Production management method based on artificial intelligence

Abstract

The application provides a production management method based on artificial intelligence, comprising the following steps: constructing a space-time correlation traceability graph based on a production network topology and taking time sequence characteristic values of micro-variable mark sequences as node and link weights according to the micro-variable mark sequences, determining a congestion diffusion conduction direction and a network topology boundary coordinate; obtaining a real-time delay of a production task transmission path, performing space-time dimension feature correlation on the congestion diffusion conduction direction, the network topology boundary coordinate and the real-time delay of the corresponding production task transmission path, and determining a congestion risk; and monitoring a path utilization rate in real time, if the path utilization rate breaks through a preset critical buffer threshold, switching a new task to a next available path in order, and obtaining a final task bearing scheme that adapts to production load demand.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a production management method based on artificial intelligence. Background Technology

[0002] In modern industrial production, AI-based dynamic scheduling systems have become a core technological support for improving resource utilization efficiency and addressing complex task demands. These systems utilize intelligent algorithms to optimize resource allocation paths in real time, ensuring efficient progress of multi-device collaborative tasks in highly dynamic and ever-changing production environments. However, traditional solutions primarily rely on real-time data collection, such as node waiting time and link occupancy rates, to select low-latency paths through threshold comparisons. But when tasks surge or multiple devices fiercely compete for resources, this data is often severely delayed due to the time lag in data collection and decision-making, failing to reflect the true state of instantaneous resource contention. This leads to the system repeatedly selecting already saturated paths, further amplifying congestion. Congestion is not an isolated event but rather a chain reaction propagating from core switches to peripheral devices through the coupling of fiber channel bandwidth saturation and node queue overflow. This propagation direction is highly dependent on real-time task flow patterns and network topology. When multiple parallel production lines operate at high loads simultaneously, congestion boundaries become blurred, expanding from a single link to the entire subnet, making it difficult for the system to distinguish between temporary fluctuations and persistent bottlenecks. Furthermore, when channel contention intensifies, queue waiting times no longer accumulate linearly but instead explode exponentially due to lock contention and retransmission mechanisms. For example, when the workload in a workshop exceeds the link bandwidth's capacity, the microsecond-level wait time of a single node quickly evolves into millisecond-level congestion, and this congestion spreads to adjacent production lines through the shared fiber optic channel. This causes the response speed of a group of originally independent equipment to drop from stable to unpredictable fluctuations within minutes. This dynamic uncertainty not only disrupts the production rhythm but also increases the risk of equipment failure, highlighting the forward-looking decision-making dilemma faced by scheduling systems in high-precision manufacturing environments. Summary of the Invention

[0003] This invention provides a production management method based on artificial intelligence, mainly including: The AI ​​feature extraction module collects queuing delay and data transmission link utilization data of production equipment nodes, extracts link jitter and queue depth change features, and obtains micro-variable marker sequences that characterize the precursors and propagation trends of link congestion. Based on the micro-variable marker sequence, a spatiotemporal correlation tracing map is constructed with the production network topology as the basis and its temporal characteristic value as the node and link weights to determine the congestion propagation path, the corresponding propagation direction, and the coordinates of the network topology boundary. The real-time latency of the production task transmission path is obtained, and the aforementioned transmission direction, boundary coordinates and real-time latency are correlated with spatiotemporal dimension features to determine the congestion risk. At the same time, the full set of alternative paths from the source node to the target node of the current production task is obtained. When the congestion risk exceeds the preset risk threshold, identify and lock the hidden bottleneck nodes on the congestion propagation path, generate a bottleneck exclusion list and update the path selection and load allocation logic of the production task scheduler; when the congestion risk does not exceed the preset risk threshold, keep the original logic of the scheduler unchanged. When the congestion risk does not exceed the preset risk threshold, the end-to-end delay value and the delay difference between paths are judged based on the full set of candidate paths to obtain the priority path ranking. When the congestion risk exceeds the preset risk threshold, hidden bottleneck nodes and associated links are excluded from the full set of candidate paths to obtain a subset of available candidate paths, and the priority path ranking is generated based on this subset. Based on the priority path ranking of the corresponding scenario, before the injection of sudden production tasks, the task traffic demand is obtained by combining historical production data and real-time link status through a preset AI prediction model, and tasks are pre-allocated to priority paths. Real-time monitoring of path utilization; if the preset critical buffer threshold is exceeded, new tasks will be switched to the next available path in the ranking, thus obtaining the final task handling scheme that adapts to the production load requirements.

[0004] Furthermore, the AI ​​feature extraction module collects queuing delay and data transmission link utilization data of production equipment nodes, extracts link jitter and queue depth change features, and obtains a micro-variable marker sequence characterizing the precursors and propagation trends of link congestion, including: Using the global synchronization clock as a reference, the AI ​​feature extraction module periodically collects queuing delay data of production equipment nodes, synchronously obtains the bandwidth occupancy ratio and instantaneous throughput of the data transmission link, and stores them in a node link synchronization sampling record table aligned with the timestamp. Based on the record table, the link jitter and queue depth change features are extracted. The link jitter feature is the dispersion of the difference in one-way transmission delay of data packets within adjacent sampling periods within a preset time window. The queue depth change feature is the increase or decrease of the node buffer occupancy within a continuous time window. When both features exceed the preset jitter threshold and the depth increase threshold respectively, it is marked as a congestion precursor interval. When neither feature exceeds the corresponding preset threshold, a micro-change marker sequence representing the stable operation state of the link is generated. For the congestion precursor interval, a sliding time window is used to perform time-series correlation operations on the two features. According to the positive and negative directions and the increment of the feature values ​​within the window, a micro-variation label sequence is generated. Each element of the sequence contains the feature trend label, amplitude value and timestamp information of the corresponding time window.

[0005] Furthermore, the step of constructing a spatiotemporal correlation tracing map based on the production network topology and using its temporal characteristic values ​​as node and link weights, according to the micro-variable marker sequence, to determine the congestion propagation path and corresponding propagation direction and network topology boundary coordinates, includes: Based on the micro-change marker sequence, the number of state switching times and the total amplitude change of the corresponding micro-change markers for each production equipment node are counted. The two are then used as time-series feature values ​​and mapped to the corresponding nodes in the physical topology of the production network to form a set of nodes labeled with time-series feature values. The number of state switching times is the cumulative number of positive and negative direction conversions of the feature trend markers within the preset observation period, and the total amplitude change is the cumulative sum of the absolute differences in amplitude values ​​between adjacent time windows during the same period. For the fiber channel connection relationship between adjacent nodes in the node set, the time difference of the occurrence of the micro-change mark between adjacent nodes is used as the transmission time difference, and the ratio of the mark amplitude before and after transmission is used as the amplitude attenuation ratio. The two are used as link weights and assigned to the corresponding connection edges. Combined with the node set, a spatiotemporal correlation tracing graph is constructed. Starting from the node with the largest total amplitude change in the graph, a transmission path chain is formed by traversing along the direction of increasing transmission time difference. When the transmission time difference continues to increase and the amplitude attenuation ratio is not higher than a preset attenuation threshold, the traversed path chain is determined to be a congestion diffusion transmission path, and the traversal direction is the extension direction of the transmission path. Based on the extension range of the transmission path chain, the connection positions between its terminal node and adjacent unmarked nodes are identified, the network addresses of the corresponding nodes and the link endpoint identifiers are marked, and the set of network topology boundary coordinates that define the scope of congestion impact is obtained.

[0006] Furthermore, the real-time latency of the production task transmission path is obtained by associating the aforementioned transmission direction, boundary coordinates, and real-time latency with spatiotemporal features to determine congestion risk. Simultaneously, a full set of alternative paths from the current production task source node to the target node is obtained, including: Obtain the transmission path information of the node sequence and link sequence through which the current production task flows, collect the round-trip time of each link data packet and accumulate them to obtain the end-to-end real-time transmission delay of the path. The congestion-affected area is identified based on the direction of congestion propagation and the coordinate set of the network topology boundary. The transmission path is compared with this area to identify the overlapping segments of the paths formed by the nodes and links falling into the area. The average delay within a preset historical period of the segment is obtained as the baseline delay. The difference between the real-time delay and the baseline delay is taken as the delay deviation value. The delay deviation value is compared with the transmission time difference in the transmission direction to obtain the spatiotemporal correlation feature vector. If the delay deviation value in the spatiotemporal correlation feature vector exceeds a preset deviation threshold and the link length of the path overlap section accounts for a proportion of the total link length of the production task transmission path exceeding a preset proportion threshold, it is determined that the production task transmission path has a congestion risk.

[0007] Furthermore, when the congestion risk exceeds a preset risk threshold, hidden bottleneck nodes on the congestion propagation path are identified and locked, a bottleneck exclusion list is generated, and the path selection and load allocation logic of the production task scheduler is updated. When the congestion risk does not exceed the preset risk threshold, the original logic of the scheduler remains unchanged, including: Only when the congestion risk exceeds a preset risk threshold, the flow convergence and queue saturation of each node are extracted from the congestion propagation path. The flow convergence is the ratio of the total flow into the node to the total flow out of the node per unit time, and the queue saturation is the ratio of the current occupancy of the node's buffer to its total capacity. If the traffic convergence of a node exceeds the preset convergence threshold and the queue saturation exceeds the preset saturation threshold, then the node is determined to be a hidden bottleneck node and is locked. Obtain the network address and associated link number of the hidden bottleneck node, register them in the order of their arrangement on the transmission path, and form a bottleneck exclusion list; Based on the bottleneck exclusion list, the corresponding node and link are set to disabled in the scheduler path selection logic, and the task allocation rights of all inbound links pointing to the hidden bottleneck node are reset to zero in the load allocation logic, thus completing the scheduling logic update. If the congestion risk does not exceed the preset risk threshold, the scheduling logic update operation will not be performed, and the original logic of the scheduler will remain unchanged.

[0008] Furthermore, when the congestion risk does not exceed a preset risk threshold, the end-to-end delay value and inter-path delay difference of each path are determined based on the full set of candidate paths to obtain a priority path ranking. When the congestion risk exceeds the preset risk threshold, hidden bottleneck nodes and associated links are excluded from the full set of candidate paths to obtain a subset of available candidate paths. A priority path ranking is generated based on this subset, including: Obtain the full set of alternative paths for production tasks, traverse each path according to the bottleneck exclusion list, and remove the path if it contains a node or link that matches the bottleneck exclusion list, thus obtaining the set of available paths after bottleneck exclusion. For each path in the available path set, the current transmission delay of the link is collected segment by segment from the starting node to the target node and accumulated to obtain the end-to-end real-time transmission delay value of the path. Sort all delay values ​​in ascending order, select the first N paths in the ascending order whose delay values ​​are lower than the preset delay upper limit threshold, and form the fastest response path subset, where N is a preset positive integer; The delay values ​​of adjacent sorted paths in the fastest response path subset are subtracted to obtain the delay difference between paths. Paths with a difference less than a preset difference threshold are assigned to the same priority level; otherwise, they are assigned to the next priority level. When comparing consecutive adjacent paths, if the delay difference between the first and last paths in the same group exceeds the preset difference threshold, the paths are split into new priority levels, and finally the priority paths are sorted from high to low according to the priority level.

[0009] Furthermore, the priority path sorting based on the corresponding scenario, before the injection of sudden production tasks, obtains the task traffic demand by combining historical production data and real-time link status through a preset AI prediction model, and pre-allocates tasks to priority paths, including: Based on priority path sorting, obtain the current remaining bandwidth capacity of a single path within each priority level. The remaining bandwidth capacity is the difference between the total bandwidth of the path and the bandwidth already occupied. At the same time, extract historical task arrival records from the historical production database that match the current time period and include task type, data volume, and arrival time interval. Long Short-Term Memory (LSTM) networks are used to extract the temporal patterns of task data volume and arrival time intervals from historical records. Combined with the real-time status of current link bandwidth utilization and queue depth, the predicted value of task traffic demand within a future preset time window is output. Calculate the proportion of the remaining bandwidth capacity of each individual path in the total remaining capacity, divide the predicted task traffic demand value according to the proportion, and the task carrying capacity of a single path in the high priority level is greater than that of a single path in the low priority level, so as to obtain the pre-allocated task volume of each path. Before a sudden task is injected, the pre-allocated task quantity is registered in the unallocated quota record of the corresponding path in the task scheduling queue. When the task arrives, the data stream is distributed according to the record value. When the actual traffic deviates from the predicted value by more than the preset deviation threshold, the pre-allocated task volume of each path is updated synchronously according to the proportion of the real-time remaining bandwidth capacity.

[0010] Furthermore, if the real-time monitoring path utilization rate exceeds a preset critical buffer threshold, newly added tasks will be switched to the next available path in the ranking order, resulting in a final task handling scheme that adapts to production load requirements, including: Real-time monitoring of the current utilization rate of each path in the priority path ranking, the utilization rate is the ratio of the currently occupied bandwidth of the path to the total bandwidth capacity; If the utilization rate of a certain path exceeds the preset critical buffer threshold, it is marked as saturated, and the available path immediately next to the current path in the priority path ranking is obtained as the switching target path. The newly added task data stream is directed to the target path for switching transmission, the unallocated quota record of each path in the task scheduling queue is updated, and the remaining unallocated quota of the saturated path is synchronously transferred to the target path for switching. If the utilization rate of the target path also exceeds the preset critical buffer threshold, then the process will be sequentially extended to the available paths in the subsequent priority levels until an available path that meets the bandwidth requirements is found. When the utilization rate of a saturated path falls below the critical buffer threshold and the congestion is relieved, the path's original priority and task allocation authority in the priority path sorting are restored, ultimately resulting in a final task-bearing scheme that adapts to the production load requirements.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an AI-based production management method, proposing a complete solution to the business scenario problems of identifying link congestion precursors, predicting congestion propagation, and avoiding hidden bottlenecks in production networks. The method collects queuing delay and link utilization data of equipment nodes through an AI feature extraction module, constructs a spatiotemporal correlation tracing map, accurately determines the direction of congestion propagation and network boundaries, and assesses congestion risk by combining real-time delay correlation analysis. When the risk exceeds a threshold, hidden bottleneck nodes are locked and the scheduling logic is updated, selecting the optimal path set from candidate paths, pre-allocating task traffic using an AI prediction model, and dynamically switching task-bearing paths by monitoring path utilization in real time. This invention significantly improves the load adaptability and stability of the production network through AI-driven congestion prediction and path optimization, effectively avoiding congestion situations and ensuring efficient task transmission. This invention can be applied to continuous production lines for power battery electrode sheets to solve the problem of sudden tasks such as roll changing / tape splicing causing a surge in bus load, exponential growth in data packet queuing delay between the coating machine and the rolling mill, and congestion propagating along the bus towards the back end for splitting and winding, resulting in speed fluctuations across the entire line. Attached Figure Description

[0012] Figure 1 This is a flowchart of an artificial intelligence-based production management method according to the present invention.

[0013] Figure 2 This is a schematic diagram of an artificial intelligence-based production management method according to the present invention.

[0014] Figure 3 This is another schematic diagram of an artificial intelligence-based production management method according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0016] like Figures 1-3 This embodiment of a production management method based on artificial intelligence may specifically include: S101. Collect queuing delay and data transmission link utilization data of production equipment nodes through the AI ​​feature extraction module, extract link jitter and queue depth change features, and obtain micro-variable marker sequences that characterize the precursors and propagation trends of link congestion.

[0017] Using an AI feature extraction module and a global synchronization clock as a reference, queuing delay data of production equipment nodes is periodically collected. The bandwidth occupancy ratio and instantaneous throughput of the data transmission link are simultaneously acquired. The queuing delay data and bandwidth occupancy ratio are stored in a timestamp-aligned manner to form a node link synchronization sampling record table. Based on this table, link jitter features and queue depth change features are extracted. Link jitter features are defined as the dispersion of the one-way transmission delay difference of data packets within a preset time window in adjacent sampling periods. First, the one-way transmission delay is calculated. Where QD is the queuing delay for data acquisition, and PT is the data packet transmission time. PS represents the data packet size, such as 1500 bytes, and TP represents the instantaneous throughput, such as 100 Mbps. The difference in one-way transmission delay between adjacent sampling periods is... The degree of dispersion is calculated using the standard deviation, and the formula is as follows: , where x i Let μ be the one-way transmission delay difference between the i-th adjacent sampling periods within the time window, μ be the arithmetic mean of all differences within the time window, N be the number of sampling points within the time window, and the queue depth change feature be the increase or decrease of the node buffer occupancy within the continuous time window. If the link jitter feature exceeds a preset jitter threshold and the increase of the queue depth change feature also exceeds a preset depth increase threshold, it is marked as a congestion precursor interval. If neither feature exceeds the corresponding preset threshold, a micro-change marker sequence representing the stable operation state of the link is generated. For the congestion precursor interval, a sliding time window method is used to perform time-series correlation calculations on the link jitter characteristics and queue depth change characteristics. Based on the positive and negative directions and increments of the feature values ​​within the window, a micro-variation marker sequence representing the link congestion precursor and propagation trend is generated. Each element of the sequence includes a feature trend marker, amplitude value, and timestamp information for the corresponding time window. The amplitude value is the weighted sum of the standard deviation of link jitter and the amplitude of queue depth change within the window, with weighting coefficients of 0.6 and 0.4, respectively. In the stationary state, the elements of the micro-variation marker sequence include a stationary state trend marker, a feature baseline value, and a timestamp information for the corresponding time window. The trend marker is fixed as a stationary trend, and the amplitude value is the baseline mean of link jitter and queue depth within the window. The sequence format is completely consistent with the micro-variation marker sequence for the congestion precursor interval. The entire sampling period adopts a unified calibration rule, with a default value of 1 second, determined by 1 / 10 of the typical execution cycle of the production task. All data acquisition, feature calculation, model inference, and status monitoring steps are executed based on this unified sampling period, ensuring full-process time synchronization.

[0018] In one implementation, the AI ​​feature extraction module is deployed at the core switching node of the production network. It establishes a communication connection with each production equipment node through a preset data acquisition interface, reads the queuing delay value and occupancy percentage of each node's buffer according to a fixed sampling interval, and obtains the current bandwidth occupancy ratio and instantaneous throughput data from the link monitoring port. The above acquisition process is completed within the same clock cycle, ensuring the time synchronization of the data.

[0019] Specifically, the node link synchronization sampling record table uses timestamps as the primary key index. Each record contains four fields: node identifier at the corresponding time, queuing delay value, bandwidth occupancy ratio, and instantaneous throughput. Each field is stored in alignment according to a unified time base, forming a structured data set that can be used for subsequent feature extraction.

[0020] For example, the process of extracting link jitter features can involve extracting the one-way transmission delay values ​​of data packets from two adjacent sampling periods within a preset time window and calculating their difference. The delay differences of all adjacent periods within this time window are then summarized, and the degree of deviation of these differences from their mean is statistically analyzed. This degree of deviation reflects the dispersion index of link transmission stability. When the link is in a stable state, the fluctuation of the delay difference between adjacent periods is small, and the dispersion value is low. When the link experiences sudden traffic or resource contention, the fluctuation of the delay difference intensifies, and the dispersion value increases accordingly, thereby identifying abnormal changes in the link status.

[0021] In one embodiment, the queue depth change feature is obtained by statistically analyzing the increase or decrease of the node buffer occupancy within a continuous time window. Specifically, the calculation method is to extract the buffer occupancy per second within the window, calculate the net change D, and then calculate the average amplitude A and the maximum amplitude M of the change per second within the window. The final feature value is the ratio of the sum of D, A, and M to the total buffer capacity C. Here, D represents the difference between the occupancy at the end and the occupancy at the beginning, with positive values ​​indicating an increase and negative values ​​indicating a decrease; A represents the average of the absolute values ​​of the changes in all adjacent seconds; and M represents the maximum absolute change. For example, if the window is 5 seconds, the initial occupancy is 100, the end is 150, and the changes per second are 2, 3, -1, and 6, then D=50, A=3, M=6, C=1000, and the feature value=0.059.

[0022] It should be noted that the determination of the congestion precursor interval adopts a dual-threshold joint triggering mechanism. When the link jitter feature exceeds a preset jitter threshold and the increase of the queue depth change feature simultaneously exceeds a preset depth increase threshold, this period is marked as a congestion precursor interval. For the congestion precursor interval, a sliding time window method is used to perform time-series correlation calculations on the link jitter feature and the queue depth change feature. The sliding time window moves on the time axis with a fixed step size, and after each move, the direction and magnitude of the numerical changes of the two types of features within the window are recalculated. The rule for determining the feature trend is as follows: when both the link jitter feature and the queue depth change feature are positively increasing within the window, it is marked as an upward trend and assigned a value of +1; when both features are negatively decreasing, it is marked as a downward trend and assigned a value of -1; in all other cases, it is marked as a stable trend and assigned a value of 0.

[0023] Understandably, the generation of the micro-change label sequence is based on the positive and negative direction and the rate of increase of the feature values ​​within the window. When the feature values ​​increase positively, it is marked as an upward trend; when the feature values ​​decrease negatively, it is marked as a downward trend. The trend labels of each time window are arranged in chronological order to form a micro-change label sequence that characterizes the precursors and propagation trends of link congestion. This sequence can intuitively reflect the dynamic trajectory of network status evolution over time.

[0024] S102. Based on the micro-variant marker sequence, construct a spatiotemporal correlation tracing map with the production network topology as the basis and the temporal characteristic value of the micro-variant marker sequence as the node and link weights, and determine the congestion propagation path, the corresponding propagation direction, and the coordinates of the network topology boundary.

[0025] Based on the micro-variant marker sequence, the number of state transitions and the total amplitude change of the corresponding micro-variant markers for each production equipment node are counted. These are then used as temporal feature values ​​and mapped to the corresponding node positions in the production network physical topology, forming a node set labeled with temporal feature values. For the fiber optic channel connection relationship between adjacent nodes in the node set, the time difference between the occurrence times of the micro-variant marker sequence on each link between adjacent nodes is obtained as the transmission time difference. The ratio of the micro-variant marker amplitude before and after transmission is obtained as the amplitude attenuation ratio. The transmission time difference and amplitude attenuation ratio are used as link weights and assigned to the corresponding channel connection edges. Combining the node set and link weights, a spatiotemporal correlation tracing graph is constructed. Specifically, when calculating the amplitude attenuation ratio, the amplitude A of the upstream node is input. up and downstream node amplitude A down Using formula Perform the calculation and assign the result to the link weight, for example, A. up For 10, A downWhen the ratio is 8, the ratio is 0.8. The lower the amplitude attenuation ratio, the less attenuated the congestion signal is during link transmission. In the spatiotemporal correlation source map, starting from the node with the largest total amplitude change, the path is traversed along the direction of increasing transmission time difference in the link weight. The nodes traversed and the links form a transmission path chain. If the transmission time difference continues to increase and the amplitude attenuation ratio is not higher than a preset attenuation threshold, the traversed path chain is determined to be a congestion propagation path, and the traversal direction is the extension direction of the propagation path. Based on the extension range of the propagation path chain, the connection position between the terminal node of the propagation path chain and the adjacent unmarked node is identified. The network address of the node at the connection position and the link endpoint identifier are marked to obtain the network topology boundary coordinates.

[0026] In one implementation, the extraction of time-series feature values ​​is based on the record of the evolution of the label state of each node in the micro-variable label sequence. By statistically analyzing the changes in micro-variable labels within a preset observation period, a quantitative indicator reflecting the degree of node congestion is obtained. The quantitative indicator includes two dimensions: the number of state switching and the total amount of amplitude change.

[0027] Specifically, the counting method for state transitions is as follows: within the observation period, traverse the sequence of small-change markers corresponding to a certain node, defining an increasing marker as +1, a decreasing marker as -1, and a stationary marker as 0. A state transition is recorded when the marker sign changes between two adjacent time points, such as from +1 to -1, +1 to 0, 0 to -1, -1 to 0, -1 to +1, or 0 to +1. The total number of all transition events is the total number of state transitions for that node, calculated using the following formula: Where t is the total length of the micro-variable marker sequence within the observation period, Let be the label value at time i in the sequence. This is an indicator function; it takes a value of 1 when the condition within the parentheses is true and 0 when it is false. The total amplitude change is calculated by extracting the amplitude values ​​of the micro-variable markers at each time point within the observation period, calculating the absolute difference between the amplitude values ​​at adjacent time points, and summing them. The summed result is the total amplitude change. The number of state transitions reflects the frequency of node state fluctuations, while the total amplitude change reflects the cumulative intensity of changes in node congestion; both together constitute the time-series characteristic value.

[0028] It should be noted that the mapping process of time-series feature values ​​to the location of production network topology nodes is to mark the number of state switching and the total amount of amplitude change of the corresponding nodes in the topology structure according to the correspondence between the physical number and network address of each production equipment node, thereby forming a set of nodes marked with time-series feature values.

[0029] In one embodiment, the process of obtaining the transmission time difference and amplitude attenuation ratio involves a time-series comparison of the micro-change marker sequences between adjacent nodes. The transmission time difference is obtained by selecting adjacent nodes at both ends of a link, extracting the first occurrence time of markers with the same trend direction in the micro-change marker sequences of both nodes, and calculating the difference between the two times. This transmission time difference reflects the time interval during which the congestion state propagates from the upstream node to the downstream node. The amplitude attenuation ratio is obtained by extracting the amplitude values ​​of the micro-change markers at the upstream node and the corresponding amplitude values ​​of the micro-change markers at the downstream node, and calculating the ratio of the downstream amplitude value to the upstream amplitude value. This ratio reflects the degree of intensity loss of the congestion signal during link transmission.

[0030] For example, the spatiotemporal correlation tracing graph is constructed based on the production network topology as its framework, upon which temporal dimension feature information is superimposed. The spatiotemporal correlation tracing graph consists of two parts: a node layer and a link layer. The node layer comprises a set of nodes labeled with temporal feature values, each node carrying attributes such as the number of state transitions and the total amplitude change. The link layer consists of the connection edges of each fiber optic channel, each connection edge carrying attributes such as transmission time difference and amplitude attenuation ratio as link weights. The node layer and the link layer are linked through physical connections, forming a composite graph with both spatial topology and temporal evolution characteristics. In this graph, the temporal feature values ​​of a node characterize the severity of congestion at that location, and the link weights characterize the propagation speed and attenuation characteristics of the congestion signal on that channel. The combination of these two factors allows tracing the complete path of congestion spreading from its source to the periphery.

[0031] Furthermore, the formation process of the transmission path chain adopts a directed traversal method. The node with the largest total change in amplitude in the spatiotemporal correlation source map is located as the starting point of the traversal. Starting from this starting point, the chain extends along the direction of increasing transmission time difference in its adjacent links. Each time a new node is reached, the node and the link it passes through are added to the transmission path chain. The above extension process is repeated until the edge of the map is reached or a link with a sudden change in transmission time difference is encountered. The rule for determining a sudden change in transmission time difference is that the relative increase of the transmission time difference of the current link and the transmission time difference of the previous link exceeds 200%, or the absolute difference exceeds the preset maximum transmission time difference threshold.

[0032] Understandably, the determination of the congestion propagation direction is based on the variation of the transmission time difference and the distribution of the amplitude attenuation ratio of each link in the propagation path. If the transmission time difference along the traversal direction shows a continuously increasing trend, it indicates that the congestion signal is propagated step by step along this direction and generates cumulative delay; if the amplitude attenuation ratio is not higher than a preset attenuation threshold, it indicates that the congestion intensity does not significantly attenuate during propagation and can stably propagate downstream. When both of the above conditions are met simultaneously, the traversal direction is determined to be the extension direction of the congestion propagation path. The preset attenuation threshold is taken as 1.2 times the average amplitude attenuation ratio of effective propagation paths in historical congestion events.

[0033] In one possible implementation, the identification of network topology boundary coordinates is based on the termination position of the extended transmission path chain. When the transmission path chain extends to a node, if there are unmarked nodes among the node's neighboring nodes that have not entered the transmission path chain, the connection position between the marked and unmarked nodes constitutes the boundary of congestion propagation. The network coordinates of the nodes and the coordinates of the link endpoints involved in the boundary position are recorded and calibrated to obtain the network topology boundary coordinates. These coordinates define the scope and propagation boundary of the current congestion situation in the production network.

[0034] S103. Obtain the real-time delay of the production task transmission path, correlate the congestion propagation direction, network topology boundary coordinates and the real-time delay of the corresponding production task transmission path with spatiotemporal dimension features to determine the congestion risk, and at the same time obtain the full set of alternative paths from the source node to the target node of the current production task.

[0035] Obtain the transmission path information of the current production task. The transmission path information includes the sequence of nodes and links through which the task data flows. Collect the round-trip time of data packets from the monitoring port of each link and accumulate them to obtain the end-to-end real-time transmission delay of the production task transmission path. The algorithm synchronously acquires a complete set of candidate paths from the source node to the target node of the current production task. This complete set of candidate paths is generated using a modified depth-first search algorithm. The algorithm inputs are a set of nodes and links represented by the network topology graph G, the source node S, the target node T, and a maximum hop count threshold of 16. The process is implemented using a recursive function DFS(current, path), where current is the current node and path is the current path stack. First, it checks if the path length exceeds 16 or if current is T. If it is T and the path is loop-free, it adds the path to the result set. Otherwise, it traverses the neighbor nodes of current. If a neighbor is not in path and is not a faulty node or a disabled link, it recursively calls DFS(neighbor, path, add neighbor). During backtracking, path stack elements are popped. During generation, invalid paths containing faulty nodes or disabled links are automatically excluded. The output is a set of all physically connected, loop-free transmission paths between the source and target nodes with a hop count not exceeding 16. The preset maximum hop count threshold is fixed at 16 hops based on the size of the production network topology. Based on the congestion propagation direction and the network topology boundary coordinate set, the spatial range of the congestion-affected area is defined. The spatial location of the production task transmission path is compared with that of the congestion-affected area. Nodes and links falling into the congestion-affected area in the transmission path are identified to form path overlap segments. The average delay of the path overlap segment within a preset historical period is obtained as the baseline delay. The real-time delay of the path overlap segment is calculated by subtracting the baseline delay to obtain the delay deviation value. The delay deviation value is compared with the transmission time difference in the congestion propagation direction to determine whether the growth trend of the delay deviation value is consistent with the cumulative direction of the transmission time difference. The Pearson correlation coefficient is used to quantify the correlation between the two growth rates. A correlation coefficient greater than 0.7 is considered positively correlated. If the two directions are consistent and the growth rates are positively correlated, the delay deviation value, the proportion of the link length of the path overlap segment to the total link length of the production task transmission path, and the time comparison results are combined to form a spatiotemporal correlation feature vector. The link length is measured in hops, and the percentage of link length in the overlapping section is calculated as: (Number of hops in the overlapping section / Total number of hops in the transmission path) × 100%. Based on the spatiotemporal correlation feature vector, if the delay deviation exceeds a preset deviation threshold and the percentage of link length in the overlapping section to the total link length of the production task transmission path exceeds a preset percentage threshold, then the production task transmission path is determined to have congestion risk, and proceeds to the subsequent risk level quantification stage. If neither condition is met simultaneously, then the transmission path is determined to have no congestion risk, and the subsequent regular path sorting process is directly executed.For transmission paths identified as having congestion risk, the congestion risk is represented by a quantitative value: Congestion risk quantification value = (delay deviation value / preset deviation threshold) × 0.6 + (path overlap segment link length ratio / preset ratio threshold) × 0.4. The preset risk threshold is fixed at 1.0. When the congestion risk quantification value is ≥ 1.0, it is determined that the congestion risk exceeds the preset risk threshold; when the congestion risk quantification value is < 1.0, it is determined that the congestion risk does not exceed the preset risk threshold.

[0036] In one implementation, the transmission path information of the production task is obtained through the task scheduling record. The transmission path information stores all intermediate nodes and link identifiers that the task data passes through from the source node to the target node in the form of an ordered list. The round-trip time of the data packet is collected by the delay detection function of each link monitoring port, and the round-trip time of all links on the transmission path is accumulated to obtain the real-time delay.

[0037] Specifically, the spatial extent of the congestion-affected area is defined by the direction of congestion propagation and the coordinates of the network topology boundary. All nodes and links covered along the propagation direction from the starting node to the boundary coordinates constitute this area. The node and link sequences in the production task transmission path are matched one by one with the node and link sets within the congestion-affected area. If a node or link in the transmission path exists within the congestion-affected area, that node or link is marked as an overlapping element. All overlapping elements are arranged in the order of the transmission path to form overlapping segments.

[0038] For example, the baseline delay is obtained based on the delay statistics of the path overlap segment within a preset historical period. The preset historical period refers to a fixed time window before the current moment. Within this time window, multiple delay samples of the path overlap segment are extracted and their arithmetic mean is calculated. The arithmetic mean is the baseline delay. The delay deviation value is obtained by subtracting the baseline delay from the real-time delay of the path overlap segment. This deviation value reflects the delay increment of the path overlap segment at the current moment relative to the historical normal state. Based on this, the delay deviation value is compared with the propagation time difference in the direction of congestion propagation. The time-series comparison determines whether the growth trend of the delay deviation value is consistent with the cumulative direction of the propagation time difference. If the two directions are consistent and the growth magnitudes are positively correlated, the delay deviation value, the proportion of the path overlap segment length, and the time-series comparison result are combined to form a spatiotemporal correlation feature vector.

[0039] It should be noted that the spatiotemporal correlation feature vector includes two parts: spatial dimension information and temporal dimension information. The spatial dimension information is represented by the ratio of the link length of the path overlap segment to the total link length of the transmission path of the production task, and the temporal dimension information is represented by the delay deviation value. The two are combined to form a comprehensive quantitative description of congestion risk.

[0040] In one embodiment, the congestion risk assessment employs a dual-threshold joint triggering mechanism. When the delay deviation value in the spatiotemporal correlation feature vector exceeds a preset deviation threshold and the proportion of link length in the path overlap segment exceeds a preset proportion threshold, the transmission path of the production task is determined to have a congestion risk. This assessment result indicates that the degree to which the current transmission path is affected by congestion spread has reached the warning level. The preset deviation threshold is twice the standard deviation of the path's historical normal delay, and the preset proportion threshold is fixed at 30%.

[0041] S104. If the congestion risk exceeds the preset risk threshold, identify and lock the hidden bottleneck nodes on the congestion propagation path, generate a bottleneck exclusion list, and update the path selection and load allocation logic of the production task scheduler through the bottleneck exclusion list. If the congestion risk does not exceed the preset risk threshold, keep the original logic of the scheduler unchanged.

[0042] When the congestion risk exceeds a preset risk threshold, the identification process for hidden bottleneck nodes is triggered. Two indicators, traffic convergence and queue saturation, are extracted from all nodes covered by the congestion propagation path. Nodes with potential bottleneck characteristics are screened using a dual-threshold joint judgment method. The traffic convergence calculation formula is C=I / O, where I is the total traffic flowing into the node per unit time (in Mbps), and O is the total outflow traffic (in Mbps). For example, when I=500Mbps and O=200Mbps, C=2.5. The queue saturation is the ratio of the current occupancy of the node's buffer to its total buffer capacity. When the traffic convergence value is greater than a preset convergence threshold and the queue saturation value exceeds a preset saturation threshold, it indicates that the node has a potential for traffic backlog and queue overflow. The node is identified as a hidden bottleneck node and locked. For each hidden bottleneck node, its network address and associated link number in the production network physical topology are obtained. The network address and associated link number of the hidden bottleneck node are registered according to their order on the congestion propagation path, forming a bottleneck exclusion list. Based on the bottleneck exclusion list, the nodes and links corresponding to the network addresses and associated link numbers in the bottleneck exclusion list are set to disabled status in the path selection logic of the production task scheduler. Simultaneously, in the load allocation logic, the task allocation weight values ​​of all inbound links pointing to the hidden bottleneck node are set to zero, completing the update of the path selection and load allocation logic. If the congestion risk does not exceed the preset risk threshold, the scheduling logic update operation described in this step is not executed, and the original scheduler logic remains unchanged.

[0043] In one implementation, when the value of the congestion risk exceeds a preset risk threshold, a process for identifying hidden bottleneck nodes is triggered. Two indicators, flow convergence and queue saturation, are extracted one by one from all nodes covered by the congestion propagation path. Nodes with potential bottleneck characteristics are screened out by a dual-threshold joint judgment method.

[0044] Specifically, traffic convergence reflects the degree to which a node performs traffic convergence in the production network. It is calculated by dividing the total traffic flowing into and out of the node per unit time. When the traffic convergence value exceeds a preset convergence threshold, it indicates that the node receives far more data than it can output, posing a risk of traffic backlog. Queue saturation reflects the congestion level of a node's buffer. It is calculated by dividing the currently occupied storage space of the node's buffer by its total storage capacity. When the queue saturation value exceeds a preset saturation threshold, it indicates that the node's buffer is about to reach its capacity limit, posing a risk of queue overflow. If a node simultaneously meets both the conditions of exceeding the preset convergence threshold and the preset saturation threshold, it is identified as a hidden bottleneck node. The preset convergence threshold is fixed at 1.5, and the preset saturation threshold is fixed at 85%.

[0045] It should be noted that the locking mark of the hidden bottleneck node is implemented by writing the bottleneck identifier in the node attribute field. The bottleneck identifier is used to distinguish the node from other normal nodes, which facilitates quick location during subsequent processing.

[0046] In one embodiment, the bottleneck exclusion list is formed by traversing all locked and marked implicit bottleneck nodes along the congestion propagation path, and sequentially writing the network address and associated link number of each implicit bottleneck node into the list data structure according to their arrangement from the start to the end of the propagation path. Further, the path selection logic of the production task scheduler is updated by reading all network addresses and associated link numbers from the bottleneck exclusion list, and setting the available status flag of the corresponding entries to disabled in the candidate node set and candidate link set for path selection. The load allocation logic is updated by modifying the task allocation weight value of adjacent links directly connected to implicit bottleneck nodes to zero, so that no new production task loads are assigned to these links during subsequent task allocation.

[0047] S105. When the congestion risk does not exceed the preset risk threshold, the end-to-end delay value and the delay difference between paths are judged based on the full set of candidate paths to obtain the priority path ranking. When the congestion risk exceeds the preset risk threshold, hidden bottleneck nodes and associated links are excluded from the full set of candidate paths to obtain a subset of available candidate paths. Based on this subset, a priority path ranking for actively avoiding congestion is generated.

[0048] When the congestion risk does not exceed the preset risk threshold, the full set of candidate paths for the production task is directly obtained as the set of paths to be screened. When the congestion risk exceeds the preset risk threshold, the full set of candidate paths for the production task is obtained. Based on the network address and associated link number recorded in the bottleneck exclusion list, the nodes and links contained in each path in the candidate path set are traversed. If there is a node or link in a certain path that matches the bottleneck exclusion list, the path is removed from the candidate path set, and the available candidate path subset after bottleneck exclusion is obtained as the set of paths to be screened. For each path in the set of paths to be filtered, the current transmission delay of each link is collected segment by segment from the starting node to the target node. The delay values ​​of each link are accumulated to obtain the end-to-end real-time transmission delay value of the path. The end-to-end real-time transmission delay values ​​of all paths in the set of paths to be filtered are sorted in ascending order of numerical value. The top N paths in the ascending order with delay values ​​lower than a preset delay upper limit threshold are selected to form the fastest response path subset, where N is a preset positive integer and the preset delay upper limit threshold is the maximum transmission delay that the production task can tolerate; for example, N=5 and the preset delay upper limit threshold=100ms can be set. Based on the end-to-end real-time transmission delay values ​​of each path in the fastest response path subset, the end-to-end delay values ​​of two paths at adjacent sorting positions are subtracted to obtain the path delay difference value. If the path delay difference value is less than the preset difference threshold, the corresponding two paths are assigned to the same priority level. If the path delay difference value is greater than or equal to the preset difference threshold, the latter path is assigned to the next priority level. When comparing consecutive adjacent paths, if the delay difference between the first and last paths in the same group exceeds a preset difference threshold, the group is split at the first adjacent path pair that exceeds the threshold, forming a new priority level. The core determination follows the formula. ,in, This is the delay value for the first path within the current group. This is the delay value for the last path to be included in the current group. The preset difference threshold is used. For example, if the input sorting delay values ​​are 10ms, 12ms, 15ms, 30ms, and 32ms and the threshold is 5ms, then the groups are as follows: Level 1: 10ms-15ms, with a first and last difference of 4ms < 5ms; Level 2: 30ms-32ms. Finally, the priority paths are sorted in descending order of priority level.

[0049] In one implementation, the candidate path set contains all physically connected transmission paths from the task source node to the target node, and each candidate path is stored in the form of a node sequence and a link sequence. The bottleneck elimination process involves reading the node sequence and link sequence of each path in the candidate path set one by one, comparing them item by item with the network address and associated link number recorded in the bottleneck elimination list, and if a match is found, the path is marked as unavailable and removed from the set.

[0050] Specifically, the paths that remain after bottleneck elimination constitute a set of available paths. Each path in the set of available paths does not pass through the identified hidden bottleneck nodes and has the basic conditions to carry out production tasks.

[0051] For example, the end-to-end latency value collection process is as follows: For a given path in the available path set, starting from the path's starting node, probe data packets are sent sequentially to downstream nodes, and the round-trip time of each link is recorded. The round-trip times of all links on the path are summed to obtain the end-to-end latency value. After collecting the end-to-end latency values ​​for all paths in the available path set, they are sorted in ascending order of value. A preset latency upper limit threshold is the maximum transmission latency that the production task can tolerate. Paths with end-to-end latency values ​​lower than this threshold are selected to form a subset of the fastest-responding paths. Paths within this subset all meet the timeliness requirements of the production task.

[0052] It should be noted that the priority level is determined based on the delay difference between paths. Paths within the subset of fastest response paths are sorted in ascending order of end-to-end delay value. The difference in end-to-end delay value is calculated between two adjacent paths. If this difference is less than a preset threshold, it indicates that the two paths have similar response speeds and are assigned to the same priority level. If the difference is greater than or equal to the preset threshold, it indicates that the response speed of the latter path is significantly lower than the former path, and the latter path is assigned to the next lower priority level. When comparing consecutive adjacent paths, if the delay difference between the first and last paths in the same group exceeds the preset threshold, the path is split from the adjacent path where the delay difference first exceeds the threshold. The first half retains the original priority level, and the second half forms a new priority level. This process continues until the delay difference between the first and last paths in all groups does not exceed the preset threshold. Finally, the priority paths are output in descending order of priority level. The preset threshold is 10% of the minimum end-to-end delay of the path.

[0053] It is understood that the priority path sorting classifies paths with similar response speeds into the same priority level. During task scheduling, paths within the same level have the same selection priority, while paths of different levels are selected in descending order of priority, thereby achieving proactive avoidance of congestion.

[0054] S106. Based on the priority path sorting, before injecting sudden production tasks, the task traffic demand is obtained by using a preset AI prediction model, combined with historical production data and real-time link status, and tasks are pre-allocated to priority paths according to the task traffic demand.

[0055] Based on the priority path sorting, the current remaining bandwidth capacity of each priority path is obtained. The remaining bandwidth capacity is the difference between the total bandwidth of the path and the bandwidth already occupied. Simultaneously, historical time period data from the historical production database, which are in the same shift and process stage as the current moment, are extracted to obtain historical task arrival records containing task type, task data volume, and task arrival time interval. A Long Short-Term Memory (LSTM) network is used to extract the temporal patterns of task data volume and arrival time interval from the historical task arrival records. Using the bandwidth occupancy rate and queue depth in the current real-time link status as auxiliary inputs, a predicted value for task traffic demand within a preset future time window is output. This predicted value represents the total amount of data transmission expected to occur when a sudden task injection occurs. For the predicted task traffic demand, the proportion of remaining bandwidth capacity of each priority path in the total remaining bandwidth capacity of all available paths is calculated. The predicted task traffic demand is then proportionally divided according to this proportion. High-priority paths are assigned a weighting coefficient of 1.2 to ensure that the task capacity of a single path within a high-priority path is greater than that of a single path within a low-priority path. Even if the remaining bandwidth proportion of a high-priority path is lower than that of a low-priority path, the weighting coefficient ensures that its basic capacity is not lower than that of the low-priority path. Low-priority paths receive a smaller task capacity than high-priority paths, resulting in the pre-allocated task volume for each path. Based on this pre-allocated task volume, before the actual injection of sudden production tasks, the pre-allocated task volume for each path is registered in the corresponding path's pending allocation record in the task scheduling queue. When a sudden task arrives, the task data stream is distributed to the corresponding priority path according to the value of the pending allocation record, completing the pre-allocation of tasks to priority paths. When the relative deviation between the actual traffic and the predicted value exceeds a preset deviation threshold, the pre-allocated task volume is immediately updated; simultaneously, the pre-allocated task volume is refreshed once every sampling period according to the proportion of the real-time remaining bandwidth capacity; the update only adjusts the allocation of newly added tasks and does not backtrack tasks that have already been allocated. The preset deviation threshold is fixed at 20%.

[0056] In one implementation, the priority path ranking is obtained based on the path priority level division result obtained from the previous processing. For each priority level path, the currently occupied bandwidth value is read from the link monitoring port, and the remaining bandwidth capacity is obtained by subtracting the occupied bandwidth from the total path bandwidth. The remaining bandwidth capacity reflects the upper limit of the current ability of each path to carry new tasks.

[0057] Specifically, historical task arrival records are extracted from the historical production database according to the time period matching principle. Priority is given to matching historical data that is in the same production shift, production process stage, product category, and workday type as the current time. The matching period is the corresponding time period within 3 months prior to the current time, and historical data under abnormal operating conditions such as equipment failure and planned downtime is automatically excluded. The task type field in the historical task arrival record identifies the production process category to which the task belongs, the task data volume field records the total number of bytes of data packets carried by a single task, and the task arrival time interval field records the time difference between two adjacent tasks entering the scheduling queue.

[0058] For example, the temporal pattern extraction process of a Long Short-Term Memory (LSTM) network is as follows: Historical task arrival records are arranged chronologically to form an input sequence. The input vector at each moment in the sequence contains two feature components: the amount of task data at that moment and the time interval between the arrival of the previous task. The LSM network internally contains three control units: a forget gate, an input gate, and an output gate. The forget gate determines which information from the previous moment's memory state needs to be discarded; the input gate determines which information from the current moment's input needs to be written into the memory state; and the output gate determines which information from the current memory state needs to be output to the next layer. Through the coordinated action of these three control units, the network gradually accumulates the periodic patterns and trend changes in task arrivals while processing the input sequence, ultimately outputting a prediction of task flow within a future time window.

[0059] It should be noted that the real-time link status is used as an auxiliary input feature by concatenating the real-time status features of the bandwidth utilization and queue depth of each priority path at the current moment with the hidden state vector output by the Long Short-Term Memory network. The concatenated vector is then input into a single fully connected layer. The output dimension of the fully connected layer is 1, and the output value is the predicted value of the task traffic demand within the future preset time window. The model performs inference to update the predicted value once every sampling period. If the predicted value exceeds 120% of the historical maximum traffic, then 120% of the historical maximum traffic is taken as the corrected predicted value to avoid abnormal prediction results. The achievable configuration of the Long Short-Term Memory (LSTM) network is as follows: the input sequence length is 24 sampling periods, the input feature dimension is 2 (including task data volume and task arrival time interval), the LSTM layer is set to 2 layers, the hidden layer dimension is 64, the activation function is Tanh, the output layer is a fully connected layer with an output dimension of 1; the training uses mean squared error (MSE) as the loss function, the Adam optimizer, the learning rate is set to 0.001, the batch size is 32, the training iterations are 100 rounds, and training is stopped early when the validation set loss does not decrease for 10 consecutive rounds; the training data is the time series data of production tasks within the past 3 months, divided into training and validation sets in an 8:2 ratio.

[0060] In one embodiment, the predicted task traffic demand is output as the total data transmission volume, representing the expected data transmission volume during a burst of tasks within a preset time window. The preset time window length T is set based on the typical execution cycle C of the production task; for example, if C = 20 minutes, T = 30 minutes. The calculation formula is as follows: C is obtained by averaging historical task data. The LSTM model takes historical traffic sequences as input and outputs predicted values. For example, if the data from the past hour is input, the predicted total volume within T is 500MB. Further, the calculation of the remaining bandwidth capacity percentage and the task allocation process involves summing the remaining bandwidth capacity of all available priority paths to obtain the total remaining capacity. The remaining bandwidth capacity of each priority path is then divided by the total remaining capacity to obtain the capacity percentage of that path. The predicted task traffic demand is then proportionally allocated according to the capacity percentage of each path. Higher priority paths with larger percentages receive larger task allocations, while lower priority paths with smaller percentages receive smaller task allocations. This allocation method, while satisfying the priority order, also considers the actual carrying capacity of each path, avoiding the allocation of too many tasks to paths with insufficient remaining capacity.

[0061] It is understandable that the registration of pre-allocated task volume is achieved by establishing a pending quota record for each priority path in the task scheduling queue. The pending quota record stores the upper limit of the amount of task data that each path has reserved before the arrival of a sudden task.

[0062] In one possible implementation, when a production burst task actually arrives at the scheduling queue, the scheduler reads the allocation quota record of each priority path and distributes the task data stream to each path in descending order of priority. The amount of task data distributed to each path does not exceed the value in its allocation quota record, until all burst tasks are distributed or the allocation quota of all paths is exhausted.

[0063] Preferably, the above-mentioned pre-allocation mechanism completes the planning and reservation of task capacity before the actual injection of emergency tasks, and distributes them directly according to the reserved capacity when the emergency task arrives, thereby reducing the response delay of scheduling decisions.

[0064] S107. Monitor path utilization in real time. If the path utilization exceeds the preset critical buffer threshold, switch the new task to the next available path in the ranking to obtain the final task carrying scheme that adapts to the production load requirements.

[0065] The system monitors the current utilization rate of each path in the priority path ranking in real time. The utilization rate is the ratio of the currently occupied bandwidth of a path to the total bandwidth capacity. If the utilization rate of a path exceeds a preset critical buffer threshold of 0.7, the path is marked as saturated. Simultaneously, the next available path immediately following the current path in the priority path ranking is selected as the switching target path. An available path is defined as a path with a utilization rate below 0.7. Based on the switching target path, the data stream of subsequent new tasks is directed to the switching target path for transmission. The unallocated quota record of each path in the task scheduling queue is updated. All remaining unused unallocated quota of the saturated path is transferred to the switching target path in one go. The quota transfer process is as follows: input the remaining quota R of the saturated path, where R = total unallocated quota - allocated quota, output the updated switching path quota as the original quota + R. If the switching target path also triggers saturation, the remaining unallocated quota of the current switching target path is transferred to the next next available path. The quota transfer only applies to the unallocated quota of new tasks and does not affect the data stream of tasks already in transmission. If the utilization rate of the target path also exceeds the preset critical buffer threshold of 0.7, the process will proceed sequentially to available paths within the subsequent priority levels until a suitable available path that meets the bandwidth requirements is found. When the utilization rate of the saturated path falls below the critical buffer threshold of 0.7 and the congestion is relieved, the path's original priority and task allocation authority in the priority path ranking will be restored, ultimately resulting in a final task handling scheme that adapts to the current production load requirements.

[0066] In one implementation, real-time monitoring of path utilization is achieved by periodically collecting the currently occupied bandwidth of each priority path. The occupied bandwidth is divided by the total bandwidth capacity of the path to obtain the utilization value. The preset critical buffer threshold is a fixed proportion of 70% of the total bandwidth capacity of the path. When the utilization value reaches or exceeds this proportion, a path state switch is triggered. Congestion is resolved when the following conditions are met simultaneously: the path utilization rate falls below the critical buffer threshold for three consecutive sampling periods; the jitter characteristic and queue depth characteristic of the corresponding link fall below the corresponding preset threshold; and there are no abnormal events such as queue overflow or packet retransmission at the path nodes. The jitter characteristic is defined as the variance of the link round-trip time, calculated using the following formula: Where V is the variance and RTT i Let V be the round-trip time (RTT) for the i-th sampling period, μ be the average RTT, n be the number of sampling periods, and the preset threshold be 50ms. The calculation process involves collecting 10 RTT samples per period to calculate V. If V is below 50ms, the latency is reduced. The queue depth feature is defined as the average percentage of occupancy of the node's buffer queue, calculated using the following formula: Where D is the depth percentage, Q iLet D be the queue length at the i-th sampling time, m be the number of samplings, C be the total queue capacity, and the preset threshold be 80%. The calculation process is to calculate the queue length D by sampling 5 times per cycle. If D is lower than 80%, the queue length will drop.

[0067] Specifically, when the utilization rate of a priority path exceeds a preset critical buffer threshold, the scheduler marks that path as saturated and suspends the allocation of new tasks to it. At the same time, it reads the second-ranked available path from the priority path ranking as the switching target path. Subsequent new task data streams will no longer enter the saturated path, but will be directed to the switching target path for transmission. This switching process is performed immediately upon task arrival, without interrupting the task data stream already in transmission.

[0068] It should be noted that after the task switch is completed, the scheduler synchronously updates the allocation quota records of each path in the task scheduling queue, sets the allocation quota of the saturated path to zero, and increases the allocation quota of the target path accordingly, thereby forming the final task carrying scheme that adapts to the current production load demand.

[0069] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A production management method based on artificial intelligence, characterized in that, The method includes: The AI ​​feature extraction module collects queuing delay and data transmission link utilization data of production equipment nodes, extracts link jitter and queue depth change features, and obtains micro-variable marker sequences that characterize the precursors and propagation trends of link congestion. Based on the micro-variable marker sequence, a spatiotemporal correlation tracing map is constructed with the production network topology as the basis and its temporal characteristic value as the node and link weights to determine the congestion propagation path, the corresponding propagation direction, and the coordinates of the network topology boundary. The real-time latency of the production task transmission path is obtained, and the aforementioned transmission direction, boundary coordinates and real-time latency are correlated with spatiotemporal dimension features to determine the congestion risk. At the same time, the full set of alternative paths from the source node to the target node of the current production task is obtained. When the congestion risk exceeds the preset risk threshold, identify and lock the hidden bottleneck nodes on the congestion propagation path, generate a bottleneck exclusion list and update the path selection and load allocation logic of the production task scheduler; when the congestion risk does not exceed the preset risk threshold, keep the original logic of the scheduler unchanged. When the congestion risk does not exceed the preset risk threshold, the end-to-end delay value and the delay difference between paths are judged based on the full set of candidate paths to obtain the priority path ranking. When the congestion risk exceeds the preset risk threshold, hidden bottleneck nodes and associated links are excluded from the full set of candidate paths to obtain a subset of available candidate paths, and the priority path ranking is generated based on this subset. Based on the priority path ranking of the corresponding scenario, before the injection of sudden production tasks, the task traffic demand is obtained by combining historical production data and real-time link status through a preset AI prediction model, and tasks are pre-allocated to priority paths. Real-time monitoring of path utilization; if the preset critical buffer threshold is exceeded, new tasks will be switched to the next available path in the ranking, thus obtaining the final task handling scheme that adapts to the production load requirements.

2. The production management method based on artificial intelligence according to claim 1, characterized in that, The AI ​​feature extraction module collects queuing delay and data transmission link utilization data of production equipment nodes, extracts link jitter and queue depth change features, and obtains a micro-variable marker sequence characterizing the precursors and propagation trends of link congestion, including: Using the global synchronization clock as a reference, the AI ​​feature extraction module periodically collects queuing delay data of production equipment nodes, synchronously obtains the bandwidth occupancy ratio and instantaneous throughput of the data transmission link, and stores them in a node link synchronization sampling record table aligned with the timestamp. Based on the record table, the link jitter and queue depth change features are extracted. The link jitter feature is the dispersion of the difference in one-way transmission delay of data packets within adjacent sampling periods within a preset time window. The queue depth change feature is the increase or decrease of the node buffer occupancy within a continuous time window. When both features exceed the preset jitter threshold and the depth increase threshold respectively, it is marked as a congestion precursor interval. When neither feature exceeds the corresponding preset threshold, a micro-change marker sequence representing the stable operation state of the link is generated. For the congestion precursor interval, a sliding time window is used to perform time-series correlation operations on the two features. According to the positive and negative directions and the increment of the feature values ​​within the window, a micro-variation label sequence is generated. Each element of the sequence contains the feature trend label, amplitude value and timestamp information of the corresponding time window.

3. The production management method based on artificial intelligence according to claim 1, characterized in that, The process of constructing a spatiotemporal correlation source map based on the production network topology and using its temporal characteristic values ​​as node and link weights, according to the micro-variant marker sequence, to determine the congestion propagation path and corresponding propagation direction and network topology boundary coordinates, includes: Based on the micro-change marker sequence, the number of state switching times and the total amplitude change of the corresponding micro-change markers for each production equipment node are counted. The two are then used as time-series feature values ​​and mapped to the corresponding nodes in the physical topology of the production network to form a set of nodes labeled with time-series feature values. The number of state switching times is the cumulative number of positive and negative direction conversions of the feature trend markers within the preset observation period, and the total amplitude change is the cumulative sum of the absolute differences in amplitude values ​​between adjacent time windows during the same period. For the fiber channel connection relationship between adjacent nodes in the node set, the time difference of the occurrence of the micro-change mark between adjacent nodes is used as the transmission time difference, and the ratio of the mark amplitude before and after transmission is used as the amplitude attenuation ratio. The two are used as link weights and assigned to the corresponding connection edges. Combined with the node set, a spatiotemporal correlation tracing graph is constructed. Starting from the node with the largest total amplitude change in the graph, a transmission path chain is formed by traversing along the direction of increasing transmission time difference. When the transmission time difference continues to increase and the amplitude attenuation ratio is not higher than a preset attenuation threshold, the traversed path chain is determined to be a congestion diffusion transmission path, and the traversal direction is the extension direction of the transmission path. Based on the extension range of the transmission path chain, the connection positions between its terminal node and adjacent unmarked nodes are identified, the network addresses of the corresponding nodes and the link endpoint identifiers are marked, and the set of network topology boundary coordinates that define the scope of congestion impact is obtained.

4. The production management method based on artificial intelligence according to claim 1, characterized in that, The process of obtaining the real-time latency of the production task transmission path involves associating the aforementioned transmission direction, boundary coordinates, and real-time latency with spatiotemporal features to determine congestion risk. Simultaneously, it obtains a complete set of alternative paths from the current production task source node to the target node, including: Obtain the transmission path information of the node sequence and link sequence through which the current production task flows, collect the round-trip time of each link data packet and accumulate them to obtain the end-to-end real-time transmission delay of the path. The congestion-affected area is identified based on the direction of congestion propagation and the coordinate set of the network topology boundary. The transmission path is compared with this area to identify the overlapping segments of the paths formed by the nodes and links falling into the area. The average delay within a preset historical period of the segment is obtained as the baseline delay. The difference between the real-time delay and the baseline delay is taken as the delay deviation value. The delay deviation value is compared with the transmission time difference in the transmission direction to obtain the spatiotemporal correlation feature vector. If the delay deviation value in the spatiotemporal correlation feature vector exceeds a preset deviation threshold and the link length of the path overlap section accounts for a proportion of the total link length of the production task transmission path exceeding a preset proportion threshold, it is determined that the production task transmission path has a congestion risk.

5. The production management method based on artificial intelligence according to claim 1, characterized in that, When the congestion risk exceeds a preset risk threshold, hidden bottleneck nodes on the congestion propagation path are identified and locked, a bottleneck exclusion list is generated, and the path selection and load allocation logic of the production task scheduler is updated. When the congestion risk does not exceed the preset risk threshold, the original logic of the scheduler remains unchanged, including: Only when the congestion risk exceeds a preset risk threshold, the flow convergence and queue saturation of each node are extracted from the congestion propagation path. The flow convergence is the ratio of the total flow into the node to the total flow out of the node per unit time, and the queue saturation is the ratio of the current occupancy of the node's buffer to its total capacity. If the traffic convergence of a node exceeds the preset convergence threshold and the queue saturation exceeds the preset saturation threshold, then the node is determined to be a hidden bottleneck node and is locked. Obtain the network address and associated link number of the hidden bottleneck node, register them in the order of their arrangement on the transmission path, and form a bottleneck exclusion list; Based on the bottleneck exclusion list, the corresponding node and link are set to disabled in the scheduler path selection logic, and the task allocation rights of all inbound links pointing to the hidden bottleneck node are reset to zero in the load allocation logic, thus completing the scheduling logic update. If the congestion risk does not exceed the preset risk threshold, the scheduling logic update operation will not be performed, and the original logic of the scheduler will remain unchanged.

6. The production management method based on artificial intelligence according to claim 5, characterized in that, When the congestion risk does not exceed a preset risk threshold, the end-to-end delay value and inter-path delay difference of each path are determined based on the full set of candidate paths to obtain a priority path ranking. When the congestion risk exceeds the preset risk threshold, hidden bottleneck nodes and associated links are excluded from the full set of candidate paths to obtain a subset of available candidate paths. A priority path ranking is generated based on this subset, including: Obtain the full set of alternative paths for production tasks, traverse each path according to the bottleneck exclusion list, and remove the path if it contains a node or link that matches the bottleneck exclusion list, thus obtaining the set of available paths after bottleneck exclusion. For each path in the available path set, the current transmission delay of the link is collected segment by segment from the starting node to the target node and accumulated to obtain the end-to-end real-time transmission delay value of the path. Sort all delay values ​​in ascending order, select the first N paths in the ascending order whose delay values ​​are lower than the preset delay upper limit threshold, and form the fastest response path subset, where N is a preset positive integer; The delay values ​​of adjacent sorted paths in the fastest response path subset are subtracted to obtain the delay difference between paths. Paths with a difference less than a preset difference threshold are assigned to the same priority level; otherwise, they are assigned to the next priority level. When comparing consecutive adjacent paths, if the delay difference between the first and last paths in the same group exceeds the preset difference threshold, the paths are split into new priority levels, and finally the priority paths are sorted from high to low according to the priority level.

7. The production management method based on artificial intelligence according to claim 6, characterized in that, The priority path sorting based on the corresponding scenario, before the injection of sudden production tasks, uses a preset AI prediction model combined with historical production data and real-time link status to obtain task traffic demand, and pre-allocates tasks to priority paths, including: Based on priority path sorting, obtain the current remaining bandwidth capacity of a single path within each priority level. The remaining bandwidth capacity is the difference between the total bandwidth of the path and the bandwidth already occupied. At the same time, extract historical task arrival records from the historical production database that match the current time period and include task type, data volume, and arrival time interval. Long Short-Term Memory (LSTM) networks are used to extract the temporal patterns of task data volume and arrival time intervals from historical records. Combined with the real-time status of current link bandwidth utilization and queue depth, the predicted value of task traffic demand within a future preset time window is output. Calculate the proportion of the remaining bandwidth capacity of each individual path in the total remaining capacity, divide the predicted task traffic demand value according to the proportion, and the task carrying capacity of a single path in the high priority level is greater than that of a single path in the low priority level, so as to obtain the pre-allocated task volume of each path. Before a sudden task is injected, the pre-allocated task quantity is registered in the unallocated quota record of the corresponding path in the task scheduling queue. When the task arrives, the data stream is distributed according to the record value. When the actual traffic deviates from the predicted value by more than the preset deviation threshold, the pre-allocated task volume of each path is updated synchronously according to the proportion of the real-time remaining bandwidth capacity.

8. The production management method based on artificial intelligence according to claim 7, characterized in that, If the real-time monitoring path utilization rate exceeds a preset critical buffer threshold, newly added tasks will be switched to the next available path in the ranking, resulting in a final task handling scheme that adapts to production load requirements, including: Real-time monitoring of the current utilization rate of each path in the priority path ranking, the utilization rate is the ratio of the currently occupied bandwidth of the path to the total bandwidth capacity; If the utilization rate of a certain path exceeds the preset critical buffer threshold, it is marked as saturated, and the available path immediately next to the current path in the priority path ranking is obtained as the switching target path. The newly added task data stream is directed to the target path for switching transmission, the unallocated quota record of each path in the task scheduling queue is updated, and the remaining unallocated quota of the saturated path is synchronously transferred to the target path for switching. If the utilization rate of the target path also exceeds the preset critical buffer threshold, then the process will be sequentially extended to the available paths in the subsequent priority levels until an available path that meets the bandwidth requirements is found. When the utilization rate of a saturated path falls below the critical buffer threshold and the congestion is relieved, the path's original priority and task allocation authority in the priority path sorting are restored, ultimately resulting in a final task-bearing scheme that adapts to the production load requirements.

Images