An artificial intelligence-based manufacturing whole-process collaborative management platform

Abstract

The application discloses a kind of based on artificial intelligence's manufacturing industry whole-process collaborative management platform, including data center, intelligent production scheduling module, intelligent visual inspection module and collaborative decision engine.Data center uniformly collects production whole-process data.Intelligent production scheduling module is based on deep reinforcement learning to dynamically schedule production orders, uses Transformer encoder to encode state, realizes local rescheduling under abnormal event through multi-objective reward function and bayesian causal inference.Intelligent visual inspection module fuses multi-station images through cross-camera attention mechanism, realizes real-time defect detection, and has illumination self-adaptation, domain generalization and few-sample rapid migration capability.Collaborative decision engine adjusts detection strategy according to scheduling result, and triggers rework rescheduling and equipment maintenance according to defect detection result, forming scheduling-inspection two-way closed loop.The application realizes whole-process collaborative management and control and intelligent optimization, effectively improves production efficiency and product quality.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent manufacturing, specifically to an AI-based collaborative management platform for the entire manufacturing process, applicable to intelligent production management in multiple industries such as machinery manufacturing, electronic equipment manufacturing, automobile manufacturing, food processing, and pharmaceutical manufacturing. Background Technology

[0002] In today's rapidly developing manufacturing industry, enterprises urgently need to improve efficiency, quality, and cost control, but traditional production management suffers from four major pain points: 1. Data silos exist across multiple business systems, with production line, equipment, and progress data scattered, making real-time global monitoring impossible, and manual statistical analysis inefficient and prone to errors; 2. Production scheduling relies on manual experience and fixed rules, which cannot quickly adapt to dynamic scenarios such as equipment failure and urgent orders, making it difficult to guarantee order delivery; 3. Industrial vision inspection relies on manually setting fixed threshold parameters, which has poor adaptability to lighting and product models, long parameter adjustment cycle for new products, and insufficient inspection accuracy and efficiency. 4. The production scheduling and quality inspection systems are disconnected from each other. Quality inspection results cannot be used to optimize production scheduling, and production scheduling plans cannot allocate testing resources, resulting in a lack of data closed-loop collaboration.

[0003] While existing technological improvements have alleviated some problems through approaches such as cloud-edge collaboration, ontology knowledge graphs, and AI-assisted production scheduling, they still have some shortcomings: they cannot connect data across the entire process of sales, planning, production, quality inspection, materials, equipment, and personnel; they do not integrate deep reinforcement learning production scheduling with deep learning visual inspection to form collaborative decision-making; the production scheduling algorithm has weak dynamic adaptability; and there is a lack of closed-loop data feedback between production scheduling and quality inspection.

[0004] In summary, the industry urgently needs an intelligent manufacturing full-process collaborative management platform that can connect data across the entire manufacturing process, support dynamic adaptive intelligent scheduling, integrate AI visual inspection, and build a closed-loop collaborative mechanism for scheduling and quality inspection. Summary of the Invention

[0005] (a) Purpose of the invention: This invention addresses the technical problems in existing manufacturing production management systems, such as severe data silos, reliance on manual experience in production scheduling lacking dynamic optimization, insufficient accuracy and adaptability of visual inspection, and disconnect between production scheduling and quality inspection lacking closed-loop collaboration. It provides an artificial intelligence-based collaborative management platform for the entire manufacturing process.

[0006] This invention aims to achieve the following objectives: 1. By building a unified data platform, we can achieve unified collection, storage and management of heterogeneous data across the entire manufacturing process (sales management, production planning, production execution, quality inspection, material management, equipment management, personnel management, inventory management and procurement management), break down data silos and form a unified business information view at the enterprise level.

[0007] 2. Through the intelligent scheduling module based on deep reinforcement learning, dynamic adaptive scheduling of production orders can be achieved, which can quickly respond to dynamic events such as equipment failure, emergency order insertion, and rework, and achieve multi-objective optimization.

[0008] 3. Through a deep learning-based intelligent visual inspection module, real-time and high-precision product defect detection is achieved, with capabilities such as lighting adaptation, cross-production line generalization, and rapid adaptation with few samples.

[0009] 4. By establishing a two-way data closed loop between production scheduling and quality inspection through a collaborative decision engine, intelligent collaboration of "production scheduling guiding inspection and inspection feedback on production scheduling" is achieved, thereby improving the overall efficiency and product quality of the entire manufacturing process.

[0010] (II) Technical Solution: An AI-based collaborative management platform for the entire manufacturing process includes: The data platform is used to uniformly collect, store, and manage heterogeneous data across the entire manufacturing process, which includes at least the following stages: sales management, production planning, production execution, quality inspection, material management, equipment management, personnel management, inventory management, and procurement management. The intelligent scheduling module is connected to the data platform and is used to dynamically schedule production orders based on deep reinforcement learning algorithms. The intelligent visual inspection module is connected to the data platform and is used to perform real-time defect detection on products on the production line based on deep learning algorithms. A collaborative decision engine is connected to the intelligent scheduling module and the intelligent visual inspection module, respectively. It is used to trigger the intelligent scheduling module to reschedule based on the defect detection results output by the intelligent visual inspection module, and to dynamically adjust the detection strategy of the intelligent visual inspection module based on the scheduling results output by the intelligent scheduling module.

[0011] Preferably, the intelligent scheduling module includes: A status encoding unit is used to encode the order status data, resource status data, and global event data provided by the data platform into a status vector. A strategy network is used to output production scheduling decision actions based on the state vector. The production scheduling decision actions include process-team allocation decision, process execution order decision, order insertion and rescheduling decision, and rework path selection decision. The reward function unit is used to calculate the reward value for production scheduling decisions based on a multi-objective optimization function, which includes maximizing the on-time delivery rate of orders, minimizing the total completion time, minimizing the shift load imbalance, and maximizing equipment utilization.

[0012] Preferably, the state encoding unit uses a multi-head attention mechanism based on the classic deep neural network Transformer architecture with self-attention mechanism to encode the order state data, resource state data and global event data, generating a unified embedding vector as the state vector; the policy network uses a deep reinforcement learning network and is trained based on the proximal policy optimization algorithm.

[0013] Preferably, the intelligent scheduling module further includes a causal reasoning unit, which is used to construct a task-resource-state causal graph based on a Bayesian network. When an abnormal event is detected, the unit identifies the propagation path of the abnormal event in the causal graph and performs local rescheduling only on the affected subgraph.

[0014] Preferably, the intelligent scheduling module further includes a human-machine collaborative interaction unit, which generates multiple candidate scheduling schemes and displays them in the form of a multi-objective Pareto front, receives user selection instructions, and uses the user selection as a feedback signal to perform weight optimization on the reward function of the policy network.

[0015] Preferably, the intelligent visual detection module includes: The image acquisition unit is used to acquire multi-source image data of the product from different workstations and different perspectives using multiple industrial cameras; A multimodal feature fusion unit is used to adaptively fuse the multi-source image data through a cross-camera attention mechanism to generate a fused feature map; The defect detection unit is used to identify and classify defects in the product based on the fused feature map, and output the defect type, defect location and defect area.

[0016] Preferably, the multimodal feature fusion unit uses a cross-camera attention mechanism to adaptively weight and fuse the multi-scale features extracted by each industrial camera. The cross-camera attention mechanism calculates the attention weights based on the spatial position relationship of each camera's corresponding workstation and the geometric constraint relationship of the detection target. The intelligent vision detection module also includes an illumination adaptive unit and a domain generalization unit. The illumination adaptive unit is used to dynamically adjust the image color space parameters to adapt to different lighting conditions. The domain generalization unit uses a domain adversarial training strategy to ensure that the defect detection model maintains the stability of detection performance across different production lines and different products.

[0017] Preferably, the intelligent visual inspection module further includes a few-shot transfer learning unit and a defect root cause analysis unit. The few-shot transfer learning unit is used to pre-train a basic inspection model on labeled data of historical products, and to achieve rapid adaptation by fine-tuning the model based on a small number of labeled samples when new products are launched. The defect root cause analysis unit is used to embed features into the detected defect images, perform similarity matching with the historical defect database, infer the process root cause of the defect, and feed the process root cause back to the intelligent scheduling module and / or equipment management module through the data platform.

[0018] Preferably, the collaborative decision engine triggers the intelligent scheduling module to reschedule based on the defect detection results output by the intelligent visual inspection module. Specifically, when the defect detection results determine that a product in a certain process needs to be reworked, the collaborative decision engine automatically inserts the rework process into the scheduling and recalculates the time window and resource allocation of the affected process.

[0019] Preferably, the data platform interacts with an external enterprise resource planning system through a standardized interface. The external enterprise resource planning system includes a business management system and a financial management system. The data platform includes a production order intermediate table management unit, which is used to manage the import status and update status of production order data imported from the external enterprise resource planning system.

[0020] (III) Beneficial Effects: Therefore, compared with the prior art, the present invention has the following beneficial effects: 1. Achieve unified management and integration of data across the entire manufacturing process. By building a unified data platform, data silos across various stages such as sales, planning, production, quality, materials, equipment, personnel, inventory, and procurement have been broken down, forming a unified business information view at the enterprise level and providing a complete, consistent, and real-time data foundation for upper-level intelligent applications.

[0021] 2. Dynamic adaptive optimization of production scheduling was achieved. Through an intelligent scheduling method based on Transformer encoders and deep reinforcement learning, the scheduling problem was modeled as a Markov decision process, enabling second-level response and adaptive adjustment to dynamic events (equipment failures, emergency orders, rework, etc.). Local rescheduling was implemented through causal inference units, avoiding the computational overhead of global recalculation. A multi-objective reward function design simultaneously optimized multiple conflicting objectives such as on-time order delivery rate, total completion time, load balancing, and equipment utilization.

[0022] 3. High accuracy and strong generalization in product defect detection were achieved. By employing a multimodal feature fusion method based on cross-camera attention, multi-view and multi-scale image information was fully utilized, improving the accuracy of defect detection. Illumination adaptation and domain adversarial training ensured stable detection performance under different lighting conditions and production line environments. Few-shot transfer learning enabled rapid adaptation to new product models, significantly shortening the debugging time for the detection system before new products went live.

[0023] 4. A closed-loop collaborative mechanism for production scheduling and quality inspection has been established. Through a collaborative decision-making engine, a two-way data loop of "production scheduling guiding inspection and inspection feedback on production scheduling" has been achieved, enabling production planning and quality control to move from being isolated to collaborative optimization, thereby improving the overall efficiency and product quality of the entire manufacturing process.

[0024] 5. It possesses excellent system compatibility and scalability. It interacts with external ERP systems such as business management and financial management through standardized interfaces, ensuring compatibility with the enterprise's existing IT infrastructure. Through a modular architecture design, each functional module can be deployed, upgraded, and expanded independently, reducing system implementation and maintenance costs.

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall architecture of an artificial intelligence-based collaborative management platform for the entire manufacturing process according to the present invention. Figure 2 This is a structural block diagram of the intelligent scheduling module of an artificial intelligence-based collaborative management platform for the entire manufacturing process according to the present invention. Figure 3 This is a structural block diagram of an intelligent visual inspection module of an artificial intelligence-based collaborative management platform for the entire manufacturing process according to the present invention. Figure 4 This is a flowchart of the production scheduling-quality inspection closed-loop collaborative mechanism of an artificial intelligence-based manufacturing end-to-end collaborative management platform according to the present invention. Figure 5 This is a schematic diagram illustrating the working principle of the Transformer-based state coding unit in the intelligent scheduling module of an AI-based collaborative management platform for the entire manufacturing process of the present invention. Figure 6 This is a schematic diagram of feature fusion in the cross-camera attention mechanism of the intelligent vision inspection module of an AI-based collaborative management platform for the entire manufacturing process. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the following description is provided in conjunction with the appendix. Figures 1-6 The embodiments of the present invention will be further described in detail below.

[0028] Example 1: Overall Platform Architecture like Figure 1 As shown, the AI-based collaborative management platform for the entire manufacturing process provided by this invention includes a data platform, an intelligent scheduling module, an intelligent visual inspection module, and a collaborative decision engine.

[0029] The data platform connects to external ERP systems (including business management systems and financial management systems) and various data acquisition terminals on the production line (including PLCs, sensors, RFID readers, industrial cameras, etc.) through standardized interfaces to collect heterogeneous data from the entire manufacturing process in real time. The data platform cleans, standardizes, stores, and manages the collected data, and provides data access services to the intelligent scheduling module, intelligent visual inspection module, and collaborative decision engine through a unified data service interface.

[0030] The intelligent scheduling module obtains order status data, resource status data, and global event data from the data platform, performs dynamic scheduling based on deep reinforcement learning algorithms, and outputs the scheduling results (including process-team allocation, process execution order, time window, etc.) to the data platform and collaborative decision engine.

[0031] The intelligent vision inspection module obtains production scheduling information from the data platform to determine inspection priorities and strategies. It acquires product images through multiple industrial cameras, performs defect detection based on deep learning algorithms, and outputs the inspection results (including defect type, defect location, defect area, yield statistics, etc.) to the data platform and collaborative decision engine.

[0032] The collaborative decision engine dynamically adjusts the detection strategy of the intelligent vision inspection module based on the production scheduling results of the intelligent production scheduling module, and triggers the intelligent production scheduling module to reschedule based on the defect detection results of the intelligent vision inspection module, thereby realizing two-way closed-loop collaboration between production scheduling and quality inspection.

[0033] Example 2: Detailed Implementation of the Intelligent Scheduling Module: like Figure 2 As shown, the intelligent scheduling module includes a state encoding unit, a policy network, a reward function unit, a causal reasoning unit, and a human-machine collaborative interaction unit.

[0034] Specific implementation methods of the state coding unit: The status coding unit retrieves order status data, resource status data, and global event data from the data platform. Taking a production workshop of an electronic equipment manufacturing company as an example, the order status data includes: a total of 156 orders currently being executed, each order containing an order number, material number, planned start time, planned completion time, current process, order status (one of 11 statuses), and order type (interval / pipe / parts box). The resource status data includes: the current load rate, available capacity, and historical efficiency of 5 production teams and 23 pieces of equipment. The global event data includes: whether there are any urgent orders, equipment malfunctions, or rework requests.

[0035] The state encoding unit uses a multi-head attention mechanism based on the Transformer architecture to encode the above data. Specifically, each order-resource-event combination is regarded as a "term". The attention weight between each term is calculated through the multi-head self-attention mechanism to capture the long-range dependencies between different data dimensions, and finally a unified embedding vector with a dimension of 512 is generated as the state vector.

[0036] Specific implementation methods for the policy network and reward function unit: The policy network employs a deep reinforcement learning network, trained based on a proximal policy optimization algorithm. The policy network takes the state vector output by the state encoding unit as input and outputs the probability distribution of scheduling decision actions. The scheduling decision action space includes: assigning a specific process of an order to a specific work group, determining the execution order of processes within a specific work group, rescheduling affected orders when interrupted orders occur, and selecting a rework path when rework occurs.

[0037] The reward function unit calculates the reward value based on the multi-objective optimization function. The specific expression of the multi-objective optimization function is: ; in, The on-time delivery rate of orders (target value ≥ 95%). It is a negative value of the total completion time (the smaller the better). This represents a negative value for the work group's load imbalance (target value < 5%). To maximize equipment utilization (target value). , , , These are adaptive weighting coefficients.

[0038] Specific implementation methods of the causal reasoning unit: The causal reasoning unit constructs a task-resource-state causal graph based on a Bayesian network. The nodes of the causal graph include task nodes (each process in each order), resource nodes (each shift, each piece of equipment), and state nodes (equipment health status, personnel attendance status, etc.). Edges represent causal relationships, such as "equipment A failure → process X delay → order Y postponement".

[0039] When a device failure event is detected, the causal reasoning unit locates the faulty device node in the causal graph, identifies its impact propagation path (which processes and orders are affected), and only performs local rescheduling on the affected subgraph, rather than recalculating the entire production schedule globally, achieving a response time within seconds.

[0040] Specific implementation methods of the human-computer collaborative interaction unit: The human-machine collaborative interaction unit generates Top-3 candidate production scheduling plans, which are visualized on the user interface in the form of a multi-objective Pareto front. Each plan is labeled with its performance on four objectives: on-time delivery rate, total completion time, load balancing, and equipment utilization. After the planner selects a plan based on business preferences, the human-machine collaborative interaction unit uses the user's selection as a feedback signal and optimizes the weight coefficients of the reward function through feedback reinforcement learning.

[0041] Example 3: Detailed implementation of the intelligent visual inspection module: like Figure 3 As shown, the intelligent visual inspection module includes an image acquisition unit, a multimodal feature fusion unit, a defect detection unit, an illumination adaptation unit, a domain generalization unit, a few-shot transfer learning unit, and a defect root cause analysis unit.

[0042] Specific implementation methods of the image acquisition unit and the multimodal feature fusion unit: Taking the visual inspection of capacitor products as an example, industrial cameras (cameras 1 to 11) are deployed at 11 workstations. Cameras 1-6 are used for side inspection, cameras 7-8 for bottom inspection, and cameras 9-11 for top and special area inspection. Each camera synchronously acquires images under the control of a trigger signal.

[0043] The multimodal feature fusion unit employs a cross-camera attention mechanism for feature fusion. Specifically, images acquired by each camera are first processed by a backbone network (an improved next-generation end-to-end real-time object detection network, YOLOv10), which extracts multi-scale feature maps. Then, the cross-camera attention mechanism calculates attention weights based on the spatial relationships between the cameras and the geometric constraints of the detected target. For example, there is a geometric constraint between the bottom depression detection camera (camera 11) and the side detection cameras (cameras 1-6)—the position of the bottom depression should be consistent with the deformation at the corresponding position on the side. The cross-camera attention mechanism utilizes this constraint to adaptively weight and fuse the feature maps from the 11 cameras, generating a unified fused feature map.

[0044] Specific implementation method of the defect detection unit: The defect detection unit identifies and classifies defects based on fused feature maps. Defect types include: dents (area, depth), protrusions (area, height), burn marks (area, grayscale), missing welds (area, diameter), incorrect identifier orientation, dimensional deviations (diameter, length), surface scratches, etc. The defect detection unit outputs the category label, bounding box position (center coordinates, width, height), and defect area (pixel equivalent converted to physical size) for each defect.

[0045] Experimental results show that the defect detection accuracy of this invention reaches over 98%, the false negative rate is less than 1%, and the detection speed reaches over 30 frames / second, meeting the real-time detection needs of industrial sites.

[0046] Specific implementation methods of the illumination adaptive unit and the domain generalization unit: The illumination adaptive unit monitors the brightness and color distribution of the current image in real time and dynamically adjusts the threshold parameters of the hue (H), saturation (S), and luminance (V) channels. For example, when the lighting is too dim, it automatically increases the gain value of the V channel; when the lighting is too warm, it automatically adjusts the hue shift of the H channel.

[0047] The domain generalization unit employs a domain adversarial training strategy. During model training, in addition to the defect classifier and location regressor, a domain discriminator is introduced. The domain discriminator attempts to distinguish which production line or product model a feature originates from, while the feature extractor attempts to generate feature representations that the domain discriminator cannot distinguish. Through this adversarial training, the feature extractor learns domain-independent defect features, enabling the model to maintain stable detection performance across different production lines and products.

[0048] Specific implementation of the few-shot transfer learning unit: When a new product (such as a new capacitor model) is launched, the few-shot transfer learning unit first loads the base detection model pre-trained on a large-scale labeled dataset of historical products. Then, based on 50-100 labeled samples of the new product, rapid adaptation is performed through model fine-tuning and dataset augmentation (including Poisson fusion, sample mixing augmentation CutMix, random rotation, scaling, etc.). The entire adaptation process is completed within 2 hours, while traditional methods require several days of manual parameter tuning.

[0049] Specific implementation of the defect root cause analysis unit: The defect root cause analysis unit maps the detected defect images into feature vectors through a feature extraction network and performs similarity matching (using cosine similarity or Euclidean distance) with feature vectors in the historical defect database. If a historical defect record with a similarity higher than the threshold is matched, the process root cause associated with that historical record (such as "abnormal temperature causing blackening" or "abnormal pressure causing dents") is read as the possible root cause inference result of the current defect.

[0050] The defect root cause analysis unit feeds back the inference results to the intelligent scheduling module and equipment management module through the data platform, triggering process parameter adjustment suggestions or equipment maintenance work orders.

[0051] Example 4: Detailed Implementation of the Collaborative Decision Engine: like Figure 4 As shown, the collaborative decision engine enables two-way closed-loop collaboration between production scheduling and quality inspection.

[0052] Production scheduling guidance and testing process: S401 The intelligent scheduling module completes the scheduling, outputting the priority of each order, the time window of each process, and the inspection task of each piece of equipment.

[0053] S402, The collaborative decision engine generates inspection strategies based on the production scheduling results, including: prioritizing the inspection tasks of high-priority orders, reasonably arranging the inspection sequence within the process time window, and balancing the distribution of inspection tasks according to the load of each inspection device.

[0054] S403 The intelligent visual inspection module performs inspection according to the inspection strategy generated by the collaborative decision engine.

[0055] Inspection feedback and production scheduling process: S411 The intelligent vision inspection module completes defect detection and outputs the defect detection results.

[0056] S412, The defect root cause analysis unit performs root cause inference on the detected defects.

[0057] S413. The collaborative decision engine determines whether the defect requires rework. If rework is not required, only the quality data is recorded; if rework is required, the following steps are executed.

[0058] S414 The collaborative decision engine automatically inserts rework processes into the production schedule, determining the location of the rework process (by default, after the current process), the rework path, and the responsible unit.

[0059] S415 The intelligent scheduling module performs local rescheduling based on the causal reasoning unit, recalculating the time window and resource allocation of the affected processes.

[0060] S416. Based on the root cause analysis results of defects, the collaborative decision engine triggers the equipment management module to generate equipment maintenance work orders, or triggers the process parameter management module to generate process parameter adjustment suggestions.

[0061] Through the aforementioned two-way closed-loop collaborative mechanism, a complete closed loop from "defect discovery - root cause localization - production scheduling adjustment - process optimization" is achieved, enabling production planning and quality control to move from being mutually isolated to collaborative optimization.

[0062] Example 5: External system data interaction based on standardized interfaces: In another specific embodiment of the present invention, the data platform interacts with an external enterprise resource planning system through a standardized interface.

[0063] The data platform includes a production order intermediate table management unit. This unit manages production order data transmitted from the business management system through an interface, including three data states: production order intermediate table (order data that has been transmitted to this system), production orders waiting to be imported (order tasks that have been transmitted but not yet issued to this system), and production orders waiting to be updated (order data that has been issued to this system but whose information has been re-maintained by the business management side and needs to be updated to this system).

[0064] Planners view order data imported from the business management system through the production order intermediate table management unit. For orders awaiting import, they perform import operations to add them to the production order list in their own system. For orders awaiting updates, they perform update operations to ensure data consistency between their system and the business management system. When material codes are created in the external financial management system, the data platform uses the material code matching function to match and associate the material codes in the financial management system with the material information on this platform, ensuring consistency of material data between the two systems.

[0065] Through the aforementioned standardized interfaces and data synchronization mechanisms, this platform can be seamlessly integrated with existing ERP systems such as business management and financial management, and is compatible with the enterprise's existing information infrastructure, reducing system switching costs and implementation difficulties.

[0066] In summary, this invention provides an AI-based collaborative management platform for the entire manufacturing process. It integrates data from sales, production, quality, materials, and equipment through a data platform; it achieves dynamic adaptive scheduling of production orders through deep reinforcement learning and a Transformer encoder, enabling rapid response to abnormal events such as equipment failures, emergency order insertions, and rework; it achieves high-precision real-time detection of product defects through deep learning visual inspection and cross-camera attention mechanisms; and it establishes a two-way closed-loop feedback mechanism between scheduling and quality inspection through a collaborative decision engine. This invention achieves end-to-end data integration, dynamic scheduling optimization, high-precision and strong generalization of detection, and collaborative linkage between scheduling and quality inspection, significantly improving production efficiency and product quality.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A collaborative management platform for the entire manufacturing process based on artificial intelligence, characterized in that, include: The data platform is used to uniformly collect, store, and manage heterogeneous data across the entire manufacturing process, which includes at least the following stages: sales management, production planning, production execution, quality inspection, material management, equipment management, personnel management, inventory management, and procurement management. The intelligent scheduling module is connected to the data platform and is used to dynamically schedule production orders based on deep reinforcement learning algorithms. The intelligent visual inspection module is connected to the data platform and is used to perform real-time defect detection on products on the production line based on deep learning algorithms. A collaborative decision engine is connected to the intelligent scheduling module and the intelligent visual inspection module, respectively. It is used to trigger the intelligent scheduling module to reschedule based on the defect detection results output by the intelligent visual inspection module, and to dynamically adjust the detection strategy of the intelligent visual inspection module based on the scheduling results output by the intelligent scheduling module.

2. The AI-based collaborative management platform for the entire manufacturing process as described in claim 1, characterized in that, The intelligent scheduling module includes: A status encoding unit is used to encode the order status data, resource status data, and global event data provided by the data platform into a status vector. A strategy network is used to output production scheduling decision actions based on the state vector. The production scheduling decision actions include process-team allocation decision, process execution order decision, order insertion and rescheduling decision, and rework path selection decision. The reward function unit is used to calculate the reward value for production scheduling decisions based on a multi-objective optimization function, which includes maximizing the on-time delivery rate of orders, minimizing the total completion time, minimizing the shift load imbalance, and maximizing equipment utilization.

3. The AI-based collaborative management platform for the entire manufacturing process as described in claim 2, characterized in that, The state encoding unit uses a multi-head attention mechanism based on the classic deep neural network Transformer architecture with self-attention mechanism to encode the order state data, resource state data and global event data, generating a unified embedding vector as the state vector; the policy network uses a deep reinforcement learning network and is trained based on the proximal policy optimization algorithm.

4. The AI-based collaborative management platform for the entire manufacturing process as described in claim 2, characterized in that, The intelligent scheduling module also includes a causal reasoning unit, which is used to construct a task-resource-state causal graph based on a Bayesian network. When an abnormal event is detected, the unit identifies the propagation path of the abnormal event in the causal graph and performs local rescheduling only on the affected subgraph.

5. The AI-based collaborative management platform for the entire manufacturing process according to claim 2, characterized in that, The intelligent scheduling module also includes a human-machine collaborative interaction unit, which generates multiple candidate scheduling schemes and displays them in the form of a multi-objective Pareto front, receives user selection instructions, and uses the user selection as a feedback signal to optimize the reward function of the policy network.

6. The artificial intelligence-based collaborative management platform for the entire manufacturing process according to claim 1, characterized in that, The intelligent visual inspection module includes: The image acquisition unit is used to acquire multi-source image data of the product from different workstations and different perspectives using multiple industrial cameras; A multimodal feature fusion unit is used to adaptively fuse the multi-source image data through a cross-camera attention mechanism to generate a fused feature map; The defect detection unit is used to identify and classify defects in the product based on the fused feature map, and output the defect type, defect location and defect area.

7. The AI-based collaborative management platform for the entire manufacturing process according to claim 6, characterized in that, The multimodal feature fusion unit uses a cross-camera attention mechanism to adaptively weight and fuse the multi-scale features extracted by each industrial camera. The cross-camera attention mechanism calculates the attention weights based on the spatial position relationship of each camera's corresponding workstation and the geometric constraint relationship of the detection target. The intelligent visual inspection module also includes an illumination adaptive unit and a domain generalization unit. The illumination adaptive unit is used to dynamically adjust the image color space parameters to adapt to different lighting conditions. The domain generalization unit adopts a domain adversarial training strategy to ensure that the defect detection model maintains the stability of detection performance across different production lines and different products.

8. The AI-based collaborative management platform for the entire manufacturing process according to claim 6, characterized in that, The intelligent visual inspection module further includes a few-shot transfer learning unit and a defect root cause analysis unit. The few-shot transfer learning unit is used to pre-train a basic inspection model on labeled data of historical products, and to achieve rapid adaptation by fine-tuning the model based on a small number of labeled samples when new products are launched. The defect root cause analysis unit is used to embed features into the detected defect images, perform similarity matching with the historical defect database, infer the process root cause of the defect, and feed the process root cause back to the intelligent scheduling module and / or equipment management module through the data platform.

9. The artificial intelligence-based collaborative management platform for the entire manufacturing process according to claim 1, characterized in that, The collaborative decision engine triggers the intelligent scheduling module to reschedule based on the defect detection results output by the intelligent vision inspection module. Specifically, when the defect detection results determine that a product in a certain process needs to be reworked, the collaborative decision engine automatically inserts the rework process into the scheduling and recalculates the time window and resource allocation of the affected process.

10. A collaborative management platform for the entire manufacturing process based on artificial intelligence according to any one of claims 1 to 9, characterized in that, The data platform interacts with an external enterprise resource planning (ERP) system through a standardized interface. The external ERP system includes a business management system and a financial management system. The data platform includes a production order intermediate table management unit, which manages the import and update status of production order data imported from the external ERP system.

Images