A warehouse management method and system supporting intelligent goods location allocation

Abstract

The application relates to the technical field of warehouse management and intelligent logistics, and discloses a warehouse management method and system supporting intelligent goods location allocation. The method divides SKUs into a high-frequency active layer, a medium-frequency stable layer and a long-tail silent layer based on time attenuation thermal values, only performs incremental thermal value updating and incremental correlation strength matrix updating on the high-frequency active layer and the SKU with layer transition, and performs incremental spectral clustering based on the last round of clustering results; a channel segment space-time conflict thermal map is generated through multi-robot path simulation, the medium and low frequency SKUs are temporarily promoted in layers based on conflict contribution score, a collaborative goods location migration scheme is generated under the constraints of migration budget and layout continuity to minimize the weighted sum of global conflict frequency and regional peak load rate variance, and finally a controlled intelligent goods location allocation instruction list is output.

Description

Technical Field

[0001] This invention relates to the field of warehouse management and intelligent logistics technology, and more specifically, to a warehouse management method and system that supports intelligent location allocation. Background Technology

[0002] In ultra-large-scale automated warehouses, the number of SKUs reaches millions, and hundreds of autonomous mobile robots perform picking tasks in parallel. Existing intelligent location allocation methods typically employ two strategies: one is to cluster all SKUs based on their outbound heatmap values ​​and map them to locations according to heatmap ranking to shorten the picking path for high-frequency SKUs; the other is to perform load balancing based on order wave prediction and regional throughput capacity, dispersing pressure on congested areas by splitting locations for popular SKUs.

[0003] However, the above methods face the following technical problems at the scale of millions of SKUs: First, the dimension of the full association strength matrix is ​​on the order of millions × millions, and the computational complexity and memory consumption of spectral clustering feature decomposition are unacceptable. Moreover, full recalculation leads to large-scale inventory relocation for each update, and the relocation cost far exceeds the benefits. Second, existing load balancing methods do not consider the spatiotemporal characteristics of path conflicts in robot scenarios, and simply distributing regional loads cannot eliminate path intersection conflicts. Third, hierarchical incremental calculation and path conflict prevention are executed independently. Low- and medium-frequency SKUs that contribute significantly to conflicts but whose heat values ​​do not change much are not included in the optimization scope for a long time. Furthermore, the location migration after incremental update and the conflict resolution migration interfere with each other, resulting in the loss of the location layout optimization effect. Summary of the Invention

[0004] This invention provides a warehouse management method and system that supports intelligent location allocation, solving the technical problems in related technologies such as lack of dynamic adaptive capability in location allocation in ultra-large-scale warehousing scenarios, excessive overhead of full recalculation, and inability to effectively cope with channel congestion and high-frequency SKU aggregation needs.

[0005] This invention discloses a warehouse management method supporting intelligent location allocation, comprising the following steps: acquiring the time-decay heatmap values ​​and warehouse aisle network topology data of all SKUs; dividing SKUs into three levels—high-frequency active layer, mid-frequency stable layer, and long-tail silent layer—based on the distribution of time-decay heatmap values, and generating an SKU level label mapping table and a aisle network diagram; updating the heatmap values ​​of actively changing SKUs based on newly added outbound order data; comparing the updated heatmap values ​​with the level interval boundaries to identify level transition SKUs and generating a level transition SKU list; and processing high-frequency active layer SKUs based on... New order data is updated with an incremental association strength matrix. Incremental spectral clustering is performed based on the previous round of clustering results to generate an incrementally updated set of high-frequency active layer association clusters. Based on the set of high-frequency active layer association clusters, storage location areas are mapped for each level of SKU. The mapping results are compared with the current actual storage locations to generate a controlled intelligent storage location allocation instruction list. The time decay heat value is obtained by weighted summation of each SKU outbound event using an exponential time decay function. Outbound events closer to the current time contribute more, while outbound events further away contribute exponentially.

[0006] Furthermore, the time decay heat value is calculated as follows: for any SKU, the time interval between each outbound event and the current time within the statistical period is substituted into an exponential function with the natural constant as the base and the negative value of the product of the time decay coefficient and the time interval as the exponent. The exponential function values ​​corresponding to each outbound event are accumulated to obtain the time decay heat value of the SKU. The time decay coefficient is a positive real number used to control the rate at which the heat contribution of historical outbound events diminishes over time.

[0007] Further, the incremental heat value update for actively changing SKUs includes: extracting the relevant SKU identifiers from the new outbound order data, summarizing and deduplicating them to form an active changing SKU set; for each SKU in the active changing SKU set, multiplying its stock time decay heat value by an exponential function value with the natural constant as the base and the negative of the product of the time decay coefficient and the time interval from the last calculation time to the current time as the exponent, to obtain a stock decay conversion value; substituting the time intervals from each outbound event of the SKU in the new outbound order to the current time into the exponential time decay function and summing them to obtain an incremental decay weight value; adding the stock decay conversion value and the incremental decay weight value to obtain the updated time decay heat value; for SKUs not in the active changing SKU set, only their stock time decay heat value is converted to stock decay.

[0008] Further, the incremental association strength matrix update and incremental spectral clustering include: calculating the incremental weighted co-occurrence frequency only for high-frequency active layer SKU pairs that appear simultaneously in the same order in new orders; for any two high-frequency active layer SKUs, decaying their existing association strength values ​​according to the elapsed time interval, and then accumulating the exponential decay weight value of the new co-occurrence event to obtain the updated association strength value; using the previous round of clustering results as the initial cluster assignment; for SKUs that newly enter the high-frequency active layer due to hierarchical transition, calculating the cosine similarity between their association strength row vector and each existing cluster center vector, and assigning them to the cluster with the highest cosine similarity; for SKUs that leave the high-frequency active layer due to hierarchical transition, removing them from their cluster and updating the cluster center vector of that cluster; wherein, the cluster center vector is the mean vector of the corresponding row vectors of each SKU in the association strength matrix within the cluster.

[0009] Furthermore, the mapping of the storage location area and the generation of the controlled intelligent storage location allocation instruction list include: sorting each associated cluster in the incrementally updated high-frequency active layer associated cluster set according to the total cluster heat value from high to low; mapping the top-ranked associated clusters to the priority storage location block closest to the outbound port, where the total cluster heat value is the sum of the time decay heat values ​​of all SKUs within the cluster after the update; mapping the mid-frequency stable layer SKUs to the storage location block at the middle distance; keeping the long-tail silent layer SKUs in the far-end storage location unchanged; comparing the difference between the layered storage location mapping result and the current actual storage location, extracting only the storage location change information corresponding to SKUs that have undergone hierarchical jumps and cluster affiliation changes, generating a list of SKUs to be migrated and their target storage locations; calculating the path improvement benefit for each SKU to be migrated; if the total number of SKUs to be migrated exceeds the preset single migration budget limit, sorting them according to the path improvement benefit from high to low and truncating them to generate a controlled intelligent storage location allocation instruction list.

[0010] Furthermore, it also includes: generating a simulated task queue based on the time-period picking demand prediction of high-frequency active SKUs, and using a multi-robot path planning algorithm to perform path planning and conflict detection on the channel network graph to generate a spatiotemporal conflict heatmap of the channel segment; wherein, the picking demand prediction adopts a long short-term memory network model, the input of which is the outbound frequency sequence of each SKU in the most recent several time periods, and the output is the predicted picking demand quantity of the SKU in the next time period; the input data is preprocessed using Z-score normalization, and the output is denormalized to restore the actual picking demand quantity; the multi-robot path planning algorithm adopts a time window-based cooperative A* path search algorithm to plan a timestamped path sequence for each simulated robot, using the Manhattan distance from the current position to the target storage location as the heuristic function estimate, and querying the time window occupancy records of planned robots to avoid conflicts when expanding candidate positions; traversing each aisle segment to detect whether the number of occupants in the same time window exceeds the upper limit of the number of robots that can pass, counting the cumulative number of spatiotemporal conflict events, and generating the spatiotemporal conflict heatmap of the channel segment.

[0011] Furthermore, it also includes: extracting several channel segments with the highest cumulative number of conflicts from the spatiotemporal conflict heatmap of the channel segments as a set of conflict hotspot channel segments; for each conflict hotspot channel segment, tracing all high-frequency picking SKUs that pass through the channel segment on the reachable path to the target location, calculating a conflict contribution score for each traced SKU, the conflict contribution score being equal to the ratio of the number of picking tasks of the SKU passing through the conflict hotspot channel segment to the total number of picking tasks of all SKUs on the conflict hotspot channel segment, multiplied by the ratio of the cumulative number of conflicts in the conflict hotspot channel segment to the upper limit of the number of robots that can pass through it; for SKUs with a conflict contribution score higher than a preset conflict contribution threshold and currently in the mid-frequency stable layer or the long-tail silent layer, temporarily upgrading their hierarchical label to the high-frequency active layer, generating a conflict-driven hierarchical upgrade SKU list and synchronously updating the SKU hierarchical label mapping table.

[0012] Furthermore, it also includes: selecting association clusters containing SKUs that drive hierarchical upgrades or whose reachable paths to target storage locations pass through conflict hotspot passage segments from the incrementally updated high-frequency active layer association cluster set, forming a conflict association cluster set; searching for candidate migration storage locations for each conflict association SKU in the conflict association cluster set, where candidate storage locations must simultaneously meet the following conditions: the reachable path does not pass through any passage segment in the conflict hotspot passage segment set, and the expected load of the area does not exceed the throughput capacity limit of the area; performing local multi-robot path simulation for each candidate migration scheme, calculating a conflict resolution evaluation value, wherein the conflict resolution evaluation value is equal to the reduction in the cumulative number of conflicts in the conflict hotspot passage segment after migration multiplied by the conflict resolution weight coefficient, minus the increase in the path length from the SKU to the outbound port after migration multiplied by the path length penalty coefficient, wherein the reduction in the cumulative number of conflicts and the increase in the path length are respectively normalized by the mean based on the range before being substituted into the calculation.

[0013] Furthermore, it also includes: taking the weighted sum of minimizing the total frequency of global channel conflicts and the variance of peak load rate in each region within the high-frequency active layer as the joint optimization objective, and under the constraints that the total number of migrations does not exceed the preset migration budget limit, the increase in path length for each SKU does not exceed the preset proportion of the path length before migration, and SKUs within the same associated cluster remain in the same storage location block or adjacent blocks after migration, a heuristic search algorithm is used to combine and optimize candidate migration schemes to generate a collaborative storage location migration scheme; the collaborative storage location migration scheme is merged with the hierarchical storage location mapping results, and if the same SKU appears in both, the target storage location specified by the collaborative storage location migration scheme is used; the merged result is compared with the current actual storage location, and a comprehensive score is calculated for each migration instruction. The comprehensive score is the weighted sum of path improvement benefits and conflict resolution evaluation value. Migration instructions are truncated from high to low according to the comprehensive score until the total migration does not exceed the budget limit, generating the final controlled smart storage location allocation instruction list.

[0014] This invention discloses a warehouse management system supporting intelligent storage location allocation, comprising: an SKU layering module, used to acquire the time-decay heatmap values ​​of all SKUs and warehouse aisle network topology data, dividing SKUs into three layers—a high-frequency active layer, a mid-frequency stable layer, and a long-tail silent layer—based on the distribution of time-decay heatmap values, and generating an SKU layer-level label mapping table and an aisle network diagram; an incremental heatmap update module, used to update the incremental heatmap values ​​of actively changing SKUs based on newly added outbound order data, comparing the updated heatmap values ​​with the layer interval boundaries, identifying layer-transition SKUs, and generating a layer-transition SKU list; an incremental clustering module, used to update the incremental association strength matrix of high-frequency active layer SKUs based on newly added order data, performing incremental spectral clustering based on the previous round of clustering results, generating an incrementally updated high-frequency active layer association cluster set; and a storage location mapping and instruction generation module, used to map storage location areas to each layer of SKUs based on the high-frequency active layer association cluster set, comparing the mapping results with the current actual storage location, and generating a controlled intelligent storage location allocation instruction list.

[0015] The beneficial effects of this invention are as follows: By dividing SKUs into three levels according to their time-decaying heat values ​​and performing incremental updates only on high-frequency active SKUs and those transitioning between levels, the dimension of the association strength matrix is ​​limited to the order of magnitude of SKUs in the high-frequency active layer. This solves the technical problems of excessive computational complexity of the full association strength matrix and large-scale inventory relocation caused by each update in scenarios with millions of SKUs. It achieves the technical effect of controlling the scale of location redistribution calculations within the high-frequency active layer and reducing unnecessary location relocation. Through a multi-robot path planning and conflict contribution scoring mechanism based on simulated task queues, it solves the technical problem of low- and medium-frequency SKUs remaining outside the path conflict optimization scope for a long time due to not triggering heat level updates. This achieves the technical effect of inversely influencing the SKU stratification strategy with path conflict information. By using a joint optimization objective function to collaboratively constrain the global conflict frequency and regional peak load rate variance, and setting continuity constraints on the association cluster layout, it solves the technical problem of mutual interference between conflict resolution migration and heat-stratified location layout. This achieves the technical effect of cooperating the two types of optimization strategies under a unified constraint framework and avoiding mutual interference. Attached Figure Description

[0016] Figure 1 This is a flowchart of the warehouse management method supporting intelligent storage location allocation according to the present invention; Figure 2 This is a representative SKU thermal value and hierarchical distribution diagram of the present invention; Figure 3 This is a comparison chart of the incremental thermal values ​​of representative SKUs before and after the update of the present invention; Figure 4 This is a SKU three-level hierarchy quantity distribution diagram of the present invention; Figure 5 This is a ranking (top 6) of the number of spatiotemporal conflicts in the channel segments according to the present invention; Figure 6 This is a graph showing the SKU conflict contribution score distribution (M-07 channel segment) of this invention; Figure 7 This is a comparison chart of conflict resolution evaluation for the SKU-A001 candidate migration scheme of the present invention; Figure 8 This is a peak load rate distribution diagram of each cargo location area in the high-frequency active layer of the present invention. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0018] This embodiment provides a warehouse management method that supports intelligent location allocation, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain full SKU heat data and warehouse channel network topology data, divide SKUs into three levels based on heat value distribution, and generate SKU level label mapping table and channel network diagram; Obtain the complete SKU list and its most recently calculated time-degradation thermodynamic value from the warehouse system, and simultaneously acquire the warehouse aisle network topology data. This aisle network topology data includes the directional attributes of each aisle, the coordinates of aisle intersection nodes, and the maximum number of robots that can simultaneously travel through each aisle.

[0019] Based on the time-degradation thermal distribution of all SKUs, SKUs are divided into three levels: a high-frequency active layer, a mid-frequency stable layer, and a long-tail silent layer. Specifically, all SKUs are sorted from highest to lowest thermal value. According to preset lower thresholds for high-frequency and mid-frequency thermal values, SKUs with thermal values ​​not lower than the high-frequency lower threshold are marked as high-frequency active layer; SKUs with thermal values ​​between the mid-frequency and high-frequency lower thresholds are marked as mid-frequency stable layer; and SKUs with thermal values ​​lower than the mid-frequency lower threshold are marked as long-tail silent layer. A corresponding level label is generated for each SKU, and these are aggregated to form an SKU level label mapping table.

[0020] Based on the warehouse aisle network topology data, the aisle intersection nodes and cargo location connection nodes are used as graph nodes, and aisle segments are used as graph edges. A directional attribute and an upper limit attribute for the number of robots that can pass through each graph edge are added to generate an aisle network graph.

[0021] It should be noted that the aforementioned time-decay heat value refers to the comprehensive heat value obtained by weighting and summing the contributions of each outbound event based on the historical outbound records of the SKU using an exponential time decay function. For any SKU, its time-decay heat value... The calculation method is as follows: in, This represents the total number of outbound events for this SKU within the statistical period. This is the sequence number of the outbound event. For the first The time interval between the next outbound event and the current time, in hours. The time decay coefficient, The value is a positive real number, used to control the rate at which the thermal contribution of historical outbound events diminishes over time. This is a natural constant. The closer the outflow event is to the current time, the greater its contribution to the heat value; conversely, the contribution of outflow events further back to the current time decays exponentially, approaching zero. Due to the time decay of the heat value... It is the sum of the exponential decay weights of each outbound event, and its dimension is a dimensionless pure numerical value. The time decay heat values ​​of different SKUs can be directly compared and sorted.

[0022] It should be noted that the determination of the lower bound thresholds for the high-frequency and mid-frequency thermal values ​​can be based on the quantiles of the thermal value distribution. For example, the lowest thermal value corresponding to the top 5% of SKUs after thermal value sorting can be used as the lower bound threshold for the high-frequency layer, and the lowest thermal value corresponding to the top 30% of SKUs can be used as the lower bound threshold for the mid-frequency layer. In other words, the high-frequency active layer includes the top 5% of SKUs in terms of thermal value sorting, the mid-frequency stable layer includes SKUs in the 5% to 30% range, and the long-tail silent layer includes the remaining SKUs. These quantile ratios can be adjusted according to actual warehouse operations.

[0023] Step 2: Based on the newly added outbound order data, update the incremental heat value of the active changing SKUs, compare the updated heat value with the hierarchical interval boundary, identify the hierarchical transition SKUs, and generate the updated SKU heat value table and hierarchical transition SKU list; Retrieve new outbound order data since the last heat value calculation, extract the relevant SKU identifiers from the new outbound order data, and form an active change SKU set after summarizing and deduplicating.

[0024] For each SKU in the set of actively changing SKUs, an incremental heatmap value is updated using an exponential time decay function. Specifically, for any SKU in the set of actively changing SKUs, its updated time-decayed heatmap value... The calculation method is as follows: in, This is the stock time decay thermodynamic value obtained from the last calculation for this SKU. This is the time interval from the last calculation to the current time, in hours. This refers to the number of times the SKU appears in new outbound orders. This is the sequence number of the newly added outbound event. For the first The time interval between the current moment and the next new outbound event, in hours. The time decay coefficient defined in step 1, is a natural constant. The first part of the above formula calculates the decay of the existing time-degradation heat value according to the elapsed time interval, and the second part calculates the incremental decay weight value of the newly added outbound event. The two are added together to obtain the updated time-degradation heat value.

[0025] For SKUs not included in the active changing SKU set, their time-decay thermal values ​​are only calculated based on stock decay, i.e. .

[0026] The updated time-decay heatmap value of each SKU is compared with the heatmap value range boundary of its current level. If the updated time-decay heatmap value of an SKU exceeds the lower bound threshold of the high-frequency layer heatmap value and its current level is not a high-frequency active layer, the SKU is marked as having undergone an upward level transition; if the updated time-decay heatmap value of an SKU is lower than the lower bound threshold of its current level heatmap value, the SKU is marked as having undergone a downward level transition. All SKUs that have undergone level transitions are summarized, a level transition SKU list is generated, and the level labels of the corresponding SKUs in the SKU level label mapping table are updated synchronously. An updated SKU heatmap value table is generated.

[0027] Step 3: Update the incremental association strength matrix of high-frequency active layer SKUs and hierarchical transition SKUs based on the new order data. Perform incremental spectral clustering based on the previous round of clustering results to generate the incrementally updated high-frequency active layer association cluster set. Based on newly added outbound order data, the incremental association strength matrix of high-frequency active SKUs (including SKUs newly entering the high-frequency active layer due to layer jumps) is updated. Specifically, only for high-frequency active SKU pairs appearing simultaneously within the same new order, the incremental weighted co-occurrence frequency is calculated, and the incremental weighted co-occurrence frequency is accumulated to the corresponding element of the existing association strength matrix. For any two high-frequency active SKUs... and The incremental update method for the elements of its correlation strength matrix is ​​as follows: in, This is the stock correlation strength value obtained from the previous calculation. For new orders of SKUs and SKU The number of orders that appeared at the same time. The sequence number of the co-occurring order. For the first The time interval between each co-occurring order and the current time, in hours. This is the time interval from the last calculation to the current time, in hours. The time decay coefficient defined in step 1, It is a natural constant. Because... It is the sum of the exponentially decaying weights of all co-occurring events, and its dimension is a dimensionless pure numerical value, which can be directly used for subsequent incremental spectral clustering operations.

[0028] Using the previous round of clustering results as the initial cluster assignment, incremental spectral clustering is performed. The input to incremental spectral clustering is the updated association strength matrix of the high-frequency active layer and the cluster affiliation labels of each SKU from the previous round; the output is the updated cluster affiliation labels of each SKU. For SKUs newly entering the high-frequency active layer due to hierarchical transitions, the cosine similarity between their association strength row vector and the existing cluster center vectors is calculated, and they are assigned to the cluster with the highest cosine similarity. For SKUs leaving the high-frequency active layer due to hierarchical transitions, they are removed from their respective clusters, and the cluster center vector of that cluster is updated. This generates the incrementally updated set of association clusters in the high-frequency active layer.

[0029] It should be noted that the cluster center in the above incremental spectral clustering refers to the mean vector of the row vectors corresponding to the correlation strength matrix of each SKU within the cluster.

[0030] Furthermore, the cosine similarity is calculated as follows: Let the row vector of the association strength of the newly entered SKU in the high-frequency active layer be... , No. The cluster center vectors of the existing clusters are Then this SKU is the same as the first The cosine similarity of each cluster is ,in Let be the dot product of two vectors. and Two vectors respectively Norm, The cluster number is used to assign the SKU to the existing clusters. All existing clusters are traversed, and the cluster with the highest cosine similarity is selected as the cluster to which the SKU belongs.

[0031] It should be noted that the incremental association strength matrix described above only maintains the association relationships between SKUs in the high-frequency active layer. Since the number of SKUs in the high-frequency active layer is much smaller than the total number of SKUs, the dimension of the association strength matrix is ​​limited to the scale of the number of SKUs in the high-frequency active layer, thus avoiding the memory consumption and computational overhead of a million × million level full matrix.

[0032] Step 4: Based on the hierarchical results and the high-frequency active layer association cluster set, perform location area mapping for each level of SKU, compare the mapping results with the current actual location, and generate a controlled intelligent location allocation instruction list; The clusters in the incrementally updated high-frequency active layer are sorted from highest to lowest based on their total cluster heat value. The top-ranked clusters are mapped to the priority storage location block closest to the outbound gate. The total cluster heat value is the sum of the time-decayed heat values ​​of all SKUs within the cluster after updates. SKUs in the mid-frequency stable layer are mapped to storage location blocks at intermediate distances. SKUs in the long-tail silent layer remain in their far-end storage locations.

[0033] The hierarchical location mapping results are compared with the current actual locations. Only the location change information corresponding to SKUs that have undergone hierarchical transitions or cluster affiliations is extracted to generate a list of SKUs to be migrated and their target locations. For each SKU in the list to be migrated, the change in the path length from the migration point to the outbound gate compared to before the migration is calculated and recorded as the path improvement benefit. If the total number of SKUs to be migrated exceeds the preset single migration budget limit, the SKUs to be migrated are sorted from highest to lowest according to the path improvement benefit, and the top-ranked SKUs are selected to ensure that the total migration amount does not exceed the budget limit. A controlled intelligent location allocation instruction list is generated, with each instruction containing an SKU identifier, source location number, and target location number.

[0034] In this embodiment of the application, in order to detect spatiotemporal conflicts on the channel network when multiple robots are picking goods in parallel, the following steps are also included in step 2: Step 5: Based on the time-period picking demand prediction of high-frequency active layer SKUs, generate a simulated task queue, and use a multi-robot path planning algorithm to perform path planning and conflict detection on the channel network graph to generate a spatiotemporal conflict heat map of the channel segment. Based on the outbound frequency trends of high-frequency active SKUs and hierarchical transition SKUs over recent time periods, the picking demand for the next time period is predicted, generating a simulated task queue. Each task record in the simulated task queue includes the target SKU identifier, target location number, and expected execution time.

[0035] Furthermore, the prediction of the picking demand quantity is implemented using a Long Short-Term Memory (LSTM) network model. The input to the LSTM model is the outbound frequency sequence of each SKU over several recent time periods. The output layer is a fully connected layer, outputting the predicted picking demand quantity (a single scalar value) for that SKU in the next time period. Before inputting the outbound frequency sequence into the LSTM model, the outbound frequency data for each time period is preprocessed using Z-score standardization to eliminate the impact of differences in the magnitude of outbound frequency between different SKUs on model training. The standardized predicted value output by the LSTM model is then de-standardized to restore the actual picking demand quantity, which serves as the basis for generating the simulated task queue. The LSTM model uses the actual outbound frequency of each time period as the supervision label, is trained using the mean squared error loss function, and is optimized using the Adam optimization algorithm.

[0036] Using a multi-robot path planning algorithm, a complete path from the current location to the target storage location and then to the exit is planned for each simulated robot on the channel network diagram. The path planning results include the estimated time window for each path segment. Each node and each lane segment in the channel network diagram is traversed to check whether the number of simulated robots occupying that node or lane segment within the same time window exceeds the upper limit of the number of robots that can pass through. Cases exceeding the upper limit are recorded as a spatiotemporal conflict event. The cumulative number of spatiotemporal conflict events for each lane segment during the simulation period is counted to generate a heatmap of spatiotemporal conflicts for the channel segments.

[0037] It should be noted that the multi-robot path planning algorithm described above adopts a time-window-based cooperative A-path search algorithm. The input of the time-window-based cooperative A-path search algorithm is the channel network diagram, the starting coordinates of each simulated robot, and the coordinates of the target cargo location. The output is the timestamped path sequence of each simulated robot.

[0038] Furthermore, the specific execution process of the above-mentioned time window-based collaborative A* path search algorithm is as follows: An independent search queue is maintained for each simulated robot. Each node in the search queue records the current position, current time, and planned path. The Manhattan distance from the current position to the target location is used as the heuristic function estimate, and the search nodes are expanded from smallest to largest based on the sum of the actual cost and the heuristic estimate. When expanding each candidate next position, the time window occupancy record of the planned robot is queried. If the position has been occupied by other robots at the corresponding time and the number of occupants reaches the upper limit of the number of passable robots, the expansion direction is marked as impassable and skipped. The path planning of all simulated robots is completed sequentially in the above order. Later planned robots must avoid the time window occupancy records of planned robots during the search, thereby achieving path conflict avoidance among multiple robots.

[0039] In this embodiment of the application, in order to feed back path conflict information to the SKU hierarchical strategy and ensure that low- and medium-frequency SKUs that contribute significantly to the conflict but whose heat values ​​do not change much are included in the incremental optimization scope in a timely manner, the following steps are also included in step 5: Step 6: Identify conflict hotspot channels from the spatiotemporal conflict heatmap, trace the high-frequency picking SKUs that pass through the conflict hotspot channels and calculate the conflict contribution score, temporarily promote the level of low- and medium-frequency SKUs that meet the conflict contribution score conditions, and generate a conflict-driven level-up SKU list. From the spatiotemporal conflict heatmap of the channel segments, extract the channel segments with the highest cumulative number of conflicts as the set of conflict hotspot channel segments.

[0040] For each conflict hotspot lane segment, trace all frequently picked SKUs whose reachable paths to the target location pass through that segment. For each traced SKU, calculate its conflict contribution score for that conflict hotspot lane segment. : in, This represents the number of picking tasks performed by this SKU during the simulated time period, traversing the conflict hotspot passageway. This is the ratio of the cumulative number of conflicts in the conflict hotspot passage segment to the maximum number of robots that can pass through the conflict hotspot passage segment. This represents the total number of SKU picking tasks across all lanes in the conflict hotspot. and The units have the same dimensions (both are frequencies), and their ratio is the proportion of dimensionless frequencies; The ratio of cumulative conflict counts to channel capacity limit is also a dimensionless numerical value; therefore, the conflict contribution score... It is a dimensionless score, which can be directly compared with the preset conflict contribution threshold.

[0041] Furthermore, the aforementioned preset conflict contribution threshold is determined as follows: in historical simulation data, the conflict contribution score of each SKU on all conflict hotspot channel segments is statistically analyzed. The distribution of the conflict contribution score is used as the threshold for conflict contribution. For example, the conflict contribution score is taken as the threshold value. The 75th percentile of the distribution is used as the threshold, meaning only the top 25% of SKUs in terms of conflict contribution score are included in the temporary tier upgrade. The above percentile ratio can be adjusted based on actual warehouse operations.

[0042] Conflict contribution rating SKUs whose conflict contribution score exceeds the preset conflict contribution threshold and are currently in the mid-frequency stable layer or long-tail silent layer will have their hierarchical labels temporarily promoted to the high-frequency active layer. A conflict-driven list of SKUs with promoted hierarchical labels will be generated, and the SKU hierarchical label mapping table will be updated simultaneously. It should be noted that this promotion is temporary. If, in a subsequent update cycle, the conflict contribution score of the SKU falls below the preset conflict contribution threshold and its time-decay thermal value still does not meet the conditions for the high-frequency active layer, its original hierarchical label will be restored.

[0043] In this embodiment of the application, in order to conduct a targeted assessment of the location migration of conflict-related SKUs, the following steps are also included in addition to steps 6 and 3: Step 7: Search for candidate migration location sets for the association clusters involving conflicting SKUs in the incrementally updated high-frequency active layer association cluster set, perform local path simulation for each candidate migration scheme and calculate the conflict resolution evaluation value, and generate the conflict resolution evaluation results for each candidate scheme; From the incrementally updated set of high-frequency active layer associated clusters, select the associated clusters containing SKUs that are part of conflict-driven hierarchical upgrades or SKUs whose reachable paths to target storage locations pass through any segment of the conflict hotspot channel segment set, and form a set of conflict associated clusters.

[0044] For each conflict-related SKU in the conflict-related cluster set, search for a set of candidate migration locations. A candidate location must meet two conditions simultaneously: first, the reachable path of the candidate location does not pass through any segment in the current conflict hotspot segment set; second, the expected load of the area where the candidate location is located does not exceed the throughput capacity limit of the area.

[0045] Furthermore, the method for determining whether the reachable path of the above candidate storage location passes through the conflict hotspot channel segment is as follows: On the channel network graph, taking the storage location connection node corresponding to the candidate storage location as the starting point and the outbound node as the ending point, a breadth-first search algorithm is used to enumerate all simple paths, and each path is checked to see if the graph edge it passes through coincides with any channel segment in the conflict hotspot channel segment set; if all enumerated paths do not pass through any channel segment in the conflict hotspot channel segment set, then the candidate storage location satisfies the first condition.

[0046] For each candidate migration scheme, after replacing the storage location of the conflict-related SKU with the candidate storage location, the local multi-robot path simulation in step 5 is re-executed. Path planning is only performed on the simulation tasks involving the conflict-related SKU and its clustered SKUs, and the change in conflict frequency at conflict hotspot passageways is statistically analyzed. The conflict resolution evaluation value for each candidate migration scheme is calculated. : in, This represents the cumulative reduction in the number of conflicts in the relocated conflict hotspot corridor segment, expressed in times. This represents the increase in path length from the conflicting SKU to the outbound port after migration, in meters. For conflict resolution weighting coefficients, This is the path length penalty coefficient. Because... and The units of measurement are different, so the weighting coefficients need to be resolved through conflict resolution. and path length penalty coefficient Perform dimensional normalization on both items: Before actual use, normalize the historical data separately. and The sample is normalized using mean normalization based on the range, scaling both values ​​to the same numerical range. The normalized values ​​are then substituted into the above formula for calculation, thus yielding the conflict resolution evaluation value. This is a dimensionless comprehensive score. Conflict resolution assessment value. The larger the value, the higher the overall benefit of the candidate solution.

[0047] In this embodiment of the application, in order to coordinately optimize conflict resolution and load balancing on a global scale and avoid conflict resolution migration from disrupting the thermally stratified storage location layout, the following steps are also included in step 7: Step 8: With the joint optimization objective of minimizing the weighted sum of the total frequency of global channel conflicts and the variance of regional peak load rate, the candidate migration schemes are combined and optimized under the constraints of migration budget and layout continuity to generate a collaborative storage location migration scheme. The joint optimization objective is to minimize the weighted sum of the total frequency of global channel conflicts and the variance of the peak load rate of each region within the high-frequency active layer. The joint optimization objective function is as follows: Defined as: in, This represents the cumulative total number of spatiotemporal conflict events across all channel segments. This represents the variance of the peak load rate of each storage location area within the high-frequency active layer. As a conflict frequency weight, For load balancing weights. The calculation method is as follows: assuming there are a total of [number] in the high-frequency active layer. Each storage area The total number of storage areas, the first The peak load rate of each storage area is , If it is the number of the storage location area, then ,in This represents the average peak load rate for each region. Defined as the ratio of the peak number of robot tasks simultaneously en route in the region during the simulation period to the upper limit of the region's throughput capacity, where the simulation period is consistent with the prediction period corresponding to the multi-robot path simulation performed in step 5. Since It is the ratio of the peak number of tasks to the upper limit of throughput capacity, with the same dimension, therefore It is a dimensionless pure numerical value. It is also a dimensionless quantity. The dimension of is the number. Since they are dimensionless quantities, and their dimensions are different, they are substituted into the joint optimization objective function. Previously, respectively for and By employing mean normalization based on the range to scale both values ​​to the same numerical range before weighted summation, the objective function can be jointly optimized. The objective value is a dimensionless comprehensive optimization value. and These are all functions that combine candidate migration schemes and vary depending on the selected migration scheme.

[0048] Set the following constraints: First, the total number of migrations shall not exceed the preset migration budget limit; Second, the increase in the path length from each SKU to the outbound port after migration shall not exceed the preset proportion of the path length before migration; Third, the high-frequency active layer associated clusters shall not be split across regions, that is, SKUs in the same associated cluster shall remain in the same storage location block or adjacent blocks after migration.

[0049] A heuristic search algorithm is used to perform combinatorial optimization on candidate migration schemes for conflict-related SKUs. The input of the heuristic search algorithm is each candidate migration scheme generated in step 7 and its conflict resolution evaluation value. The output is the combination of candidate schemes that satisfies the above constraints and minimizes the joint optimization objective function value. The specific solution process is as follows: Candidate schemes are traversed sequentially in descending order of conflict resolution evaluation value, and candidate schemes are added to the current scheme combination step by step; after each scheme is added, it is verified whether the constraints are satisfied and the change in the joint optimization objective function value is calculated; if the joint optimization objective function value decreases after adding the scheme and the constraints are satisfied, the scheme is retained; otherwise, the scheme is skipped; the search is terminated when all candidate schemes have been traversed or the migration budget is exhausted, and the coordinated storage location migration scheme is output.

[0050] Furthermore, the calculation method for the change in the above-mentioned joint optimization objective function value is as follows: after each candidate solution is added to the current solution combination, based on the cargo location allocation status corresponding to the current solution combination, the global multi-robot path simulation in step 5 is re-executed, and the total number of spatiotemporal conflict events in all channel segments is counted to update the value. And recalculate the peak load rate of each storage area within the high-frequency active layer to update The updated version and After normalization, the sub-function is input into the joint optimization objective function. The difference between the joint optimization objective function value and the value before the addition of the scheme is calculated to obtain the change in the joint optimization objective function value; if the change is negative (i.e. the joint optimization objective function value decreases) and the constraint conditions are met, the scheme is retained.

[0051] In this embodiment, to unify and integrate the collaborative storage location migration scheme with the hierarchical storage location mapping results and ensure that the number of migration instructions output is controllable, the controlled intelligent storage location allocation instruction generation process in step 4 is optimized based on steps 8 and 4. Specifically, the collaborative storage location migration scheme is merged with the hierarchical storage location mapping results in step 4. During the merging process, if the same SKU appears in both the hierarchical storage location mapping results and the collaborative storage location migration scheme, the target storage location specified by the collaborative storage location migration scheme shall prevail. The merged results are compared with the current actual storage locations, and only the storage location migration instructions corresponding to the SKUs involved in hierarchical jumps, cluster affiliation changes, and conflict resolution migrations are output. If the total migration amount exceeds the preset single migration budget limit, a comprehensive score is calculated for each migration instruction. The comprehensive score is the weighted sum of the path improvement benefit and the conflict resolution evaluation value. Migration instructions are truncated from high to low according to the comprehensive score until the total migration amount does not exceed the budget limit, generating the final controlled intelligent storage location allocation instruction list.

[0052] The following is an example of an application of the present invention, such as... Figure 2-8As shown, the implementation process is as follows: A large e-commerce fulfillment center (hereinafter referred to as "Warehouse Center A") is responsible for the storage, picking, and shipping of goods in East China. The warehouse has a total area of ​​approximately 80,000 square meters, storing 1.2 million SKUs, and deploying 280 autonomous mobile robots (AMRs) to perform picking tasks in parallel. The warehouse aisle network consists of 12 main longitudinal aisles and 36 connecting transverse aisles, forming 187 intersection nodes. The outbound exit is located in the central area on the south side of the warehouse. During peak hours (10:00 AM to 12:00 PM), the concurrent order volume can reach 18,000 orders per hour. Historical data shows that the frequency of robot path conflicts remains high in some main aisle sections during peak hours.

[0053] At 09:58 AM on a certain morning in 20XX, the warehouse management system triggered a location reallocation calculation cycle, which was initiated after the previous calculation time (09:00). Minutes (approximately 0.967 hours). System time decay factor is set. The lower limit threshold of the thermal value of the high-frequency layer corresponds to the lowest value in the top 5% of the thermal value ranking, and the lower limit threshold of the thermal value of the mid-frequency layer corresponds to the lowest value in the top 30% of the thermal value ranking. The maximum budget for a single migration is 200 entries.

[0054] The system reads the previous time decay heat values ​​of all 1.2 million SKUs from the warehouse management system and reads the warehouse aisle network topology data. After sorting the heat values ​​from high to low, according to the quantile rules: the top 5% (approximately 60,000 SKUs) are assigned to the high-frequency active layer, 5% to 30% (approximately 300,000 SKUs) are assigned to the mid-frequency stable layer, and the rest (approximately 840,000 SKUs) are assigned to the long-tail silent layer.

[0055] Taking six representative SKUs as examples, the results of heat value calculation and hierarchical classification are presented. Taking SKU-A001 as an example, it had four outbound events within the statistical period, and the time-decay heat value was calculated as follows: Table 1. Calculation results and hierarchical classification of representative SKU thermal values The network diagram uses 187 intersection nodes and cargo docking nodes as nodes, and each lane segment as edges. Each edge is assigned a directional attribute (one-way / two-way) and a maximum number of robots that can pass through it. Taking the main lane segment M-07 as an example, it is a two-way passage, and the maximum number of robots that can pass through it is 3.

[0056] The system retrieved 1,423 new outbound orders from the order management system between 09:00 and 09:58, involving 8,714 active SKUs after deduplication.

[0057] The incremental update calculation process is illustrated using SKU-B112 and SKU-A001 as examples. SKU-B112 appears twice in the new order, with the first outbound event occurring a time after the current time. Hours, the second time since the current time Hourly stored heat value , Hour: After the SKU-B112 update, the thermal value jumped from 1.654 to 3.518, exceeding the lower threshold of the thermal value of the high-frequency layer (the threshold for this cycle is 3.102), triggering an upward transition from the mid-frequency stable layer to the high-frequency active layer.

[0058] Table 2. Update results of incremental heat values ​​for representative SKUs A total of 312 hierarchical transition SKUs were identified this period, including 287 upward transitions (mid frequency → high frequency) and 25 downward transitions (high frequency → mid frequency). A hierarchical transition SKU list was generated and the SKU hierarchical label mapping table was updated synchronously.

[0059] The total number of high-frequency active SKUs this period (including 287 newly added SKUs) is approximately 60,287. The system only maintains a correlation strength matrix among approximately 60,000 SKUs, rather than a full matrix of 1.2 million × 1.2 million.

[0060] Taking SKU-A001 and SKU-A047 as examples, they co-occur in 14 new orders, and the average time interval between the m-th co-occurring order and the current time is approximately 0.51 hours. (Inventory correlation strength) , Hour: For SKU-B112, which has newly entered the high-frequency active layer, the cosine similarity between its association strength row vector and the center vectors of each existing cluster is calculated. The cosine similarity between SKU-B112 and cluster C-03 is 0.847, the cosine similarity with cluster C-07 is 0.312, and the cosine similarity with the other clusters is all below 0.3. Therefore, SKU-B112 is assigned to cluster C-03, which has the highest cosine similarity.

[0061] Table 3. Incremental spectral clustering results (some representative SKUs) This period's high-frequency active layer associated cluster set contains a total of 1,847 associated clusters. After the incremental update, the cluster center vector of cluster C-03 is updated synchronously with the addition of SKU-B112 and the removal of SKU-E021.

[0062] The 1,847 associated clusters are sorted from highest to lowest based on their total cluster heat value. The top-ranked clusters are mapped to the nearest storage location block in Zone A. Taking cluster C-03 as an example, its total cluster heat value is the sum of the updated heat values ​​of all SKUs within the cluster: Cluster C-03 ranks 3rd in total thermal value and is mapped to storage location block 01 in area A (approximately 18 meters from the exit). The mid-frequency stabilization layer SKU is mapped to storage location block B (approximately 45 meters from the exit), while the long-tail quiescent layer SKU remains unchanged at the far end of storage location in area C.

[0063] After comparing the differences, a total of 487 SKUs were identified in this cycle that needed to be migrated, of which 357 were caused by hierarchical jumps and 130 were caused by changes in cluster affiliation. Since 487 exceeded the budget limit of 200 per migration, the top 200 were selected from highest to lowest path improvement benefit to generate an initial list of controlled smart storage location allocation instructions (which will be replaced and optimized by the collaborative solution in step 8 later).

[0064] Based on the outbound frequency sequences of approximately 60,000 SKUs in the high-frequency active layer over the most recent 8 time periods (each 30 minutes), the LSTM model (input sequence length 8, output layer is a fully connected layer outputting a single scalar) is used to predict the picking demand for the next time period (10:00 to 10:30). Taking SKU-A001 as an example, its outbound frequency sequences over the past 8 time periods are standardized by Z-score and then input into the LSTM. The standardized predicted value is de-standardized to obtain a predicted picking demand of 7 times, generating 7 simulated task records.

[0065] After the system aggregates all SKUs in the high-frequency active layer, the simulation task queue for this period contains a total of 43,218 task records, which are assigned to 280 simulated robots for execution.

[0066] A time-window-based collaborative A* path search algorithm was used to plan complete paths for 280 simulated robots. Taking robot AMR-112 as an example, it starts from its current position node N-043, targets the storage location of SKU-A001 (storage location connection node N-156), and then proceeds to the outbound node N-OUT. The path planning result is: N-043→N-071→N-098→N-127 (passing through main aisle section M-07)→N-156→N-127→N-098→N-OUT. The estimated time window occupied by each segment has been recorded in the common time window table.

[0067] Traverse all passage segments and count the number of events exceeding the maximum number of robots that can pass through within each time window.

[0068] Table 4. Heat map of spatiotemporal conflicts in the channel segment (top 6 in terms of conflict frequency) The set of conflict hotspots was determined to be the top 3 in terms of cumulative conflict frequency: M-07 (87 times), M-03 (74 times), and X-14 (61 times).

[0069] For the conflict hotspot section M-07, all high-frequency picking SKUs whose reachable paths to the target location pass through M-07 were traced, totaling 214 SKUs. Taking SKU-B209 as an example (currently in the mid-frequency stable layer), the number of picking tasks it performed through M-07 during the simulated period was recorded. The ratio of the cumulative number of conflicts to the maximum passable limit for M-07 Total number of SKU picking tasks on M-07 ,but: The system is based on historical simulation data statistics. The distribution is used, with the 75th percentile as the conflict contribution threshold; the threshold for this period is 0.187. (SKU-B209) Currently, it is in the mid-frequency stable layer, triggering a temporary upgrade to the high-frequency active layer.

[0070] Table 5 Calculation results of conflict contribution score (M-07 channel segment, some SKUs) Based on the three conflict hotspot channels, a conflict-driven SKU upgrade list was generated in this cycle, involving a total of 38 SKUs, and the SKU level tag mapping table was updated simultaneously.

[0071] From the high-frequency active layer associated cluster set, we select the associated clusters containing SKUs that have conflict-driven hierarchical upgrades (such as SKU-B209) or whose accessible paths to the storage location pass through any of the M-07, M-03, or X-14 channel segments, forming a conflict associated cluster set. This cycle involves a total of 127 associated clusters.

[0072] Taking SKU-A001 (currently located in block 01 of area A, accessible via M-07) as an example, candidate relocation locations are searched. A breadth-first search is used to enumerate paths, filtering out candidate locations whose accessible paths do not pass through M-07, M-03, or X-14, and whose area load does not exceed the upper limit. Three candidate locations are found: LOC-A08 (block 08 of area A, approximately 23 meters from the exit), LOC-A11 (block 11 of area A, approximately 26 meters from the exit), and LOC-B03 (block 03 of area B, approximately 41 meters from the exit).

[0073] For candidate solution LOC-A08, after replacing SKU-A001 with LOC-A08, the local path simulation involving cluster C-03 related tasks is re-executed, and the cumulative number of M-07 conflicts is reduced. The increase in path length from SKU-A001 to the outbound port Meters. After range normalization, , Substitute: Table 6 Conflict Resolution Evaluation Results for SKU-A001 Candidate Migration Schemes LOC-A08 Conflict Resolution Assessment Value The highest-ranked candidate is SKU-A001. The system performs the same evaluation on all conflict-related SKUs within the conflict-related cluster set, generating a set of conflict resolution evaluation results for each candidate solution.

[0074] The joint optimization objective function is defined as: This week's schedule , The high-frequency active layer is divided into multiple layers. The current peak load rates for each of the three storage areas are as follows: Area A, storage areas 01 to 08 have load rates of 0.91, 0.88, 0.84, 0.79, 0.73, 0.67, 0.62, and 0.58 respectively, with an average of: variance: The heuristic search iterates through candidate solutions in descending order of conflict resolution evaluation value. SKU-A001 is migrated to LOC-A08. After adding the scheme combination, re-execute the global path simulation. The number of times decreased from 364 to 333. The value of the joint optimization objective function decreased from 0.01310 to 0.01187 after normalization, and the total number of migrations (1) did not exceed the budget, so this scheme was retained. The process was then iterated step by step until the cumulative number of migration schemes reached the budget limit of 200 or all candidate schemes were traversed. Finally, a total of 163 collaborative warehouse location migration schemes were output.

[0075] The 163 collaborative warehouse relocation schemes were merged with the initial instruction list from Step 4 (487 difference comparison results): For SKUs appearing in both, the target warehouse location specified by the collaborative scheme was used as the criterion. After merging and deduplication, 521 candidate relocation instructions were obtained. Since 521 instructions exceeded the budget limit of 200, a comprehensive score (a weighted sum of path improvement benefits and conflict resolution evaluation values) was calculated for each instruction. The top 200 instructions were selected from high to low based on their comprehensive scores to generate the final controlled smart warehouse allocation instruction list.

[0076] Table 7. List of final controlled smart storage location allocation instructions (first 5 examples) Throughout the implementation process, the data starts with the heatmap values ​​of 1.2 million SKUs obtained in Step 1. Step 2 incrementally updates and identifies 312 SKUs that have undergone hierarchical transitions. Step 3 integrates these transition SKUs into a high-frequency active layer association strength matrix of approximately 60,000 to complete incremental spectral clustering. Step 4 generates an initial list of location mapping differences based on the clustering results. Step 5 uses LSTM to predict picking demand and generates 43,218 simulated tasks. Multi-robot path planning identifies three conflict hotspot channels: M-07, M-03, and X-14. Step 6 performs temporary hierarchical upgrades on 38 low-to-medium frequency SKUs, injecting path conflict information back into the hierarchical system. Step 7 evaluates candidate migration schemes for each conflict-related SKU in 127 conflict-related clusters. Step 8 uses a joint optimization objective function to collaboratively solve the problem. Finally, the hierarchical mapping results are merged with conflict resolution schemes, and the top 200 schemes with the best overall scores are extracted to output the final controlled intelligent location allocation instruction list. This achieves a complete closed-loop data flow from heatmap data to conflict detection and collaborative optimization.

[0077] It is understood that data preprocessing methods known to those skilled in the art include data cleaning, data transformation, and data reduction. Data transformation includes type conversion and normalization and standardization. Although the dimensions and types of data were omitted in the description of the preceding embodiments, data preprocessing is a technical knowledge known to those skilled in the art and a prerequisite step in data processing. Therefore, the previously described well-known data preprocessing steps were not described independently.

[0078] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A warehouse management method supporting intelligent location allocation, characterized in that, Includes the following steps: Obtain the time decay heat values ​​and warehouse channel network topology data of all SKUs. Based on the distribution of time decay heat values, divide the SKUs into three levels: high-frequency active layer, mid-frequency stable layer, and long-tail silent layer. Generate SKU level label mapping table and channel network diagram. Based on newly added outbound order data, the incremental heat value of active and changing SKUs is updated. The updated heat value is compared with the hierarchical interval boundary to identify hierarchical transition SKUs and generate a hierarchical transition SKU list. For high-frequency active layer SKUs, the incremental association strength matrix is ​​updated based on new order data. Incremental spectral clustering is performed using the previous round of clustering results as initialization to generate the set of high-frequency active layer association clusters after incremental update. Based on the high-frequency active layer associated cluster set, the storage location area is mapped for each level of SKU. The mapping result is compared with the current actual storage location to generate a controlled intelligent storage location allocation instruction list. The time decay heat value is obtained by weighted summation of each SKU outbound event using an exponential time decay function. Outbound events closer to the current time contribute more, while outbound events further away contribute exponentially.

2. The warehouse management method supporting intelligent location allocation according to claim 1, characterized in that, The time decay heat value is calculated as follows: For any SKU, the time interval between each outbound event and the current time within the statistical period is substituted into an exponential function with the natural constant as the base and the negative value of the product of the time decay coefficient and the time interval as the exponent. The exponential function values ​​corresponding to each outbound event are accumulated to obtain the time decay heat value of the SKU. The time decay coefficient is a positive real number used to control the rate at which the heat contribution of historical outbound events decays over time.

3. The warehouse management method supporting intelligent location allocation according to claim 1, characterized in that, The incremental heatmap update for actively changing SKUs includes: Extract the relevant SKU identifiers from the newly added outbound order data, and after summarizing and deduplicating, form a set of active and changing SKUs; For each SKU in the active changing SKU set, its stock time decay heat value is multiplied by an exponential function with a base of the natural constant and an exponent of the negative of the product of the time decay coefficient and the time interval from the last calculation time to the current time, to obtain the stock decay converted value; the time intervals from each outbound event of this SKU in the new outbound order to the current time are substituted into the exponential time decay function and accumulated to obtain the incremental decay weight value; the stock decay converted value and the incremental decay weight value are added to obtain the updated time decay heat value; For SKUs that are not in the set of active changing SKUs, only their stock time decay thermal values ​​are converted into stock decay values.

4. The warehouse management method supporting intelligent location allocation according to claim 1, characterized in that, The incremental correlation strength matrix update and incremental spectral clustering include: The incremental weighted co-occurrence frequency is calculated only for high-frequency active SKUs that appear simultaneously in the same new order. For any two high-frequency active SKUs, their existing association strength values ​​are decayed according to the elapsed time interval, and then the exponential decay weight value of the new co-occurrence event is added to obtain the updated association strength value. Using the clustering results from the previous round as the initial cluster assignment, for SKUs that newly enter the high-frequency active layer due to hierarchical transition, the cosine similarity between their association strength row vector and the existing cluster center vectors is calculated, and they are assigned to the cluster with the highest cosine similarity; for SKUs that leave the high-frequency active layer due to hierarchical transition, they are removed from their clusters and the cluster center vector of that cluster is updated; wherein, the cluster center vector is the mean vector of the row vectors corresponding to the association strength matrix of each SKU in the cluster.

5. The warehouse management method supporting intelligent location allocation according to claim 1, characterized in that, The generation of the location area mapping and the controlled intelligent location allocation instruction list includes: The associated clusters in the high-frequency active layer associated cluster set after incremental updates are sorted from high to low according to the total cluster heat value. The associated clusters with higher rankings are mapped to the priority storage location block closest to the outbound gate. The total cluster heat value is the sum of the time decay heat values ​​of all SKUs in the cluster after the update. The mid-frequency stable layer SKUs are mapped to storage location blocks with intermediate distance. The long-tail silent layer SKUs remain unchanged at the far storage location. The hierarchical location mapping results are compared with the current actual locations. Only the location change information corresponding to SKUs that have undergone hierarchical transitions and cluster affiliations is extracted to generate a list of SKUs to be migrated and their target locations. The path improvement benefit is calculated for each SKU to be migrated. If the total number of SKUs to be migrated exceeds the preset single migration budget limit, they are sorted from high to low according to the path improvement benefit and truncated to generate a controlled intelligent location allocation instruction list.

6. The warehouse management method supporting intelligent location allocation according to claim 1, characterized in that, Also includes: Based on the time-period picking demand prediction of high-frequency active layer SKUs, a simulated task queue is generated. A multi-robot path planning algorithm is used to perform path planning and conflict detection on the channel network graph to generate a spatiotemporal conflict heat map of the channel segment. The picking demand prediction adopts a long short-term memory network model. The input of the long short-term memory network model is the outbound frequency sequence of each SKU in the most recent several time periods, and the output is the predicted picking demand quantity of the SKU in the next time period. The input data is preprocessed by Z-score standardization, and the output is restored to the actual picking demand quantity by inverse standardization transformation. The multi-robot path planning algorithm adopts a time window-based collaborative A* path search algorithm to plan a timestamped path sequence for each simulated robot. The Manhattan distance from the current position to the target storage location is used as the heuristic function estimate. When expanding candidate positions, the time window occupancy records of the planned robots are queried to avoid conflicts. The system iterates through each lane segment to check whether the number of occupied robots exceeds the upper limit of the number of robots that can pass through within the same time window, counts the cumulative number of spatiotemporal conflict events, and generates a spatiotemporal conflict heatmap of the channel segment.

7. The warehouse management method supporting intelligent location allocation according to claim 6, characterized in that, Also includes: Extract the channel segments with the highest cumulative number of conflicts from the spatiotemporal conflict heatmap of the aforementioned channel segments as a set of conflict hotspot channel segments; For each conflict hotspot channel segment, trace all high-frequency picking SKUs that pass through the channel segment on the path to the target location. Calculate a conflict contribution score for each traced SKU. The conflict contribution score is equal to the ratio of the number of picking tasks for that SKU in the conflict hotspot channel segment to the total number of picking tasks for all SKUs in the conflict hotspot channel segment, multiplied by the ratio of the cumulative number of conflicts in the conflict hotspot channel segment to the upper limit of the number of robots that can pass through it. For SKUs whose conflict contribution score is higher than the preset conflict contribution threshold and are currently in the mid-frequency stable layer or the long-tail silent layer, their hierarchical labels are temporarily promoted to the high-frequency active layer, generating a conflict-driven hierarchical promotion SKU list and synchronously updating the SKU hierarchical label mapping table.

8. The warehouse management method supporting intelligent location allocation according to claim 7, characterized in that, Also includes: From the incrementally updated set of high-frequency active layer associated clusters, select the associated clusters containing SKUs that are involved in conflict-driven hierarchical upgrades or whose target storage locations are accessible via conflict hotspot passages, and form a set of conflict associated clusters. For each conflict-related SKU in the conflict-related cluster set, search for candidate migration locations. The candidate locations must simultaneously meet the following conditions: the reachable path does not pass through any channel segment in the conflict hotspot channel segment set, and the expected load of the area does not exceed the throughput capacity limit of the area. For each candidate migration scheme, a local multi-robot path simulation is performed to calculate the conflict resolution evaluation value. The conflict resolution evaluation value is equal to the reduction in the cumulative number of conflicts in the conflict hotspot channel segment after migration multiplied by the conflict resolution weight coefficient, minus the increase in the path length from the SKU to the outbound port after migration multiplied by the path length penalty coefficient. The reduction in the cumulative number of conflicts and the increase in the path length are respectively processed by mean normalization based on the range before being substituted into the calculation.

9. The warehouse management method supporting intelligent location allocation according to claim 8, characterized in that, Also includes: With the joint optimization objective of minimizing the weighted sum of the total frequency of global channel conflicts and the variance of the peak load rate of each region within the high-frequency active layer, a heuristic search algorithm is used to combine and optimize candidate migration schemes to generate a collaborative storage location migration scheme, under the constraints that the total number of migrations does not exceed the preset migration budget limit, the increase in the path length of each SKU does not exceed the preset proportion of the path length before migration, and the SKUs in the same associated cluster remain in the same storage location block or adjacent blocks after migration. The collaborative storage location migration scheme is merged with the hierarchical storage location mapping results. If the same SKU appears in both, the target storage location specified by the collaborative storage location migration scheme shall prevail. The merged result is compared with the current actual storage location. A comprehensive score is calculated for each migration instruction. The comprehensive score is the weighted sum of the path improvement benefit and the conflict resolution evaluation value. Migration instructions are extracted from high to low according to the comprehensive score until the total migration amount does not exceed the budget limit, and the final controlled smart storage location allocation instruction list is generated.

10. A warehouse management system supporting intelligent storage location allocation, used to execute the warehouse management method supporting intelligent storage location allocation as described in any one of claims 1 to 9, characterized in that, include: The SKU layering module is used to obtain the time decay heat values ​​of all SKUs and the warehouse channel network topology data. Based on the distribution of time decay heat values, the SKUs are divided into three layers: high-frequency active layer, mid-frequency stable layer and long-tail silent layer, generating an SKU layer label mapping table and channel network diagram. The incremental heatmap update module is used to update the heatmap values ​​of active SKUs based on newly added outbound order data, compare the updated heatmap values ​​with the tier interval boundaries, identify tier transition SKUs, and generate a tier transition SKU list. The incremental clustering module is used to update the incremental association strength matrix of high-frequency active layer SKUs based on new order data. It initializes the previous round of clustering results and performs incremental spectral clustering to generate the incrementally updated set of high-frequency active layer association clusters. The location mapping and instruction generation module is used to map the location areas of each level SKU based on the high-frequency active layer association cluster set, compare the mapping results with the current actual location, and generate a controlled intelligent location allocation instruction list.

Images