Automated sorting method and system based on artificial intelligence

Abstract

The present application relates to the field of automatic sorting control technology, and discloses an automated sorting method and system based on artificial intelligence. The method includes: performing multi-dimensional data collection and processing on packages to obtain feature vectors; inputting the feature vectors into a hierarchical control system for analysis to obtain classification instructions; identifying target channels and timings based on classification instructions and channel loads; assigning tasks to robotic arms according to channel numbers and timings; verifying and calculating the matching degree through dual sensors, triggering reclassification when it is lower than a threshold; and adjusting control parameters based on evaluation data to generate a closed-loop control set. The present application constructs an automated sorting method based on artificial intelligence to achieve a closed-loop control mechanism for multimodal data collection and fusion, hierarchical classification control, dynamic task allocation, real-time quality monitoring, and parameter self-optimization, thereby significantly improving the accuracy, efficiency, and adaptability of the sorting system.

Description

Technical Field

[0001] The present application relates to the field of automatic sorting control technology, and in particular to an automatic sorting method and system based on artificial intelligence. Background Art

[0002] In modern logistics and e-commerce, efficient and accurate parcel sorting systems are crucial for ensuring delivery efficiency and customer satisfaction. Traditional parcel sorting methods rely primarily on manual operations or simple mechanical automation equipment, such as sorting belts, chutes, and pushers. These methods suffer from inefficiencies when handling large quantities of diverse parcels. With technological advancements, semi-automatic sorting systems based on barcode recognition are becoming increasingly popular. These systems determine sorting routes by scanning barcode information on packages, improving basic sorting efficiency. In recent years, some advanced sorting centers have begun adopting machine vision-based sorting technology, using camera arrays to identify package features and then classify and process them based on pre-set rules. While these technologies have made some progress in standard parcel processing, their intelligent decision-making capabilities in complex environments remain limited.

[0003] However, existing sorting technologies have many shortcomings. First, most systems lack multimodal perception capabilities and rely solely on single sensor data (such as barcodes or images). They are unable to fully perceive the physical characteristics of packages, which makes it easy to make mistakes when processing packages with irregular shapes, unclear labels, or damaged packaging. Second, existing technologies generally use static rules or simple algorithms to make decisions and lack intelligent learning and adaptive capabilities. When faced with dynamic factors such as fluctuations in logistics flow and changes in package types, they are unable to effectively adjust sorting strategies, resulting in large fluctuations in system efficiency. Third, traditional sorting systems lack effective quality monitoring and closed-loop optimization mechanisms. Once a sorting error occurs, it is difficult to automatically discover and correct it, and it is even more impossible to learn and improve from the error. In addition, existing technologies also have shortcomings in the collaborative control of multiple robotic arms, making it difficult to achieve load balancing and efficient task allocation, resulting in some equipment being overloaded while other equipment is idle, and low resource utilization. Summary of the Invention

[0004] The present application provides an artificial intelligence-based automated sorting method and system, which is used to realize a closed-loop control mechanism of multimodal data acquisition and fusion, hierarchical classification control, dynamic task allocation, real-time quality monitoring, and parameter self-optimization by constructing an artificial intelligence-based automated sorting method, thereby significantly improving the accuracy, efficiency, and adaptability of the sorting system.

[0005] In a first aspect, the present application provides an artificial intelligence-based automated sorting method, the artificial intelligence-based automated sorting method comprising: collecting and processing multi-dimensional data of packages on a conveyor belt using a high-resolution camera array, a depth sensor, and a weight sensor to obtain a package feature vector;

[0006] The package feature vector is input into a hierarchical control system for analysis of control parameters to obtain a package classification control instruction; sorting status identification is performed based on the package classification control instruction and the current load status of each sorting channel to obtain a target sorting channel number and its execution timing; according to the target sorting channel number and execution timing, a task allocation algorithm is used to allocate a grasping task to an industrial robot arm; the package transferred by the industrial robot arm is verified using a dual sensor array at the entrance and exit, and a weighted sum of feature similarities is calculated as a matching score. When the matching score is lower than a preset threshold, a lightweight neural network is triggered for reclassification to obtain sorting quality assessment data; and the feedback gain parameter, decision threshold, and execution control timing of the hierarchical control system are adjusted according to the sorting quality assessment data to generate a closed-loop control parameter set.

[0007] In a second aspect, the present application provides an artificial intelligence-based automated sorting system, the artificial intelligence-based automated sorting system comprising:

[0008] The acquisition module is used to collect and process multi-dimensional data of packages on the conveyor belt using a high-resolution camera array, depth sensor, and weight sensor to obtain package feature vectors;

[0009] An input module, configured to input the package feature vector into a hierarchical control system to analyze control parameters and obtain a package classification control instruction;

[0010] an identification module, configured to identify the sorting status based on the parcel classification control instruction and the current load status of each sorting channel, and obtain the target sorting channel number and its execution sequence;

[0011] An allocation module is used to allocate grasping tasks to the industrial robot arm through a task allocation algorithm according to the target sorting channel number and execution sequence;

[0012] A classification module is used to verify the parcels transferred by the industrial robot arm through dual sensor arrays at the entrance and exit, calculate the weighted sum of feature similarities as a matching score, and trigger a lightweight neural network to reclassify when the matching score is lower than a preset threshold to obtain sorting quality assessment data;

[0013] The control module is used to adjust the feedback gain parameters, decision thresholds and execution control timing of the hierarchical control system according to the sorting quality evaluation data to generate a closed-loop control parameter set.

[0014] In a third aspect, an artificial intelligence-based automated sorting device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the artificial intelligence-based automated sorting device executes the above-mentioned artificial intelligence-based automated sorting method.

[0015] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned artificial intelligence-based automated sorting method.

[0016] The technical solution provided in this application uses multi-dimensional data acquisition and processing technology to obtain package feature vectors, achieving comprehensive perception of the package's physical characteristics and significantly improving classification and identification accuracy. The step of inputting standardized package feature vectors into the hierarchical control system to analyze control parameters fully leverages the advantages of artificial intelligence algorithms in feature extraction and pattern recognition, enabling the system to learn optimal classification strategies from large amounts of historical data and adapt to the characteristic differences of different types of packages. The design of sorting status recognition based on package classification control instructions and the current load status of each sorting channel introduces a dynamic load balancing mechanism, which solves the problem of uneven resource allocation in traditional sorting systems and improves overall throughput. The method of assigning grasping tasks to industrial robotic arms through a task allocation algorithm realizes the collaborative operation and intelligent scheduling of multiple robotic arms, reduces equipment idle time, and improves hardware resource utilization. The quality monitoring mechanism of parcels transferred by the industrial robotic arms is verified by a dual sensor array, establishing real-time error correction capabilities, significantly reducing the sorting error rate and ensuring sorting quality. The closed-loop control design that dynamically adjusts the parameters of the hierarchical control system based on sorting quality assessment data gives the system the ability to continuously self-optimize, enabling it to adapt to changes in the logistics environment and fluctuations in package characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 This is a schematic diagram of an embodiment of an automated sorting method based on artificial intelligence in an embodiment of the present application;

[0019] Figure 2 This is a schematic diagram of an embodiment of an automated sorting system based on artificial intelligence in an embodiment of the present application;

[0020] Figure 3It is a schematic block diagram of the structure of an automated sorting device based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The embodiments of the present application provide an automated sorting method and system based on artificial intelligence. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0022] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the automated sorting method based on artificial intelligence includes:

[0023] Step S101: Use a high-resolution camera array, depth sensor, and weight sensor to collect and process multi-dimensional data of the package on the conveyor belt to obtain a package feature vector;

[0024] Step S102: Input the package feature vector into the hierarchical control system to analyze the control parameters and obtain the package classification control instructions;

[0025] Step S103: Identify the sorting status based on the parcel classification control instruction and the current load status of each sorting channel to obtain the target sorting channel number and its execution sequence;

[0026] Step S104: assigning a grabbing task to the industrial robot arm using a task assignment algorithm according to the target sorting channel number and execution sequence;

[0027] Step S105: Parcels transferred by the industrial robot are verified by dual sensor arrays at the entrance and exit, and a weighted sum of feature similarities is calculated as a matching score. When the matching score falls below a preset threshold, a lightweight neural network is triggered for reclassification to obtain sorting quality assessment data.

[0028] Step S106: Adjust the feedback gain parameters, decision thresholds, and execution control timing of the hierarchical control system according to the sorting quality evaluation data to generate a closed-loop control parameter set.

[0029] It is understandable that the execution subject of this application can be an automated sorting system based on artificial intelligence, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0030] Specifically, a high-resolution camera array, comprised of at least three industrial cameras, captures multi-angle images of the package from the top and sides. A depth sensor emits infrared structured light and receives reflected signals to generate three-dimensional point cloud data of the package. This point cloud records the spatial coordinates of every point on the package's surface. A weight sensor measures the package's mass using the strain gauge principle. This raw data undergoes Gaussian filtering to remove noise, and perspective transformation to correct for image viewing angle deviation. Feature extraction is then performed, including edge features, color histograms, and texture features. These features are then fused into a multidimensional feature vector representing the package's physical characteristics. A dual-light source illumination system uses a two-to-one brightness ratio between the primary and auxiliary light sources to ensure image clarity. As a typical express package passes through the conveyor belt, the image data captured by the camera, the point cloud data obtained by the depth sensor, and the mass data measured by the weight sensor undergo data registration, associating these heterogeneous data with the same package based on timestamps and spatial location information.

[0031] The package feature vector is input into the hierarchical control system for analysis. The hierarchical control system consists of a main control engine and secondary specialized controllers. The main control engine is designed based on a residual connection structure and includes a spatial pyramid pooling module to process package features of varying sizes. After receiving the feature vector, the main control engine maps it to the primary classification space and outputs a probability distribution for the primary classification category. For example, when the system processes a package, the main control engine classifies it as an electronics product with a probability confidence of 0.95. The system then selects a secondary specialized controller dedicated to electronics for further analysis. The secondary controller uses a densely connected structure to perform more detailed classification of the package, such as further classifying electronics into mobile phones, tablets, etc. The control system generates classification confidence scores by combining a weighted fusion of cross entropy and focal loss, assigning higher weights to difficult-to-classify samples. These scores are then compensated by Bayesian probability to calibrate the classification results and reduce classification bias.

[0032] Sorting decisions are made based on package classification control instructions and sorting channel load status. First, each sorting channel is monitored in real time, collecting the current package quantity, processing rate, and buffer occupancy rate of each channel to construct a sorting channel load status matrix. Package category attributes and processing priority are extracted based on the package classification control instructions, and an initial mapping relationship between packages and potential target channels is established. A comprehensive sorting score is calculated for each candidate channel, taking into account package category matching, channel load rate, and historical sorting success rate. A Bayesian decision rule is used to select the channel with the highest score as the sorting target. The estimated time for the package to arrive at the sorting point is calculated based on the conveyor belt speed. Combined with the robot arm's motion time constraints, a sorting execution window is generated, and a timing optimization algorithm is applied to resolve potential sorting conflicts.

[0033] Grasping tasks are assigned to industrial robotic arms based on the target sorting channel number and execution sequence. First, the status of the industrial robotic arms in the sorting system is detected, and the current position, load, and task queue of each robotic arm are collected to obtain robotic arm status data. The estimated time of arrival of the package at each robotic arm's workspace is calculated, and grasping constraints are constructed based on the physical properties of the package. The task allocation cost is calculated, taking into account time matching, energy consumption, and robotic arm load balance. A task allocation algorithm is applied to the task allocation cost matrix for global optimization, selecting the robotic arm with the lowest cost for each package to be sorted. The robotic arm allocation plan is judged. When the current load of the robotic arm exceeds the threshold, the task is placed in the waiting queue and reallocation is triggered. The robotic arm trajectory planning data is generated and converted into a sequence of motion instructions in the joint space.

[0034] Quality verification is performed on parcels transferred by industrial robotic arms. The first sensor array at the entrance of the sorting channel collects package features and calculates the cosine similarity with the original features to obtain the entrance matching score. The second sensor array at the exit collects exit feature data and calculates the exit matching score. The weighted average of the entrance and exit matching scores is used to obtain a comprehensive matching score. When the matching score falls below the preset threshold, a lightweight neural network is triggered for reclassification. The lightweight neural network consists of three convolutional layers, two pooling layers, and two fully connected layers. Each convolutional layer uses a 3×3 convolution kernel, and the fully connected layers contain 128 and 64 neurons. Sorting quality assessment data is generated based on the consistency of the reclassification results with the original classification results, the comprehensive matching score, and the operating status of the sorting channel.

[0035] Adjust control system parameters based on sorting quality assessment data. Classify and compile sorting quality assessment data, extracting classification error rates, feature deviation values, and response delay values. Build a performance parameter association model and train the association model using operational data. Perform matrix decomposition on the parameter influence matrix to identify the key parameters influencing control performance. Numerically adjust the feedback gain parameters in the hierarchical control system, setting different gain values for different classification scenarios. Correct the decision thresholds in the hierarchical control system, focusing on correcting the thresholds in areas with misclassified samples. Combine the adjusted gain parameters and decision thresholds with the sorting channel load data, and redistribute the control execution timing to form a closed-loop control system. This closed-loop control method continuously improves the accuracy and efficiency of the sorting system through continuous self-adjustment. In actual operation, the sorting system continuously optimizes control parameters based on the sorting results of each batch of packages, forming an intelligent sorting control closed loop.

[0036] In the embodiments of the present application, multi-dimensional data acquisition and processing technology is used to obtain package feature vectors, achieving a comprehensive perception of the package's physical characteristics and significantly improving the accuracy of classification and identification. The step of inputting standardized package feature vectors into the hierarchical control system to analyze control parameters fully leverages the advantages of artificial intelligence algorithms in feature extraction and pattern recognition, enabling the system to learn optimal classification strategies from large amounts of historical data and adapt to the characteristic differences of different types of packages. The design of sorting status identification based on package classification control instructions and the current load status of each sorting channel introduces a dynamic load balancing mechanism, which solves the problem of uneven resource allocation in traditional sorting systems and improves overall throughput. The method of assigning grasping tasks to industrial robotic arms through a task allocation algorithm realizes the collaborative operation and intelligent scheduling of multiple robotic arms, reduces equipment idle time, and improves hardware resource utilization. A quality monitoring mechanism for parcels transferred by industrial robotic arms, verified by a dual sensor array, establishes real-time error correction capabilities, significantly reduces the sorting error rate, and ensures sorting quality. The closed-loop control design that dynamically adjusts the parameters of the hierarchical control system based on sorting quality assessment data gives the system the ability to continuously self-optimize, enabling it to adapt to changes in the logistics environment and fluctuations in package characteristics.

[0037] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0038] The packages moving on the conveyor belt are illuminated by a dual-light source lighting system, with the brightness ratio of the main light source to the auxiliary light source set at two to one to obtain the lighting environment;

[0039] Use a triangulation laser rangefinder to scan the package surface in the lighting environment, construct multi-point three-dimensional point cloud data, and obtain the shape feature data of the package;

[0040] The temperature distribution of the package is scanned by an infrared thermal imaging sensor, and a multi-pixel thermal map is collected to obtain the temperature distribution map of the package;

[0041] Use OCR technology to identify and decode the barcode and text information on the package surface, extract the package's destination code and classification mark, and obtain the package's identification information;

[0042] Perform multimodal feature fusion on shape feature data, temperature distribution map and identification information, generate a multi-dimensional fusion feature matrix through tensor calculation, and obtain multi-dimensional description data of the package;

[0043] Nonlinear dimensionality reduction and statistical filtering are performed on the multidimensional description data, and the high-dimensional features are converted into low-dimensional vector space representation through the t-SNE algorithm to obtain the package feature vector.

[0044] Specifically, regional lighting is achieved through a dual-light source lighting system, which refers to a lighting device consisting of a main light source and an auxiliary light source. The main light source usually uses a high-brightness LED lamp group, installed in the center above the conveyor belt, responsible for providing the main lighting; the auxiliary light source is installed at a 45-degree angle on the side and uses a scattered LED lamp group to fill the shadows caused by the main light source. The brightness ratio of the main light source to the auxiliary light source is set to two to one. The specific operation is to adjust the input power of the two light sources through the PWM (pulse width modulation) controller so that the illumination value of the main light source is twice that of the auxiliary light source. The lighting environment generated by this configuration has two characteristics: one is that the light intensity is moderate to avoid overexposure; the other is that the shadow levels are rich, which is conducive to extracting the three-dimensional information of the package.

[0045] The package surface is scanned using a triangulation laser rangefinder. This precision measuring device operates based on the principle of optical triangulation. A laser emitter emits a laser beam, which strikes the package surface, forming a light spot. This spot is detected by a CCD or CMOS sensor mounted at a fixed angle. Variations in the package surface's height cause the image position of the light spot on the sensor to shift. Trigonometric functions are used to calculate the precise distance from the package surface to the rangefinder based on the light spot offset, the distance between the emitter and receiver, and the receiver's focal length. During the actual scanning process, the laser rangefinder performs a raster scan along the package surface in the XY plane, recording the three-dimensional coordinates (x, y, z) of each grid point to form a point cloud. For a standard-sized package, typically 3,000 to 5,000 points are collected, with a spacing of approximately 5 mm between points, to form a complete point cloud representation of the package surface. This point cloud data is then processed through filtering, denoising, hole filling, and normal vector calculation to convert it into a data structure describing the package's shape.

[0046] An infrared thermal imaging sensor is a device that can detect infrared radiation from the surface of an object and convert it into a temperature image. The sensor contains a microwave-long infrared photodetector array, typically 384×288 or 640×480 pixels, with each pixel corresponding to a temperature measurement point. When a package passes through the sensor, the intensity of infrared radiation emitted at each point on the surface is recorded and converted into absolute temperature values through an internal algorithm to form a temperature distribution heat map. Different types of items have significantly different temperature distribution patterns: normal temperature items exhibit a uniform ambient temperature distribution; electronic devices exhibit hot spots at specific locations; and temperature-controlled pharmaceuticals exhibit characteristics below ambient temperature. The temperature distribution heat map collected by the infrared sensor is a two-dimensional matrix, in which each element represents the temperature value at the corresponding location. These data supplement the physical characteristics of the package.

[0047] Barcodes and text on the package surface are recognized and decoded using OCR technology. OCR (Optical Character Recognition) converts text in images into editable text. In the sorting system, a high-resolution camera first captures a clear image of the package surface. This image is then preprocessed: grayscale conversion to a single channel, binarization to black and white, and denoising to remove any interfering information. After preprocessing, an image segmentation algorithm is used to locate the text and barcode areas. For the barcode area, a barcode decoding algorithm is used to identify the encoded content, primarily including bar width calculation, encoding rule determination, and checksum verification. For the text area, a convolutional neural network model, pretrained to recognize text in various fonts and languages, is used for character recognition. The recognition results include the package's destination code (such as distribution center number or region code) and classification tags (such as item type and handling priority). This textual information is directly linked to the package's routing requirements.

[0048] Multimodal feature fusion of shape feature data, temperature distribution maps, and identification information is a key data processing step. Multimodal feature fusion refers to the process of integrating heterogeneous data from different sensory channels into a unified representation. First, the data is preprocessed and standardized: shape feature data is converted into shape descriptors, such as curvature histograms and volume ratios; temperature distribution maps are converted into temperature feature sequences through downsampling and feature extraction; and identification information is converted into numerical vectors using text embedding techniques. Feature fusion is then performed using tensor computation, a mathematical technique for processing multidimensional data arrays. In this method, a tensor product operation is used to expand feature vectors from different modalities into a common high-dimensional feature space, preserving the interaction between the modalities. The resulting multidimensional feature matrix comprehensively represents the physical, temperature, and semantic characteristics of the package, providing a data foundation for subsequent classification. Nonlinear dimensionality reduction and statistical filtering are then performed on the multidimensional descriptive data, and the high-dimensional features are converted into low-dimensional vectors using the t-SNE algorithm. t-SNE (t-distributed Stochastic Neighbor Embedding) is a nonlinear dimensionality reduction algorithm particularly well suited for visualizing high-dimensional data in a low-dimensional space. The algorithm first calculates the similarity between pairs of data points in the high-dimensional space and converts the distances into conditional probabilities using a Gaussian distribution. It then constructs a similar probability distribution in the low-dimensional space, but using a t-distribution instead of a Gaussian distribution. Finally, it uses gradient descent to minimize the KL divergence (Kullback-Leibler divergence) between the two distributions, optimizing the positions of the data points in the low-dimensional space. Through t-SNE, the fused feature matrix is mapped from a high-dimensional space of potentially hundreds of dimensions to a low-dimensional space of tens of dimensions, preserving the local structural relationships of the data while filtering out redundancy and noise.

[0049] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0050] The package feature vector is input into the main control engine, which contains a residual connection structure and a spatial pyramid pooling module. It performs primary classification mapping on the feature vector and obtains the probability distribution of the main classification category.

[0051] Based on the probability distribution of the main classification category, the corresponding secondary specialized controller is selected. The secondary specialized controller adopts a dense connection structure to classify the feature vector and obtain the probability distribution of the subdivided category.

[0052] Perform a weighted fusion calculation of cross entropy and focal loss on the main classification category probability distribution and the subdivision category probability distribution to obtain the classification confidence score;

[0053] Bayesian probability compensation is performed based on the classification confidence score and historical classification accuracy, and the classification result is calibrated to obtain the calibrated classification result;

[0054] The calibrated classification results are matched and analyzed with the current system physical constraints, including the capacity limits of each sorting channel and the outbound loading capacity, to obtain preliminary sorting instructions.

[0055] The parameters of the preliminary sorting instructions are adjusted through the control rule engine of the hierarchical control system, including instruction priority setting, execution timing arrangement and exception handling plan, to obtain the package classification control instructions.

[0056] Specifically, the main control engine is the core neural network structure used to handle the primary classification of packages. It includes a residual connection structure and a spatial pyramid pooling module. The residual connection structure refers to adding cross-layer connections in a deep neural network, allowing gradients to flow directly through, thus solving the gradient vanishing problem in deep networks. The computational logic of the residual connection is to directly add the input features to the output of several convolutional layers, forming an identity map plus a residual. The spatial pyramid pooling module is a multi-scale feature extraction technology that converts feature maps of different sizes into fixed-length representations by performing pooling operations at different scales. When the 32-dimensional package feature vector is input to the main control engine, it first undergoes multi-layer convolution processing, each layer including residual connections. The spatial pyramid pooling module then captures features at multiple scales. Finally, a fully connected layer maps the features to a predefined primary classification space, outputting the primary classification probability distribution, that is, the probability values of each primary category.

[0057] Selecting the corresponding secondary specialized controller based on the probability distribution of the main classification category is a key step in hierarchical classification. A secondary specialized controller is a dedicated neural network model that performs fine-grained classification on a large category of items. The secondary controller uses a densely connected structure, in which each layer is directly connected to all subsequent layers, enhancing feature reuse and parameter efficiency. Unlike residual connections, dense connections directly connect (concatenate) the features of the previous layer to the input of the subsequent layer instead of adding them. Once the main classification is determined, a corresponding secondary controller is selected from multiple secondary controllers, and the package feature vector and the main classification information are input into the secondary controller. The secondary controller processes these inputs to generate more refined classification results and outputs the probability distribution of the subdivided categories, that is, the probability values of each subcategory under the subdivided category.

[0058] Performing a weighted fusion of cross-entropy and focal loss on the probability distributions of the main classification class and the sub-classes is a key step in assessing classification quality. Cross-entropy loss measures the difference between the predicted distribution and the true distribution and is calculated as the negative of the log-probability of the true label. Focal loss is a variation of cross-entropy that downweights easy-to-classify examples, allowing the model to focus more on difficult-to-classify examples. In this weighted fusion calculation, the cross-entropy loss for the main classification and the focal loss for the sub-classes are weighted averaged according to preset weights to produce a combined loss. The inverse of this loss is converted to a classification confidence score, with higher scores indicating more reliable classifications.

[0059] Bayesian probability compensation based on classification confidence scores and historical classification accuracy is a calibration technique. Bayesian probability compensation is a posterior probability adjustment method based on Bayes' theorem, taking into account both the prior probability (historical accuracy) and the likelihood (current confidence). The calculation combines the current classification confidence with the historical accuracy of the category to adjust the probability of the initial classification result. For example, if a category has been frequently misclassified historically (with low historical accuracy), even if the current confidence is high, the probability value after Bayesian compensation will be appropriately lowered. This calibration mechanism effectively reduces classification bias and produces more reliable calibrated classification results.

[0060] Matching and analyzing the calibrated classification results with the current system's physical constraints is a step to ensure the feasibility of the sorting scheme. Physical constraints include the capacity limits of each sorting channel and the outbound loading capacity, which refers to the actual processing capacity limits of the sorting system's hardware facilities. The sorting channel capacity limit refers to the maximum number of packages that can be processed per channel per unit time; the outbound loading capacity refers to the upper limit of the loading rate of each outbound station. During the matching analysis process, the ideal sorting channel for the package is first determined based on the calibrated classification results; then the current load status of the channel is checked. If it is close to or exceeds the capacity limit, alternative channels are considered; at the same time, the outbound loading capacity is considered to ensure that the sorted packages can be loaded in a timely manner. This process generates preliminary sorting instructions, including the target channel number and approximate timing requirements.

[0061] The final decision-making step involves parameter tuning of preliminary sorting instructions through the hierarchical control system's control rules engine. The control rules engine is a set of program modules that handle business logic and control strategies, refining and optimizing sorting instructions through a series of predefined rules. The parameter tuning process includes three aspects: instruction priority setting, which assigns a priority value to each sorting instruction based on package type and timeliness requirements; execution scheduling, which calculates specific execution times and resolves potential timing conflicts; and exception handling, which develops alternative solutions to address possible execution anomalies. Through these tuning steps, preliminary sorting instructions are converted into detailed package classification control instructions, including execution information.

[0062] For example, a wrapped 32-dimensional feature vector is input to the main control engine and processed by a deep network consisting of five residual blocks, each of which includes multiple convolutional layers and direct connections. The feature vector is abstracted layer by layer in the network, while the original information is preserved through residual connections. Midway through the network, a spatial pyramid pooling module performs pooling operations on the feature map at four scales (1×1, 2×2, 3×3, and 6×6) to generate a multi-scale feature representation. After projection through a fully connected layer, the output is a probability distribution for 10 primary classification categories, with the electronics category having the highest probability of 0.85. Based on this result, a dedicated secondary controller for electronics products is selected for fine-grained classification. This secondary controller employs a four-layer densely connected architecture, with the output of each layer directly connected to the inputs of all subsequent layers. The resulting output probability for the mobile phone and accessories subcategory is 0.78. The primary and fine-grained classification results are calculated using cross-entropy and focal loss, resulting in a combined loss of 0.32, which translates to a classification confidence score of 0.68. A query of historical data revealed a historical accuracy of 0.92 for the mobile phone and accessories category. Using the Bayesian formula, combined with the current confidence level and historical accuracy, the calibrated classification probability was calculated to be 0.83. Physical constraints were checked and the current load of Channel 12 in the Electronics Zone was 65%, within capacity limits. The outbound loading point was also operating normally. A preliminary sorting instruction was generated, specifying the target channel as Channel 12 in the Electronics Zone. Finally, parameters were adjusted using the control rule engine: Because mobile phones are valuable items, a high priority was set; the execution sequence was scheduled to 15 seconds after the current time point; and the exception handling plan was set to the backup channel, Channel 14 in the Electronics Zone. The resulting package classification control instruction, containing sorting execution information, was passed to the subsequent robotic arm execution system to achieve accurate package sorting.

[0063] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0064] Monitor each sorting channel in real time, collect the current number of packages, processing rate, and buffer occupancy rate of each channel, and obtain the sorting channel load status matrix;

[0065] Extract the category attributes and processing priority of the package according to the package classification control instruction, establish the initial mapping relationship between the package and the potential target channel, and obtain the candidate channel set;

[0066] Calculate the comprehensive sorting score for each channel in the candidate channel set. The comprehensive sorting score is obtained by weighted integration of package category matching, channel load rate and historical sorting success rate to obtain the channel score vector.

[0067] The channel score vector is input into the decision fusion framework, and the channel with the highest score is selected as the sorting target through the Bayesian decision rule to obtain the target sorting channel number;

[0068] The estimated time for the package to arrive at the sorting point is calculated based on the target sorting channel number and the current conveyor belt speed, and the sorting execution window is obtained by considering the time constraint of the robot arm movement;

[0069] A timing optimization algorithm is applied to the sorting execution window to remove sorting conflicts and generate grabbing time points and placing time points to obtain the target sorting channel number and its execution timing.

[0070] Specifically, a sorting channel refers to a physical channel or exit in a logistics sorting system that receives specific categories of packages. Real-time monitoring collects channel operation data by installing a sensor array in each sorting channel, including counting sensors, speed sensors, and occupancy sensors. Counting sensors record the number of packages in the channel; speed sensors measure the rate at which packages move through the channel in pieces per minute; and occupancy sensors monitor the space occupied in the channel buffer. This sensor data is collected at a fixed sampling frequency (typically 5-10 Hz) and transmitted to the data processing unit. The data processing unit integrates this raw data into a sorting channel load status matrix, an n×3 matrix, where n is the total number of sorting channels and 3 represents the three monitoring indicators (number of packages, processing rate, and occupancy). Each row in the matrix represents the current status of a sorting channel, and each column represents the value of a monitoring indicator.

[0071] The package's category attributes and processing priority are extracted based on the package classification control instructions. The package classification control instructions are a data structure generated by the hierarchical control system in the previous stage, which contains the package's classification information and processing requirements. The category attribute refers to the content type of the package, such as electronic products, clothing, books, etc.; the processing priority refers to the timeliness requirement of the package, such as ordinary, expedited, and special urgent. The extraction process is to parse the data structure of the package classification control instructions, read the category field and priority field therein, and match these attributes with the sorting channel function. The sorting channel function is pre-configured and specifies the package type that each channel is suitable for processing. By querying the sorting channel function table, all channels with matching functions are found for the package, and an initial mapping relationship between the package and the potential target channel is established to form a candidate channel set. This set contains all the channel numbers that may receive the package.

[0072] Calculating a comprehensive sorting score for each candidate channel is a multi-factor evaluation process. The comprehensive sorting score is a numerical indicator of channel suitability, calculated by weighted fusion of three key factors: package category matching, channel load rate, and historical sorting success rate. Package category matching indicates the degree of compatibility between the package type and the channel's expertise, calculated by comparing the similarity between package category attributes and channel functionality. The channel load rate reflects the channel's current busyness and is calculated from data in the channel load status matrix. The historical sorting success rate records the past success rate of parcels of that category being processed in a specific channel. These three factors are weighted and summed to produce a comprehensive sorting score for each candidate channel. The scores of all candidate channels are combined into a vector, the channel score vector. The decision fusion framework is an algorithmic structure that comprehensively considers multiple decision factors, with the Bayesian decision rule as its core algorithm, making optimal decisions based on Bayesian probability theory. Specifically, the channel score vector is treated as an observation and combined with prior knowledge (such as the historical performance of each channel) to calculate the posterior probability. The Bayesian decision rule selects the option with the highest posterior probability, or the channel with the highest score, as the final sorting target. This process effectively takes into account the influence of historical data on top of the comprehensive scoring, making the decision more robust. The target sorting channel number determined in this way represents the optimal choice after considering various factors.

[0073] The estimated time for a package to arrive at the sorting point is calculated based on the target sorting lane number and the current conveyor speed. The sorting point is the physical location where the robot arm will grab the package. The estimated arrival time is the time required for the package to move from its current location to the sorting point. This calculation is done by measuring the distance from the package's current location to the sorting point and dividing it by the conveyor speed (in meters per second) to obtain the time value (in seconds). The robot's motion time constraint is also considered: the minimum time required for the robot arm to complete a single pick-and-place operation. By combining the estimated arrival time with the robot's motion time, a feasible time range for sorting the package is determined. This range is called the sorting execution window. The start time of the window is the earliest time at which the package can be grabbed, and the end time is the latest time at which the package must be grabbed.

[0074] Applying a timing optimization algorithm to the sorting execution window is the final step in ensuring smooth sorting operations. A timing optimization algorithm is a computational method for resolving multi-task scheduling conflicts. In a sorting system, it is primarily used to coordinate the sorting sequence of multiple packages and avoid robotic arm conflicts. The algorithm first detects overlap between execution windows to identify potential sorting conflicts. It then adjusts the specific execution time for each package based on package priority and arrival order. Finally, it determines the precise pickup and placement times for each package. Thus, the target sorting lane number, along with its execution sequence, constitutes the sorting execution instruction.

[0075] For example, when a batch of different types of packages needs to be sorted, all channels in the sorting system are first monitored in real time. Assume the system has 10 sorting channels. Current data is collected through the sensor network of each channel: Channel 1 currently has 15 packages, a processing rate of 20 packages / minute, and a buffer occupancy of 30%; Channel 2 currently has 8 packages, a processing rate of 15 packages / minute, and an occupancy of 20%. This process continues for all channels, forming a 10×3 load state matrix. When a package classified as "Electronics - Cell Phone - Expedited" enters the sorting system, the category attribute "Electronics - Cell Phone" and the processing priority "Expedited" are extracted from its classification control instructions. A query of the sorting channel function table reveals that channels 1, 3, and 7 all support electronics processing, forming the candidate channel set {1, 3, 7}. Next, the comprehensive sorting score for each candidate channel is calculated: Channel 1 has a high category match (specializing in mobile phones) but a high load rate, a historical success rate of 92%, and a comprehensive score of 0.78; Channel 3 has a medium category match (handling various electronic products), a low load rate, a historical success rate of 88%, and a comprehensive score of 0.72; Channel 7 has a low category match (handling various small items), a very low load rate, a historical success rate of 85%, and a comprehensive score of 0.65. These scores are combined into a channel score vector [0.78, 0.72, 0.65] and input into the Bayesian decision framework. Considering historical data indicating that Channel 1 is particularly reliable for handling expedited mobile phone packages, the Bayesian decision rule selects Channel 1 as the target sorting channel. Assuming the current package is 5 meters from the sorting point and the conveyor speed is 1 m / s, the package is expected to arrive at the sorting point in 5 seconds. Given that the robot arm takes at least 2 seconds to complete a pick-and-place operation, the sorting execution window is determined to be [5, 7] seconds (relative to the current time). The timing optimization algorithm checks and finds another package that needs to be sorted within the time window, creating a conflict. Because this package's priority is expedited, it has a higher priority than the other package. Therefore, this package is picked up at 5.5 seconds and placed in channel 1 at 7 seconds, while the other package's sorting time is postponed. The final output target sorting channel number and execution sequence are channel 1, with a pick-up time of 5.5 seconds and a placement time of 7 seconds.

[0076] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0077] Perform status detection on the industrial robotic arms in the sorting system, collect the current position, load status and task queue of each robotic arm, and obtain the robotic arm status data;

[0078] The estimated time point when the package arrives at each robotic arm's work area is calculated based on the target sorting channel number, and the grasping constraints are constructed based on the physical properties of the package to obtain the task's spatiotemporal constraint matrix.

[0079] The task allocation cost is calculated based on the task spatiotemporal constraint matrix and the robot state data. The task allocation cost comprehensively considers time matching, energy consumption and robot load balance to obtain the task allocation cost matrix.

[0080] Apply the task allocation algorithm to the task allocation cost matrix for global optimization, select the robot arm with the minimum cost for each package to be sorted, and obtain the robot arm allocation plan;

[0081] According to the robot arm allocation plan, if the current load of the robot arm exceeds the preset upper limit, the task will be placed in the waiting queue and the reallocation process will be triggered. Otherwise, the robot arm trajectory planning data will be directly generated.

[0082] The robot arm trajectory planning data is converted into a motion instruction sequence in the joint space, and the target gripper selection and gripping force control parameters are added to obtain the gripping task instructions of the industrial robot arm.

[0083] Specifically, an industrial robotic arm is a multi-jointed robot used for grasping and placing packages. It typically has six degrees of freedom and can accurately position itself in three-dimensional space. State detection involves collecting real-time information about the robotic arm's operating status through various sensors and encoders. Position detection first uses the angle encoders at each joint to obtain the current angle values of each joint. Then, forward kinematics is used to calculate the position and posture of the end effector (the robotic arm's hand) in space. Load refers to the weight currently carried by the robotic arm, measured by joint torque sensors or end torque sensors. The task queue is a list of tasks assigned to the robotic arm but not yet executed, including the task number, estimated execution time, and operation type. This data is aggregated and processed by the data acquisition module to form structured robotic arm status data, describing the real-time operating status of each robotic arm.

[0084] The estimated time for the package to arrive at each robot's work area is calculated based on the target sorting lane number. The target sorting lane number represents the package's destination, determined in the previous stage. The work area refers to the spatial range within which the robot can grasp objects. The calculation process first determines the package's current position on the conveyor belt, measures the distance from this position to each robot's work area, and then divides it by the conveyor belt's operating speed to determine the estimated time for the package to arrive at each work area. Furthermore, grasping constraints are constructed based on the package's physical properties (such as size, weight, shape, and fragility). Grasping constraints are rules that must be followed to ensure safe grasping. For example, for packages weighing over a certain threshold, a specific gripper must be used; for fragile items, the grasping force must not exceed a certain value; and for packages with unusual shapes, the grasping point must be located at a specific location. This temporal information and constraints are combined to form the task's spatiotemporal constraint matrix. Each row represents a package, and each column represents a constraint or time value, fully describing the spatiotemporal constraints for task execution.

[0085] The task allocation cost refers to the overall cost or expense of assigning a specific package to a specific robot arm for execution. A lower cost represents a better value. The calculation takes into account three key factors: time alignment, which refers to the proximity between the robot arm's idle time after completing its current task queue and the package's arrival time; a smaller time difference indicates a higher fit; energy consumption, which refers to the energy required to move the robot arm from its current position to the grasping position and then to the placement position, is typically proportional to the distance moved and the weight of the load; and arm load balance, which refers to the balance of workload between the arms, preventing some arms from being overloaded while others are idle. These three factors are combined in a weighted manner to determine the allocation cost for each package-robot pairing, forming a task allocation cost matrix. This matrix has a number of rows equal to the number of packages to be sorted and a number of columns equal to the number of available robots. Each element in the matrix represents the cost of assigning a package to a specific robot arm.

[0086] The task allocation algorithm is applied to the task allocation cost matrix for global optimization. The task allocation algorithm is a computational method for solving many-to-many matching problems. It is used to determine the optimal parcel and robot arm pairing scheme in the sorting system. The commonly used method is the Hungarian algorithm, which is a classic combinatorial optimization algorithm that can find the optimal match in polynomial time. The input of the algorithm is the task allocation cost matrix, and the output is the allocation scheme with the minimum total cost. The specific operation process includes: matrix preprocessing (row subtraction, column subtraction), finding independent zero elements, drawing lines to cover all zero elements, adjusting the values of elements not covered by the lines, and iterating repeatedly until a perfect match is found. The result of the optimization solution is to assign the most suitable robot arm to each parcel to be sorted, forming a robot arm allocation scheme, which contains a list of parcel-robot arm pairing relationships.

[0087] Judging based on the robot arm allocation plan is a protective measure to ensure the safe operation of the system. The current load of the robot arm refers to the number of tasks or work intensity that the robot arm has been assigned, and the preset upper limit is the maximum safe workload of the robot arm set by the system. The judgment process determines whether it exceeds the safety range by comparing the total load of the robot arm after allocation with the preset upper limit. If it exceeds the upper limit, the current task is placed in the waiting queue. The waiting queue is a task buffer sorted by priority, which temporarily stores tasks that cannot be executed immediately; at the same time, the reallocation process is triggered to re-search for a suitable execution robot arm for the task. If it does not exceed the upper limit, it goes directly to the next step to generate the robot arm trajectory planning data, which is a data structure that describes the robot arm's motion path, including the starting point, end point, intermediate path points, and corresponding speed and acceleration parameters.

[0088] Converting the robot's trajectory planning data into a sequence of motion commands in joint space bridges high-level task planning and low-level control. Joint space describes the angles of each joint in the robot arm, distinct from Cartesian space (which describes the end-effector's position and posture). This conversion process uses an inverse kinematics algorithm to convert the end-effector's target position and posture in three-dimensional space into the angles of each joint. Inverse kinematics is a key algorithm in robotics, solving the problem of determining the angles of each joint given the end-effector's position. The converted data represents the angles of each joint at a series of time points, forming the motion trajectory. Simultaneously, based on the physical characteristics of the package, an appropriate end-effector (such as a vacuum cup or robotic gripper) is selected as the target gripper, and corresponding gripping force control parameters are set to ensure safe and reliable grasping of different package types. This data is combined to form the industrial robot's grasping task instructions, which contain execution information and are directly transmitted to the robot's control system to perform the actual grasping and placement operations.

[0089] For example, in a sorting system with six industrial robotic arms, the status of each arm is monitored. Arm #1 is located at coordinates (2.5m, 1.8m, 1.2m), currently loaded with two tasks, and its task queue contains two pending grasping tasks. Arm #2 is located at (5.3m, 1.8m, 1.0m), currently loaded with zero tasks, and its task queue is empty. Status data for the other robotic arms is collected similarly. When a package classified as an electronics-cell phone-expedited item is sorted to aisle 12, the estimated time for the package to reach each robotic arm's workspace is calculated based on the package's current position on the conveyor (1.0m, 0.5m) and the conveyor speed (1m / s). The estimated time it takes to reach arm #1's workspace is 3 seconds, arm #2's workspace is 5 seconds, and so on. Furthermore, because the package is identified as an electronic product and is lightweight (0.3kg), the grasping constraint is set to use a vacuum cup with a medium gripping force. This information is combined to form a task spatiotemporal constraint matrix, which includes information such as the package ID, estimated arrival time, required gripping tool, and force requirements. Next, the task allocation cost is calculated: For robot arm #1, since its task queue has two tasks and it is expected to be idle in 4 seconds, while the package arrives in 3 seconds, the time alignment is poor. The travel distance is short, energy consumption is low, and the current load is moderate, resulting in a comprehensive cost of 7.5. For robot arm #2, its task queue is empty, making it immediately available and having good time alignment. However, the travel distance is long, energy consumption is high, and the current load is 0, resulting in good load balance, resulting in a comprehensive cost of 6.2. Similarly, the costs for all robot arms are calculated to form a cost matrix. The Hungarian algorithm is applied to optimize the cost matrix, resulting in the selection of robot arm #2 as the optimal allocation solution. After receiving the new task, the load of robot arm #2 is checked to see if it exceeds the preset upper limit (e.g., each robot arm can handle a maximum of three tasks simultaneously). It is not, so trajectory planning data is generated directly. The trajectory planning forms a spatial path from the current position of robot arm #2 (5.3m, 1.8m, 1.0m) to the package grasping position (approximately 6.0m, 0.5m, 0.2m), and then to the entrance position of channel 12 (7.2m, 2.5m, 0.5m). Using an inverse kinematics algorithm, this spatial path is converted into a sequence of six joint angles: for example, joint 1 rotates from 30° to 45°, joint 2 from 60° to 75°, and so on. At the same time, based on the characteristics of the package, a vacuum suction cup is specified as the gripper, and the gripping force is set to medium (corresponding to a vacuum degree of -40kPa). The resulting industrial robot arm grasping task command, which includes the motion trajectory, gripper selection, and force control parameters, is sent directly to the control system of robot arm #2 for actual operation.

[0090] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0091] The first sensor array collects image, depth, and weight data of the parcel at the entrance of the sorting channel, and performs feature extraction to obtain entrance feature data;

[0092] Calculate the cosine similarity between the entry feature data and the original features in the package classification control instruction, set different weight coefficients for each feature dimension, and obtain the entry matching score;

[0093] The second sensor array collects image, depth, and weight data of the parcel at the exit of the sorting channel, and performs feature extraction to obtain exit feature data;

[0094] Calculate the export matching score based on the export feature data and the original feature, and perform a weighted average with the import matching score to obtain the comprehensive matching score;

[0095] A threshold is applied to the comprehensive matching score. When it is lower than the preset threshold, the package features are input into a lightweight neural network for reclassification. The lightweight neural network consists of three convolutional layers, two pooling layers, and two fully connected layers. Each convolutional layer uses a 3×3 convolution kernel, and the fully connected layers contain 128 and 64 neurons, respectively, to obtain the reclassification results.

[0096] Sorting quality evaluation data is generated based on the reclassification results, comprehensive matching scores and sorting channel operation status.

[0097] Specifically, quality control begins with the first sensor array collecting image, depth, and weight data for packages at the sorting channel entrance. This first sensor array comprises a combination of sensors installed at the sorting channel entrance, including a high-resolution camera, a depth sensor, and a weight sensor. The high-resolution camera captures 2D image information of the package, including color, appearance, and surface markings; the depth sensor acquires 3D structural information, recording volume and shape characteristics; and the weight sensor accurately measures the package's mass. When a package enters the sorting channel entrance, these three sensors are synchronously triggered to collect raw data. Feature extraction is then performed to convert the raw data into structured feature vectors. Feature extraction includes image processing (such as color histogram extraction and edge feature extraction), depth data processing (such as volume calculation and principal axis orientation extraction), and weight data normalization. These processed features are combined into a uniformly formatted feature vector to form the entry feature data, which is used for subsequent matching calculations.

[0098] Calculating cosine similarity between entry feature data and the original features in the package sorting control instructions is a mathematical method for assessing classification accuracy. The original features in the package sorting control instructions refer to the package feature vector generated during the sorting decision phase, which serves as a reference. Cosine similarity measures the degree of directional similarity between two vectors. It calculates the cosine of the angle between the two vectors, ranging from -1 to 1, with values closer to 1 indicating greater similarity. During the calculation process, different weight coefficients are assigned to each feature dimension. The weight coefficients reflect the importance of each feature to the classification result. For example, for fragile items, shape may be weighted higher than weight. Specifically, the entry feature vector and the original feature vector are weighted together for the corresponding dimensions, and the sum is divided by the product of the weighted moduli of the two vectors to obtain the entry match score. This score reflects the degree of match between the features of the package entering the sorting channel and the expected classification features.

[0099] The second line of defense in the dual verification mechanism is to collect image, depth, and weight data from packages at the exit of the sorting channel using a second sensor array. Similar to the first sensor array, this second sensor array is installed at the exit of the sorting channel to provide final confirmation before the package leaves the sorting system. This array also includes a high-resolution camera, depth sensor, and weight sensor to collect physical characteristic data of the package as it leaves the sorting channel. Similar to the entry processing, feature extraction is performed on the collected raw data and converted into a standardized feature vector, the exit feature data. The format of the exit feature data is consistent with the entry feature data, facilitating subsequent comparison and analysis.

[0100] Calculating the exit match score based on the exit feature data and the original features, and taking a weighted average of this score with the entry match score, is a key step in comprehensively evaluating sorting quality. The exit match score is calculated in a similar way to the entry match score, using the cosine similarity method to calculate the similarity between the exit features and the original features. The entry match score and the exit match score are then weighted averaged to form a comprehensive match score. Weighted averaging is an averaging method that takes the importance of each factor into account. Typically, the exit match score is weighted slightly higher than the entry match score because the exit features better reflect the final sorting results. The comprehensive match score is a value between 0 and 1, representing the quality assessment result of the entire sorting process. Values closer to 1 indicate that the sorting results are more in line with expectations.

[0101] Applying a threshold to the overall match score is the key decision point for determining whether additional verification is required. The threshold is a preset critical value for the system, typically determined through historical data analysis and system tuning. When the overall match score falls below the threshold, the reliability of the current sorting result is insufficient and further verification is required. In this case, the package features are fed into a lightweight neural network for reclassification. A lightweight neural network is a simplified deep learning model optimized for fast inference, with low computational complexity while maintaining comparable classification accuracy. The network architecture consists of three convolutional layers, two pooling layers, and two fully connected layers. The convolutional layers use 3×3 convolutional kernels to extract features from the input. Each convolution operation is followed by an activation function (typically ReLU) to add nonlinearity. The pooling layer reduces the data dimensionality by downsampling, typically using 2×2 max pooling. The fully connected layers contain 128 and 64 neurons, respectively, mapping the extracted features to the output space. The network input is the package's feature vector, and the output is a probability distribution over each class. The class with the highest probability is the reclassification result.

[0102] The final output of quality monitoring is sorting quality assessment data, generated based on the reclassification results, the overall matching score, and the sorting channel operating status. This data is a structured data set consisting of multiple indicators: the classification accuracy indicator records the consistency between the original classification and the reclassification; the matching indicator records the overall matching score; the anomaly type indicator identifies possible causes of errors; and the system status indicator reflects the operating status of the sorting channel. This data is integrated into a standard format and provided to the subsequent control system for parameter adjustment and optimization, forming a closed-loop quality control loop.

[0103] For example, when a package classified as an electronics product / mobile phone is assigned to Aisle 12 in the electronics zone, it first passes through the first sensor array at the entrance. A high-resolution camera captures the package's image, recording its rectangular shape and the product logo on the packaging surface. A depth sensor scans and obtains 3D point cloud data of the package, measuring its dimensions to approximately 15 x 8 x 2 cm. A weight sensor detects the package's weight as 302 grams. The feature extraction module processes this raw data and extracts a 128-dimensional feature vector as the entry feature data. This entry feature vector is then compared to the feature vector used in the original classification process using a cosine similarity calculation, assigning a weight of 0.4 to shape features, 0.3 to weight features, and 0.3 to surface features. The resulting entry match score is 0.92, indicating that the package's features at the entrance are highly consistent with the expected values. As the package passes through the sorting aisle and reaches the exit, data is collected again by the second sensor array. Due to slight movement of the package during transport, the image and depth data collected at the exit differ slightly from those at the entrance, but the weight remains essentially the same. The same feature extraction process yields exit feature data, which, when compared with the original features, yields an exit match score of 0.89. The weighted average of the entry match score (weighted 0.4) and the exit match score (weighted 0.6) yields an overall match score of 0.90. The system's preset threshold is 0.85. Since the overall match score exceeds the threshold, the sorting result is reliable and no reclassification is required. However, if the overall match score were 0.82, which falls below the preset threshold, a lightweight neural network would be triggered for reclassification. The package features are then fed into the neural network, where they undergo three layers of convolution (using 16, 32, and 64 3×3 kernels per layer), two layers of pooling (2×2 max pooling), and finally mapped to the output layer via two fully connected layers (containing 128 and 64 neurons, respectively). The reclassification result still points to the Electronics - Mobile Phones category, consistent with the original classification, but with improved confidence. The system generates sorting quality assessment data, records information such as package ID, original classification results, matching scores, reclassification results (if any), and channel operation status, forming quality monitoring records.

[0104] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0105] Classify and count the sorting quality assessment data, extract the classification error rate, feature deviation value and response delay value, and obtain the control performance index;

[0106] The control performance indicators are trained through real-time collected operating data to obtain the parameter influence matrix;

[0107] Perform matrix decomposition on the parameter influence matrix to obtain the parameter adjustment type;

[0108] Numerically adjusting a feedback gain parameter in the hierarchical control system based on the parameter adjustment type to obtain a correction gain parameter;

[0109] Correcting the decision threshold in the hierarchical control system according to the correction gain parameter to obtain a target decision threshold;

[0110] The correction gain parameter and target decision threshold are combined with the sorting channel load data, and the control execution timing is redistributed to obtain the closed-loop control parameter set.

[0111] Specifically, the classification and statistics of sorting quality assessment data is the starting point of self-optimization. This process systematically processes the assessment data generated in the quality monitoring stage through data analysis algorithms. Sorting quality assessment data refers to a data set containing quality information of package sorting results, including matching scores, classification results, verification results and other information. Classification statistics is to classify and calculate these data according to different dimensions, and extract three key indicators: the classification error rate refers to the frequency of the sorting system sorting packages into the wrong channel, which is calculated by counting the proportion of inconsistencies between the reclassification results and the original classification; the feature deviation value refers to the average degree of deviation between the actual measured package features and the expected features, which is obtained by calculating the average absolute difference of each dimension of the feature vector; the response delay value represents the average time delay from the package entering the system to the completion of sorting, which is calculated by the timestamp difference. These three indicators together constitute the control performance indicators, which reflect the accuracy, stability and efficiency of the sorting system.

[0112] Training control performance indicators using real-time collected operational data is the process of establishing a model for the relationship between performance and control parameters. Operational data includes the various control parameter settings for the sorting system (such as gain values, thresholds, timing parameters, etc.) and the corresponding control performance indicators. The training process uses a supervised learning approach, taking the control parameters as feature inputs and the control performance indicators as target outputs, and establishing a mapping relationship between the two through a regression algorithm. In specific implementation, the historical operational data is first divided into training and validation sets in chronological order, and then the model is trained using algorithms such as multivariate linear regression, support vector regression, or neural networks. The training results generate a parameter influence matrix, a numerical matrix that describes the relationship between control parameters and performance indicators. Each element in the matrix represents the degree of influence of a specific control parameter on a specific performance indicator.

[0113] Matrix decomposition of the parameter influence matrix is a mathematical process for identifying key parameters and adjustment directions. Matrix decomposition is the process of breaking down a complex matrix into the product of multiple, simpler matrices. Common methods include singular value decomposition (SVD), principal component analysis (PCA), and non-negative matrix factorization (NMF). In this method, the SVD algorithm is applied to the parameter influence matrix, decomposing it into the product of three matrices: a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. The left singular vectors correspond to the dominant modes of the control parameters, the right singular vectors correspond to the dominant modes of the performance indicators, and the singular values indicate the importance of each mode. By analyzing the singular values and corresponding singular vectors, the parameter combinations and adjustment directions that most significantly affect system performance are determined, resulting in the parameter adjustment type, which categorizes the parameter adjustment strategy, such as gain adjustment, threshold adjustment, or timing adjustment.

[0114] Numerical adjustment of the feedback gain parameters in a hierarchical control system based on the parameter adjustment type is a direct measure to optimize control effectiveness. A hierarchical control system refers to a layered control architecture consisting of a main control engine and secondary specialized controllers. The feedback gain parameters are numerical coefficients in the control loop that determine the response strength. Depending on the parameter adjustment type, a corresponding adjustment strategy is adopted: for gain adjustment, the proportional coefficient, integral coefficient, and differential coefficient in the feedback loop are directly adjusted; for controllers with a large number of parameters, the gradient descent method can be used to make incremental adjustments along the direction indicated by the parameter influence matrix. The adjustment range is dynamically determined based on the size of the performance deviation and the historical adjustment effect to prevent over-adjustment from causing system instability. After adjustment, the corrected gain parameters are obtained, which is a new optimized set of gain values that will be applied to each control link in the hierarchical control system.

[0115] Correcting the decision threshold in a hierarchical control system based on correction gain parameters is a key step in optimizing classification accuracy. Decision thresholds are critical values used in control systems for classification decisions, such as the similarity threshold in matching calculations and the deviation threshold in error detection. The correction process uses the correction gain parameters and historical performance data to determine the optimal value through threshold sensitivity analysis. For areas with large deviations, the threshold is appropriately lowered to increase sensitivity; for areas with high accuracy, the threshold is appropriately raised to reduce false positives. This correction calculation typically employs Bayesian optimization or grid search methods, testing the performance of different threshold values within a certain range and selecting the optimal result as the target decision threshold. Combining the correction gain parameters and target decision threshold with sorting channel load data and reallocating control execution timing is the final step in achieving a complete closed-loop control system. Sorting channel load data provides real-time information describing the current operating status of each sorting channel, including the number of packages in the channel, processing rate, and buffer occupancy. This combination involves combining the optimized control parameters with real-time load conditions to adjust task execution schedules. Timing reallocation utilizes discrete event simulation and scheduling optimization algorithms to predict system performance under different execution sequences and select the optimal execution plan. Finally, a closed-loop control parameter set is generated. This is a complete set of control parameter configurations, including correction gain parameters, target decision thresholds, and optimized execution timing. It will be directly applied to the next operation cycle of the sorting system to achieve a closed-loop control.

[0116] The above describes the automated sorting method based on artificial intelligence in the embodiment of the present application. The following describes the automated sorting system based on artificial intelligence in the embodiment of the present application. Figure 2 In the embodiments of the present application, an embodiment of an automated sorting system based on artificial intelligence includes:

[0117] The acquisition module 201 is used to collect and process multi-dimensional data of the packages on the conveyor belt using a high-resolution camera array, a depth sensor, and a weight sensor to obtain a package feature vector;

[0118] An input module 202 is configured to input the package feature vector into a hierarchical control system to analyze control parameters and obtain a package classification control instruction;

[0119] Identification module 203, configured to identify the sorting status based on the parcel classification control instruction and the current load status of each sorting channel, and obtain the target sorting channel number and its execution sequence;

[0120] The allocation module 204 is used to allocate the grabbing task to the industrial robot arm through a task allocation algorithm according to the target sorting channel number and execution sequence;

[0121] The classification module 205 is used to verify the parcels transferred by the industrial robot arm through the dual sensor arrays at the entrance and exit, calculate the weighted sum of feature similarities as a matching score, and trigger the lightweight neural network to reclassify when the matching score is lower than a preset threshold to obtain sorting quality assessment data;

[0122] The control module 206 is configured to adjust the feedback gain parameters, decision thresholds, and execution control timing of the hierarchical control system according to the sorting quality evaluation data to generate a closed-loop control parameter set.

[0123] Through the collaborative efforts of these components, package feature vectors are acquired through multi-dimensional data acquisition and processing techniques, achieving comprehensive perception of the package's physical characteristics and significantly improving classification and identification accuracy. The step of inputting standardized package feature vectors into the hierarchical control system for analysis of control parameters fully leverages the advantages of artificial intelligence algorithms in feature extraction and pattern recognition, enabling the system to learn optimal classification strategies from extensive historical data and adapt to the characteristic differences between different types of packages. A design that identifies sorting status based on package classification control instructions and the current load status of each sorting channel introduces a dynamic load balancing mechanism, addressing the uneven resource allocation problem in traditional sorting systems and improving overall throughput. A task allocation algorithm assigns grasping tasks to industrial robotic arms, enabling collaborative operation and intelligent scheduling of multiple robotic arms, reducing equipment idle time and improving hardware resource utilization. A quality monitoring mechanism, verified by a dual sensor array for packages transferred by the industrial robotic arms, establishes real-time error correction capabilities, significantly reducing the sorting error rate and ensuring sorting quality. A closed-loop control design that dynamically adjusts the hierarchical control system parameters based on sorting quality assessment data empowers the system with continuous self-optimization capabilities, enabling it to adapt to changes in the logistics environment and fluctuations in package characteristics.

[0124] above Figure 2 The artificial intelligence-based automated sorting system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The artificial intelligence-based automated sorting equipment in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0125] Figure 3The figure is a schematic diagram of the structure of an AI-based automated sorting device provided by an embodiment of the present invention. The AI-based automated sorting device 300 may vary significantly depending on its configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions for operating the AI-based automated sorting device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, executing the series of instructions stored in the storage medium 330 on the AI-based automated sorting device 300 to implement the steps of the aforementioned AI-based automated sorting method.

[0126] The artificial intelligence-based automated sorting device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the artificial intelligence-based automated sorting equipment shown does not constitute a limitation on the artificial intelligence-based automated sorting equipment provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0127] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the artificial intelligence-based automated sorting method.

[0128] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0129] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling an artificial intelligence-based automated sorting device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An automated sorting method based on artificial intelligence, characterized in that: The method comprises: The packages on the conveyor belt are collected and processed in multiple dimensions using a high-resolution camera array, depth sensor, and weight sensor to obtain a package feature vector. Inputting the package feature vector into a hierarchical control system to analyze control parameters and obtain package classification control instructions; Perform sorting status identification based on the parcel classification control instruction and the current load status of each sorting channel to obtain the target sorting channel number and its execution sequence; According to the target sorting channel number and execution sequence, the task allocation algorithm is used to allocate the grasping task to the industrial robot arm; The parcels transferred by the industrial robot arm are verified by the dual sensor arrays at the entrance and exit, and the weighted sum of feature similarities is calculated as a matching score. When the matching score is lower than a preset threshold, a lightweight neural network is triggered to perform reclassification to obtain sorting quality evaluation data, including: collecting image, depth and weight data of the parcels at the entrance of the sorting channel through the first sensor array, and performing feature extraction to obtain entrance feature data; performing cosine similarity calculation on the entrance feature data and the original features in the parcel classification control instruction, setting different weight coefficients for each feature dimension, and obtaining the entrance matching score; collecting image, depth and weight data of the parcels at the exit of the sorting channel through the second sensor array Data is collected and feature extracted to obtain export feature data; an export matching score is calculated based on the export feature data and the original features, and a weighted average is performed on the export matching score and the import matching score to obtain a comprehensive matching score; a threshold is determined for the comprehensive matching score, and when it is lower than a preset threshold, the package features are input into a lightweight neural network for reclassification. The lightweight neural network consists of three convolutional layers, two pooling layers, and two fully connected layers. Each convolutional layer uses a 3×3 convolution kernel, and the fully connected layers contain 128 and 64 neurons, respectively, to obtain a reclassification result; sorting quality assessment data is generated based on the reclassification result, the comprehensive matching score, and the operating status of the sorting channel; The feedback gain parameters, decision thresholds and execution control timing of the hierarchical control system are adjusted according to the sorting quality evaluation data to generate a closed-loop control parameter set.

2. The automated sorting method based on artificial intelligence according to claim 1, characterized in that: The package on the conveyor belt is collected and processed in multiple dimensions using a high-resolution camera array, a depth sensor, and a weight sensor to obtain a package feature vector, including: The packages moving on the conveyor belt are illuminated by a dual-light source lighting system, with the brightness ratio of the main light source to the auxiliary light source set at two to one to obtain the lighting environment; Scanning the surface of the package in the illumination environment using a triangulation laser rangefinder to construct multi-point three-dimensional point cloud data to obtain shape feature data of the package; Scanning the temperature distribution of the package using an infrared thermal imaging sensor to collect a multi-pixel thermal map to obtain a temperature distribution map of the package; Using OCR technology to identify and decode the barcode and text information on the surface of the package, extract the destination code and classification mark of the package, and obtain the identification information of the package; Performing multimodal feature fusion on the shape feature data, temperature distribution map and identification information, generating a multi-dimensional fusion feature matrix through tensor calculation, and obtaining multi-dimensional description data of the package; Nonlinear dimensionality reduction and statistical filtering are performed on the multidimensional description data, and the high-dimensional features are converted into low-dimensional vector space representation through the t-SNE algorithm to obtain the package feature vector.

3. The automated sorting method based on artificial intelligence according to claim 1, characterized in that: The step of inputting the package feature vector into a hierarchical control system to analyze control parameters and obtain a package classification control instruction includes: Inputting the package feature vector into a main control engine, the main control engine including a residual connection structure and a spatial pyramid pooling module, performing primary classification mapping on the feature vector to obtain a main classification category probability distribution; Selecting a corresponding secondary specialized controller based on the primary classification probability distribution, wherein the secondary specialized controller uses a densely connected structure to perform classification processing on the feature vector to obtain a subdivided classification probability distribution; Performing a weighted fusion calculation of cross entropy and focal loss on the main classification category probability distribution and the subdivision category probability distribution to obtain a classification confidence score; Performing Bayesian probability compensation according to the classification confidence score and the historical classification accuracy, calibrating the classification result, and obtaining a calibrated classification result; Matching and analyzing the calibrated classification results with the current system physical constraints, including the capacity limits of each sorting channel and the outbound loading capacity, to obtain preliminary sorting instructions; The preliminary sorting instructions are parameterized by the control rule engine of the hierarchical control system, including instruction priority setting, execution timing arrangement and exception handling plan, to obtain the package classification control instructions.

4. The artificial intelligence-based automated sorting method according to claim 1, characterized in that: The sorting state identification is performed based on the parcel classification control instruction and the current load state of each sorting channel to obtain the target sorting channel number and its execution sequence, including: Monitor each sorting channel in real time, collect the current number of packages, processing rate, and buffer occupancy rate of each channel, and obtain the sorting channel load status matrix; Extracting the category attributes and processing priority of the package according to the package classification control instruction, establishing an initial mapping relationship between the package and the potential target channel, and obtaining a set of candidate channels; Calculating a comprehensive sorting score for each channel in the candidate channel set, wherein the comprehensive sorting score is obtained by weighted integration of the package category matching degree, the channel load rate, and the historical sorting success rate to obtain a channel scoring vector; The channel score vector is input into the decision fusion framework, and the channel with the highest score is selected as the sorting target through the Bayesian decision rule to obtain the target sorting channel number; Calculate the estimated time for the package to arrive at the sorting point based on the target sorting channel number and the current conveyor belt speed, and consider the time constraint of the robot arm movement to obtain the sorting execution window; A timing optimization algorithm is applied to the sorting execution window to delete sorting conflicts and generate grabbing time points and placing time points, thereby obtaining a target sorting channel number and its execution timing.

5. The automated sorting method based on artificial intelligence according to claim 1, characterized in that: The method of allocating a grabbing task to the industrial robot arm through a task allocation algorithm according to the target sorting channel number and the execution sequence includes: Perform status detection on the industrial robotic arms in the sorting system, collect the current position, load status and task queue of each robotic arm, and obtain the robotic arm status data; The estimated time point when the package arrives at each robotic arm's work area is calculated based on the target sorting channel number, and the grasping constraint conditions are constructed based on the physical properties of the package to obtain the task spatiotemporal constraint matrix; Calculate the task allocation cost according to the task spatiotemporal constraint matrix and the robot arm state data, wherein the task allocation cost comprehensively considers time matching, energy consumption and robot arm load balance to obtain a task allocation cost matrix; Applying a task allocation algorithm to the task allocation cost matrix for global optimization, selecting a robotic arm with the minimum cost for each package to be sorted, and obtaining a robotic arm allocation plan; According to the robot arm allocation plan, if the current load of the robot arm exceeds the preset upper limit, the task is placed in the waiting queue and the reallocation process is triggered; otherwise, the robot arm trajectory planning data is directly generated; The robot arm trajectory planning data is converted into a motion instruction sequence in the joint space, and the target gripper selection and gripping force control parameters are added to obtain the gripping task instructions of the industrial robot arm.

6. The artificial intelligence-based automated sorting method according to claim 1, characterized in that: The step of adjusting the feedback gain parameters, decision thresholds, and execution control timing of the hierarchical control system according to the sorting quality evaluation data to generate a closed-loop control parameter set includes: Performing classification statistics on the sorting quality assessment data, extracting classification error rate, feature deviation value and response delay value, and obtaining control performance indicators; The control performance index is trained using real-time collected operating data to obtain a parameter influence matrix; Performing matrix decomposition on the parameter influence matrix to obtain a parameter adjustment type; numerically adjusting a feedback gain parameter in the hierarchical control system based on the parameter adjustment type to obtain a correction gain parameter; Correcting the decision threshold in the hierarchical control system according to the correction gain parameter to obtain a target decision threshold; The correction gain parameter and the target decision threshold are combined with the sorting channel load data, and the control execution timing is redistributed to obtain a closed-loop control parameter set.

7. An automated sorting system based on artificial intelligence, characterized in that: For implementing the artificial intelligence-based automated sorting method according to any one of claims 1 to 6, the artificial intelligence-based automated sorting system comprises: The acquisition module is used to collect and process multi-dimensional data of packages on the conveyor belt using a high-resolution camera array, depth sensor, and weight sensor to obtain package feature vectors; An input module, configured to input the package feature vector into a hierarchical control system to analyze control parameters and obtain a package classification control instruction; an identification module, configured to identify the sorting status based on the parcel classification control instruction and the current load status of each sorting channel, and obtain the target sorting channel number and its execution sequence; An allocation module is used to allocate grasping tasks to the industrial robot arm through a task allocation algorithm according to the target sorting channel number and execution sequence; The classification module is used to verify the parcels transferred by the industrial robot arm through the dual sensor arrays at the entrance and exit, calculate the weighted sum of feature similarities as a matching score, and trigger the lightweight neural network to reclassify when the matching score is lower than a preset threshold to obtain sorting quality evaluation data, including: collecting image, depth and weight data of the parcels at the entrance of the sorting channel through the first sensor array, and performing feature extraction to obtain entrance feature data; performing cosine similarity calculation on the entrance feature data and the original features in the parcel classification control instruction, setting different weight coefficients for each feature dimension, and obtaining the entrance matching score; collecting image, depth and weight data of the parcels at the exit of the sorting channel through the second sensor array, and extracting the features to obtain entrance feature data; performing cosine similarity calculation on the entrance feature data and the original features in the parcel classification control instruction, and setting different weight coefficients for each feature dimension to obtain the entrance matching score; and weight data, and perform feature extraction to obtain export feature data; calculate the export matching score based on the export feature data and the original features, and perform weighted average with the import matching score to obtain a comprehensive matching score; perform threshold judgment on the comprehensive matching score, and when it is lower than the preset threshold, input the package features into a lightweight neural network for reclassification. The lightweight neural network consists of three convolutional layers, two pooling layers, and two fully connected layers. Each convolutional layer uses a 3×3 convolution kernel, and the fully connected layers contain 128 and 64 neurons, respectively, to obtain a reclassification result; generate sorting quality assessment data based on the reclassification result, the comprehensive matching score, and the operating status of the sorting channel; The control module is used to adjust the feedback gain parameters, decision thresholds and execution control timing of the hierarchical control system according to the sorting quality evaluation data to generate a closed-loop control parameter set.

8. An automated sorting device based on artificial intelligence, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the artificial intelligence-based automated sorting method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the artificial intelligence-based automated sorting method according to any one of claims 1 to 6.

Images