Wood cutting control method based on artificial intelligence

Abstract

The invention discloses a wood cutting control method based on artificial intelligence, and belongs to the technical field of wood cutting control. The method comprises the following steps: acquiring a high-resolution surface image, X-ray tomography data and sound wave propagation characteristic data of wood to be cut; performing distributed feature extraction on the multi-dimensional detection data through an edge computing node cluster to generate a primary feature set; inputting the primary feature set into a federated learning framework for cross-node feature fusion, and generating global feature representation; based on global feature representation, a multi-objective optimization cutting scheme including a cutting path, a cutting depth, a feeding speed and a spindle rotating speed is generated through a hierarchical decision model; and the scheme is dynamically verified and adjusted, a final cutting control instruction is generated, and the instruction correspondingly controls a main shaft driving part, a feeding transmission part and a saw blade posture adjusting part of the wood cutting saw.

Description

Technical Field

[0001] This invention relates to the field of wood cutting control technology, specifically to a wood cutting control method based on artificial intelligence. Background Technology

[0002] In the timber processing industry, timber cutting saws are generally used to cut timber. The cutting process directly affects the utilization rate of timber resources, processing efficiency, and the quality of the final product. However, traditional timber cutting saws mostly rely on manual experience for operation. Operators judge the cutting position and path based on their own observation of the appearance of the wood. This method has obvious limitations. Since manual observation cannot deeply understand the internal structure of the wood, such as the grain direction, whether there are defects such as insect holes or cracks, it often leads to unreasonable cutting plans, which can easily waste timber resources. At the same time, it is also difficult to guarantee the dimensional accuracy and surface quality of the cut wood. With the development of technology, some wood processing enterprises have begun to introduce some detection-based wood cutting saws to assist in cutting control. For example, they use ordinary cameras to acquire images of the wood surface and use simple image analysis to help determine the cutting position. However, this method can only obtain limited information about the wood surface and cannot accurately grasp the internal structural characteristics of the wood. It is difficult to identify hidden defects inside the wood, which leads to a significant degree of uncertainty in the cutting plan. In traditional wood cutting control, data processing is often concentrated on a single wood saw or a single node. When processing large amounts of wood inspection data, the data processing capacity of a single node is limited, easily leading to data processing delays. This prevents timely and accurate control commands from being provided to the wood saw, affecting cutting efficiency. Furthermore, data from different processing stages or different wood saws is difficult to share and integrate effectively. Data from each stage remains relatively isolated, failing to fully realize its value and hindering the formation of a globally optimal cutting scheme. When generating cutting schemes, planning can only focus on a single objective, such as yield, without fully considering the performance characteristics of the wood saw itself, such as the power of the spindle drive motor, the speed range of the feed drive motor, and the adjustment precision of the saw blade posture adjustment cylinder.

[0003] Traditional cutting schemes typically consider only a single objective, such as maximizing wood utilization or focusing solely on cutting efficiency, while neglecting other crucial factors. They fail to optimize parameters based on the characteristics of different types of saws—such as the saw blade tension in band saws, the matching relationship between the saw blade feed rate and spindle speed in circular saws, and the synchronicity of saw blades in multi-blade collaborative cutting systems. This single-objective approach easily leads to saw blade breakage in band saws, overheating in circular saws, and saw blade collisions in multi-blade systems, increasing equipment maintenance costs. Furthermore, it overlooks factors such as the quality of the cut wood and the energy consumption of the wood cutting saw. Such single-objective cutting schemes cannot meet the multifaceted performance requirements of the modern wood processing industry, hindering sustainable development and market competitiveness for enterprises. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based wood cutting control method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an artificial intelligence-based wood cutting control method, applied in a wood cutting control system. The wood cutting control system includes a wood cutting saw, which is equipped with a spindle drive motor, a feed drive motor, and a saw blade attitude adjustment cylinder. The method includes the following steps: Acquire multi-dimensional detection data of the wood to be cut, including high-resolution surface images, X-ray tomography data, and acoustic wave propagation characteristic data; Distributed feature extraction is performed on the multi-dimensional detection data through an edge computing node cluster to generate a primary feature set. The primary feature set is input into the federated learning framework for cross-node feature fusion to generate a global feature representation. Based on the global feature representation, a multi-objective optimized cutting scheme is generated through a hierarchical decision model. The scheme includes the cutting path, cutting depth, feed rate, and spindle speed of the wood cutting saw. The multi-objective optimized cutting scheme is dynamically verified and adjusted to generate the final cutting control command, which controls the spindle drive motor, feed transmission motor and saw blade posture adjustment cylinder of the wood cutting saw.

[0006] Preferably, the distributed feature extraction of the multi-dimensional detection data through an edge computing node cluster includes: Multi-scale convolution operations are performed on the high-resolution surface image to extract surface texture feature maps, which are used to determine the initial contact angle of the saw blade of the wood cutting saw. The X-ray tomography data is divided into three-dimensional meshes to construct a voxelized structural model and mark the spatial coordinates of internal defects in the wood to plan cutting avoidance paths. The sound wave propagation characteristic data is spatiotemporally aligned with the voxelized structural model to generate a material property distribution map. The surface texture feature map and the material property distribution map are integrated by a feature coding network to output a primary feature set.

[0007] Preferably, the primary feature set is input into the federated learning framework for cross-node feature fusion, including: In each edge computing node, a local feature extraction model is used to process the primary feature set; The feature extraction results from multiple nodes are verified and weighted using a blockchain consensus mechanism; The verified feature representations are integrated using a federated aggregation algorithm to generate the global feature representation.

[0008] Preferably, a multi-objective optimized cutting scheme is generated based on the global feature representation through a hierarchical decision model, including: The global feature representation is input into a multi-layer decision network that incorporates an attention mechanism; The optimal cutting direction sequence is determined through the first decision-making layer; The second decision layer calculates the combination of saw blade parameters for each cutting segment; The multi-objective optimized cutting scheme, which includes timing control instructions, is generated through the third decision layer.

[0009] Preferably, the multi-objective optimized cutting scheme is dynamically verified and adjusted, including: The multi-objective optimization cutting scheme is distributed to the verification node cluster. Each verification node performs a feasibility assessment of the cutting scheme based on its local knowledge base, collects the assessment results, calculates the scheme correction parameters, and iteratively optimizes the cutting scheme according to the scheme correction parameters.

[0010] Preferably, the method further includes: Real-time acquisition of multimodal sensing data during the cutting process, including cutting force waveform data, saw blade vibration spectrum and infrared thermal imaging data; The multimodal sensing data is input into the anomaly detection network for real-time analysis; When an abnormal pattern is detected, an online adjustment process is triggered.

[0011] Preferably, the process of triggering online adjustment includes: Analyze the type and impact range of abnormal modes, calculate the adjustment range of the current cutting parameters, and generate a compensation control signal containing real-time adjustment instructions.

[0012] Preferably, the method further includes: After the cutting process is completed, comprehensive quality inspection data of the finished wood is collected. The comprehensive quality inspection data is compared and analyzed with the digital twin model, and a model optimization dataset is generated based on the analysis results.

[0013] Preferably, the step of comparing and analyzing the comprehensive quality inspection data with the digital twin model includes: Extract the geometric precision characteristics, surface roughness distribution, and internal defect characteristics of the finished wood; Calculate the deviation matrix between the actual processing results and the expected target, evaluate the compliance level of each quality indicator, and generate a multi-dimensional quality analysis report.

[0014] Preferably, the step of generating a model-optimized dataset based on the analysis results includes: The multidimensional quality analysis report is converted into feature-label pair training data; The parameters of the local feature extraction model and the hierarchical decision model are updated using a federated learning framework, and the updated model parameters are synchronized to each edge computing node through a blockchain network.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This AI-based wood cutting control method acquires multi-dimensional detection data, including high-resolution surface images, X-ray tomography data, and acoustic wave propagation characteristics, to comprehensively understand the surface features and internal structure of the wood. This includes surface texture and defects, as well as internal texture direction and latent defects. Compared to traditional methods relying solely on limited surface information or human experience, this method provides a more accurate understanding of the wood's actual condition, offering a comprehensive and reliable basis for subsequent cutting plans and reducing cutting errors caused by insufficient understanding of the wood's condition. It can output customized data for different types of saws—providing data on the wood texture angle and surface knot distribution at the saw blade contact point for band saws to assist in precise adjustment of the saw blade guide mechanism; providing data on internal wood density changes for circular saws to optimize the saw blade cutting path; and providing defect coordinates for corresponding areas of each saw blade for multi-blade collaborative cutting equipment to support dynamic calibration of the saw blade spacing.

[0016] By leveraging edge computing node clusters to perform distributed feature extraction on multi-dimensional detection data, the computing resources of multiple nodes can be fully utilized to accelerate data processing speed, avoid latency issues that occur when a single node processes large amounts of data, and generate a preliminary feature set in a timely manner. This ensures the smooth progress of subsequent cutting-related processes, improves the overall cutting control response efficiency, and meets the real-time requirements of wood processing. By inputting the initial feature set into the federated learning framework for cross-node feature fusion to generate a global feature representation, the isolation of data from different nodes can be broken, enabling effective sharing and fusion of data from each node. This fully explores the value of data from each node, allowing the generated global feature representation to more comprehensively and accurately reflect the overall characteristics of the timber. It avoids feature bias caused by data limitations, provides a better data foundation for subsequent hierarchical decision-making models to formulate cutting schemes, and helps to formulate more reasonable cutting strategies from a global perspective. Based on global feature representation, a hierarchical decision model is used to generate multi-objective optimized cutting schemes. This model can comprehensively consider multiple objectives in the wood processing process, such as wood utilization, cutting efficiency, wood quality, and wood cutting saw energy consumption. It is no longer limited to the traditional single-objective scheme formulation mode, but can achieve a balance among multiple objectives. This meets the multi-faceted performance requirements of the modern wood processing industry, helps enterprises reduce processing costs and enhance their overall competitiveness while ensuring wood processing quality and efficiency.

[0017] The system dynamically verifies and adjusts multi-objective optimized cutting schemes to generate final cutting control commands. This allows for the verification of the feasibility and rationality of the cutting scheme before implementation, and timely adjustments based on the verification results. This prevents problems in the cutting process caused by flawed schemes, ensuring that the final generated cutting control commands are accurate and reliable. This guarantees that the wood cutting saw operates according to the optimal scheme, improves the quality stability of the cut wood, reduces waste of wood resources, and also helps reduce wear and tear on the wood cutting saw caused by executing unreasonable schemes, thus extending the service life of the wood cutting saw. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based wood cutting control method described in this invention. Figure 2 A flowchart for distributed feature extraction of an edge computing node cluster; Figure 3 A schematic diagram of the overall structure of a wood cutting saw; Figure 4 A flowchart for generating a multi-objective optimized cutting scheme for global feature representation.

[0019] In the diagram: 1. Spindle drive motor; 2. Feed drive motor; 3. Saw blade posture adjustment cylinder. Detailed Implementation

[0020] The technical solutions of 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.

[0021] Please see Figure 1 This invention provides an artificial intelligence-based wood cutting control method, comprising: fusing multi-dimensional sensor data with distributed artificial intelligence computing to achieve precise modeling and optimized control of the wood cutting process. The method first acquires multi-dimensional detection data consisting of high-resolution surface images, X-ray tomographic data, and acoustic wave propagation characteristic data of the wood to be cut. This data is synchronously acquired by a high-speed sensor array deployed at the acquisition terminal and transmitted to an edge computing node cluster. The edge computing node cluster performs distributed preprocessing and feature extraction on the raw data, generating a primary feature set containing information on surface texture, internal structure, and material properties. This feature set is input into a federated learning framework, where secure aggregation and fusion computation across nodes generates a global feature representation that comprehensively reflects the physical properties of the wood. Based on this global feature representation, a hierarchical decision model performs multi-objective optimization calculations to generate a cutting scheme that simultaneously considers yield, cutting efficiency, and saw blade wear. This scheme is then evaluated and its parameters are adjusted through dynamic verification nodes, ultimately generating a sequence of control commands to drive the wood cutting saw.

[0022] Example 1: See Figure 2The distributed feature extraction and cross-node feature fusion process for multi-dimensional detection data is completed collaboratively by an edge computing node cluster and a federated learning framework. High-resolution surface image acquisition in the multi-dimensional detection data is achieved using a linear CCD camera paired with a ring-shaped polarized light source array. The camera resolution is set to 4096×2160 pixels, and the polarized light source is incident at a 45-degree angle to eliminate surface reflection interference. The image acquisition frame rate is synchronously controlled with the conveyor belt speed to ensure that at least three sets of surface image data from different angles are obtained for each piece of wood. When the raw images are transmitted to the edge computing node cluster, a lossless compression algorithm is used to reduce bandwidth consumption. The image processing nodes are equipped with dedicated GPU computing cards, and their multi-scale convolutional operations are implemented through a parallel convolutional neural network architecture. This network contains three parallel convolutional branches, using 3×3, 5×5, and 7×7 convolutional kernels for feature extraction, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The extracted feature maps are fused at multiple scales through a spatial pyramid pooling layer, ultimately generating a 256-channel surface texture feature map. When performing multi-scale convolution operations on high-resolution surface images to extract surface texture feature maps, different convolution processing strategies are adapted to the saw blade type of the wood cutting saw. For band saws, the focus is on extracting detailed features of the edge texture of knots and the direction of annual rings on the wood surface. Multi-scale convolution kernels capture the contour information of knots of different sizes. Based on these feature maps, the optimal initial contact angle between the band saw blade and the wood surface is determined to avoid sawing deviation or saw jamming at knots due to improper contact angles. For circular saws, in addition to extracting the surface texture distribution, the density and direction of the texture are analyzed to determine the initial entry angle of the circular saw blade, ensuring that the saw blade can cut efficiently along the wood grain direction during the cutting process, reducing tearing of the wood surface. Figure 3 As shown, the wood cutting saw mainly includes a spindle drive motor 1, a feed transmission motor 2, and a saw blade attitude adjustment cylinder 3. The cutting mechanism of this wood cutting saw has a power-driven sawing component, which can perform precise cutting operations on wood according to the planned cutting avoidance path and the determined initial contact angle of the saw blade. The various parts of the structure work together to ensure the efficiency and precision of the wood cutting process.

[0023] X-ray computed tomography (CT) data acquisition was achieved using a scanning system consisting of a mobile X-ray source and a flat panel detector. The X-ray tube voltage was set to 150 kV, the current to 5 mA, and the detector pixel size to 0.1 mm. During the scanning process, the wood passed through the scanning area at a constant speed, with the slice thickness accuracy controlled within 0.5 mm. The acquired two-dimensional projection sequence was reconstructed into three-dimensional volume data using a filtered back-projection algorithm. The reconstructed volume data was then transmitted to a three-dimensional processing node in an edge computing node cluster. This node used an octree structure for three-dimensional mesh generation, with a minimum cell size of 0.2 mm × 0.2 mm × 0.2 mm. Each voxel stored a grayscale value and the calculated material density value. A trilinear interpolation algorithm was used during voxelization to ensure data continuity. The final voxelized structural model contained the spatial coordinates of the material's internal density distribution and potential defect areas. When generating a three-dimensional mesh and constructing a voxelized structural model from the X-ray computed tomography data, the mesh generation accuracy was set according to the processing range and cutting accuracy requirements of different wood cutting saws. For multi-blade collaborative cutting equipment, since multiple saw blades need to operate simultaneously, the three-dimensional mesh division will be more refined to accurately mark the spatial coordinates of defects inside the wood, such as insect holes and cracks. This clarifies the specific location and extension range of each defect in the three-dimensional space of the wood. Based on this, a dedicated cutting avoidance path is planned for each saw blade to prevent multiple saw blades from simultaneously touching the same defect area due to untimely avoidance during the cutting process. For single-blade band saws or circular saws, the mesh density will be reasonably set based on the cutting stroke and avoidance flexibility of the equipment, while ensuring that the spatial coordinates of the main internal defects are clearly marked. The overall cutting avoidance path of the saw blade is planned according to the defect coordinates to ensure that the saw blade avoids all internal defects throughout the cutting process, thereby improving the integrity of the finished wood.

[0024] Acoustic wave propagation characteristic data were acquired using a 32-element ultrasonic transducer array with a center frequency of 100kHz. The array was mounted in a cross-shaped layout below the scanning platform, with a transmit pulse width of 10μs and a receive sampling rate of 2MHz. Acoustic data preprocessing included signal filtering, gain compensation, and temporal alignment. The sound wave propagation time was calculated using a cross-correlation algorithm, and the propagation velocity field and attenuation coefficient field of the sound wave within the wood were calculated based on the array's geometric parameters. Signal processing nodes in the edge computing node cluster performed spatiotemporal alignment of the acoustic data with the voxelized structural model. The alignment process employed a feature-point-based registration algorithm. First, density abrupt change points in the voxel model were extracted as spatial feature points, and simultaneously, velocity abrupt change points in the acoustic wave data were extracted as temporal feature points. The transformation matrix between spatial coordinates and the time series was calculated using the least squares method, accurately mapping the acoustic characteristic parameters to three-dimensional mesh nodes and generating a material property distribution map containing the elastic modulus and damping coefficient.

[0025] The feature encoding network employs a variational autoencoder with a skip connection structure. The encoder consists of five fully connected layers, with 512, 256, 128, 64, and 32 neurons per layer, respectively. The decoder uses a symmetric structure. Surface texture feature maps and material property distribution maps are first converted into one-dimensional vectors through a flattening layer. After layer normalization, these vectors are input into the encoder network. The latent space dimension is set to 256. During training, a weighted sum of the KL divergence loss function and the reconstruction loss function is used as the optimization objective. The final output primary feature vector contains comprehensive information on the surface and interior properties of the wood. The primary feature vectors generated by each edge node are uploaded to the data aggregation node through an encrypted channel, forming a standardized primary feature set.

[0026] Cross-node fusion of the initial feature set is achieved through a federated learning framework. Each edge computing node maintains a training copy of its local feature extraction model, which is a three-layer fully connected neural network with 128 and 64 neurons in the hidden layers, respectively. Training employs a stochastic gradient descent algorithm with a learning rate of 0.01 and a batch size of 32. The federated learning framework maintains a distributed consensus mechanism through a blockchain network. Each node records the hash values ​​of its locally trained model parameters on the blockchain and verifies them using a practical Byzantine fault-tolerant algorithm. Nodes that pass verification receive weights based on their data quality scores, with weight calculations taking into account historical accuracy, data volume, and network latency. The federated aggregation algorithm uses a weighted average strategy, with the weight allocation formula dynamically adjusted based on node contributions. The aggregated global model parameters are protected against noise using differential privacy technology to ensure data security. Each node processes its local initial feature set using the updated global model. The output feature representation is uploaded to the federated server for principal component analysis. By retaining 95% of the variance ratio in the feature dimension, the data is compressed to 128 dimensions. The final global feature representation contains cross-node fusion information, providing a unified data representation for subsequent decision-making. The entire process utilizes a distributed computing framework for task scheduling, with a heartbeat mechanism between nodes to maintain communication synchronization, and a faulty node automatically triggering a data retransmission mechanism to ensure continuous processing.

[0027] Example 2: See Figure 4This paper describes the construction of a hierarchical decision-making model and the generation and verification of a multi-objective optimization segmentation scheme. The global feature representation serves as the input data for the hierarchical decision-making model, which employs a multi-layer feedforward neural network architecture incorporating an attention mechanism. The number of neurons in the input layer is consistent with the dimension of the global feature representation, i.e., 128. The input data is first processed by layer normalization to eliminate dimensional differences. The attention mechanism module uses a scaled dot product attention calculation method, calculating feature weights through linear transformations of the query vector, key vector, and value vector. The query vector is generated from the global feature representation through a fully connected layer, while the key and value vectors are extracted from the same input by different fully connected layers. The temperature coefficient of the attention weights is set to 0.1, and the dimension of the output weighted feature vector remains unchanged.

[0028] The first decision layer receives the weighted feature vectors and processes them through a fully connected layer with a softmax activation function. The output dimension of this layer corresponds to all possible combinations of cutting direction sequences. The cutting direction is defined in polar coordinates, including two parameters: radial angle θ and axial angle φ, with an angle resolution of 1 degree. The decision logic is based on a Markov decision process, calculating the expected return value of each direction sequence through a value iteration algorithm. The return function comprehensively considers material utilization, cutting time, and saw blade load factors, and finally selects the cutting direction sequence with the highest expected return as the output. This sequence includes the coordinates of the cutting start point, the cutting direction angles of each segment, and the cutting depth information.

[0029] The second decision layer, based on the cutting direction sequence output by the first decision layer, calculates the saw blade parameter combinations for each cutting segment using a regression neural network. The regression network employs a three-layer fully connected structure with 64 and 32 neurons in the hidden layers, and the output layer contains parameter values ​​in three dimensions. The feed rate parameter is dynamically adjusted based on wood density and cutting length, taking into account the saw blade's ultimate cutting capability. The spindle speed parameter is determined based on material hardness and cutting depth, using an adaptive control algorithm to maintain stable cutting force. The cutting angle parameter is optimized based on the angle between the wood grain direction and the cutting direction, with an angle adjustment accuracy of 0.1 degrees. All parameter calculations utilize constrained optimization methods to ensure that the parameter combinations are within the working range of the wood cutting saw and meet processing quality requirements. The third decision layer integrates the output information from the first two layers and generates an instruction sequence containing execution time points through a timing control module. The timing arrangement is based on critical path analysis, considering the dependencies between cutting segments and the kinematic constraints of the wood cutting saw. The instruction sequence uses a standardized encoding format, with each instruction containing a timestamp, motion parameters, and execution conditions. The timestamp accuracy is at the millisecond level, the motion parameters include coordinate position, velocity, and acceleration information, and the execution conditions include sensor triggering conditions and safety verification conditions. The generated multi-objective optimized cutting scheme is encapsulated in JSON format, containing complete processing parameters and timing control information.

[0030] The multi-objective optimized cutting scheme is distributed to the verification node cluster via a message middleware. The verification node cluster consists of three heterogeneous nodes, each configured with a different professional knowledge base. The wood mechanical properties verification node stores elastic modulus, bending strength, and failure strain data for various wood materials. This node uses finite element simulation to calculate stress distribution during the cutting process and assesses whether the scheme has a risk of stress exceeding limits. The saw blade performance verification node contains a saw blade life model and a cutting force database. It assesses saw blade wear by calculating the saw blade load spectrum and predicts the remaining service life of the saw blade. The historical scheme verification node stores a library of past wood cutting saw processing cases. It uses similarity matching to find similar processing scenarios and compares the parameter settings in historical schemes.

[0031] Each verification node independently evaluates the cutting scheme based on its local knowledge base. The evaluation results adopt a unified risk scoring format, with a score range of 0-100, where a lower score indicates lower risk. Evaluation results are collected to the verification coordinator via a publish-subscribe model. The coordinator integrates the evaluation results using a weighted voting mechanism, with weights dynamically adjusted based on the historical evaluation accuracy of each node. When the overall risk score exceeds a threshold, the scheme correction module initiates a genetic algorithm for parameter optimization. The algorithm population size is set to 50, the number of iterations is 100 generations, the mutation probability is 0.01, and the crossover probability is 0.8. The corrected scheme is redistributed to the verification nodes for a second evaluation until the overall risk score falls below the threshold. The final generated cutting control command is transmitted to the wood cutting saw controller via a secure protocol. The entire verification process is executed asynchronously and in parallel, with a timeout control mechanism ensuring the evaluation process is completed within a specified time.

[0032] Taking the processing of a 2000mm×300mm×60mm red pine board as an example, the wood has a knot hole with a diameter of approximately 40mm at 700mm from one end, and a slight longitudinal bend in the 1200mm to 1500mm section. The global feature representation data contains a feature vector with 128 dimensions. Dimensions 23 to 35 are significantly higher than the average, corresponding to internal density anomalies; dimensions 67 to 72 have lower values, reflecting material stiffness characteristics; dimensions 89 to 94 show fluctuations in specific patterns, associated with surface texture features. This feature vector is input into a hierarchical decision model for processing. The model employs a three-layer fully connected neural network architecture. The input layer contains 128 neurons, and the output layer has variable output dimensions depending on the decision task. After receiving the weighted feature vector, the first decision layer of the hierarchical decision model calculates that the 27th, 69th, and 91st dimensions have the highest weight coefficients. These dimensions correspond to the knot hole region, the bend section, and the texture direction features, respectively. Based on weighted feature representation, the first decision layer calculates the cutting direction sequence using a softmax classifier, outputting a scheme containing five cutting segments: the first segment, from the starting end to 600mm, uses longitudinal cutting; the second segment, the 600-800mm section of the joint area, uses radial cutting; the third segment, from 800-1200mm, resumes longitudinal cutting; the fourth segment, the 1200-1500mm curved section, uses oblique cutting at 45 degrees; and the fifth segment, from 1500mm to the end, uses longitudinal cutting. Each cutting segment is labeled with its starting coordinates, cutting direction, and depth parameters.

[0033] The second decision layer calculates the saw blade parameters for each segment using a regression neural network based on the cutting sequence output by the first decision layer. For the knot area segment, the feed speed is reduced from the usual 1200 mm / min to 800 mm / min, the spindle speed is increased from 8000 rpm to 10000 rpm, and the cutting angle is adjusted to 90 degrees perpendicular to the grain direction. For the curved segment, the feed speed is set to 1000 mm / min, the spindle speed is maintained at 8000 rpm, and a 45-degree bevel cutting angle is used. The remaining regular segments use standard parameters: feed speed 1200 mm / min, spindle speed 8000 rpm, and cutting angle 0 degrees. All parameters take into account the saw blade's withstand limit, ensuring that the cutting force does not exceed 200 N and the cutting temperature is controlled below 80℃. The third decision layer integrates the output information from the first two layers to generate a cutting scheme that includes time control instructions. The timing arrangement is based on the kinematic characteristics of the wood cutting saw, with a 50 mm gap reserved between the acceleration and deceleration phases, a 30% reduction in the cutting speed in the knot area, and continuous feed to avoid interruptions in the curved segment. The entire scheme contains 28 control commands, each command including a timestamp, coordinate position, movement speed and execution conditions. The timestamp accuracy reaches the millisecond level, the coordinate positioning accuracy is 0.1mm, and the speed control accuracy is 1mm / min.

[0034] The multi-objective optimized cutting scheme was distributed to a cluster of verification nodes for dynamic verification. The wood mechanical properties verification node, through finite element analysis, found that the cutting force in the knot area could exceed 220N under normal parameters, posing a risk of saw blade breakage. The saw blade performance verification node, based on historical data, calculated that saw blade life would decrease by 15% at a speed of 10,000 rpm. The historical scheme verification node retrieved similar cases, suggesting that the beveling angle be adjusted to 40 degrees to improve surface quality. The verification coordinator collected the evaluation results from the three nodes, with a comprehensive risk score of 65 points, exceeding the threshold of 50 points. The scheme correction module initiated genetic algorithm optimization, obtaining improved parameters after 50 iterations: the speed in the knot area was adjusted to 9500 rpm, and the feed rate to 850 mm / min; the beveling angle in the bending section was changed to 40 degrees; and the coolant flow rate in the knot area was increased by 20%. The corrected scheme was re-verified, and the comprehensive risk score decreased to 45 points, reaching an acceptable range. The final generated cutting control commands are transmitted to the CNC controller of the wood cutting saw via the OPCUA protocol. These commands include all motion parameters, cutting parameters, and auxiliary function commands. The command format conforms to the ISO6983 standard. The entire verification and adjustment process is completed within 8 seconds. Specifically, the cutting control command precisely matches the control logic of each part with the cutting scheme parameters: the spindle drive motor 1 receives the spindle speed parameters of the corresponding cutting segment in the command (e.g., 9500 rpm in the joint area and 8000 rpm in the normal section), and adjusts the output power of the spindle drive motor 1 through the frequency converter driver to ensure that the speed fluctuation error during the cutting process does not exceed ±50 rpm; the feed drive motor 2, based on the feed speed in the command (e.g., 1000 mm / min in the bending section and 850 mm / min in the joint area) and the path trajectory coordinates, drives the ball screw of the wood cutting saw table through the servo motor to achieve linear movement, with the position positioning accuracy controlled within 0.1 mm and the speed adjustment response time less than 100 ms; the saw blade posture adjustment cylinder 3, based on the cutting angle parameters in the command (e.g., 40 degrees in the bending section and 0 degrees in the normal section), adjusts the spatial posture of the wood cutting saw blade through the multi-axis linkage mechanism, with the angle adjustment accuracy reaching 0.1 degrees, and achieves a smooth transition of posture when switching cutting segments to avoid mechanical impact. Control signal transmission for all components is achieved via a real-time Ethernet bus, ensuring the synchronization and accuracy of command execution. In the verification node cluster, the verification nodes for band saw processing equipment evaluate whether the saw blade cutting path and cutting depth in the cutting scheme are suitable for the band saw's processing capabilities, based on local knowledge base information such as band saw blade performance and saw blade drive mechanism parameters. The verification nodes for circular saw processing equipment evaluate whether the feed speed and spindle speed in the scheme meet the requirements of the circular saw, taking into account the circular saw blade speed range and feed transmission system characteristics. The verification nodes for multi-blade saws evaluate whether the parameters of each saw blade in the scheme are coordinated, based on the multi-blade synchronous control parameters, ensuring the scheme is compatible with different types of cutting equipment.

[0035] Example 3: Real-time acquisition and analysis of multimodal sensor data during cutting, and online adjustment process for abnormal situations. Multimodal sensor data acquisition is achieved through sensor arrays distributed at key parts of the wood cutting saw. Cutting force waveform data is acquired using a quartz triaxial dynamic force sensor installed at the front end of the spindle. The sensor range is 0-5kN, the natural frequency is not less than 50kHz, and the sampling frequency is set to 10kHz to capture the high-frequency dynamic components of the cutting force. The acquired raw data is digitized after anti-aliasing filtering and amplifier conditioning via a 24-bit high-precision ADC module. Saw blade vibration spectrum data is acquired using a triaxial MEMS accelerometer. The sensor is installed near the tool holder clamping area, with a measurement range of ±500g, a frequency response of 0.5-10kHz, and a sampling rate of 20kHz. A data acquisition synchronous triggering mechanism ensures time alignment with the cutting force data. After window function processing and spectral leakage compensation, the vibration signal is converted into a power spectral density distribution within the 0-10kHz frequency range using a fast Fourier transform. Infrared thermal imaging data were acquired using an uncooled infrared focal plane array camera with a resolution of 640×480, a spectral range of 8-14μm, a temperature measurement accuracy of ±1℃, and an acquisition frame rate of 30fps. The camera was installed inside a protective housing and the mirror surface was kept clean by an air curtain. After non-uniformity correction and radiometric calibration, the thermal image data generated a temperature distribution matrix.

[0036] Multimodal sensor data is aligned using a time synchronization module. The synchronization signal is triggered by the zero-position pulse of the spindle encoder. Timestamps are added to each sensor data point with microsecond-level accuracy. The aligned data forms a 256×256×3 dimensional tensor structure. The anomaly detection network employs a deep fusion architecture, incorporating a hybrid structure of Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNNs). The network input layer receives three channels of sensor data. The convolutional branch uses four convolutional layers to extract spatial features, with kernel sizes of 7×7, 5×5, 3×3, and 3×3, and channel numbers of 32, 64, 128, and 256 respectively. Each convolutional layer is followed by batch normalization and a LeakyReLU activation function. The LSM branch processes time-series data, with 128 hidden units. A sigmoid activation function is used for the forget gate. Finally, the spatiotemporal features are concatenated by a feature fusion layer before being input into a fully connected layer. The anomaly probability is calculated using the sigmoid output layer. When the output value exceeds a threshold of 0.85, an online adjustment process is triggered. This threshold can be dynamically adjusted according to processing requirements.

[0037] After the online adjustment process is initiated, the anomaly type identification module first classifies the anomaly patterns. This module employs a support vector machine multi-classifier, selecting a radial basis function as the kernel function. The feature vector includes 12 dimensions such as cutting force harmonic components, vibration dominant frequency offset, and temperature gradient distribution. The classification results output anomaly type identifiers, including categories such as saw blade wear, foreign material, and vibration instability, along with a confidence score. The impact range assessment module performs calculations using a simplified finite element model. The model establishes a thermo-mechanical coupling equation based on the current cutting depth and feed rate to calculate the regional impact coefficient. , in: Indicates the thermo-coupling effect coefficient. The number of affected grid cells, This represents the rate of temperature change of the i-th unit. Indicates the magnitude of the cutting force. Material density, Specific heat capacity Unit volume.

[0038] The calculation results generate the three-dimensional coordinate boundary of the affected area.

[0039] The adjustment range of the current cutting parameters is calculated by a fuzzy inference system. The system input variables include the severity of the anomaly (0-1) and the processing progress percentage (0-100%), while the output variables are the feed rate adjustment coefficient and the spindle speed correction. The fuzzy rule base contains 25 rules, employing the Mamdani inference method, and defuzzification is performed using the centroid method. The adjustment instruction generation module generates compensation control signals based on the calculation results. The signal format adopts the EtherCAT standard protocol, including the target parameter value, adjustment slope, and execution time. The adjustment slope is limited to the allowable acceleration range of the wood cutting saw. The compensation signal is transmitted to the driver via a real-time Ethernet bus, with a transmission cycle ensured to be within 1ms. The driver uses a feedforward control algorithm to smoothly execute the parameter adjustment, while simultaneously monitoring the actual execution effect and feeding it back to the anomaly detection system to form a closed-loop control. The entire adjustment process is logged, including the anomaly trigger time, adjustment parameter values, and execution results, for system performance analysis and model optimization.

[0040] Taking the real-time monitoring and adjustment of a 2500mm×400mm×80mm oak board during processing as an example, the wood encounters a hidden resin pocket area when processed to approximately 1300mm. A multimodal sensing system collects processing data in real time. Cutting force waveform data is acquired using a Kistler 9257B triaxial force gauge mounted on the spindle, with a sampling frequency of 10kHz and a measurement range of ±8kN. During normal cutting, the force signal amplitude fluctuates within the range of 1200N±200N. Saw blade vibration spectrum data is acquired using a PCB352C33 accelerometer mounted on the tool holder, with a sampling frequency of 20kHz and a frequency response range of 0.5-10kHz. Under normal processing conditions, the dominant vibration frequency is concentrated around 850Hz, with the amplitude maintained at 2.5m / s². Infrared thermal imaging data was acquired using a FLIRA6750sc thermal imager with a resolution of 640×512, a frame rate of 30fps, and a temperature measurement range of 20-150℃. The temperature distribution in the normal cutting area was uniform, maintained between 65-75℃.

[0041] When the saw blade reached the 1280mm position, the sensor data began to show abnormal changes. The cutting force waveform amplitude suddenly increased to 1850N and fluctuated violently in the 1400-2100N range. The vibration spectrum's main frequency shifted to 1250Hz and multiple high-frequency harmonic components appeared, with the amplitude jumping to 4.8m / s². Infrared thermography showed that the temperature in the cutting area rose from 70℃ to 92℃ within 0.5 seconds, forming a localized high-temperature region. After the multimodal data was aligned by the time synchronization module, it was formed into a 256×256×3 tensor structure. The convolutional branch of the anomaly detection network extracted high-frequency impact features from the force signal, broadband resonance features from the vibration spectrum, and gradient abrupt change features from the thermography. The long short-term memory branch identified a time series pattern in which the force signal amplitude increased by more than 50% within 0.2 seconds, and the network output anomaly probability value reached 0.91, exceeding the set threshold of 0.85. The online adjustment process was immediately initiated. The anomaly type identification module extracted 12-dimensional feature vectors for analysis: the amplitude of the third harmonic in the cutting force harmonic component increased by 2.3 times, the vibration dominant frequency offset rate increased by 400Hz per second, and the temperature gradient reached 15℃ / mm. The support vector machine classifier, based on the radial basis function kernel, calculated and output the anomaly type as "material foreign object" with a confidence level of 0.87, and the secondary possibility as "saw blade wear" with a confidence level of 0.13. The influence range assessment module established a thermo-mechanical coupling model, calculated the influence coefficient of the resin capsule area to be 0.78, and determined the affected area to be the 1280-1320mm segment, with a depth-direction influence range of 8-15mm.

[0042] The current cutting parameters are adjusted via a fuzzy inference system. The input variable, anomaly severity, is set to 0.82, and the machining progress percentage is 51.2%. 25 rules in the fuzzy rule base are activated, outputting a feed rate adjustment coefficient of 0.65 and a spindle speed correction of +8%. The adjustment command generation module calculates the specific parameters: feed rate decreases from 1200 mm / min to 780 mm / min, spindle speed increases from 8000 rpm to 8640 rpm, and coolant flow rate increases by 25%. The compensation control signal is encapsulated using the EtherCAT protocol, containing the target parameter value, adjustment slope, and execution time point. The adjustment slope is limited to an acceleration range of 300 mm / min².

[0043] The compensation signal was transmitted to the Beckhoff AX5000 series driver via real-time Ethernet with a transmission cycle of 0.5ms. The driver employed a feedforward control algorithm to smoothly execute parameter adjustments. Within 0.8 seconds after execution, the cutting force amplitude gradually decreased to 1550N±150N, the vibration frequency stabilized at 980Hz, the amplitude decreased to 3.2m / s², and the temperature distribution returned to the range of 75-82℃. The entire adjustment process lasted 3.2 seconds. The system recorded an operation log including the abnormal trigger time, adjustment parameter values, and execution results. This data was synchronously uploaded to a cloud database for subsequent analysis. During the quality inspection after machining, the surface quality of the resin capsule area met acceptable standards, with no chipping or burning defects observed.

[0044] Example 4: The process of comprehensive quality inspection of finished timber after the cutting process and comparison analysis with the digital twin model is illustrated using a 2000mm×200mm×50mm red pine board as an example. The finished timber is transported to an automated inspection station via conveyor belt. This station is equipped with three inspection systems: a laser scanning system using a binocular laser scanner with a scanning accuracy of 0.02mm, which performs a full-coverage scan of the six surfaces of the timber during the uniform movement of the conveyor belt, generating point cloud data containing 2 million points; a contact roughness inspection system using a multi-probe measuring arm, with each probe having a measuring stroke of 50mm and a measuring pressure of 4mN, performing a measurement every 100mm along the length of the timber; and an ultrasonic flaw detection system using a 32-element phased array probe with a probe frequency of 5MHz, a scanning angle range of ±45°, a scanning interval of 0.5mm, and water spray coupling.

[0045] Laser scanning point cloud data is registered and denoised using the ICP algorithm, and then geometric accuracy features are extracted. Length is calculated using the maximum distance between the two end faces. Width and thickness are measured at 100mm intervals. Flatness is calculated by fitting a reference plane using the least squares method to calculate the maximum deviation. Parallelism is calculated by comparing the angle between the fitted planes of two opposite faces. Contact roughness measurement data is filtered to extract profile curves, and the average of five points—arithmetic mean deviation Ra, maximum profile height Rz, and profile support length ratio Rmr—is calculated. Ultrasonic testing data is used to reconstruct the internal structure using a synthetic aperture focusing algorithm, identifying defect types including knots, cracks, and decay areas, and calculating the defect size, location, and density variation coefficients. The digital twin model includes the ideal geometric model, surface quality standards, and internal structural requirements for this specification of red pine lumber. The ideal geometric model is based on design drawings, with dimensions of 2000±0.5mm × 200±0.3mm × 50±0.2mm, flatness requirement ≤0.3mm, and parallelism requirement ≤0.4mm. The surface roughness standard requires Ra≤6.3μm and Rz≤25μm. The internal quality requirement is no defects exceeding 2mm and a density variation coefficient ≤0.15.

[0046] The comparative analysis process is executed using specialized analysis software, comparing actual test data with the digital twin model item by item. Geometric dimensional deviations are calculated by the difference between the actual measured value and the standard value; flatness and parallelism deviations are calculated by comparing the positional relationship between the actual fitted plane and the ideal plane. Surface roughness deviation is calculated as the ratio of the actual measured value to the standard value. Internal defects are assessed by comparing the actual defect distribution with the degree of conformity to the allowable defect standard. The compliance levels of each quality indicator are graded according to industry standards: dimensional accuracy is divided into three levels according to ISO 13002 standard, surface quality is divided into five levels according to ASTM D1666 standard, and internal quality is divided into four levels according to EN 1310 standard.

[0047] The analysis results generate a multi-dimensional quality analysis report, stored in structured XML format, which includes deviation values, compliance levels, and spatial distribution information. Table 1 shows the comparison between measured data and standard values ​​at some test locations of the red pine board.

[0048] Table 1: Comparison of Finished Timber Quality Inspection Data with Standards

[0049] After the report is generated, the system automatically performs quality trend analysis, plots the distribution curves of various quality indicators along the length of the timber, and identifies areas exceeding the allowable range. Simultaneously, a 3D visualization model is generated, using different colors to indicate the quality level of each part: red indicates exceeding the standard, yellow indicates a critical state, and green indicates compliance. The final report includes three parts: textual analysis, data tables, and graphs. It is transmitted to the quality management department and production control system via the company's internal network, serving as the basis for product grading and process optimization. All test data is stored in the quality database, linked to production batches, timber cutting saw parameters, and environmental conditions, providing a complete data chain for quality traceability.

[0050] Example 5: The entire process of converting quality analysis data into model training data and updating the system model through a federated learning framework. The multidimensional quality analysis report is first input into the data parsing module. This module uses an XMLSchema-based parser to extract the structured data from the report. It extracts numerical fields such as length deviation, width deviation, thickness deviation, flatness deviation, and parallelism deviation from the geometric accuracy section; Ra, Rz, and Rmr values ​​for each measurement point from the surface quality section; and defect size, location coordinates, and density variation coefficient from the internal quality section. Each data field is mapped to a specific dimension of the feature vector, ultimately forming an original feature vector containing 32 dimensions. Simultaneously, the deviation values ​​of each quality indicator are used as label vectors. The data standardization module preprocesses the original feature vector. Length-type size deviations are transformed to the [-1,1] range using a min-max normalization method. Surface roughness data undergoes Z-score standardization after logarithmic transformation. Internal defect data is one-hot encoded according to defect type before normalization. The processed feature dimensions are uniformly scaled to the [0,1] range, and the label data is divided into positive and negative bias categories based on the bias direction and processed separately. All standardized parameters are stored in the configuration file for unified use in subsequent data processing. The training data construction module combines the standardized features and labels into feature-label pairs. Each data pair contains metadata such as timestamp, batch number, and ambient temperature. The final generated dataset is stored in the TFRecord format in a distributed file system.

[0051] When the federated learning framework initiates the model optimization task, each edge node obtains the training dataset from the file system, and the local feature extraction model is updated using a transfer learning strategy. The model is based on a pre-trained ResNet-34 architecture, keeping the weights of the bottom convolutional layers fixed and only fine-tuning the top fully connected layers. During training, a momentum-driven stochastic gradient descent algorithm is used, with a momentum coefficient set to 0.9. The learning rate is gradually decayed from 0.01 to 0.0001 using cosine annealing scheduling, the batch size is set to 32, and the training period is 50 epochs. The loss function is a weighted sum of mean squared error loss and an L2 regularization term, with a regularization coefficient set to 0.001. The hierarchical decision model is updated using a policy gradient-based reinforcement learning algorithm, using the quality bias value as the reward signal. The reward function is designed as a negative weighted sum of squared biases. The decision network parameters are updated using a proximal policy optimization algorithm. Both the policy network and the value network are three-layer fully connected structures with hidden layer dimensions of 256, 128, and 64, respectively. Each update uses the most recent 1000 training samples and performs 10 policy iterations. Each iteration uses KL divergence constraints to ensure the stability of the policy update, and the KL objective value is set to 0.01.

[0052] The updated model parameters are uploaded to the federated aggregation server via an encrypted channel. The server uses secure multi-party computation technology for parameter aggregation. The feature extraction model's parameter aggregation uses the FedAvg algorithm, assigning weights based on the amount of data at each node; the decision model's parameter aggregation uses an importance-weighted averaging method, considering the historical performance metrics of each node. Gaussian noise is added during the aggregation process to achieve differential privacy protection, with a noise standard deviation set to 0.001. The new generation of global model parameters is synchronized to each edge computing node via the blockchain network. The blockchain smart contract manages the parameter distribution process, and each parameter update transaction includes the model version number, hash value, and digital signature. After receiving the parameters, nodes first verify the digital signature and hash value, then use cross-validation to evaluate the new model's performance on the local test set. Only when the performance improvement exceeds a threshold is the local model updated; otherwise, it rolls back to the previous version and reports an anomaly. 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 wood-cutting saw that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or wood-cutting saw.

[0053] 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. An artificial intelligence-based wood cutting control method, applied in a wood cutting control system, characterized in that, The wood cutting control system includes a wood cutting saw, which is equipped with a spindle drive motor (1), a feed drive motor (2), and a saw blade attitude adjustment cylinder (3). The method includes the following steps: Acquire multi-dimensional detection data of the wood to be cut, including high-resolution surface images, X-ray tomography data, and acoustic wave propagation characteristic data; Distributed feature extraction is performed on the multi-dimensional detection data through an edge computing node cluster to generate a primary feature set. The primary feature set is input into the federated learning framework for cross-node feature fusion to generate a global feature representation. Based on the global feature representation, a multi-objective optimized cutting scheme is generated through a hierarchical decision model. The scheme includes the cutting path, cutting depth, feed rate, and spindle speed of the wood cutting saw. The multi-objective optimized cutting scheme is dynamically verified and adjusted to generate the final cutting control command, which controls the main shaft drive motor (1), feed transmission motor (2) and saw blade posture adjustment cylinder (3) of the wood cutting saw.

2. The artificial intelligence-based wood cutting control method according to claim 1, characterized in that, The distributed feature extraction of the multi-dimensional detection data through an edge computing node cluster includes: Multi-scale convolution operations are performed on the high-resolution surface image to extract surface texture feature maps, which are used to determine the initial contact angle of the saw blade of the wood cutting saw. The X-ray tomography data is divided into three-dimensional meshes to construct a voxelized structural model. The spatial coordinates of internal defects in the wood are marked to plan cutting avoidance paths. The acoustic wave propagation characteristic data is spatiotemporally aligned with the voxelized structural model to generate a material property distribution map. The surface texture feature map and the material property distribution map are integrated by a feature coding network to output a primary feature set.

3. The artificial intelligence-based wood cutting control method according to claim 2, characterized in that, The primary feature set is input into the federated learning framework for cross-node feature fusion, including: In each edge computing node, a local feature extraction model is used to process the primary feature set; The feature extraction results from multiple nodes are verified and weighted using a blockchain consensus mechanism; The verified feature representations are integrated using a federated aggregation algorithm to generate the global feature representation.

4. The artificial intelligence-based wood cutting control method according to claim 3, characterized in that, Based on the global feature representation, a multi-objective optimized cutting scheme is generated through a hierarchical decision model, including: The global feature representation is input into a multi-layer decision network that incorporates an attention mechanism; The optimal cutting direction sequence is determined through the first decision-making layer; The second decision layer calculates the combination of saw blade parameters for each cutting segment; The multi-objective optimized cutting scheme, which includes timing control instructions, is generated through the third decision layer.

5. The artificial intelligence-based wood cutting control method according to claim 4, characterized in that, The multi-objective optimized cutting scheme is dynamically verified and adjusted, including: The multi-objective optimization cutting scheme is distributed to the verification node cluster. Each verification node performs a feasibility assessment of the cutting scheme based on its local knowledge base, collects the assessment results, calculates the scheme correction parameters, and iteratively optimizes the cutting scheme according to the scheme correction parameters.

6. The artificial intelligence-based wood cutting control method according to claim 5, characterized in that, Also includes: Real-time acquisition of multimodal sensing data during the cutting process, including cutting force waveform data, saw blade vibration spectrum and infrared thermal imaging data; The multimodal sensing data is input into the anomaly detection network for real-time analysis; When an abnormal pattern is detected, an online adjustment process is triggered.

7. The artificial intelligence-based wood cutting control method according to claim 6, characterized in that, The process of triggering online adjustment includes: Analyze the type and impact range of abnormal modes, calculate the adjustment range of the current cutting parameters, and generate a compensation control signal containing real-time adjustment instructions.

8. The artificial intelligence-based wood cutting control method according to claim 7, characterized in that, Also includes: After the cutting process is completed, comprehensive quality inspection data of the finished wood is collected. The comprehensive quality inspection data is compared and analyzed with the digital twin model, and a model optimization dataset is generated based on the analysis results.

9. The artificial intelligence-based wood cutting control method according to claim 8, characterized in that, The comparative analysis of comprehensive quality inspection data with digital twin models includes: Extract the geometric precision characteristics, surface roughness distribution, and internal defect characteristics of the finished wood; Calculate the deviation matrix between the actual processing results and the expected target, evaluate the compliance level of each quality indicator, and generate a multi-dimensional quality analysis report.

10. The artificial intelligence-based wood cutting control method according to claim 9, characterized in that, The dataset for model optimization based on the analysis results includes: The multidimensional quality analysis report is converted into feature-label pair training data; The parameters of the local feature extraction model and the hierarchical decision model are updated using a federated learning framework, and the updated model parameters are synchronized to each edge computing node through a blockchain network.

Images