Intelligent furniture board sorting system
By constructing probabilistic digital genome vectors through multimodal perception and high-dimensional confidence inference modules, and combining adaptive sorting decision-making and reinforcement learning, the perception and decision-making problems of automated furniture board sorting systems are solved, achieving high-precision, adaptive sorting and quality traceability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG CHENGLONG HOME IND CO LTD
- Filing Date
- 2025-12-19
- Publication Date
- 2026-06-19
AI Technical Summary
Existing automated sorting systems for furniture boards have a single perception dimension, making it difficult to identify three-dimensional geometric and internal structural defects. Their rigid decision-making logic leads to a high misjudgment rate, poor adaptability, and inability to meet production needs.
A probabilistic digital genome vector is constructed using a multimodal perception module. Combined with a high-dimensional confidence inference module and an adaptive sorting decision module, reinforcement learning and chain-based tracing mechanisms are introduced to form a multi-objective optimization decision and achieve adaptive sorting.
It significantly improves the accuracy of sheet material sorting, enhances the system's adaptability and resource utilization, reduces the misjudgment rate, and provides highly reliable quality traceability and production process stability.
Smart Images
Figure CN122244404A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of furniture board sorting technology, specifically to an intelligent furniture board sorting system. Background Technology
[0002] In the modern, large-scale furniture manufacturing industry, the quality of boards, as the most basic building block, directly determines the quality, durability, and aesthetics of the final product. Therefore, before the boards enter the finishing process (such as cutting, edge banding, and drilling), efficient and accurate automated sorting is crucial to removing defective products and classifying boards of different grades, ensuring production efficiency and product qualification rate.
[0003] Currently, automated sorting technology for furniture boards has replaced traditional manual visual inspection to some extent, but it still faces deep-seated technical bottlenecks. Existing automated systems mainly rely on machine vision technology based on two-dimensional industrial cameras. This technology can effectively identify some obvious surface defects such as scratches, stains, and color differences. However, this perception method is relatively one-dimensional, and its limitations are becoming increasingly apparent. For three-dimensional geometric defects such as flatness, warping, and uneven thickness of boards, two-dimensional vision is difficult to quantify and judge accurately. More importantly, for internal structural defects that affect the mechanical properties of boards and the stability of subsequent processing, such as internal micro-cracks, voids in the adhesive layer, and uneven density, traditional vision technology is completely powerless, which constitutes a huge quality hazard.
[0004] At the decision-making logic level, existing sorting systems generally adopt a judgment mechanism based on preset rules and fixed thresholds. When the production line changes to different batches, textures, or materials of boards, it is often necessary to recalibrate parameters and adjust rules in a tedious manner. In addition, this deterministic decision-making logic cannot effectively handle fuzzy, complex, or critical defects. This judgment method often leads to a high rate of false positives and false negatives, which not only wastes qualified boards but also allows some defective boards to flow into subsequent processes, thus failing to meet the needs of users. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides an intelligent sorting system for furniture boards, which solves the problems of low sorting accuracy and poor adaptability caused by the single perception dimension, rigid decision-making logic, and lack of feedback loop in existing intelligent sorting systems for furniture boards.
[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent sorting system for furniture boards, comprising: The multimodal sensing module is used to collect multimodal information on furniture boards during the transportation process and construct a probabilistic digital genome vector that characterizes the macroscopic and microscopic features of the board. A high-dimensional confidence inference module, connected to the multimodal perception module, is used to receive the probabilistic digital genome vector and, based on a preset probabilistic inference model, evaluate and derive the posterior probability distribution of the quality status of the board material. An adaptive sorting decision module, connected to the high-dimensional confidence inference module, is used to receive the posterior probability distribution of the quality status, generate sorting instructions, and make adaptive strategy adjustments based on preset optimization objectives and / or feedback data from downstream processes. The execution module, connected to the adaptive sorting decision module, is used to perform physical sorting operations on the board material according to the sorting instructions.
[0007] Preferably, the multimodal sensing module includes a heterogeneous sensor array, which includes at least one of a multispectral industrial camera, a 3D structured light scanner, and an acoustic or ultrasonic sensor array; and the multimodal sensing module constructs the probabilistic digital genome vector by performing spatiotemporal feature decoupling and deep latent space encoding on the multimodal information.
[0008] Preferably, the multimodal perception module is further used to acquire the internal microstructure information of the board material, and to model the internal microstructure information through a graph neural network to generate a part of the probabilistic digital genome vector.
[0009] Preferably, the high-dimensional confidence inference module employs a Bayesian network or a Gaussian process regression model to infer the uncertainty information contained in the probabilistic digital genome vector.
[0010] Preferably, the adaptive sorting decision module generates the sorting instructions based on a dynamic multi-objective optimization algorithm, wherein the optimization objectives include at least one of the following: qualified product throughput, sorting error rate, subsequent processing cost, and system energy consumption.
[0011] Preferably, the adaptive sorting decision module includes a reinforcement learning agent that uses feedback data from the downstream process as a reward signal and employs a reinforcement learning algorithm to update and optimize the sorting strategy.
[0012] Preferably, the system further includes a chain traceability module for associating the probabilistic digital genome vector with the sorting decision of the board and the performance of downstream processes to form an immutable traceability record and provide the feedback data to the adaptive sorting decision module.
[0013] Preferably, the chain traceability module is implemented based on blockchain or distributed ledger technology.
[0014] Preferably, the execution module includes at least one of a multi-axis industrial robot, a palletizing robot, and an autonomous mobile robot.
[0015] Preferably, the system further includes a digital twin module for constructing a real-time virtual model of the system to simulate and verify the sorting strategy generated by the adaptive sorting decision module and to perform predictive maintenance on the physical equipment of the execution module.
[0016] This invention provides an intelligent sorting system for furniture boards. It has the following beneficial effects: 1. This invention constructs a probabilistic "digital genome" vector that includes macroscopic surface features, geometric features, and internal microscopic structural features, and uses deep variational autoencoders (VAE) and graph neural networks (GNNs) for modeling. This fundamentally improves the depth and breadth of material state perception. Compared with traditional methods that rely solely on surface vision or single physical quantity detection, this system can form a comprehensive and three-dimensional understanding of the material, structure, and defects of the material, and quantify the uncertainty of the perception results. This significantly enhances the ability to identify composite defects, hidden defects, and microscopic defects that indicate future processing risks, thereby improving the sorting accuracy of the material.
[0017] 2. This invention analyzes probabilistic digital genomes using a high-dimensional confidence inference module and constructs sorting decisions as a dynamic multi-objective optimization problem. This makes the system decision no longer a simple classification based on deterministic rules, but an intelligent trade-off aimed at maximizing production efficiency. The system can comprehensively evaluate the direct and indirect costs that sorting errors may bring based on the posterior probability distribution of the board quality status, and find the optimal balance point among multiple objectives such as throughput, energy consumption, and subsequent process costs, making decisions that are more economically reasonable and risk-avoiding, thereby effectively improving resource utilization and the economy of the overall production process.
[0018] 3. This invention introduces reinforcement learning agents and combines real feedback data from downstream processes provided by the chain traceability module as key reward signals to achieve adaptive evolution and forward-looking optimization of sorting strategies. The system can learn from the entire life cycle history of each board and automatically discover the deep correlation between "digital genome" features and the final product qualification rate. This enables sorting decisions to transcend the limitations of the current workstation and predictively handle boards with potential processing risks. This realizes a paradigm shift from "correct sorting" to "ensuring the success of the final product" and endows the system with long-term intelligence for continuous self-improvement and response to process changes.
[0019] 4. This invention uses a chain-based traceability module based on blockchain or distributed ledger technology to create a unique and tamper-proof "digital identity file" for each board throughout its entire lifecycle. This mechanism not only provides highly reliable traceability evidence for quality issues down to the individual board, solving the problems of difficult quality traceability and ambiguous responsibility definition in traditional production, but also transforms massive amounts of data in the production process into structured and usable digital assets, providing a solid data foundation for process optimization, supply chain management, and continuous improvement of product quality.
[0020] 5. By constructing a digital twin model that is synchronized with the physical system in real time, this invention significantly improves the operational stability and maintainability of the entire intelligent sorting system. On the one hand, the digital twin provides a low-cost, risk-free simulation verification environment for the deployment of new sorting strategies, ensuring the safety and reliability of algorithm iteration. On the other hand, through continuous analysis of equipment operation data and machine learning modeling, the system can achieve predictive maintenance of key components and provide early warnings of potential faults, thereby minimizing unplanned downtime and ensuring the continuity and efficiency of the entire production process. Attached Figure Description
[0021] Figure 1 This is one of the system flowcharts of the present invention; Figure 2 This is the second system flowchart of the present invention. Detailed Implementation
[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] Please see the appendix Figure 1 - Appendix Figure 2 This invention provides an intelligent sorting system for furniture boards, comprising: A multimodal sensing module is used to collect multimodal information of furniture boards during the transportation process and construct a probabilistic digital genome vector characterizing the macroscopic and microscopic features of the boards. The multimodal sensing module includes a heterogeneous sensor array, which includes at least one of a multispectral industrial camera, a 3D structured light scanner, and an acoustic or ultrasonic sensor array. Furthermore, the multimodal sensing module constructs the probabilistic digital genome vector by decoupling the spatiotemporal features of the multimodal information and encoding it with deep latent space. The multimodal sensing module is also used to acquire the internal microstructure information of the boards and model the internal microstructure information through a graph neural network to generate a portion of the probabilistic digital genome vector. Specifically, the system sets up a sensing gantry above the automated conveyor line (such as a roller or belt conveyor). When the sheet material passes through, a heterogeneous sensor array deployed on the gantry synchronously starts collecting data. To ensure the accuracy of data fusion, the system first uses a visual inertial odometry (VIO) module to calculate the six-degree-of-freedom pose matrix of the sheet material in the world coordinate system in real time. All data collected by the sensors is accompanied by a precise timestamp. and the corresponding pose matrix To provide a benchmark for subsequent spatiotemporal alignment, the heterogeneous sensor array specifically includes: a multispectral industrial camera: used to capture two-dimensional images of the board material in multiple spectral bands such as visible light (RGB) and near-infrared (NIR). Different bands have different sensitivities to different materials or defects (such as glue spots, slight color differences), and multispectral information can provide surface details beyond the human eye; a high-precision 3D structured light scanner: reconstructs high-density three-dimensional point cloud data of the board surface by projecting a specifically coded grating pattern and collecting its deformation. This data is used to accurately measure the geometric dimensions of the board material, such as length, width, and thickness, and to detect geometric defects such as flatness, warping, chipping, and indentation; and an acoustic / ultrasonic sensor array: emitting acoustic waves or ultrasonic waves of a specific frequency to the board material in a non-contact manner and receiving its transmitted or reflected signals. By analyzing the signal attenuation, time delay, and spectral changes, it can non-destructively detect whether there are internal structural defects such as delamination, voids, and uneven density inside the board material. The spatiotemporally aligned multispectral image sequence, 3D point cloud sequence, and acoustic spectrum sequence in S1 are used to construct a high-dimensional data tensor T. The dimensions of this tensor T(x,y,t,λ,z,s,…) include spatial coordinates (x,y), time t, spectral band λ, depth information z, and acoustic properties s.
[0024] Spatiotemporal feature decoupling: To extract the latent, independent essential features from the data, a high-order tensor decomposition algorithm (such as Tucker decomposition) is used to process the tensor T. This decomposition approximates T as the n-modulus product of a core tensor G and a series of factor matrices U(n). The column vectors of the factor matrix U(n) form an orthogonal basis for each dimension (e.g., texture, geometry), representing the main patterns of the macroscopic features of the board material, thus achieving effective decoupling of different physical property features.
[0025] Deep Latent Space Encoding: The decoupled factor matrix U(n) set is input into a pre-trained deep variational autoencoder (VAE) network. The VAE maps high-dimensional features to a low-dimensional probabilistic latent space through an encoder network. Unlike traditional autoencoders that output a deterministic vector, the VAE encoder outputs Gaussian distributed parameters: a mean vector μmacro and a covariance matrix Σmacro. The mean vector represents the most likely condensed description of the macroscopic features of the board, while the covariance matrix quantifies the uncertainty or confidence of this description. Together, these two constitute the macroscopic "digital fingerprint" of the board. When the system is equipped with a terahertz (THz) or micro-focused X-ray computed tomography (micro-CT) module, it can acquire images of the microstructure inside the board. The system identifies key units in these microscopic images (such as fiber bundles, adhesive clusters, and tiny voids) as nodes and defines the physical adjacency or interaction relationships between them as edges, thereby constructing a graph structure G=(V,E).
[0026] Graph Neural Networks (GNNs) are used to model the graph. By iteratively passing messages and aggregating features on the graph, GNNs can learn the topological features and patterns of the microstructure, effectively identifying problems such as microcracks or uneven bonding that may affect the long-term stability of the board. The final output of the GNN is a vector representing the health status of the microstructure. and its confidence score Together, they constitute a microscopic "digital fingerprint," which in turn integrates with the macroscopic "digital fingerprint." With micro digital fingerprints ( Cascaded to form the final digital genome vector of the board. This vector is a comprehensive, multi-layered probabilistic description of the state of the sheet material, which inherently contains a quantification of uncertainty.
[0027] The high-dimensional confidence inference module, connected to the multimodal perception module, is used to receive probabilistic digital genome vectors and, based on a preset probabilistic inference model, evaluate and derive the posterior probability distribution of the board's quality status. The high-dimensional confidence inference module uses a Bayesian network or a Gaussian process regression model to infer the uncertainty information contained in the probabilistic digital genome vectors. Specifically, the system uses a pre-trained Bayesian network as the probabilistic inference model. The structure of this network is predefined based on expert knowledge and historical data, and its nodes include representative... Key feature variable nodes, and those representing the final quality category. Decision nodes (e.g., qualified products, repairable defective products, and scrap products) are defined by directed edges between nodes, representing their conditional dependencies. When a specific decision node is received... When a vector is input, the system uses it as evidence into the Bayesian network. By employing inference algorithms such as belief propagation, the network calculates the evidence given... Under the conditions, quality category Posterior probability distribution of different values This probabilistic output provides richer information for subsequent risk assessment and decision optimization compared to traditional deterministic classification. As an alternative, Gaussian process regression (GPR) models can also be used to predict continuous quality indicators and provide their uncertainty ranges.
[0028] An adaptive sorting decision module, connected to a high-dimensional confidence inference module, receives the posterior probability distribution of quality status, generates sorting instructions, and adjusts its adaptive strategy based on preset optimization objectives and / or feedback data from downstream processes. The adaptive sorting decision module generates sorting instructions based on a dynamic multi-objective optimization algorithm. Optimization objectives include at least one of the following: qualified product throughput, sorting error rate, subsequent processing cost, and system energy consumption. The adaptive sorting decision module includes a reinforcement learning agent that uses feedback data from downstream processes as a reward signal and employs a reinforcement learning algorithm to update and optimize the sorting strategy. The system also includes a chain-based traceability module, which associates probabilistic digital genome vectors with the sorting decisions of the board material and the performance of downstream processes to form an immutable traceability record and provides feedback data to the adaptive sorting decision module. The chain-based traceability module is implemented based on blockchain or distributed ledger technology. Specifically, the system considers multiple conflicting optimization objectives in real time, such as: (Maximize the throughput of qualified products per unit time) (Minimize based on) Calculated sorting error cost) (Minimize the expected subsequent repair costs caused by sorting to a specific defective line) (Minimize system energy consumption caused by robot movement and conveyor line start / stop). Then, an evolutionary multi-objective optimization algorithm is used. When each piece of material reaches the decision point, the algorithm is adjusted based on the current system state (e.g., the capacity of each exit buffer) and the material's... The system solves this optimization problem in real time, generating a set of Pareto-optimal sorting strategies that achieve different trade-offs among the objectives. To achieve long-term evolution of the decision-making strategies, a reinforcement learning (RL) agent is introduced. This agent's role is to select the currently optimal strategy from the Pareto-optimal policy set based on learned experience. The reward function of the RL agent... This is key to its learning, and the reward consists of two parts: one is an immediate reward based on the results of multi-objective optimization; the other, more important, is a delayed reward from S6 feedback. This delayed reward reflects the actual performance of boards with specific "digital genomes" in downstream processes after sorting (such as whether it leads to processing failure or final product downgrade). Through algorithms such as Q-Learning or PolicyGradient, the RL agent continuously updates its policy network based on the cumulative rewards. This allows the system to learn deep-seated correlations, such as "a microscopic defect that seems acceptable during sorting is likely to cause edge breakage in the subsequent edge banding process." This allows the system to adjust the sorting strategy in advance, diverting such boards to the pre-processing or manual re-inspection line, achieving true predictive and adaptive decision-making.
[0029] The execution module, connected to the adaptive sorting decision module, is used to perform physical sorting operations on the boards according to the sorting instructions. The execution module includes at least one of the following: multi-axis industrial robot, palletizing robot, and autonomous mobile robot. Specifically, the central controller schedules the execution module based on the output optimal sorting instructions. This module typically consists of a multi-axis industrial robot, a palletizing robot, and... The system comprises components that work together to precisely grasp, gently place, intelligently palletize, and automatically transfer the boards. Simultaneously with the sorting process, a chain traceability module is activated, which records the board's unique ID and complete digital genome. The sorting decision details, execution timestamp, and a reserved downstream feedback data field are packaged into a data block. Based on blockchain or distributed ledger technology (DLT), this data block is linked to the existing historical data chain of the board through cryptographic hash operation, forming an immutable and traceable full life cycle "digital traceability chain".
[0030] The system also includes a digital twin module, which is used to build a real-time virtual model of the system. This module can simulate and verify the sorting strategies generated by the adaptive sorting decision module and perform predictive maintenance on the physical equipment of the execution module. Specifically, after the board material is processed in downstream processes (such as CNC machining, edge banding, and coating), the intelligent equipment or quality inspection systems in these processes generate data on the processing performance of the board material (such as processing accuracy, edge banding quality, and yield). This feedback data from downstream processes is transmitted back through the Industrial Internet of Things (IIoT) and updated to the corresponding data block of the board material in the chain traceability system. The system automatically extracts this feedback data and quantifies it into the delayed reward signal required by the RL agent. This forms a complete closed loop of "perception-decision-execution-feedback-relearning," driving the sorting strategy to continuously optimize towards improving the efficiency of the entire process. The digital twin module can be used to conduct large-scale, low-cost simulation testing and verification of the new strategies learned by the RL agent in a virtual environment, ensuring the safety and effectiveness of the new strategies before deploying them to the physical system. The digital twin model receives and analyzes the operating data of each physical device in the execution module (such as robot joint motor current, servo driver temperature, and conveyor line vibration) in real time. Through time-series data prediction models such as Long Short-Term Memory Network (LSTM), the system can predict the remaining useful life (RUL) and potential failure risks of key components of the equipment, generate maintenance work orders in advance, realize the transformation from passive response to predictive maintenance, and maximize the stable operating time of the system.
[0031] In summary, this invention provides an intelligent sorting system for furniture boards. By constructing a probabilistic "digital genome" vector that includes macroscopic surface features, geometric features, and internal microscopic structural features, and by using deep variational autoencoders (VAEs) and graph neural networks (GNNs) for modeling, the system fundamentally improves the depth and breadth of board state perception. Compared with traditional methods that rely solely on surface vision or single physical quantity detection, this system can form a comprehensive and three-dimensional understanding of the board material, structure, and defects, thereby improving the sorting accuracy of the boards.
[0032] 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 smart sorting system for furniture boards, characterized in that, include: The multimodal sensing module is used to collect multimodal information on furniture boards during the transportation process and construct a probabilistic digital genome vector that characterizes the macroscopic and microscopic features of the board. A high-dimensional confidence inference module, connected to the multimodal perception module, is used to receive the probabilistic digital genome vector and, based on a preset probabilistic inference model, evaluate and derive the posterior probability distribution of the quality status of the board material. An adaptive sorting decision module, connected to the high-dimensional confidence inference module, is used to receive the posterior probability distribution of the quality status, generate sorting instructions, and make adaptive strategy adjustments based on preset optimization objectives and / or feedback data from downstream processes. The execution module, connected to the adaptive sorting decision module, is used to perform physical sorting operations on the board material according to the sorting instructions.
2. The intelligent sorting system for furniture boards according to claim 1, characterized in that, The multimodal sensing module includes a heterogeneous sensor array, which includes at least one of a multispectral industrial camera, a 3D structured light scanner, and an acoustic or ultrasonic sensor array; and the multimodal sensing module constructs the probabilistic digital genome vector by performing spatiotemporal feature decoupling and deep latent space encoding on the multimodal information.
3. The intelligent sorting system for furniture boards according to claim 2, characterized in that, The multimodal perception module is also used to acquire the internal microstructure information of the board material, and to model the internal microstructure information through a graph neural network to generate a part of the probabilistic digital genome vector.
4. The intelligent sorting system for furniture boards according to claim 1, characterized in that, The high-dimensional confidence inference module employs a Bayesian network or a Gaussian process regression model to infer the uncertainty information contained in the probabilistic digital genome vector.
5. The intelligent sorting system for furniture boards according to claim 1, characterized in that, The adaptive sorting decision module generates the sorting instructions based on a dynamic multi-objective optimization algorithm. The optimization objectives include at least one of the following: qualified product throughput, sorting error rate, subsequent processing cost, and system energy consumption.
6. The intelligent sorting system for furniture boards according to claim 1, characterized in that, The adaptive sorting decision module includes a reinforcement learning agent that uses feedback data from the downstream process as a reward signal and employs a reinforcement learning algorithm to update and optimize the sorting strategy.
7. The intelligent sorting system for furniture boards according to claim 6, characterized in that, The system also includes a chain traceability module, which is used to associate the probabilistic digital genome vector with the sorting decision of the board and the performance of downstream processes to form an immutable traceability record and provide the feedback data to the adaptive sorting decision module.
8. The intelligent sorting system for furniture boards according to claim 7, characterized in that, The chain-based traceability module is implemented based on blockchain or distributed ledger technology.
9. The intelligent sorting system for furniture boards according to claim 1, characterized in that, The execution module includes at least one of the following: a multi-axis industrial robot, a palletizing robot, and an autonomous mobile robot.
10. The intelligent sorting system for furniture boards according to claim 1, characterized in that, The system also includes a digital twin module for constructing a real-time virtual model of the system to simulate and verify the sorting strategy generated by the adaptive sorting decision module and to perform predictive maintenance on the physical equipment of the execution module.