Intelligent logistics commodity scheduling management method and system based on feature recognition

Abstract

This application discloses a smart logistics commodity scheduling and management method and system based on feature recognition, relating to the field of logistics sorting technology. The method includes: assessing feature consistency based on initial physical feature information and establishing a first feature recognition mechanism; performing barcode recognition on packages to be sorted on the conveyor line; if barcode recognition passes, performing package image recognition using the first feature recognition mechanism; if barcode recognition fails, performing image recognition on packages that failed barcode recognition using a second feature recognition mechanism; predicting and obtaining the probability of package congestion on each conveyor line within a preset time zone; performing conveyor path planning; and managing logistics commodity scheduling within the preset time zone. This solves the technical problems of low commodity feature recognition accuracy, insufficient sorting efficiency, and high reliance on manual labor in existing logistics sorting methods, making it difficult to meet the automation requirements of smart logistics.

Description

Technical Field

[0001] This application relates to the field of logistics sorting technology, specifically to a smart logistics commodity scheduling and management method and system based on feature recognition. Background Technology

[0002] With the rapid development of e-commerce and intelligent manufacturing, the logistics industry faces multiple challenges, including a surge in parcel volume, increased demands for sorting efficiency, and the need to control operating costs. Traditional logistics sorting systems mainly rely on manual scanning or single barcode recognition technology for parcel classification and route allocation, which exposes many limitations when dealing with the massive parcel processing demands.

[0003] Barcodes on package surfaces are prone to recognition failure due to wear, dirt, wrinkles, or poor printing quality. For packages that fail to be recognized, manual intervention or return to the exception handling channel is usually required, which severely restricts the continuity of sorting efficiency and the level of automation. In addition, barcodes only provide identification information for packages and cannot directly reflect the physical attributes of packages, such as size, weight, and fragility.

[0004] On the other hand, existing visual recognition assistance solutions mostly adopt fixed feature extraction strategies, which fail to fully consider the differences in the distribution of physical features of packages at different times and in different batches. This results in a mismatch between the feature recognition model and the actual package features, making it difficult to guarantee recognition accuracy. Visual recognition systems with fixed parameters often have a high false recognition rate, requiring frequent manual calibration and parameter adjustments, which increases system maintenance costs. Summary of the Invention

[0005] This application provides a smart logistics commodity scheduling and management method and system based on feature recognition, which solves the technical problems of low commodity feature recognition accuracy, insufficient sorting efficiency, and high reliance on manual labor in existing logistics sorting methods, making it difficult to meet the automation requirements of smart logistics.

[0006] The technical solution to the above-mentioned technical problems in this application is as follows:

[0007] Firstly, this application provides a smart logistics commodity scheduling and management method based on feature recognition, the method comprising:

[0008] Based on the initial physical characteristics of several packages recorded in the database within a preset time period, a characteristic consistency assessment is performed, and a first characteristic recognition mechanism is established.

[0009] The packages to be sorted on the conveyor line are identified by barcodes. If the barcode identification is successful, the package image is identified using the first feature recognition mechanism to obtain multimodal information of the package.

[0010] If the barcode recognition fails, the second feature recognition mechanism is used to perform image recognition on the package that failed the barcode recognition to obtain the package's multimodal information;

[0011] Based on the multimodal information set of parcels on the conveyor line in the current time period, combined with the current conveyor status information of each conveyor line, the probability of parcel congestion on each conveyor line in the preset time zone is predicted.

[0012] With the constraint that the probability of package congestion on each conveyor line does not exceed a preset threshold, and with the dual optimization objectives of minimizing the total package transportation time and minimizing the variance of the predicted congestion probability of each conveyor line, the conveyor path is planned based on the preset conveyor line association characteristics, the package multimodal information set, and the current conveyor status information of each conveyor line, and an optimized sorting and scheduling strategy is obtained to manage the logistics goods scheduling within the preset time zone.

[0013] Secondly, this application provides a feature-based intelligent logistics commodity scheduling and management system, including:

[0014] The feature evaluation module is used to evaluate the consistency of features based on the initial physical feature information of several packages recorded in the database within a preset time period, and to formulate a first feature recognition mechanism.

[0015] The barcode recognition module is used to recognize the barcodes of packages to be sorted on the conveyor production line. If the barcode recognition is successful, the first feature recognition mechanism is used to recognize the package image and obtain the package multimodal information.

[0016] The image recognition module is used to perform image recognition on packages that fail barcode recognition through a second feature recognition mechanism to obtain multimodal information about the packages if barcode recognition fails.

[0017] The probability prediction module is used to predict the probability of package congestion on each conveyor line within a preset time zone based on the multimodal information set of packages on the conveyor line in the current time period and the current conveying status information of each conveyor line.

[0018] The path optimization module is used to plan the transport path based on the constraints that the probability of package congestion on each conveyor line does not exceed a preset threshold, and to minimize the total transport time of packages and the variance of the predicted congestion probability of each conveyor line as dual optimization objectives. It plans the transport path according to the preset conveyor line association characteristics, the package multimodal information set, and the current transport status information of each conveyor line, and obtains an optimized sorting and scheduling strategy for the logistics goods scheduling management within the preset time zone.

[0019] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0020] This application provides a smart logistics commodity scheduling and management method and system based on feature recognition. First, a first feature recognition mechanism is dynamically formulated based on the consistency evaluation results of the physical characteristics of packages within a preset time period. This allows the feature extraction strategy to adaptively match the actual feature distribution of the current batch of packages, avoiding the problem of decreased recognition accuracy caused by the mismatch between the fixed parameter model and the actual package characteristics. Second, differentiated feature recognition mechanisms are adopted for packages that pass and fail barcode recognition. This ensures rapid processing of high-confidence packages and allows for in-depth physical feature analysis of abnormal packages through a second feature recognition mechanism, reducing the need for manual intervention. Third, by fusing multimodal information of packages with the real-time status of the conveyor line to predict congestion probability, and constructing an optimization model with transportation efficiency and load balancing as dual objectives, dynamic planning and load balancing scheduling of the conveyor path are achieved. This effectively alleviates systemic congestion caused by local conveyor line overload and improves the automation level and operational stability of the overall sorting system.

[0021] Through the above technical solutions, this application effectively solves the technical bottlenecks in traditional logistics sorting systems, such as insufficient robustness of barcode recognition, rigid visual feature extraction strategies, and lack of foresight in conveying path planning. It realizes the intelligent upgrade of the entire process from package feature perception to conveying path decision-making, and improves sorting efficiency, recognition accuracy, and system adaptability in smart logistics scenarios. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the intelligent logistics commodity scheduling and management method based on feature recognition provided in the embodiments of this application;

[0024] Figure 2 This is a schematic diagram of the structure of the intelligent logistics commodity scheduling and management system based on feature recognition provided in the embodiments of this application.

[0025] The components represented by each number in the attached diagram are explained below:

[0026] Feature evaluation module 11, barcode recognition module 12, image recognition module 13, probability prediction module 14, path optimization module 15. Detailed Implementation

[0027] This application provides a smart logistics commodity scheduling and management method and system based on feature recognition, which addresses the technical problems of low accuracy of commodity feature recognition, insufficient sorting efficiency, and high dependence on manual labor in existing logistics sorting methods, making it difficult to meet the automation requirements of smart logistics.

[0028] Example 1, as Figure 1 As shown in the embodiments of this application, a smart logistics commodity scheduling and management method based on feature recognition is provided, including:

[0029] S10: Based on the initial physical characteristic information of several packages recorded in the database within a preset time period, perform characteristic consistency assessment and formulate a first characteristic identification mechanism;

[0030] Furthermore, encoders are installed at predetermined positions on the conveyor line to record the displacement distance of the package on the conveyor line in real time using encoder pulse signals;

[0031] When a package passes through the barcode recognition station, the value of the first encoder is recorded and the barcode information of the package is obtained.

[0032] When a package passes through the visual recognition station, the value of the second encoder is recorded and a visual image of the package is captured. The visual recognition station is set at a preset distance behind the barcode recognition station.

[0033] The displacement distance of the package between the two workstations is calculated based on the difference between the values ​​of the first encoder and the second encoder. The correspondence between the barcode information and the visual image information is determined based on the displacement distance, so as to ensure that the identification feature information and physical feature information of the same package are aligned in time and space.

[0034] In this embodiment, the encoder can be an incremental photoelectric encoder, the resolution of which is determined according to the conveyor line's running speed and positioning accuracy requirements, typically set to 10 to 100 pulses per millimeter. The encoder is coaxially mounted with the conveyor line drive roller to ensure a linear correspondence between the pulse signal and the package displacement. When a package enters the initial station of the conveyor line, a unique timing identifier is assigned to each package. This identifier is bound to the cumulative number of encoder pulses, forming a position tracking reference for the package throughout the entire conveyor line journey.

[0035] The barcode recognition station is equipped with a dual-mode recognition device consisting of a high-speed industrial camera and a laser scanner. The industrial camera captures images of the package surface to assist in barcode positioning, while the laser scanner quickly decodes one-dimensional or two-dimensional barcodes. When a package triggers the photoelectric sensor at the barcode recognition station, the current encoder pulse count value is latched as the first encoder value, and the barcode recognition process is initiated. If the laser scanner successfully decodes the barcode within a preset time window, the decoding result is associated with and stored in relation to the first encoder value. If decoding fails, the package is marked as having an abnormal barcode status, and the type of abnormality is recorded, including missing barcode, damaged barcode, barcode wrinkles, or barcode obstruction.

[0036] The visual recognition station is equipped with a multi-view image acquisition array, including a top front-view camera, a side-view camera, and an optional bottom-view camera. Each camera uses a synchronous triggering method to ensure spatiotemporal consistency of the multi-view images. The preset distance between the visual recognition station and the barcode recognition station is determined based on the conveyor line speed and image processing latency, typically set to the displacement distance corresponding to 2 to 5 seconds of the package traveling on the conveyor line, ensuring that barcode recognition results are promptly fed back to the processing logic of the visual recognition station. When a package triggers the photoelectric sensor at the visual recognition station, the current encoder pulse count value is latched as the second encoder value, and multi-view image acquisition is initiated.

[0037] The displacement distance is calculated by multiplying the encoder pulse difference by the pulse equivalent. The pulse equivalent is determined by the joint calibration of the encoder resolution and the roller circumference. A dynamic package queue is maintained, and each entry in the queue includes a package sequence identifier, a first encoder value, barcode information, a second encoder value, visual image data, and a processing status marker. The continuity of package movement between two recognitions is verified by comparing the difference between the first and second encoder values ​​with the theoretical number of pulses for a preset distance. If the difference exceeds the allowable error range, a package tracking anomaly alarm is triggered, initiating manual verification or a return processing procedure.

[0038] Furthermore, a feature consistency assessment is conducted based on the initial physical characteristic information of several packages recorded in the database within a preset time period. The preset time period is selected based on the statistical characteristics of the conveyor line's historical operation data, typically set as the time interval from the start time of the current shift to the current time, or the time span of the N most recent completed sorting batches, where N ranges from 50 to 200 batches. The core of the feature consistency assessment lies in quantifying the degree of deviation between the current time period's package feature distribution and the historical baseline distribution.

[0039] Specifically, step S10 in the method includes:

[0040] The initial physical feature information of several packages within a preset time period is obtained from the database, wherein the initial physical feature information includes size features, shape features and texture features;

[0041] Based on the initial physical feature information of the several packages, the size feature consistency coefficient, shape feature consistency coefficient and texture feature consistency coefficient of the multiple packages within the preset time period are calculated respectively.

[0042] Based on the relationship between the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient and the corresponding preset threshold, a first feature recognition mechanism is formulated.

[0043] In this embodiment, firstly, a sample set of packages within a preset time period is extracted from the database. The sample size is determined based on statistical significance requirements, typically not less than 500 package records. Size features include measurements of the package's length, width, and height. Shape features are parametrically represented using Fourier descriptors or Zernike moments of the package's contour. Texture features are extracted based on statistics derived from the gray-level co-occurrence matrix or local binary patterns of the package's surface image. For each package, its initial physical feature information has been manually verified or obtained using precision measurement equipment during the data entry process, possessing high confidence and serving as the benchmark for subsequent feature consistency evaluation.

[0044] The coefficient of variation method is used to calculate the consistency coefficient of size features. The coefficients of variation of package length, width and height in the sample set are calculated separately, which are the ratios of standard deviation to mean. Then, the weighted average of the coefficients of variation of the three dimensions is taken as the consistency coefficient of size features. The weights are set according to the degree of influence of each dimension on sorting decisions. Usually, length and width have higher weights, while height has a relatively lower weight.

[0045] The calculation of the shape feature consistency coefficient is based on the clustering density of the shape descriptor vectors. The Fourier descriptor of the contour point set is used to encode the package shape into a low-dimensional feature vector. The average cosine similarity of the feature vectors of each package shape in the sample set or the clustering contour coefficient based on the Gaussian mixture model is calculated.

[0046] The texture feature consistency coefficient is calculated by quantifying the distribution dispersion of texture feature vectors. The chi-square distance or Earth movement distance between texture histograms extracted by local binary mode is used as a measure of distribution difference, and then the texture consistency index of the sample set is calculated.

[0047] Furthermore, based on the relationship between the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient and the corresponding preset threshold, specific rules for the first feature recognition mechanism are formulated.

[0048] Specifically, based on the relationship between the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient and their corresponding preset thresholds, a first feature recognition mechanism is established, including:

[0049] Preset size consistency threshold, shape consistency threshold, and texture consistency threshold;

[0050] When the size consistency coefficient is less than the size consistency threshold, only the recognition mechanism of the size feature extraction unit is activated;

[0051] When the size consistency coefficient is greater than or equal to the size consistency threshold and the shape consistency coefficient is less than the shape consistency threshold, the recognition mechanism of the size feature extraction unit and the shape feature extraction unit is activated.

[0052] When the size consistency coefficient is greater than or equal to the size consistency threshold and the shape consistency coefficient is greater than or equal to the shape consistency threshold, the recognition mechanism of the size feature extraction unit, the shape feature extraction unit and the texture feature extraction unit is activated, wherein the size feature extraction unit, the shape feature extraction unit and the texture feature extraction unit are constructed based on a convolutional neural network.

[0053] In this embodiment of the application, firstly, the initial values ​​of three consistency thresholds are determined based on historical data statistical analysis. The size consistency threshold is usually set in the range of 0.15 to 0.25, the shape consistency threshold is set in the range of 0.20 to 0.30, and the texture consistency threshold is set in the range of 0.25 to 0.35. The thresholds are not fixed, but are adaptively adjusted. The adjustment cycle is usually set to fine-tune the thresholds based on the recognition accuracy feedback after every 1,000 packages are sorted.

[0054] Furthermore, when the size consistency coefficient is less than the size consistency threshold, it indicates that the current batch of packages has a high degree of consistency in size dimension, and the difference in length, width and height between packages is small. At this time, it is only necessary to activate the size feature extraction unit to complete reliable package differentiation.

[0055] Specifically, the size feature extraction unit employs a lightweight convolutional neural network structure. The input layer receives a top-view image of the package, and after feature extraction via two 3×3 convolutional kernels, the predicted length, width, and height of the package are output through a fully connected layer. During training, the network uses a large number of labeled package images for supervised learning. The loss function is a weighted sum of the mean square error and relative error between the predicted and actual sizes, ensuring the model's prediction accuracy for packages of various sizes. The strategy of activating only the size feature extraction unit significantly reduces computational resource consumption, and the processing latency for a single package image can be controlled within 50 milliseconds, meeting the real-time requirements of high-speed transport lines.

[0056] Furthermore, when the size consistency coefficient is greater than or equal to the size consistency threshold, it indicates that there is significant variation in the size dimension of the current batch of packages, and accurate differentiation cannot be achieved solely based on size features. In this case, it is necessary to further examine the consistency of shape features. If the shape consistency coefficient is less than the shape consistency threshold, it indicates that the shape features of the packages are relatively consistent, and the size variation mainly comes from different specifications of standard-shaped packages. At this time, the joint recognition mechanism of the size feature extraction unit and the shape feature extraction unit is activated.

[0057] Specifically, the shape feature extraction unit, based on the intermediate feature map of the size feature extraction unit, adds edge detection and contour encoding branches to extract the contour curvature features and corner distribution features of the package. The feature vectors of the two units are concatenated in the fusion layer and then input into the joint classifier, which outputs the category confidence score of the package. The computational complexity of the joint recognition mechanism is about 40% higher than that of single-size recognition, but the recognition accuracy can be improved by 8% to 12%, which has a significant advantage for batches of packages with regular shapes but diverse sizes.

[0058] Furthermore, when both the size consistency coefficient and the shape consistency coefficient are greater than or equal to the corresponding threshold, it indicates that the current batch of packages exhibits high heterogeneity in both size and shape dimensions, and may include irregularly shaped packages, soft and deformable packages, or packages with significantly different surface packaging. In this case, the complete three-feature extraction unit identification mechanism must be activated.

[0059] Specifically, the texture feature extraction unit adopts a multi-scale convolution structure, which captures the detailed texture and global texture pattern of the wrapped surface by setting dilated convolutions with different dilation rates in parallel. At the same time, an attention mechanism is introduced to weight the salient texture areas and suppress background noise interference.

[0060] The output features of the three units are adaptively weighted and fused in a deep fusion network. The weight parameters are dynamically adjusted according to the current batch feature distribution to ensure that the contribution of each feature matches its actual distinguishing ability. Although the complete recognition mechanism has the largest computational overhead, with a single package processing latency of approximately 150 to 200 milliseconds, it can achieve an accuracy rate of over 98% for complex and heterogeneous package batches, which is more than 15 percentage points higher than the traditional fixed strategy.

[0061] S20: Barcode recognition is performed on the packages to be sorted on the conveyor line. If the barcode recognition is successful, the package image is recognized using the first feature recognition mechanism to obtain multimodal information of the package.

[0062] In this embodiment, the barcode recognition result determines the subsequent feature recognition mechanism's invocation strategy. Packages awaiting sorting on the conveyor line first pass through the barcode recognition station. A laser scanner performs a full-coverage scan of the package surface at a scanning frequency of 500 to 1000 times per second, while an industrial camera simultaneously captures images of the package surface for barcode area positioning and decoding assistance. After the laser scanner successfully decodes the barcode, it extracts structured data such as the package's logistics order number, destination code, and weight information, and associates the recognition result with the first encoder value, writing it into the dynamic package queue.

[0063] Secondly, after the barcode recognition is successful, the corresponding image recognition process is triggered according to the currently active first feature recognition mechanism configuration. If the mechanism is configured to activate only the size feature extraction unit, the top front-view camera of the visual recognition station is controlled to acquire single-view images at the highest frame rate, while the side-view camera and the bottom perspective camera are in standby power-saving mode. If the configuration is dual-unit or triple-unit joint recognition, the multi-view image acquisition array is triggered synchronously, and each camera ensures that the exposure time deviation is less than 1 millisecond through hardware synchronization signals.

[0064] Specifically, the multimodal information of a package comprises three layers: the base layer consists of identification information obtained from barcode parsing, including sender code, recipient code, service type code, and timeliness level marker; the feature layer consists of physical feature information extracted from image recognition, which, depending on the activation unit, includes size prediction values, shape description vectors, texture feature maps, and their weighted fusion representations; and the association layer consists of spatiotemporal alignment information, including package time sequence identifiers, encoder position tracking data, processing timestamps at each workstation, and recognition confidence scores. This multimodal information is stored in a distributed caching system in the form of structured data packets, supporting real-time querying and batch reading by downstream sorting decision modules.

[0065] S30: If the barcode recognition fails, the second feature recognition mechanism is used to perform image recognition on the package that failed the barcode recognition to obtain the package's multimodal information;

[0066] In this embodiment, barcode recognition failures include situations where the laser scanner fails to decode any valid barcode within a preset time window, or the decoding result verification fails, such as incorrect verification bits or non-compliant format. When a package is marked as having an abnormal barcode status, a second feature recognition mechanism is triggered. The core difference between this mechanism and the first feature recognition mechanism lies in the aggressiveness and fault-tolerant design of the recognition strategy.

[0067] The second feature recognition mechanism is used to perform image recognition on packages that have not passed barcode recognition, including:

[0068] A physical feature extraction model library is established, wherein the model library includes size feature extraction models, shape feature extraction models, and texture feature extraction models built based on convolutional neural networks;

[0069] Based on the physical feature information of multiple packages within the current time period, the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient are calculated, and the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient are weighted and evaluated to obtain the comprehensive physical feature consistency coefficient.

[0070] The number of physical feature extraction models to be called is determined based on the comprehensive consistency coefficient of the physical features. The higher the comprehensive consistency coefficient of the physical features, the more physical feature extraction models are called.

[0071] In this embodiment, firstly, the physical feature extraction model library adopts a modular architecture design, with each model trained and version managed independently, supporting hot-update deployment without interrupting production line operation. The size feature extraction model uses a lightweight ResNet-18 backbone network, with the input image resolution normalized to 512×512 pixels, and outputs normalized predicted values ​​of the length, width, and height. The shape feature extraction model is built based on the HRNet high-resolution network, maintaining the full-resolution representation of the feature map to accurately capture contour details, and outputs a 128-dimensional shape description vector. The texture feature extraction model adopts the EfficientNet-B0 architecture, combined with a feature pyramid network to achieve multi-scale texture perception, and outputs a 256-dimensional texture embedding vector.

[0072] Secondly, the calculation of the physical feature consistency coefficient adopts a weighted fusion strategy, and the weight coefficient reflects the distinguishing contribution of each feature dimension in the current time period.

[0073] For example, the weight of the size feature consistency coefficient is set to 0.35, the weight of the shape feature consistency coefficient is set to 0.40, and the weight of the texture feature consistency coefficient is set to 0.25. The weight configuration is adjusted annually based on the information gain ratio of each feature dimension in the historical sorting data. The value of the overall consistency coefficient ranges from 0 to 1. A higher value indicates a greater dispersion in the distribution within the multi-dimensional feature space during the current period, making it difficult for a single feature model to cover all variation patterns.

[0074] Secondly, the number of models called is determined based on the overall consistency coefficient of physical features. Specifically, when the overall consistency coefficient is below 0.30, a single size feature extraction model is called. At this time, the package feature distribution is highly concentrated, and the size dimension can effectively distinguish the package. When the overall consistency coefficient is between 0.30 and 0.60, the size feature extraction model and the shape feature extraction model are called in parallel, and the output features of the two models are fused in the early stage at the decision layer. When the overall consistency coefficient is above 0.60, all three physical feature extraction models are activated, and the recognition results of the three models are fused in a multimodal manner to obtain the comprehensive recognition result of the package.

[0075] S40: Based on the multimodal information set of packages on the conveyor line in the current time period, combined with the current conveying status information of each conveyor line, predict and obtain the probability of package congestion on each conveyor line in the preset time zone.

[0076] In this embodiment, destination information and physical feature classification information of the conveyor line in the current time period are first obtained. Combined with the current conveying status information of each conveyor line, a congestion prediction model is built and trained based on a deep learning physical model, so as to predict the probability of package congestion of each conveyor line in a preset time zone.

[0077] Specifically, step S40 in the method includes:

[0078] Obtain destination information and physical characteristic classification information of all packages within the current time period, and count the number of packages and the distribution of package types for each conveyor line;

[0079] Obtain real-time conveying status information for each conveyor line, wherein the real-time conveying status information includes current load rate, equipment operating status, and historical congestion data;

[0080] The number of packages, the distribution of package types, and the real-time delivery status information are input into the congestion prediction model to analyze and obtain the predicted congestion probability of each delivery line in the preset time zone.

[0081] The steps for constructing the congestion prediction model include:

[0082] Based on the historical operation records of each conveyor line, several sample package quantities, several package type distributions, and several real-time conveyor status information distributions were collected. The proportion of congestion events that occurred on each conveyor line within the historical time zone was statistically analyzed as the sample predicted congestion probability, resulting in several sample predicted congestion probability distributions.

[0083] Using the number of sample packages, the distribution of package types, and the distribution of real-time delivery status information as inputs, and using the predicted congestion probability distribution of the sample packages as supervision, a deep learning model is trained until convergence to generate a congestion prediction model.

[0084] In this embodiment, firstly, the parsing of destination information relies on the destination encoding field in the barcode decoding result. This field follows the postal code standard or the logistics company's custom partitioning encoding system, dividing the service area into several levels of delivery grids. The physical feature classification information comes from the output of the aforementioned feature recognition mechanism, including the package's size level, shape category label, and texture type identifier. The classification information is used to estimate the space occupancy characteristics of the package during transportation and the smoothness of its passage through bottleneck stations.

[0085] Secondly, real-time conveying status information is collected through a multi-source sensor network deployed at key nodes of the conveyor line. The current load rate is calculated using fusion data from a photoelectric counting array and a weighing module, reflecting the packing density per unit length of the conveyor belt; equipment operating status includes drive motor current fluctuations, belt tension monitoring values, and response delay indicators of the sorting actuator; historical congestion data is extracted from an event log database, recording the timestamps, durations, and cause classifications of congestion that occurred on each conveyor line in the past 30 days.

[0086] Furthermore, the training data for the congestion prediction model follows a spatiotemporal sliding window strategy, using a 15-minute time granularity. Samples are extracted from historical operation records, and each sample includes statistics on the number of packages at the start of that time zone, a histogram of package type distribution, and a real-time state feature vector. The sample label is a binary indicator variable indicating whether a congestion event occurred within 30 minutes after the end of that time zone, or a normalized value of the congestion duration. The deep learning model employs a hybrid architecture of temporal convolutional networks and long short-term memory networks. The temporal convolutional layers capture multi-scale periodic patterns of transport status, the LSTM layers model the dynamic cumulative effect of package inflow, and the output layer generates congestion probability predictions through a sigmoid activation function.

[0087] Furthermore, a congestion prediction model is constructed based on a deep learning network model. First, multi-source operation data of each transport line are collected in continuous historical periods to construct a spatiotemporal feature matrix as the model input. The data collection range covers the complete operation cycle of the past 90 days, and the time resolution is uniformly set to 5 minutes to ensure that the model can learn multiple time scale features such as intraday peak fluctuations, intraweek cycle patterns, and seasonal trends.

[0088] Specifically, the construction of the spatiotemporal feature matrix includes three dimensions: the time dimension features include the hour index of the current moment, the week type, whether it is a holiday, and the offset from the nearest peak period; the spatial dimension features include the topological location of the conveyor line, the upstream and downstream connection relationship, the sorting port capacity configuration, and the coupling strength index with other conveyor lines; the state dimension features cover the dynamic sequence of the aforementioned package quantity, type distribution, and real-time conveying status information, with the sequence length set to 12 time steps, which includes the complete state evolution trajectory of the past hour.

[0089] Furthermore, the deep learning network model adopts an encoder-decoder architecture. The encoder consists of three stacked spatiotemporal graph convolutional networks. Each graph convolutional layer uses the topology of the transport lines as its adjacency matrix, aggregating the state information of adjacent transport lines to achieve spatial propagation modeling of congestion risk. In the temporal dimension, the outputs of each graph convolutional layer are sequentially encoded through gated recurrent units to extract the long-term dependencies of state evolution. The decoder employs a multi-step prediction structure enhanced by an attention mechanism. For the next six time steps (i.e., a preset time zone of 30 minutes), it outputs congestion probability predictions and calculates the uncertainty estimates for each time step prediction.

[0090] During model training, the loss function is designed as a combination of weighted binary cross-entropy and prediction interval coverage. Positive samples are given higher weights to solve the class imbalance problem, while the calibration of prediction probability is constrained to avoid overconfident extreme predictions. The optimal model on the validation set achieves a recall rate of 85% for congestion events while keeping the false positive rate below 12%, meeting the requirements of prediction reliability in the production environment.

[0091] S50: With the constraint that the probability of package congestion on each conveyor line does not exceed a preset threshold, and with the dual optimization objectives of minimizing the total package transportation time and minimizing the variance of the predicted congestion probability of each conveyor line, the conveyor path is planned based on the preset conveyor line association characteristics, the package multimodal information set, and the current conveyor status information of each conveyor line, and an optimized sorting and scheduling strategy is obtained to manage the logistics goods scheduling within the preset time zone.

[0092] In this embodiment, the goal of minimizing the total parcel transportation time includes the total time cost of a parcel traveling from its current location through the sorting network to its destination, including the cumulative time of conveyor line transmission, sorting mechanism execution time, and node queuing waiting time. The goal of minimizing the variance of the predicted congestion probability for each conveyor line focuses on global load balancing, avoiding resource allocation imbalances where some conveyor lines are overloaded while others are idle. The variance calculation is based on the congestion probability distribution of all candidate conveyor lines within the preset time zone. The preset threshold is set based on the conveyor line's design throughput capacity and historical peak load, typically ranging from 0.70 to 0.85. Conveyor lines exceeding this threshold are prohibited from receiving new parcels within the time zone.

[0093] Specifically, the steps for constructing the preset conveyor line association features include:

[0094] Construct the topology of the conveyor network and record the set of destinations that each conveyor line can reach;

[0095] For each destination, identify all transport lines that can reach that destination and establish a set of transport line substitution relationships;

[0096] Calculate the substitution cost between the alternative conveyor line and the original conveyor line, wherein the substitution cost includes the path length increment, the expected transportation time increment, and the number of transfers;

[0097] The conveyor line substitution relationships and corresponding substitution costs are stored as a conveyor line association feature matrix, and a preset conveyor line association feature is generated.

[0098] In this embodiment, the conveyor network topology is first constructed based on a digital model of the physical layout of the logistics sorting center. This model is created using a fusion of laser ranging and BIM technology, recording the spatial coordinates, direction angles, connection node positions, and interface relationships with other automated equipment for each conveyor line. Each conveyor line is assigned a unique identifier and its engineering parameters, such as design operating speed, maximum load capacity, and effective conveying length, are labeled. The destination set is determined based on the reachability of the sorting port or transfer channel connected to the end of the conveyor line. A path search algorithm traverses the sorting network to calculate the codes of all delivery grids that can be covered starting from the conveyor line.

[0099] Secondly, the set of alternative conveyor line relationships is established using a reverse indexing strategy. For each destination code, all conveyor line identifiers reachable from that destination are queried from the network topology database, forming a list of candidate conveyor lines for that destination. The validity of the alternative relationships must meet real-time status verification, meaning that the candidate conveyor line is not currently in a fault shutdown or maintenance lockout state, and its predicted congestion probability is lower than a preset threshold.

[0100] Furthermore, the quantitative calculation of substitution costs comprehensively considers three types of cost factors: path length increment is measured by the difference in the physical length of the conveyor line, reflecting additional equipment wear and energy costs; the estimated transportation time increment is calculated based on the conveyor line's operating speed and path length, adding the estimated waiting time from the current queue length; and the number of transfers is counted to determine the number of cross-sorting or manual handling nodes that a package must pass through to switch from the original conveyor line to the alternative line. A fixed time penalty coefficient is introduced for each transfer to reflect operational complexity and error risk. After normalization, the three types of costs are combined into a comprehensive substitution cost using a weighted summation method. The weighting is dynamically adjusted according to the operational strategy, with increased weighting for time-sensitive businesses and increased weighting for cost-sensitive businesses. For example, when a batch of fresh cold chain packages needs to be prioritized, the time increment weight can be temporarily increased to 0.60, while the standard e-commerce package weighting remains at 0.30. The substitution cost calculation results are stored in a sparse matrix structure, retaining only the non-zero elements corresponding to effective substitution relationships to reduce the storage and computational overhead of subsequent path planning algorithms.

[0101] The process involves several key elements: First, a predefined threshold is set as the constraint that the probability of package congestion on each conveyor line does not exceed a preset threshold. Second, the dual optimization objectives are minimizing the total package transportation time and minimizing the variance of the predicted congestion probability for each conveyor line. Third, the transport path is planned based on preset conveyor line association characteristics, the package multimodal information set, and the current transport status information of each conveyor line.

[0102] With the constraint that the predicted congestion probability of each conveyor line does not exceed a preset threshold, conveyor lines that meet the constraint are selected as candidate paths.

[0103] For all packages within a preset time zone, the packages are assigned to candidate delivery lines that meet the constraints, and multiple candidate path allocation schemes are constructed.

[0104] Calculate the estimated total transportation time for all packages under each candidate scheme, wherein the estimated total transportation time is determined based on the current location of the packages, the length of the conveyor line, and the operating speed of the conveyor line;

[0105] Calculate the variance of the predicted congestion probability for each transport line under each candidate scheme;

[0106] A multi-objective optimization algorithm is used to select the optimal route allocation scheme from the candidate schemes that simultaneously minimizes the total parcel transportation time and the variance of the predicted congestion probability, while satisfying the constraints.

[0107] In this embodiment, firstly, the hard boundary handling of constraints adopts a dynamic threshold adjustment mechanism. The preset threshold value is not a fixed constant, but is adaptively corrected according to the real-time operational situation. When it is detected that the current load rate of a certain conveyor line has exceeded 80% of its design capacity, the congestion probability threshold of that line is temporarily lowered by 0.05 to 0.10, forming a preventive protection barrier. Conversely, for conveyor lines that have just completed maintenance or are in off-peak hours, the threshold can be appropriately raised to make full use of idle capacity. The candidate path screening process simultaneously performs feasibility verification, excluding conveyor lines temporarily closed due to equipment failure, emergency maintenance, or manual intervention, ensuring the timeliness and effectiveness of the candidate set.

[0108] Secondly, the candidate route allocation scheme is constructed using a hierarchical combination strategy. The first layer pre-allocates high-priority packages based on their timeliness commitment level, prioritizing same-day delivery and next-morning delivery packages to the transport routes with the lowest congestion probability, forming the basic allocation framework. The second layer allocates standard time-sensitive packages using a fill-in allocation within the remaining capacity space, initializing multiple feasible schemes through heuristic rules as the starting point for multi-objective optimization. Each candidate scheme is encoded in the form of a package-transport line allocation matrix, where the matrix elements are binary decision variables indicating whether a specific package is allocated to a specific transport line. The constraint requires that each row has exactly one non-zero element, meaning that each package must be allocated to one and only one transport line.

[0109] Secondly, the calculation of the estimated total transportation time introduces a dynamic time estimation model, which differs from the static theoretical transmission time. This model comprehensively considers the current queue length on the conveyor line, the parcel inflow rate during the predicted period, and the processing capacity bottleneck of the sorting execution mechanism. Specifically, the waiting time of parcels on the conveyor line is approximately estimated using the M / M / c queuing model, where the number of service stations c corresponds to the parallel processing capacity of the sorting station, the arrival rate λ is taken from the parcel inflow prediction output by the congestion prediction model, and the service rate μ is obtained based on historical data statistics. For paths that require passing through transfer nodes, an additional transfer waiting time is added. This time follows an empirical distribution based on historical data, and the 95th percentile value is taken as a conservative estimate to ensure the reliability of fulfilling the time commitment.

[0110] Furthermore, the calculation of the predicted congestion probability variance adopts a weighted form, using the designed throughput capacity of each transport line as the weight, to avoid excessive impact of probability fluctuations of low-capacity lines on the overall balance index. The goal of minimizing variance is essentially to promote the evolution of parcel allocation towards probability flattening, so that the congestion risk of each transport line tends to be consistent, rather than excessively concentrating the load on a few low-probability lines.

[0111] The multi-objective optimization algorithm employs a decomposition-based evolutionary algorithm framework, transforming the bi-objective optimization problem into several single-objective sub-problems. A neighborhood cooperation mechanism is used to achieve uniform search of the Pareto front. To address the real-time requirements of logistics scheduling scenarios, the algorithm sets dual termination conditions: a maximum number of iterations and the number of objective function evaluations. Under typical configurations, it outputs an approximate Pareto optimal solution set after 500 generations of evolution or 100,000 evaluations.

[0112] Furthermore, a multi-objective optimization algorithm is employed to select the optimal route allocation scheme from the candidate schemes that simultaneously minimizes the total parcel transportation time and the variance of the predicted congestion probability, while satisfying the constraints. This includes:

[0113] Weighting coefficients are set for the objectives of minimizing the total parcel delivery time and minimizing the variance of the predicted congestion probability.

[0114] Multiple candidate path allocation schemes are generated through iterative search, and the search direction is adjusted in each iteration based on the optimization target value of the current scheme.

[0115] After each iteration, evaluate whether the current scheme satisfies the constraints and eliminate any scheme that causes the predicted congestion probability of any conveyor line to exceed a preset threshold.

[0116] When the preset iteration termination condition is met, the optimal path allocation scheme is selected from all candidate schemes that meet the constraints, based on the scheme with the best comprehensive optimization objective value.

[0117] In this embodiment, firstly, the setting of weight coefficients incorporates a human-machine collaborative decision-making mechanism, providing recommended weight configurations based on historical operational data. For example, during high-traffic promotional periods, the system automatically switches to a time-priority mode, setting the transportation time weight to 0.70 and the variance weight to 0.30; while during regular operating periods, a balanced mode is adopted, with each objective weighting 0.50. Operations managers can adjust the weight allocation in real time through a visual interface, receiving immediate feedback on changes in the Pareto frontier after adjustment, assisting decision-makers in understanding the trade-offs under different preference orientations. The dynamic adjustment of weight coefficients also incorporates a reinforcement learning module, continuously optimizing the weight configuration strategy based on feedback from key performance indicators after actual scheduling execution, forming a closed-loop learning mechanism of "setting-execution-evaluation-correction".

[0118] Secondly, the iterative search process employs an adaptive variable neighborhood search strategy, dynamically switching search operators in different regions of the solution space. When the algorithm detects that the current search region is trapped in a local optimum, it automatically triggers a large-span perturbation operator, randomly exchanging package allocation relationships or reconstructing the allocation combination of the entire transport line to escape the local convergence trap. In the promising region, a fine-grained search operator is activated to perform neighborhood fine-tuning for package migration on the bottleneck transport line. The adjustment of the search direction is based on the gradient information of the dual objective function. The dimensional differences are eliminated through normalization, and a comprehensive improvement direction vector is calculated to guide the search toward the Pareto optimal front.

[0119] Furthermore, the constraint satisfaction assessment employs a two-tiered mechanism combining pre-screening and fine-tuning. The pre-screening phase utilizes the rapid reasoning capabilities of the congestion prediction model to batch-screen candidate solutions, eliminating those that clearly violate threshold constraints. The fine-tuning phase calls upon a discrete event simulation engine to perform micro-level dynamic simulations of the pre-screened solutions, simulating the actual flow process within the transport network and capturing transient congestion phenomena that the prediction model might overlook. The simulation time window covers the entire preset time zone, progressing in 1-minute increments, recording the instantaneous load rate change curves of each transport line. Any overload observed at any point indicates that the solution is infeasible. This dual verification mechanism of "prediction + simulation" significantly improves the reliability of constraint assessment, reducing the probability of infeasible solutions entering the actual execution phase to below 2%.

[0120] When the average Hamming distance of the population over several generations is lower than the threshold, it indicates that the population has converged significantly and further iterations are unlikely to yield significant improvements. The search can be terminated early, and the optimal path allocation scheme can be selected from all candidate schemes that meet the constraints, based on the optimal comprehensive optimization objective value.

[0121] In summary, compared with existing technologies, this application realizes a paradigm shift from passive response to proactive prevention in scheduling by constructing a multimodal feature fusion-based commodity recognition system and a spatiotemporally coupled congestion prediction mechanism.

[0122] In summary, the embodiments of this application have at least the following technical effects:

[0123] This application provides a smart logistics commodity scheduling and management method based on feature recognition. First, a first feature recognition mechanism is dynamically formulated based on the consistency evaluation results of the physical characteristics of packages within a preset time period. This allows the feature extraction strategy to adaptively match the actual feature distribution of the current batch of packages, avoiding the problem of decreased recognition accuracy caused by the mismatch between the fixed parameter model and the actual package features. Second, differentiated feature recognition mechanisms are adopted for packages that pass and fail barcode recognition. This ensures the rapid processing of high-confidence packages and allows for in-depth physical feature analysis of abnormal packages through a second feature recognition mechanism, reducing the need for manual intervention. Third, by fusing multimodal information of packages with the real-time status of the conveyor line to predict congestion probability, and constructing an optimization model with transportation efficiency and load balancing as dual objectives, dynamic planning and load balancing scheduling of the conveyor path are realized. This effectively alleviates systemic congestion caused by local conveyor line overload and improves the automation level and operational stability of the overall sorting system.

[0124] Through the above technical solutions, this application effectively solves the technical bottlenecks in traditional logistics sorting systems, such as insufficient robustness of barcode recognition, rigid visual feature extraction strategies, and lack of foresight in conveying path planning. It realizes the intelligent upgrade of the entire process from package feature perception to conveying path decision-making, and improves sorting efficiency, recognition accuracy, and system adaptability in smart logistics scenarios.

[0125] Example 2, as Figure 2 As shown, based on the same inventive concept as the feature-recognition-based intelligent logistics commodity scheduling and management method provided in Embodiment 1, this application also provides a feature-recognition-based intelligent logistics commodity scheduling and management system, including:

[0126] The feature evaluation module 11 is used to evaluate the consistency of features based on the initial physical feature information of several packages recorded in the database within a preset time period, and to formulate a first feature recognition mechanism.

[0127] The barcode recognition module 12 is used to recognize the barcodes of packages to be sorted on the conveyor production line. If the barcode recognition is successful, the first feature recognition mechanism is used to recognize the package image and obtain the package multimodal information.

[0128] Image recognition module 13 is used to perform image recognition on packages that fail barcode recognition through a second feature recognition mechanism if barcode recognition fails, and to obtain multimodal information of the packages.

[0129] The probability prediction module 14 is used to predict the probability of package congestion on each conveyor line within a preset time zone based on the multimodal information set of packages on the conveyor line in the current time period and the current conveying status information of each conveyor line.

[0130] The path optimization module 15 is used to plan the transport path based on the preset transport line association characteristics, the package multimodal information set, and the current transport status information of each transport line, with the constraint that the package congestion probability of each transport line does not exceed a preset threshold value, and with the dual optimization objectives of minimizing the total transport time of packages and minimizing the variance of the predicted congestion probability of each transport line. The optimized sorting and scheduling strategy is used to manage the logistics goods scheduling within the preset time zone.

[0131] Furthermore, in one embodiment of the application, the method further includes:

[0132] An encoder is installed at a predetermined position on the conveyor line, and the displacement distance of the package on the conveyor line is recorded in real time by the encoder pulse signal;

[0133] When a package passes through the barcode recognition station, the value of the first encoder is recorded and the barcode information of the package is obtained.

[0134] When a package passes through the visual recognition station, the value of the second encoder is recorded and a visual image of the package is captured. The visual recognition station is set at a preset distance behind the barcode recognition station.

[0135] The displacement distance of the package between the two workstations is calculated based on the difference between the values ​​of the first encoder and the second encoder. The correspondence between the barcode information and the visual image information is determined based on the displacement distance, so as to ensure that the identification feature information and physical feature information of the same package are aligned in time and space.

[0136] In one embodiment, the feature evaluation module 11 is specifically used for:

[0137] The initial physical feature information of several packages within a preset time period is obtained from the database, wherein the initial physical feature information includes size features, shape features and texture features;

[0138] Based on the initial physical feature information of the several packages, the size feature consistency coefficient, shape feature consistency coefficient and texture feature consistency coefficient of the multiple packages within the preset time period are calculated respectively.

[0139] Based on the relationship between the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient and the corresponding preset threshold, a first feature recognition mechanism is formulated.

[0140] Furthermore, in one embodiment, a first feature recognition mechanism is established based on the relationship between the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient and corresponding preset thresholds, including:

[0141] Preset size consistency threshold, shape consistency threshold, and texture consistency threshold;

[0142] When the size consistency coefficient is less than the size consistency threshold, only the recognition mechanism of the size feature extraction unit is activated;

[0143] When the size consistency coefficient is greater than or equal to the size consistency threshold and the shape consistency coefficient is less than the shape consistency threshold, the recognition mechanism of the size feature extraction unit and the shape feature extraction unit is activated.

[0144] When the size consistency coefficient is greater than or equal to the size consistency threshold and the shape consistency coefficient is greater than or equal to the shape consistency threshold, the recognition mechanism of the size feature extraction unit, the shape feature extraction unit and the texture feature extraction unit is activated, wherein the size feature extraction unit, the shape feature extraction unit and the texture feature extraction unit are constructed based on a convolutional neural network.

[0145] Furthermore, in one embodiment, image recognition of packages that fail barcode recognition is performed using a second feature recognition mechanism, including:

[0146] A physical feature extraction model library is established, wherein the model library includes size feature extraction models, shape feature extraction models, and texture feature extraction models built based on convolutional neural networks;

[0147] Based on the physical feature information of multiple packages within the current time period, the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient are calculated, and the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient are weighted and evaluated to obtain the comprehensive physical feature consistency coefficient.

[0148] The number of physical feature extraction models to be called is determined based on the comprehensive consistency coefficient of the physical features. The higher the comprehensive consistency coefficient of the physical features, the more physical feature extraction models are called.

[0149] Furthermore, the probability prediction module 14 is specifically used for:

[0150] Obtain destination information and physical characteristic classification information of all packages within the current time period, and count the number of packages and the distribution of package types for each conveyor line;

[0151] Obtain real-time conveying status information for each conveyor line, wherein the real-time conveying status information includes current load rate, equipment operating status, and historical congestion data;

[0152] The number of packages, the distribution of package types, and the real-time delivery status information are input into the congestion prediction model to analyze and obtain the predicted congestion probability of each delivery line in the preset time zone.

[0153] The steps for constructing the congestion prediction model include:

[0154] Based on the historical operation records of each conveyor line, several sample package quantities, several package type distributions, and several real-time conveyor status information distributions were collected. The proportion of congestion events that occurred on each conveyor line within the historical time zone was statistically analyzed as the sample predicted congestion probability, resulting in several sample predicted congestion probability distributions.

[0155] Using the number of sample packages, the distribution of package types, and the distribution of real-time delivery status information as inputs, and using the predicted congestion probability distribution of the sample packages as supervision, a deep learning model is trained until convergence to generate a congestion prediction model.

[0156] Furthermore, the steps for constructing the preset conveyor line association features include:

[0157] Construct the topology of the conveyor network and record the set of destinations that each conveyor line can reach;

[0158] For each destination, identify all transport lines that can reach that destination and establish a set of transport line substitution relationships;

[0159] Calculate the substitution cost between the alternative conveyor line and the original conveyor line, wherein the substitution cost includes the path length increment, the expected transportation time increment, and the number of transfers;

[0160] The conveyor line substitution relationships and corresponding substitution costs are stored as a conveyor line association feature matrix, and a preset conveyor line association feature is generated.

[0161] Furthermore, in one embodiment, with the constraint that the probability of package congestion on each conveyor line does not exceed a preset threshold, and with the dual optimization objectives of minimizing the total package transportation time and minimizing the variance of the predicted congestion probability of each conveyor line, transport path planning is performed based on preset conveyor line association characteristics, the package multimodal information set, and the current transport status information of each conveyor line, including:

[0162] With the constraint that the predicted congestion probability of each conveyor line does not exceed a preset threshold, conveyor lines that meet the constraint are selected as candidate paths.

[0163] For all packages within a preset time zone, the packages are assigned to candidate delivery lines that meet the constraints, and multiple candidate path allocation schemes are constructed.

[0164] Calculate the estimated total transportation time for all packages under each candidate scheme, wherein the estimated total transportation time is determined based on the current location of the packages, the length of the conveyor line, and the operating speed of the conveyor line;

[0165] Calculate the variance of the predicted congestion probability for each transport line under each candidate scheme;

[0166] A multi-objective optimization algorithm is used to select the optimal route allocation scheme from the candidate schemes that simultaneously minimizes the total parcel transportation time and the variance of the predicted congestion probability, while satisfying the constraints.

[0167] Furthermore, a multi-objective optimization algorithm is employed to select the optimal route allocation scheme from the candidate schemes that simultaneously minimizes the total parcel transportation time and the variance of the predicted congestion probability, while satisfying the constraints. This includes:

[0168] Weighting coefficients are set for the objectives of minimizing the total parcel delivery time and minimizing the variance of the predicted congestion probability.

[0169] Multiple candidate path allocation schemes are generated through iterative search, and the search direction is adjusted in each iteration based on the optimization target value of the current scheme.

[0170] After each iteration, evaluate whether the current scheme satisfies the constraints and eliminate any scheme that causes the predicted congestion probability of any conveyor line to exceed a preset threshold.

[0171] When the preset iteration termination condition is met, the optimal path allocation scheme is selected from all candidate schemes that meet the constraints, based on the scheme with the best comprehensive optimization objective value.

Claims

1. A smart logistics commodity scheduling and management method based on feature recognition, characterized in that, The methods include: Based on the initial physical characteristics of several packages recorded in the database within a preset time period, a characteristic consistency assessment is performed, and a first characteristic recognition mechanism is established. The packages to be sorted on the conveyor line are identified by barcodes. If the barcode identification is successful, the package image is identified using the first feature recognition mechanism to obtain multimodal information of the package. If the barcode recognition fails, the second feature recognition mechanism is used to perform image recognition on the package that failed the barcode recognition to obtain the package's multimodal information; Based on the multimodal information set of packages on the conveyor lines during the current time period, combined with the current conveyor status information of each conveyor line, the probability of package congestion on each conveyor line within a preset time zone is predicted, including: Obtain destination information and physical characteristic classification information of all packages within the current time period, and count the number of packages and the distribution of package types for each conveyor line; Obtain real-time conveying status information for each conveyor line, wherein the real-time conveying status information includes current load rate, equipment operating status, and historical congestion data; The number of packages, the distribution of package types, and the real-time delivery status information are input into the congestion prediction model to analyze and obtain the predicted congestion probability of each delivery line in the preset time zone. The steps for constructing the congestion prediction model include: Based on the historical operation records of each conveyor line, several sample parcel quantities, several parcel type distributions, and several real-time conveying status information distributions were collected. The proportion of congestion events that occurred on each conveyor line within the historical time zone was statistically analyzed as the sample predicted congestion probability, resulting in several sample predicted congestion probability distributions. Using the distribution of the number of sample packages, the distribution of package types, and the distribution of real-time delivery status information as inputs, and using the distribution of predicted congestion probability of the sample packages as supervision, a deep learning model is trained until convergence to generate a congestion prediction model. With the constraint that the probability of package congestion on each conveyor line does not exceed a preset threshold, and with the dual optimization objectives of minimizing the total package transportation time and minimizing the variance of the predicted congestion probability of each conveyor line, the conveyor path is planned based on the preset conveyor line association characteristics, the package multimodal information set, and the current conveyor status information of each conveyor line, and an optimized sorting and scheduling strategy is obtained to manage the logistics goods scheduling within the preset time zone. The steps for constructing the preset conveyor line association features include: Construct the topology of the conveyor network and record the set of destinations that each conveyor line can reach; For each destination, identify all transport lines that can reach that destination and establish a set of transport line substitution relationships; Calculate the substitution cost between the alternative conveyor line and the original conveyor line, wherein the substitution cost includes the path length increment, the expected transportation time increment, and the number of transfers; The conveyor line substitution relationships and corresponding substitution costs are stored as a conveyor line association feature matrix, and a preset conveyor line association feature is generated.

2. The intelligent logistics commodity scheduling and management method based on feature recognition according to claim 1, characterized in that, The method also includes: An encoder is installed at a predetermined position on the conveyor line, and the displacement distance of the package on the conveyor line is recorded in real time by the encoder pulse signal; When a package passes through the barcode recognition station, the value of the first encoder is recorded and the barcode information of the package is obtained. When a package passes through the visual recognition station, the value of the second encoder is recorded and a visual image of the package is captured. The visual recognition station is set at a preset distance behind the barcode recognition station. The displacement distance of the package between the two workstations is calculated based on the difference between the values ​​of the first encoder and the second encoder. The correspondence between the barcode information and the visual image information is determined based on the displacement distance, so as to ensure that the identification feature information and physical feature information of the same package are aligned in time and space.

3. The intelligent logistics commodity scheduling and management method based on feature recognition according to claim 1, characterized in that, Based on the initial physical characteristic information of several packages recorded in the database within a preset time period, a feature consistency assessment is performed, and a first feature recognition mechanism is established, including: The initial physical feature information of several packages within a preset time period is obtained from the database, wherein the initial physical feature information includes size features, shape features and texture features; Based on the initial physical feature information of the several packages, the size feature consistency coefficient, shape feature consistency coefficient and texture feature consistency coefficient of the multiple packages within the preset time period are calculated respectively. Based on the relationship between the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient and the corresponding preset threshold, a first feature recognition mechanism is formulated.

4. The intelligent logistics commodity scheduling and management method based on feature recognition according to claim 3, characterized in that, Based on the relationship between the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient and the corresponding preset thresholds, a first feature recognition mechanism is established, including: Preset size consistency threshold, shape consistency threshold, and texture consistency threshold; When the size consistency coefficient is less than the size consistency threshold, only the recognition mechanism of the size feature extraction unit is activated; When the size consistency coefficient is greater than or equal to the size consistency threshold and the shape consistency coefficient is less than the shape consistency threshold, the recognition mechanism of the size feature extraction unit and the shape feature extraction unit is activated. When the size consistency coefficient is greater than or equal to the size consistency threshold and the shape consistency coefficient is greater than or equal to the shape consistency threshold, the recognition mechanism of the size feature extraction unit, the shape feature extraction unit and the texture feature extraction unit is activated, wherein the size feature extraction unit, the shape feature extraction unit and the texture feature extraction unit are constructed based on a convolutional neural network.

5. The intelligent logistics commodity scheduling and management method based on feature recognition according to claim 1, characterized in that, Image recognition of packages that fail barcode scanning is performed using a second feature recognition mechanism, including: A physical feature extraction model library is established, wherein the model library includes size feature extraction models, shape feature extraction models, and texture feature extraction models built based on convolutional neural networks; Based on the physical feature information of multiple packages within the current time period, the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient are calculated, and the size feature consistency coefficient, shape feature consistency coefficient, and texture feature consistency coefficient are weighted and evaluated to obtain the comprehensive physical feature consistency coefficient. The number of physical feature extraction models to be called is determined based on the comprehensive consistency coefficient of the physical features. The higher the comprehensive consistency coefficient of the physical features, the more physical feature extraction models are called.

6. The intelligent logistics commodity scheduling and management method based on feature recognition according to claim 1, characterized in that, With the constraint that the probability of package congestion on each conveyor line does not exceed a preset threshold, and with the dual optimization objectives of minimizing the total package transportation time and minimizing the variance of the predicted congestion probability for each conveyor line, the conveyor path planning is performed based on preset conveyor line association characteristics, the package multimodal information set, and the current conveying status information of each conveyor line, including: With the constraint that the predicted congestion probability of each conveyor line does not exceed a preset threshold, conveyor lines that meet the constraint are selected as candidate paths. For all packages within a preset time zone, the packages are assigned to candidate delivery lines that meet the constraints, and multiple candidate path allocation schemes are constructed. Calculate the estimated total transportation time for all packages under each candidate scheme, wherein the estimated total transportation time is determined based on the current location of the packages, the length of the conveyor line, and the operating speed of the conveyor line; Calculate the variance of the predicted congestion probability for each transport line under each candidate scheme; A multi-objective optimization algorithm is used to select the optimal route allocation scheme from the candidate schemes that simultaneously minimizes the total parcel transportation time and the variance of the predicted congestion probability, while satisfying the constraints.

7. The intelligent logistics commodity scheduling and management method based on feature recognition according to claim 6, characterized in that, A multi-objective optimization algorithm is used to select the optimal route allocation scheme from the candidate schemes that simultaneously minimizes the total parcel transportation time and the variance of the predicted congestion probability, while satisfying the constraints. This includes: Weighting coefficients are set for the objectives of minimizing the total parcel delivery time and minimizing the variance of the predicted congestion probability. Multiple candidate path allocation schemes are generated through iterative search, and the search direction is adjusted in each iteration based on the optimization target value of the current scheme. After each iteration, evaluate whether the current scheme satisfies the constraints and eliminate any scheme that causes the predicted congestion probability of any conveyor line to exceed a preset threshold. When the preset iteration termination condition is met, the optimal path allocation scheme is selected from all candidate schemes that meet the constraints, based on the scheme with the best comprehensive optimization objective value.

8. A smart logistics commodity scheduling and management system based on feature recognition, characterized in that: The method for implementing the intelligent logistics commodity scheduling and management method based on feature recognition as described in any one of claims 1-7 includes: The feature evaluation module is used to evaluate the consistency of features based on the initial physical feature information of several packages recorded in the database within a preset time period, and to formulate a first feature recognition mechanism. The barcode recognition module is used to recognize the barcodes of packages to be sorted on the conveyor line. If the barcode recognition is successful, the first feature recognition mechanism is used to recognize the package image and obtain the package multimodal information. The image recognition module is used to perform image recognition on packages that fail barcode recognition through a second feature recognition mechanism to obtain multimodal information about the packages if barcode recognition fails. The probability prediction module is used to predict the probability of package congestion on each conveyor line within a preset time zone based on the multimodal information set of packages on the conveyor line in the current time period and the current conveying status information of each conveyor line. This includes: Obtain destination information and physical characteristic classification information of all packages within the current time period, and count the number of packages and the distribution of package types for each conveyor line; Obtain real-time conveying status information for each conveyor line, wherein the real-time conveying status information includes current load rate, equipment operating status, and historical congestion data; The number of packages, the distribution of package types, and the real-time delivery status information are input into the congestion prediction model to analyze and obtain the predicted congestion probability of each delivery line in the preset time zone. The steps for constructing the congestion prediction model include: Based on the historical operation records of each conveyor line, several sample parcel quantities, several parcel type distributions, and several real-time conveying status information distributions were collected. The proportion of congestion events that occurred on each conveyor line within the historical time zone was statistically analyzed as the sample predicted congestion probability, resulting in several sample predicted congestion probability distributions. Using the distribution of the number of sample packages, the distribution of package types, and the distribution of real-time delivery status information as inputs, and using the distribution of predicted congestion probability of the sample packages as supervision, a deep learning model is trained until convergence to generate a congestion prediction model. The path optimization module is used to plan the transport path based on the constraints that the probability of package congestion on each conveyor line does not exceed a preset threshold, and to minimize the total transport time of packages and the variance of the predicted congestion probability of each conveyor line as dual optimization objectives. It plans the transport path based on the preset conveyor line association characteristics, the package multimodal information set, and the current transport status information of each conveyor line, and obtains an optimized sorting and scheduling strategy for the logistics goods scheduling management within the preset time zone. The steps for constructing the preset conveyor line association features include: Construct the topology of the conveyor network and record the set of destinations that each conveyor line can reach; For each destination, identify all transport lines that can reach that destination and establish a set of transport line substitution relationships; Calculate the substitution cost between the alternative conveyor line and the original conveyor line, wherein the substitution cost includes the path length increment, the expected transportation time increment, and the number of transfers; The conveyor line substitution relationships and corresponding substitution costs are stored as a conveyor line association feature matrix, and a preset conveyor line association feature is generated.

Images