Enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent scheduling, in particular to an enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence, which comprises a multi-source heterogeneous data fusion unit, a priority resource coupling decision unit, a bottleneck prediction and tracing unit and a scheduling instruction generation unit, the multi-source heterogeneous data fusion unit collects multi-dimensional data such as equipment vibration, temperature, order delivery time and the like in real time and constructs a joint feature vector, and the priority resource coupling decision unit dynamically adjusts task priority and resource allocation through a dual-channel depth Q network to cope with order insertion tasks and equipment health degree fluctuation. The bottleneck prediction and tracing unit predicts production bottlenecks and traces root causes by using a process dependency graph, a multi-modal fusion model and a causal discovery algorithm, and supports preventive maintenance and dynamic scheduling, and the scheduling instruction generation unit synthesizes a preorder result to generate an adaptive scheduling instruction. And enterprise production equipment utilization rate and production efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent scheduling technology, and in particular to an enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence. Background Art

[0002] Intelligent scheduling is an important technology. In the production process of enterprises, multi-order mixed production scenarios are becoming increasingly common, which puts higher demands on the flexibility and accuracy of production scheduling systems. However, because traditional systems rely on fixed priority rules and manual experience, they lack the real-time perception and intelligent processing capabilities of dynamic factors such as equipment status fluctuations and urgent order changes. Traditional production scheduling systems have significant defects in this scenario.

[0003] It is difficult to respond in real time to dynamic factors such as equipment status fluctuations and urgent order changes. When sudden equipment failures hinder production progress or urgent order insertions require readjustment of the production sequence, traditional systems, due to their over-reliance on fixed priority rules and manual experience, are unable to adjust scheduling strategies in a timely manner based on real-time data such as equipment health and order urgency. This can easily cause equipment to be idle or overloaded, resulting in low equipment utilization and reduced production efficiency. To address this technical problem, we provide an AI-based enterprise production real-time monitoring and intelligent scheduling system. Summary of the Invention

[0004] The purpose of the present invention is to provide an enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence to solve the problems raised in the above background technology.

[0005] 1. Because traditional systems rely on fixed priority rules and manual experience, they struggle to respond promptly to fluctuations in equipment status, resulting in low equipment utilization. Therefore, this case uses a multi-source heterogeneous data fusion unit to collect equipment data. The priority resource coupling decision unit utilizes a dual-channel deep Q network to output a dynamic scheduling solution. This enables real-time perception of equipment status, optimized scheduling strategies, and improved equipment utilization.

[0006] 2. Because traditional systems cannot respond to urgent order changes in real time, which can easily cause production chaos, this case establishes a dynamic weight adjustment mechanism in the priority resource coupling decision unit. This generates preemptive weight coefficients based on order delivery dates to adjust the ranking. This can quickly respond to order changes, ensure orderly production, and improve production efficiency.

[0007] To achieve the above objectives, an artificial intelligence-based enterprise production real-time monitoring and intelligent scheduling system is provided, including;

[0008] The multi-source heterogeneous data fusion unit collects real-time vibration, temperature, and current time series data of production line equipment through edge computing nodes, and aligns them with the order delivery data of the manufacturing execution system and the material completeness data of the warehouse management system in time and space to construct a joint feature vector.

[0009] The priority resource coupling decision unit inputs the joint feature vector into a dual-channel deep Q network for multi-objective decision-making, outputs a dynamic scheduling baseline plan, and marks resource conflict risk points. The joint feature vector includes the order urgency score, the equipment health index, and the material readiness coefficient. The dual channels include a priority channel and a resource allocation channel. The resource conflict risk point is when a high-load device is assigned an urgent order.

[0010] The bottleneck prediction and tracing unit receives the dynamic scheduling baseline plan and equipment health index, constructs a process dependency graph and bottleneck probability calculation to output the blocking probability of each node, marks bottleneck equipment with probabilities exceeding the threshold, and finally inputs vibration spectrum data and material readiness coefficient into the causal discovery algorithm to determine the root cause.

[0011] The scheduling instruction generation unit generates scheduling instructions according to the dynamic scheduling baseline plan, bottleneck prediction results and root causes.

[0012] As a further improvement of this technical solution, the priority channel of the dual-channel deep Q network in the priority resource coupling decision unit includes a dynamic weight adjustment mechanism:

[0013] When an insertion task is detected, a preemptive weight coefficient is generated based on the ratio of the remaining delivery time of the order to the current highest priority task, which is used to adjust the insertion task sorting;

[0014] If the device health index is lower than the preset threshold, an exponential decay penalty term is applied to the device's assignable tasks to avoid assigning critical tasks to high-risk devices.

[0015] As a further improvement of this technical solution, the load balancing strategy of the resource allocation channel in the priority resource coupling decision unit includes:

[0016] The load balancing factor is calculated based on the variance of the remaining capacity of the equipment group. If the variance exceeds the threshold, the task migration mechanism is triggered, and the tasks of high-load equipment are reversely allocated to low-load equipment according to the health index. Combined with the time-of-use electricity price strategy, tasks are preferentially allocated to low-power equipment groups during peak electricity price periods.

[0017] As a further improvement of this technical solution, the resource conflict risk point marking method in the priority resource coupling decision unit includes:

[0018] Define the conflict level, where red represents high load + emergency order, yellow represents single indicator overrun, when red warning, automatically split the task to adjacent equipment group, and adjust the material distribution path synchronously.

[0019] As a further improvement of the technical solution, the training method of the double-channel deep Q network in the priority resource coupling decision unit comprises:

[0020] The priority channel adopts a customer level weighted reward function, and the VIP order delivery reward value is increased by 2 times. The resource allocation channel introduces load balancing and energy consumption penalty term, and gradually increases multi-objective constraints through curriculum learning.

[0021] As a further improvement of the technical solution, the construction method of the process dependency graph in the bottleneck prediction and tracing unit comprises:

[0022] Based on the graph neural network, the pre-post relationship of the process in the process route document is embedded, the node represents the device type and the historical failure rate, and the edge weight is dynamically corrected by the historical average switching time consumption and temperature change rate.

[0023] As a further improvement of the technical solution, the blocking probability calculation method in the bottleneck prediction and tracing unit comprises:

[0024] A multi-modal fusion prediction model is adopted, in which a time series convolution network analyzes device sensor data, and a deep learning model based on attention mechanism captures order task sequence dependency relationship;

[0025] Based on the historical failure mode library matching the same scene, a probability correction coefficient is applied to the prediction result, wherein when the failure rate of the same type of equipment is greater than 30%, the blocking probability is increased by 15%.

[0026] As a further improvement of the technical solution, the causal discovery algorithm in the bottleneck prediction and tracing unit comprises:

[0027] A causal analysis algorithm and a domain knowledge fusion framework are adopted, and three types of root cause hypotheses, device degradation, material shortage and process conflict, are defined.

[0028] If the vibration spectrum data is abnormal for more than 3 sampling periods and the material preparation degree is greater than the threshold, the material factor is excluded, and the device performance degradation is locked as the root cause.

[0029] As a further improvement of the technical solution, the root cause determination rule in the bottleneck prediction and tracing unit comprises:

[0030] The device degradation root cause triggers a preventive maintenance work order and enables redundant devices, the material shortage root cause dynamically adjusts the automatic guided vehicle path priority and releases safety stock, and the process conflict root cause rearranges the production sequence and inserts buffer time.

[0031] As a further improvement of this technical solution, the bottleneck prediction model in the bottleneck prediction and tracing unit includes an online update mechanism:

[0032] When the deviation between the actual blocking event and the predicted result is greater than 20%, the incremental learning process is triggered to update the weights of the temporal convolutional network model. The bottleneck prediction model size is compressed through knowledge distillation technology to adapt to the resource limitations of edge computing nodes.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] In the real-time monitoring and intelligent scheduling system for enterprise production based on artificial intelligence, the priority resource coupling decision unit makes multi-objective decisions through a dual-channel deep Q network, dynamically adjusts task priorities and resource allocation, and can not only cope with fluctuations in insertion tasks and equipment health, but also achieve load balancing and energy consumption optimization, effectively improving equipment utilization and production efficiency. The bottleneck prediction and tracing unit uses process dependency graphs, multimodal fusion models and causal discovery algorithms to accurately predict blocking probabilities and trace root causes, and combines online update mechanisms to continuously optimize the model, achieving early warning and rapid disposal of bottlenecks, thereby improving the company's production efficiency and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A collaboration diagram of the priority resource coupling decision unit and the bottleneck prediction and tracing unit of the present invention;

[0036] Figure 2 It is an overall block diagram of the present invention.

[0037] The meaning of each number in the figure is:

[0038] 1. Multi-source heterogeneous data fusion unit; 2. Priority resource coupling decision unit; 3. Bottleneck prediction and tracing unit; 4. Scheduling instruction generation unit. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] This invention provides an enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence, please refer to Figure 1-Figure 2 Shown, including:

[0041] The multi-source heterogeneous data fusion unit 1 collects the production line equipment vibration, temperature, and current time series data in real time through the edge computing node, and performs space-time alignment with the order delivery date data of the manufacturing execution system and the material completeness data of the warehouse management system, constructs a joint feature vector, as follows:

[0042] A unified time reference is established, and the network time protocol is used to synchronize the clocks of the edge computing node, the manufacturing execution system, and the warehouse management system. For order delivery date data, the order start execution time is used as the time anchor point, which is matched with the equipment data collection time. The material completeness data is aligned with the equipment data according to the material demand planning time. The dynamic time warping algorithm is used to perform elastic alignment of the equipment time series data and the order and material data in the time sequence, solving the problem of inconsistent data collection frequency, making different types of data correspond in the time dimension, eliminating data time differences, and achieving precise matching of multi-source data in the time dimension. Time domain features are extracted from the equipment vibration data, frequency domain features are obtained through fast Fourier transform, the mean and standard deviation of temperature and current data are extracted, the remaining time and urgency level of the order delivery date data are extracted, the number of missing material types and the material delay probability of the material completeness data are extracted, the features of the above different types of data are normalized, the minimum-maximum normalization method is used to map all feature values to the [0, 1] interval, and then combined into a joint feature vector in a predetermined order, providing high-quality data input for the production decision analysis model.

[0043] The priority resource coupling decision unit 2 inputs the joint feature vector into the double-channel deep Q network for multi-objective decision, outputs the dynamic scheduling baseline scheme and labels the resource conflict risk points, the joint feature vector includes order urgency score, equipment health index, and material preparation coefficient, the double-channel includes priority channel and resource allocation channel, and the resource conflict risk point is a high-load equipment assigned to an urgent order.

[0044] The priority channel of the double-channel deep Q network in the priority resource coupling decision unit 2 contains a dynamic weight adjustment mechanism:

[0045] When an order insertion task is detected, a pre-emptive weight coefficient is generated according to the ratio of the order delivery remaining time to the current highest priority task, which is used to adjust the order insertion task ranking, as follows:

[0046] A due date ratio factor R is defined, where R = Tinsertion / Thighest, where Tinsertion is the remaining duration of the insertion task and Thighest is the remaining duration of the current highest priority task. A nonlinear mapping function Wpreemption is designed, where Wpreemption = e(k・(1-R)), where k is the adjustment coefficient and ranges from [3, 5], and e is a constant. When R < 1, it indicates that the insertion task is urgent, in which case Wpreemption > 1; when R > 1, Wpreemption < 1. The order attribute factors are integrated to obtain the final weight, which is: basic weight × Wpreemption × order value coefficient × customer level coefficient. An independent "urgent task evaluation layer" is added to the deep Q network. The input is the order feature vector, and the output is the adjusted priority score. A dual time scale update strategy is adopted, where the main network is updated every 100 steps and the urgent task evaluation layer is updated every 10 steps. This ensures a rapid response to the insertion task. The dual time scale update balances decision stability and real-time performance.

[0047] If the device health index is lower than the preset threshold, an exponential decay penalty term is applied to the device's assignable tasks to avoid assigning critical tasks to high-risk devices. The details are as follows:

[0048] A multi-parameter evaluation index HI is constructed, where HI = 0.4 × vibration severity + 0.3 × temperature anomaly + 0.2 × current fluctuation coefficient + 0.1 × cumulative operating time. Three health thresholds are set, including green (HI < 0.3), yellow (0.3 ≤ HI < 0.7), and red (HI ≥ 0.7). A penalty function P is defined, where P = y(α・(HI - H threshold)), α is the attenuation coefficient with a value range of [2, 3], and y is a constant. When HI ≥ H threshold, P ≥ 1, and task allocation is adjusted. The device allocation probability = basic probability / P. An allocation probability attenuation of ≥ 75% is applied to red devices. A "device risk perception layer" is added to the priority channel. The device health characteristics and task characteristics are cross-attentionally calculated. A risk-sensitive loss function L is designed, L = LQ + λ・∑(HIi・Pi), where LQ is the standard Q learning loss and λ is the risk. Weight coefficient, HIi is the evaluation index of a certain parameter, Pi is the penalty function of a certain parameter, and the exponential penalty function imposes strong constraints on high-risk devices. The cross-attention mechanism achieves intelligent matching of tasks and devices. Single mechanism optimization has limitations, and collaborative optimization can achieve global optimal resource allocation. The comprehensive objective function J is defined, where J = w1・∑ completion rate + w2・∑(1 / average delay) - w3 × ∑ device risk. Among them, w1, w2, and w3 are weight coefficients. The weights are adjusted dynamically: w1 = 0.5 + 0.3 × order urgency index, w2 = 0.3 - 0.1 × order urgency index, and w3 = 0.2. The underlying feature extraction layer is shared in the dual-channel network, and parameter distillation technology is used to transfer device risk knowledge to the order insertion decision module. When a drastic change in device health is detected, the learning rate is increased by 50%. When the order insertion frequency exceeds the threshold, the exploration rate is reduced to 0.05.

[0049] The load balancing strategy of the resource allocation channel in the priority resource coupling decision unit 2 includes:

[0050] The load balancing factor is calculated based on the variance of the remaining capacity of the equipment group, as follows:

[0051] Establish an equipment surplus capacity assessment model, comprehensively consider the equipment's current task volume, rated production capacity, and expected idle time, and calculate the surplus capacity of each device. Residual capacity = rated capacity - currently assigned task volume - expected subsequent natural consumption capacity. The expected subsequent natural consumption capacity is predicted based on historical data. The variance of surplus capacity is calculated by equipment group, and dynamic thresholds are set. The thresholds are adjusted according to the size of the equipment group and the characteristics of the production task. For example, the threshold for small equipment groups is set to 20% of the average surplus capacity, and for large equipment groups it is set to 15%. When the variance exceeds the threshold, the equipment group is deemed to have an unbalanced load, triggering the task migration mechanism. The dynamic threshold adapts to different production scenarios and accurately identifies the equipment groups that need to be adjusted.

[0052] If the variance exceeds the threshold, the task migration mechanism is triggered, and the tasks of high-load devices are reversely allocated to low-load devices according to the health index. Combined with the time-of-use electricity price strategy, tasks are preferentially allocated to low-power device groups during peak electricity price periods. The details are as follows:

[0053] Build an equipment health index evaluation system that integrates equipment vibration, temperature, and energy consumption. Calculate the health index through weighted summation: health index = 0.4 × vibration normality score + 0.3 × temperature normality score + 0.2 × energy consumption normality score + 0.1 × historical failure frequency score. The higher the score, the healthier the equipment. Filter tasks to be migrated from high-load equipment and assign them to low-load equipment in descending order of equipment health index, taking into account both load balancing and equipment health, prioritize protecting key equipment, reduce the risk of equipment failure, and access the power department's real-time electricity price data interface. Obtain accurate time-of-use electricity price information, including peak and valley periods and corresponding electricity prices, mark the power consumption level for each equipment group, calculate the average power consumption through the equipment's historical energy consumption data and rated power, and divide the equipment into high, medium, and low power consumption groups. During peak electricity price periods, adjust the task allocation strategy and give priority to flexibly scheduled tasks to low-power equipment groups. For tasks that must be completed by high-power consumption equipment, try to arrange them for execution during valley periods. At the same time, establish a task scheduling priority queue, combine the task urgency and equipment status, and comprehensively determine the task execution order to optimize energy use and reduce enterprise electricity costs.

[0054] The resource conflict risk point marking method in the priority resource coupling decision unit 2 includes:

[0055] Define conflict levels, where red represents high load + urgent orders, defined as equipment load rate > 90% and associated urgent orders, where urgent orders are orders with a ratio of remaining delivery time to estimated production time < 1.2; yellow represents a single indicator exceeding the limit, equipment load rate > 80%, order urgency reaches the highest level, material shortage risk exceeds the threshold, and there is a potential risk of production interruption. Real-time equipment load data is collected through IoT sensors to calculate the proportion of equipment operating time, order urgency information is obtained from the manufacturing execution system, and material status data is obtained from the warehouse management system based on order priority tags, remaining delivery time, and a streaming computing engine is deployed to scan data once per second. The conflict level is quickly determined based on preset rules and real-time risk annotations are generated. When a red alert is issued, the task is automatically split to adjacent equipment groups, and the material distribution path is adjusted synchronously. Based on process divisibility analysis, the task is broken down into minimum execution units. A genetic algorithm is used to optimize the splitting plan, with the goal of minimizing task completion time and balancing equipment load, and the optimal split ratio and sub-tasks are calculated. The system combines task allocation and construction, builds a capacity assessment model for adjacent equipment groups, and comprehensively considers the equipment's available capacity, processing accuracy, and current load to screen out 3-5 optimal equipment groups for the task. Task migration requests are sent to the receiving equipment groups, and confirmation of receipt is obtained from the equipment groups. This fully utilizes idle capacity and reduces the risk of overloading a single piece of equipment. A factory logistics network model is established, including warehouses, equipment groups, and transfer point nodes and corresponding paths, with the distance of each path marked. The Dijkstra algorithm, a classic algorithm for finding the shortest path in graph theory, is used to replan the optimal material distribution path for the migration task. Through the data interface between the manufacturing execution system and the warehouse management system, task migration information and new path instructions are synchronized to the material distribution scheduling center. After receiving the instructions, the delivery vehicle automatically updates the navigation path and feedbacks the execution status through the IoT terminal. If the path is blocked, secondary path planning is triggered, achieving coordinated adjustment of tasks and material distribution to avoid the impact of material delays on production progress.

[0056] The training method of the dual-channel deep Q network in the priority resource coupling decision unit 2 includes:

[0057] The priority channel uses a customer-level weighted reward function, and the VIP order delivery reward value is doubled. The resource allocation channel introduces load balancing and energy consumption penalty items, as follows:

[0058] Customers are divided into four levels: ordinary, silver, gold, and VIP. Order delivery reward coefficients are set to 1, 1.5, 1.8, and 2 respectively. When order data is accessed, each order is labeled with the corresponding customer level label, and a task completion reward function is designed. Task completion reward function = customer level reward coefficient × (on-time delivery reward baseline value - delayed delivery penalty value). Among them, a reward of 5 times the corresponding coefficient will be deducted for a delay of 1 hour. When a VIP order is delivered on time, the reward value is increased by 2 times compared to an ordinary order. The model's priority decision on high-value orders is strengthened. In the early stage of training, the weight of the penalty item is reduced, and the focus is on rewarding positive delivery behavior, so that the model can be fast. Quickly learn the basic order scheduling logic. As the training progresses, gradually increase the weight of the penalty item to improve the model's ability to avoid delivery risks. Define the equipment group load balancing penalty function, load balancing penalty function = penalty coefficient × equipment group remaining capacity variance. When the variance exceeds the set threshold, impose a penalty on the decision action to encourage the model to assign tasks to low-load equipment. Construct an energy consumption penalty function, energy consumption penalty function = energy consumption penalty coefficient × the expected energy consumption of the equipment to perform the task. Among them, the energy consumption penalty coefficient is dynamically adjusted according to the time-of-use electricity price. The coefficient of the peak electricity price period is increased by 50%. Giving priority to low-power equipment to perform tasks can reduce penalties and guide the model. The model reduces production energy consumption during peak electricity price periods, combines the two penalty items with the task completion reward into a total reward function, the total reward function = task completion reward function - energy consumption penalty function - load balancing penalty function, updates the network parameters through the gradient descent algorithm, enables the model to automatically weigh task progress, load balancing and energy consumption costs during training, and promotes the model to generate a better resource allocation plan. Through course learning, multi-objective constraints are gradually added, and the training process is divided into three stages. In the first stage, only basic task completion rewards are enabled, and the training model masters the basic decision-making logic. In the second stage, a single penalty item is introduced, such as adding load balancing penalties first, so that the model learns After the initial multi-objective trade-off, all rewards and penalties are combined in the third stage to fully train the model's ability to handle complex multi-objective problems. A training round threshold is set for each stage. After completing the preset rounds of the previous stage and meeting the performance indicators, such as when the order delivery accuracy is greater than 85%, it automatically enters the next stage. At the same time, the penalty coefficient is dynamically adjusted, gradually increasing from the initial small value to avoid the model from being difficult to converge due to the sudden increase in constraints. When the stage transitions, the network layer parameters that were stable in the previous stage training are frozen, and only the network layers related to the newly added constraints are updated. The learned knowledge is used to accelerate the learning of new goals, prevent catastrophic forgetting, and improve the multi-objective optimization effect.

[0059] The bottleneck prediction and tracing unit 3 receives the dynamic scheduling baseline plan and the equipment health index, and outputs the blocking probability of each node by constructing a process dependency graph and bottleneck probability calculation, marking the bottleneck equipment with a probability exceeding the threshold, and finally inputs the vibration spectrum data and material readiness coefficient into the causal discovery algorithm to determine the root cause.

[0060] The method for constructing the process dependency graph in the bottleneck prediction and tracing unit 3 includes:

[0061] Based on the graph neural network embedded in the process route document, the process relationship between the previous and next steps is expressed. The nodes represent the equipment type and historical failure rate. The edge weight is dynamically adjusted based on the historical average switching time and temperature change rate. The details are as follows:

[0062] The BERT-CRF model is used to perform named entity recognition on process route documents, extract process names, equipment types, and process parameters. The training data set contains 2,000 annotated process documents, and builds a process relationship template library, which includes relationship description patterns such as "immediately afterwards", "need to be completed first", and "performed in parallel". The pre- and post-process relationships are automatically identified through semantic similarity matching, and each process is mapped to a node in the graph. The node feature vector contains the equipment type and historical failure rate encoded in a unique hot encoding. The initial edge is generated based on the identified process relationship, and the default edge weight is set to 1, indicating the basic dependency strength. Traditional graph representation methods cannot fully capture the complex relationship between equipment type and failure rate. Graph neural networks can learn more expressive nodes. Point embedding adopts a two-layer graph convolution structure. The first layer aggregates information of directly adjacent nodes, and the second layer aggregates information of second-order neighbors. The attention mechanism is introduced to assign different weights to different neighbor nodes. Static edge weights cannot reflect the actual difficulty and risk of process switching. Dynamic correction can improve the accuracy of bottleneck prediction. Production logs for the past 12 months are collected to record the actual switching time and equipment temperature change data between each pair of processes. The sampling frequency is 1 minute. The average switching time and temperature change rate are calculated, and outliers are filtered. The average switching time and temperature change rate are recalculated after every 100 process switches. When equipment maintenance or process adjustment is detected, the edge weight is adaptively adjusted by ±15% of the original weight.

[0063] The blocking probability calculation method in the bottleneck prediction and tracing unit 3 includes:

[0064] A multimodal fusion prediction model is used, in which a temporal convolutional network analyzes device sensor data, and a deep learning model based on an attention mechanism captures the order task sequence dependencies, as follows:

[0065] The input layer receives equipment vibration, temperature, and current data at a sampling frequency of 10 Hz. The window length is set to 3000 points, covering 5 minutes of data. An expanded convolution structure is used with expansion coefficients of 1, 2, 4, and 8, respectively, to capture features at different time scales. Each convolution layer is followed by a BatchNorm and ReLU activation function. The output layer generates a device state feature vector with a dimension of 64. A bidirectional long-term and short-term attention network is used to process order task sequences. The input includes order priority, process complexity, and estimated time, and a self-attention mechanism is applied. The output layer uses a Sigmoid function to predict the basic blocking probability. Then, based on the historical failure mode library, the same scenario is matched and a probability correction factor is applied to the prediction result. When the failure rate of similar equipment is greater than 30%, the blocking probability is increased by 15%, as follows:

[0066] Equipment failure records from the past three years are collected, and the equipment status characteristics, order task characteristics, and environmental parameters before the failure are extracted to form a failure mode library for the case. The cosine similarity between the current scenario and the historical pattern is calculated. When the similarity exceeds the threshold, it is determined to be a successful match. For the matched historical pattern, its corresponding failure probability and equipment type failure rate are extracted. When the equipment type failure rate is greater than 30%, an additional 15% blocking probability is added as a safety margin to enhance risk response capabilities.

[0067] The causal discovery algorithms in Bottleneck Prediction and Tracing Unit 3 include:

[0068] The causal analysis algorithm and the domain knowledge fusion framework are adopted, three types of root cause hypotheses including equipment degradation, material shortage and process conflict are defined, a knowledge graph is constructed, for example: "equipment vibration anomaly→equipment performance degradation→production bottleneck", "low material set rate→material shortage→production stagnation", the confidence and impact weight of each relationship are labeled, the parameters are determined through expert evaluation and historical data verification, the causal structure learning is performed by using the causal analysis algorithm, the significance level of conditional independence test is set, the Granger causality test is combined to analyze the causal relationship between time series variables, the lag order is set to 10, corresponding to 5 minutes of data, the wavelet packet decomposition technology is used to decompose the vibration signal into 8 frequency bands, the energy proportion of each frequency band is calculated, an anomaly detector is constructed, the training set contains 10000 normal samples, an abnormal score threshold is set, a duration counter is designed, when an anomaly is detected for 3 consecutive sampling periods, an equipment degradation warning is triggered, each period is 30 seconds, a material preparation degree index is defined, material preparation degree index = actual arrival quantity / planned usage quantity × 0.6 + qualified inspection quantity / actual arrival quantity × 0.4, a threshold is set, when the material preparation degree index is greater than the threshold, it is determined that the material is sufficient, and a decision rule is designed, if the vibration spectrum data anomaly lasts more than 3 sampling periods and the material preparation degree is greater than the threshold, the material factor is excluded, and the equipment performance degradation is locked as the root cause, when multiple root cause hypotheses are satisfied at the same time, the hypothesis with the highest score is selected preferentially, and other hypotheses are recorded as secondary causes, the explicit decision rule improves the automation degree, the exclusion mechanism reduces the misjudgment, a 1-minute update cycle is set, the causal strength is recalculated based on the latest sensor data and production state, when a new abnormal event is detected, the local recalculation of the causal chain is triggered, only the affected part is updated, a force-directed graph layout algorithm is used, the node size represents the abnormal severity, the edge thickness represents the causal strength, the time slider function is realized, the historical causal chain evolution process is supported to be viewed, the time granularity is 5 minutes, natural language explanations are automatically generated based on the causal chain, for example: "equipment X vibration spectrum anomaly and abnormal score 0.85 for 180 seconds, and material preparation degree 0.92, the root cause is determined to be equipment performance degradation".

[0069] The root cause determination rules in the bottleneck prediction and tracing unit 3 include:

[0070] The equipment degradation root cause triggers a preventive maintenance work order and enables redundant equipment, the material shortage root cause dynamically adjusts the automatic guided vehicle path priority and releases safety stock, the process conflict root cause rearranges the production sequence and inserts buffer time, which are as follows:

[0071] A device health assessment model is built to calculate a health index based on vibration, temperature, and current data. When the health index is below 40 points and persists for one hour, a preventive maintenance work order is automatically generated. The work order includes the device number, details of abnormal parameters, and recommended maintenance measures. The work order is then pushed to the maintenance team's mobile device through the enterprise asset management system. A 30-minute response time limit is set for reminders. A redundant device resource pool is established to monitor the availability of each redundant device in real time, including maintenance cycles and remaining capacity. When a degradation warning is triggered for a primary device, the system prioritizes a redundant device of the same type with a health score above 80 and closest to the primary device. Control instructions are automatically switched using the Industrial Internet of Things protocol, completing the device switch within five minutes. Equipment allocation information is also updated in the production scheduling system, reducing manual intervention costs. A real-time material demand map is constructed, integrating production plans, material inventory locations, the current location of automated guided vehicles, and task queue information. When a material shortage is detected (i.e., inventory below the safety threshold and replenishment delays greater than two hours), the system prioritizes the transportation of the corresponding material to the highest priority. A path planning algorithm is used to replan the AGV's route to prioritize the delivery of the shortfall, while also sending avoidance commands to other AGVs. To ensure smooth transportation routes, a dynamic safety stock threshold is set for each material. When the actual inventory reaches the threshold, the system automatically generates a safety stock release application. After the application is approved online by the production supervisor, the picking process of the warehouse management system is triggered. The picking instructions are simultaneously pushed to the automatic guided vehicle scheduling system, which prioritizes the allocation of vehicles to transport the released safety stock materials to the production line to ensure uninterrupted production. Based on the process dependency graph and the current production progress, a genetic algorithm is used to optimize the production sequence, with order delivery time, equipment utilization, and process switching cost as optimization targets. The crossover probability is set to 0.8 and the mutation probability is set to 0.2. , iterate 100 times to generate the optimal production sequence, and the rearrangement result is sent to each workstation terminal through the manufacturing execution system, synchronously updating the task progress and equipment scheduling information of the production board, calculating the total duration deviation of the task chain involved in the process conflict, and dynamically inserting buffer time according to the deviation value, where buffer time = deviation value × 1.2, and the minimum insertion is 15 minutes. In the rearranged production sequence, the buffer time is evenly distributed to the critical path nodes by adjusting the task start time or extending the process interval. At the same time, adjustment notifications are sent to the affected upstream and downstream processes to ensure the synchronization of production rhythm and enhance the flexibility of production plan.

[0072] The bottleneck prediction model in Bottleneck Prediction and Traceability Unit 3 includes an online update mechanism:

[0073] When the actual blocking event deviates from the predicted result by more than 20%, the incremental learning process is triggered to update the temporal convolutional network model weights. The bottleneck prediction model size is compressed through knowledge distillation technology to adapt to the resource constraints of edge computing nodes. The details are as follows:

[0074] The deviation between the prediction result and the actual blocking event is counted every hour. When the deviation between the actual blocking event and the prediction result is greater than 20%, it is marked as an anomaly and the abnormal data sample is recorded, including input features such as device status, order tasks and actual blocking results. An abnormal data cache queue is built. When the number of abnormal samples in the queue reaches 50, the incremental learning process is triggered. Before the incremental learning starts, the new samples are cleaned and the temporal convolutional network model is divided into a feature extraction layer and a prediction layer. During incremental learning, the weights of the first three layers of the feature extraction layer are frozen first, and only the parameters of the last two layers and the prediction layer are updated to avoid forgetting the original knowledge due to over-learning of new samples. The newly collected abnormal samples and historical training data are compared by 1: A new training dataset is constructed by mixing the two models in a three-fold ratio. Auxiliary data such as equipment maintenance records and order priority changes is also introduced to enrich the feature dimensions of the training data. During training, the prediction accuracy on the validation set is calculated after every two training cycles. If the accuracy improvement is less than 1%, the learning rate is reduced and training continues to prevent the model from falling into a local optimum. The updated full-time convolutional network model is used as the teacher model, and a lightweight student model with one-third the parameter size of the teacher model is constructed. This model uses the same network architecture but reduces the number of convolution kernels and hidden layer neurons. A distillation loss function is defined, and a temperature parameter is used to soften the probability distribution of the teacher model's output, allowing the student model to learn richer knowledge. During training, each batch of data is first passed through the teacher model to generate soft labels, which are then combined with the true labels to guide the student model's training. The training process monitors the difference in prediction accuracy between the student model and the teacher model on the test set. Training is terminated when the gap is less than 5%. If the gap persists above 10%, the student model structure is adjusted and training is restarted.

[0075] The scheduling instruction generation unit 4 generates a scheduling instruction according to the dynamic scheduling baseline plan, the bottleneck prediction result and the root cause.

[0076] In the present invention, the multi-source heterogeneous data fusion unit 1 collects multi-dimensional data such as equipment vibration, temperature, order delivery time, etc. in real time and constructs a joint feature vector. The priority resource coupling decision unit 2 dynamically adjusts task priority and resource allocation through a dual-channel deep Q network to cope with insertion tasks and fluctuations in equipment health, and achieves load balancing and energy consumption optimization. The bottleneck prediction and tracing unit 3 uses process dependency graphs, multimodal fusion models and causal discovery algorithms to predict production bottlenecks and trace the root causes, support preventive maintenance and dynamic scheduling, and the scheduling instruction generation unit 4 generates adaptive scheduling instructions based on previous results to improve the utilization rate and production efficiency of enterprise production equipment.

[0077] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence is characterized by: include: The multi-source heterogeneous data fusion unit (1) collects the vibration, temperature, and current time series data of the production line equipment in real time through the edge computing node, and aligns them with the order delivery data of the manufacturing execution system and the material completeness data of the warehouse management system in time and space to construct a joint feature vector; The priority resource coupling decision unit (2) inputs the joint feature vector into a dual-channel deep Q network for multi-objective decision-making, outputs a dynamic scheduling baseline plan and marks resource conflict risk points, wherein the joint feature vector includes an order urgency score, an equipment health index, and a material readiness coefficient, the dual channels include a priority channel and a resource allocation channel, and the resource conflict risk point is that a high-load equipment is assigned an urgent order; The bottleneck prediction and tracing unit (3) receives the dynamic scheduling baseline plan and the equipment health index, and outputs the blocking probability of each node by constructing a process dependency graph and bottleneck probability calculation, marking the bottleneck equipment with a probability exceeding the threshold, and finally inputs the vibration spectrum data and material readiness coefficient into the causal discovery algorithm to determine the root cause; The scheduling instruction generation unit (4) generates scheduling instructions according to the dynamic scheduling baseline plan, bottleneck prediction results and root causes.

2. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The priority channel of the dual-channel deep Q network in the priority resource coupling decision unit (2) includes a dynamic weight adjustment mechanism: When an insertion task is detected, a preemptive weight coefficient is generated based on the ratio of the remaining delivery time of the order to the current highest priority task, which is used to adjust the insertion task sorting; If the device health index is lower than the preset threshold, an exponential decay penalty term is applied to the device's assignable tasks to avoid assigning critical tasks to high-risk devices.

3. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The load balancing strategy of the resource allocation channel in the priority resource coupling decision unit (2) includes: The load balancing factor is calculated based on the variance of the remaining capacity of the equipment group. If the variance exceeds the threshold, the task migration mechanism is triggered, and the tasks of high-load equipment are reversely allocated to low-load equipment according to the health index. Combined with the time-of-use electricity price strategy, tasks are preferentially allocated to low-power equipment groups during peak electricity price periods.

4. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The resource conflict risk point marking method in the priority resource coupling decision unit (2) includes: Define conflict levels, where red indicates high load + urgent orders, and yellow indicates a single indicator exceeds the limit. When a red warning is issued, the task is automatically split to adjacent equipment groups, and the material distribution path is adjusted synchronously.

5. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The training method of the dual-channel deep Q network in the priority resource coupling decision unit (2) includes: The priority channel adopts a customer level weighted reward function, the VIP order delivery reward value is increased by 2 times, the resource allocation channel introduces load balancing and energy consumption penalty items, and multi-objective constraints are gradually added through course learning.

6. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The method for constructing the process dependency graph in the bottleneck prediction and tracing unit (3) includes: Based on the graph neural network, the process front-end and back-end relationships in the process route document are embedded. The nodes represent the equipment type and historical failure rate, and the edge weights are dynamically modified based on the historical average switching time and temperature change rate.

7. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The blocking probability calculation method in the bottleneck prediction and tracing unit (3) includes: A multimodal fusion prediction model is used, in which a temporal convolutional network analyzes device sensor data, and a deep learning model based on an attention mechanism captures the order task sequence dependencies; Based on the historical failure mode library and matching the same scenarios, a probability correction coefficient is applied to the prediction results. When the failure rate of similar equipment is greater than 30%, the blocking probability increases by 15%.

8. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The causal discovery algorithm in the bottleneck prediction and tracing unit (3) includes: Using a causal analysis algorithm and a domain knowledge fusion framework, we defined three root cause hypotheses: equipment degradation, material shortages, and process conflicts. If the vibration spectrum data abnormality persists for more than three sampling cycles and the material readiness is greater than the threshold, eliminate the material factor and identify equipment performance degradation as the root cause.

9. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The root cause determination rules in the bottleneck prediction and tracing unit (3) include: The root cause of equipment degradation is the triggering of preventive maintenance work orders and the activation of redundant equipment. The root cause of material shortage is the dynamic adjustment of automated guided vehicle path priorities and the release of safety stock. The root cause of process conflict is the rescheduling of production sequences and the insertion of buffer time.

10. The enterprise production real-time monitoring and intelligent scheduling system based on artificial intelligence according to claim 1 is characterized in that: The bottleneck prediction model in the bottleneck prediction and tracing unit (3) includes an online update mechanism: When the deviation between the actual blocking event and the predicted result is greater than 20%, the incremental learning process is triggered to update the weights of the temporal convolutional network model. The bottleneck prediction model size is compressed through knowledge distillation technology to adapt to the resource limitations of edge computing nodes.

Images