Intelligent panel grouping method and system based on machine learning

Abstract

This application provides a machine learning-based intelligent cargo loading method and system, applied in the fields of artificial intelligence and air logistics technology. By collecting and processing multi-dimensional feature data of cargo, feature vectors and physical constraint parameters are constructed. With the help of a pre-trained deep learning model group, the stackability, fragility, and center of gravity stability of cargo after stacking are intelligently predicted. By combining the prediction results and constraint parameters, a multi-objective optimization model is established and the optimal stacking scheme that takes into account space, safety, and efficiency is solved. During execution, a real-time safety verification and dynamic backtracking mechanism are introduced to ensure the stability and reliability of the loading process. Through an online learning mechanism, actual operation data is accumulated to continuously update model parameters and optimization weights, realizing the system's self-evolution and closed-loop optimization. This achieves a fundamental transformation from unreplicable personal experience to sustainable evolution of system intelligence, significantly improving the intelligence level, safety reliability, and long-term operational efficiency of air cargo loading.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and aviation logistics technology, specifically to an intelligent board assembly method and system based on machine learning. Background Technology

[0002] In the field of air logistics, the palletizing of air container (Unit Load Device, ULD) is a critical link affecting transportation efficiency, cost, and safety. Traditional palletizing methods mainly rely on manual experience for placement and combination, which is not only inefficient but also makes it difficult to accurately control the stability and center of gravity distribution of the stack, easily leading to safety hazards such as center of gravity shift, cargo damage, or even pallet collapse.

[0003] To improve the automation level of palletizing operations, some palletizing schemes have been developed that employ heuristic or rule-based algorithms, such as using simple rules like volume priority or weight priority for decision-making. However, these schemes typically only consider a single or a few static dimensions, and their algorithmic logic is rigid. They cannot learn and optimize stacking strategies autonomously from massive amounts of historical operational data, resulting in problems such as inflated space utilization but insufficient stability, loss of center of gravity, and violations of special cargo safety regulations in practical applications. These issues make it difficult to meet the high standards of safety, timeliness, and intelligence required by modern air logistics.

[0004] Therefore, a new intelligent board assembly solution is needed. Summary of the Invention

[0005] In view of this, the embodiments of this specification provide a machine learning-based intelligent palletizing method and system, which is applicable to cargo stacking planning and operation optimization in scenarios such as aviation, ports and warehousing logistics. By introducing machine learning and combinatorial optimization algorithms, automated and intelligent palletizing decisions are achieved, improving ULD space utilization and palletizing efficiency.

[0006] The embodiments in this specification provide the following technical solutions:

[0007] This specification provides an intelligent board assembly method based on machine learning, including:

[0008] Obtain multidimensional feature data of the cargo to be loaded;

[0009] The multidimensional feature data is combined to generate a feature vector for each cargo, and a compatibility matrix and physical constraint parameters between cargoes are established based on the multidimensional feature data.

[0010] The cargo feature vector is input into a pre-trained deep learning model set to obtain a prediction result; wherein, the deep learning model set includes:

[0011] Stackability prediction model, used to predict the stacking level and order between any two goods;

[0012] A vulnerability regression model is used to predict the vulnerability score for each item.

[0013] The center of gravity stability prediction model is used to predict the overall center of gravity offset and stability coefficient after stacking, based on the current stacking state and the characteristics of the goods to be placed.

[0014] Based on the prediction results, the cargo compatibility matrix, and physical constraint parameters, a multi-objective optimization model is constructed with space utilization, center of gravity balance, stability, and operational efficiency as objectives. The multi-objective optimization model is then solved to obtain the stacking scheme.

[0015] The stacking scheme is subjected to real-time safety verification of its overall center of gravity and stability. If the verification passes, the stacking scheme is converted into a work instruction and sent to the execution device to perform loading. If the verification fails, a backtracking mechanism is triggered to replan the stacking scheme.

[0016] The actual operation data of the execution device during the loading process is collected, and the parameters of the deep learning model group and the multi-objective optimization model are incrementally updated based on the actual operation data.

[0017] This specification also provides an intelligent board assembly system based on machine learning, comprising:

[0018] The data acquisition module is used to acquire multidimensional feature data of the goods to be loaded;

[0019] The feature engineering module is used to combine features from the multidimensional feature data to generate a feature vector for each cargo, and to establish a compatibility matrix and physical constraint parameters between cargoes based on the multidimensional feature data.

[0020] The learning prediction module is used to input the cargo feature vector into a pre-trained deep learning model group to obtain the prediction result; wherein, the deep learning model group includes: a stackability prediction model, used to predict the stacking level and order between any two cargoes; a vulnerability regression model, used to predict the vulnerability score of each cargo; and a center of gravity stability prediction model, used to predict the overall center of gravity offset and stability coefficient after stacking based on the current stacking state and the characteristics of the cargo to be placed.

[0021] The pallet planning module is used to construct a multi-objective optimization model with the objectives of space utilization, center of gravity balance, stability and operation efficiency based on the prediction results, the compatibility matrix between goods and physical constraint parameters, and to solve the multi-objective optimization model to obtain the stacking scheme.

[0022] The safety verification module is used to perform real-time safety verification of the overall center of gravity and stability of the stacking scheme. If the verification passes, the stacking scheme is converted into a work instruction and sent to the execution device to perform loading; if the verification fails, a backtracking mechanism is triggered to replan the stacking scheme.

[0023] The execution feedback module is used to collect the actual operation data of the execution device during the loading process, and to incrementally update the parameters of the deep learning model group and the multi-objective optimization model based on the actual operation data.

[0024] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least:

[0025] By collecting multi-dimensional feature data, intelligent prediction, optimized planning, online verification, and self-learning feedback, a complete closed loop of "perception-prediction-optimization-execution-learning" is formed, realizing automated decision-making and dynamic correction of cargo during the palletizing process. It automatically generates stacking schemes with high space utilization, stable center of gravity, and excellent operational efficiency, significantly improving loading space utilization, stacking stability, and operational efficiency. Furthermore, through continuous self-learning, it adapts to diverse freight scenarios, significantly improving the intelligence level, safety, reliability, and long-term operational efficiency of air cargo loading. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of a machine learning-based intelligent board assembly method in this application;

[0028] Figure 2 This is a diagram showing the fusion of machine learning and optimization algorithms in this application;

[0029] Figure 3 This is an overall architecture diagram of a machine learning-based intelligent board assembly system in this application. Detailed Implementation

[0030] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0031] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0033] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0034] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0035] Traditional air container palletizing operations rely heavily on manual experience, resulting in low efficiency, poor operational consistency, large fluctuations in space utilization, and safety hazards such as imbalance of center of gravity, cargo damage, and stack collapse.

[0036] In view of this, the inventors discovered through research and improvement that existing palletizing solutions partially employ heuristic or rule-based algorithms, such as making automated decisions based on simple rules like "volume priority" and "heavy loads should not be placed on light loads." However, these solutions are essentially rule-driven automation and do not achieve true intelligence. They cannot handle irregular goods, cannot constrain from multiple dimensions, and cannot learn autonomously and continuously optimize from historical data, resulting in extremely poor flexibility and adaptability, making it difficult to meet the complex and ever-changing actual operational needs.

[0037] Based on this, the embodiments of this specification propose an intelligent cargo loading method and system based on machine learning. This method can decompose and transform the experience-based decision-making process of human experts into computable and optimizable technical modules. The overall approach is as follows: First, by collecting and processing multi-dimensional feature data of cargo, feature vectors and physical constraint parameters are constructed. Then, a pre-trained deep learning model group is used to intelligently predict the stackability, fragility, and center of gravity stability of the cargo after stacking. Next, by combining these prediction results and constraints, a multi-objective optimization model is established to solve for the optimal stacking scheme that balances space, safety, and efficiency. During the execution process, a real-time safety verification and dynamic backtracking mechanism are introduced, and actual operation data is accumulated through an online learning mechanism to continuously update model parameters and optimization weights, enabling the system to self-evolve and forming a closed-loop design of "perception-prediction-optimization-execution-learning". This achieves a fundamental transformation from unreplicable personal experience to sustainable evolutionary system intelligence, significantly improving the intelligence level, safety reliability, and long-term operational efficiency of air cargo loading.

[0038] The technical solutions provided by the various embodiments of this application are described below with reference to the accompanying drawings.

[0039] like Figure 1 As shown in the embodiments of this specification, an intelligent board assembly method based on machine learning is provided, including:

[0040] Step S100: Data Acquisition: Obtain multidimensional feature data of the goods to be loaded.

[0041] In implementation, step S101, cargo identification and registration: the identification process is automatically started when the cargo enters the site. For example, the geometric shape of the cargo is scanned non-contactly using a 3D laser scanner, industrial camera and RFID tag identification equipment to record the cargo number, batch number and basic attributes.

[0042] Step S102, Size and Mass Acquisition: Automatically read the length L, width W, and height H of the goods, obtain the mass m through electronic weighing equipment, and automatically calculate the volume V = L×W×H and the density ρ = m / V;

[0043] Step S103, Packaging and Material Inspection: Use image recognition and infrared scanning equipment to identify the packaging material s (such as wooden box, plastic box, metal box, etc.), and match the friction coefficient μ, rotatable set R and bearing pressure coefficient σ from the database;

[0044] Step S104, Safety Attribute Collection: Input or scan the cargo safety label on the operating terminal to identify the dangerous goods category DG, cold chain temperature zone t, and vulnerability level f;

[0045] Step S105, Center of Gravity Data Measurement: The force distribution of the cargo at different support points can be obtained through a dual-platform weighing device, and the center of gravity coordinates g = (g_x, g_y, g_z) can be calculated. After the data is verified by the edge nodes, it is uploaded to the database to form a cargo feature table.

[0046] Specifically, place the goods smoothly onto the dual-platform weighing device, ensuring that the bottom surface of the goods is in complete contact with the platforms. The two weighing platforms are located in the coordinate system. and The position and platform height are consistent, with z=0 as the reference plane.

[0047] Read the supporting forces F1 and F2 of the two platforms, and the total mass of the cargo m (already obtained from S102). Take the gravitational acceleration as g = 9.8 m / s², then the total gravity G = m × g.

[0048] Calculation of centroid coordinates in the X direction:

[0049] Assume the two platforms are symmetrically distributed along the X-axis with a distance of L between them. x According to the principle of torque balance:

[0050] F1× L x = G × (L x g x )

[0051] Solving for:

[0052] g x = L x × (1 F1 / G)

[0053] Or equivalent:

[0054] g x = (F2× x1+ F1× x2) / (F1+ F2) (if the platform positions are not equidistant).

[0055] Calculation of barycenter coordinates in the Y direction:

[0056] Similarly, if the two platforms are arranged along the Y-axis or if repeated measurements are taken by rotating 90°, we can obtain:

[0057] g_y = L_y × (1 F1_y / G)

[0058] Alternatively, synchronous measurement via a dual-axis platform can be used for direct decoupling calculations.

[0059] Calculation of centroid coordinates in the Z direction:

[0060] Since the cargo is a rigid body and placed on a horizontal platform, the center of gravity in the z-direction can be estimated using empirical formulas or structural modeling:

[0061] If the goods are regular geometric shapes (such as cuboids), then g_z = H / 2;

[0062] For irregular bodies, the system combines 3D scan data with density distribution and uses the volume integral method for calculation:

[0063] g_z = ∫∫∫ z·ρ(x,y,z) dV / ∫∫∫ ρ(x,y,z) dV;

[0064] Where ρ(x,y,z) is a local density function, obtained by interpolating the density ρ of S102 and the material distribution of S103.

[0065] The calculated results are compared with historical center of gravity data for similar goods. If the deviation exceeds a preset threshold (e.g., ±5%), manual verification or a second measurement is triggered. The final output center of gravity coordinates g = (g x The data (g_y, g_z) are verified by edge nodes and then uploaded to the database to form a cargo feature table.

[0066] Step S200, Feature and Constraint Modeling: Combine the features of the multidimensional feature data to generate the feature vector of each cargo, and establish the compatibility matrix and physical constraint parameters between cargoes based on the multidimensional feature data.

[0067] During implementation, the collected multidimensional feature data is systematically processed, and the multidimensional feature data of each cargo are combined to construct a structured feature vector.

[0068] For example, features such as geometric dimensions, mass, density, packaging material, safety attributes, and centroid coordinates can be combined into a feature vector F_i.

[0069] F_i = [L, W, H, m, ρ, s, μ, R, DG, t, f, g_x, g_y, g_z];

[0070] Meanwhile, a compatibility matrix representing the stackable relationship between goods is established based on the multidimensional feature data of the goods, and physical constraint parameters for ensuring stacking safety and stability are derived, providing a quantitative basis for subsequent optimization and safety verification.

[0071] Step S300, Intelligent Prediction: Input the cargo feature vector into a pre-trained deep learning model set to obtain the prediction result; wherein, the deep learning model set includes:

[0072] Stackability prediction model (StackabilityNet) is used to predict the stacking level and order between any two items;

[0073] The vulnerability regression model (FragilityReg) is used to predict the vulnerability score for each cargo.

[0074] The Center of Gravity Stability Prediction Model (CGPredictor) is used to predict the overall center of gravity offset and stability coefficient after stacking, based on the current stacking state and the characteristics of the goods to be placed.

[0075] Step S400: Based on the prediction results, the compatibility matrix between goods, and the physical constraint parameters, construct a multi-objective optimization model with the objectives of space utilization, center of gravity balance, stability, and operational efficiency, and solve the multi-objective optimization model to obtain the stacking scheme.

[0076] Specifically, a multi-objective optimization model is established based on the prediction results, with space utilization rate U, center of gravity deviation ΔCG, instability penalty P and operation cost T as objective functions, and the optimal stacking scheme is solved under constraints such as geometric feasibility, pressure limit, hazardous materials isolation, and cold chain independence.

[0077] Step S500: Perform a real-time safety check on the overall center of gravity and stability of the stacking scheme. If the check passes, the stacking scheme is converted into a work instruction and sent to the execution device to perform loading. If the check fails, a backtracking mechanism is triggered to replan the stacking scheme.

[0078] Specifically, the overall center of gravity coordinates are calculated in real time after each placement of goods. If the center of gravity deviates from the safety threshold or the layer bearing capacity exceeds the limit, a backtracking mechanism is triggered to replan the stacking scheme to ensure stacking stability and safety. If successful, the optimal stacking scheme is converted into an executable operation instruction. The instruction includes: goods number, placement coordinates (x, y, z), rotation angle θ, reinforcement method, and strapping path, etc., and is directly sent to the automated palletizing robot arm or loading control system. For example, the operation instruction is sent to the automated robot arm for execution through the OPC-UA interface.

[0079] Step S600, Data Acquisition and Feedback: Collect actual operation data of the execution device during the loading process, and incrementally update the parameters of the deep learning model group and the multi-objective optimization model based on the actual operation data.

[0080] Specifically, during the loading process, real-time data is collected, such as operation time, weighting rate, and rework status. The model is then adaptively updated based on this feedback data. An incremental learning mechanism can be used to dynamically adjust model parameters and target weights. Figure 2 As shown, by repeatedly executing steps S300-S600, the online model can be dynamically optimized, thereby adapting to different cargo types, flights, and seasonal scenarios and maintaining long-term performance stability.

[0081] This application achieves intelligent, interpretable, and safe optimization of cargo stacking through multi-model learning, real-time verification, and self-learning feedback, which has significant practical value and promotional significance in air logistics and intelligent warehousing.

[0082] In some embodiments, the multidimensional feature data is preprocessed before feature combination, including:

[0083] Unit standardization and anomaly detection are used to remove missing values ​​and outliers.

[0084] Specifically, by standardizing units, the measurement units of cargo feature data (such as size, weight, etc.) from different sensors or data sources are uniformly converted into the system's preset standard units, and the data is normalized to eliminate the benchmark deviation caused by differences in measurement systems, ensuring that each feature has comparability and balanced weight in the algorithm.

[0085] Furthermore, by using an anomaly detection mechanism to identify and remove missing values ​​and anomalies that clearly do not conform to physical laws or business logic, the consistency and reliability of the data input to the feature combination and prediction model are ensured, providing a high-quality data foundation for subsequent intelligent decision-making.

[0086] In some embodiments, establishing the inter-cargo compatibility matrix and physical constraint parameters based on the multi-dimensional feature data includes:

[0087] Based on the geometric dimensions and packaging materials in the multidimensional feature data, the layer bearing capacity threshold of each item is obtained;

[0088] The coefficient of friction between the stacked goods is obtained based on the packaging material of the contact surfaces of the goods.

[0089] By combining the layer bearing capacity threshold, the friction coefficient, and the preset safety isolation rules, a compatibility matrix between goods is generated; wherein, the elements in the compatibility matrix are used to characterize whether stacking is allowed between corresponding pairs of goods.

[0090] Specifically, the layer bearing capacity (i.e. maximum bearing capacity) σmax(i) of each cargo i is determined by its packaging material s and geometry. The system obtains the corresponding unit area compressive strength σ_material(s) (in kPa) from the pre-stored material database of various packaging compressive strengths based on the packaging material, and calculates the maximum allowable bearing value of the cargo based on the bottom area Ai of the cargo, thereby ensuring that the lower layer cargo will not be damaged due to the pressure exceeding its limit during the stacking process.

[0091] The formula for calculating the layer bearing capacity threshold σmax(i) is as follows:

[0092] σmax(i) = σ_material(s) × Ai;

[0093] Where: σmaterial(s) represents the compressive strength per unit area of ​​material s (such as wooden boxes, plastic boxes, metal boxes), which is obtained by looking up a table;

[0094] Ai represents the base area (effective pressure-bearing area) of cargo i. For cargo with a rectangular base, Ai = Li × Wi, where Li represents the length of cargo i and Wi represents the width of cargo i. For cargo with irregular shapes (such as cylinders or irregular parts), the minimum projected area of ​​the cargo is taken as the effective pressure-bearing area.

[0095] for example:

[0096] If the goods are wooden crates with a compressive strength of 80 kPa and a base area of ​​0.5 m², then the layer bearing capacity of the wooden crate is: σmax(i) = 40 kN;

[0097] If the goods are in a plastic box with a compressive strength of 30 kPa and a bottom area of ​​0.3 m², then the layer bearing capacity of the plastic box is: σmax(i) = 9 kN.

[0098] For any two cargo pairs (i, j) that can potentially form a stacking relationship, such as cargo i stacked on cargo j, their static friction coefficient μ(i, j) is jointly determined by the packaging materials of the contact surfaces of both cargo pairs. By querying a pre-defined database of material pair friction coefficients, the stability of the upper cargo against slippage during stacking can be evaluated.

[0099] The formula for calculating the friction matrix μ(i,j) is as follows:

[0100] μ(i,j) = μ_surface(s_i, s_j);

[0101] in:

[0102] s_i: Packaging material of goods i;

[0103] s_j: Packaging material of goods j;

[0104] μ_surface(s_i, s_j): A pre-stored table of friction coefficients for material pairs (e.g., wood-wood=0.4, plastic-metal=0.2, cardboard-plastic=0.3).

[0105] If there are additional materials such as anti-slip mats or tape on the contact surface, a correction coefficient k is added according to the type of additional material. Usually, k ∈ [1.0, 1.5]. The corrected friction matrix is:

[0106] μ(i,j) = μ_surface(s_i, s_j) × k;

[0107] for example:

[0108] Cardboard boxes (s_i) are stacked on top of plastic boxes (s_j). From the table, we find that μ_surface = 0.3, which means μ(i,j) = 0.3.

[0109] If the surface of the plastic box is covered with an anti-slip mat, and the correction factor k = 1.2 is taken, then μ(i,j) = 0.3 × 1.2 = 0.36.

[0110] Finally, considering the pressure threshold of the integrated layer, the friction coefficient, and the preset safety isolation rules, such as the isolation requirements for dangerous goods and general goods, and the temperature zone independence principle for cold chain goods, a compatibility matrix M(i,j) is established between goods. Each element in the compatibility matrix corresponds to a pair of goods. The value of M(i,j) is used to indicate whether the corresponding pairs of goods are allowed to be stacked. For example, M(i,j)=1 means that goods i can be stacked on j.

[0111] It should be noted that: σmax(i) is a one-dimensional array with a length equal to the total number of goods N, and each element corresponds to the maximum pressure value of goods i; μ(i,j) is an N×N matrix, where μ(i,j) represents the friction coefficient of goods i stacked on j. If i cannot be stacked on j, then μ(i,j)=0.

[0112] In some embodiments, feature data is passed to a deep learning model group as an input sample set via a message bus.

[0113] In some embodiments, the stacking prediction model is a trained deep neural network model, trained using the following method:

[0114] Collect a training dataset, which includes positive samples and negative samples. The positive samples are cargo pairs that were successfully stacked in historical operations, and the negative samples are cargo pairs that failed to stack in historical operations.

[0115] Label each sample in the training dataset with a stacking level label;

[0116] Using the feature vectors and relative pose features of the cargo pair as input, and the corresponding stacking level label as the supervision target, the stacking prediction model is trained using the weighted cross-entropy loss function to obtain the trained stacking prediction model.

[0117] The weighted cross-entropy loss function is configured such that a higher misclassification penalty weight is assigned to the prohibited stacking category than to other stackable categories.

[0118] In practice, a training dataset containing positive and negative samples is first constructed. Positive samples are derived from successfully stacked cargo pairs in historical operations, while negative samples are derived from cargo pairs that have experienced failures such as collapse, slippage, or overpressure in historical operations. This ensures that the model can learn from both positive and negative experiences.

[0119] Next, each sample in the training dataset is assigned a stacking level label. The labels are usually in a 1 to 5-level hierarchy, where level 1 means stacking is prohibited and level 5 means optimal stacking. The labels can be annotated by domain experts or generated by a simulation system.

[0120] The deep neural network model is trained under supervised supervision using the feature vectors F_i and F_j of the cargo pairs, as well as their relative pose features, such as rotation angle and offset. The corresponding stacking level labels serve as the supervision targets. During training, a weighted cross-entropy loss function is used. This function assigns a significantly higher misclassification penalty weight to the "No Stacking" category than to other stackable categories, forcing the model to prioritize avoiding misclassifying prohibited cargo pairs as stackable during training, thereby reducing the risk of stacking accidents.

[0121] After training, the stackability prediction model can predict the stacking level S_i of any pair of goods, thereby determining whether goods i can be stacked on goods j, and providing the optimal stacking order for overall pallet planning.

[0122] In some embodiments, the stacking prediction model is constructed using the following methods:

[0123] Extract the geometric dimensions of the goods using convolutional neural network layers;

[0124] Multidimensional structured features of goods are fused through the Transformer encoder layer;

[0125] The stacking level probability is output based on the fused features through a fully connected layer and a Softmax function.

[0126] Specifically, the front end uses CNN to extract geometric shape features (L, W, H, contour map);

[0127] The intermediate layer uses a Transformer encoder to fuse structural features such as material, density, pressure, and friction.

[0128] The output layer is a 5-class fully connected layer, with Softmax outputting the stacking level probability.

[0129] In some embodiments, the vulnerability regression model is a trained multilayer perceptron model, trained in the following manner:

[0130] Collect a training dataset, which includes at least: multidimensional feature data of the cargo and historical damage records;

[0131] Based on the historical damage records, a normalized vulnerability score label is generated;

[0132] Using the multidimensional feature data as input and the corresponding vulnerability score label as the supervision target, the multilayer perceptron model is trained using the mean squared error loss function or the Hubble loss function to obtain the trained vulnerability regression model.

[0133] In practice, a training dataset is first constructed, which mainly consists of two parts: one is multidimensional feature data of goods extracted from historical operations (such as geometric dimensions, weight, packaging materials, etc.); the other is historical transportation damage records corresponding to these goods, which are used to reflect their actual damage in past operations.

[0134] Based on historical damage records, a label f ​​is generated for each cargo sample for supervised learning. Specifically, the historical damage rate is obtained by aggregating the number of historical damages / total number of shipments according to the cargo number, and the historical damage rate is normalized to a preset numerical range, such as [0,1], thereby obtaining a quantified vulnerability score label f, where f = damage rate normalized to [0,1].

[0135] Next, the multidimensional feature data of the goods (such as packaging material s, historical damage rate, size ratio (H / W), center of gravity height g_z, whether it is cold chain t, whether it is dangerous goods DG, etc.) are used as model input, and the corresponding vulnerability score label f ​​is used as the regression prediction target. The multilayer perceptron model is trained under supervision. During the training process, the mean squared error (MSE) loss function or Huber loss function is selected to avoid the influence of outliers.

[0136] The network structure of the vulnerability regression model includes a multilayer perceptron (MLP) structure, with an input layer, three hidden layers (512, 256, 128), and an output layer, where the output layer uses the sigmoid activation function.

[0137] After training, the vulnerability regression model can predict the vulnerability score f of the input cargo based on the cargo's feature data. The vulnerability score f∈[0,1], thereby determining whether the cargo is allowed to be subjected to pressure.

[0138] In actual palletizing decisions, if the vulnerability score f of a certain item is greater than 0.6, it is marked as prohibited from being subjected to pressure by the lower layer, thereby proactively avoiding the risk of pressure damage in the stacking planning.

[0139] In some embodiments, an attention mechanism is incorporated into the vulnerability regression model to highlight key features that affect vulnerability (such as historical damage rate) and improve the model’s predictive accuracy.

[0140] In some embodiments, the center of gravity stability prediction model is a neural network model based on the Transformer architecture, trained using the following method:

[0141] Multiple cargo stacking scenarios are generated based on a physics simulation engine, and the overall center of gravity offset and stability coefficient of the stack before and after the cargo is placed in each stacking scenario are recorded.

[0142] Using the current stacking state of the stacking scenario, the characteristics of the goods to be placed, and the current stacking center of gravity as inputs, and the center of gravity offset and the stability coefficient as supervision targets, the center of gravity offset branch is trained using the mean absolute error loss function, and the stability coefficient branch is trained using the binary cross-entropy loss function or the mean square error loss function, to obtain a trained center of gravity stability prediction model.

[0143] In practice, a large number of stacking scenes are first generated using a physics simulation engine (such as Unity + PhysX or PyBullet). In each scene, the change in the center of gravity of the entire stack before and after the goods to be loaded are recorded to obtain the actual center of gravity offset ΔCG = (Δx, Δy, Δz). Based on mechanical principles, such as torque balance and overturning angle, the corresponding stability coefficient S is obtained, thus forming a training sample set with accurate physical labels.

[0144] Then, the current stacking state (e.g., the pose, mass, and center of gravity of the placed goods), the pose (x, y, z, θ), mass m_i, effective bearing area A_i of the goods to be placed i, and the current stack center of gravity CG_current are used as inputs. The actual center of gravity offset ΔCG and stability coefficient S are used as dual-path supervision targets to train the neural network based on the Transformer architecture end-to-end. The "current stacking state" is used as the context encoding. The network outputs two prediction results, namely the offset of the overall center of gravity ΔCG = (Δx, Δy, Δz) and the stability coefficient S ∈ [0,1].

[0145] During training, the mean absolute error loss function (L1 Loss) is used to optimize the centroid offset prediction branch; for the stability coefficient prediction branch, the binary cross-entropy loss function (such as binarized "stable / unstable" classification) or the mean squared error loss function (regression MSE) continuous value regression is selected for optimization according to the modeling method.

[0146] After training, the center of gravity stability prediction model can predict the offset ΔCG of the overall center of gravity and the stability coefficient S after the cargo is stacked, based on the cargo's pose, mass and contact area. Generally, S < 0.8 indicates an unstable state, and S ≥ 0.8 indicates a stable state. When S < 0.8, the weight is reduced.

[0147] In some embodiments, the outputs of the stackability prediction model, vulnerability regression model, and center of gravity stability prediction model are further integrated to generate a normalized prior weight vector, as shown in the formula:

[0148] Weights = [w_1·S_i + w_2·(1 f) + w_3·S];

[0149] Where Weights represents the overall weights, S_i represents the stacking level output by the stacking prediction model, f represents the vulnerability score output by the vulnerability regression model, S represents the stability coefficient output by the centroid stability prediction model, and w_1, w_2, and w_3 represent the weight coefficients of the corresponding items.

[0150] This comprehensive weighting integrates the intelligent evaluation results of cargo in three dimensions: stackability, compressive vulnerability, and stacking stability. These results are used as prior knowledge input into the subsequent multi-objective optimization model to obtain a stacking scheme that achieves the optimal balance between space utilization, safety, and operational efficiency.

[0151] In some embodiments, a hybrid optimization algorithm is used when solving the multi-objective optimization model. The hybrid optimization algorithm executes the following stages sequentially:

[0152] In the reinforcement learning pre-search phase, reinforcement learning is used to search the action space consisting of cargo placement actions to generate an initial set of candidate stacking schemes.

[0153] In the heuristic local optimization stage, the schemes in the initial candidate stacking scheme set are locally adjusted based on preset heuristic rules to obtain an optimized candidate stacking scheme set; wherein, the local adjustment includes at least: swapping the positions of adjacent goods, adjusting the pose of goods, and adjusting the positions of goods whose center of gravity height is higher than a preset threshold to balance the overall center of gravity.

[0154] In the mathematical programming refinement stage, the stacking problem is modeled as a mixed integer programming model, and the optimized candidate stacking scheme set is used as the initial solution to solve the problem and output the final stacking scheme.

[0155] In implementation, the objective function is established using the following formula:

[0156] J = α·U β·|ΔCG| γ·P δ·T;

[0157] Where: U is space utilization, ΔCG is center of gravity deviation, P is instability penalty, and T is operation cost.

[0158] The constraints include:

[0159] Geometric constraint: Goods must not overlap;

[0160] Pressure constraint: Interlayer pressure ≤ σmax;

[0161] Center of gravity constraint: |ΔCG| ≤ εx, εy;

[0162] Safety constraints: Dangerous goods and cold chain goods are separated into different zones;

[0163] Binding constraint: The binding path is continuous and does not intersect.

[0164] During the solution process, a joint solver combining reinforcement learning, heuristics, and mathematical programming (MIP) is invoked to generate the optimal stacking scheme. Specifically:

[0165] Phase 1: Reinforcement Learning Pre-Search

[0166] Input: Cargo feature vector F_i, comprehensive weight W, current stacking state (list of placed cargo, remaining space grid);

[0167] Action space: For each item to be placed, define the action set as {position(x,y,z), rotation angle θ, whether to flip};

[0168] Status representation: The current loading space occupancy status is encoded using a three-dimensional voxel mesh (e.g., 100×100×50 elements);

[0169] Reward function:

[0170] Positive reward: Space utilization rate U increases → +α·ΔU;

[0171] Negative Reward: Center of Gravity Shift |ΔCG| Increase → β·|ΔCG|;

[0172] Negative Reward: Violation of stress / safety constraints → γ·P;

[0173] Negative reward: Increased complexity of the job path → δ·T;

[0174] Algorithm selection: Use the Proximal Policy Optimization (PPO) or Soft Actor-Critic (SAC) algorithm to train the policy network to output the optimal action probability distribution;

[0175] Output: Generate a set of candidate stacking schemes (such as Top-K sorting) as the initial solution for the next stage.

[0176] Phase Two: Heuristic Local Refinement

[0177] Input: Top-K solutions for reinforcement learning output;

[0178] Optimization goal: While maintaining the overall structure, fine-tune the position and angle of the goods to further improve space utilization and stability;

[0179] Algorithm implementation:

[0180] A "greedy swap" strategy is adopted: try swapping the positions of adjacent goods, and accept the swap if the objective function J increases;

[0181] A "local perturbation" strategy is adopted: the coordinates and angles of a single item are randomly fine-tuned within a range of ±5mm, while retaining the optimal solution;

[0182] Introducing the "center of gravity alignment" heuristic: forcing high-center-of-gravity goods to move towards the central area, reducing overall offset;

[0183] Constraint checks: After each adjustment, immediately verify hard constraints such as geometric overlap, bearing capacity, and safety zoning;

[0184] Output: The optimized set of candidate solutions, retaining the top N optimal solutions, for example, the top 3.

[0185] Phase 3: Mathematical Programming Exact Solver (MIP)

[0186] Input: The heuristically optimized Top-3 solution;

[0187] Modeling method:

[0188] Model the stacking problem as a mixed integer programming (MIP) problem;

[0189] Define decision variables:

[0190] x_i, y_i, z_i: The placement coordinates of cargo i;

[0191] r_i: rotation state (e.g., 0~3 represents 0°, 90°, 180°, 270°);

[0192] b_ij: Whether cargo i is stacked on cargo j (0 / 1);

[0193] Objective function: Maximize J = α·U β·|ΔCG| γ·P δ·T;

[0194] Constraints:

[0195] Geometric constraints: i≠j, goods i and j do not overlap in three-dimensional space;

[0196] Pressure constraint: ∑_{j∈below(i)} m_j·g ≤ σmax(i);

[0197] Center of gravity constraint: |CG_x CG_0| ≤ ε_x, |CG_y CG_0| ≤ ε_y;

[0198] Safety constraints: Dangerous goods and cold chain goods must be located in separate zones (restrained by zone number).

[0199] Binding constraints: Path continuity is modeled using graph theory (such as TSP variants);

[0200] Solver selection: Use a commercial solver (such as Gurobi, CPLEX) or an open-source solver (such as SCIP).

[0201] Solution strategy:

[0202] Using a heuristic solution as the initial solution accelerates convergence;

[0203] Set a time limit (e.g., 30 seconds), and if convergence is not achieved, return the current optimal solution;

[0204] Output: The final optimal stacking scheme, including the precise coordinates, rotation angle, stacking relationship, and binding path of each item.

[0205] The objective weight coefficients (α:β:γ:δ) in the multi-objective optimization model can be dynamically adjusted based on actual work performance to optimize stacking schemes for different work scenarios. When space utilization is low, α (space utilization weight) is increased; when stability or center of gravity deviation exceeds the limit, β (center of gravity penalty weight) is increased; if the rework rate is high, γ (instability penalty weight) is increased; if the work takes too long, δ (work cost weight) is increased. Through an adaptive weight update mechanism, a long-term balance between efficiency, safety, and stability is ensured.

[0206] In some embodiments, the backtracking mechanism is a local backtracking, including:

[0207] When the real-time safety check fails, the stacking scheme for unloaded goods is replanned based on the actual state of the currently loaded goods.

[0208] During implementation, the physical simulation module calculates the overall center of gravity using the following formula:

[0209] CG_x = Σ(m_i·x_i) / Σm_i, CG_y = Σ(m_i·y_i) / Σm_i; where m_i represents the mass of the goods, and x_i and y_i represent the position coordinates of the goods, respectively.

[0210] If the deviation is > ±30mm, a local backtracking is triggered, specifically:

[0211] First, try to keep most of the loaded goods unchanged and only replan the unloaded goods. If a feasible solution can be generated under the constraints, then continue to execute the subsequent loading.

[0212] If it fails, the backtracking scope can be gradually expanded. For example, the pose of the last N placed goods can be adjusted or moved back to the queue to be planned, and then replanned together with the unloaded goods. If successful, loading can continue.

[0213] If a feasible solution cannot be generated through local adjustments within the time limit, it indicates that the current accumulated error or constraint conflict cannot be resolved through local adjustments, and a completely new global optimization process can be carried out based on the original data of the goods.

[0214] The backtracking mechanism in this application can be adaptively executed according to the degree of deviation and constraint conflict of the actual operation, so as to maintain the continuity of operation as much as possible while ensuring the safety and feasibility of the stacking scheme.

[0215] In some embodiments, during the execution phase, data from the robotic arm's sensors are collected to monitor placement error ≤3mm, time consumption, and temperature, and to record and store indicators such as space utilization, weight deviation, rework rate, and stability.

[0216] Incremental learning is used to update model parameters, including cargo stacking level and compatibility prediction parameters; the vulnerability regression model is updated based on the latest damage and packaging samples; and the center of gravity prediction and stability model are optimized by combining the actual center of gravity shift data after the operation.

[0217] In some implementations, reinforcement learning normalizes the action selection probability, making the algorithm more biased towards historically high-yield actions.

[0218] The structure, inputs, outputs, and training objectives of the three sub-models in the deep learning model group in this application are all designed for the "Aerospace ULD Assembly" task.

[0219] This application uses machine learning algorithms to automatically generate stacking strategies, which can improve warehouse space utilization by 10%–20%.

[0220] This application significantly reduces the risk of imbalance and collapse through a center of gravity balance optimization algorithm, which meets aviation transport safety standards;

[0221] This application can automatically recommend the best stacking order and placement, reducing the average board assembly time by more than 30%;

[0222] This application can continuously iterate and optimize the model based on historical operational results to adapt to different routes and cargo types;

[0223] This application can also be applied to other air cargo centers, port warehouses, and smart logistics hubs.

[0224] Based on the same inventive concept, this application also provides a machine learning-based intelligent board assembly system, comprising:

[0225] The data acquisition module is used to acquire multidimensional feature data of the goods to be loaded;

[0226] The feature engineering module is used to combine features from the multidimensional feature data to generate a feature vector for each cargo, and to establish a compatibility matrix and physical constraint parameters between cargoes based on the multidimensional feature data.

[0227] The learning prediction module is used to input the cargo feature vector into a pre-trained deep learning model group to obtain the prediction result; wherein, the deep learning model group includes: a stackability prediction model, used to predict the stacking level and order between any two cargoes; a vulnerability regression model, used to predict the vulnerability score of each cargo; and a center of gravity stability prediction model, used to predict the overall center of gravity offset and stability coefficient after stacking based on the current stacking state and the characteristics of the cargo to be placed.

[0228] The pallet planning module is used to construct a multi-objective optimization model with the objectives of space utilization, center of gravity balance, stability and operation efficiency based on the prediction results, the compatibility matrix between goods and physical constraint parameters, and to solve the multi-objective optimization model to obtain the stacking scheme.

[0229] The safety verification module is used to perform real-time safety verification of the overall center of gravity and stability of the stacking scheme. If the verification passes, the stacking scheme is converted into a work instruction and sent to the execution device to perform loading; if the verification fails, a backtracking mechanism is triggered to replan the stacking scheme.

[0230] The execution feedback module is used to collect the actual operation data of the execution device during the loading process, and to incrementally update the parameters of the deep learning model group and the multi-objective optimization model based on the actual operation data.

[0231] Specifically, such as Figure 3 As shown, the data acquisition module (corresponding to the data acquisition layer in the system architecture) is used to acquire geometric, physical, packaging, safety, and center of gravity parameters of the goods. For example, through size sensors, weight sensors, and label recognition devices, the system automatically extracts the goods' length L, width W, height H, mass m, material s, fragility f, hazardous goods category DG, cold chain temperature zone t, and center of gravity coordinates g=(gx, gy, gz), providing the input basis for subsequent algorithms.

[0232] It can also obtain environmental parameters such as temperature, humidity, and air pressure according to actual needs.

[0233] The feature engineering module and the learning prediction module together constitute the intelligent learning layer of the system.

[0234] The feature engineering module normalizes and combines features from the collected data to generate cargo feature vectors and compatibility matrices. At this stage, the system calculates the layer bearing capacity threshold table and friction coefficient matrix, providing prior physical constraints for the model.

[0235] The learning and prediction module is the core intelligent unit of the system, containing three deep learning sub-models:

[0236] Stackability prediction model (StackabilityNet) is used to predict the stacking level and order of goods;

[0237] The vulnerability regression model (FragilityReg) is used to regress and predict the vulnerability score f of cargo.

[0238] The center of gravity stability prediction model (CGPredictor) is used to predict the center of gravity offset ΔCG and stability coefficient S during the stacking process.

[0239] Each model is trained using a combination of supervised learning and reinforcement learning, and its prediction results are used as the weight input for the optimization module.

[0240] The pallet planning module corresponds to the optimization planning layer. Based on the prediction results, a multi-objective optimization model is established, with space utilization U, center of gravity deviation ΔCG, instability penalty P, and operational cost T as objective functions. The optimal stacking scheme is solved under constraints such as geometric feasibility, pressure limitations, hazardous materials isolation, and independent cold chain. The solution employs a hybrid algorithm combining machine learning prediction with heuristic search, mathematical programming, and reinforcement learning, balancing global optimality and real-time performance.

[0241] The security verification module and the execution feedback module together constitute the system's verification execution layer.

[0242] The safety verification module is used to calculate the overall center of gravity coordinates in real time after each cargo placement.

[0243] If the center of gravity deviates from the safety threshold or the layer bearing capacity exceeds the limit, the system immediately triggers the backtracking mechanism to re-plan the local scheme in order to ensure the stability and safety of the stack.

[0244] The execution feedback module transforms the optimization results into executable work instructions, outputting information such as: cargo number, placement coordinates, rotation angle, reinforcement method, and strapping path, which are then directly sent to the automated palletizing robot arm or loading control system. Simultaneously, the system records operation time, weight imbalance rate, and rework status.

[0245] Finally, the online learning module corresponds to the system's feedback self-learning layer. Through monitoring the work process and data feedback, it drives the model's online learning and parameter updates, enabling continuous system optimization. The system employs an incremental learning mechanism to dynamically adjust the model parameters and target weights (α:β:γ:δ), allowing the algorithm to continuously optimize under different cargo types, flight schedules, and seasonal scenarios, maintaining long-term performance stability.

[0246] The overall structure of this application adopts a closed-loop design of "data acquisition --> intelligent learning --> optimized planning --> verification and execution --> feedback self-learning". It consists of multiple functional modules, which are interconnected through a unified data interface and control logic to form a continuous information flow and decision flow, forming a complete "perception-learning-decision-execution-feedback" closed-loop intelligent system. This not only significantly improves the efficiency and safety of airport cargo loading, but also adapts to complex and ever-changing cargo scenarios through continuous self-learning, possessing the system's self-evolution capability, and providing important technical support for smart logistics and aviation ground handling automation.

[0247] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0248] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A machine learning-based intelligent board assembly method, characterized in that, include: Obtain multidimensional feature data of the cargo to be loaded; The multidimensional feature data is combined to generate a feature vector for each cargo. Based on the multidimensional feature data, a compatibility matrix and physical constraint parameters between cargoes are established, including: Based on the geometric dimensions and packaging materials in the multidimensional feature data, the layer bearing capacity threshold of each item is obtained; The coefficient of friction between the stacked goods is obtained based on the packaging material of the contact surfaces of the goods. By combining the layer bearing capacity threshold, the friction coefficient, and the preset safety isolation rules, a compatibility matrix between goods is generated; wherein, the elements in the compatibility matrix are used to characterize whether stacking is allowed between corresponding pairs of goods; The cargo feature vector is input into a pre-trained deep learning model set to obtain a prediction result; wherein, the deep learning model set includes: A stackability prediction model is used to predict the stacking level and order between any two goods; wherein the stacking level adopts a hierarchical system, and different levels are used to characterize the suitability of goods for stacking from prohibited to optimal. A vulnerability regression model is used to predict the vulnerability score for each item. The center of gravity stability prediction model is used to predict the overall center of gravity offset and stability coefficient after stacking, based on the current stacking state and the characteristics of the goods to be placed. Based on the prediction results, the cargo compatibility matrix, and physical constraint parameters, a multi-objective optimization model is constructed with space utilization, center of gravity balance, stability, and operational efficiency as objectives. The multi-objective optimization model is then solved to obtain the stacking scheme. The stacking scheme is subjected to real-time safety verification of its overall center of gravity and stability. If the verification passes, the stacking scheme is converted into a work instruction and sent to the execution device to perform loading. If the verification fails, a backtracking mechanism is triggered to replan the stacking scheme. The actual operation data of the execution device during the loading process is collected, and the parameters of the deep learning model group and the multi-objective optimization model are incrementally updated based on the actual operation data.

2. The intelligent board assembly method based on machine learning according to claim 1, characterized in that, Before performing feature combination on the multidimensional feature data, the multidimensional feature data is preprocessed, including: Unit standardization and anomaly detection are used to remove missing values ​​and outliers.

3. The intelligent board assembly method based on machine learning according to claim 1, characterized in that, The stacking prediction model is a trained deep neural network model, which is trained using the following method: Collect a training dataset, which includes positive samples and negative samples. The positive samples are cargo pairs that were successfully stacked in historical operations, and the negative samples are cargo pairs that failed to stack in historical operations. Label each sample in the training dataset with a stacking level label; Using the feature vectors and relative pose features of the cargo pair as input, and the corresponding stacking level label as the supervision target, the stacking prediction model is trained using the weighted cross-entropy loss function to obtain the trained stacking prediction model. The weighted cross-entropy loss function is configured such that a higher misclassification penalty weight is assigned to the prohibited stacking category than to other stackable categories.

4. The intelligent board assembly method based on machine learning according to claim 1 or 3, characterized in that, The stacking prediction model is constructed using the following method: Extract the geometric dimensions of the goods using convolutional neural network layers; Multidimensional structured features of goods are fused through the Transformer encoder layer; The stacking level probability is output based on the fused features through a fully connected layer and a Softmax function.

5. The intelligent board assembly method based on machine learning according to claim 1, characterized in that, The vulnerability regression model is a trained multilayer perceptron model, which is trained in the following way: Collect a training dataset, which includes at least: multidimensional feature data of the cargo and historical damage records; Based on the historical damage records, a normalized vulnerability score label is generated; Using the multidimensional feature data as input and the corresponding vulnerability score label as the supervision target, the multilayer perceptron model is trained using the mean squared error loss function or the Hubble loss function to obtain the trained vulnerability regression model.

6. The intelligent board assembly method based on machine learning according to claim 1, characterized in that, The center of gravity stability prediction model is a neural network model based on the Transformer architecture, which is trained using the following method: Multiple cargo stacking scenarios are generated based on a physics simulation engine, and the overall center of gravity offset and stability coefficient of the stack before and after the cargo is placed in each stacking scenario are recorded. Using the current stacking state of the stacking scenario, the characteristics of the goods to be placed, and the current stacking center of gravity as inputs, and the center of gravity offset and the stability coefficient as supervision targets, the center of gravity offset branch is trained using the mean absolute error loss function, and the stability coefficient branch is trained using the binary cross-entropy loss function or the mean square error loss function, to obtain a trained center of gravity stability prediction model.

7. The intelligent board assembly method based on machine learning according to claim 1, characterized in that, When solving the multi-objective optimization model, a hybrid optimization algorithm is used, which executes the following stages sequentially: In the reinforcement learning pre-search phase, reinforcement learning is used to search the action space consisting of cargo placement actions to generate an initial set of candidate stacking schemes. In the heuristic local optimization stage, the schemes in the initial candidate stacking scheme set are locally adjusted based on preset heuristic rules to obtain an optimized candidate stacking scheme set; wherein, the local adjustment includes at least: swapping the positions of adjacent goods, adjusting the pose of goods, and adjusting the positions of goods whose center of gravity height is higher than a preset threshold to balance the overall center of gravity. In the mathematical programming refinement stage, the stacking problem is modeled as a mixed integer programming model, and the optimized candidate stacking scheme set is used as the initial solution to solve the problem and output the final stacking scheme.

8. The intelligent board assembly method based on machine learning according to claim 1, characterized in that, The backtracking mechanism is a local backtracking, including: When the real-time safety check fails, the stacking scheme for unloaded goods is replanned based on the actual state of the currently loaded goods.

9. A machine learning-based intelligent board assembly system, characterized in that, include: The data acquisition module is used to acquire multidimensional feature data of the goods to be loaded; The feature engineering module is used to combine features from the multidimensional feature data to generate a feature vector for each cargo, and based on the multidimensional feature data, to establish a compatibility matrix and physical constraint parameters between cargoes, including: Based on the geometric dimensions and packaging materials in the multidimensional feature data, the layer bearing capacity threshold of each item is obtained; The coefficient of friction between the stacked goods is obtained based on the packaging material of the contact surfaces of the goods. By combining the layer bearing capacity threshold, the friction coefficient, and the preset safety isolation rules, a compatibility matrix between goods is generated; wherein, the elements in the compatibility matrix are used to characterize whether stacking is allowed between corresponding pairs of goods; The learning prediction module is used to input the cargo feature vector into a pre-trained deep learning model set to obtain prediction results. The deep learning model set includes: a stackability prediction model for predicting the stacking level and order between any two cargoes, wherein the stacking level adopts a hierarchical system, with different levels characterizing the suitability of cargoes for stacking from prohibited to optimal; a vulnerability regression model for predicting the vulnerability score of each cargo; and a center of gravity stability prediction model for predicting the overall center of gravity offset and stability coefficient after stacking based on the current stacking state and the characteristics of the cargo to be placed. The pallet planning module is used to construct a multi-objective optimization model with the objectives of space utilization, center of gravity balance, stability and operation efficiency based on the prediction results, the compatibility matrix between goods and physical constraint parameters, and to solve the multi-objective optimization model to obtain the stacking scheme. The safety verification module is used to perform real-time safety verification of the overall center of gravity and stability of the stacking scheme. If the verification passes, the stacking scheme is converted into a work instruction and sent to the execution device to perform loading; if the verification fails, a backtracking mechanism is triggered to replan the stacking scheme. The execution feedback module is used to collect the actual operation data of the execution device during the loading process, and to incrementally update the parameters of the deep learning model group and the multi-objective optimization model based on the actual operation data.

Images