A production process status monitoring and scheduling optimization method based on real-time data acquisition

Abstract

This invention discloses a production process status monitoring and scheduling optimization method based on real-time data acquisition, belonging to the field of manufacturing process scheduling technology, to solve the problem of insufficient real-time performance and stability in process scheduling. This invention constructs a process monitoring mechanism based on real-time acquisition and closed-loop scheduling, forming time-series structured data frames under a unified time base. It extracts operational stability and quality deviation features, generates data credibility labels and current process status factor vectors, and inputs them into an optimized scheduling model to identify task conflicts and resource bottlenecks. Combining scheduling priority mapping and resource path graphs, it implements path compression and sequence adjustment. During operation, it constructs scheduling deviation vectors based on task response delays and process interruptions, triggering online updates of rules and parameters and local rearrangements, which are then solidified after verification within the prediction window. This reduces equipment idling and switching fragmentation during production, improving the real-time performance, resource utilization, and system stability of the production process.

Description

Technical Field

[0001] This invention relates to the field of manufacturing process scheduling technology, and more specifically, to a method for optimizing the scheduling of production process status based on real-time data acquisition. Background Technology

[0002] In discrete manufacturing industries that primarily involve multiple product types and small batches (such as electronic assembly lines, injection molding processing lines, and precision hardware parts workshops), there are frequent switching and resource competition among various processes. The production scheduling system needs to make high-frequency production scheduling decisions based on multi-source information such as equipment operating status, processing cycle time, and quality inspection feedback.

[0003] The shortcomings of existing technologies: Current production data processing is mostly based on the arrival time transmitted from equipment or the control layer for scheduling logic judgment, lacking a unified event time benchmark and reliable labeling mechanism. For example, on an electronic surface mount production line, the same component placement task is executed on two placement machines of different brands. Although the equipment logs show "completed on time", due to the inconsistent event time, the system judges that the two tasks are in the correct order and the process is correct. In fact, subsequent quality inspection shows that the welding defect rate has soared. The scheduling system cannot identify such data risks of out-of-order and late reporting, nor can it perform compensation and rescheduling based on event trends. In addition, the current scheduling system lacks the ability to continuously observe and analyze trends such as task response delay, sudden increase in resource occupation, and process blockage. Even if there is a clear bottleneck trend (such as the continuous increase in the queue length of a certain machine), the system is difficult to form an effective rule update and local rescheduling mechanism and can only rely on manual intervention. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the following solution is proposed to solve the problem of insufficient real-time performance and stability of process scheduling in the background art.

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

[0006] A production process status monitoring and scheduling optimization method based on real-time data acquisition includes the following steps:

[0007] S1. Collect equipment operating status parameters, process execution parameters, and quality inspection feedback parameters deployed at key process locations on the production line, and perform data calibration based on a unified time to construct a time-series structured data frame sequence;

[0008] S2. Extract the operational stability features, resource load features, and quality offset features of each process from the time-series structured data frame sequence. Combine the timeliness and completeness of the collected data to construct a reliable data label and generate the current process state factor vector. The operational stability features are used to reflect the frequency of equipment state switching and the duration of operation. The resource load features are used to reflect the trend of changes in the equipment's resource load. The quality offset features are used to reflect the fluctuation range of the pass rate in the quality inspection feedback data and the presence of abnormal quality state segments.

[0009] S3. Input the current process status factor vector and resource allocation parameters into the scheduling optimization model, identify nodes with task conflicts or resource bottlenecks, generate the initial task scheduling sequence in combination with the scheduling priority mapping rules, and construct a resource path graph to perform path compression and sequence adjustment operations.

[0010] S4. During the scheduling execution process, continuously collect scheduling response behavior data, extract task response latency, sudden increase in resource consumption and process interruption events, construct scheduling deviation vector and classify deviation types;

[0011] S5. Based on the changing trends of the scheduling deviation vector and the current process state factor vector, update the scheduling priority mapping rules and resource allocation parameters, implement the rearrangement of scheduling paths, and form an adaptive scheduling strategy with a closed loop of state and response.

[0012] Further, step S1 includes:

[0013] Using the production line's main clock as a unified time reference, clock calibration is performed on each data acquisition node;

[0014] A sliding time window based on event time is used to align the collected data. When the deviation between the timestamp of the collected data and the unified time base exceeds the preset deviation threshold, time synchronization correction is triggered and the affected data is reordered according to event time.

[0015] Construct data frames in a time-series structured data frame sequence. Each data frame must contain at least the event timestamp, process number, equipment identifier, task label, parameter type, and parameter value fields.

[0016] Furthermore, the operational stability characteristics, resource load characteristics, and quality offset characteristics of each process are extracted from the time-series structured data frame sequence. Specific steps include:

[0017] A start-stop state sequence is constructed based on the equipment's operating status parameters. The state switching frequency, continuous running time, and abnormal shutdown segments are detected, and the detection results are summarized as operational stability characteristics.

[0018] Based on the collected equipment uptime, task queue length and resource utilization, the resource load change trend is calculated as a resource load feature.

[0019] The fluctuation range of the pass rate is calculated based on continuous quality inspection feedback data, and quality status segments with abnormal changes are identified as quality deviation features.

[0020] Furthermore, a reliable data label is constructed by combining the timeliness and completeness of the collected data, and a current process status factor vector is generated. Specific steps include:

[0021] The event timestamp of each data frame is compared with a unified time base. When the difference between the arrival time and the event timestamp is not greater than a preset first threshold, it is judged as timely. When the difference is greater than the preset first threshold but not greater than a preset second threshold, it is judged as delayed. When the difference is greater than the preset second threshold, it is judged as expired.

[0022] When time disorder is detected within the sliding time window, an disorder marker is added to form a timeliness marker;

[0023] Perform a completeness check on the essential fields of the data frame, and verify the continuity of the process sequence and the consistency of cross-device constraint parameters within the current sliding time window, obtaining a completeness result of qualified, missing, or conflicting;

[0024] According to the decision rules, the timeliness mark and the completeness result are combined for judgment. If the timeliness is timely and the completeness is qualified, a high-confidence label is generated. If there is a lag or non-critical field missing, a medium-confidence label is generated. If it is expired, out of order, or there is a cross-device conflict, a low-confidence label is generated.

[0025] Under high-confidence or medium-confidence labeling conditions, interval mapping and threshold segmentation are performed on the operational stability characteristics, resource load characteristics and quality offset characteristics respectively, and the current process state factor vector is generated by connecting them in a fixed-dimensional order.

[0026] When the trusted label is medium trusted, only the feature dimensions that have passed the consistency check are included, and the rest are marked as placeholders.

[0027] When the trust label is low trust, the feature does not update the vector dimension with a new value in the current cycle. Instead, it uses the most recent valid value of the same dimension that passed the consistency check in the previous sliding time window as a placeholder, and binds the data trust label with the current process state factor vector for output.

[0028] Furthermore, the current process state factor vector and resource allocation parameters are input into the scheduling optimization model to identify nodes with task conflicts or resource bottlenecks, including the following steps:

[0029] Load the current process state factor vector and resource allocation parameters into the scheduling optimization model. Determine the start conditions and necessary prerequisite processes for each task based on the current process state factor vector. Determine the available equipment set, resource parallelism limit, minimum switching interval and minimum effective processing time based on the resource allocation parameters.

[0030] Based on the pre- and post-task relationships and the available resource time window, the start and end time intervals of the plan are derived for each task. The time intervals are determined by the intersection of the expected processing time of the task and the corresponding available resource time, and a task time interval table is generated.

[0031] On each resource dimension, the task time interval table is checked for overlap. When the time intervals of two or more tasks overlap on the same resource and the resource parallel restriction does not allow them to be occupied at the same time, the corresponding task is marked as a task conflict node. When the resource allows parallelism but needs to meet the minimum switching interval or the minimum effective processing period and cannot be met at the same time, it is also marked as a task conflict node.

[0032] Within the sliding time window, based on resource allocation parameters, the duration of unmet resource requests, the growth trend of queue length, and the degree of fragmentation of idle time caused by switching are statistically analyzed. When any statistical result continuously exceeds the preset judgment condition, the corresponding time period of the resource is marked as a resource bottleneck node.

[0033] The identified task conflict nodes and resource bottleneck nodes are used as the constraint outputs of the scheduling optimization model.

[0034] Furthermore, based on resource allocation parameters, within the sliding time window, the duration of unmet resource requests, the trend of queue length growth, and the degree of idle time fragmentation caused by switching are statistically analyzed, including the following steps:

[0035] The duration of unmet resource requests is calculated as follows: for each resource request, the continuous waiting time from registration to the first allocation is recorded as the unmet duration. The maximum value is accumulated within the sliding time window. When the accumulated value or the maximum value exceeds the preset condition, it is marked as a candidate bottleneck.

[0036] The queue length growth trend is determined by sampling the queue length at fixed intervals, comparing adjacent sampled values, statistically analyzing the length of continuously growing segments and comparing the difference between the start and end of the sliding time window. If any of these values ​​exceeds a preset condition, a continuous growth trend is determined.

[0037] The assessment of the fragmentation of idle time is as follows: extract the idle interval of resources, record the intervals shorter than the minimum effective processing time as fragments, count the number and the proportion of the total idle time, and determine that idle time fragments are formed within the sliding time window when the preset conditions are exceeded.

[0038] Furthermore, an initial task scheduling sequence is generated by combining scheduling priority mapping rules, and a resource path graph is constructed to perform path compression and order adjustment operations, including the following steps:

[0039] The task set is topologically sorted according to the scheduling priority mapping rules and combined with pre- and post-dependencies to obtain an executable task set. Tasks are selected sequentially from the executable task set to generate the initial task scheduling sequence. When there are ties in priority, they are resolved by the earliest start time and the shortest preparation time in turn.

[0040] The earliest start time refers to the start time of the feasible time interval obtained by the intersection of the available time window of resources corresponding to the available equipment set and the expected processing time of the task.

[0041] The estimated preparation time refers to the sum of the switching time required to switch from the initial execution path to the current task on the target device corresponding to the candidate insertion position and the process parameter calibration time.

[0042] A resource path graph is constructed using devices and buffers as nodes and reachability and switching relationships as edges, with each edge having switching time interval and minimum switching interval attributes. The initial sequence of task scheduling is then mapped to the initial execution path.

[0043] Path compression is implemented on the initial execution path, including moving available idle segments forward without violating the minimum switching interval and minimum effective processing time, and merging adjacent tasks of the same type on the same device to reduce the cumulative switching time when adjacent tasks are executed by the same device and the controlled vocabulary is marked as continuously processable.

[0044] For adjacent tasks on the same resource that have additional switching overhead, perform partial swapping or reordering. If the minimum switching interval and minimum effective processing time are met and no new conflicts are introduced, the order adjustment is adopted; otherwise, it reverts to the original order.

[0045] Further, step S4 includes:

[0046] Calculate the task response delay. For each task, take the time difference between the actual start time and the planned start time as the response delay. If the task has not started within the current sliding time window, the time difference between the end time of the sliding time window and the planned start time is included in the delay.

[0047] To detect sudden increases in resource usage, the usage rate sequence of the same resource is read at a fixed sampling interval. The segment where the difference between adjacent sampling values ​​continuously exceeds a preset change threshold is recorded as a sudden increase segment, and the length of the longest sudden increase segment and the peak increase are recorded.

[0048] Perform process interruption event identification. When an executable task enters a waiting state due to failure to meet prerequisite dependencies or unavailability of target resources, and the waiting time is not less than the minimum effective processing time, it is recorded as an interruption event. The number of events and the cumulative duration are counted.

[0049] Deviation vectors are constructed and classified. Task response latency indicators, resource consumption surge indicators, and blocking event indicators are connected in a fixed order to generate scheduling deviation vectors. Based on the duration, recurrence interval, and whether blocking events are included, the deviation types are classified as persistent, burst, periodic, or blocking.

[0050] Furthermore, based on the changing trends of the scheduling deviation vector and the current process state factor vector, the scheduling priority mapping rules and resource allocation parameters are updated, and a local rearrangement of the scheduling path is implemented, including the following steps:

[0051] Perform sliding time window trend analysis on the scheduling deviation vector. When any component shows a monotonically increasing or periodically recurring trend within a continuous sliding time window, and occurs simultaneously with an anomaly in the corresponding dimension of the current process status factor vector, trigger rule and parameter updates.

[0052] Rule and parameter updates refer to adjusting the task categories associated with the triggering node according to a preset upgrade and downgrade relationship in the scheduling priority mapping rules, and performing range adjustments on resource allocation parameters according to the deviation type. Range adjustments include at least one of tightening parallelism, extending the minimum switching interval, or expanding the available time window.

[0053] The triggering node is either a task conflict node or a resource bottleneck node;

[0054] In the resource path graph, the adjacent range is selected with the trigger node as the center. Under the condition of not violating the pre- and post-dependencies and resource constraints, the task is moved forward, inserted into the gap, or partially swapped to form a rearranged local execution path.

[0055] Furthermore, the expected latency and conflict count are calculated for the rearranged local execution paths within the prediction window. If any metric does not decrease, the rules and parameters before the update are reverted. If the metric decreases, the change is fixed and the updated task sequence is output.

[0056] The technical effects and advantages of the production process status monitoring and scheduling optimization method based on real-time data acquisition of this invention are as follows:

[0057] This invention achieves dynamic optimization in the production process by constructing a process status monitoring mechanism based on real-time acquisition and closed-loop scheduling. The system collects equipment operating status parameters, process execution parameters, and quality inspection feedback parameters at key process locations, and calibrates them based on a unified time reference to form a traceable time-series structured data frame sequence. On this basis, it extracts operating stability features, resource load features, and quality offset features, and generates reliable data labels by combining timeliness and completeness, and constructs a process status factor vector to ensure the completeness and reliability of status modeling.

[0058] The scheduling optimization model uses process status factor vectors and resource allocation parameters to identify task conflict nodes and resource bottleneck nodes in real time. It generates an initial scheduling sequence by combining scheduling priority mapping rules and performs path compression and sequence adjustment on the resource path graph to reduce equipment idling and switching overhead. During execution, it continuously collects task response latency, sudden increases in resource consumption, and process interruption events, constructs and classifies scheduling deviation vectors, and triggers rule and parameter updates when deviations correspond to abnormal trends in process status factor vectors, and performs local rearrangement in the neighborhood of the trigger node.

[0059] By verifying the rearrangement effect within the prediction window and solidifying the effective adjustments, production line fluctuations can be quickly detected, and task priorities and resource allocation can be adjusted online to avoid long-term bottlenecks or the spread of conflicts, thus significantly improving the real-time performance of task scheduling, resource utilization, and the stability of the production system. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a production process status monitoring and scheduling optimization method based on real-time data acquisition, according to the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] In order to achieve the above objectives, Figure 1 A schematic diagram of a production process status monitoring and scheduling optimization method based on real-time data acquisition according to the present invention is provided, which specifically includes the following steps;

[0063] S1. Collect equipment operating status parameters, process execution parameters, and quality inspection feedback parameters deployed at key process locations on the production line, and perform data calibration based on a unified time to construct a time-series structured data frame sequence;

[0064] S2. Extract the operational stability features, resource load features, and quality offset features of each process from the time-series structured data frame sequence. Combine the timeliness and completeness of the collected data to construct a reliable data label and generate the current process state factor vector. The operational stability features are used to reflect the frequency of equipment state switching and the duration of operation. The resource load features are used to reflect the trend of changes in the equipment's resource load. The quality offset features are used to reflect the fluctuation range of the pass rate in the quality inspection feedback data and the presence of abnormal quality state segments.

[0065] S3. Input the current process status factor vector and resource allocation parameters into the scheduling optimization model, identify nodes with task conflicts or resource bottlenecks, generate the initial task scheduling sequence in combination with the scheduling priority mapping rules, and construct a resource path graph to perform path compression and sequence adjustment operations.

[0066] S4. During the scheduling execution process, continuously collect scheduling response behavior data, extract task response latency, sudden increase in resource consumption and process interruption events, construct scheduling deviation vector and classify deviation types;

[0067] S5. Based on the changing trends of the scheduling deviation vector and the current process state factor vector, update the scheduling priority mapping rules and resource allocation parameters, implement the rearrangement of scheduling paths, and form an adaptive scheduling strategy with a closed loop of state and response.

[0068] Step 1: Collect equipment operating status parameters, process execution parameters, and quality inspection feedback parameters deployed at key process locations on the production line, and perform data calibration based on a unified time frame to construct a time-series structured data frame sequence. The specific implementation is as follows:

[0069] On a discrete manufacturing production line containing multiple critical process locations, a data acquisition node is deployed at each critical process location to collect equipment operating status parameters, process execution parameters, and quality inspection feedback parameters. The data acquisition node accesses a time synchronization service through the production line communication network. The production line master clock serves as a unified time reference, provided by a high-precision time source and periodically broadcast throughout the network. Each data acquisition node performs clock calibration during system power-on initialization, performs calibration verification at fixed time intervals during operation, and immediately performs on-demand calibration when the deviation between the local clock and the unified time reference exceeds a preset deviation threshold. After clock calibration is completed, the data acquisition node generates an event timestamp locally for each collected raw record. The event timestamp is used to indicate the actual time when the parameter is collected or generated on-site, which is different from the arrival time when the record is received by the network or written to storage.

[0070] To avoid time drift caused by the accumulation of local timing errors from different device controllers, the acquisition node will mark the original records generated between the last calibration and the current calibration as potentially affected data after each on-demand calibration.

[0071] During the data alignment phase, a sliding time window is used to align records from multiple acquisition nodes based on the event time. The length of the sliding time window is set to be no less than the cycle time of the key process and an integer multiple thereof, and the sliding step size is set to be no greater than the length of the sliding time window to ensure that there are coverage areas between windows to accommodate late data.

[0072] Each time the window is swiped, the event timestamps of all records within the current window are checked. When the difference between the event timestamp of any record and the unified time reference exceeds a preset deviation threshold, time synchronization correction is triggered. The preset deviation threshold is used to determine the allowable deviation between the local clock of the acquisition node and the unified time reference. Time synchronization correction includes two parts: first, performing on-demand calibration on the acquisition node that generated the record; and second, reordering the affected records by event time. The handling of late data follows the following strategy:

[0073] Late data arriving within the tolerance interval of the sliding time window is merged into the current window for alignment. Late data arriving after the tolerance interval is played back into the window. Within the playback range, it is inserted into the corresponding historical window sequence according to the event time. If the historical window has been fixed and playback is not allowed, the late data is marked as expired data and enters the abnormal record queue, which is then handled by a dedicated remediation process according to the rules. For time disorder detected within the same sliding time window, the sliding time window is marked as disordered and the records in the window are rearranged according to the event time during the data frame construction stage. For dense records caused by high-frequency acquisition within the minimum switching interval or minimum effective processing period, they are retained in chronological order of the event time and are not merged or discarded.

[0074] It should be noted that the tolerance interval is the time range for accepting late data. This range is set during the deployment phase based on the cycle time of key processes, network transmission latency, and storage persistence strategy, and remains unchanged during operation with the same naming convention.

[0075] During the data frame construction phase, records that have undergone time alignment and necessary reordering are used as input to generate a sequence of time-series structured data frames. Each data frame contains at least the following fields: event timestamp, process number, equipment identifier, task tag, parameter type, and parameter value. Their meanings are as follows:

[0076] The event timestamp uniquely identifies the point in time when the data occurred on-site; the process number indicates the specific process to which the data belongs, using a unique process number code in the process routing; the equipment identifier identifies the specific equipment or tooling that generated the data, using a unique equipment identifier code in the production line asset ledger; the task tag associates the data with the production task being executed, with each task tag corresponding one-to-one with the work order or batch identifier in the manufacturing execution system; the parameter type distinguishes between three main categories: equipment operating status parameters, process execution parameters, and quality inspection feedback parameters, as well as their subcategories. A controlled thesaurus defines the value range and meaning of parameter types to ensure that parameters with the same name maintain consistent semantics across different processes; the parameter value carries the actual measurement value or judgment result and is stored in the unit of measurement corresponding to the parameter type.

[0077] When generating a time-series structured data frame sequence, a uniqueness check is performed on combinations of the same event timestamp, the same process number, the same equipment identifier, and the same task tag. When duplicates are found, the parameter types and parameter values ​​are compared to see if the content is completely consistent.

[0078] When records are completely consistent, they are considered duplicate reports of the same event and only one record is retained. When records are inconsistent, the entire set of records is written to provide input for subsequent consistency verification processes. Finally, the time-series structured data frame sequence is output in order of event timestamps, ensuring that accurate retrieval can be performed at any time based on process number, equipment identifier, and task tag. It also provides consistent, complete, and traceable input for extracting operational stability features, resource load features, and quality offset features from the time-series structured data frame sequence, as well as for constructing reliable data tags by combining the timeliness and completeness of the collected data and generating the current process status factor vector.

[0079] Step 2: Extract the operational stability features, resource load features, and quality offset features of each process from the time-series structured data frame sequence. Combine the timeliness and completeness of the collected data to construct a reliable data label and generate a current process state factor vector. Operational stability features reflect the frequency of equipment state switching and runtime; resource load features reflect the trend of equipment resource load changes; and quality offset features reflect the fluctuation range of the pass rate and the presence of abnormal quality state segments in the quality inspection feedback data. The specific implementation is as follows:

[0080] After obtaining the time-series structured data frame sequence, the data within the same sliding time window are aggregated according to the process number, equipment identifier, and task label to form a window sample for feature extraction. The original fields for feature extraction include event timestamp, process number, equipment identifier, task label, parameter type, and parameter value. The parameter type is limited to three categories: equipment operating status parameters, process execution parameters, and quality inspection feedback parameters.

[0081] The preset first threshold and preset second threshold used for threshold determination are set during the deployment phase based on the process cycle time, equipment sampling frequency and network latency. The preset first threshold is used to distinguish between timely and delayed arrival, and the preset second threshold is used to distinguish between delayed and expired arrival. The preset first threshold and preset second threshold are used to classify and determine the timeliness of data arrival as timely, delayed and expired. The two types of thresholds are respectively applicable to clock synchronization and data arrival determination, and are semantically independent and do not replace each other.

[0082] Operational stability characteristics reflect the frequency of equipment state switching and runtime; resource load characteristics reflect the trend of equipment resource load changes; quality offset characteristics reflect the fluctuation range of pass rate and the presence of abnormal quality state segments in quality inspection feedback data. The specific steps for obtaining these characteristics are as follows:

[0083] The extraction of operational stability features is based on constructing a start-stop state sequence using equipment operating status parameters. First, equipment operating status parameters corresponding to the same process number and equipment identifier are selected from the time-series structured data frame sequence. Records representing the equipment's running and stopping states are sorted by event timestamps to obtain the start-stop state sequence. Subsequently, three types of detection are performed on the start-stop state sequence: First, the detection of state switching frequency involves counting the number of transitions between "running to stopping" and "stopping to running" within the current sliding time window, and comparing this number with the allowed number of state switching times for the process. When the number of transitions falls within the lower, middle, or upper range of the allowed range, it is marked as low switching frequency, medium switching frequency, and high switching frequency, respectively, such as 0-5. The first switch is the lower interval, 5-10 switches are the middle interval, and the remaining switches are the upper interval; secondly, the detection of continuous running time refers to finding the longest continuous running segment in the start-stop state sequence and comparing it with the minimum effective processing time required by the process. When the longest continuous running segment is shorter than, close to or longer than the minimum effective processing time, it is marked as short continuous running, medium continuous running, and long continuous running, respectively; thirdly, the detection of abnormal shutdown segments refers to identifying the start and end times of shutdown. When the duration of a single shutdown exceeds the planned maintenance allowable range or occurs during unplanned maintenance periods, the segment is marked as an abnormal shutdown segment, and the number and cumulative duration of abnormal shutdown segments are recorded. The above three detection results together constitute the operational stability characteristics.

[0084] Resource load characteristics are extracted based on three observables: device uptime, task queue length, and resource utilization. Device uptime is obtained by accumulating the total time the device is in operation within the current sliding time window and comparing it with the total time of the sliding time window. When the ratio falls into a preset low, medium, or high range, it is marked as low uptime, medium uptime, and high uptime, respectively. Task queue length is obtained by reading the queue length sequence of the same resource at fixed sampling intervals and comparing adjacent sample values. When most adjacent sample values ​​show an increasing trend and the difference between the queue length at the start and end of the window is positive and exceeds the queue growth judgment threshold, it is marked as an increasing trend. When adjacent sample values ​​alternate between increasing and decreasing and the difference between the start and end points is in the middle range of the queue growth judgment threshold, it is marked as a stable trend. When most adjacent sample values ​​show a decreasing trend and the difference between the start and end points is negative and exceeds the queue decrease judgment threshold, it is marked as a decreasing trend.

[0085] The resource utilization rate is obtained by statistically analyzing the percentage of sampling points where resources are occupied at fixed sampling intervals within the current sliding time window, and marking them as low resource utilization rate, medium resource utilization rate, and high resource utilization rate according to the preset low, medium, or high intervals. The queue length trend category, operation time category, and resource utilization rate category are output as resource load features.

[0086] The extraction of quality deviation features is based on the correspondence between quality inspection feedback parameters and quality inspection batches. Continuous quality inspection feedback data with the same process number and task label are integrated. First, the number of qualified and unqualified records is calculated within the current sliding time window. When both qualified and unqualified records exist, the ratio of the qualified quantity to the total inspection quantity is used as the pass rate for the current window. To obtain the pass rate fluctuation range, the current sliding time window is divided into at least two consecutive time segments, and the pass rate within each time segment is calculated. When the difference in pass rates between different time segments is within a preset low range... When the range is in the middle or high range, it is marked as low fluctuation amplitude, medium fluctuation amplitude, and high fluctuation amplitude, respectively. If the quality inspection feedback parameter is a continuous measurement value, the measurement values ​​of adjacent inspection records are compared with the process tolerance as a reference. When the measurement value continuously exceeds the upper or lower limit of the process tolerance in a short period of time, or when a continuous transition from qualified to unqualified state occurs, the continuous record is marked as an abnormal change quality state segment and its start and end time and the number of batches involved are recorded. The three items of qualified rate fluctuation category, abnormal change existence mark, and abnormal change location description are output as quality offset features.

[0087] The generation of data reliability tags and current process status factor vectors follows a unified timeliness and integrity decision-making rule. The timeliness determination process is as follows:

[0088] For each data frame, an event timestamp and arrival time are recorded simultaneously (both represented using a unified time base). The data is compared with the unified time base. If the difference between the data arrival time and the event timestamp is not greater than a preset first threshold, it is considered timely. If the difference is greater than the preset first threshold but not greater than a preset second threshold, it is considered delayed. If the difference is greater than the preset second threshold, it is considered expired. If the event timestamp does not satisfy the condition of monotonically decreasing within the same sliding time window, the window is marked with a time disorder mark.

[0089] The integrity determination process is as follows: First, check whether each data frame has the event timestamp, process number, device identifier, task label, parameter type, and parameter value fields. If any one is missing, it is recorded as missing. Second, verify the continuity of the process sequence within the current sliding time window. If there is a jump in the process number or repeated rollback, it is recorded as discontinuous. Third, perform consistency verification on the constraint parameters across devices. When the same task label appears on mutually exclusive device identifiers at the same time, it is recorded as conflict.

[0090] Generate data reliability labels based on the timeliness and completeness results according to the following rules:

[0091] A high-confidence label is generated when the timeliness is timely and the completeness is qualified; a medium-confidence label is generated when there is lag or missing non-critical fields; and a low-confidence label is generated when there is expiration, time disorder, or cross-device conflict. Subsequently, under high-confidence or medium-confidence label conditions, interval mapping and threshold segmentation are performed on the operational stability feature, resource load feature, and quality offset feature, respectively. The interval mapping method involves writing the category labels obtained above into vector positions in a fixed-dimensional order, where the fixed-dimensional order is the sequential connection of the operational stability feature sub-dimensional set, the resource load feature sub-dimensional set, and the quality offset feature sub-dimensional set. When the data confidence label is a medium-confidence label, only the sub-dimensions that have passed the consistency check are written into the vector, and the remaining vector positions are filled with placeholders marked "not accepted." When the data confidence label is a low-confidence label, no new vector is generated in the current period; instead, the valid value from the previous period (i.e., the most recent valid value that passed the consistency check in the same dimension in the previous sliding time window is used as the placeholder) or the entire value is marked "not accepted." Simultaneously, the data confidence label is bound to the current process state factor vector for output.

[0092] A concrete example is used to illustrate the implementation of generating trusted data tags and the current process state factor vector from a time-series structured data frame sequence. The process is as follows:

[0093] Within the same sliding time window, data with process number A1, equipment identifier M1, and task tag T-001 arrives. The event timestamp of each data frame is read and compared with a unified time benchmark. Most records show a difference of no more than a preset first threshold (determined as timely), a small number of records show a difference between the preset first threshold and a preset second threshold (determined as delayed), and no records exceed the preset second threshold (no expired). At the same time, no time disorder is detected in this sliding time window (no disorder marker). In terms of integrity, all data frames contain "event timestamp, process number, equipment identifier, task tag, parameter type, and parameter value", and the process sequence within this sliding time window is continuous, and the cross-equipment constraint parameters are consistent (no missing or conflicting data). Based on this, a high-confidence label is generated according to the decision rules.

[0094] Subsequently, three types of features were extracted: operational stability features (the start-stop state sequence shows that the frequency of state switching is in the middle range of the allowable range, the longest continuous operation segment is not shorter than the minimum effective processing time, and there are no abnormal downtime segments); resource load features (equipment uptime is in the middle range, the queue length increases slightly from the start to the end of the sliding time window but does not exceed the growth judgment threshold, and the resource utilization rate is in the middle range); and quality deviation features (the difference in the pass rate of continuous quality inspection feedback between different time segments is in the low range, and no abnormal sudden quality status segments are found). Under the condition of high-confidence labeling, the above category labels and discrete descriptions such as event counts are mapped to intervals and segmented by thresholds according to a predetermined fixed dimension order, and written bit by bit to generate the current process state factor vector. Since the data confidence label is a high-confidence label, the vector does not need to occupy the "unaccepted" position.

[0095] Step 3: Input the current process state factor vector and resource allocation parameters into the scheduling optimization model, identify nodes with task conflicts or resource bottlenecks, generate an initial task scheduling sequence based on scheduling priority mapping rules, and construct a resource path graph to perform path compression and sequence adjustment operations. The specific implementation is as follows:

[0096] After receiving the current process status factor vector and resource allocation parameters, the scheduling optimization model first performs task executability and resource availability determination. The task executability determination is performed by verifying the startable conditions corresponding to the prerequisite dependency satisfaction, operational stability characteristics and quality deviation characteristics in the current process status factor vector. Any task that does not simultaneously meet the requirements of the completion of the necessary prerequisite processes and the establishment of the startable conditions is marked as unexecutable and removed from the current scheduling round.

[0097] Resource availability is determined based on resource allocation parameters, which define the set of available devices, resource parallelism limits, minimum switching interval, and minimum effective processing time for each task. Simultaneously, the start and end boundaries of the resource availability time window are read for subsequent task feasibility calculation. Then, based on the task sequence relationships and the resource availability time window, the model derives the planned start and end time intervals for each task.

[0098] First, determine the expected processing time period of the task, and then find the intersection with the corresponding available time window of the resource to obtain the feasible time interval. When the intersection is empty, the task is marked as unschedulable in the current loop. All feasible time intervals are summarized into a task time interval table with tasks as rows and resources as columns, which serves as the input for conflict and bottleneck identification.

[0099] At each resource dimension, the scheduling optimization model performs overlap checks on the task time interval table to identify conflicting task nodes, specifically:

[0100] When the feasible time intervals of two or more tasks on the same resource overlap and the resource parallelism limit does not allow them to occupy the resource at the same time, mark the group of tasks as task conflict nodes one by one.

[0101] When resource parallelism limits allow parallel usage but require both minimum switching interval and minimum effective processing time to be met, and after local sorting it is still impossible to leave time no less than the minimum switching interval between adjacent tasks or to provide continuous processing time no less than the minimum effective processing time for any task, the relevant tasks will also be marked as task conflict nodes.

[0102] After completing the identification of task conflict nodes, the scheduling optimization model performs three statistics on resource operation status within the sliding time window to identify resource bottleneck nodes: First, it counts the duration of unmet resource requests, that is, it accumulates the continuous waiting time from the registration of each resource request to the completion of the first allocation, and records the maximum continuous waiting time within the sliding time window. When the accumulated waiting time or the maximum continuous waiting time exceeds the preset judgment condition, the resource corresponding to the time period is marked as a candidate bottleneck.

[0103] Second, the queue length growth trend is determined by reading the queue length sequence at a fixed sampling interval, comparing adjacent sample values ​​point by point to count the length of the continuous growth segment, and comparing it with the queue length difference between the start and end points of the sliding time window. If any result exceeds the preset judgment condition, it is determined that there is a continuous growth trend.

[0104] Third, the fragmentation degree of idle time is assessed. This involves extracting idle intervals on the resource timeline, marking idle intervals shorter than the minimum effective processing period as fragments, and counting the number of fragments and their proportion of the total idle time. When the number or proportion exceeds a preset judgment condition, the degree of fragmentation of idle time is judged to be high. If any of the above three statistics are met, the resource is marked as a resource bottleneck node in the corresponding time period.

[0105] The scheduling optimization model aggregates task conflict nodes and resource bottleneck nodes into scheduling constraint outputs.

[0106] When generating the initial sequence of task scheduling, the scheduling optimization model performs topological sorting of the task set according to the scheduling priority mapping rules and the pre- and post-dependencies to first obtain the set of executable tasks; then, tasks are selected in sequence from the set of executable tasks according to the scheduling priority mapping rules to form the initial sequence of task scheduling.

[0107] When there are equal priorities, the scheduling optimization model uses the earliest start time and the shortest preparation time as the resolution conditions in turn until a definite selection order is obtained. For example, when there are equal priorities, the earliest start time is given priority for resolution. If they are still equal, the expected preparation time is given priority. The expected preparation time is the sum of the minimum switching time and the necessary parameter adjustment time in the set of available equipment for the task, under the conditions of resource parallelism and minimum switching interval.

[0108] It should be noted that the earliest start time is the start time of the feasible time interval obtained by intersecting the available resource time window corresponding to the available devices for the task and the expected processing time of the task, under the conditions of satisfying pre- and post-dependencies, resource parallelism constraints, minimum switching interval, and minimum effective processing period. When there are multiple available devices, the smallest start time among the feasible time intervals is taken as the earliest start time of the task. When the target device has no preceding task at the candidate position, the earliest start time is still determined by the start time of the intersection interval. The estimated preparation time is the time taken on the target device corresponding to the candidate insertion position. The estimated preparation time is the sum of the switching time required to switch from the preceding task in the initial execution path to this task and the process parameter adjustment time. The switching time is determined based on the switching time attribute of the corresponding edge in the resource path graph, and the process parameter adjustment time is determined based on the settings of the differences in relevant process execution parameters in the controlled vocabulary. When there are multiple available devices or multiple candidate insertion positions, the minimum of the above sums is taken as the estimated preparation time for this task, provided that the resource parallelism limit and the minimum switching interval are met. When there is no preceding task for the target device at the candidate position, the switching time is recorded as zero, and the estimated preparation time is determined only by the process parameter adjustment time.

[0109] To accommodate equipment switching and logistics connectivity constraints, the model constructs a resource path graph with equipment and buffers as nodes and reachability and switching relationships as edges. Edges are accompanied by switching time and minimum switching interval attributes, representing the time and minimum interval required to switch from one task to the next. The scheduling optimization model maps the initial task scheduling sequence onto the resource path graph to obtain the initial execution path. If the initial execution path overlaps with task conflict nodes or resource bottleneck nodes in the aforementioned scheduling constraint output, the relevant tasks are marked as candidate tasks requiring priority adjustment without modifying the definitions of pre- and post-dependencies and resource parallelism constraints. This provides a target set for subsequent path compression and sequence adjustment.

[0110] During the path compression phase, the scheduling optimization model, under the premise of not violating pre- and post-dependencies, resource parallelism constraints, minimum switching interval, and minimum effective processing time, performs two types of operations on the initial execution path: First, it moves the available idle segment forward, that is, when there is a continuous idle time in front of the target task that meets the minimum switching interval and minimum effective processing time, the task is moved forward to the nearest available time position to shorten the waiting time; Second, it merges adjacent tasks of the same type on the same device, that is, when adjacent tasks are executed by the same device and the parameter type and process execution parameters allow continuous processing, unnecessary gaps in the middle are removed to reduce the accumulation of switching time.

[0111] After path compression is completed, the scheduling optimization model attempts to adjust the order of adjacent tasks on the same resource that have additional switching overhead: if the minimum switching interval, minimum effective processing time and no new conflict are simultaneously satisfied through local swapping or insertion rearrangement, the order adjustment is adopted; if any of the conditions are not met, the original order is rolled back and remains unchanged. The task sequence after path compression and order adjustment and its mapping result on the resource path graph are used as the output of this step.

[0112] Step 4: Continuously collect scheduling response behavior data during the scheduling execution process, extract task response latency, sudden increases in resource consumption, and process interruption events, construct a scheduling deviation vector, and classify the deviation types. The specific implementation is as follows:

[0113] After obtaining the initial task scheduling sequence and resource path map, the system continuously records scheduling response behavior data using a sliding time window as the collection and statistics unit. For each task, the system reads the planned start time from the initial task scheduling sequence and the actual start time and resource occupancy trajectory from the execution layer log. For each resource, an occupancy rate sequence is generated at a fixed sampling interval, and the occupancy rate of a single sample is the percentage of time the resource is occupied within that sampling interval. For each process, the system records the waiting state caused by unmet prerequisites or unavailable target resources.

[0114] It should be noted that all of the above records include event timestamps, process numbers, equipment identifiers, and task tags to ensure consistency with the time-series structured data frame sequence and traceability.

[0115] The task response latency calculation process is as follows: For each task, the system uses the time difference between the actual start time and the planned start time as the task response latency. When the task has not started within the current sliding time window, the time difference between the end time of the sliding time window and the planned start time is temporarily included in the task response latency to characterize the delay status within the window. Two types of descriptive values ​​are retained at the same time: one type marks the task response latency as short latency, medium latency, or long latency according to the preset latency interval boundary for subsequent classification and quick judgment; the other type is the longest latency and cumulative latency associated with the task label, used for continuous and trend comparisons between multiple windows. If the task is paused and restarted multiple times, the system uses the first actual start time to participate in the task response latency calculation and reflects the subsequent impact from pause to restart in resource consumption surges and process interruption events to avoid double counting.

[0116] The detection process for sudden increases in resource occupancy is as follows: the occupancy rate sequence of the same resource is read at a fixed sampling interval, and the difference between adjacent sampling values ​​is calculated point by point. When the difference between adjacent sampling values ​​continuously exceeds the preset change threshold and continuously reaches the preset continuous counting threshold, the continuous segment is marked as an occupancy surge segment, and the start and end sampling points, the longest continuous length, and the maximum increase in occupancy rate in the segment are recorded. The preset change threshold and continuous counting threshold are set during the deployment phase based on the resource switching time interval, minimum switching interval, and process cycle time. To avoid false alarms caused by brief fluctuations, before marking the occupancy surge segment, it is required that the resource availability time window in the segment has not undergone any planned artificial contraction, and there are no short-term concurrent peaks caused by path compression in the segment. If these exist, the segment is recorded as a planned increase in occupancy and removed from the surge statistics.

[0117] The process for identifying process interruption events is as follows: When an executable task, after satisfying its pre- and post-dependent dependencies and having the planned start-up time in the initial task scheduling sequence, still enters a waiting state due to the unavailability of the target resource, and the waiting time is not less than the minimum effective processing period, this waiting is recorded as a process interruption event and bound to the corresponding task tag, equipment identifier, and time range. When an interruption event spans multiple sliding time windows, it is merged according to the continuity of the event occurrence. The cumulative duration and number of events are only calculated at the end of the event, and a placeholder registration is performed in each involved sliding time window to ensure the continuity of subsequent trend determination.

[0118] For waiting caused by unmet prerequisites, only when the waiting continues due to resource unavailability after the prerequisites are met and the minimum effective processing time is reached, it is counted as a process blocking event, thus distinguishing between dependency waiting and blocking caused by resource bottlenecks.

[0119] The process of constructing and classifying scheduling deviation vectors is as follows:

[0120] The task response latency index, resource usage surge index, and process interruption event index are connected in a fixed order to generate a scheduling deviation vector. The task response latency index includes three items: latency category, longest latency, and cumulative latency; the resource usage surge index includes three items: whether there is a surge in usage, the length of the longest surge segment, and the peak increase; and the process interruption event index includes two items: the number of events and the cumulative duration.

[0121] When classifying, the judgment is based on three categories of rules:

[0122] First, persistent deviations are defined as follows: when any component in the scheduling deviation vector remains in the same abnormal state within a continuous sliding time window and reaches the threshold of the number of consecutive windows, it is marked as persistent.

[0123] Secondly, the sudden deviation is defined as follows: when the scheduling deviation vector is abnormal within a single sliding time window but recovers to the normal range within two adjacent sliding time windows, it is marked as sudden.

[0124] Third, the determination of periodic deviation is that when the anomaly of the scheduling deviation vector recurs at approximately equal intervals in multiple sliding time windows, and the recurrence interval is within the preset recurrence interval tolerance range, it is marked as periodic.

[0125] When a process blocking event occurs within any sliding time window, it is simultaneously marked as blocking to indicate the direct impact of resource unavailability on the execution path. Finally, the classification results are bound together with the scheduling deviation vector, the initial task scheduling sequence, and the resource path graph as input for updating the scheduling priority mapping rules and resource allocation parameters, as well as implementing local reordering of scheduling paths.

[0126] Step 5: Based on the changing trends of the scheduling deviation vector and the current process state factor vector, update the scheduling priority mapping rules and resource allocation parameters, implement the rearrangement of scheduling paths, and form an adaptive scheduling strategy with a closed loop of state and response. Specifically, the implementation is as follows:

[0127] During scheduling execution, the sliding time window is used as the analysis unit to determine the trend of the scheduling deviation vector and to perform linkage verification with the current process status factor vector: when any component in the scheduling deviation vector shows a monotonically increasing trend within a continuous sliding time window (i.e., the degree of abnormality of the component does not decrease in adjacent windows) or repeats at approximately equal intervals (periodic recurrence, with the interval falling within the preset recurrence interval tolerance range), and the dimension corresponding to the semantics of the component in the current process status factor vector shows an anomaly within the same time span (e.g., resource load characteristics show continuous high occupancy, operation stability characteristics show high switching frequency, quality offset characteristics show high fluctuation), it is determined that the trigger update condition is met; then, the trigger node is located on the resource path graph. The trigger node is a task conflict node or a resource bottleneck node, and the trigger time, the involved task tags, and the equipment identifier are recorded;

[0128] After the triggering condition is met, the system executes rule and parameter updates, and adjusts the task category associated with the triggering node according to the preset promotion and demotion relationship based on the scheduling priority mapping rule:

[0129] If the deviation type is persistent or periodic and is triggered by a resource bottleneck node, the priority of the affected task category will be raised to a higher level so that it will be placed earlier in the subsequent selection order;

[0130] If the deviation type is sudden and is triggered by a task conflict node, the priority of this type of task will only be increased within a local time range, and the original level will be automatically restored outside this range.

[0131] If the deviation is related to quality offset characteristics and is not caused by resource bottlenecks, then the task will be downgraded to free up capacity.

[0132] For resource allocation parameters, adjustments are made in intervals and take effect within the adjacent range of the trigger node:

[0133] When excessive parallelism causes congestion, the parallelism is reduced to a smaller range; when frequent switching leads to fragmentation, the minimum switching interval is extended to a larger range.

[0134] When insufficient available time causes a blockage, the start and end boundaries of the available time window are expanded to provide continuous processing time.

[0135] After updating the rules and parameters, a local rearrangement is performed in the resource path graph, centered on the trigger node, selecting the adjacent range.

[0136] The adjacency range is determined using a dual constraint approach: first, a hop count constraint on the graph structure, selecting candidate tasks only within a limited number of hops adjacent to the trigger node; and second, a local window constraint on time, attempting to rearrange tasks that overlap or are adjacent to the trigger node in time.

[0137] The order of operations for local rearrangement is as follows:

[0138] Prioritize moving tasks forward. When there is a continuous free interval ahead of the target task that meets the minimum switching interval and minimum effective processing time, move it forward to the position with the nearest available time.

[0139] If moving forward is not feasible, try inserting into gaps, inserting shorter-duration tasks into gaps that meet the minimum switching interval and do not generate new conflicts; if neither of the above is feasible, try local swapping, swapping adjacent tasks on the same resource that have additional switching overhead in sequence, provided that the swapping still meets the requirements of pre- and post-dependencies, resource parallelism limits, minimum switching interval and minimum effective processing time, and does not introduce new task conflict nodes or resource bottleneck nodes.

[0140] Each reordering attempt generates a candidate local execution path and retains the scheduling priority mapping rules and resource allocation parameter version numbers on which the candidate path was generated, facilitating backtracking and rollback.

[0141] To verify the effect of local reordering, a rapid simulation was performed on each candidate local execution path within a set prediction window. The prediction window was set as a short time range for quickly evaluating the effect of local reordering, and was set according to the process cycle time and scheduling response delay during the deployment phase.

[0142] Calculate two metrics: estimated latency and collision count. The estimated latency is calculated as follows:

[0143] Within the prediction window, the estimated delay of a single task is obtained by subtracting the planned start time of that task in the initial sequence of task scheduling from the estimated start time of each task in the candidate local execution path. The estimated delays of the single tasks of related tasks are then combined and statistically analyzed within the prediction window to obtain the estimated delay of the candidate path.

[0144] The collision count is calculated as follows:

[0145] Within the prediction window, candidate local execution paths are arranged along a resource timeline. The overlapping relationships of feasible time intervals for tasks on the resource dimension and the satisfaction of minimum switching interval and minimum effective processing time are checked one by one. Any violation of parallel constraints, switching interval, or effective processing time is recorded as a conflict event and accumulated as a conflict count.

[0146] The expected latency and conflict count of the candidate local execution path are compared with the expected latency and conflict count of the corresponding range before the change. If either indicator does not decrease (i.e. remains unchanged or increases), the candidate path is deemed invalid and reverts to the scheduling priority mapping rules and resource allocation parameters before the update. If both indicators decrease, the candidate path is deemed valid and the change is solidified. The updated task sequence, scheduling priority mapping rules, and resource allocation parameters are used as the new effective version, and the corresponding resource path graph mapping results are output.

[0147] To ensure the continuous operation of the closed loop, the basis for this strategy evaluation is recorded after solidification or rollback, including the trigger node type, deviation type, involved task tags and device identifiers, the version of the scheduling priority mapping rule and resource allocation parameter adopted, the category of local reordering operation, and the comparison results of two indicators in the prediction window. This record, together with the scheduling deviation vector and the current process status factor vector, is archived for trend determination and re-triggering in subsequent sliding time windows. The archived record, together with the scheduling deviation vector and the current process status factor vector, serves as the trend analysis input in the next sliding time window, ensuring the continuous operation of the closed loop from collection to identification to adjustment to evaluation and back to collection.

[0148] For example, in the placement process on an electronic assembly line, the placement machine identified as M2 experienced task delays within multiple consecutive sliding time windows. The scheduling deviation vector showed that the task response delay component continuously increased, the resource usage surge component recurred periodically, and a process interruption event also occurred. Simultaneously, the resource load characteristics in the current process status factor vector also showed high utilization and a continuously increasing queue length trend. These two aspects showed an abnormal dimensional correspondence within the same time period, thus the system determined that the triggering condition was met.

[0149] The system marks M2 as a resource bottleneck node during this time period. In the scheduling priority mapping rules, the priority of the critical task categories bound to M2 is increased to ensure that they are executed earlier in the subsequent scheduling sequence. At the same time, resource allocation parameters are adjusted, such as tightening parallelism and extending the minimum switching interval, to avoid fragmentation caused by frequent resource changes. Subsequently, in the resource path graph, with M2 as the trigger node, adjacent tasks are attempted to be moved to the nearest idle time period. If this is not feasible, short tasks are placed in suitable slots. If this is still not feasible, the order of adjacent tasks is partially swapped. Within the prediction window, the system calculates the expected latency and conflict count of the candidate paths. It finds that both of these indicators have decreased compared to the original. Therefore, this update is solidified, and the updated task sequence and rule parameters are output, forming a successful adaptive scheduling closed loop.

[0150] It should be noted that the threshold information in this embodiment was set in advance by professionals and will not be explained in detail here.

[0151] This invention achieves dynamic optimization in the production process by constructing a process status monitoring mechanism based on real-time acquisition and closed-loop scheduling. The system collects equipment operating status parameters, process execution parameters, and quality inspection feedback parameters at key process locations, and calibrates them based on a unified time reference to form a traceable time-series structured data frame sequence. On this basis, it extracts operating stability features, resource load features, and quality offset features, and generates reliable data labels by combining timeliness and completeness, and constructs a process status factor vector to ensure the completeness and reliability of status modeling.

[0152] The scheduling optimization model uses process status factor vectors and resource allocation parameters to identify task conflict nodes and resource bottleneck nodes in real time. It generates an initial scheduling sequence by combining scheduling priority mapping rules and performs path compression and sequence adjustment on the resource path graph to reduce equipment idling and switching overhead. During execution, it continuously collects task response latency, sudden increases in resource consumption, and process interruption events, constructs and classifies scheduling deviation vectors, and triggers rule and parameter updates when deviations correspond to abnormal trends in process status factor vectors, and performs local rearrangement in the neighborhood of the trigger node.

[0153] By verifying the rearrangement effect within the prediction window and solidifying the effective adjustments, production line fluctuations can be quickly detected, and task priorities and resource allocation can be adjusted online to avoid long-term bottlenecks or the spread of conflicts, thus significantly improving the real-time performance of task scheduling, resource utilization, and the stability of the production system.

[0154] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0155] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0156] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0157] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0158] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A production process state monitoring scheduling optimization method based on real-time data acquisition, characterized in that: The method comprises the following steps: S1, collecting equipment running state parameters, process execution parameters and quality inspection feedback parameters deployed at key process positions of the production line, calibrating data based on a unified time, and constructing a time-sequenced structured data frame sequence; S2, extracting running stability features, resource load features and quality deviation features of each process from the time-sequenced structured data frame sequence, constructing a data credibility label based on the timeliness and integrity of the collected data, and generating a current process state factor vector; the running stability features are used to reflect the switching frequency and running time of the equipment state; the resource load features are used to reflect the resource load change trend of the equipment; the quality deviation features are used to reflect the fluctuation amplitude of the qualified rate in the quality inspection feedback data and the abnormal quality state section; S3, inputting the current process state factor vector and the resource allocation parameter into a scheduling optimization model, identifying nodes with task conflicts or resource bottlenecks, generating an initial task scheduling sequence according to the scheduling priority mapping rule, and constructing a resource path graph to perform path compression and sequence adjustment operations; S4, continuously collecting scheduling response behavior data during scheduling execution, extracting task response time delay, resource occupation sudden increase and process blocking events, constructing a scheduling deviation vector and classifying the deviation types; S5, updating the scheduling priority mapping rule and the resource allocation parameter according to the change trend of the scheduling deviation vector and the current process state factor vector, rearranging the scheduling path, and forming a state and response closed-loop adaptive scheduling strategy; The data credibility label is constructed based on the timeliness and integrity of the collected data, and the current process state factor vector is generated, and the specific steps include: Comparing the event timestamp of each data frame with the unified time reference, when the difference between the arrival time and the event timestamp is not greater than a preset first threshold, it is determined as timely, when the difference is greater than the preset first threshold and not greater than a preset second threshold, it is determined as lag, and when the difference is greater than the preset second threshold, it is determined as expired; When time out-of-order is detected in the sliding time window, an out-of-order mark is marked, and the timeliness mark is formed; The completeness of the necessary fields of the data frame is checked, and the continuity of the process sequence and the consistency of the cross-device constraint parameters in the current sliding time window are verified, and the completeness result is qualified, missing or conflict; According to the decision rule, the timeliness mark and the completeness result are combined to determine, if the timeliness is timely and the completeness is qualified, a high credibility label is generated, if there is lag or missing of non-key fields, a medium credibility label is generated, and if it is expired, out-of-order or cross-device conflict occurs, a low credibility label is generated; Under the condition of high credibility or medium credibility label, interval mapping and threshold segmentation are performed on the running stability features, resource load features and quality deviation features, respectively, and the current process state factor vector is generated by connecting in the fixed dimension order; When the credibility label is medium, only the feature dimensions that pass the consistency check are included, and the rest are marked with placeholders. When the trust label is low, the feature does not update the vector dimension with a new value in the current cycle, and the last effective value of the same dimension in the last sliding time window is used as the placeholder through consistency verification, and the data trust label is bound with the current process state factor vector and output.

2. The production process state monitoring and scheduling optimization method based on real-time data collection according to claim 1, characterized in that: Step S1 comprises: Taking the production line master clock as the unified time reference, the clock of each collection node is calibrated; Using the sliding time window based on event time to align the collected data, when the deviation of the collection data timestamp from the unified time reference exceeds the preset deviation threshold, trigger time synchronization correction and reorder the affected data according to the event time; Constructing the data frame in the time-structured data frame sequence, each data frame at least contains event timestamp, process number, device identification, task label, parameter type and parameter value fields.

3. The production process state monitoring and scheduling optimization method based on real-time data collection according to claim 2, characterized in that: Input the current process state factor vector and resource allocation parameter into the scheduling optimization model to identify nodes with task conflicts or resource bottlenecks, including the following steps: Load the current process state factor vector and resource allocation parameter into the scheduling optimization model, determine the startable conditions and necessary pre-processes of each task according to the current process state factor vector, and determine the available device set, resource parallel restriction, minimum switching interval and minimum effective processing period according to the resource allocation parameter; Based on the task pre-post relationship and resource available time window, derive the plan start and end time interval for each task, and generate a task time interval table, the time interval is determined by the intersection of the task expected processing period and the corresponding resource available period; On each resource dimension, check the overlap of the task time interval table, when the time intervals of at least two tasks on the same resource overlap and the resource parallel restriction does not allow simultaneous occupation, mark the corresponding task as a task conflict node, when the resource allows parallel but needs to meet the minimum switching interval or minimum effective processing period and cannot meet at the same time, also mark it as a task conflict node; Within the sliding time window, based on the resource allocation parameter, respectively, the duration of unmet resource requests, the growth trend of queue length, and the fragmentation degree of idle time caused by switching are counted, when any of the statistical results continuously exceeds the preset judgment condition, the corresponding time period of the resource is marked as a resource bottleneck node; Output the identified task conflict nodes and resource bottleneck nodes as constraints of the scheduling optimization model.

4. The production process state monitoring and scheduling optimization method based on real-time data collection according to claim 3, characterized in that: According to the resource allocation parameter, the duration of unmet resource requests, the growth trend of queue length, and the fragmentation degree of idle time caused by switching are counted within the sliding time window, including the following steps: The duration of unmet resource requests is counted: for each resource request, the continuous waiting time from registration to the first allocation is recorded as the unmet duration, the maximum value is accumulated and recorded within the sliding time window, when the accumulated value or the maximum value exceeds the preset condition, it is marked as a candidate bottleneck; The queue length growth trend is determined: the queue length is sampled at fixed intervals, the adjacent sampling values are compared, the length of the continuous growth segment is counted and compared with the sliding time window start and end difference, when any of them exceeds the preset condition, it is determined that there is a continuous growth trend; The idle time fragmentation degree evaluation is evaluated as follows: extracting the resource idle interval, recording the interval shorter than the minimum effective processing period as fragmentation, counting the number and its proportion in the total idle time, and determining the idle time fragmentation in the sliding time window when the preset condition is exceeded.

5. The production process state monitoring and scheduling optimization method based on real-time data collection according to claim 4, characterized in that: The task scheduling initial sequence is generated according to the scheduling priority mapping rule, and a resource path graph is constructed to perform path compression and sequence adjustment operations, including the following steps: The task set is topologically sorted according to the scheduling priority mapping rule and combined with the pre-post dependency to obtain an executable task set, and the initial sequence of task scheduling is generated by sequentially selecting tasks from the executable task set, and when there are parallel priorities, the earliest start-up time and the shortest preparation time are used as conditions for elimination; The earliest start-up time refers to the start time of the feasible time interval obtained by the intersection of the available time window of the available device set and the expected processing period of the task; The expected preparation time refers to the sum of the switching time and the process parameter adjustment time required for switching from the initial execution path to the current task on the target device at the candidate insertion position; A resource path graph is constructed with devices and buffers as nodes and reachable and switching relationships as edges, and the switching time and minimum switching interval attributes are attached to the edges, and the initial execution path is mapped to the initial execution path; Path compression is performed on the initial execution path, including moving the available idle segment forward without violating the minimum switching interval and the minimum effective processing period, and when adjacent tasks are executed by the same device and the controlled word table is marked as continuous processing, adjacent tasks of the same type on the same device are merged to reduce the cumulative switching time. Step S4 includes:

6. The production process state monitoring and scheduling optimization method based on real-time data collection according to claim 5, characterized in that: The task response delay is calculated, for each task, the time difference between the actual start time and the planned start-up time is taken as the response delay, and when the task is not started in the current sliding time window, the time difference between the sliding time window end time and the planned start-up time is counted into the delay; Resource occupation sudden increase detection is performed, the occupation rate sequence of the same resource is read at a fixed sampling interval, the difference between adjacent sampling values is continuously greater than the preset variation threshold, and the section is recorded as a sudden increase section, and the longest sudden increase section length and peak amplitude are recorded; Process blocking event identification is performed, when the executable task enters the waiting state due to the unmet pre-dependency or the unavailability of the target resource, and the waiting time is not less than the minimum effective processing period, it is recorded as a blocking event, and the number of events and the cumulative time are counted; The bias vector is constructed and classified, the task response delay index, the resource occupation sudden increase index and the blocking event index are connected in a fixed order to generate a scheduling bias vector, and the bias types are classified into continuous, burst, periodic or blocking types according to the duration, repetition interval and whether the blocking event is included. According to the change trend of the scheduling bias vector and the current process state factor vector, the scheduling priority mapping rule and the resource allocation parameter are updated, and the local rearrangement of the scheduling path is implemented, including the following steps:

7. The production process state monitoring and scheduling optimization method based on real-time data collection according to claim 6, characterized in that: ​ A sliding time window trend analysis is performed on the scheduling bias vector. When any component is monotonically increasing or periodically recurring within a continuous sliding time window, and simultaneously appears with an anomaly in the corresponding dimension of the current process state factor vector, the rule and parameter update is triggered; The rule and parameter update refers to adjusting the level of the task category associated with the trigger node according to the preset promotion and demotion relationship in the scheduling priority mapping rule, and performing interval adjustment on the resource allocation parameters according to the bias type. The interval adjustment includes at least one of tightening the parallelism, extending the minimum switching interval, or expanding the available time window; The trigger node is a task conflict node or a resource bottleneck node; In the resource path graph, a neighbor range is selected centered on the trigger node. Without violating the pre-post dependency and resource constraints, the task is preferentially executed by moving forward, inserting a space, or locally exchanging to form a rearranged local execution path.

8. The production process state monitoring and scheduling optimization method based on real-time data collection according to claim 7, characterized in that: The predicted delay and conflict count are calculated for the rearranged local execution path within the set prediction window. When any indicator does not decrease, the rule and parameters before the update are returned to. When the indicators decrease, the current changes are fixed and the updated task sequence is output.

Images