Intelligent warehouse optimization and dynamic scheduling method based on deep learning

Abstract

The present application relates to the technical field of warehouse management, and more particularly to an intelligent warehouse optimization and dynamic scheduling method based on deep learning. The method comprises the following steps: obtaining warehouse storage location data and cargo warehousing data, and performing time alignment to determine the current distribution of each cargo type; dividing the storage units corresponding to each cargo type based on the current distribution of each cargo type; assigning the picking paths corresponding to each storage unit; converting each picking path into a path planning instruction and transmitting it to a pre-deployed automated guided vehicle to execute the instruction; when the automated guided vehicle reaches the cargo loading position in the path planning instruction, performing dynamic loading and locking of the cargo according to the cargo type in the storage unit; and when the cargo loading capacity of the automated guided vehicle reaches the preset loading threshold, controlling the automated guided vehicle to execute the remaining path planning instruction. The present application can improve the utilization rate of warehouse space, reduce the length of cargo handling paths and operational conflicts, and improve the efficiency of cargo picking and transportation.

Description

Technical Field

[0001] This invention relates to the field of warehouse management technology, and in particular to an intelligent warehouse optimization and dynamic scheduling method based on deep learning. Background Technology

[0002] Traditional warehouse management relies on manual experience for location allocation, inventory counting, and order picking scheduling. This inherently suffers from rigid resource allocation and slow response times, making it difficult to adapt to the multi-category, high-frequency, and short-cycle order demands of e-commerce retail and smart manufacturing. To overcome the bottlenecks of traditional models, the industry has gradually introduced automated warehousing equipment (such as AGVs and robotic arms) and information management systems, achieving partial automation of warehousing operations. However, these solutions still have significant limitations: location allocation is mostly based on fixed rules, failing to fully consider product correlation, dynamic order changes, and equipment operating status, leading to redundant picking paths and low inventory turnover efficiency; scheduling strategies are mostly static pre-configured, making it difficult to respond in real-time to dynamic scenarios such as order surges, equipment failures, and inventory fluctuations, easily causing operational congestion or resource idleness; simultaneously, the collection and analysis of inventory and operational data lack deep integration capabilities, making it impossible to accurately predict inventory demand and operational load, resulting in the dual dilemma of insufficient warehouse resource utilization and high operating costs. Summary of the Invention

[0003] Therefore, the present invention needs to provide a deep learning-based intelligent warehouse optimization and dynamic scheduling method to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a deep learning-based intelligent warehouse optimization and dynamic scheduling method includes the following steps:

[0005] Step S1: Obtain warehouse location data and goods receipt data, and align the execution time of the warehouse location data and goods receipt data to determine the current distribution of each type of goods;

[0006] Step S2: Divide the storage units corresponding to each type of goods based on the current distribution of each type of goods;

[0007] Step S3: Allocate picking paths for each storage unit based on the pre-acquired order operation information for each item; convert each picking path into a path planning instruction and transmit it to the pre-deployed automated guided vehicle for execution.

[0008] Step S4: When the automated guided vehicle arrives at the loading position in the path planning instruction, it performs dynamic loading and locking of the cargo according to the cargo type in the storage unit;

[0009] Step S5: When the cargo loading capacity of the automated guided vehicle reaches the preset loading threshold, control the automated guided vehicle to execute the remaining path planning instructions.

[0010] This application achieves real-time acquisition of cargo type distribution through multi-source sensing and temporal alignment, and optimizes warehouse layout by utilizing a cargo type coding library and storage unit partitioning. Combined with an improved multi-layer spatiotemporal convolutional network, it dynamically optimizes picking paths, generating optimal path planning instructions while considering time constraints and aisle congestion. Automated guided vehicles (AGVs) execute dynamic loading and locking of cargo according to the path planning instructions, adjusting loading and transportation sequences in real time to achieve loading threshold control and phased unloading closed-loop management. This application significantly improves warehouse space utilization, reduces cargo handling path length and operational conflicts, and enhances cargo picking and transportation efficiency. Simultaneously, it ensures cargo safety and warehouse system operational stability through cargo load balancing and packaging stability, adapting to dynamic warehousing environments with multiple batches, multiple types, and high frequencies, achieving intelligent and refined warehouse management. Attached Figure Description

[0011] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0012] Figure 1 This is a flowchart illustrating the steps of the intelligent warehouse optimization and dynamic scheduling method based on deep learning of the present invention.

[0013] Figure 2 These are photos of a warehouse scene in an embodiment of the present invention;

[0014] Figure 3 This is a schematic diagram of the structure of the automated guided vehicle in an embodiment of the present invention;

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0017] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0018] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0019] To achieve the above objectives, please refer to Figures 1 to 3 This invention provides a deep learning-based intelligent warehouse optimization and dynamic scheduling method, which includes the following steps:

[0020] Step S1: Obtain warehouse location data and goods receipt data, and align the execution time of the warehouse location data and goods receipt data to determine the current distribution of each type of goods;

[0021] In this embodiment, if the warehouse scheduling is in operation and a new batch of goods is detected to be entering the warehouse, the pressure sensing units and the inbound scanning and identification terminals deployed in the warehouse storage area collect warehouse storage location data including storage location pressure change data and goods location information, and goods inbound data including goods code scan data and goods appearance scan data, respectively. Each pressure sensing unit outputs the storage location pressure value in real time with a sampling period of 0.5 seconds, and marks the corresponding time node when a pressure change is detected. The inbound scanning terminal simultaneously reads the goods barcode and goods type information, and generates an inbound record table containing the goods number, inbound time, and goods type fields. The pressure change timestamp and the inbound timestamp are aligned in time sequence, and an association matching mechanism with a time deviation of no more than 2 seconds is used to establish the correspondence between goods and storage locations, generating a goods type distribution matrix. Each row of this matrix corresponds to a single storage location number, each column corresponds to a goods type, and the matrix element value represents the real-time occupancy ratio of the corresponding type of goods in that storage location, thereby determining the current distribution of each goods type.

[0022] Step S2: Divide the storage units corresponding to each type of goods based on the current distribution of each type of goods;

[0023] In a further embodiment, if the proportion of the same type of goods among adjacent storage locations in the goods type distribution matrix is ​​greater than 0.6, then the continuous storage location set is determined as a potential storage unit for the same type of goods. The carrying capacity and occupancy rate of each storage location are further extracted to calculate the adaptability of the continuous storage location set. When the adaptability is between 70% and 90% of the rated carrying capacity of each goods type and the space utilization rate exceeds 0.75, the continuous interval is confirmed to meet the stable storage conditions and is defined as an effective storage unit. For storage location sets that do not meet the conditions, storage location merging and division adjustments will be performed based on channel connectivity to form a storage unit set that adapts to the weight and size requirements of different goods types.

[0024] Step S3: Allocate picking paths for each storage unit based on the pre-acquired order operation information for each item; convert each picking path into a path planning instruction and transmit it to the pre-deployed automated guided vehicle for execution.

[0025] In a further embodiment, the target goods number and outbound time limit are extracted from the order operation information. Then, based on the aforementioned storage unit structure table, the set of coordinates of the target goods' location is calculated, generating the initial shortest path for goods picking. This path is represented as a sequence of channel nodes, with the weights between nodes determined by the travel distance and real-time channel congestion. A pre-defined improved multi-layer spatiotemporal convolutional network structure can be used, taking the channel node sequence, historical travel delays, and the current channel state graph as input, and outputting a multi-objective weighted optimized path. After model optimization, the picking path is transformed into a path planning instruction, which includes the target node order, expected arrival time, and path priority weights, and is transmitted to a pre-deployed automated guided vehicle for control and execution.

[0026] It is worth noting that the path optimization model adopts an improved multi-layer spatiotemporal convolutional network structure, which includes an input layer, a dual-channel spatiotemporal encoding layer, a constraint mapping layer, and a path decision layer. The input layer receives composite input data consisting of a picking order time series and a channel congestion spatial matrix. The time series is represented by a vector of the difference in target outbound time limits for each storage unit, and the spatial matrix is ​​a two-dimensional adjacency graph composed of channel node connectivity weights and real-time congestion. The dual-channel spatiotemporal encoding layer encodes temporal and spatial constraint features using temporal convolutional units and graph convolutional units respectively, and dynamically fuses the two types of features through gated residual connections between layers. The constraint mapping layer dynamically adjusts the temporal constraint weights and spatial connectivity weights based on a self-attention mechanism to form a path constraint tensor. The path decision layer takes the path constraint tensor as input and outputs the passage priority coefficient of each node and the minimized path cost function result through Softmax normalization, thereby determining the optimal access sequence for each node in the path optimization result.

[0027] The model construction process includes four stages: path sample extraction, feature encoding standardization, constraint feature training, and performance verification. The path sample extraction stage extracts path samples containing time and space constraints from historical picking task logs. The feature encoding standardization stage normalizes the outbound time difference vector to the [0,1] interval and transforms the channel congestion matrix into a sparse adjacency matrix. The constraint feature training stage uses a weighted loss function, taking the average path travel time and congestion cost as dual-objective constraints, and performs multi-round parameter iteration training using the Adam optimizer. The performance verification stage calculates the deviation rate between the predicted path and the actual executed path based on real-time path feedback data; when the deviation rate is less than 5%, the model is considered to have reached stable convergence.

[0028] It is worth noting that converting the picking path into path planning instructions involves three steps: path encoding, timing constraint embedding, and execution instruction generation. First, the channel nodes are encoded according to the picking order to form a path node timing table. Second, the expected arrival time of each node is written into the path node timing table to form a path planning cache table. Finally, based on the instruction template of the automated guided vehicle (AGV), the path planning cache table is converted into a path planning instruction file. The instruction file adopts a standardized task instruction format, including path number, node sequence, target arrival time, steering parameters, and speed control threshold fields. This path planning instruction is transmitted to the AGV control unit via a wireless communication module, where it is parsed and executed to achieve automated execution of the goods picking path.

[0029] Step S4: When the automated guided vehicle arrives at the loading position in the path planning instruction, it performs dynamic loading and locking of the cargo according to the cargo type in the storage unit;

[0030] In a further embodiment, if the guided vehicle reaches the loading location specified by the path planning instruction, the vehicle control unit calls the cargo location recognition camera to read the cargo location number and queries the corresponding storage unit table to confirm the cargo type. When the cargo type matches the current picking task type, the guided vehicle starts the robotic arm to perform cargo loading. The loading operation controls the gripping force and lifting speed according to the cargo's dimensions and weight parameters, setting the maximum gripping pressure to 150N and the minimum lifting speed to 0.25m / s. After loading is completed, the guided vehicle performs a locking and fixing operation based on packaging stability conditions, including a cargo center offset of no more than 5mm, a tilt angle of less than 3°, and a packaging strap tension between 180 and 220N. The micro-motion response of the cargo is detected by a dual-axis accelerometer. If the stability coefficient K value is greater than 0.92, the locking is considered successful, and the corresponding state is recorded.

[0031] Step S5: When the cargo loading capacity of the automated guided vehicle reaches the preset loading threshold, control the automated guided vehicle to execute the remaining path planning instructions.

[0032] In a further embodiment, after the automated guided vehicle (AGV) completes several loading operations consecutively, the total loaded weight and the number of picked goods are calculated in real time. If the total loaded weight reaches 85% of the vehicle's rated load capacity, the loading threshold control logic is triggered. The control immediately pauses the current picking task, calls the remaining path planning cache table in the path planning instruction (excluding the path from which the AGV reaches the loading position in the path planning instruction), and performs path segment switching, instructing the vehicle to proceed along the shortest safe path to the outbound area to complete the phased unloading. After unloading, the vehicle continues to execute the unfinished picking path task according to the remaining path planning instruction, realizing closed-loop control of loading and transportation tasks.

[0033] Optionally, aligning the warehouse location data with the goods code entry data in step S1 includes:

[0034] Compare the timestamps of pressure change in each storage location in the warehouse location data with the timestamps of the inbound scan of each cargo code in the cargo inbound data, and generate the correspondence between storage locations and storage location codes based on the timestamp comparison results;

[0035] In this embodiment, if the pressure value of the storage location pressure sensing unit in the warehouse monitoring system increases by more than a set sudden change threshold (e.g., 20 kPa) within a certain sampling period compared to the previous period, it is determined that an inbound event has occurred at that storage location. At this time, the timestamp corresponding to the pressure sudden change is recorded. Meanwhile, the barcode scanning terminal of the inbound management system records the barcode scanning timestamp when goods enter the warehouse. and corresponding cargo code information. Timestamp and Within the allowable error range Matching is performed within ±2 seconds (e.g., if...) Then, a one-to-one correspondence is established between the storage location number and the cargo code, generating a data table corresponding to the storage location and the storage location code.

[0036] Based on the preset cargo type coding library, the correspondence between cargo location and cargo location code is converted into a mapping relationship between cargo location and cargo type, and the current distribution of each cargo type is statistically analyzed based on this mapping relationship.

[0037] In a further embodiment, the cargo code field in the cargo type code library is searched and matched with the cargo code in the corresponding data table, thereby retrieving the corresponding cargo type in the type feature matrix. Upon successful matching, a mapping table between storage locations and cargo types is generated. Subsequently, based on the distribution records of the same cargo type in different storage locations in this mapping table, the current distribution of each cargo type within the storage area is statistically analyzed. This distribution can be output in the form of a distribution matrix. The row index of the distribution matrix represents the cargo type number, the column index represents the storage location number, and the value of each matrix element represents the proportion of that type of cargo in that storage location.

[0038] Optionally, methods for obtaining the cargo type code library include:

[0039] Link the barcode scan data and appearance inspection data under the same timestamp in the goods warehousing data to form the basic attribute data of the goods;

[0040] In this embodiment, goods are detected upon entry into the warehouse via an entry scanning and identification terminal. The unique barcode information for each item is recorded at a 20ms sampling interval, and corresponding shape detection data is simultaneously acquired by a binocular camera device. The shape detection data includes length, width, height, surface color feature matrix, and edge contour curves. Through a timestamp matching mechanism, barcode scanning data and shape detection data from the same sampling moment are associated one-to-one to generate basic attribute data for the goods.

[0041] The basic attribute data of goods is compared with the barcode information of goods type and appearance description information in the order operation information of each pre-acquired goods, and the similarity of goods attributes is calculated based on the comparison results.

[0042] In a further embodiment, the basic cargo attribute data is compared with the cargo movement information already registered in the order movement information. The order movement information is obtained through the warehouse management platform and consists of a cargo type barcode information table, a shape description information table, and a target storage information table. During the comparison, the similarity between the barcode scan data in the basic cargo attribute data and the cargo type barcode information is calculated using cosine distance, and the similarity between the shape detection data in the basic cargo attribute data and the shape description information is also calculated using cosine distance. The two similarities are then combined into a cargo attribute similarity score.

[0043] If the similarity of the goods attributes is less than the preset similarity threshold, the corresponding barcode scan data will be removed, and the goods type barcode information corresponding to the shape description information with the highest similarity to the shape detection data will be used as the replacement barcode.

[0044] In one embodiment, if both similarity scores in the cargo attribute similarity are determined to be below a similarity threshold (e.g., 80%), then while removing barcode scan data, the record most similar to the detected shape data is retrieved from the shape description information table, the shape description information with the highest similarity is returned, and the cargo type barcode information corresponding to this shape description information is identified as the replacement barcode. This replacement barcode is stored in the replacement record table, along with the similarity parameter value and the retrieval timestamp.

[0045] If the similarity of goods attributes is greater than or equal to the similarity threshold, the goods type barcode information corresponding to the barcode scan data is used as the barcode to be written, and the replacement barcode and the barcode to be written are classified together to obtain the goods type coding library.

[0046] In another embodiment, if both similarity scores in the cargo attribute similarity are determined to be greater than or equal to a similarity threshold (e.g., 80%), the cargo type barcode information corresponding to the barcode scan data is identified as the barcode to be written, and a classification operation is performed at the database level with the aforementioned replacement barcode. Classification is accomplished through a barcode aggregation function based on key-value mapping, merging and encoding replacement barcodes and the barcode to be written under the same cargo type to form a cargo type encoding library.

[0047] Optionally, the storage units corresponding to each type of goods in step S2 include:

[0048] Based on the current distribution of each type of goods, the continuous distribution intervals of goods types between adjacent storage locations are divided. If the length of the continuous distribution interval is greater than or equal to the preset minimum concentration threshold, the continuous distribution interval is determined as a potential storage unit candidate area; if the length of the continuous distribution interval is less than the minimum concentration threshold, the continuous distribution interval is divided into boundaries to obtain independent distribution units.

[0049] In this embodiment, the Euclidean spatial distance between adjacent storage locations is calculated based on the current distribution of each cargo type, and consecutive storage locations with the same cargo type are defined as continuous distribution intervals. Each continuous distribution interval consists of a starting storage location number, an ending storage location number, and a type coding structure, serving as the input basis for potential storage unit candidate areas. When the Euclidean spatial distance of a continuous distribution interval is greater than or equal to the minimum concentration threshold (3 storage locations), these continuous distribution intervals are determined as potential storage unit candidate areas.

[0050] In another embodiment, if the length of a continuous distribution interval is less than the minimum concentration threshold of 3 storage locations, a boundary segmentation procedure is triggered. Segmentation is completed by detecting changes in the type of the interval endpoints, using the midpoint between adjacent dissimilar storage locations as boundary nodes to generate several independent distribution units. Each independent distribution unit is represented by a storage location number sequence, and the cargo type labels before and after segmentation, as well as the segmentation time, are recorded.

[0051] Evaluate the suitability of each potential storage unit candidate region to identify suitable and unsuitable storage units;

[0052] In a further embodiment, cargo load balancing parameters are calculated based on the load pressure of each potential storage unit candidate area and the historical occupancy rate of the storage location. Suitable and unsuitable storage units are then classified based on the comparison between the preset load requirements of each cargo type in the order operation information and the cargo load balancing parameters.

[0053] The execution unit is merged based on the channel connectivity between the independent distributed unit and the unsuitable storage unit, thereby obtaining the reorganized storage unit;

[0054] In a further embodiment, connectivity analysis is performed on the independent distributed units and the mismatched storage units. The channel connection path is calculated using a breadth-first search. If the length of the channel connection path is less than the set maximum merging step size of 5 storage locations and the corresponding cargo types are consistent, the independent distributed units and the mismatched units are merged to generate a recombined storage unit.

[0055] The adaptation storage unit and the reorganization storage unit are bound to the corresponding cargo type and numbered to obtain the storage unit corresponding to each cargo type.

[0056] In a further embodiment, the storage units are bound and numbered according to the coding rules in the cargo type coding library. The numbering format consists of a cargo type prefix and a storage unit serial number, such as "T01-SU03" representing the 3rd storage unit of type T01. After binding, a cargo type-to-storage unit mapping table is generated and written to the database index.

[0057] It is worth noting that the minimum concentration threshold is set at 3 storage locations, which is based on a comprehensive consideration of the stability and operational efficiency of centralized storage of goods in the warehousing system. If there are fewer than 3 storage locations for consecutive similar goods, frequent repositioning and increased handling paths are likely to occur during picking operations, which is not conducive to forming a stable storage and retrieval area. By setting the threshold to 3, it is ensured that the formed continuous distribution area has basic operational concentration in space, while avoiding the misjudgment of a single storage location type as a local concentration area. In addition, the maximum merging step size is set at 5 storage locations to balance aisle connectivity and operational safety distance when merging independent distribution units and incompatible units. If the merging span is too large, it may cross high-frequency operation aisles, leading to path interference and storage location congestion; while if the step size is too small, it will cause fragmentation of the merging and an excessive number of units. Experiments have determined that a step size of 5 storage locations can ensure the minimum turning radius and avoidance space of picking vehicles in the aisle, while maintaining the physical continuity between storage locations, thereby making the layout of reorganized storage units more reasonable and the structure more stable.

[0058] Optionally, evaluating the suitability of each potential memory cell candidate region includes:

[0059] Acquire load pressure sensor data and historical occupancy data of each storage location in each potential storage unit candidate area to calculate cargo load balance parameters.

[0060] In this embodiment, the load-bearing pressure sensor data and historical occupancy data of each storage location are obtained through the warehouse management platform, and these data are then linked to potential storage unit candidate areas. Based on the load-bearing pressure sensor data, historical occupancy data, and preset weighting coefficients, a smoothing weighted method is used to calculate the cargo load balance parameter. The weighting coefficients are set to a pressure average weight of 0.6 and an occupancy weight of 0.4, respectively, to reflect the overall load-bearing stability of the storage location.

[0061] Based on the preset load requirements of each type of goods in the order operation information, and compared with the load balancing parameters of the corresponding storage location, if the load balancing parameters meet the load requirements of the corresponding type of goods, then the potential storage unit candidate area is determined to be a suitable storage unit; otherwise, the potential storage unit candidate area is determined to be an unsuitable storage unit.

[0062] In a further embodiment, preset load-bearing requirement parameters for each type of goods, including the maximum static load limit, are extracted from the order operation information. Permissible dynamic load fluctuation range and the lower limit of recommended market share Subsequently, the maximum static load limit, allowable dynamic load fluctuation range, and recommended occupancy lower limit in the preset load requirements parameters are mapped to three normalized sub-scores (static load adaptation sub-score, dynamic load adaptation sub-score, and occupancy adaptation sub-score). These are then weighted and summed to obtain the target adaptation score, with weighting coefficients of 0.5 for the static load adaptation sub-score, 0.3 for the dynamic load adaptation sub-score, and 0.2 for the occupancy adaptation sub-score. The load balancing parameters of each storage location within the potential storage unit candidate area are compared with the target adaptation score for the corresponding cargo type. If the load balancing parameters of all storage locations are greater than the target adaptation score, the candidate area is confirmed as a suitable storage unit; if any storage location is less than the target adaptation score, it is determined to be an unsuitable storage unit.

[0063] It is worth noting that mapping to normalized sub-scores also requires obtaining location observations (average load factor of locations) through the warehouse management platform. Bearing standard deviation Actual market share Static load adapter score ,like but Dynamic load adapter rating ,like but Market share fit sub-score ; of which actual occupancy rate It refers to the proportion of a storage location or unit that is actually occupied by goods, that is, the area or weight of the goods occupied divided by the maximum usable area or load-bearing capacity of the storage location.

[0064] Optionally, the execution unit merging includes:

[0065] If the load balancing parameters of the unmatched storage unit meet the carrying requirements of any cargo type, then establish the association between the unmatched storage unit and that cargo type.

[0066] In this embodiment, the load balancing parameter calculated for each unsuitable storage unit is compared with the target adaptation score of each type of goods in the order operation information. If the load balancing parameter is greater than the target adaptation score of any type of goods, the type of goods is recorded as a potential adaptation type of the storage unit and an association table entry is established.

[0067] Based on the association between the mismatched storage unit and the cargo type, units of the same cargo type are selected from the independent distributed units as candidate merging units;

[0068] In a further embodiment, all independent distributed units are traversed, their corresponding cargo types are matched, and units whose cargo types are consistent with the associated types of incompatible storage units are selected as candidate merging units.

[0069] Determine the channel connectivity between mismatched storage units and candidate merging units, and merge mismatched storage units with the same cargo type and connected channels into a recombined storage unit.

[0070] In a further embodiment, the path between channel nodes in the storage channel topology table is calculated. If the node path is continuous and not blocked, the path is considered connected. The automated guided vehicle (AGV) replaces the cargo type in the incompatible storage unit with the cargo type that has established a relationship with the incompatible storage unit, and merges it with the candidate merging unit into a reorganized storage unit. The storage channel topology table is generated from the storage planning and design drawings. It is a data table that represents the spatial relationship between storage locations and channels in a graph structure. The nodes record the storage location or intersecting channel number and coordinates, and the edges record the channel connection relationship between nodes, the passage distance, width, capacity, and real-time congestion coefficient, and also include attributes such as directionality and historical delay.

[0071] Optionally, after establishing the association between the unsuitable storage unit and the type of goods, the process also includes:

[0072] Determine the connectivity between incompatible storage units. If the connectivity between incompatible storage units is such that the packaging of the corresponding goods types meets the preset packaging stability conditions, then use an automated guided vehicle to replace the goods types between incompatible storage units with goods types that are associated with the incompatible storage units.

[0073] In this embodiment, the warehouse channel topology table is read, and the connectivity between nodes is determined by traversing the node numbers and edge information between incompatible storage units. If the shortest connecting path exists and its width is greater than 0.8 meters, and the channel capacity is greater than 1.2 times the volume of a single automated guided vehicle (AGV), then the channel is considered connected. Subsequently, the shape detection data of the incompatible storage unit is obtained, and the packaging flatness, anti-tilt angle, and shape integrity are calculated. Preset thresholds (flatness ≥ 0.9, anti-tilt angle ≥ 15°, integrity ≥ 0.95) are used to determine whether the packaging stability meets the requirements. If all requirements are met, the goods type replacement operation is confirmed to be feasible. During the replacement operation, the AGV receives control commands, moves along the connecting channel to the target storage unit, accurately docks with the shipping interface using the positioning and identification unit, grabs and unloads the goods, and replaces the goods with the goods type associated with the incompatible storage unit. During the operation, the loading speed is controlled at 0.2–0.5 m / s to ensure that the handling process is not overloaded and the packaging is stable.

[0074] Of particular importance are the methods for obtaining packaging stability conditions, including:

[0075] The flatness, tilt resistance, and shape integrity of the packaging for each type of goods are determined based on the shape inspection data corresponding to each type of goods.

[0076] The flatness, tilt resistance, and shape integrity of the packaging for each type of goods are compared with the corresponding stability thresholds, and the shape detection data that all meet the stability thresholds are taken as the packaging stability conditions for that type of goods.

[0077] In this embodiment, the packaging flatness, anti-tilt angle and shape integrity are calculated using the shape detection data obtained by the warehouse scanning and recognition terminal. The preset stability thresholds (flatness ≥ 0.9, anti-tilt angle ≥ 15°, integrity ≥ 0.95) are used to determine whether the packaging stability meets the requirements. If all are met, the packaging shape feature parameters in the corresponding shape detection data are used as the packaging stability conditions for this type of goods.

[0078] Optionally, the picking path allocated to each storage unit in step S3 includes:

[0079] Determine the target outbound time limit corresponding to the type of goods based on the order operation information of each goods;

[0080] In this embodiment, the target outbound time limit for each item is extracted from the target storage information table in the order operation information. For example, the target outbound time limit for fragile items is set to within 30 minutes, and for standard items it is within 60 minutes.

[0081] Determine the channel connectivity of the storage unit, and generate the shortest picking path for each type of goods based on the spatial distribution information of the storage unit and the channel connectivity.

[0082] In a further embodiment, based on the coordinate information in the warehouse channel topology table and the spatial distribution information of the storage units, the set of shortest paths between storage units is calculated. Each path is represented by a channel node sequence, and the weight between nodes is determined by the travel distance and the real-time channel congestion coefficient. For example, the travel distance is assigned a value of 1 per meter, and the congestion coefficient is weighted from 0 to 1 to the total weight.

[0083] Path optimization is performed on the shortest picking path for each type of goods, and the path optimization results are matched with the corresponding goods type to obtain the picking path for each storage unit.

[0084] In a further embodiment, the initial path is input into the path optimization model (an improved multi-layer spatiotemporal convolutional network), which consists of three spatiotemporal convolutional layers and two graph attention layers. The input includes the channel node sequence, historical travel delay, and real-time channel state graph. The output is the optimal path node sequence and the estimated arrival time. The optimization result is mapped to generate a path planning instruction.

[0085] Optionally, execution path optimization includes:

[0086] Determine the channel congestion between storage units based on the channel connectivity of the storage units;

[0087] In this embodiment, the congestion level between storage units is determined based on the warehouse aisle topology table and storage unit coordinate information. The congestion level ranges from 0 to 1, where 0 indicates unobstructed access and 1 indicates complete blockage. For example, for an area with a main aisle width of 2 meters, if the number of automated guided vehicles reaches 4 in real time, the congestion level is set to 0.8.

[0088] The picking order of each storage unit is assigned based on the target outbound time limit corresponding to the type of goods and the passage direction between storage units;

[0089] In this embodiment, picking order is assigned to each storage unit based on the target outbound time limit for each type of goods. For example, the target outbound time limit for fragile goods is 30 minutes, and for standard goods it is 60 minutes. Storage units corresponding to goods with lower target outbound time limits are prioritized for picking. If the target outbound time limits are the same, the picking order is assigned according to the distance between the corresponding storage unit and the sorting starting point, from closest to furthest. The picking order is stored in the form of a list of storage unit numbers.

[0090] The picking order of each storage unit is used as a time constraint, the congestion of the passage between storage units is used as a spatial constraint, and a preset path optimization model is used to apply path constraints to the shortest picking path for each type of goods to obtain the path optimization results.

[0091] In a further embodiment, the picking order of each storage unit is used as a time constraint, and the channel congestion is used as a spatial constraint. An improved multi-layer spatiotemporal convolutional network is used to perform path constraint optimization. The network input includes an initial shortest path node sequence, a channel congestion matrix, and a time constraint vector. The output is an optimal path node sequence and the estimated arrival time of each node. After mapping, the final path optimization result is generated and updated to the path planning instruction for execution by the automated guided vehicle.

[0092] Optionally, performing dynamic cargo loading in step S4 includes:

[0093] The positioning and identification unit of the automated guided vehicle is invoked to detect the storage unit number corresponding to the cargo loading location, and the type of cargo to be loaded in the storage unit is read according to the storage unit number;

[0094] If the type of goods to be loaded is found to match the target type of goods in the real-time updated order operation information, the loading device of the automated guided vehicle and the shipping interface of the storage unit are controlled to complete the docking within the same time window to execute the loading of goods; the target type of goods is the type of goods corresponding to the storage unit with the highest priority in the picking order.

[0095] In this embodiment, the positioning and identification unit of the automated guided vehicle (AGV) employs a high-precision lidar and visual sensor fusion positioning system. It acquires three-dimensional coordinates by scanning the storage racks and matches them with a pre-stored storage unit number and location mapping table to determine the storage unit number corresponding to the loading location in real time. Subsequently, based on the acquired storage unit number, it reads the type of goods to be loaded from the storage unit. This information is stored in the goods inventory table of the warehouse management system, with fields including goods type code, quantity, and batch number. To ensure synchronization, the inventory table data is refreshed every 5 seconds. In the comparison step, the type of goods to be loaded is matched item by item with the target goods type in the real-time updated order operation information. When a match is detected, the AGV controls the loading robotic arm to dock with the storage unit's shipping interface within the same time window, which can be set to 3 seconds. Here, the target goods type corresponds to the storage unit with the highest priority in the picking order.

[0096] Of particular importance, the implementation of dynamic cargo loading also includes:

[0097] If the type of goods to be loaded is found to be inconsistent with the target type of goods in the real-time updated order operation information, the automated guided vehicle is controlled to move along the storage unit with the second highest priority in the picking order according to the path planning instructions in order to perform goods loading.

[0098] In this embodiment, if the comparison results show inconsistencies, the path adjustment logic will be triggered. Based on the pre-generated path planning instructions and the storage unit picking order, the automated guided vehicle (AGV) determines the second-priority storage unit number and obtains its three-dimensional coordinates and channel connectivity information. Then, the AGV moves along the channel specified by the path planning instructions to the second-priority storage unit. Upon arrival at the storage unit, the AGV reads the type of goods in that unit and performs a loading operation. The time window is set to 3 seconds, and the loading task is completed synchronously with the loading robotic arm.

[0099] Figure 3 This is a schematic diagram of the structure of the automated guided vehicle in an embodiment of the present invention; as shown below. Figure 3 As shown, the automated guided vehicle includes a controller and a loading device disposed in the inner shell of the vehicle.

[0100] The positioning and identification unit, mounted on the top of the automated guided vehicle (AGV), may include a lidar sensor, an RGB-D camera, and an inertial measurement unit (IMU). It can perceive the vehicle's absolute position and orientation in real time, while simultaneously identifying obstacles and cargo location numbers in the surrounding environment. The positioning and identification unit is used to acquire vehicle location, store unit location, and path information in real time, enabling high-precision navigation, dynamic obstacle avoidance, and cargo location identification. It then transmits the data to the controller for path adjustment and loading command generation.

[0101] The controller, an embedded computing unit, includes a CPU, GPU, and storage module. It is capable of running path planning, scheduling strategies, and real-time data interaction programs with sensors. Functionally, the controller is responsible for receiving data from the positioning and identification unit, calculating the optimal driving path, controlling the vehicle's motion state (speed, steering), and coordinating the synchronous operation of the loading device and the cargo interface.

[0102] The loading device, installed at the front or side of the automated guided vehicle (AGV), includes a robotic arm, a multi-degree-of-freedom gripping mechanism, and inductive cargo-grabbing sensors. It is extendable to accommodate different cargo heights. Functionally, this loading device is used to grasp, transport, place, and lock cargo after the vehicle arrives at the designated storage unit. Simultaneously, it can execute synchronized actions according to controller commands to ensure safe loading and rapid unloading of cargo.

[0103] The automated guided vehicle can be equipped with different accessories according to the actual needs of the scenario, and this application does not impose any restrictions on this.

[0104] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0105] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A deep learning-based intelligent warehouse optimization and dynamic scheduling method, characterized in that, Includes the following steps: Step S1: Obtain warehouse location data and goods receipt data, and align the execution time of the warehouse location data and goods receipt data to determine the current distribution of each type of goods; Step S2: Divide the storage units corresponding to each type of goods based on the current distribution of each type of goods; Step S3: Allocate picking paths for each storage unit based on the pre-acquired order operation information for each item; convert each picking path into a path planning instruction and transmit it to the pre-deployed automated guided vehicle for execution. Step S4: When the automated guided vehicle arrives at the loading position in the path planning instruction, it performs dynamic loading and locking of the cargo according to the cargo type in the storage unit; Step S5: When the cargo loading capacity of the automated guided vehicle reaches the preset loading threshold, control the automated guided vehicle to execute the remaining path planning instructions; Step S1, which aligns the warehouse location data with the goods code entry data in terms of execution sequence, includes: Compare the timestamps of pressure change in each storage location in the warehouse location data with the timestamps of the inbound scan of each cargo code in the cargo inbound data, and generate the correspondence between storage locations and storage location codes based on the timestamp comparison results; Based on the preset cargo type coding library, the correspondence between cargo location and cargo location code is converted into a mapping relationship between cargo location and cargo type, and the current distribution of each cargo type is statistically analyzed based on this mapping relationship; Methods for obtaining the cargo type code library include: Link the barcode scan data and appearance inspection data under the same timestamp in the goods warehousing data to form the basic attribute data of the goods; The basic attribute data of goods is compared with the barcode information of goods type and appearance description information in the order operation information of each pre-acquired goods, and the similarity of goods attributes is calculated based on the comparison results. If the similarity of the goods attributes is less than the preset similarity threshold, the corresponding barcode scan data will be removed, and the goods type barcode information corresponding to the shape description information with the highest similarity to the shape detection data will be used as the replacement barcode. If the similarity of goods attributes is greater than or equal to the similarity threshold, the goods type barcode information corresponding to the barcode scan data will be used as the barcode to be written, and the replacement barcode and the barcode to be written will be classified together to obtain the goods type coding library. Step S2 involves dividing the storage units corresponding to each type of goods, including: Based on the current distribution of each type of goods, the continuous distribution intervals of goods types between adjacent storage locations are divided. If the length of the continuous distribution interval is greater than or equal to the preset minimum concentration threshold, the continuous distribution interval is determined as a potential storage unit candidate area; if the length of the continuous distribution interval is less than the minimum concentration threshold, the continuous distribution interval is divided into boundaries to obtain independent distribution units. Evaluate the suitability of each potential storage unit candidate region to identify suitable and unsuitable storage units; The execution unit is merged based on the channel connectivity between the independent distributed unit and the unsuitable storage unit, thereby obtaining the reorganized storage unit; The adaptable storage unit and the reorganized storage unit are bound to the corresponding cargo type and numbered to obtain the storage unit corresponding to each cargo type; The assessment of the suitability of each potential memory cell candidate region includes: Acquire load pressure sensor data and historical occupancy data of each storage location in each potential storage unit candidate area to calculate cargo load balance parameters. Based on the preset load requirements of each type of goods in the order operation information, and compared with the load balancing parameters of the corresponding storage location, if the load balancing parameters meet the load requirements of the corresponding type of goods, then the potential storage unit candidate area is determined to be a suitable storage unit; otherwise, the potential storage unit candidate area is determined to be an unsuitable storage unit.

2. The intelligent warehouse optimization and dynamic scheduling method based on deep learning according to claim 1, characterized in that, Execution unit merging includes: If the load balancing parameters of the unmatched storage unit meet the carrying requirements of any cargo type, then establish the association between the unmatched storage unit and that cargo type. Based on the association between the mismatched storage unit and the cargo type, units of the same cargo type are selected from the independent distributed units as candidate merging units; Determine the channel connectivity between mismatched storage units and candidate merging units, and merge mismatched storage units with the same cargo type and connected channels into a recombined storage unit.

3. The intelligent warehouse optimization and dynamic scheduling method based on deep learning according to claim 2, characterized in that, After establishing the association between the incompatible storage unit and the type of goods, the following steps are also included: Determine the connectivity between incompatible storage units. If the connectivity between incompatible storage units is such that the packaging of the corresponding goods types meets the preset packaging stability conditions, then use an automated guided vehicle to replace the goods types between incompatible storage units with goods types that are associated with the incompatible storage units.

4. The intelligent warehouse optimization and dynamic scheduling method based on deep learning according to claim 1, characterized in that, Step S3, which assigns the picking path to each storage unit, includes: Determine the target outbound time limit corresponding to the type of goods based on the order operation information of each goods; Determine the channel connectivity of the storage unit, and generate the shortest picking path for each type of goods based on the spatial distribution information of the storage unit and the channel connectivity. Path optimization is performed on the shortest picking path for each type of goods, and the path optimization results are matched with the corresponding goods type to obtain the picking path for each storage unit.

5. The intelligent warehouse optimization and dynamic scheduling method based on deep learning according to claim 4, characterized in that, Execution path optimization includes: Determine the channel congestion between storage units based on the channel connectivity of the storage units; The picking order of each storage unit is assigned based on the target outbound time limit corresponding to the type of goods and the passage direction between storage units; The picking order of each storage unit is used as a time constraint, the congestion of the passage between storage units is used as a spatial constraint, and a preset path optimization model is used to apply path constraints to the shortest picking path for each type of goods to obtain the path optimization results.

6. The intelligent warehouse optimization and dynamic scheduling method based on deep learning according to claim 1, characterized in that, Step S4, which involves dynamic loading of cargo, includes: The positioning and identification unit of the automated guided vehicle is invoked to detect the storage unit number corresponding to the cargo loading location, and the type of cargo to be loaded in the storage unit is read according to the storage unit number; If the type of goods to be loaded is found to match the target type of goods in the real-time updated order operation information, the loading device of the automated guided vehicle and the shipping interface of the storage unit are controlled to complete the docking within the same time window to execute the loading of goods; the target type of goods is the type of goods corresponding to the storage unit with the highest priority in the picking order.

Images