Artificial Intelligence-Based Furniture Board Optimization Method and System

Abstract

This invention relates to the fields of artificial intelligence and intelligent manufacturing technology, specifically disclosing an AI-based method and system for optimizing furniture board materials. The system includes a data perception and preprocessing module, a multi-constraint fusion modeling module, a deep reinforcement learning decision-making module, a post-optimization and simulation module, and a system control and execution module. By collecting and standardizing multi-source production data, it constructs an optimization model embedded with multiple constraints, uses deep reinforcement learning to iteratively generate a layout scheme that approximates the global optimum, performs local optimization and cutting simulation verification, and finally converts it into equipment command execution. This invention achieves end-to-end intelligent board material optimization, improving material utilization and production adaptability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent manufacturing technology, specifically relating to an artificial intelligence-based method and system for optimizing furniture boards. Background Technology

[0002] In the fields of furniture manufacturing and materials science, the optimized design and efficient utilization of wood panels are core technologies for improving production efficiency, reducing costs, and achieving sustainable development. Furniture panel optimization methods aim to maximize raw material utilization and reduce waste generation through scientific calculation and planning, while meeting product structural strength, aesthetic requirements, and production constraints.

[0003] Among them, panel layout and cutting technology based on computer-aided design and optimization algorithms is an important research direction in this field. The basic goal of this technology is to automatically or semi-automatically generate a layout scheme with high material utilization and reasonable cutting path based on the size and quantity of the furniture parts to be processed and the specifications of the panels.

[0004] Existing technologies primarily rely on human experience or fixed, heuristic nesting algorithms. These methods exhibit significant limitations when dealing with complex and varied furniture component shapes and sizes, as well as dynamic combinations of production orders. Manual nesting heavily depends on operator skill, resulting in low efficiency and difficulty in ensuring optimal material utilization. Traditional fixed algorithms, on the other hand, lack self-learning and adaptive capabilities, failing to adjust strategies based on historical production data and real-time demands. Consequently, material waste remains a prominent issue when dealing with non-standard components or small-batch, multi-order orders.

[0005] Furthermore, existing methods often struggle to effectively integrate and optimize various constraints in actual production, such as texture orientation, processing defect avoidance, and equipment cutting capabilities, with the layout scheme. This results in insufficient adaptability of the final solution to the production site, impacting overall production efficiency and cost control. Therefore, achieving intelligent board material optimization with high material utilization and high production adaptability has become an urgent technical challenge. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for optimizing furniture boards based on artificial intelligence, so as to solve the technical contradictions of low material utilization, lack of self-learning and self-adaptation capabilities, and difficulty in effectively integrating multiple production constraints in the existing technology.

[0007] The artificial intelligence-based furniture board optimization method and system provided by this invention comprises an integrated intelligent optimization architecture. This system architecture is collaboratively composed of a data perception and preprocessing module, a multi-constraint fusion modeling module, a deep reinforcement learning decision-making module, a post-optimization and simulation module, and a system control and execution module.

[0008] The data perception and preprocessing module is responsible for collecting and standardizing historical order data from the enterprise resource planning system, geometric data of furniture components from the computer-aided design system, real-time equipment status data from the manufacturing execution system, and surface quality data of boards obtained through the visual inspection system.

[0009] The multi-constraint fusion modeling module is used to transform the collected data into a unified mathematical model of the optimization problem. This model simultaneously embeds the material utilization objective function as well as various production constraints, including texture direction constraints, processing defect avoidance area constraints, minimum cutting spacing constraints, and equipment cutting capability constraints.

[0010] As the core of the system, the deep reinforcement learning decision module receives the optimization problem model from the multi-constraint fusion modeling module and, based on a pre-trained and online fine-tuned deep neural network strategy, efficiently explores and utilizes the huge solution space to iteratively generate a plate layout scheme that approximates the global optimum.

[0011] The post-scheme optimization and simulation module performs local fine-tuning of the preliminary layout scheme output by the deep reinforcement learning decision module, and performs cutting path simulation and production time evaluation based on the physics engine to verify the feasibility and economy of the scheme.

[0012] The system control and execution module is ultimately responsible for coordinating the workflow of each module and converting the verified optimized nesting scheme into instruction codes that can be recognized by the CNC cutting equipment for execution.

[0013] Furthermore, the specific data processing flow of the data perception and preprocessing module is as follows. This module first connects to the enterprise resource planning system to extract information on furniture component types, quantities, dimensions, and the types of boards used from historical orders.

[0014] Meanwhile, the module reads two-dimensional vector graphic data of furniture components from the computer-aided design system interface and automatically identifies key geometric features of the components, including the outer contour polygon and the location and shape of internal openings.

[0015] For equipment status data from the manufacturing execution system, this module focuses on analyzing the type of currently available cutting equipment, maximum processing size, number of cutting heads, and tool magazine configuration.

[0016] The surface quality data of the board material provided by the visual inspection system is used to identify the specific location and boundaries of processing defects such as knots and cracks on the board material.

[0017] After all the aforementioned multi-source heterogeneous data are collected, they undergo unified format conversion and coordinate system unification by the data standardization unit within this module, and are stored in the system's dedicated real-time database, providing a consistent and reliable data foundation for subsequent optimization calculations.

[0018] Furthermore, the mathematical model construction process of the multi-constraint fusion modeling module is as follows. This module defines the material utilization objective function as the ratio of the total area of ​​all arranged furniture parts to the total area of ​​the boards used, and the optimization objective is to maximize this ratio.

[0019] Texture orientation constraints are modeled as imposing a permissible range of rotation angles on each furniture component, determined by the texture orientation requirements of the component during the design phase. Defect avoidance area constraints are implemented by transforming visually inspected defect areas into prohibited placement areas during the nesting process.

[0020] The minimum cutting spacing constraint requires that the shortest distance between any two adjacent layout component outlines must not be less than a preset safety value, which is determined by the physical characteristics of the cutting equipment.

[0021] The equipment cutting capability constraint is further refined into a maximum cutting length constraint and a minimum inner corner radius constraint, ensuring that the generated nesting path can be physically executed by the specific equipment. This module integrates all the above objectives and constraints into a combinatorial optimization problem with multiple constraints, and expresses it in a state-action space form that the deep reinforcement learning decision module can handle.

[0022] Furthermore, the core algorithm implementation of the deep reinforcement learning decision module is as follows. This module adopts a deep reinforcement learning algorithm based on the actor-critic framework. Its state space is defined as the remaining available area of ​​the current board material, the current state of the set of furniture parts to be arranged, and the layout of the arranged parts.

[0023] The action space is defined as selecting a component from the set of components to be arranged and determining its specific position and rotation angle on the board. The actor network is responsible for outputting the probability distribution of an action based on the current state, i.e., the policy function; the critic network is responsible for evaluating the long-term expected cumulative reward obtained by performing an action in a given state, i.e., the value function.

[0024] The design of the reward signals is closely centered around the optimization objectives, and mainly includes instant material utilization increments, penalties for violating constraints, and rewards for encouraging the exploration of novel layout patterns.

[0025] In the offline phase, this module uses large-scale historical nesting data for pre-training to learn basic nesting strategies. After online deployment, it continuously collects new nesting experience data to fine-tune the network parameters through interaction with the environment, i.e., the post-scheme optimization and simulation modules, thereby achieving adaptive evolution of the strategy and performance improvement.

[0026] Furthermore, the fine-tuning and verification process of the post-scheme optimization and simulation module is as follows. After receiving the preliminary layout scheme generated by the deep reinforcement learning decision module, this module first starts a local search algorithm. This algorithm focuses on merging and optimizing the tiny gaps between components in the layout scheme, attempting to fill these gaps by fine-tuning the position or angle of individual components, thereby further improving material utilization.

[0027] After completing local optimization, the module calls the built-in physical simulation engine to dynamically simulate the actual motion trajectory of the CNC cutting head based on the layout scheme and preset equipment cutting parameters.

[0028] The simulation process accurately calculates the total cutting path length, estimates the processing time, and automatically detects whether there is tool interference or illegal operation beyond the equipment's travel range.

[0029] The simulation results, along with the optimized layout scheme, are fed back to the system control and execution module as the basis for the final scheme decision and execution.

[0030] Furthermore, the working mechanism of the system control and execution module is as follows. As the system's central control unit, this module is responsible for initializing the data acquisition tasks of the data perception and preprocessing module, triggering the model building of the multi-constraint fusion modeling module, scheduling the deep reinforcement learning decision-making module to search for sampling schemes, and managing the verification process of the scheme optimization and simulation module.

[0031] Upon receiving a feasible layout scheme verified through simulation, the code generation unit within this module converts the component contour coordinates, cutting sequence, and process parameters in the layout scheme into G-code or other standard CNC instructions conforming to a specific CNC system format. These instructions are then sent directly to the CNC cutting equipment control cabinet in the workshop via industrial communication protocols, driving the equipment to complete the automated cutting operation of the sheet metal.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. This invention achieves end-to-end automated sheet metal optimization from order data to equipment instructions by constructing a complete system architecture integrating data perception, multi-constraint modeling, deep reinforcement learning decision-making, post-scheme optimization and simulation, and execution control. The application of the deep reinforcement learning decision-making module enables the system to learn from massive amounts of historical data and online interaction, continuously optimizing its nesting strategy. This surpasses traditional methods that rely on fixed heuristic rules or human experience, demonstrating superior adaptability and optimization capabilities when facing complex component shapes and dynamic order combinations, thereby fundamentally improving material utilization.

[0034] 2. This invention, through a multi-constraint fusion modeling module, systematically and accurately embeds key production constraints such as texture direction, processing defect avoidance, and equipment cutting capabilities into the optimization mathematical model. This ensures that the generated layout scheme not only has high material utilization in theory but also closely matches specific production conditions, process requirements, and quality standards in practice. This avoids the problem of the scheme being out of touch with actual production, reduces rework and adjustments caused by impractical schemes, and improves the first-time success rate and overall efficiency of production.

[0035] 3. This invention introduces a post-design optimization and simulation module to perform local fine-tuning and physical simulation verification of the initial layout scheme. Local optimization further explores the potential for improving material utilization, while physical simulation rehearses the cutting process before the scheme is put into actual production. It can identify and avoid potential operational risks such as equipment interference and path exceeding limits, ensuring the safety and feasibility of the production process, reducing equipment wear and production failure rate, and providing a solid guarantee for stable and efficient intelligent manufacturing. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall technical solution architecture of the furniture board optimization method and system based on artificial intelligence proposed in this invention;

[0037] Figure 2 This is a schematic diagram of the core principle framework of the deep reinforcement learning decision module in this invention;

[0038] Figure 3 This is a logical flowchart of the data perception and preprocessing module in this invention;

[0039] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the multi-constraint fusion modeling module and the deep reinforcement learning decision module in this invention;

[0040] Figure 5 This is a flowchart illustrating the fine-tuning and verification logic of the post-optimization and simulation module in this invention. Detailed Implementation

[0041] Please refer to the attached document. Figure 1 This embodiment details the specific implementation architecture and workflow of an AI-based furniture board optimization system. This system is a highly integrated hardware and software collaborative platform. Its core lies in transforming the original furniture production requirements into a high-efficiency board layout scheme that can directly drive CNC cutting equipment through a series of sophisticated data processing and intelligent decision-making modules. The system is deployed on an industrial server cluster and interacts with the enterprise resource planning system server, computer-aided design system workstations, manufacturing execution system servers, vision inspection system industrial computers, and CNC cutting equipment control cabinets within the workshop via standard industrial Ethernet.

[0042] The data perception and preprocessing module is the system's data entry point; its specific implementation details are as follows. Please refer to the appendix. Figure 3 This module comprises four independent data acquisition sub-units and one central data standardization and storage sub-unit. The enterprise resource planning system data acquisition sub-unit connects to the enterprise resource planning system's order database via an open database interconnect interface. This sub-unit is configured with a scheduled task that automatically queries newly added or modified order records every 5 minutes.

[0043] The query command precisely filters out fields that include a list of furniture parts, the quantity of parts, the size specifications of parts, and the specified board material model.

[0044] For example, for a standard 2440mm×1220mm sheet, the system will record its unique identifier, thickness specification, and inventory quantity.

[0045] The data acquisition subunit of the computer-aided design system communicates with the computer-aided design software through an application programming interface (API). This interface supports directly parsing the internal data structure of the design file and extracting the two-dimensional contour vector information of each furniture component.

[0046] Vector information is represented by a series of ordered two-dimensional coordinate points. The system's built-in geometric feature recognition algorithm will traverse these coordinate sequences, automatically construct the outer contour polygon of the component, and detect whether there are openings inside.

[0047] For the detected openings, the algorithm will also generate their boundary polygons and record their positions relative to the origin of the component.

[0048] The data acquisition subunit of the manufacturing execution system continuously monitors the status messages of available cutting equipment in the workshop through the real-time data service provided by the manufacturing execution system.

[0049] The message parser will extract the device identifier, current worktable size, number of cutting heads, diameter and type of loaded tools, and whether the device is currently idle or faulty.

[0050] The visual inspection system's data acquisition subunit is integrated with an industrial camera system deployed at the sheet material loading station. The camera immediately captures high-resolution surface images after each sheet is loaded. The image processing algorithm, based on a pre-trained deep learning model, identifies knots and cracks in the images and outputs the bounding box coordinates of these defective areas. The coordinates have been converted to an absolute coordinate system with the lower left corner of the sheet as the origin.

[0051] The central data standardization and storage subunit receives all data from the above four subunits. This subunit first transforms the coordinate systems from different sources to a global coordinate system with the lower left corner of the board as the origin, and the unit is unified to mm.

[0052] All geometric data, including part outlines and defect areas, are converted into standard scalable vector graphics format fragments.

[0053] Text and numerical data, such as component type, quantity, and device parameters, are serialized into predefined key-value pair structures.

[0054] The processed data packets are timestamped and stored in the system's real-time database. The database adopts a distributed architecture, and the main tables include an order table, a component geometry table, an equipment status table, and a sheet metal defect table. Each table is indexed by the order number and the sheet metal number.

[0055] The multi-constraint fusion modeling module is responsible for constructing a mathematical model from standardized data that can be used for intelligent decision-making. The specific implementation process is as follows.

[0056] Please refer to the attached document. Figure 4 This module starts after receiving a data ready signal from the data sensing and preprocessing module.

[0057] The module first initializes an optimization problem instance, the core of which is the material utilization objective function.

[0058] The objective function is mathematically defined as the ratio of the total area of ​​the board material minus the sum of the areas of all laid-out component outline polygons to the total area of ​​the board material. The goal of system optimization is to maximize this ratio, that is, to minimize the waste area.

[0059] Subsequently, the module loads and instantiates each production constraint one by one.

[0060] The texture orientation constraint loader reads the texture orientation metadata for each part from the part geometry table. This metadata is specified by the designer during the computer-aided design phase and is typically expressed as a range of allowed rotation angles, such as 0° or 90°, or a continuous range such as 0° to 180°.

[0061] The system creates a rotation angle constraint object for each component, which restricts the placement orientation of the component during the nesting decision process.

[0062] The processing defect avoidance area constraint processor extracts all identified defect area polygons from the sheet defect table and marks these polygon areas as absolutely prohibited placement areas in the nesting process.

[0063] The system generates a tangent circle within the restricted area for each defect region, which is used for rapid collision detection in the early stages of the nesting algorithm.

[0064] The minimum cutting spacing constraint setter reads a preset safety distance value from the equipment status table based on the physical parameters of the currently selected cutting equipment. This value is usually between 2mm and 5mm, depending on the cutter diameter and the material of the sheet.

[0065] The constraint is modeled as requiring that the shortest Euclidean distance between any two nesting component outline polygons must be greater than or equal to this safety value.

[0066] The equipment cutting capability constraint evaluator is further refined into two sub-constraints.

[0067] The maximum cutting length constraint checks whether the length of any side in the polygonal outline of the component is less than or equal to the effective travel of the equipment table.

[0068] The minimum inner corner radius constraint analyzes all concave corners of the component profile to ensure that the radius of curvature at the concave corner is greater than or equal to the minimum turning radius that the cutting head of the equipment can achieve, typically 1 mm.

[0069] Finally, this module integrates all defined objective functions and constraints to construct a complete constraint satisfaction problem model.

[0070] The state of the model is represented as a multi-dimensional vector, including the geometric description of the remaining available area of ​​the board, the current state vector of the list of parts to be laid out, and the position and orientation matrix of the already laid-out parts. The action space is defined as selecting a part index from the set of parts to be laid out, and assigning it a two-dimensional coordinate within the placeable area of ​​the board and a rotation angle within an allowable range.

[0071] This complete optimization model is encapsulated as a standard environment interface and output to the deep reinforcement learning decision module.

[0072] The deep reinforcement learning decision-making module is the intelligent core of the system, and its specific algorithm and operating mechanism are as follows. Please refer to the appendix. Figure 2 This module is implemented using the asynchronous advantage actor critic algorithm framework.

[0073] The module is deployed on a dedicated computing server equipped with two graphics processors to accelerate the training and inference process of deep neural networks.

[0074] The actor-critic framework comprises two deep neural networks that work together: the actor network and the critic network.

[0075] The actor network employs an encoder structure that combines convolutional neural networks and long short-term memory networks.

[0076] Its input state is received from the state vector output by the multi-constraint fusion modeling module, and first passes through a fully connected layer for dimension unification and feature extraction.

[0077] The feature vectors are then fed into a multilayer perceptron with three hidden layers, each containing 512 neurons, and activated using a modified linear unit activation function.

[0078] The network's final output layer is a flexible maximum layer, whose output dimension equals the number of all possible discretized actions, which is the product of the number of parts to be sorted and the number of position and angle discretization grid points. Each element in the output vector represents the probability of selecting the corresponding action in the current state.

[0079] The critic network has an encoder structure similar to, but not exactly the same as, the actor network. Its input is also a state vector, but its output layer is a single linear unit used to estimate the long-term value function of the current state, that is, the expected cumulative reward that can be obtained from the current state by following the current policy.

[0080] The design of the reward function is crucial to the algorithm's success, and its calculation is performed after each sorting step. The immediate reward consists of three parts.

[0081] The first part is the material utilization increment, which is calculated as the increase in the area of ​​the sheet metal already used after the component is placed in the current step.

[0082] The second part is the constraint violation penalty. If the placement action causes the part to overlap with the already laid-out part, intrude into the defect area, or violate the texture direction or spacing constraint, a penalty signal with a value of -10 is applied.

[0083] The third part is the exploration reward, which encourages the network to try patterns that have not yet been fully explored. Its calculation is based on the frequency of action access; the lower the access frequency, the higher the exploration reward. The entire deep reinforcement learning decision-making module operates in two phases.

[0084] During the offline pre-training phase, the system uses more than 100,000 sets of historical layout scheme data stored in the database as an experience playback buffer to conduct large-scale supervised learning and reinforcement learning training on the actor network and the critic network. The training cycle is at least 500,000 iterations, with the goal of enabling the network to initially master effective layout strategies.

[0085] During the online fine-tuning phase, when the system processes real production orders, the deep reinforcement learning decision-making module and the post-optimization and simulation module form a closed-loop learning environment.

[0086] Each generated layout scheme and its subsequent simulation evaluation results are recorded as new empirical data. This data is added to the empirical replay buffer in real time and triggers a small batch of network parameter updates. The update frequency is usually once every 10 orders processed, thereby achieving continuous online optimization and adaptation of the strategy.

[0087] The actor network is responsible for generating the layout strategy, which is a probability distribution that defines the probability of selecting any component and its placement position and angle given the current board state and the set of components to be laid out.

[0088] The policy function is a parameterized function based on a deep neural network.

[0089] The critic network is responsible for evaluating the state value, which is a scalar value representing the expected total reward that can be obtained starting from the current sorting state and following the strategy generated by the actor network.

[0090] The calculation of the value function also relies on a deep neural network.

[0091] The design of the reward signal is crucial for guiding the agent's learning; its calculation is performed immediately after each sorting action. The reward consists of three parts.

[0092] The first part is the instant material utilization bonus, which is equal to the ratio of the area of ​​the component placed this time to the total area of ​​the board.

[0093] The second part is the constraint penalty item. If the placement action causes a conflict with any production constraints, such as the component rotation angle exceeding the allowable range, the component overlapping with a defective area, or the distance between the component and the already laid-out component being less than the safety value, a fixed negative value will be deducted from the reward. This negative value is usually set to 10 times the component area reward.

[0094] The third part is the exploration reward, which encourages the algorithm to try new, underexplored sorting patterns. Its value is inversely proportional to the historical frequency of action selection.

[0095] The post-design optimization and simulation module is responsible for refining and verifying the feasibility of the initial layout plan. Its specific implementation process is as follows. Please refer to the appendix. Figure 5 This module is activated after receiving the preliminary sorting scheme output by the deep reinforcement learning decision module.

[0096] The module first initiates a local fine-tuning submodule, which implements a local search strategy based on the simulated annealing algorithm.

[0097] The algorithm uses the preliminary solution as its initial solution, defining its neighborhood action as randomly selecting a pre-arranged component and performing a small translation or rotation on it. The translation distance is limited to ±5mm, and the rotation angle is limited to ±1°. The objective function is to further improve the material utilization rate of the current solution while ensuring that all production constraints are still met.

[0098] The algorithm searches for a better solution within 2000 iterations, and replaces the original solution if found. After completing local optimization, the solution enters the physical simulation verification submodule.

[0099] This submodule integrates a physics engine-based path cutting simulator.

[0100] The simulator first loads the optimized nesting scheme and the current cutting equipment parameters obtained from the manufacturing execution system, including cutting speed, idle speed, tool diameter and piercing time.

[0101] The simulator dynamically calculates the motion trajectory of the cutting head based on a preset cutting sequence strategy, such as the nearest neighbor method.

[0102] The trajectory calculation is accurate to the millimeter level, and the acceleration, constant speed and deceleration process of the cutting head are simulated.

[0103] The simulation process calculates and accumulates the total cutting path length in real time, and estimates the total processing time based on the cutting speed and idle movement speed.

[0104] Meanwhile, the collision detection algorithm continues to run, checking whether the cutting head will interfere with the fixture, the cut part or other equipment structure during the movement, and verifying whether the cutting trajectory exceeds the physical travel limit of the equipment table.

[0105] After the simulation is completed, a detailed simulation report is generated, which includes the final material utilization rate, total cutting path length, estimated processing time, and any detected interference or over-limit warnings.

[0106] The report, along with the final optimized layout scheme data package, was sent to the system control and execution module.

[0107] The system control and execution module serves as the central scheduling and command distribution center of the system, and its specific working mechanism is as follows.

[0108] This module runs on the system's main control server. Its core is a state machine engine, responsible for coordinating the timing and data flow of all modules. The state machine is initially in an idle state.

[0109] When a new production order arrives, the state machine first triggers the data sensing and preprocessing module to start the data acquisition task and waits for it to return a data ready signal.

[0110] Upon receiving the signal, the state machine activates the multi-constraint fusion modeling module to optimize model construction.

[0111] After the model is built, the state machine scheduling deep reinforcement learning decision module starts the sorting scheme search process.

[0112] The search process may take anywhere from a few seconds to a few minutes, depending on the complexity and number of the parts to be sorted.

[0113] Once the deep reinforcement learning decision module outputs a solution, the state machine immediately calls the solution optimization and simulation module for further processing.

[0114] After the state machine waits for the simulation verification to complete and receives the final solution package, its internal code generation unit begins to work.

[0115] This unit analyzes the outline of each component in the layout scheme, and generates corresponding CNC instructions based on the final coordinates and orientation of the component on the sheet metal, combined with the specific CNC system syntax rules obtained from the manufacturing execution system.

[0116] Taking the common G-code as an example, the generator will generate a series of G00 quick positioning instructions, G01 straight line cutting instructions, and G02 or G03 arc cutting instructions for each part contour.

[0117] The instructions precisely include cutting speed, spindle speed, and tool compensation parameters. The generated complete G-code program is sent to the control cabinet of the target CNC cutting equipment via a file transfer protocol.

[0118] After receiving and verifying the code, the control cabinet returns an acknowledgment signal to the main control server, and then drives the cutting equipment to perform automated processing operations.

[0119] Meanwhile, the system control and execution module records all the data from this optimization process, including the final layout plan, material utilization rate, simulation report, and equipment execution status, into the system log database for subsequent analysis and report generation.

[0120] This embodiment, through the detailed modular design and in-depth technical integration described above, constructs a fully automated, self-learning, and high-precision furniture board optimization cutting system, from data acquisition to equipment execution. The system not only improves raw material utilization but also ensures the high feasibility and safety of the production plan through rigorous constraint modeling and physical simulation, greatly promoting the intelligent upgrading of the furniture manufacturing industry.

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A furniture board optimization system based on artificial intelligence, characterized in that, include: The data perception and preprocessing module is used to collect and standardize historical order data from the enterprise resource planning system, geometric data of furniture components from the computer-aided design system, real-time equipment status data from the manufacturing execution system, and surface quality data of boards obtained through the visual inspection system. The multi-constraint fusion modeling module is used to transform the collected data into a unified mathematical model of the optimization problem. This model simultaneously embeds the material utilization objective function as well as a variety of production constraints, including texture direction constraints, processing defect avoidance area constraints, minimum cutting spacing constraints, and equipment cutting capability constraints. The deep reinforcement learning decision module receives the optimization problem model from the multi-constraint fusion modeling module and explores and utilizes the solution space based on the pre-trained and online fine-tuned deep neural network strategy to iteratively generate a plate layout scheme that approximates the global optimum. The post-scheme optimization and simulation module is used to make local fine adjustments to the preliminary layout scheme output by the deep reinforcement learning decision module, and to perform cutting path simulation and production time evaluation based on the physics engine to verify the feasibility and economy of the scheme. The system control and execution module is used to coordinate the workflow of each module and convert the verified optimized nesting scheme into instruction codes that can be recognized by the CNC cutting equipment for execution. The data sensing and preprocessing module includes: The data acquisition unit of the enterprise resource planning system is used to extract information on furniture component types, quantities, dimensions, and the types of boards used in historical orders. The data acquisition unit of the computer-aided design system is used to read the two-dimensional vector graphic data of furniture parts and automatically identify the key geometric features of the parts, including the outer contour polygon and the location and shape of the internal openings; The manufacturing execution system data acquisition unit is used to analyze the type of currently available cutting equipment, maximum processing size, number of cutting heads, and tool magazine configuration; The visual inspection system's data acquisition unit is used to identify the specific location and boundaries of processing defects such as knots and cracks on the board material. The data standardization unit is used to perform unified format conversion and coordinate system unification on the collected multi-source heterogeneous data, and store it in the system's dedicated real-time database; The mathematical model construction process of the multi-constraint fusion modeling module is as follows: the objective function of material utilization rate is defined as the ratio of the total area of ​​all arranged furniture parts to the total area of ​​the board used, and the optimization objective is to maximize this ratio; Texture orientation constraints are modeled as imposing a range of permissible rotation angles on each furniture component; The processing defect avoidance area constraint is achieved by converting the defect areas identified by visual inspection into prohibited placement areas during the nesting process; The minimum cutting spacing constraint requires that the shortest distance between any two adjacent layout component outlines must not be less than a preset safety value; The equipment cutting capacity constraint is refined into a maximum cutting length constraint and a minimum inner corner radius constraint. All objectives and constraints are integrated into a combinatorial optimization problem with multiple constraints, and expressed in a state-action space form that can be handled by the deep reinforcement learning decision module; The deep reinforcement learning decision module adopts a deep reinforcement learning algorithm based on the actor-critic framework; The state space is defined as the remaining available area of ​​the current board, the current state of the set of furniture components to be laid out, and the layout of the already laid-out components; The motion space is defined as selecting components from the set of components to be laid out and determining their specific positions and rotation angles on the board. The actor network is responsible for outputting the probability distribution of actions based on the current state; The critic network is responsible for evaluating the long-term expected cumulative reward that can be obtained by performing an action in a given state; The incentive signal design includes instant material utilization increments, penalties for violating constraints, and rewards for exploring novel layout patterns; This module is pre-trained using historical sorting data during the offline phase, and after online deployment, it continuously collects new sorting experience data through interaction with the environment to fine-tune the network parameters. The post-optimization and simulation module of the proposed scheme includes: The local fine-tuning submodule is used to implement a local search strategy based on the simulated annealing algorithm. It takes the preliminary layout scheme as the initial solution and further improves the material utilization rate by randomly selecting the layout components for slight translation or rotation. The physical simulation verification submodule is used to integrate a physics engine-based cutting path simulator, load the optimized layout scheme and cutting equipment parameters, dynamically calculate the motion trajectory of the cutting head, calculate the total cutting path length and estimated processing time in real time, and perform collision detection and equipment travel range verification. The system control and execution module includes: The state machine engine is used to coordinate the timing and data flow of all modules, sequentially triggering the data perception and preprocessing module to start data acquisition, activating the multi-constraint fusion modeling module to optimize model construction, scheduling the deep reinforcement learning decision module to start the sorting scheme search, and calling the scheme optimization and simulation module for subsequent processing; The code generation unit is used to parse the component outline in the nesting scheme, and generate corresponding CNC instructions based on the final coordinates and orientation of the component on the sheet metal, combined with the specific CNC system syntax rules obtained from the manufacturing execution system.

2. The furniture board optimization system based on artificial intelligence according to claim 1, characterized in that, The actor network adopts an encoder structure that combines convolutional neural networks and long short-term memory networks; The input state is processed by a fully connected layer for dimensionality unification and feature extraction, and then fed into a multilayer perceptron containing multiple hidden layers. The final output layer is a flexible maximum layer, whose output dimension is equal to the number of all possible discretization actions; The critic network has an encoder structure similar to, but not exactly the same as, the actor network. Its output layer is a single linear unit used to estimate the long-term value function of the current state.

3. The furniture board optimization system based on artificial intelligence according to claim 2, characterized in that, The CNC instructions generated by the code generation unit include G00 rapid positioning instructions, G01 linear cutting instructions, and G02 or G03 circular cutting instructions; the instructions precisely include cutting speed, spindle speed, and tool compensation parameters; the generated complete G-code program is sent to the control cabinet of the target CNC cutting equipment via a file transfer protocol.

4. The furniture board optimization system based on artificial intelligence according to claim 3, characterized in that, The system operates within a multi-timescale hierarchical framework, including a strategic layer, a tactical layer, and an operational layer. The strategic layer is used to set long-term supply and demand goals and game theory frameworks on a monthly timescale; The tactical layer is used to plan major logistics routes and inventory strategies on a weekly timescale. The operation layer is used to perform the search for the layout scheme output by the deep reinforcement learning decision module and the verification of the scheme output by the scheme optimization and simulation module on a daily time scale.

5. A method for optimizing furniture boards based on artificial intelligence, characterized in that, Furniture board optimization is achieved using the AI-based furniture board optimization system described in any one of claims 1-4.

Images