An edge gateway risk constraint cloud edge scheduling closed-loop control method and system

Abstract

The application discloses an edge gateway risk constraint cloud edge scheduling closed-loop control method and system. The edge gateway collects time sequence data of the controlled equipment, perceives CPU occupancy rate, NPU occupancy rate, network bandwidth and residual control time limit, calculates a joint scheduling index in combination with a fault risk value and a risk elimination effective degree of the last period, and selects a mode in local lightweight reasoning, double reasoning and cloud auxiliary reasoning; cloud reasoning alone is prohibited when the risk is high or the time limit is insufficient, and a local conservative action is executed when the time is out. After the control suggestion is mapped through process coupling constraints and equipment operation envelope, it is written into a PLC or a frequency converter and read back to confirm, and then the next period priority, sampling rate and unloading threshold are updated with the elimination effective degree. The application can complete safe closed-loop intervention within the control time limit under edge overload and network jitter.

Description

Technical Field

[0001] This invention relates to the fields of industrial Internet of Things and industrial control technology, specifically to an edge gateway risk-constrained cloud-edge scheduling closed-loop control method and system, used for predictive maintenance and local intervention of equipment in factory production lines, and is particularly suitable for rotating machinery and thermal equipment such as fans, pumps, and heating furnaces. Background Technology

[0002] Industrial sites commonly use sensors to collect time-series data such as vibration, current, and temperature, and then use edge gateways or cloud models for fault prediction. Existing solutions have the following main shortcomings.

[0003] First, cloud-edge collaborative inference scheduling typically optimizes based on computing load, network bandwidth, or service priority, without taking device failure risk levels and allowable intervention deadlines as hard constraints. When edge nodes are overloaded, high-risk tasks may still be offloaded to the cloud; if network jitter occurs, inference results cannot be returned within the control time limit, and on-site intervention can only issue an alarm but cannot be completed.

[0004] Secondly, predictive maintenance results often stop at alarms or work orders. Even if the inference results are written as control quantities and sent to the PLC or frequency converter, they are mostly written in an open loop. There is a lack of back-read verification of the actual set value or operating status within the confirmation time limit, which poses a risk that the command is not effective but is regarded as successful.

[0005] Third, performing the highest-risk action on a single piece of equipment (such as directly stopping the pump) could disrupt the production line's process coupling, leading to secondary exceedances such as overheating of related thermal equipment. Furthermore, the output of artificial intelligence is probabilistic; directly overriding PLC safety interlocks would not meet the deterministic safety requirements of industrial control.

[0006] Fourth, existing closed-loop feedback is mostly used for model parameter updates or general resource reallocation. It does not quantify the physical resolution effect after control execution into the effectiveness of risk resolution, and adjust the inference task priority, sampling rate and cloud-edge offloading threshold for the next control cycle accordingly. Therefore, it cannot automatically encrypt monitoring and tighten local inference when intervention is ineffective.

[0007] Therefore, there is a need for a closed-loop intervention method and system that can ensure high-risk devices complete safe, verifiable, and non-disruptive process coupling within the control time limit under conditions of edge overload and network jitter. Summary of the Invention

[0008] The technical problem to be solved by the present invention is: at industrial sites where the edge gateway has limited computing power and network bandwidth fluctuates, how to adaptively select local lightweight inference or cloud-assisted inference according to the fault risk, the remaining control time limit and the control effect of the previous cycle, map the inference result to executable instructions through process coupling constraints and equipment operation envelopes, complete the safety closed-loop control of the PLC or frequency converter through read-after-write, and update the next round of scheduling parameters with the risk resolution effectiveness at the same time.

[0009] To solve the above technical problem, the present invention adopts the following technical solution.

[0010] In a first aspect, an edge gateway risk-constrained cloud-edge scheduling closed-loop control method is provided, comprising: an edge gateway collects multi-source time series data of controlled equipment, and collects the CPU occupancy, NPU occupancy, available network bandwidth and remaining control time limit T_left; obtaining a fault risk value R according to the time series data, and calculating a joint scheduling index J in combination with the risk resolution effectiveness η of the previous cycle; selecting the current mode from the three modes of local lightweight inference, local and cloud dual inference, and cloud-assisted inference, and prohibiting the use of only cloud-assisted inference when R≥R_crit or T_left<T_min; in the dual inference mode, the local side outputs a conservative control suggestion first, and the cloud result is discarded if it is overdue; mapping the control suggestion to a target instruction in the allowable action set restricted by process coupling constraints and equipment operation envelopes; after writing to the PLC or frequency converter, performing read-back within the time limit τ, and retrying or rolling back according to the fail-safe policy if inconsistency occurs; after the instruction takes effect, calculating η, and updating the inference task priority, sampling rate and cloud-edge offloading threshold for the next cycle

[0011] Further, the joint scheduling index J=w1·L_cpu+w2·L_npu+w3·(1-B / B_max)+w4·R+w5·(1-η)+w6·f(T_left). When the estimated cloud round-trip delay is greater than T_left, cloud-assisted inference is not selected even if the bandwidth is sufficient. The dual inference cut-off duration takes the smaller value of T_left-T_min and the cloud timeout upper limit, and the conservative action gives priority to load reduction rather than shutdown.

[0012] Further, the risk resolution effectiveness η=1-||x_after-x_safe|| / (||x_before-x_safe||+ε). If η is lower than the threshold, the priority of the device is increased, the sampling period is shortened, and the offloading threshold is tightened toward the local side. Scheduling, envelope verification and read-after-write are bound to the CPU real-time core, neural network inference is bound to the NPU, and data is exchanged through a lock-free ring buffer.

[0013] Furthermore, when the circulating pump and the heating furnace are process-coupled and the furnace temperature slope has reached the upper limit of the envelope, the pump's permissible action set excludes safe shutdown, retaining only load reduction to the lower frequency limit, and prohibiting any additional load on the furnace. Write-back is performed using Modbus TCP, OPC UA, or PROFINET, with commands bearing unique identifiers; if continuous readback fails, the system switches to a safe setting, shutting down only in cases of critical risk and when interlocking permits.

[0014] Secondly, a corresponding system is provided, including a field device layer, an edge gateway, and a cloud inference service; the edge gateway includes a time-series acquisition module, a resource awareness module, a joint scheduler, a local lightweight inference engine, a cloud collaboration interface, a security action mapping module, a protocol writing and reading module, and a resolution effectiveness evaluation module.

[0015] Thirdly, an edge computing gateway is provided, including a memory, CPU, NPU, and industrial fieldbus interface, for performing the above methods.

[0016] The beneficial effects of this invention are as follows:

[0017] (1) Write the fault risk, remaining control time limit and risk mitigation effectiveness into the scheduling indicators, and set a hard constraint that prohibits cloud-based inference only for high-risk / short-time limits, so as to avoid the critical intervention window being eaten up by network latency.

[0018] (2) Dual inference adopts local conservative actions first and cloud results can be upgraded but expired, achieving an feasible compromise between accuracy and determinism.

[0019] (3) The target instruction must fall within the equipment operation envelope and be subject to process coupling constraints to avoid the optimal action of a single device causing related equipment to exceed the limit.

[0020] (4) The write-back and failure safety strategy changes the closed loop from "issuing instructions" to "confirming that the physical side has responded", reducing open loop malfunctions.

[0021] (5) Use η to feed back the priority, sampling rate and unloading threshold of the next cycle, so that the equipment that is ineffective in intervention can automatically obtain higher monitoring density and local computing power, forming a complete technical closed loop of acquisition-scheduling-inference-control-feedback.

[0022] (6) The CPU real-time core is isolated from the NPU computing power to avoid the inference load crowding out the control loop and improve the real-time performance inside the edge gateway. Attached Figure Description

[0023] To more clearly illustrate the embodiments of the present invention, the accompanying drawings will be briefly described below.

[0024] Figure 1 A schematic diagram of the overall structure of the system of this invention.

[0025] Figure 2 Diagram showing the internal modules and computing power isolation of the edge gateway.

[0026] Figure 3 A schematic diagram of the method flow of this invention.

[0027] Figure 4 Schematic diagram of reasoning mode selection decision.

[0028] Figure 5 Target instruction write and readback timing diagram.

[0029] Figure 6 Schematic diagram of process coupling constraints between circulating pump and heating furnace.

[0030] Figure 7 A schematic diagram illustrating how the effectiveness of risk mitigation feeds back into the scheduling process.

[0031] Figure 8 A schematic diagram of the method steps of this invention.

[0032] In the diagram: 1 - Field Equipment Layer; 11 - Fan; 12 - Pump Body; 13 - Heating Furnace; 14 - PLC; 15 - Frequency Converter; 2 - Edge Gateway; 21 - Timing Acquisition Module; 22 - Resource Awareness Module; 23 - Joint Scheduler; 24 - Local Lightweight Inference Engine; 25 - Cloud Collaboration Interface; 26 - Safety Action Mapping Module; 27 - Protocol Writing and Readback Module; 28 - Resolution Effectiveness Evaluation Module; 29 - CPU Real-Time Core; 210 - NPU; 211 - Lockless Circular Buffer; 3 - Cloud Inference Service; 4 - Process Coupling Constraint Library; 5 - Equipment Operation Envelope Library. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the invention. Those skilled in the art will understand that conventional substitutions of details made without departing from the principles of the present invention are all within the scope of protection of the present invention.

[0034] Example 1: System Composition.

[0035] See Figure 1 and Figure 2 The system of this invention includes a field device layer 1, an edge gateway 2, and a cloud inference service 3. The field device layer 1 includes at least one type of controlled device selected from fans 11, pumps 12, and furnaces 13, as well as a PLC 14 and / or a frequency converter 15 connected thereto. Sensor signals enter the edge gateway 2 via Modbus RTU / TCP, OPC UA, or PROFINET.

[0036] Edge gateway 2 includes a timing acquisition module 21, a resource awareness module 22, a joint scheduler 23, a local lightweight inference engine 24, a cloud collaboration interface 25, a security action mapping module 26, a protocol write and readback module 27, and a resolution effectiveness evaluation module 28. The joint scheduler 23, the security action mapping module 26, and the protocol write and readback module 27 run on the CPU real-time core 29; the local lightweight inference engine 24 runs on the NPU 210; the two exchange fault risk values ​​R and control recommendations through a lock-free ring buffer. Edge gateway 2 also connects to the process coupling constraint library 4 and the device operation envelope library 5.

[0037] The cloud-based inference service 3 deploys more accurate time-series fault prediction models, receiving feature or compressed samples and returning risk estimates and recommended actions only under the mode permitted by the joint scheduler. Cloud results must arrive before the deadline to replace local conservative recommendations; otherwise, they are discarded to ensure the determinism of the control loop.

[0038] Example 2: Methods and Parameters

[0039] See Figure 3 , Figure 4 and Figure 8 The method operates according to the control cycle; Figure 8 The overall process from steps one to seven is shown. At the beginning of each cycle, the timing acquisition module 21 reads the multi-source timing data of the controlled device; the resource awareness module 22 reads the CPU utilization rate, NPU utilization rate, available bandwidth B, and the remaining control time limit T_left of the device. T_left is obtained by subtracting the acquisition and preprocessing time consumed in this cycle from the maximum intervention delay allowed by the process.

[0040] The input to the local lightweight model is the temporal features within a sliding window, such as the effective value of vibration, peak frequency of the spectrum, current imbalance, temperature, and their first-order differences; the output is the normalized fault risk value R∈[0,1] and the confidence level. The model can be trained offline in the cloud using historical fault samples and then quantized and deployed on the NPU. This invention does not limit the specific network structure, as long as the input and output meet the above conventions.

[0041] The joint scheduling index is calculated using the following formula: J = w1·L_cpu + w2·L_npu + w3·(1-B / B_max) + w4·R + w5·(1-η) + w6·f(T_left). Where L_cpu and L_npu are normalized occupancy rates; η is the resolution effectiveness of the previous cycle, which can be 1 for the first cycle; f(T_left) = T_min / (T_left + ε_t), where a smaller T_left indicates greater urgency. A set of implementable weights is w1 = 0.15, w2 = 0.15, w3 = 0.10, w4 = 0.25, w5 = 0.20, w6 = 0.15, which can be calibrated on-site.

[0042] The mode selection rules are as follows. If R≥R_crit or T_left<T_min, cloud-only assisted inference is prohibited, R_crit is preferably 0.85, T_min for rotating machinery is preferably 80ms, and T_min for thermal process equipment is preferably 500ms. If the estimated cloud round-trip latency is greater than T_left, cloud assistance will not be selected even if the bandwidth is sufficient. In other cases: select local lightweight inference when J is high or the bandwidth is insufficient; cloud-assisted inference is optional for medium risk and sufficient bandwidth; select dual inference when between the two cases. The cut-off duration for dual inference is min(T_left-T_min, T_cloud_max), and T_cloud_max is preferably 150ms. The local conservative suggestion is to prioritize load reduction rather than shutdown in the set of allowed actions.

[0043] See Figure 6 , the safety action mapping module 26 reads the process coupling constraint library 4 and the equipment operating envelope library 5. The equipment operating envelope at least includes the upper and lower limits of the frequency of the frequency converter, the upper limit of the load change rate, and the upper limit of the thermal temperature slope. If the control suggestion exceeds the envelope, the next-level conservative action is selected instead. The process coupling constraints are represented by an equipment dependency graph: if an action will cause the associated equipment to further exceed the limit, the action is deleted from the set of allowed actions.

[0044] See Figure 5 , the protocol writing and readback module 27 assigns a unique instruction identifier to each target instruction, and writes through ModbusTCP to hold registers, or through OPC UA to nodes, or through PROFINET to process data. The confirmation time limit τ is preferably 40ms. It is regarded as successful if the deviation between the readback value and the target value falls within the preset tolerance. When consecutive failures reach a preset number of times (for example, 3 times), the frequency converter or PLC is switched to a preset safety set value and an alarm is reported; safety shutdown shall be executed only when R≥R_crit and the PLC safety interlock allows, and artificial intelligence suggestions shall not bypass the interlock.

[0045] See Figure 7 , after the instruction is confirmed to take effect, an evaluation window is entered (preferably 2s for rotating machinery, and 10s for thermal processes). A state vector x is constructed, whose components include the vibration effective value, current unbalance, and temperature change rate. η=1-||x_after-x_safe|| / (||x_before-x_safe||+ε). If η<η_low (preferably 0.3), the inference priority of the equipment is increased by at least one level, the sampling period is changed from 100ms to 20ms, and the cloud-edge offloading threshold is tightened toward the local side, making it more difficult to select cloud assistance in the next cycle. If η is relatively high, the inference frequency can be reduced, and cloud assistance can be re-enabled when the time limit allows, so as to avoid occupying the NPU for a long time.

[0046] Embodiment 3 Fan predictive maintenance and local load reduction

[0047] For the workshop induced draft fan, bearing housing vibration and motor current were collected. In a certain period, the local model output R = 0.78, T_left = 120ms, NPU utilization 55%, and available bandwidth 2Mbps. Because R did not reach 0.85 but the time limit was short, the joint scheduler selected dual inference: the local system immediately provided a conservative suggestion to reduce the inverter's given frequency from 45Hz to 38Hz and wrote it; if the cloud returned a response within min(120-80,150) = 40ms with higher confidence, the frequency could be fine-tuned within the envelope; otherwise, it remained at 38Hz. The frequency was read back as 38.0Hz within 40ms, confirming success. After 2 seconds, the effective vibration value decreased, η = 0.62 > 0.3, allowing 100ms of sampling to resume in the next period, and cloud assistance was enabled when bandwidth was sufficient. This embodiment illustrates that sensor input, risk-constrained scheduling, inference, inverter physical output, and effect feedback constitute a complete closed loop, and within high-risk windows, it does not rely on guaranteed cloud response.

[0048] Example 4: Coupling Constraints Between Circulating Pump and Heating Furnace

[0049] The circulating pump R = 0.91, reaching the critical threshold, and should be shut down according to single-device logic. The process coupling constraint library records that the furnace cycle depends on this pump, and the current furnace temperature slope is close to the upper envelope limit. Therefore, the set of allowed actions excludes safe shutdowns, retaining only load reduction to the lower frequency limit, and simultaneously prohibiting the furnace fuel valve from being loaded. The protocol write and readback modules respectively read back the pump frequency and fuel valve position; only when both are within tolerance is the cycle considered closed-loop successful. If η is still below 0.3, the next cycle forces the pump to have the highest local priority and shortest sampling period, instead of offloading the task back to the cloud to wait. This embodiment illustrates the difference between this invention and the existing "AI direct shutdown" approach: closed-loop actions must be performed through coupling constraints and interlocking, rather than using probabilistic outputs to cover deterministic safety.

[0050] Example 5: Computational Power Isolation and Failure Security

[0051] When the CPU real-time core utilization exceeds 70%, the joint scheduler will no longer open new dual inference sessions, but will only maintain tasks already in transit to ensure that write-back can still be completed within τ. When the NPU is overloaded, priority will be given to ensuring the local lightweight model of devices with higher R values, while the inference frequency of low-risk devices will be reduced. In the event of a network interruption, the cloud-based collaborative interface will mark all in-transit requests as timed out, and the on-site operations will all fall within the local conservative actions and envelope constraints. In the event of consecutive readback failures, a safe frequency will be switched and an alarm will be triggered to avoid repeatedly writing shutdown commands under communication failure conditions.

[0052] It should be noted that the fan, pump, and heating furnace are only typical examples of the controlled equipment. This invention can be implemented in any industrial equipment that has sensor timing input, permissible intervention time limits, and load that can be changed via PLC or frequency converter. The specific values ​​mentioned above are intended to enable those skilled in the art to implement this invention and can be calibrated based on equipment inertia and process safety margins; they are not intended to limit the scope of protection.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A closed-loop control method for cloud-edge scheduling under risk constraints of edge gateways, characterized in that, comprising the following steps: In S1, the edge gateway collects multi-source time series data of at least one controlled device through an industrial fieldbus protocol, and collects the CPU occupancy, NPU occupancy, available network bandwidth of the edge gateway, and the remaining control time limit T_left of the controlled device from the allowed intervention deadline; the controlled devices include at least one of rotary mechanical equipment and thermal process equipment; In S2, the current fault risk value R is obtained according to the multi-source time series data, and combined with the risk resolution effectiveness η of the previous control cycle, the joint scheduling index J is determined jointly by the normalized CPU load, normalized NPU load, bandwidth sufficiency, the fault risk value R, the risk resolution effectiveness η, and the time limit urgency; In S3, the current inference mode is selected from a preset inference mode set according to the joint scheduling index J, the fault risk value R and the remaining control time limit T_left; the preset inference mode set includes a local lightweight inference mode, a local and cloud dual inference mode, and a cloud-assisted inference mode; when R≥R_crit or T_left<T_min, selecting only the cloud-assisted inference mode is prohibited; where R_crit is a preset critical risk threshold, and T_min is a preset minimum control time limit; In S4, when the local and cloud dual inference mode is selected, the local lightweight model deployed on the edge gateway first outputs a conservative control suggestion and starts a deadline timer; if the cloud inference result is returned before the deadline and its confidence is higher than the confidence of the local lightweight model, the cloud inference result is used to replace the conservative control suggestion, otherwise the cloud inference result is discarded and the conservative control suggestion is maintained; In S5, the current control suggestion is mapped to a target instruction in the allowed action set according to process coupling constraints and equipment operation envelopes; the allowed action set includes at least two of maintaining operation, load reduction and safe shutdown; the process coupling constraint is used to prohibit performing an action on the controlled device that will cause the associated device to further exceed the limit when the associated device is in an over-limit state; In S6, the target instruction is written into the corresponding register or node of the PLC or frequency converter through the field protocol, and the actual set value or operation state is read back within the confirmation time limit τ; if the readback result is inconsistent with the target instruction, retry or rollback is performed according to the fail-safe strategy; In S7, within the evaluation window after the target instruction takes effect, the risk resolution effectiveness η of this cycle is calculated according to the sensor feedback, and the η is used to update the inference task priority, sampling rate and cloud-edge offloading threshold of the controlled device in the next control cycle.

2. The method according to claim 1, characterized in that, The joint scheduling index J in step S2 is calculated in the following manner: J=w1·L_cpu+w2·L_npu+w3·(1-B / B_max)+w4·R+w5·(1-η)+w6·f(T_left) Where L_cpu and L_npu are the normalized CPU utilization and NPU utilization, respectively, B is the currently available network bandwidth, B_max is the rated bandwidth, f(T_left) is the time urgency function that increases as T_left decreases, and w1 to w6 are all positive weight coefficients; when the expected round-trip time of cloud-assisted inference is greater than T_left, the cloud-assisted inference mode will not be selected even if the available network bandwidth is higher than the preset bandwidth threshold.

3. The method according to claim 1, characterized in that, The duration of the cutoff timer in step S4 is the smaller of the difference between T_left and T_min and the preset cloud timeout limit; the conservative control suggestion is to prioritize load reduction rather than safe shutdown in the set of allowed actions.

4. The method according to claim 1, characterized in that, The risk mitigation effectiveness η mentioned in step S7 is calculated as follows: η=1-||x_after-x_safe|| / (||x_before-x_safe||+ε) Where x_before and x_after are state vectors composed of vibration effective value, current imbalance degree and temperature change rate before the instruction takes effect and at the end of the evaluation window, respectively; x_safe is the preset safe state vector; ε is a positive number to prevent division by zero; if η < η_low, the priority of the inference task of the controlled device is increased by at least one level, the sampling period is shortened, and the cloud-edge offloading threshold is tightened to the local side; where η_low is the preset low effective threshold.

5. The method according to claim 1, characterized in that, The edge gateway binds the scheduling logic, process coupling constraint verification, and write-back read-back of steps S3 to S6 to the CPU real-time core for execution, and binds the neural network inference to the NPU for execution. The two exchange fault risk values ​​and control suggestions through a lock-free ring buffer. The equipment operation envelope includes at least the upper and lower limits of the inverter frequency, the upper limit of the load change rate, and the upper limit of the temperature slope of the thermal equipment. The target command must fall within the equipment operation envelope; otherwise, the next most conservative action in the set of allowed actions will be selected.

6. The method according to claim 1, characterized in that, When the controlled device is a circulating pump and its associated device is a heating furnace, if the temperature change rate of the heating furnace has reached the upper limit of the temperature slope of the operating envelope of the device, then the safe shutdown is removed from the set of allowed actions for the circulating pump, and only the load reduction action of reducing the frequency to the lower limit is retained, and the load increase action of the heating furnace is prohibited at the same time; when the controlled device is a rotating machinery, the multi-source timing data includes at least vibration signal and motor current signal, and the load reduction specifically means reducing the given frequency of the frequency converter.

7. The method according to claim 1, characterized in that, The field protocol mentioned in step S6 is one of Modbus TCP, OPC UA, or PROFINET; when writing, a unique instruction identifier is assigned to the target instruction, and the readback confirmation is successful if the deviation between the register value or node value and the target instruction falls within the preset tolerance; if the readback fails for a preset number of times within the time limit τ, the frequency converter or PLC is switched to the preset safety setting value and an alarm is reported, and a safety shutdown is only performed when R≥R_crit and the safety interlock allows.

8. A risk-constrained cloud-edge scheduling closed-loop control system for edge gateways, characterized in that, This includes the field device layer, edge gateway, and cloud inference services; The field device layer includes at least one controlled device, and a PLC and / or a frequency converter connected to the controlled device; The edge gateway comprises: a time sequence acquisition module, configured to acquire multi-source time sequence data of the controlled device through an industrial field bus protocol; a resource awareness module, configured to acquire a CPU occupancy rate, an NPU occupancy rate, an available network bandwidth and a remaining control time limit T_left; a joint scheduler, configured to determine a joint scheduling index J according to a fault risk value R, a risk resolution effectiveness η as well as computing power, bandwidth and time limit, select a current inference mode from a local lightweight inference mode, a local and cloud dual inference mode and a cloud-assisted inference mode, and prohibit the use of only the cloud-assisted inference mode when R≥R_crit or T_left<T_min; a local lightweight inference engine, configured to output a fault risk value and a conservative control suggestion locally at the edge gateway; a cloud collaboration interface, configured to submit an auxiliary inference request to the cloud inference service and receive a cloud inference result; a safety action mapping module, configured to map the control suggestion to a target instruction in an allowed action set according to process coupling constraints and a device operation envelope; a protocol writing and readback module, configured to write the target instruction into the PLC or the frequency converter, read back an actual set value or an operation state within a confirmation time limit τ, and retry or roll back according to a fail-safe strategy when inconsistency occurs; a resolution effectiveness assessment module, configured to calculate the risk resolution effectiveness η within an assessment window after the instruction takes effect, and update the inference task priority, sampling rate and cloud-edge offloading threshold of the next control cycle.

9. The system according to claim 8, characterized in that, The edge gateway further comprises a CPU real-time core and an NPU, wherein the joint scheduler, the safety action mapping module and the protocol writing and readback module run on the CPU real-time core, and the local lightweight inference engine runs on the NPU; The edge gateway is further connected with a process coupling constraint library and a device operation envelope library, wherein the process coupling constraint library stores the process dependency relationship between the controlled device and associated devices, and the device operation envelope library stores the allowable ranges of frequency, load change rate and temperature slope.

10. An edge computing gateway, characterized in that, Comprising a memory, a CPU, an NPU and an industrial field bus interface; wherein the memory stores a computer program, and the CPU and the NPU execute the computer program to implement the method according to any one of claims 1 to 7.

Images