An artificial intelligence-based assay pipeline dynamic scheduling optimization method, device and computer program product

Abstract

The application discloses a kind of based on artificial intelligence's verification pipeline dynamic scheduling optimization method, device and computer program product, method includes: acquisition key station multi-source operating data and pre-processing;Through attention mechanism bidirectional long short-term memory network extraction degradation feature, fusion data-driven mode matching and physical stress-life curve calculation equipment comprehensive health degree index, below threshold when generating early warning;Real-time construction dynamic resource constraint matrix;Build heterogeneous graph neural network, to matrix and task attribute as input extraction global state embedding vector;By double intelligent agent reinforcement learning model under the guidance of reward function containing production efficiency, load balancing penalty and health incentive, collaborative generation task allocation and path planning instruction and execute.The application forms closed loop with equipment health real-time perception and adaptive scheduling, realizes the change from passive maintenance to active protection, effectively reduces unplanned downtime, improves pipeline continuous operation ability and comprehensive efficiency.

Description

Technical Field

[0001] This invention relates to the field of automated verification technology for intelligent metering equipment, specifically to a dynamic scheduling optimization method, device, and computer program product for verification production lines based on artificial intelligence. Background Technology

[0002] With the comprehensive advancement of smart grid construction, the demand for verification of smart metering equipment such as smart meters and instrument transformers has exploded. Automated verification lines have become the core equipment for power metering institutions to ensure verification efficiency. Currently, they generally adopt a centralized control architecture based on PLC, achieving large-scale automated process verification based on fixed production cycles.

[0003] However, when dealing with complex tasks such as multi-variety variable batch production, emergency order insertion, and performance degradation such as mechanical wear and electrical characteristic drift caused by long-term equipment operation, the above-mentioned fixed cycle time and static control mode gradually exposed the following technical limitations: In terms of equipment maintenance, existing methods are mostly reactive repairs or periodic maintenance, lacking the ability to monitor equipment health status in real time and predict performance degradation trends. Unplanned downtime is frequent, resulting in low overall efficiency. The massive amounts of status data generated during operation have not been effectively mined and utilized, making it difficult to form a data-driven closed loop for status assessment, remaining life prediction, and fault early warning.

[0004] In terms of production scheduling, existing scheduling strategies are mostly based on static rules or offline optimization, assuming constant equipment capacity. In actual operation, equipment health status deteriorates dynamically, and disturbances such as sudden deterioration and emergency order insertion need to be addressed. Existing mechanisms cannot integrate real-time equipment health status with dynamic task load, and cannot adaptively adjust task allocation, process paths, and resource ratios. This results in equipment continuing to operate at full capacity in a deteriorated state, exacerbating its decline, or maintenance windows conflicting with production peaks, failing to achieve coordinated optimization of equipment maintenance and production efficiency.

[0005] In summary, how to leverage artificial intelligence technology to deeply mine multi-source heterogeneous operational data, construct a real-time monitoring and fault early warning model for equipment health status, and on this basis, establish an adaptive dynamic scheduling mechanism that integrates real-time health status and dynamic task requirements to achieve coordinated optimization of maintenance and scheduling has become a key technical challenge that urgently needs to be addressed to improve the intelligence level of intelligent metering equipment calibration lines. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a dynamic scheduling optimization method, device and computer program product for calibration production line based on artificial intelligence, so as to realize the proactive prediction of the health status of calibration production line equipment and integrate it into dynamic scheduling in real time, thereby achieving adaptive and collaborative optimization of equipment maintenance and production efficiency.

[0007] To address the aforementioned technical problems, this invention provides a dynamic scheduling optimization method for an inspection pipeline based on artificial intelligence, comprising: Step S1: Collect and preprocess the operating data of key stations in the inspection line, then input the data into a bidirectional long short-term memory network based on the attention mechanism to extract deep degradation features, and fuse data-driven pattern matching and physical stress-life curves to calculate the comprehensive health index of each device. Step S2: When the comprehensive health index is lower than a preset threshold or the remaining lifespan is insufficient, an early warning message is generated; based on the early warning message, the attributes of the task queue to be scheduled, and the current status of each device, a dynamic resource constraint matrix is ​​constructed in real time; wherein, the current status includes the device busy / idle status, the availability status calculated based on the comprehensive health index, the procedure compatibility, and the equivalent load rate weighted by health. Step S3: Construct a heterogeneous graph neural network, taking the dynamic resource constraint matrix and task attributes as input, and extract a global state embedding vector that integrates device status, task load and topological constraint relationship; Step S4: Using the global state embedding vector as input, the dual-agent reinforcement learning model collaboratively generates task allocation and path planning instructions under the guidance of a reward function that includes a production efficiency reward term, a critical equipment load balancing penalty term, and a health maintenance incentive term, and then issues them for execution.

[0008] Preferably, step S1 specifically includes: Step S11 involves synchronously collecting vibration waveforms, temperature rise curves, motor current, robotic arm torque, and PLC operation logs through vibration sensors, temperature sensors, current transformers, torque sensors, and PLC controllers deployed on the inspection assembly line's loading robot, conveyor line, connection and disconnection device, inspection platform, and sorting mechanism. These data are then preprocessed by edge computing nodes to perform timestamp alignment, outlier removal, and normalization, forming a multi-dimensional time-series data stream. Step S12: Input the multidimensional time-series data stream into the bidirectional long short-term memory network based on the attention mechanism, and extract the deep degradation feature vector that represents the evolution trend of the device health status by weighting the time-series features through the self-attention layer. Step S13: Input the deep degradation feature vector into the degradation trend encoder, use the dynamic time warping alignment layer to calculate its matching degree with the typical degradation pattern center obtained by clustering historical failure data, and generate weights reflecting the degree of correlation with the historical behavior of the device through the external memory network; sum the matching degree and weights by weight and normalize to obtain the data-driven health component. Step S14: Calculate a rough estimate of the physical remaining life based on the equipment stress-life curve, generate a physical prior correction term through Sigmoid transformation; multiply the data-driven health component with the physical prior correction term, and map it to the [0,1] interval through the Sigmoid function to obtain the comprehensive health index.

[0009] Preferably, the step S2, which involves generating an early warning message when the comprehensive health index is below a preset threshold or when the remaining lifespan is insufficient, specifically includes: When the overall health index of a device falls below a preset warning threshold for the first time, or when the rough estimate of the remaining physical lifespan is less than the minimum time required from the current moment to the planned next maintenance, a structured warning message is generated. The structured warning message includes the device ID, the current overall health index, the predicted failure mode category, and the recommended latest switchover time.

[0010] Preferably, in step S2, the dynamic resource constraint matrix, for any device Its elements include: Current busy / idle status ; Availability status calculated based on the health status ; Embedded representation of the set of equipment-compatible verification procedure types ;as well as Health-weighted equivalent load factor ,in, For real-time load rate, This is the rated maximum processing capacity. This refers to the comprehensive health index.

[0011] Preferably, in step S3, constructing the heterogeneous graph neural network and extracting the global state embedding vector specifically includes: Construct a heterogeneous graph containing device nodes and task nodes. The initial features of the device nodes are taken from the corresponding row vectors in the dynamic resource constraint matrix. The initial features of the task nodes consist of the task priority weight, the type code of the meter to be inspected, the embedded representation of the verification procedure number, and the estimated verification time. Based on the compatibility of the procedures, establish the processable edges between equipment nodes and task nodes, and based on the physical path connection relationship of the verification pipeline, establish the material flow edges between equipment nodes; Information is propagated and aggregated through a multi-layer graph attention mechanism. Each node calculates the attention importance coefficient of its neighbors and aggregates the features of its neighbors in a weighted manner to update its own representation. After L-layer graph attention aggregation, the final representations of all device nodes and task nodes are concatenated, and a fixed-dimensional global state embedding vector is generated by a global readout function that combines mean pooling and max pooling.

[0012] Preferably, in step S4, the dual-agent reinforcement learning model includes a first agent responsible for task allocation and a second agent responsible for path planning; the policy networks of the first agent and the second agent both take the global state embedding vector as input, output the probability distribution of task allocation actions and the probability distribution of path planning actions respectively, and obtain specific scheduling actions through joint sampling.

[0013] Preferably, in step S4, the reward function includes production efficiency rewards, critical equipment load balancing penalties, and health maintenance incentives, and the reward function is defined by the following formula:

[0014] in, It is the scheduling engine at all times The reward function; This is a collection of completed tasks. It is a dynamic adjustment coefficient for the return on production efficiency; For the total set of tasks; Assign task weights; It is the dynamic adjustment coefficient of the load balancing penalty; It is a collection of key equipment whose health level is below the warning threshold. It is the number of elements in the set; This represents the average equivalent load rate of all critical equipment. It is a health maintenance indicator function; It is a fixed coefficient for health maintenance incentives.

[0015] Preferably, in step S4, the dual-agent reinforcement learning model is collaboratively trained using an entropy regularized belief region policy optimization algorithm, and differentiable discrete action sampling is achieved using the Gumbel-Softmax technique; the generated scheduling instructions are sent to the PLC control system for execution via industrial Ethernet; after execution, the comprehensive health index of each device is recalculated based on the updated sensor data, and the execution effect is fed back to the model for online updates.

[0016] The present invention also provides an artificial intelligence-based dynamic scheduling optimization device for a testing pipeline, comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to execute the AI-based verification pipeline dynamic scheduling optimization method.

[0017] The present invention also provides a computer program product, including computer instructions, which instruct a computer device to perform the operation corresponding to the AI-based dynamic scheduling optimization method for the verification pipeline.

[0018] The beneficial effects of this invention are as follows: By constructing a closed-loop intelligent architecture of "perception-decision-execution-feedback," it systematically solves the technical problem of the disconnect between equipment maintenance and production scheduling in traditional calibration production lines. At the perception level, this invention integrates multi-source sensor data with a deep neural network based on an attention mechanism to evaluate the overall health index of the equipment in real time and predict its remaining lifespan. When the health level drops to a warning threshold, an early warning is automatically triggered, transforming equipment health status from passive post-event maintenance to proactive pre-failure prediction. This allows resource switching and task migration to be completed before equipment failure, effectively avoiding unplanned downtime and significantly improving the continuous operation capability of the production line. At the decision-making level, this invention innovatively introduces a collaborative scheduling architecture based on heterogeneous graph neural networks and dual-agent reinforcement learning. Through heterogeneous graphs, it unifies and globally encodes equipment states, task attributes, and physical topology constraints. The scheduling strategy is then collaboratively generated by the task allocation agent and the path planning agent. The reward function design incorporates three factors: production efficiency rewards, critical equipment load balancing penalties, and health maintenance incentives. This forces the scheduling strategy to automatically balance short-term output with long-term equipment reliability under dynamic operating conditions: when the health of critical equipment declines, tasks can be automatically diverted to healthy equipment to protect deteriorated equipment; when the urgency of tasks increases, the system can moderately tolerate load imbalances to prioritize delivery, thereby endowing the production line with adaptive response capabilities to sudden abnormal operating conditions and avoiding overall efficiency losses caused by local bottlenecks. At the constraint modeling level, the dynamic resource constraint matrix constructed in this invention uniformly encodes multi-dimensional information such as equipment busy / idle status, health-weighted equivalent load rate, and procedure compatibility, providing the agent with an accurate view of equipment availability. By introducing a penalty for the dispersion of equivalent load rate into the reward function, the scheduling strategy is guided to minimize the load difference between key equipment, suppress the overuse and accelerated aging of high-value equipment, and reduce the idleness of inefficient equipment, thereby achieving a significant improvement in the load balance of all equipment, the extension of the lifespan of key equipment, and the overall verification efficiency. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an artificial intelligence-based dynamic scheduling optimization method for a testing pipeline according to Embodiment 1 of the present invention. Detailed Implementation

[0021] The following description of the embodiments is taken with reference to the accompanying drawings, which illustrate specific embodiments in which the invention can be implemented.

[0022] Please refer to Figure 1 As shown, Embodiment 1 of the present invention provides a dynamic scheduling optimization method for an inspection pipeline based on artificial intelligence, comprising: Step S1: Collect and preprocess the operating data of key stations in the inspection line, then input the data into a bidirectional long short-term memory network based on the attention mechanism to extract deep degradation features, and fuse data-driven pattern matching and physical stress-life curves to calculate the comprehensive health index of each device. Step S2: When the comprehensive health index is lower than a preset threshold or the remaining lifespan is insufficient, an early warning message is generated; based on the early warning message, the attributes of the task queue to be scheduled, and the current status of each device, a dynamic resource constraint matrix is ​​constructed in real time; wherein, the current status includes the device busy / idle status, the availability status calculated based on the comprehensive health index, the procedure compatibility, and the equivalent load rate weighted by health. Step S3: Construct a heterogeneous graph neural network, taking the dynamic resource constraint matrix and task attributes as input, and extract a global state embedding vector that integrates device status, task load and topological constraint relationship; Step S4: Using the global state embedding vector as input, the dual-agent reinforcement learning model collaboratively generates task allocation and path planning instructions under the guidance of a reward function that includes a production efficiency reward term, a critical equipment load balancing penalty term, and a health maintenance incentive term, and then issues them for execution.

[0023] As can be seen from the above steps, the embodiments of the present invention integrate multi-source sensor data and deep neural networks to evaluate equipment health in real time and predict failure risks in advance. The remaining lifespan and health index of the equipment are directly incorporated into the constraints of scheduling decisions, and the task allocation and path planning are coordinated and optimized through a dual-agent reinforcement learning architecture, thereby maximizing equipment availability and life-cycle health status while ensuring production efficiency.

[0024] Specifically, in this embodiment of the invention, step S1 further includes steps S11 to S14.

[0025] In step S11, to achieve real-time monitoring and fault prediction of equipment status, multi-source heterogeneous sensors are deployed at key stations of the calibration line. These key stations include a loading robot, conveyor line, connection / disconnection device, calibration platform, and sorting mechanism. Vibration waveforms, temperature rise curves, motor current, robotic arm torque, and PLC internal register operation log data are simultaneously collected via high-speed data acquisition interfaces from vibration sensors, temperature sensors, current transformers, torque sensors, and the PLC controller. The collected raw data is first preprocessed through an edge computing node. Preprocessing operations include timestamp alignment, outlier removal, and normalization, ultimately forming a unified format multi-dimensional time-series data stream.

[0026] In step S12, the preprocessed multidimensional time-series data stream is input into a bidirectional long short-term memory network based on an attention mechanism. This network weights the time-series features through a self-attention layer, adaptively focusing on the time segments most critical to representing the device's health status, thereby extracting deep degradation feature vectors that characterize the evolution trend of the device's health status. This feature vector It represents all health-related information of the device from the start of operation to the current moment.

[0027] In step S13, the deep degradation feature vector is... The input is fed into the degradation trend encoder module, which calculates the health index by fusing data-driven pattern matching and physical prior knowledge. Specifically, a dynamic time warping alignment layer is first used to align this deep degradation feature vector. A similarity measure is performed between the data and the centers of typical degradation patterns formed by clustering historical failure data to obtain a matching degree. Simultaneously, an external memory network dynamically generates weights based on current operating conditions; these weights reflect the correlation between the current degradation pattern and the equipment's historical behavior records. The matching degree and weights are then weighted and summed, and divided by a globally maximum normalization factor to eliminate biases caused by differences in units or operating conditions between different devices, resulting in a data-driven health component.

[0028] In step S14, a rough estimate of the physical remaining life is calculated based on the inherent stress-life curve of the equipment. This rough estimate is then transformed using a Sigmoid function to generate a physical prior correction term. The data-driven health component obtained in step S13 is multiplied by this physical prior correction term, and the product is then mapped to the [0,1] interval using the Sigmoid function to obtain the comprehensive health index.

[0029] The complete expression for the overall health index is as follows:

[0030] in, Represents the device at any time The health index has a value range strictly limited to between 0 and 1 by the sigmoid function, where 0 represents complete failure and 1 represents a perfect health state. It is a moment The deep degenerate feature vector; It is the first cluster obtained from pre-training clustering Feature centers of typical degradation patterns; This refers to the capacity of the pattern library; parameters It is a positive scalar that controls the bandwidth of the radial basis kernel function. The larger the value, the higher the sensitivity of feature matching; Is with the first The dynamic weight coefficients corresponding to the typical degradation patterns reflect the weights of the first class. The degree of correlation between the degradation pattern and the current device's past behavior; It is the square of the Euclidean distance, measuring the current feature and the... The smaller the difference between the pattern centers, the higher the matching degree. It is based on each model center When used as a baseline, it is the sum of its weighted matching degrees with all pattern centers (including itself), and all... The maximum value in the corresponding sum serves as a global normalization, eliminating the influence of dimensions between different devices. The latter part is an S-shaped function based on the physical model correction. It is a rough estimate of the remaining life calculated based on the equipment stress-life curve, parameters Controlling the steepness of the curve, The offset represents the median point of the physical lifetime estimate. The two parts are combined through multiplication, and then the final health score is output via sigmoid, thus integrating data-driven deep feature matching and physical degradation patterns to achieve a quantitative assessment of equipment health status that is both sensitive and robust.

[0031] In step S2, the overall health index of each key device is continuously monitored. When the overall health index of a device first falls below a preset warning threshold, or when the rough estimate of the remaining physical life calculated in step S14 is less than the minimum time required from the current moment until the next planned maintenance, the system immediately generates a structured warning message. This warning message includes the device ID, the current overall health index, the predicted failure mode category, and the suggested latest switchover time, and is pushed to the scheduling decision-making stage in real time.

[0032] After the scheduling engine starts, it first acquires all status information of the current pipeline, including the queue of tasks to be inspected and the real-time health of all devices. Each task in the queue of tasks to be inspected has multiple attributes: task ID, type of meter to be inspected, corresponding verification procedure number, task priority, and expected inspection duration. Simultaneously, the scheduling engine reads the latest health of all devices from the database and, combined with the aforementioned warning information, constructs a dynamic resource constraint matrix in real time. The rows of this matrix correspond to each device, and the columns correspond to various real-time attributes of the device.

[0033] Specifically, regarding equipment The elements in the matrix include: the current busy / idle status. Availability status calculated based on health status Embedded representation of the set of equipment-compatible verification procedure types ; and the equivalent load factor weighted by health status. ,in For real-time load rate, This represents the rated maximum processing capacity. The construction of this matrix can be formalized as follows:

[0034] in, It is a moment The resource constraint matrix; Indicates transpose; It is an embedding vector of device type and location; For equipment collection, Let be the total number of devices. A health-weighted equivalent load rate is introduced, so that when a device's health decreases, its equivalent load rate increases accordingly, thereby reducing the tendency to assign new tasks to that device in scheduling decisions.

[0035] After constructing the resource constraint matrix, step S3 operates a spatiotemporal co-evolutionary reinforcement learning decision architecture through a scheduling engine. This architecture employs a dual-agent collaborative graph attention reinforcement learning model, with the two agents responsible for task allocation and path planning, respectively. To achieve shared environmental perception between the two agents, a heterogeneous graph neural network is first constructed to extract global state embedding vectors. The node set of a heterogeneous graph consists of two types of nodes: device nodes. and task nodes The initial feature vector of each device node is directly taken from the corresponding row vector in the resource constraint matrix; the initial feature vector of each task node consists of the attributes of the task in the task queue, including the task priority weight, the type code of the meter to be inspected, the embedded representation of the verification procedure number, and the estimated verification duration. The edges in the heterogeneous graph include two types: "processable" edges between devices and tasks, indicating that the task can be executed by the device (based on procedure compatibility); and "material flow" edges between devices, representing the preset physical path connection relationship on the verification pipeline.

[0036] After the heterogeneous graph is constructed, information propagation and aggregation are performed through a multi-layer graph attention mechanism. At each layer, each node calculates the importance coefficient of its neighboring nodes using the attention mechanism, and then weights and aggregates neighbor features accordingly to update its own representation. After the graph attention layer, the final representations of all device nodes and task nodes are concatenated and input into a global readout function. This function uses mean pooling and max pooling to fuse the node features of the entire graph into a fixed-dimensional global state embedding vector. It condenses the current state of the equipment, the workload, and the topological constraints between them in the entire pipeline, and serves as the common input for the subsequent two reinforcement learning agent policy networks to generate scheduling actions.

[0037] In step S4, the decision architecture employs a dual-agent collaborative graph attention reinforcement learning model, with the two agents responsible for task allocation and path planning, respectively. The policy networks of both the first agent (task allocation agent) and the second agent (path planning agent) use the global state embedding vector generated in step S3. As input, the algorithm outputs the probability distributions of task allocation actions and path planning actions, respectively. Specific scheduling actions are obtained by joint sampling from these two probability distributions, and differentiable discrete action sampling is achieved using the Gumbel-Softmax technique to support end-to-end gradient optimization training.

[0038] The agent's strategy is optimized through a reward function, which unifies the conflicting objectives of maximizing production efficiency and maintaining equipment health in traditional methods within a single optimization framework. This allows the agent to make scheduling strategies in dynamic environments that balance current output with long-term equipment reliability. Reward function It consists of three components, corresponding to the production efficiency reward, the critical equipment load balancing penalty, and the health maintenance incentive, and their specific expressions are as follows:

[0039] in, It is the scheduling engine at all times The reward function is used to evaluate the merits of the current scheduling action and guide the policy update of the reinforcement learning agent; the first term Represents the return on production efficiency; This is a collection of completed tasks. It is a dynamic adjustment coefficient for production efficiency returns, with a value range of [0.5, 3]. For the total set of tasks; This represents the task weight; a higher ratio indicates higher value created per unit of time. (Second item) This is a penalty item for equipment load balancing; It is the dynamic adjustment coefficient of the load balancing penalty, with a value range of [0.2, 2]. It is a collection of key equipment whose health level is below the warning threshold. It is the number of elements in the set; The average equivalent load rate of all critical equipment is given, and the standard deviation quantifies the degree of load imbalance. Subtracting this term forces tasks to flow from equipment with low health to equipment with high health. (Third term) Incentives for health maintenance; This is a health maintenance indicator function, taking the value 0 or 1. It is valid only if, after the current scheduling action is executed, in the next health assessment, the health index of at least one critical device increases or the downward trend in health stops (e.g., the difference between two consecutive values ​​is less than a certain minimum positive number). It is 1 if it is true, otherwise it is 0. It is a fixed coefficient for health maintenance incentives, with a value range of [0.1, 0.5].

[0040] Based on the aforementioned reward function, the two agents are collaboratively trained using an entropy-regularized confidence region policy optimization algorithm. During training, a policy entropy regularization term is added to the policy optimization objective to encourage the agents to maintain a certain level of exploration capability and avoid premature convergence to local optima. At each decision time, the two agents embed their global state vectors... Each strategy distribution is output, and joint sampling yields specific scheduling actions. Differentiable discrete action sampling is achieved using the Gumbel-Softmax technique. The generated scheduling instructions specifically include at least one of the following operations: changing material flow direction, adjusting the load distribution of the calibration platform, or calling on backup equipment to replace a failing workstation. These scheduling instructions are sent to the PLC control system for execution via industrial Ethernet.

[0041] After the instruction is executed, the equipment load changes, the sensor data stream is updated synchronously, and the system returns to step S1 to recalculate the comprehensive health index of each device and the task completion status. All execution effects (including the new state and the reward value calculated based on the new state) are fed back to the dual-agent reinforcement learning model for online model updates, thereby enabling the scheduling strategy to continuously adapt to the dynamic evolution of the pipeline state, forming a complete closed loop of perception-decision-execution-feedback.

[0042] Corresponding to the AI-based dynamic scheduling optimization method for a testing pipeline in Embodiment 1 of the present invention, Embodiment 2 of the present invention also provides an AI-based dynamic scheduling optimization device for a testing pipeline, comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to execute the AI-based verification pipeline dynamic scheduling optimization method.

[0043] Corresponding to the AI-based dynamic scheduling optimization method for the verification pipeline in Embodiment 1 of the present invention, Embodiment 3 of the present invention also provides a computer program product, including computer instructions, which instruct a computer device to perform the operation corresponding to the AI-based dynamic scheduling optimization method for the verification pipeline.

[0044] Preferably, the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the device, connecting various parts of the device through various interfaces and lines.

[0045] The memory mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard drive, a SmartMedia Card (SMC), a Secure Digital (SD) card, and a Flash Card, or other volatile solid-state storage devices.

[0046] It should be noted that the above-mentioned devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art.

[0047] For the working principle and process of the above embodiments, please refer to the description of Embodiment 1 of the present invention, which will not be repeated here.

[0048] Compared with existing technologies, this invention has the following significant advantages: By constructing a closed-loop intelligent architecture of "perception-decision-execution-feedback," it systematically solves the technical problem of the disconnect between equipment maintenance and production scheduling in traditional calibration production lines. At the perception level, this invention integrates multi-source sensor data with a deep neural network based on an attention mechanism to evaluate the overall health index of the equipment in real time and predict its remaining lifespan. When the health level drops to a warning threshold, an early warning is automatically triggered, transforming equipment health status from passive post-event maintenance to proactive pre-failure prediction. This allows resource switching and task migration to be completed before equipment failure, effectively avoiding unplanned downtime and significantly improving the continuous operation capability of the production line. At the decision-making level, this invention innovatively introduces a collaborative scheduling architecture based on heterogeneous graph neural networks and dual-agent reinforcement learning. Through heterogeneous graphs, it unifies and globally encodes equipment states, task attributes, and physical topology constraints. The scheduling strategy is then collaboratively generated by the task allocation agent and the path planning agent. The reward function design incorporates three factors: production efficiency rewards, critical equipment load balancing penalties, and health maintenance incentives. This forces the scheduling strategy to automatically balance short-term output with long-term equipment reliability under dynamic operating conditions: when the health of critical equipment declines, tasks can be automatically diverted to healthy equipment to protect deteriorated equipment; when the urgency of tasks increases, the system can moderately tolerate load imbalances to prioritize delivery, thereby endowing the production line with adaptive response capabilities to sudden abnormal operating conditions and avoiding overall efficiency losses caused by local bottlenecks. At the constraint modeling level, the dynamic resource constraint matrix constructed in this invention uniformly encodes multi-dimensional information such as equipment busy / idle status, health-weighted equivalent load rate, and procedure compatibility, providing the agent with an accurate view of equipment availability. By introducing a penalty for the dispersion of equivalent load rate into the reward function, the scheduling strategy is guided to minimize the load difference between key equipment, suppress the overuse and accelerated aging of high-value equipment, and reduce the idleness of inefficient equipment, thereby achieving a significant improvement in the load balance of all equipment, the extension of the lifespan of key equipment, and the overall verification efficiency.

[0049] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A dynamic scheduling optimization method for an inspection production line based on artificial intelligence, characterized in that, include: Step S1: Collect and preprocess the operating data of key stations in the inspection line, then input the data into a bidirectional long short-term memory network based on the attention mechanism to extract deep degradation features, and fuse data-driven pattern matching and physical stress-life curves to calculate the comprehensive health index of each device. Step S2: When the comprehensive health index is lower than a preset threshold or the remaining lifespan is insufficient, an early warning message is generated; based on the early warning message, the attributes of the task queue to be scheduled, and the current status of each device, a dynamic resource constraint matrix is ​​constructed in real time; wherein, the current status includes the device busy / idle status, the availability status calculated based on the comprehensive health index, the procedure compatibility, and the equivalent load rate weighted by health. Step S3: Construct a heterogeneous graph neural network, taking the dynamic resource constraint matrix and task attributes as input, and extract a global state embedding vector that integrates device status, task load and topological constraint relationship; Step S4: Using the global state embedding vector as input, the dual-agent reinforcement learning model collaboratively generates task allocation and path planning instructions under the guidance of a reward function that includes a production efficiency reward term, a critical equipment load balancing penalty term, and a health maintenance incentive term, and then issues them for execution.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: Step S11 involves synchronously collecting vibration waveforms, temperature rise curves, motor current, robotic arm torque, and PLC operation logs through vibration sensors, temperature sensors, current transformers, torque sensors, and PLC controllers deployed on the inspection assembly line's loading robot, conveyor line, connection and disconnection device, inspection platform, and sorting mechanism. These data are then preprocessed by edge computing nodes to perform timestamp alignment, outlier removal, and normalization, forming a multi-dimensional time-series data stream. Step S12: Input the multidimensional time-series data stream into the bidirectional long short-term memory network based on the attention mechanism, and extract the deep degradation feature vector that represents the evolution trend of the device health status by weighting the time-series features through the self-attention layer. Step S13: Input the deep degradation feature vector into the degradation trend encoder, use the dynamic time warping alignment layer to calculate its matching degree with the typical degradation pattern center obtained by clustering historical failure data, and generate weights reflecting the degree of correlation with the historical behavior of the device through the external memory network; sum the matching degree and weights by weight and normalize to obtain the data-driven health component. Step S14: Calculate a rough estimate of the physical remaining life based on the equipment stress-life curve, generate a physical prior correction term through Sigmoid transformation; multiply the data-driven health component with the physical prior correction term, and map it to the [0,1] interval through the Sigmoid function to obtain the comprehensive health index.

3. The method according to claim 2, characterized in that, The step S2, which involves generating an early warning message when the comprehensive health index falls below a preset threshold or when the remaining lifespan is insufficient, specifically includes: When the overall health index of a device falls below a preset warning threshold for the first time, or when the rough estimate of the remaining physical lifespan is less than the minimum time required from the current moment to the next planned maintenance, a structured warning message is generated. The structured early warning information includes the device ID, the current comprehensive health index, the predicted failure mode category, and the recommended latest switchover time.

4. The method according to claim 1, characterized in that, In step S2, the dynamic resource constraint matrix, for any device Its elements include: Current busy / idle status ; Availability status calculated based on the health status ; Embedded representation of the set of equipment-compatible verification procedure types ;as well as Health-weighted equivalent load factor ,in, For real-time load rate, This is the rated maximum processing capacity. This refers to the comprehensive health index.

5. The method according to claim 1, characterized in that, In step S3, the construction of the heterogeneous graph neural network and the extraction of the global state embedding vector specifically includes: Construct a heterogeneous graph containing device nodes and task nodes. The initial features of the device nodes are taken from the corresponding row vectors in the dynamic resource constraint matrix. The initial features of the task nodes consist of the task priority weight, the type code of the meter to be inspected, the embedded representation of the verification procedure number, and the estimated verification time. Based on the compatibility of the procedures, establish the processable edges between equipment nodes and task nodes, and based on the physical path connection relationship of the verification pipeline, establish the material flow edges between equipment nodes; Information is propagated and aggregated through a multi-layer graph attention mechanism. Each node calculates the attention importance coefficient of its neighbors and aggregates the features of its neighbors in a weighted manner to update its own representation. After L-layer graph attention aggregation, the final representations of all device nodes and task nodes are concatenated, and a fixed-dimensional global state embedding vector is generated by a global readout function that combines mean pooling and max pooling.

6. The method according to claim 1, characterized in that, In step S4, the dual-agent reinforcement learning model includes a first agent responsible for task allocation and a second agent responsible for path planning; the policy networks of the first agent and the second agent both take the global state embedding vector as input, output the probability distribution of task allocation actions and the probability distribution of path planning actions respectively, and obtain specific scheduling actions through joint sampling.

7. The method according to claim 1 or 6, characterized in that, In step S4, the reward function includes production efficiency rewards, critical equipment load balancing penalties, and health maintenance incentives. The reward function is defined by the following formula: in, It is the scheduling engine at all times The reward function; This is a collection of completed tasks. It is a dynamic adjustment coefficient for the return on production efficiency; For the total set of tasks; Assign task weights; It is the dynamic adjustment coefficient of the load balancing penalty; It is a collection of key equipment whose health level is below the warning threshold. It is the number of elements in the set; This represents the average equivalent load rate of all critical equipment. It is a health maintenance indicator function; It is a fixed coefficient for health maintenance incentives.

8. The method according to claim 1, characterized in that, In step S4, the dual-agent reinforcement learning model is collaboratively trained using the entropy regularized belief region policy optimization algorithm, and differentiable discrete action sampling is achieved using the Gumbel-Softmax technique. The generated scheduling instructions are sent to the PLC control system for execution via industrial Ethernet; after execution, the comprehensive health index of each device is recalculated based on the updated sensor data, and the execution effect is fed back to the model for online updates.

9. A dynamic scheduling and optimization device for an artificial intelligence-based inspection production line, characterized in that, include: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the AI-based dynamic scheduling optimization method for verification pipelines as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, It includes computer instructions that instruct a computer device to perform operations corresponding to the dynamic scheduling optimization of the verification pipeline based on artificial intelligence as described in any one of claims 1 to 8.

Images