A method for STK account material self-checking and abnormal self-healing based on device state comparison

Abstract

The application discloses a kind of STK account material self-checking and abnormal self-healing method based on equipment state comparison, belongs to industrial automation control and intelligent warehousing technical field, and the specific steps of the self-healing method are as follows: I: the real-time state data of each device is collected and preprocessed, and according to each state data after processing, the health degree score and state entropy value of each device are calculated;The present application can identify the sub-health and unstable state of equipment in advance, which not only ensures the production safety in case of serious failure, but also avoids unnecessary downtime caused by slight abnormalities, significantly reduces the abnormal interruption in the task execution process, and improves the accuracy and traceability of material management;At the same time, the present application can find potential problems in advance, avoid task out-of-order execution, dependency conflict and deadlock problem, effectively prevent the preemption of concurrent tasks on process resources, and significantly optimize Crane walking path, reduce the overall energy consumption of warehouse, and greatly shorten the abnormal processing cycle.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation control and intelligent warehousing technology, and in particular to a STK self-verification and anomaly self-healing method based on equipment status comparison. Background Technology

[0002] With the rapid development of intelligent manufacturing technology, automated storage and retrieval equipment (STK) has become an indispensable core equipment in discrete manufacturing production lines such as semiconductor wafer manufacturing and electronic component assembly. It undertakes the tasks of storing, transferring and handling work-in-process, semi-finished products and finished products. Its operating efficiency, reliability and inventory management accuracy directly determine the capacity level and product quality of the entire production line. At present, the industry has widely used PLC and industrial Ethernet-based automated control systems to realize the basic operation control of STK equipment, and combined with MES / MCS systems to realize the issuance of production tasks and feedback of execution status.

[0003] Existing STK (Site Management Kit) self-verification and anomaly self-healing methods cannot identify sub-health and unstable equipment states in advance, leading to frequent abnormal interruptions during task execution and poor accuracy and traceability of material management. These methods are also prone to out-of-order task execution, dependency conflicts, and deadlocks, as well as concurrent tasks vying for process resources. Furthermore, the Crane travel path is not optimized, resulting in high overall energy consumption in the warehouse. Therefore, we propose a STK self-verification and anomaly self-healing method based on equipment status comparison. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a STK data self-verification and anomaly self-healing method based on equipment status comparison.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for STK data self-verification and anomaly self-healing based on equipment status comparison, the specific steps of which are as follows:

[0007] Ⅰ: Collect and preprocess the real-time status data of each device, and calculate the health score and status entropy value of each device based on the processed status data;

[0008] II: Based on the calculated health scores and state entropy values ​​of each device, determine whether to trigger the self-healing process and dynamic task avoidance;

[0009] Ⅲ: Construct a task dependency chain graph, mark the priority of all tasks according to the task dependency chain graph, and then determine whether to send a resource reservation signal to the corresponding target device;

[0010] IV: Obtain the multi-type characteristics of all candidate storage sites, calculate the selection suitability score of each candidate storage site, and then select the optimal storage site based on the calculation results;

[0011] V: Obtain the status data of each device's inventory records, compare it with the corresponding real-time status data, and then trigger the exception handling process based on the comparison results.

[0012] As a further aspect of the present invention, the specific steps for calculating the health score and state entropy value of each device in step I are as follows:

[0013] S1.1: Collect real-time status data of each device in the storage equipment STK, and remove real-time status data that does not meet the preset standards. Then check whether the logical relationship between each real-time status data is valid. If the logical relationship is not valid, remove the corresponding real-time status data. Then map the remaining real-time status data to the interval [0, 1]. Each device in the storage equipment STK includes the conveying mechanism Crane, the interface Port, and the storage position. The real-time status data includes the frequency of equipment failure, task execution time, IO status changes, and communication quality.

[0014] S1.2: Extract multiple statistical feature values ​​for each device from the processed real-time status data, and convert each extracted statistical feature value into a single health score between 0 and 100%. Then, multiply all single health scores of each device by the corresponding preset weight coefficient, and sum all the calculated products to obtain the health score of the device.

[0015] S1.3: Obtain the IO state change record table in the local database, and extract all state change events of each device in the past 10 seconds. Arrange all state change events into a continuous device state sequence according to the time sequence. Then, according to the preset state coding rules of each device PLC, convert each state into a corresponding code. Then, count the total number of times different state codes appear in each device state sequence, and divide the total number of times each state code appears by the total length of the device state sequence to obtain the probability of occurrence of the corresponding state. Then, multiply the probability of occurrence of each state by the logarithm of the probability to the base 2, and then add all the products and take the negative number to obtain the state entropy value of each device.

[0016] As a further aspect of the present invention, the specific steps for determining whether the self-healing process and dynamic task avoidance are triggered in step II are as follows:

[0017] S2.1: Read the health score and state entropy value of each device, and compare them with the preset warning threshold, fault threshold, and instability threshold respectively. If the health score is lower than the preset fault threshold, trigger a fault alarm and emergency shutdown; if the health score is higher than or equal to the preset warning threshold and the state entropy value is lower than the preset instability threshold, maintain the normal operation of the device; if the health score is lower than the preset warning threshold but higher than the preset fault threshold, trigger the self-healing process; if the state entropy value is higher than or equal to the preset instability threshold, trigger dynamic avoidance.

[0018] S2.2: When the self-healing process is triggered, the corresponding device status is marked as sub-healthy in the built-in device status table, and a temporary interception rule is added at the task allocation entry of the built-in task scheduler. Then, according to the added temporary interception rule, all new task requests pointing to sub-healthy devices are intercepted. At the same time, the current task queue waiting to be executed is obtained and traversed, all tasks whose target devices are sub-healthy devices are filtered out, and all filtered tasks are removed and then transferred to the preset temporary scheduling queue.

[0019] S2.3: Query all devices of the same type that are in normal status, obtain the multi-type factors of each normal device, sort them according to preset criteria, select the device ranked first as the replacement device, if there is a usable replacement device, update the target device of the task to the selected replacement device, recalculate the task execution path, and add the task to the normal task scheduling queue; otherwise, update the target location of the task to the preset dedicated buffer storage location and temporarily store the goods in the buffer area.

[0020] S2.4: Extract the accounting data of all sub-healthy devices from the built-in local material management database, and perform integrity and consistency verification on each extracted accounting data. At the same time, based on the accounting data that passes the verification, generate the corresponding accounting snapshot file. Then, monitor the current task status of each sub-healthy device in real time. When the sub-healthy device completes the current task, no new tasks will be assigned to it, the device will be set to the maintenance mode, and no new tasks will be issued to the device.

[0021] S2.5: When dynamic avoidance is triggered, all new tasks pointing to devices with state entropy values ​​higher than the preset instability threshold are marked as pending observation and placed in the preset pending observation task queue. All tasks in the pending observation task queue are suspended for a preset backoff time. During the backoff time, tasks pointing to other stable devices are scheduled first. After the backoff time ends, the latest state entropy value of the device is obtained. If the latest state entropy value is lower than the preset instability threshold, the tasks in the pending observation task queue are added back to the normal scheduling queue.

[0022] As a further aspect of the present invention, the specific steps for determining whether to send a resource reservation signal to the corresponding target device in step III are as follows:

[0023] S3.1: Listen to the production orders issued by the upper-level MCS / MES system, then parse the complete process route information in the production orders, and extract the multi-type data and sequence constraints of each process task from them. Then, based on the parsed task sequence constraints, construct a task dependency chain graph, while continuously monitoring the status changes of all tasks, and updating the task dependency chain graph according to the status changes of all tasks.

[0024] S3.2: Obtain the status of all tasks and mark tasks whose status meets the preset criteria as executable tasks. At the same time, obtain the starting device and ending device of each executable task and check whether the starting device and ending device of each executable task are in the preset highest priority device list. If the starting device and ending device of any executable task are in the preset highest priority device list, mark the task as the highest priority task and trigger the resource reservation process. Then, extract the corresponding target device from each highest priority task and send a resource reservation signal to the PLC of the target device. Then update the status of the corresponding highest priority task to "resource reserved".

[0025] S3.3: When the highest priority task with reserved resources is scheduled, a task execution instruction is sent to the target device. At the same time, after receiving the task execution instruction, the PLC of the target device verifies whether the task ID and station number in the task execution instruction match the current reserved information. If they match, the reserved flag is cleared and the task execution begins; otherwise, the task execution is refused and a reservation mismatch error is indicated. Subsequently, when the target device completes the task, the status of the corresponding task is updated to completed, and a reserved resource release instruction is sent to the target device. After receiving the instruction, the PLC of the target device clears any residual reserved flags of the corresponding station and restores the station to an available state.

[0026] As a further aspect of the present invention, the specific steps for selecting the optimal storage location based on the calculation results in step IV are as follows:

[0027] S4.1: During the task scheduling and execution process, continuously monitor the task execution conditions. When it is detected that the target device of the task is in a sub-healthy state and there is no available alternative device, trigger the temporary storage location selection process. At the same time, read the real-time status information of all storage locations from the built-in local storage location management database, and exclude storage locations with abnormal lock status. Then exclude storage locations whose distance from the Crane location of the currently executing task exceeds the preset maximum allowable distance, and then integrate all remaining storage locations to generate a list of candidate storage locations.

[0028] S4.2: Extract the feature data of each storage location from the list of candidate storage locations, and input the feature data of each candidate storage location into the trained score inference model according to a preset batch. The model generates the suitability score of each candidate storage location through forward propagation calculation. Then, sort the suitability scores of each candidate storage location from high to low, select the storage location ranked first as the target temporary storage location for this task, and mark the selected storage location as reserved. The feature data includes the physical distance of the storage location, the task popularity of the area where the storage location is located, the path congestion coefficient, and the number of tasks expected to arrive in the future.

[0029] S4.3: Obtain all Crane numbers, task queue lengths, identifiers of goods to be stored, and target storage location numbers. Generate task execution instructions based on the acquired data and send them to the corresponding Crane controller. After receiving the execution instructions, the Crane controller collects the real-time operating data of the Crane itself, monitors the IO status of the target storage location and the operating status of peripheral devices, and compares the real-time operating data of the Crane itself with the preset task plan. Calculate the deviation between the actual task progress and the preset task plan. If the deviation exceeds the preset threshold, a progress warning is triggered and the abnormal event is recorded.

[0030] As a further aspect of the present invention, the specific training steps of the fractional reasoning model described in S4.2 are as follows:

[0031] P1.1: Collect complete execution data of all temporary tasks within 3 months, and simultaneously collect auxiliary data for each temporary task. Then integrate all the collected data to form the original training dataset. Subsequently, remove data in the original training dataset that exceeds 3 times the normal range standard deviation, and map the remaining data in the original training dataset to the interval [0, 1]. Then, randomly divide the processed original training dataset into training set, validation set and test set in a ratio of 7:2:1.

[0032] P1.2: Based on the essential characteristics of the STK storage location selection task, the constraints of embedded hardware resources, and the real-time requirements of industry, a fractional inference model is constructed. At the same time, the Xavier uniform distribution method is used to initialize the model weights, and the mean squared error is selected as the loss function. Then, the Adam optimizer is selected for parameter update, and the initial learning rate, β1 parameter, and β2 parameter are set.

[0033] P1.3: The training set data is input into the score inference model in preset batches for iterative training. In each training batch, forward propagation is performed first to obtain the model's predicted fitness score. Then, the mean squared error loss function between the predicted fitness score and the actual label is calculated. Next, the gradient of the mean squared error loss function with respect to the weights and biases of each layer is calculated using the backpropagation algorithm. Subsequently, the model parameters are updated according to the gradient using the Adam optimizer. After each complete training set iteration, the model is evaluated using the validation set data, and the mean squared error on the validation set is calculated. When the validation set loss does not decrease for 20 consecutive iterations, training is stopped, and the training of the score inference model is completed.

[0034] As a further aspect of the present invention, the specific steps of triggering the exception handling process based on the comparison result in step V are as follows:

[0035] S5.1: Retrieves the status data of each device from the built-in local material management database, and retrieves the real-time status data of each device. It then compares the real-time status data of each device with the recorded status data one by one to check for inconsistencies. If inconsistencies are found, the status of the same device is repeatedly verified in the next three comparison cycles. If the results of three consecutive comparisons show the same inconsistent status, an abnormal event is confirmed to have occurred. Subsequently, the context feature vector of each abnormal event is extracted. The context feature data includes device type features, recent fault records, IO status fluctuation features, and task execution history data.

[0036] S5.2: Input all extracted context feature vectors sequentially into the input layer of the trained anomaly attribution model. The model generates three probability values ​​in the output layer through forward propagation and performs a summation check to ensure that the sum of the probability values ​​is 1. At the same time, sort the three probability values ​​from high to low and select the anomaly category with the highest probability value as the judgment result of this anomaly. Then, based on the judgment result of this anomaly, select the corresponding preset processing strategy. If it is determined to be a sensor false alarm, the processing strategy is to perform a virtual clearing operation and not generate a physical handling task; if it is determined to be residual goods, the processing strategy is to perform association and release operations, and at the same time attempt to find and complete the residual task; if it is determined to be an unknown intrusion, the processing strategy is to perform security isolation and manual intervention operations, and transfer the goods to a dedicated anomaly storage location.

[0037] S5.3: Based on the selected processing strategy, execute the corresponding operations. For sensor false alarms, update the corresponding record in the local material management database, modify the material status to match the actual equipment status, and then clear the abnormal lock mark. For task-related residual goods, search the task database for an incomplete task record that matches the location and goods. If found, resume task execution. Otherwise, move the goods to the designated normal storage location. For unknown intrusions, generate a dedicated abnormal handling task, schedule an idle Crane to move the goods to the preset abnormal isolation storage location, and lock the storage location.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This invention collects real-time status data of each device within the STK, removes real-time status data that does not meet preset standards or whose logical relationships are invalid, maps the remaining real-time status data to the [0,1] interval, and extracts statistical feature values ​​of each device. These are then converted into single-item health scores from 0 to 100%, and a weighted sum is used to obtain the health score for each device. Simultaneously, the IO status change events of each device over the past 10 seconds are extracted, sorted by time to form a status sequence, and converted into PLC preset codes. The probability of each status occurrence is calculated, and the state entropy value of each device is obtained using the Shannon entropy formula. These two indicators are compared with preset warning thresholds, fault thresholds, and instability thresholds. If the health score is lower than the preset fault threshold, a fault alarm and emergency shutdown are triggered. If the health score is higher than or equal to the preset warning threshold and the state entropy value is lower than the preset instability threshold, normal operation is maintained. If the health score is lower than the preset warning threshold but higher than the preset fault threshold, a self-healing process is triggered. The device is first marked as sub-healthy, and an interception rule is added at the scheduling entry point. Tasks targeting this device are transferred to a temporary queue. The system simultaneously queries and sorts normal equipment of the same type to select the best alternative. After updating the task objective and path, it adds the equipment to the normal scheduling queue. If there is no alternative equipment, the goods are temporarily stored in a dedicated buffer storage location. At the same time, it extracts and verifies the inventory data of sub-healthy equipment to generate a snapshot. After the equipment completes the current task, it is set to a maintenance-pending mode. If the state entropy value is higher than or equal to the preset instability threshold, dynamic avoidance is triggered. New tasks pointing to the equipment are marked as pending observation and suspended for a preset backoff time. During this period, other stable equipment tasks are scheduled first. After the backoff ends and the state entropy value returns to normal, the pending observation task is added back to the normal scheduling queue. This system can identify sub-healthy and unstable equipment states in advance, ensuring production safety in the event of serious failures and avoiding unnecessary downtime caused by minor anomalies. It significantly reduces abnormal interruptions during task execution and improves the accuracy and traceability of material management.

[0040] 2. This invention monitors production orders issued by the upper-level MCS / MES system, parses complete process route information, extracts the type data and sequence constraints of each process task, constructs a task dependency chain diagram, and then obtains the status of all tasks. Tasks whose status meets the preset criteria are marked as executable tasks. The origin and destination devices are checked to see if they are in the preset highest priority device list. If they are, they are marked as the highest priority task and the resource reservation process is triggered. At the same time, a reservation signal containing the task ID and workstation information is sent to the target device PLC, and the task status is updated to "resources reserved". When the task is scheduled, an execution command is sent. The PLC verifies the reservation information and clears the flag bit to start execution. If the match fails, execution is rejected and an error is returned. After the task is completed, a resource release command is sent to restore the workstation to its available state. If the target device is found to be in a sub-healthy state and there is no available alternative device during task execution, a temporary storage location selection process is triggered to filter and exclude abnormal locking and over-distance storage locations to generate candidate locations. The system extracts feature data from each storage location and inputs it into a trained score inference model to obtain a suitability score. The storage location with the highest score is selected and marked for reservation. Simultaneously, the system sends instructions to the corresponding controller. The system also collects real-time operational data from the Crane itself, monitors the I / O of the target storage location and the status of surrounding equipment, compares the actual progress with the planned progress to calculate the deviation, and triggers an alert and records the anomaly if the deviation exceeds a preset threshold. At the same time, the system compares the real-time status of the equipment with the status of the inventory records. If there are three consecutive inconsistencies, an anomaly event is confirmed. The system then extracts the contextual features of the anomaly event and inputs them into the anomaly attribution model to generate three probability values. Finally, the system executes the corresponding strategy based on the anomaly category with the highest probability, such as virtual cleanup, restoring residual tasks, or transferring goods to the anomaly storage location. This approach can detect potential problems in advance, avoid out-of-order task execution, dependency conflicts, and deadlock issues, effectively prevent concurrent tasks from competing for process resources, significantly optimize the Crane's movement path, reduce overall warehouse energy consumption, and greatly shorten the anomaly handling cycle. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0042] Figure 1 This is a flowchart of an STK data self-verification and anomaly self-healing method based on device status comparison proposed in this invention. Detailed Implementation

[0043] Example 1, referring to Figure 1 A method for STK data self-verification and anomaly self-healing based on equipment status comparison, the specific steps of which are as follows:

[0044] Collect and preprocess real-time status data of each device, and calculate the health score and status entropy value of each device based on the processed status data.

[0045] Specifically, real-time status data of each device within the storage equipment STK is collected, and real-time status data that does not meet preset standards is removed. The logical relationships between the real-time status data are then checked; if no logical relationship exists, the corresponding real-time status data is removed. The remaining real-time status data are then uniformly mapped to the [0, 1] interval. Multiple statistical feature values ​​for each device are extracted from the processed real-time status data, and each extracted statistical feature value is converted into a single health score between 0 and 100%. All single health scores for each device are then multiplied by their corresponding preset weight coefficients, and all calculated products are summed to obtain the health score of the device. The scoring process involves retrieving the IO state change record table from the local database and extracting all state change events for each device within the past 10 seconds. These events are then arranged into a continuous device state sequence in chronological order. Based on the PLC's preset state coding rules for each device, each state is converted into a corresponding code. The total number of occurrences of different state codes in each device's state sequence is then counted, and the total number of occurrences of each state code is divided by the total length of the device's state sequence to obtain the probability of occurrence for that state. Finally, the probability of occurrence for each state is multiplied by the logarithm of that probability (base 2), and all products are summed and the negative number is taken to obtain the state entropy value for each device.

[0046] It should be further explained that the equipment in the storage device STK includes the conveying mechanism Crane, the interface Port, and the storage location; the real-time status data includes the frequency of equipment failure, task execution time, IO status changes, and communication quality.

[0047] It should be further explained that the specific formula for calculating the device's health score is as follows:

[0048]

[0049] In the formula, This indicates the health score of the device; Indicates feature index; Indicates the first Individual health scores for each feature; Indicates the first Preset weighting coefficients for each feature; Indicates the total number of features;

[0050] The specific formula for calculating the state entropy value of the device is as follows:

[0051]

[0052] In the formula, This represents the device's state entropy value; Indicates the total number of device states; State code The probability of occurrence; It represents the logarithm to the base 2.

[0053] Based on the calculated health scores and state entropy values ​​of each device, it is determined whether to trigger the self-healing process and dynamic task avoidance.

[0054] Specifically, the system reads the health score and state entropy value of each device and compares them with preset warning thresholds, fault thresholds, and instability thresholds. If the health score is lower than the preset fault threshold, a fault alarm and emergency shutdown are triggered. If the health score is higher than or equal to the preset warning threshold and the state entropy value is lower than the preset instability threshold, the device maintains normal operation. If the health score is lower than the preset warning threshold but higher than the preset fault threshold, a self-healing process is triggered. If the state entropy value is higher than or equal to the preset instability threshold, dynamic avoidance is triggered. When the self-healing process is triggered, the corresponding device is first added to the built-in device status table. The status is marked as sub-healthy, and a temporary interception rule is added to the task allocation entry point of the built-in task scheduler. Based on this rule, all new task requests targeting sub-healthy devices are intercepted. Simultaneously, the current task queue is retrieved and iterated through, filtering out all tasks whose target devices are sub-healthy. These filtered tasks are then removed and transferred to a preset temporary scheduling queue. All devices of the same type with a normal status are queried, and multiple factors for each normal device are obtained. These are then sorted according to preset criteria, and the top-ranked device is selected as a replacement. If a usable replacement device exists, the task is... The target device for the task is updated to the selected alternative device, the task execution path is recalculated, and the task is added to the normal task scheduling queue. Conversely, the target location of the task is updated to a preset dedicated buffer storage location, the goods are temporarily stored in the buffer area, the accounting data of all sub-healthy devices is extracted from the built-in local material management database, and the integrity and consistency of each extracted accounting data are verified. At the same time, based on each accounting data that passes the verification, a corresponding accounting snapshot file is generated. Subsequently, the current task status of each sub-healthy device is monitored in real time. When a sub-healthy device completes its current task, no further tasks are assigned to it. A new task is assigned, and the device is set to maintenance mode. At the same time, no new tasks are assigned to the device. When dynamic avoidance is triggered, all new tasks pointing to devices with state entropy values ​​higher than the preset instability threshold are marked as being under observation and placed in a preset queue of tasks to be observed. All tasks in the queue of tasks to be observed are suspended for a preset backoff time. During the backoff time, tasks pointing to other stable devices are scheduled first. After the backoff time ends, the latest state entropy value of the device is obtained. If the latest state entropy value is lower than the preset instability threshold, the tasks in the queue of tasks to be observed are added back to the normal scheduling queue.

[0055] Example 2, refer to Figure 1 A method for STK data self-verification and anomaly self-healing based on equipment status comparison, the specific steps of which are as follows:

[0056] Construct a task dependency chain graph, mark the priority of all tasks according to the task dependency chain graph, and then determine whether to send a resource reservation signal to the corresponding target device.

[0057] Specifically, the system listens to production orders issued by the upper-layer MCS / MES system, parses the complete process route information in the production orders, and extracts multi-type data and sequence constraints for each task. Then, based on the parsed task sequence constraints, it constructs a task dependency chain graph, continuously monitors the status changes of all tasks, and updates the task dependency chain graph accordingly. It obtains the status of all tasks and marks tasks whose status meets preset criteria as executable tasks. Simultaneously, it obtains the start and end devices of each executable task and checks whether these devices are in a preset list of highest-priority devices. If both the start and end devices of any executable task are in this list, the task is marked as a highest-priority task, and resource reservation is triggered. The process involves extracting the corresponding target device from each highest priority task and sending a resource reservation signal to the target device's PLC. The status of the corresponding highest priority task is then updated to "resource reserved." When the highest priority task with reserved resources is scheduled, a task execution instruction is sent to the target device. Upon receiving the task execution instruction, the target device's PLC verifies whether the task ID and workstation number in the instruction match the current reservation information. If they match, the reservation flag is cleared, and task execution begins; otherwise, task execution is refused, and a reservation mismatch error is reported. When the target device completes the task, the corresponding task's status is updated to "completed," and a reserved resource release instruction is sent to the target device. Upon receiving the instruction, the target device's PLC clears any remaining reservation flags for the corresponding workstation, restoring the workstation to a usable state.

[0058] Obtain the multi-type characteristics of all candidate storage sites, calculate the selection suitability score for each candidate storage site, and then select the optimal storage site based on the calculation results.

[0059] Specifically, during task scheduling and execution, task execution conditions are continuously monitored. When the target device is detected to be in a sub-optimal state and no alternative device is available, a temporary storage location selection process is triggered. Simultaneously, the real-time status information of all storage locations is read from the built-in local storage location management database, excluding storage locations with abnormally locked status. Storage locations whose distance from the Crane location of the currently executing task exceeds the preset maximum allowable distance are then excluded. All remaining storage locations are then integrated to generate a candidate storage location list. Feature data for each storage location is extracted from this list, and this data is input into a pre-trained score inference model in a preset batch. The model calculates suitability scores for each candidate storage location through forward propagation. Finally, the suitability scores for each candidate storage location are... The data is sorted from highest to lowest, and the top-ranked storage location is selected as the target temporary storage location for this task. The selected storage location is marked as reserved. All Crane numbers, task queue lengths, identifiers of goods to be stored, and target storage location numbers are obtained. Based on the obtained data, a task execution instruction is generated and sent to the corresponding Crane controller. After receiving the execution instruction, the Crane controller collects the Crane's operating data in real time, monitors the IO status of the target storage location and the operating status of peripheral devices, and compares the real-time Crane's operating data with the preset task plan. The deviation between the actual task progress and the preset task plan is calculated. If the deviation exceeds a preset threshold, a progress warning is triggered and the abnormal event is recorded.

[0060] It should be further explained that the specific formula for calculating the task progress deviation is as follows:

[0061]

[0062] In the formula, Indicates a deviation in task progress; Indicates the actual execution time of the task; Indicates the scheduled execution time of the task.

[0063] It should be further noted that the feature data includes the physical distance of the storage site, the task popularity of the area where the storage site is located, the path congestion coefficient, and the number of tasks expected to arrive in the future.

[0064] It should be further explained that the specific training steps for the fractional reasoning model are as follows:

[0065] Complete execution data of all temporary tasks within 3 months were collected, along with auxiliary data for each temporary task. All collected data were then integrated to form the original training dataset. Data exceeding three times the normal range standard deviation in the original training dataset were removed, and the remaining data were mapped to the [0, 1] interval. The processed original training dataset was then randomly divided into training, validation, and test sets in a 7:2:1 ratio. Based on the essential characteristics of the STK storage location selection task, embedded hardware resource constraints, and industrial real-time requirements, a fractional inference model was constructed. The Xavier uniform distribution method was used to initialize the model weights, and mean squared error was selected as the loss function. Adam optimization was then employed. The algorithm updates the parameters, then sets the initial learning rate, β1 parameter, and β2 parameter. The training set data is input into the score inference model in preset batches for iterative training. In each training batch, forward propagation is performed first to obtain the model's predicted fitness score. Then, the mean squared error loss function between the predicted fitness score and the actual label is calculated. Next, the gradient of the mean squared error loss function with respect to the weights and biases of each layer is calculated using the backpropagation algorithm. Then, the model parameters are updated according to the gradient by the Adam optimizer. After each complete training set iteration, the model is evaluated using validation set data, and the mean squared error on the validation set is calculated. When the validation set loss does not decrease for 20 consecutive iterations, training stops, and the training of the score inference model is completed.

[0066] Obtain the status data of each device's inventory records, compare it with the corresponding real-time status data, and then trigger the exception handling process based on the comparison results.

[0067] Specifically, the system retrieves the status data recorded for each device from the built-in local material management database, and simultaneously retrieves the real-time status data for each device. The real-time status data is then compared with the recorded status data for each device to check for inconsistencies. If inconsistencies are found, the status of the same device is repeatedly verified in the next three comparison cycles. If the same inconsistency is shown in three consecutive comparisons, an anomaly is confirmed. The system then extracts the context feature vectors for each anomaly and sequentially inputs them into the input layer of the trained anomaly attribution model. The model generates three probability values ​​at the output layer through forward propagation and performs a summation check to ensure the sum of the probabilities is 1. The three probability values ​​are then sorted from highest to lowest, and the anomaly category with the highest probability value is selected as the anomaly determination. Finally, based on the anomaly determination, the corresponding preset treatment is selected. The handling strategy is as follows: If the sensor alarm is determined to be a false alarm, the strategy is to perform a virtual clearing operation and not generate a physical handling task; if the task is determined to be residual goods, the strategy is to perform association and release operations, while attempting to find and complete the residual task; if the unknown intrusion is determined to be an intrusion, the strategy is to perform security isolation and manual intervention operations, and transfer the goods to a dedicated abnormal storage location. Based on the selected handling strategy, the corresponding operations are performed. For sensor false alarms, the corresponding record in the local material management database is updated, the book material status is modified to match the actual equipment status, and then the abnormal lock mark is cleared; for task residual goods, the task database is searched for an incomplete task record that matches the location and goods. If found, the task execution is resumed; otherwise, the goods are moved to the designated normal storage location; for unknown intrusions, a dedicated abnormal handling task is generated, an idle Crane is scheduled to move the goods to a preset abnormal isolation storage location, and the storage location is locked.

[0068] It should be further explained that the contextual feature data includes device type features, recent fault records, IO status fluctuation features, and task execution history data.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A method for STK data self-verification and anomaly self-healing based on equipment status comparison, characterized in that, The specific steps of this self-healing method are as follows: Ⅰ: Collect and preprocess the real-time status data of each device, and calculate the health score and status entropy value of each device based on the processed status data; II: Based on the calculated health scores and state entropy values ​​of each device, determine whether to trigger the self-healing process and dynamic task avoidance; Ⅲ: Construct a task dependency chain graph, mark the priority of all tasks according to the task dependency chain graph, and then determine whether to send a resource reservation signal to the corresponding target device; IV: Obtain the multi-type characteristics of all candidate storage sites, calculate the suitability score for each candidate storage site, and then select the optimal storage site based on the calculation results; the specific steps are as follows: S4.1: During the task scheduling and execution process, continuously monitor the task execution conditions. When it is detected that the target device of the task is in a sub-healthy state and there is no available alternative device, trigger the temporary storage location selection process. At the same time, read the real-time status information of all storage locations from the built-in local storage location management database, and exclude storage locations with abnormal lock status. Then exclude storage locations whose distance from the Crane location of the currently executing task exceeds the preset maximum allowable distance, and then integrate all remaining storage locations to generate a list of candidate storage locations. S4.2: Extract the feature data of each storage location from the list of candidate storage locations, and input the feature data of each candidate storage location into the trained score inference model according to a preset batch. The model generates the suitability score of each candidate storage location through forward propagation calculation. Then, sort the suitability scores of each candidate storage location from high to low, select the storage location ranked first as the target temporary storage location for this task, and mark the selected storage location as reserved. The feature data includes the physical distance of the storage location, the task popularity of the area where the storage location is located, the path congestion coefficient, and the number of tasks expected to arrive in the future. S4.3: Obtain all Crane numbers, task queue length, identifiers of goods to be stored, and target storage location numbers. Generate task execution instructions based on the acquired data and send them to the corresponding Crane controller. After receiving the execution instructions, the Crane controller collects the operation data of the Crane itself in real time, monitors the IO status of the target storage location and the operation status of peripheral devices, and compares the real-time collected operation data of the Crane itself with the preset task plan. Calculate the deviation between the actual task progress and the preset task plan. If the deviation exceeds the preset threshold, a progress warning is triggered and abnormal events are recorded. V: Obtain the status data of each device's inventory records, compare it with the corresponding real-time status data, and then trigger the exception handling process based on the comparison results.

2. The STK data self-verification and anomaly self-healing method based on equipment status comparison according to claim 1, characterized in that, The specific steps for calculating the health score and state entropy value of each device in step I are as follows: S1.1: Collect real-time status data of each device in the storage equipment STK, and remove real-time status data that does not meet the preset standards. Then check whether the logical relationship between each real-time status data is valid. If the logical relationship is not valid, remove the corresponding real-time status data. Then map the remaining real-time status data to the interval [0, 1]. Each device in the storage equipment STK includes the conveying mechanism Crane, the interface Port, and the storage position. The real-time status data includes the frequency of equipment failure, task execution time, IO status changes, and communication quality. S1.2: Extract multiple statistical feature values ​​for each device from the processed real-time status data, and convert each extracted statistical feature value into a single health score between 0 and 100%. Then, multiply all single health scores of each device by the corresponding preset weight coefficient, and sum all the calculated products to obtain the health score of the device. S1.3: Obtain the IO state change record table in the local database, and extract all state change events of each device in the past 10 seconds. Arrange all state change events into a continuous device state sequence according to the time sequence. Then, according to the preset state coding rules of each device PLC, convert each state into a corresponding code. Then, count the total number of times different state codes appear in each device state sequence, and divide the total number of times each state code appears by the total length of the device state sequence to obtain the probability of occurrence of the corresponding state. Then, multiply the probability of occurrence of each state by the logarithm of the probability to the base 2, and then add all the products and take the negative number to obtain the state entropy value of each device.

3. The STK data self-verification and anomaly self-healing method based on equipment status comparison according to claim 2, characterized in that, The specific steps for determining whether the self-healing process and dynamic task avoidance are triggered in step II are as follows: S2.1: Read the health score and state entropy value of each device, and compare them with the preset warning threshold, fault threshold, and instability threshold respectively. If the health score is lower than the preset fault threshold, trigger a fault alarm and emergency shutdown; if the health score is higher than or equal to the preset warning threshold and the state entropy value is lower than the preset instability threshold, maintain the normal operation of the device; if the health score is lower than the preset warning threshold but higher than the preset fault threshold, trigger the self-healing process. If the state entropy value is higher than or equal to the preset instability threshold, dynamic avoidance is triggered. S2.2: When the self-healing process is triggered, the corresponding device status is marked as sub-healthy in the built-in device status table, and a temporary interception rule is added at the task allocation entry of the built-in task scheduler. Then, according to the added temporary interception rule, all new task requests pointing to sub-healthy devices are intercepted. At the same time, the current task queue waiting to be executed is obtained and traversed, all tasks whose target devices are sub-healthy devices are filtered out, and all filtered tasks are removed and then transferred to the preset temporary scheduling queue. S2.3: Query all devices of the same type that are in normal status, obtain the multi-type factors of each normal device, sort them according to preset criteria, select the device ranked first as the replacement device, if there is a usable replacement device, update the target device of the task to the selected replacement device, recalculate the task execution path, and add the task to the normal task scheduling queue; otherwise, update the target location of the task to the preset dedicated buffer storage location and temporarily store the goods in the buffer area. S2.4: Extract the accounting data of all sub-healthy devices from the built-in local material management database, and perform integrity and consistency verification on each extracted accounting data. At the same time, based on the accounting data that passes the verification, generate the corresponding accounting snapshot file. Then, monitor the current task status of each sub-healthy device in real time. When the sub-healthy device completes the current task, no new tasks will be assigned to it, the device will be set to the maintenance mode, and no new tasks will be issued to the device. S2.5: When dynamic avoidance is triggered, all new tasks pointing to devices with state entropy values ​​higher than the preset instability threshold are marked as pending observation and placed in the preset pending observation task queue. All tasks in the pending observation task queue are suspended for a preset backoff time. During the backoff time, tasks pointing to other stable devices are scheduled first. After the backoff time ends, the latest state entropy value of the device is obtained. If the latest state entropy value is lower than the preset instability threshold, the tasks in the pending observation task queue are added back to the normal scheduling queue.

4. The STK data self-verification and anomaly self-healing method based on equipment status comparison according to claim 1, characterized in that, The specific steps for determining whether to send a resource reservation signal to the corresponding target device in step III are as follows: S3.1: Listen to the production orders issued by the upper-level MCS / MES system, then parse the complete process route information in the production orders, and extract the multi-type data and sequence constraints of each process task from them. Then, based on the parsed task sequence constraints, construct a task dependency chain graph, while continuously monitoring the status changes of all tasks, and updating the task dependency chain graph according to the status changes of all tasks. S3.2: Obtain the status of all tasks and mark tasks whose status meets the preset criteria as executable tasks. At the same time, obtain the starting device and ending device of each executable task and check whether the starting device and ending device of each executable task are in the preset highest priority device list. If the starting device and ending device of any executable task are in the preset highest priority device list, mark the task as the highest priority task and trigger the resource reservation process. Then, extract the corresponding target device from each highest priority task and send a resource reservation signal to the PLC of the target device. Then update the status of the corresponding highest priority task to "resource reserved". S3.3: When the highest priority task with reserved resources is scheduled, a task execution instruction is sent to the target device. At the same time, after receiving the task execution instruction, the PLC of the target device verifies whether the task ID and station number in the task execution instruction match the current reserved information. If they match, the reserved flag is cleared and the task execution begins; otherwise, the task execution is refused and a reservation mismatch error is indicated. Subsequently, when the target device completes the task, the status of the corresponding task is updated to completed, and a reserved resource release instruction is sent to the target device. After receiving the instruction, the PLC of the target device clears any residual reserved flags of the corresponding station and restores the station to an available state.

5. The STK data self-verification and anomaly self-healing method based on equipment status comparison according to claim 4, characterized in that, The specific training steps for the fractional reasoning model described in S4.2 are as follows: P1.1: Collect complete execution data of all temporary tasks within 3 months, and simultaneously collect auxiliary data for each temporary task. Then integrate all the collected data to form the original training dataset. Subsequently, remove data in the original training dataset that exceeds 3 times the normal range standard deviation, and map the remaining data in the original training dataset to the interval [0, 1]. Then, randomly divide the processed original training dataset into training set, validation set and test set in a ratio of 7:2:

1. P1.2: Based on the essential characteristics of the STK storage location selection task, the constraints of embedded hardware resources, and the real-time requirements of industry, a fractional inference model is constructed. At the same time, the Xavier uniform distribution method is used to initialize the model weights, and the mean squared error is selected as the loss function. Then, the Adam optimizer is selected for parameter update, and the initial learning rate, β1 parameter, and β2 parameter are set. P1.3: The training set data is input into the score inference model in preset batches for iterative training. In each training batch, forward propagation is performed first to obtain the model's predicted fitness score. Then, the mean squared error loss function between the predicted fitness score and the actual label is calculated. Next, the gradient of the mean squared error loss function with respect to the weights and biases of each layer is calculated using the backpropagation algorithm. Subsequently, the model parameters are updated according to the gradient using the Adam optimizer. After each complete training set iteration, the model is evaluated using the validation set data, and the mean squared error on the validation set is calculated. When the validation set loss does not decrease for 20 consecutive iterations, training is stopped, and the training of the score inference model is completed.

6. The STK data self-verification and anomaly self-healing method based on equipment status comparison according to claim 2, characterized in that, The specific steps for triggering the exception handling process based on the comparison results, as described in step V, are as follows: S5.1: Retrieves the status data of each device from the built-in local material management database, and retrieves the real-time status data of each device. It then compares the real-time status data of each device with the recorded status data one by one to check for inconsistencies. If inconsistencies are found, the status of the same device is repeatedly verified in the next three comparison cycles. If the results of three consecutive comparisons show the same inconsistent status, an abnormal event is confirmed to have occurred. Subsequently, the context feature vector of each abnormal event is extracted. The context feature data includes device type features, recent fault records, IO status fluctuation features, and task execution history data. S5.2: Input all extracted context feature vectors sequentially into the input layer of the trained anomaly attribution model. The model generates three probability values ​​in the output layer through forward propagation and performs a summation check to ensure that the sum of the probability values ​​is 1. At the same time, sort the three probability values ​​from high to low and select the anomaly category with the highest probability value as the judgment result of this anomaly. Then, based on the judgment result of this anomaly, select the corresponding preset processing strategy. If it is determined to be a sensor false alarm, the processing strategy is to perform a virtual clearing operation and not generate a physical handling task; if it is determined to be residual goods, the processing strategy is to perform association and release operations, and at the same time attempt to find and complete the residual task; if it is determined to be an unknown intrusion, the processing strategy is to perform security isolation and manual intervention operations, and transfer the goods to a dedicated anomaly storage location. S5.3: Based on the selected processing strategy, execute the corresponding operations. For sensor false alarms, update the corresponding record in the local material management database, modify the material status to match the actual equipment status, and then clear the abnormal lock mark. For task-related residual goods, search the task database for an incomplete task record that matches the location and goods. If found, resume task execution. Otherwise, move the goods to the designated normal storage location. For unknown intrusions, generate a dedicated abnormal handling task, schedule an idle Crane to move the goods to the preset abnormal isolation storage location, and lock the storage location.

Images