Production resource management system based on AI neural network

Abstract

The invention relates to the technical field of production resource management, and discloses a production resource management system based on an AI neural network. The system comprises a multi-source data acquisition module for acquiring multi-source heterogeneous data of a production environment; the multi-dimensional feature extraction module is used for extracting feature vectors of each data source; the heterogeneous data fusion module is used for generating a resource state semantic representation based on a cross-modal attention mechanism; the resource topological graph construction module is used for constructing a double-layer topological library; and the intelligent optimization decision module is used for generating a real-time resource scheduling strategy and a task allocation optimization scheme based on the result. In addition, the system also has the functions of anomaly detection, path re-planning, resource aging prediction and the like. The system can comprehensively process production data, optimize resource scheduling and task allocation, improve the resource utilization rate, shorten the task completion time, enhance the stability of the production process and the scientificity of equipment maintenance, and improve the production management level of enterprises.

Description

Technical Field

[0001] The present invention relates to the technical field of production resource management, and specifically to a production resource management system based on an AI neural network. Background Art

[0002] In the modern production and manufacturing field, with the continuous expansion of production scale and the increasing complexity of production processes, production resource management faces many severe challenges. The traditional production resource management methods can no longer meet the needs of enterprises for efficient operation, and a series of problems have gradually emerged.

[0003] In terms of data collection, the data in the production environment shows the characteristics of multi-source heterogeneity. Various types of data such as the time-series signals generated by equipment sensors, the text documents of production plans, the images of logistics transportation trajectories, the operation monitoring video streams, and the resource scheduling operation logs have huge differences in format and type. The previous collection methods often can only collect single-type data and cannot comprehensively obtain all types of information in the production process, resulting in a one-sided understanding of the production status and being unable to provide complete data support for management decisions.

[0004] From the perspective of feature extraction, due to the lack of effective technical means, it is difficult to accurately extract valuable features from these complex multi-source heterogeneous data. For example, for the time-series signals of equipment sensors, it is impossible to deeply mine their multi-scale operation mode features and timely discover potential operation fault hidden dangers of the equipment; when parsing the text documents of production plans, it is impossible to accurately grasp the dependency relationships and priorities between tasks, which easily leads to unreasonable production task arrangements and affects the production progress.

[0005] In terms of data fusion, due to the differences in format and semantics of data from different data sources, it is difficult to fuse them into a unified framework for analysis. This makes it impossible for enterprises to accurately judge the status of production resources from a global perspective, lack a precise basis for resource allocation, and reduces the utilization efficiency of resources. For example, in the logistics transportation link, due to the inability to effectively fuse the transportation trajectory image data with other production data, it is difficult to optimize the allocation of logistics resources, resulting in an increase in transportation costs.

[0006] The construction of resource topology relationships has always been a difficult point in production resource management. Traditional methods are difficult to comprehensively and dynamically construct production resource association graphs and supply chain dependency graphs, and cannot clearly present the topological relationships between equipment resources, the matching rules between tasks and resources, and the constraint conditions of historical scheduling paths. This makes enterprises lack references to historical experience and global relationships when formulating resource scheduling strategies and task allocation plans, and the decision-making process has a large degree of blindness.

[0007] In addition, the existing resource scheduling and task allocation methods usually rely on manual experience or simple algorithm models and cannot adapt to complex and changeable production environments. These methods often cannot effectively shorten the task completion time while ensuring resource utilization. In the face of emergencies during the production process, such as equipment failures and order changes, they cannot adjust the scheduling strategy in a timely manner, resulting in production interruptions and increased costs. Summary of the Invention

[0008] The purpose of the present invention is to provide a production resource management system based on an AI neural network to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: A production resource management system based on an AI neural network, the system includes: Multi-source data acquisition module: used to collect multi-source heterogeneous data of the production environment, the multi-source heterogeneous data includes device sensor time-series signals, production plan text documents, logistics transportation trajectory images, operation monitoring video streams, and resource scheduling operation logs; Multi-dimensional feature extraction module: perform multi-dimensional feature extraction on the multi-source heterogeneous data to generate feature vectors corresponding to each data source, the feature vectors include: device operation status features, text task priority features, image spatial distribution features, video behavior recognition features, and scheduling behavior sequence features; Heterogeneous data fusion module: map the feature vectors to a unified representation space based on a cross-modal attention mechanism to generate a fused resource state semantic representation; Resource topology map construction module: construct a two-layer topology library of a production resource association map and a supply chain dependence map, and the two-layer topology library stores device resource topology relationships, task-resource matching rules, and historical scheduling path constraint conditions; Intelligent optimization decision module: based on the resource state semantic representation and the two-layer topology library, generate a real-time resource scheduling strategy and a task allocation optimization plan through a dynamic programming algorithm.

[0010] Preferably, the multi-dimensional feature extraction of the multi-source heterogeneous data includes: Adopt a deep convolutional network to extract multi-scale operation mode features from the device sensor time-series signals; Adopt a hierarchical semantic encoder of a pre-trained language model to parse task dependence relationships for the production plan text documents; Adopt a region segmentation algorithm combined with a spatial pyramid network to extract cargo distribution heat map features from the logistics transportation trajectory images.

[0011] Preferably, the multi-dimensional feature extraction further includes: Extract spatio-temporal features of operation behaviors from the operation monitoring video stream using a 3D residual network, and calculate the continuity index of personnel action trajectories through a motion estimation model; Model the scheduling decision pattern for the resource scheduling operation log using a sequence-to-sequence model to generate the hidden state transition matrix of the operation sequence.

[0012] Preferably, the construction of the double-layer topology library of the production resource association graph and the supply chain dependence graph includes: Construct a dynamically expanding production resource association graph using a graph convolutional network based on the mapping relationship between device attribute entities and task resources; Generate the node embedding representation of the supply chain dependence graph using a temporal graph attention mechanism according to historical scheduling records and resource utilization rate indicators.

[0013] Preferably, the intelligent optimization decision module adopts a co-evolution algorithm framework, including: Define the optimization objective as a multi-objective function of resource utilization rate and task completion time, and the decision variable as the candidate set of scheduling paths; Generate a scheduling plan through a population iteration strategy, and design a fitness evaluation function based on the resource load balancing constraint.

[0014] Preferably, the intelligent optimization decision module further includes: using a simulated annealing algorithm to perform local optimization on the scheduling path, calculating the robustness index of the task assignment plan by combining fuzzy logic reasoning, and pruning invalid decision branches through the supply chain topology connectivity rule.

[0015] Preferably, the system further includes: performing pattern anomaly detection on the resource scheduling operation log, and using a density clustering algorithm to identify operation sequences that deviate from the conventional scheduling strategy; when an abnormal scheduling is detected, trigger the path re-planning mechanism of the intelligent optimization decision module to regenerate a scheduling adjustment plan adapted to the current resource state.

[0016] Preferably, the path re-planning mechanism models the historical scheduling sequence using a bidirectional gated recurrent unit network, including: Extract the task execution intention features from the forward sequence and the resource conflict features from the backward sequence; Generate a corrected scheduling path representation through feature fusion and update the edge weight parameters of the supply chain dependence graph.

[0017] Preferably, the system further includes: constructing a resource aging prediction model, which is based on the device historical operation duration and failure rate data, and dynamically adjusts the trigger threshold and priority of the maintenance strategy using an exponential smoothing function.

[0018] Preferably, the resource aging prediction model further includes: Align the equipment life cycle curve with the characteristics of real-time vibration signals based on the wavelet transform algorithm, generate an adaptive maintenance plan and embed it into the real-time resource scheduling strategy.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: In terms of data processing, the multi-source data acquisition module can comprehensively collect multi-source heterogeneous data in the production environment, covering equipment sensor time series signals, production plan text documents, logistics transportation trajectory images, operation monitoring video streams, and resource scheduling operation logs, etc. This provides a rich and complete data basis for subsequent analysis, enabling managers to comprehensively understand all aspects of the production process. The multi-dimensional feature extraction module adopts advanced technologies such as deep convolutional networks and pre-trained language models according to the characteristics of different types of data to accurately extract various feature vectors, such as equipment operation status features, text task priority features, etc. Compared with traditional methods, it can dig deeper into the data value, discover potential equipment failures in a timely manner, accurately grasp task priorities, and provide strong support for production decisions.

[0020] The heterogeneous data fusion module maps different feature vectors to a unified representation space based on the cross-modal attention mechanism to generate a fused resource status semantic representation. This fusion method breaks the gap between data, enabling enterprises to accurately judge the resource status from a global perspective, providing accurate basis for resource allocation, and effectively improving resource utilization efficiency. For example, in the material distribution link, it can achieve accurate distribution according to the real-time demand of production equipment and material inventory, reducing material backlog and waste.

[0021] The production resource association graph and the double-layer topology library of the supply chain dependence graph constructed by the resource topology graph construction module store equipment resource topology relationships, task-resource matching rules, and historical scheduling path constraint conditions. Through the graph convolutional network and the temporal graph attention mechanism, the graph can be dynamically expanded and updated to comprehensively and dynamically reflect the internal connections of production resources and the supply chain. This provides rich historical experience and global relationship references for enterprises to formulate resource scheduling strategies and task allocation plans, avoiding decision-making blindness and optimizing the production process.

[0022] Based on the resource status semantic representation and the double-layer topology library, the intelligent optimization decision module generates a real-time resource scheduling strategy and a task allocation optimization plan through the dynamic programming algorithm. Adopting the co-evolution algorithm framework, taking resource utilization rate and task completion time as multi-objective functions for optimization, and designing a fitness evaluation function considering resource load balance constraints. This enables the system to quickly find the optimal resource scheduling and task allocation methods in a complex production environment, significantly improving resource utilization rate and shortening task completion time. For example, in the order production process, it can reasonably arrange production tasks and allocate resources according to the order urgency, equipment production capacity, and material supply situation to ensure the on-time delivery of orders while reducing production costs.

[0023] In addition, the system also has a pattern anomaly detection and path replanning mechanism. By detecting pattern anomalies in the resource scheduling operation logs, it can promptly identify operation sequences that deviate from the conventional scheduling strategy and trigger the path replanning mechanism. The bidirectional gated recurrent unit network is used to model the historical scheduling sequences, and a scheduling adjustment plan adapted to the current resource status is regenerated to effectively handle emergencies during the production process and ensure the continuity and stability of production. The resource aging prediction model is based on the historical operation duration and failure rate data of the equipment. It dynamically adjusts the triggering thresholds and priorities of the maintenance strategy using the exponential smoothing function and generates an adaptive maintenance plan in combination with the wavelet transform algorithm to be embedded in the real-time resource scheduling strategy, achieving preventive maintenance of the equipment, reducing the impact of equipment failures on production, and extending the service life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is the working principle diagram of the production resource management system based on the AI neural network according to the present invention; Figure 2 is the flowchart of the construction of the resource topology map; Figure 3 is the flowchart of other strategies of the intelligent optimization decision module; Figure 4 is the flowchart of anomaly detection and path replanning. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0026] Please refer to Figures 1-4 , the present invention relates to a production resource management system based on an AI neural network, and its specific implementation will be elaborated in detail below.

[0027] Multi-source data acquisition module: This module is responsible for collecting multi-source heterogeneous data in the production environment, which is the basis for the operation of the entire system. Among them, the time-series signals of equipment sensors are collected in real time by various sensors installed on production equipment, covering data on the variation of operating parameters such as temperature, pressure, and rotational speed of the equipment over time; the production plan text document contains detailed information about production tasks, such as product types, production quantities, delivery times, etc.; the logistics transportation trajectory image records the location information of goods during transportation and is presented in the form of images; the operation monitoring video stream is used to monitor the work conditions of operators in real time; and the resource scheduling operation log records past resource scheduling decisions and operation processes.

[0028] Multi-dimensional Feature Extraction Module: For multi-source heterogeneous data, this module performs multi-dimensional feature extraction to generate feature vectors corresponding to each data source. These feature vectors can more accurately reflect the internal features of the data and provide support for subsequent analysis.

[0029] Heterogeneous Data Fusion Module: Based on the cross-modal attention mechanism, this module maps the feature vectors generated by the multi-dimensional feature extraction module to a unified representation space. In this way, different types of data can be fused within the same space to generate a fused semantic representation of the resource state, enabling the system to comprehensively consider various factors and gain a more comprehensive understanding of the state of production resources.

[0030] Resource Topology Map Construction Module: Constructs a two-layer topology library of the production resource association map and the supply chain dependency map. In the production resource association map, the topological relationships of device resources are stored, which describe the physical connections and logical relationships between production devices, as well as the task-resource matching rules that determine the types and quantities of resources required for different production tasks. The supply chain dependency map contains historical scheduling path constraint conditions. By analyzing past scheduling records, the scheduling limitations and optimization directions in different situations are summarized.

[0031] Intelligent Optimization Decision Module: Based on the semantic representation of the resource state and the two-layer topology library, this module generates a real-time resource scheduling strategy and a task allocation optimization plan through the dynamic programming algorithm. The dynamic programming algorithm can find the optimal resource scheduling and task allocation methods under various constraint conditions to improve production efficiency and reduce costs.

[0032] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0033] Embodiment 1:

[0034] In an actual production scenario, taking a large-scale electronic device manufacturing factory as an example, the feature extraction process of the multi-dimensional feature extraction module for the time-series signals of device sensors, the production plan text documents, and the logistics transportation trajectory images will be described in detail.

[0035] For the timing signals of device sensors, various production devices in the factory, such as high-precision chip mounters and automatic detection devices, are equipped with a large number of sensors to monitor the operating conditions of the devices in real time. The timing signals collected by these sensors contain rich device operation information. When using a deep convolutional network for multi-scale operation mode feature extraction, the continuously collected timing signals are first divided into multiple data segments at a fixed time interval (such as 10 seconds). Each data segment serves as the input to the deep convolutional network, and different-sized convolutional kernels are set in the convolutional layers of the network. A smaller convolutional kernel, such as a 3×1 convolutional kernel, can capture the subtle changes in device operation parameters in a short period of time when sliding and convolving on the data segment, like the pressure fluctuation during the chip mounting instant of the chip mounter. A larger convolutional kernel, such as a 9×1 convolutional kernel, can obtain the trend characteristics of device operation parameters over a longer period of time, such as the temperature change trend of the detection device within a production batch. Through the combined operations of multiple convolutional layers and pooling layers, the device operation state feature vector is gradually extracted from the original timing signals. This feature vector comprehensively reflects the operation modes of the device at different time scales, providing a key basis for subsequent judgment of whether the device is operating normally and performance optimization.

[0036] For the production plan text document, the production plan of this factory is recorded in the form of a detailed text document, containing information on numerous production tasks, such as product models, production quantities, the sequence of each production link, and delivery times, etc. When using the hierarchical semantic encoder of the pre-trained language model to analyze task dependencies, the entire production plan text document is input into the pre-trained language model. The hierarchical semantic encoder first analyzes the text word by word. Through the self-attention mechanism, the model can pay attention to the semantic associations between different words in the text. For example, when analyzing production tasks, it can identify the sequential dependency relationship between "producing 1000 pieces of product model A" and "needing to complete the assembly of component B first". By analyzing the entire text, the complex dependency relationships between various production tasks are parsed, and thus the text task priority feature vector is extracted. This feature vector clarifies the importance and urgency of each production task in the entire production process, which is crucial for the reasonable arrangement of production resources and the scheduling of production tasks.

[0037] For the logistics transportation trajectory image, during the transportation process of raw materials and finished products in the factory, the logistics transportation trajectory image is obtained through cameras installed at key positions of transportation vehicles and warehouses. When using the region segmentation algorithm combined with the spatial pyramid network to extract the heat map features of the cargo distribution, the region segmentation algorithm first processes the image, separates the cargo area from the complex background, and accurately determines the position and scope of the cargo. Then, the spatial pyramid network performs multi-scale feature extraction on the segmented cargo area. It divides the cargo area into multiple sub-regions through pooling operations at different scales, and each sub-region represents a different spatial resolution. Calculate the distribution density of the cargo on each sub-region. For example, count the number of cargos or the proportion of the occupied area in each sub-region. According to these calculation results, a heat map of the cargo distribution is generated. The darker the color in the heat map, the denser the cargo distribution. Extract the image spatial distribution feature vector from the heat map, which provides intuitive data support for optimizing the logistics transportation route and reasonably arranging the storage space of the warehouse.

[0038] Example 2:

[0039] Taking a certain automobile manufacturing factory as an example, the feature extraction process of the operation monitoring video stream and the resource scheduling operation log in the multi-dimensional feature extraction module is elaborated in depth.

[0040] In the production workshop of an automobile manufacturing factory, the operation monitoring video stream is used to monitor the operation process of workers and the running status of equipment in real time. A three-dimensional residual network is used to extract the spatio-temporal features of operation behaviors, and the continuity index of the personnel movement trajectory is calculated through a motion estimation model. The three-dimensional residual network divides the monitoring video stream into multiple video segments in chronological order, and each segment contains a continuous number of frames of images. For example, 10 consecutive frames of images are selected as a segment. The convolutional layer of the network performs convolutional operations on the video segment in both the spatial and temporal dimensions. In the spatial dimension, it can capture the action postures of workers, such as reaching, grasping, installing, etc. when workers assemble automobile parts; in the temporal dimension, it can track the sequence and duration of these actions. The motion estimation model, based on the features extracted by the three-dimensional residual network, obtains the movement trajectory of workers by calculating the changes in the position and posture of workers between adjacent frames. For example, by comparing the positions of the worker's hands in two adjacent frames, the movement path is determined. Further calculate the continuity index of the movement trajectory, which measures the smoothness of the movement trajectory. If the worker's actions suddenly pause or deviate significantly from the normal trajectory, the continuity index will decrease, which may indicate that the worker's operation has made a mistake or encountered a problem. Integrating the spatio-temporal features of operation behaviors and the continuity index of the personnel movement trajectory into a video behavior recognition feature vector helps to detect abnormal operations in the production process in a timely manner and ensure production safety and product quality.

[0041] For the resource scheduling operation logs, the automobile manufacturing plant has detailed records, including information such as the time of each resource scheduling, the type of resources scheduled (such as parts, production equipment, etc.), the quantity scheduled, and the corresponding production tasks. The sequence-to-sequence model is used to model the scheduling decision pattern to generate the hidden state transition matrix of the operation sequence. The sequence-to-sequence model consists of an encoder and a decoder. The operation sequence in the resource scheduling operation log is input into the encoder in chronological order. The encoder encodes the operation sequence and converts a series of scheduling operations into a fixed-length hidden state vector. This hidden state vector contains the overall information of the operation sequence, such as the order of resource scheduling, the demand pattern of different tasks for resources, etc. Based on the hidden state vector output by the encoder, the decoder gradually generates a new operation sequence. During the generation process, the model learns the dependency relationships and transition rules between operation sequences. For example, if a specific part is frequently scheduled for a certain production line within a certain time period, the model will learn this scheduling pattern and reflect it in the hidden state transition matrix. The elements in the hidden state transition matrix represent the probability of transitioning from one operation state to another. By analyzing this matrix, the patterns and rules of past scheduling decisions can be summarized, providing important references for current resource scheduling and serving as a key component of the scheduling behavior sequence feature vector.

[0042] Embodiment 3:

[0043] Taking a large clothing manufacturing enterprise as an example, the construction processes of the production resource association graph and the supply chain dependency graph in the resource topology graph construction module are described in detail.

[0044] When constructing the production resource association graph, the production resources of this clothing manufacturing enterprise include various production equipment, such as sewing machines, cutting machines, etc., as well as different production tasks, such as tasks of producing different styles of clothing. Based on the mapping relationship between the equipment attribute entities and the task resources, a graph convolutional network is adopted. First, the nodes in the production resource association graph are determined. The equipment nodes are defined according to attributes such as the type and number of the equipment. Each sewing machine has a unique number, and the node information includes the production speed of the equipment, the types of fabrics that can be processed, etc. The task nodes are defined according to the tasks in the production plan. For example, the task node of "producing 500 summer dresses" includes information such as the types of fabrics, colors, and size distributions required for the task. The edges in the graph are determined according to the mapping relationship between the equipment attribute entities and the task resources. For example, if a specific model of sewing machine is required to sew the summer dresses, an edge is established between this task node and the corresponding sewing machine equipment node, and a corresponding weight is assigned to this edge. The size of the weight is determined according to the degree of dependence of the task on the equipment. For example, if this style of dress has high requirements for the special functions of the sewing machine, the weight is set to be relatively large. The constructed graph is input into the graph convolutional network. The graph convolutional network continuously updates the feature representation of the nodes by learning the node features and edge information. With the introduction of new equipment or the change of production tasks, the graph can be automatically adjusted and updated. For example, when the enterprise purchases a new type of high-speed sewing machine, the graph will include it in a timely manner and recalculate the information of relevant nodes and edges to reflect the latest production resource association situation.

[0045] For the supply chain dependency graph, the supply chain of this clothing enterprise involves multiple links such as fabric suppliers, accessory suppliers, production workshops, warehouses, and logistics distribution companies. According to historical scheduling records and resource utilization rate indicators, a temporal graph attention mechanism is used to generate node embedding representations. First, historical scheduling records are collected, including information such as the time of each scheduling, the resources scheduled (such as the types and quantities of fabrics and accessories), the completion status of production tasks, and the time and cost of logistics distribution. At the same time, resource utilization rate indicators are calculated, such as the utilization rate of sewing machines and the inventory turnover rate of fabrics. These information are integrated into the supply chain dependency graph, where the nodes in the graph represent each link in the supply chain. The edges represent relationships such as logistics, information flow, and capital flow between each link. For example, the edge between the fabric supplier and the production workshop represents the supply relationship of fabrics, including information such as supply time and supply quantity. The temporal graph attention mechanism can automatically adjust the degree of attention to each node and edge according to information at different time points. When calculating the node embedding representation, the model will focus on historical information and resource utilization rate related to the current scheduling decision. For example, when formulating the production plan for the next season's clothing, the model will pay more attention to the supply stability of the fabric supplier and the inventory turnover rate of fabrics during the production of the same type of clothing in the previous season, so as to generate more accurate and representative node embedding representations. These node embedding representations can clearly reflect the dependency relationships and importance degrees between each link in the supply chain, providing strong support for resource scheduling and supply chain optimization.

[0046] Example 4:

[0047] This example details the specific implementation of the co-evolution algorithm framework in the intelligent optimization decision module, including the optimization objective, decision variables, population iteration strategy, and fitness evaluation function.

[0048] The intelligent optimization decision module adopts a co-evolution algorithm framework. Define the optimization objective as a multi-objective function of resource utilization rate and task completion time. Let the resource utilization rate be R and the task completion time be T. The optimization objective function F can be expressed as: , where, and are weight coefficients, and . and The values of and are adjusted according to actual production requirements. For example, if the current production pays more attention to resource utilization rate, the value of can be appropriately increased; if more attention is paid to task completion time, the value of is increased. R is obtained by calculating the ratio of the actually used resource amount to the total resource amount, and T is calculated based on the start time and end time of the task.

[0049] The decision variable is the candidate set of scheduling paths. During the production process, there are multiple possible resource scheduling paths, which constitute the candidate set of scheduling paths. Each candidate set of scheduling paths contains a series of resource allocation decisions. For example, which device is used to execute which task, the transportation route of materials, etc.

[0050] The scheduling plan is generated through the population iteration strategy. First, an initial population is randomly generated, and each individual represents a scheduling plan. In each generation of iteration, each individual in the population is evaluated and selected. The selection operation is based on the fitness evaluation function, and the fitness evaluation function is designed based on the resource load balancing constraint. Let the resource load balance degree be B, and its calculation formula is: , where n is the number of types of resources, is the actual load of the i-th type of resource, is the average load of all resources. The smaller the value of B, the more balanced the resource load. The fitness evaluation function G can be expressed as: , where, is the adjustment coefficient, which is used to balance the importance of the optimization objective and the resource load balance degree. The value of

[0051] Example 5:

[0052] This example further elaborates on the simulated annealing algorithm, fuzzy logic reasoning, and supply chain topology connectivity rules adopted in the intelligent optimization decision-making module.

[0053] The intelligent optimization decision-making module uses the simulated annealing algorithm to perform local optimization on the scheduling path. The simulated annealing algorithm is a heuristic search algorithm that simulates the physical annealing process and adjusts the search strategy by controlling the temperature parameter. In the initial stage, the temperature is high, and the algorithm has a large search range and can accept worse solutions to avoid falling into local optimal solutions. As the iteration progresses, the temperature gradually decreases, and the algorithm gradually focuses on the local optimal region.

[0054] Specifically, based on the current scheduling path, a neighborhood solution is randomly generated. Calculate the difference between the neighborhood solution and the current solution of the objective function. If , that is, if the neighborhood solution is better, then accept the neighborhood solution as the new current solution; if , then accept the neighborhood solution with a certain probability, which is determined by the Metropolis criterion, and the formula is: , where P is the probability value obtained through calculation, and T is the current temperature. As the temperature T decreases, the probability of accepting a worse solution gradually decreases. Through continuous iteration, the simulated annealing algorithm can jump out of the local optimal solution to a certain extent and find a better scheduling path.

[0055] Combine fuzzy logic reasoning to calculate the robustness index of the task assignment scheme. Fuzzy logic reasoning can handle uncertain and fuzzy information. First, determine the factors affecting the robustness of the task assignment scheme, such as the urgency of the task, the reliability of the resources, etc. Fuzzify these factors and convert them into fuzzy linguistic variables, such as "high", "medium", "low". Establish a fuzzy rule base based on practical experience and expert knowledge. For example, if the task urgency is "high" and the resource reliability is "low", then the robustness of the task assignment scheme is "low". Through the fuzzy inference engine, according to the input fuzzy linguistic variables and the fuzzy rule base, obtain the robustness index of the task assignment scheme. This robustness index can help evaluate the stability and reliability of the task assignment scheme in the face of various uncertain factors.

[0056] Prune invalid decision branches through the supply chain topology connectivity rule. The supply chain topology connectivity rule is used to judge whether the scheduling decision conforms to the actual structure and logic of the supply chain. If a certain scheduling decision causes a certain link in the supply chain to be unable to be normally connected or there is a resource disconnection, etc., then this decision branch is considered invalid. For example, if the materials required for a certain task cannot be transported from the supplier to the production workshop in time, resulting in production interruption, then this scheduling decision is invalid. By checking the supply chain topology connectivity, promptly prune these invalid decision branches, reduce the search space, and improve the efficiency of the optimization decision.

[0057] Example 6:

[0058] This example introduces the pattern anomaly detection, path re-planning mechanism and resource aging prediction model for the resource scheduling operation log in the system.

[0059] The system performs pattern anomaly detection on the resource scheduling operation logs and uses the density clustering algorithm to identify the operation sequences that deviate from the conventional scheduling strategies. The density clustering algorithm discovers clusters based on the density distribution of data points. First, according to the operation sequences in the resource scheduling operation logs, the similarity between each operation sequence and other operation sequences is calculated. The similarity can be measured by calculating the distance between the feature vectors of the operation sequences, such as the Euclidean distance or cosine similarity. According to the set density threshold and minimum sample number, the operation sequences are divided into different clusters. If an operation sequence does not belong to any cluster or is far from the core point of the cluster it belongs to, then this operation sequence is considered abnormal. When an abnormal scheduling is detected, the path re-planning mechanism of the intelligent optimization decision module is triggered.

[0060] The path re-planning mechanism models the historical scheduling sequences using a bidirectional gated recurrent unit network. The bidirectional gated recurrent unit network can learn the forward and backward information of the historical scheduling sequences simultaneously. The task execution intention features are extracted from the forward sequence, and the resource conflict features are extracted from the backward sequence. For example, the forward sequence can reflect the execution order and priority of tasks, while the backward sequence can reveal the resource conflict situations that occur during the scheduling process. Through feature fusion, the task execution intention features and resource conflict features are integrated to generate a corrected representation of the scheduling path. At the same time, according to the corrected representation of the scheduling path, the edge weight parameters of the supply chain dependency graph are updated to reflect the latest scheduling situation and resource dependency relationships.

[0061] The system constructs a resource aging prediction model. This model is based on the device's historical running duration and failure rate data and uses an exponential smoothing function to dynamically adjust the triggering threshold and priority of the maintenance strategy. Let the device's historical running duration be t, and the exponential smoothing function is: , where, is the smoothed value at time t, is the smoothed value at time t - 1, is the actual observed value at time t (such as the failure rate), is the smoothing coefficient, and its value range is (0, 1). By processing the failure rate data with the exponential smoothing function, a more stable predicted value of the failure rate is obtained. According to the predicted failure rate, the triggering threshold and priority of the maintenance strategy are dynamically adjusted. If the predicted failure rate is high, then the triggering threshold of the maintenance strategy is lowered, and the maintenance priority is increased to maintain the device in a timely manner and avoid the impact of device failures on production.

[0062] The resource aging prediction model also aligns the equipment life cycle curve and the real-time vibration signal characteristics based on the wavelet transform algorithm, generates an adaptive maintenance plan and embeds it into the real-time resource scheduling strategy. The wavelet transform algorithm can perform multi-resolution analysis on signals, decomposing the equipment life cycle curve and the real-time vibration signal into components of different frequencies. By comparing the characteristics of different frequency components, the similarities and differences between the equipment life cycle curve and the real-time vibration signal are found. Based on these analysis results, an adaptive maintenance plan is generated, such as determining the maintenance time and content of the equipment. The adaptive maintenance plan is embedded into the real-time resource scheduling strategy to ensure that when resource scheduling is carried out, the maintenance requirements of the equipment are fully considered, production tasks are reasonably arranged, and the overall reliability and stability of the production system are improved.

[0063] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0064] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A production resource management system based on AI neural network, characterized in that: include: Multi-source data acquisition module: used to collect multi-source heterogeneous data in the production environment, including equipment sensor timing signals, production plan text documents, logistics transportation trajectory images, operation monitoring video streams and resource scheduling operation logs; Multidimensional feature extraction module: extract multidimensional features from the multi-source heterogeneous data to generate feature vectors corresponding to each data source, wherein the feature vectors include: equipment operation status features, text task priority features, image space distribution features, video behavior recognition features, and scheduling behavior sequence features; Heterogeneous data fusion module: maps the feature vectors to a unified representation space based on a cross-modal attention mechanism to generate a fused semantic representation of resource status; Resource topology map construction module: constructs a double-layer topology library of production resource association map and supply chain dependency map, which stores equipment resource topology relationship, task-resource matching rules and historical scheduling path constraints; Intelligent optimization decision module: Based on the resource status semantic representation and the two-layer topology library, a real-time resource scheduling strategy and task allocation optimization solution are generated through a dynamic programming algorithm.

2. The production resource management system according to claim 1, characterized in that: The extracting multi-dimensional features from the multi-source heterogeneous data includes: Extracting multi-scale operation mode features from the device sensor timing signals using a deep convolutional network; Parsing task dependencies for the production plan text document using a hierarchical semantic encoder of a pre-trained language model; The regional segmentation algorithm combined with the spatial pyramid network is used to extract the cargo distribution heat map features from the logistics transportation trajectory image.

3. The production resource management system according to claim 2, characterized in that: The multidimensional feature extraction also includes: A three-dimensional residual network is used to extract the spatiotemporal characteristics of the operation behavior from the operation monitoring video stream, and a continuity index of the personnel's action trajectory is calculated through a motion estimation model; A sequence-to-sequence model is used to model the scheduling decision mode of the resource scheduling operation log to generate a hidden state transfer matrix of the operation sequence.

4. The production resource management system according to claim 1, characterized in that: The two-layer topology library for constructing the production resource association graph and the supply chain dependency graph includes: Based on the mapping relationship between equipment attribute entities and task resources, a graph convolutional network is used to build a dynamically expanded production resource association graph; According to historical scheduling records and resource utilization indicators, the time-series graph attention mechanism is used to generate node embedding representation of the supply chain dependency graph.

5. The production resource management system according to claim 1, characterized in that: The intelligent optimization decision module adopts a collaborative evolutionary algorithm framework, including: The optimization objective is defined as a multi-objective function of resource utilization and task completion time, and the decision variable is the candidate set of scheduling paths; The scheduling scheme is generated through the population iteration strategy, and the fitness evaluation function is designed based on the resource load balancing constraint.

6. The production resource management system according to claim 5, characterized in that: The intelligent optimization decision module also includes: using a simulated annealing algorithm to perform local optimization on the scheduling path, combining fuzzy logic reasoning to calculate the robustness index of the task allocation plan, and pruning invalid decision branches through supply chain topology connectivity rules.

7. The production resource management system according to claim 1, characterized in that: The system also includes: performing pattern anomaly detection on the resource scheduling operation log, and using a density clustering algorithm to identify operation sequences that deviate from conventional scheduling strategies; when abnormal scheduling is detected, triggering a path replanning mechanism of an intelligent optimization decision module, and regenerating a scheduling adjustment plan that adapts to the current resource status.

8. The production resource management system according to claim 7, characterized in that: The path replanning mechanism uses a bidirectional gated recurrent unit network to model the historical scheduling sequence, including: Extract task execution intention features from the forward sequence and resource conflict features from the backward sequence; The revised scheduling path representation is generated through feature fusion, and the edge weight parameters of the supply chain dependency graph are updated.

9. The production resource management system according to claim 1, characterized in that: The system also includes: constructing a resource aging prediction model, wherein the resource aging prediction model is based on historical equipment operation time and failure rate data, and uses an exponential smoothing function to dynamically adjust the triggering threshold and priority of the maintenance strategy.

10. The production resource management system according to claim 9, characterized in that: The resource aging prediction model also includes: Based on the wavelet transform algorithm, the equipment life cycle curve is aligned with the real-time vibration signal characteristics, and an adaptive maintenance plan is generated and embedded in the real-time resource scheduling strategy.

Images