to the physical actuators in the industrial field. It runs on an embedded real-time operating system and features hard real-time fixed-cycle cyclic scanning. It is granted unique control over the physical actuators within the system and is prohibited from responding to external hardware interrupt requests or drive commands directly issued by external systems. Within each fixed scan cycle, the fast-flow control subsystem actively retrieves currently stored process parameter suggestions from the intermediate data interaction register module. Within the same scan cycle, it performs a deterministic safety check on the retrieved process parameter suggestions. If the check passes, the process parameter suggestions are processed into the final control target value. This final control target value is then substituted into the built-in closed-loop control algorithm to generate a drive signal, which is sent to the corresponding physical actuator to complete the control action. If the check fails, the safe control target value from the previous scan cycle is maintained, and an anomaly alarm is triggered.

2. The industrial control parameter optimization system based on the neural-symbolic dual-flow architecture according to claim 1, wherein, The multimodal operating condition data acquired by the slow flow cognitive subsystem includes time-series operating condition data collected by industrial field sensors, process procedure text data, equipment maintenance record data, and experience record data of on-site operators. When the slow-flow cognitive subsystem performs global optimization calculations of process parameters, it simultaneously satisfies the preset process quality constraints.

3. The industrial control parameter optimization system based on the neural-symbolic dual-flow architecture according to claim 1, wherein, The intermediate data interaction register module adopts a non-volatile data storage area inside the industrial controller, or an independent dual-port RAM module mounted on the industrial real-time bus; the intermediate data interaction register module only grants one-way write permission to the specified storage address to the slow flow cognitive subsystem, and only grants read permission to the corresponding storage address to the fast flow control subsystem.

4. The industrial control parameter optimization system based on the neural-symbolic dual-flow architecture according to claim 1, wherein, The industrial controller used in the fast flow control subsystem includes a programmable logic controller, a distributed control system controller, or an independent motion controller; the fixed cyclic scanning period of the fast flow control subsystem is 1 millisecond to 20 milliseconds, the time deviation of the scanning period is controlled at the microsecond level, and its internal control logic is hard-coded in accordance with international standards.

5. The industrial control parameter optimization system based on the neural-symbolic dual-flow architecture according to claim 1, wherein, The deterministic safety checks performed by the fast flow control subsystem on the recommended values ​​of process parameters include static boundary interval checks and dynamic rate of change checks. The static boundary interval checks are used to determine whether the recommended values ​​of process parameters are within a preset safe physical range, and the dynamic rate of change checks are used to determine whether the parameter change rate corresponding to the recommended values ​​of process parameters is within a preset safe rate of change range.

6. The industrial control parameter optimization system based on the neural-symbolic dual-flow architecture according to claim 5, characterized in that, When the fast flow control subsystem performs static boundary interval verification and dynamic rate of change verification, if any one of the verifications fails, the currently retrieved process parameter suggestion value is determined to be invalid. The system maintains the safety control target value of the previous scan cycle, clears the invalid process parameter suggestion value in the intermediate data interaction register module, and triggers the corresponding type of abnormal alarm. If both checks pass, the recommended process parameter values ​​are deemed valid. Smoothing approximation processing is then performed on the valid recommended process parameter values ​​to generate the final control target values.

7. The industrial control parameter optimization system based on a neural symbolic dual-stream architecture according to claim 1, characterized in that, The slow-flow cognitive subsystem cannot directly access the input / output image area of ​​the fast-flow control subsystem and cannot directly issue any drive commands to the physical actuator; the fast-flow control subsystem only executes the drive signals generated based on the final control target value within its own scan cycle and does not execute any drive commands directly issued by external systems.

8. The industrial control parameter optimization system based on a neural symbolic dual-stream architecture according to claim 1, characterized in that, When the system is deployed in an existing industrial control system, there is no need to modify the core closed-loop regulation and control program code built into the fast flow control subsystem. Only the corresponding storage address needs to be allocated to the intermediate data interaction register module, and the one-way write permission of the corresponding storage address needs to be opened to the slow flow cognition subsystem to complete the non-intrusive deployment of the system.

9. The industrial control parameter optimization system based on the neural-symbolic dual-flow architecture according to claim 1, wherein, The operating cycle of the slow-flow cognitive subsystem is not constrained by the fixed scanning cycle of the fast-flow control subsystem. The inference delay and computational load fluctuations of the slow-flow cognitive subsystem are buffered through the intermediate data interaction register module, which does not interfere with the fixed scanning cycle of the fast-flow control subsystem.

10. An industrial control parameter optimization method based on a neural symbolic dual-stream architecture, characterized in that, The industrial control parameter optimization system based on the neural symbolic dual-stream architecture as described in any one of claims 1 to 9 is implemented, and the industrial control parameter optimization method based on the neural symbolic dual-stream architecture includes: Step 1. The slow flow cognitive subsystem acquires multimodal operating condition data from the industrial site through bypass monitoring, combines it with built-in historical work order data and process expert experience data, performs global optimization calculation of process parameters, and generates suggested values ​​for process parameters. Step 2. The slow flow cognitive subsystem silently writes the generated process parameter suggestions into the intermediate data interaction register module. The writing action does not trigger a hardware interrupt in the fast flow control subsystem. Step 3. Within each fixed scan cycle, the fast flow control subsystem actively retrieves the currently stored process parameter suggestion values ​​from the intermediate data interaction register module; Step 4. Within the same scan cycle, the fast flow control subsystem completes the deterministic safety verification of the pulled process parameter suggestion values. If the verification passes, the process parameter suggestion values ​​are processed into the final control target values, substituted into the built-in closed-loop control algorithm to generate drive signals, and sent to the corresponding physical actuators to complete the control actions. If the verification fails, the safety control target value of the previous scan cycle is maintained and an abnormal alarm is triggered.

Images