Intelligent storage rack and using method thereof

Abstract

The invention relates to the technical field of storage racks, and provides an intelligent storage rack and a using method thereof.The intelligent storage rack comprises a storage rack body, a controller and at least one feature obtaining module used for obtaining feature information of goods to be stored, and the storage rack body is provided with a plurality of storage positions; the controller is used for storing historical access data of the storage positions, the feature obtaining module is connected with the controller and feeds feature information back to the controller, and the controller is further used for obtaining the heat value of each storage position according to the historical access data and constructing an access heat map according to the heat values of the multiple storage positions; according to the historical access data and the feature information, obtaining the cargo comprehensive popularity of the to-be-stored cargo; and allocating an optimal storage position according to the access heat map and the cargo comprehensive heat. The goods storage path can be optimized, the operation time is shortened, the space utilization rate is increased, the comprehensive cost is reduced, and the intelligent degree of the storage rack can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of storage racks, and more particularly, to an intelligent storage rack and a method for using the same. Background Art

[0002] Storage racks are widely used in various industries. For example, the shelves used in shopping malls, the bookshelves used in libraries or bookstores, and the storage shelves used in warehouses all belong to storage racks. With the development of intelligent and robotic technologies, more and more shelves / storage racks are integrated with intelligent or robotic technologies to improve the access efficiency of storage racks.

[0003] For example, the Chinese invention patent with the application number CN201810188194.0 discloses an intelligent shelf and a control method for the intelligent shelf. An image acquisition device is arranged on the shelf body and / or the shelf board to acquire the image information of the target object. The control device can adjust the height of the intelligent shelf according to the height of the shopping person based on the image information of the target object. Another Chinese invention patent with the application number CN202111364038.3 can automatically allocate space according to the size of the shelf, efficiently utilize the storage shelf, automatically enter and exit the warehouse, and record the shelf information in detail, improving the space utilization rate of the shelf. Still another Chinese invention patent with the application number CN202410010597.1 discloses a shelf recognition method, a robot control method, a terminal device, and a computer storage medium, which adopt a combination of three-dimensional point cloud rough recognition and two-dimensional point cloud fine recognition to accurately identify the position of the center of the shelf.

[0004] Therefore, for intelligent storage racks, there are still many technical problems to be solved urgently in their actual applications, and many solutions have not been proposed yet. Summary of the Invention

[0005] Based on this, in order to improve the intelligence level of the storage rack, the present invention provides an intelligent storage rack and a method for using the same. The specific technical solutions are as follows:

[0006] An intelligent storage rack includes a storage rack body, a controller, and at least one feature acquisition module for acquiring the feature information of the goods to be stored. A plurality of storage positions are arranged on the storage rack body. The controller is used to store the historical access data of the storage positions, and the feature acquisition module is connected to the controller and feeds back the feature information to the controller. The controller is further used for:

[0007] Obtaining the heat value of each storage position according to the historical access data, and constructing an access heat map based on the heat values of the plurality of storage positions;

[0008] Obtaining the comprehensive heat of the goods to be stored according to the historical access data and the feature information;

[0009] and allocating an optimal storage location according to the access heat map and the comprehensive heat of the goods.

[0010] By constructing a heat map and obtaining the comprehensive heat of the goods, and allocating an optimal storage location according to the heat map and the comprehensive heat of the goods, the intelligent storage rack can optimize the access path of the goods, shorten the operation time, improve the space utilization rate and reduce the comprehensive cost, which is beneficial to improving the intelligence level of the storage rack.

[0011] Preferably, the controller includes:

[0012] a distance acquisition module, configured to acquire the three-dimensional Euclidean distance between the storage location and each access location in the historical access events;

[0013] a contribution degree acquisition module, configured to acquire the contribution degree of each historical access event to the storage location according to the three-dimensional Euclidean distance;

[0014] a heat map construction module, configured to perform weighted superposition on the contribution degrees to obtain the heat value, and construct an access heat map according to the heat value.

[0015] Preferably, the controller further includes:

[0016] an access frequency acquisition module, configured to acquire the comprehensive access frequency and the recent access frequency of the goods to be stored;

[0017] an associated heat acquisition module, configured to acquire the heat of associated goods;

[0018] a comprehensive heat acquisition module, configured to acquire the comprehensive heat of the goods according to the comprehensive access frequency, the recent access frequency and the heat of associated goods.

[0019] Preferably, the heat map construction module constructs an access heat map according to the formula ;

[0020] where H(x, y, z) represents the heat value of the three-dimensional coordinate point (x, y, z) of the storage location, N represents the total number of historical access events, d i represents the three-dimensional Euclidean distance between the three-dimensional coordinate point (x, y, z) of the storage location and the access location of the i-th time, σ represents the diffusion coefficient of the Gaussian kernel, μ represents the distance offset, e represents the natural constant, represents the contribution degree of the i-th historical access event to the storage location.

[0021] Preferably, the comprehensive heat acquisition module acquires the comprehensive heat of the goods according to the formula H i =α1·F Com_i +α2·F Recent_i +α3·R i ;

[0022] Among them, F Com_i and F Recent_i and R i respectively represent the comprehensive access frequency, recent access frequency, and associated goods popularity, and α1, α2, and α3 respectively represent the weight coefficients of the comprehensive access frequency, recent access frequency, and associated goods popularity.

[0023] Preferably, the intelligent storage rack further includes an access robotic arm that is signal - connected to the controller. The access robotic arm is used to respond to control instructions for order sorting. The controller further includes:

[0024] A path planning module, which is used to, after receiving a sorting order, obtain a picking sequence and plan a path according to the Q - function and generate a control instruction according to the picking sequence and the planned path;

[0025] Among them, s represents the current system state, s t represents the operating state of the system at time t, s represents the currently selected action, and s t represents the specific action of the access robotic arm at time t. γ represents the discount factor, r t+k represents the immediate reward of the system at time t + k. E represents the expectation operator, k represents the future time step index starting from the current time t, and r t represents the reward function, including time - efficiency reward, energy - consumption penalty, and collision - risk penalty.

[0026] A method for using an intelligent storage rack, which includes:

[0027] Obtain the characteristic information of the goods to be stored and the historical access data of the storage positions;

[0028] Obtain the popularity value of each storage position according to the historical access data, and construct an access popularity map based on the popularity values of multiple storage positions;

[0029] Obtain the comprehensive goods popularity of the goods to be stored according to the historical access data and the characteristic information;

[0030] Allocate the optimal storage position according to the access popularity map and the comprehensive goods popularity.

[0031] Preferably, the specific method for constructing the access popularity map includes:

[0032] Obtain the three - dimensional Euclidean distance d i between the storage position and each access position in the historical access events;

[0033] Obtain the contribution degree of each historical access event to the storage position according to the three - dimensional Euclidean distance

[0034] The contribution degrees are weighted and superimposed to obtain the heat value, and an access heat map is constructed based on the heat value;

[0035] wherein, represents the heat value of the three-dimensional coordinate point (x, y, z) of the storage location, N represents the total number of historical access events, σ represents the diffusion coefficient of the Gaussian kernel, μ represents the distance offset, and e represents the natural constant.

[0036] Preferably, the specific method for obtaining the comprehensive heat of the goods to be stored includes:

[0037] Obtain the comprehensive access frequency F of the goods to be stored Com_i , the recent access frequency F Recent_i and the associated goods heat R i ;

[0038] Obtain the comprehensive heat H of the goods according to the comprehensive access frequency, the recent access frequency and the associated goods heat i = α1·F Com_i + α2·F Recent_i + α3·R i .

[0039] wherein, α1, α2, and α3 respectively represent the weight coefficients of the comprehensive access frequency, the recent access frequency, and the associated goods heat.

[0040] Preferably, the method for using the intelligent storage rack further includes:

[0041] After receiving the sorting order, obtain the picking sequence and plan the path according to the Q function , and generate a control instruction according to the picking sequence and the planned path;

[0042] The access manipulator responds to the control instruction to perform order sorting;

[0043] wherein, s represents the current system state, s t represents the operating state of the system at time t, s represents the currently selected action, s t represents the specific action of the access manipulator at time t, γ represents the discount factor, r t+k represents the immediate reward of the system at time t + k, E represents the expectation operator, k represents the future time step index starting from the current time t, r t represents the reward function, including time efficiency reward, energy consumption penalty, and collision risk penalty. Description of the Drawings

[0044] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but the emphasis is on showing the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0045] Figure 1 is a schematic diagram of the overall structure of the intelligent storage rack in an embodiment of the present invention;

[0046] Figure 2 is a schematic diagram of the overall process of the method for using the intelligent storage rack in an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of the process of the specific method for constructing the access heat map in an embodiment of the present invention;

[0048] Figure 4 is a schematic diagram of the process of the specific method for obtaining the comprehensive heat of the goods to be stored in an embodiment of the present invention.

[0049] Figure 5 is a schematic diagram of the process of the method for using the intelligent storage rack in another embodiment of the present invention Figure 1 ;

[0050] Figure 6 is a schematic diagram of the process of the method for using the intelligent storage rack in another embodiment of the present invention Figure 2 ;

[0051] Figure 7 is a schematic diagram of the structure of the intelligent storage rack in an embodiment of the present invention;

[0052] Figure 8 is a schematic diagram of the functional modules of the controller in an embodiment of the present invention Figure 1 ;

[0053] Figure 9 is a schematic diagram of the functional modules of the controller in an embodiment of the present invention Figure 2 .

[0054] Explanation of reference numerals: 1. Storage rack body; 2. Controller. Detailed implementation manners

[0055] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with its embodiments. It should be understood that the specific implementation manners described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0056] It should be noted that when an element is referred to as "fixed to" another element, it can be directly on the other element or there may also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation.

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this invention belongs. The terms used in the specification of this invention are only for the purpose of describing specific embodiments and are not intended to limit the invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0058] The "first" and "second" mentioned in this invention do not represent specific quantities and sequences, but are only used for name distinction.

[0059] Before specifically elaborating on the embodiments of the present invention, a brief introduction to the prior art is given first.

[0060] The allocation of storage positions on a storage rack is related to the efficiency of goods in and out of the warehouse, or in other words, the storage and retrieval efficiency. By optimizing the allocation of storage positions on the storage rack, the efficiency of goods storage and retrieval can often be improved, and the comprehensive operation cost can be reduced. In the prior art, the comprehensive popularity of goods often depends on the experience of technicians or managers to judge, so it is difficult to effectively ensure the accuracy of the comprehensive popularity of goods. If the allocation of goods storage positions is realized according to the comprehensively judged popularity of goods by humans, it will not only reduce the intelligent level of the storage rack, but also reduce the efficiency of goods storage and retrieval.

[0061] Prior arts such as "An intelligent three-dimensional storage rack storage position allocation system and method" with the application number CN202211552812.8, "A storage position allocation method for a mobile rack storage system for cold storage" with the application number CN201911138483.0, "An optimization method for storage position allocation in a three-dimensional warehouse for non-shaped racks" with the application number CN202211594440.5, and "A storage rack allocation method, device, electronic device and readable storage medium" with the application number CN202010568008.3 all involve the allocation of storage positions on the storage rack. Although these storage position allocation methods for storage racks can improve the speed of goods in and out of the warehouse, they do not consider the integration of the popularity of storage positions on the storage rack and the comprehensive popularity of goods to be stored when allocating storage positions, and there is still room for further improvement in their intelligent level and warehousing efficiency.

[0062] In order to improve the efficiency of goods storage and retrieval and the intelligent level of the storage rack, an embodiment of the present invention provides an intelligent storage rack, such asFigure 1 and Figure 7 As shown in Figure 7 , it includes a storage rack body 1, a controller 2, and at least one feature acquisition module for acquiring the feature information of the goods to be stored. A plurality of storage positions are arranged on the storage rack body. The controller is used to store the historical access data of the storage positions. The feature acquisition module is connected to the controller and feeds back the feature information to the controller.

[0063] The storage rack body is placed in a warehouse or a goods sorting workshop. The storage rack body can be a multi-layer parallel shelf, and a plurality of storage positions for storing goods are arranged on each layer of the shelf. The controller includes but is not limited to a server, a processing center, and a client. It is signal-connected to the feature acquisition module through a cable or a wireless communication module, and is used to receive the feature information of the goods to be stored acquired by the feature acquisition module, and analyze the feature information to judge the type of the goods to be stored.

[0064] The feature acquisition module includes but is not limited to an RGB-D camera (for example, with a depth resolution of 1280×720), a weight sensor (for example, with a measuring range of 0-50 kg and an accuracy of ±0.1 g), and an RFID scanning module (for example, with an identification distance of 1.5 m). Correspondingly, the feature information includes but is not limited to image information, weight information, and label information. Through the RGB-D camera, the weight sensor, and the RFID scanning module, a joint perception network of three-dimensional space-weight-identification can be formed, so as to facilitate the controller to identify the type of the goods to be stored according to the feature information, and to obtain the access data of the storage positions according to the feature information and generate historical access data.

[0065] The RGB-D camera, the weight sensor, and the RFID scanning module can be installed at appropriate positions (such as the entrance and exit) on the storage rack body or even in the warehouse or the goods sorting workshop according to actual needs.

[0066] Preferably, the feature acquisition module covers the static storage area, the dynamic operation area, and the interaction interface area of the storage rack. The specific installation scheme is as follows:

[0067] I. The central area at the top of the storage rack

[0068] 1. The RGB-D main camera is fixedly installed at the top beam of the storage rack body, tilting downward at a depression angle of 15°. A plurality of cameras are provided and evenly distributed along the length direction of the beam. It is used for global monitoring of the storage state of the shelf, generating a 1280×720 depth point cloud (accuracy ±2 mm), and eliminating blind spots (such as occlusion at the edge of a high-level shelf) through multi-view stitching.

[0069] Using the RGB-D camera, real-time obstacle avoidance for the path planning of the robotic arm can be supported, such as detecting sudden offset of goods.

[0070] 2. The RFID scanning module is coaxially installed with the RGB-D camera, and the antenna faces the entrance direction of the storage rack. It is used to identify the unique ID of the tagged goods within the range. By fusing with the visual data, it can solve the confusion problem of goods with similar appearances.

[0071] By setting the RGB-D main camera + RFID scanning module in the center of the storage rack, the storage rack can be globally monitored to solve the problem of "seeing but not accurately identifying".

[0072] II. Shelf Layered Structure Area

[0073] 1. The embedded node weight sensor is installed at the bottom of the support column of each layer of the shelf and adopts a four-point weighing structure. It is used to real-time monitor the weight change of goods from 0 to 50 kg (with an accuracy of less than 0.1%), and detect abnormal access, such as partial picking without resetting, etc.

[0074] 2. The layered RGB-D edge camera is installed 5 cm below the front edge of each layer of the shelf, and the horizontal view covers the storage cells of this layer. It is used to assist the RGB-D main camera to complete high-precision positioning (such as millimeter-level coordinate calibration of small parts), and independently record the operation logs of each layer.

[0075] By setting the embedded node weight sensor + layered RGB-D edge camera in the shelf layered structure area, the goods can be accurately positioned and weight verified to prevent misplacement / missing picking.

[0076] III. Entrance / Exit of the Area where the Storage Rack is Located

[0077] 1. A channel-type weight detection platform composed of weight sensors is embedded in the ground at the entrance and exit of the area where the storage rack is located, with a size of 80 cm × 120 cm. It is used to verify the total weight of the incoming and outgoing goods (compare with the data of the embedded node weight sensor on the shelf layer, and an alarm will be triggered if the error is greater than the preset weight value).

[0078] 2. An RFID gate array composed of RFID scanning modules sets the entrance and exit into an arch structure, and 3 groups of directional antennas are deployed on each side of the arch. It is used to batch read all the goods tags in the goods turnover box.

[0079] By setting the channel-type weight detection platform + RFID gate array at the entrance and exit of the area where the storage rack is located, double verification of incoming and outgoing can be achieved.

[0080] The controller is also used to obtain the heat value of each storage location according to the historical access data, and construct an access heat map based on the heat values of multiple storage locations; obtain the comprehensive heat of the goods to be stored according to the historical access data and the feature information; and allocate the optimal storage location according to the access heat map and the comprehensive heat of the goods.

[0081] Preferably, as Figure 8 shown, the controller includes a distance acquisition module, a contribution degree acquisition module, and a heat map construction module.

[0082] The distance acquisition module is used to acquire the three-dimensional Euclidean distance between the storage location and each access location in the historical access events. Specifically, the three-dimensional Euclidean distance characterizes spatial proximity, which represents the three-dimensional Euclidean distance between the three-dimensional coordinate point (x, y, z) of the storage location and the three-dimensional coordinate point of the i-th access location, and can be labeled as (x i , y i , z i ) represents the three-dimensional coordinate point of the i-th access location.

[0083] The contribution degree acquisition module is used to acquire the contribution degree of each historical access event to the storage location according to the three-dimensional Euclidean distance. Specifically, the heat contribution of each historical access event to the surrounding space follows a normal distribution, and the contribution degree of the i-th historical access event to the storage location can be expressed as It can be understood as the contribution degree of the i-th historical access event at the three-dimensional coordinate point (x, y, z) of the storage location.

[0084] The heat map construction module is used to perform weighted superposition on the contribution degrees to obtain the heat value, and construct an access heat map according to the heat value. Here, the heat value is obtained according to the formula and is used to reflect the access frequency or demand intensity of the three-dimensional coordinate point (x, y, z) of the storage location.

[0085] Preferably, the heat map construction module constructs an access heat map according to the formula where H(x, y, z) represents the heat value of the three-dimensional coordinate point (x, y, z) of the storage location; N represents the total number of historical access events, reflecting the time span and sample size of historical access data points; σ represents the diffusion coefficient of the Gaussian kernel, used to control the influence range of a single access event on the surrounding area; μ represents the distance offset, used to correct the sensitivity of the model to access events (usually set to 0, indicating centered on the event location); and e represents the natural constant.

[0086] Specifically, by obtaining the heat value of each storage location and using professional data analysis tools (such as Matplotlib and Seaborn libraries in Python), a heat map is drawn based on the heat value of each storage location. In the heat map, different colors can represent different heat levels, and the darker the color, the higher the heat. According to the heat map, a heat threshold is set, and the area with a heat value higher than the heat threshold is determined as the high-frequency area (usually the area where goods are most frequently in and out of storage).

[0087] The larger the diffusion coefficient σ of the Gaussian kernel, the wider the influence range of a single access event, but the peak intensity decreases; the smaller the diffusion coefficient σ of the Gaussian kernel, the more concentrated the influence range and the more prominent the local peak. For the distance offset μ, if μ ≠ 0, the directional offset of the access behavior can be simulated (for example, the goods closer to the aisle are more likely to be accessed frequently).

[0088] The contributions of all historical access events can be superimposed to form a continuous heat distribution surface. The H value in the high-frequency access area (such as the entrance of the shelf) is significantly higher than that in the low-frequency area (such as the corner). Based on the access heat map, warehouse optimization can be guided to allocate the optimal storage location for the goods to be stored. For example: 1. Place the goods with high demand and high heat (large H value) in the peak area of H(x, y, z) to shorten the picking path; 2. Identify the local extreme points of the H value, divide the "hot areas" and "cold areas" for storing goods, and optimize the shelf partition design; 3. Perform dynamic parameter adjustment, adjust the diffusion coefficient σ according to the turnover rate of the goods. For example, for seasonal goods, increase the diffusion coefficient σ during the promotion period to expand the stocking range.

[0089] By weighting the contributions of historical access events through the Gaussian distribution function, the access demand intensity of spatial positions can be dynamically quantified; by analyzing the heat distribution during the peak order period, the location demand of popular goods can be predicted to achieve automatic replenishment.

[0090] Preferably, as Figure 9 shown, the controller further includes an access frequency acquisition module, a related heat acquisition module, and a comprehensive heat acquisition module.

[0091] The access frequency acquisition module is used to acquire the comprehensive access frequency and the recent access frequency of the goods to be stored. Specifically, the comprehensive storage frequency The recent access frequency where C i ' represents the total number of accesses of the goods to be stored i within the statistical period D'. It can be statistically obtained through the RFID scanning module or the RGB-D camera image recognition log. The statistical period D' can be understood as the total number of days of system operation or the time span of the goods being put on the shelf. For example, if the goods are on the shelf for 30 days, then the statistical period D' = 30. C i ” represents the number of accesses of the goods to be stored i in the recent D”. The recent D” is a preset time window (such as the most recent 7 days). By calculating the recent access frequency, the problem of heat lag of long-tail goods can be solved.

[0092] The related heat acquisition module is used to acquire the related goods heat. Specifically, the related goods heat

[0093] where M represents the number of related goods j, C ijDenotes the co-occurrence access times of goods i and j, which can be understood as the total number of times goods i and j are accessed and stored simultaneously within the statistical period (e.g., when the user accesses and stores two goods in one operation). The co-occurrence frequency can be statistically obtained through the user behavior log (such as the record of the RFID scanning module), C ij Used to quantify the association strength between two goods. The higher the co-occurrence times, the stronger the association. For example, the high co-occurrence times between a screwdriver and screws reflect a functional association.

[0094] C total Denotes the total access times of goods i, which can be understood as the total independent access times of goods i within the statistical period. It can be obtained by directly counting the access records of goods i in the log, C total Used to normalize the co-occurrence frequency and avoid the excessive influence of high-frequency goods on the association heat. For example, if goods i are frequently accessed and stored, their accidental co-occurrence with multiple goods will be weakened.

[0095] e -λΔt Denotes the time decay factor, which can be understood as measuring the time decay effect of co-occurrence events based on Newton's law of cooling. λ represents the decay coefficient (which can be set by technicians according to experience), and Δt represents the time difference between the current time and the most recent co-occurrence (calculated in days). The larger the value of λ, the faster the decay rate (e.g., when λ = 0.01, the co-occurrence weight 30 days ago drops to e -0.3 ≈0.74), and the smaller the Δt (recent co-occurrence), the higher its weight. The time decay factor e -λΔt Is used to emphasize the recent association and can solve the problem of outdated historical co-occurrence data. For example, the association of seasonal goods changes significantly over time.

[0096] Sim(i, j) represents the semantic similarity weight, which can be understood as the semantic association strength between goods i and j. It is calculated based on text descriptions or knowledge graphs at the second degree. For example, based on the word embedding model Word2Vec, vectors are generated from the text descriptions of goods, and the cosine similarity is calculated. Or based on the knowledge graph, if goods belong to the same category or have an upstream-downstream relationship (such as "coffee cup" and "coffee machine"), a higher weight is assigned. The role of Sim(i, j) is to supplement the deficiencies of behavioral co-occurrence data and identify implicit associations. For example, even if the co-occurrence times are few, "battery" and "remote control" may be associated due to semantic association.

[0097] Specifically, if calculating the semantic similarity weight based on the cosine similarity of word vectors, then v i 、v j Represent the word vectors of goods i and j respectively, which can be generated through models such as Word2Vec and BERT. ||·|| represents the vector norm. If calculating the semantic similarity weight based on the path-weighted similarity of the knowledge graph, then K represents the number of the shortest paths between goods i and j in the knowledge graph. Relscore(p k ) and Length(p k ) respectively represent the weight of the relational edge and the path length (number of nodes) in path p k .

[0098] H j represents the independent popularity of good j. When the popularity of the associated good itself is higher, its contribution to R i is greater. For example, if the associated good j is a popular product, even if the co-occurrence times are average, its high popularity will increase the value of R i . For the independent popularity of good j, it can be calculated according to the formula where C j ' represents the total access times of good j within the statistical period D'.

[0099] The independent popularity of the said good j where S j , D j , P j respectively represent the search popularity, discussion popularity and propagation popularity of good j, and w1, w2, w3 respectively represent the weight coefficients of the search popularity, discussion popularity and propagation popularity of good j, which can be set by technical personnel. Preferably, the LSTM neural network is used to predict the user behavior trend and dynamically adjust the weight coefficients w1, w2, w3, that is, based on time series prediction, the weight coefficients are dynamically allocated.

[0100] The search popularity can be calculated based on the search volume on platforms such as Baidu Index and Google Trends. Preferably, a regional weighting coefficient (which can be set by technical personnel or managers according to experience) is introduced. By obtaining the platform search volume of the keywords of good j in different regions and combining the regional platform search volume with the regional weighting coefficient, the weighted value is calculated to distinguish the search popularity differences in different regions.

[0101] The discussion popularity can be calculated based on the number of social media comments and forum posts, combined with the user influence weight (such as the comment weight of Weibo big V is higher). Through user influence weighting, the interference of low-quality discussion brushing can be avoided. The propagation popularity can be calculated according to the number of forwards and shares, combined with the propagation path depth (such as the third-level propagation weight in the circle of friends is higher). By introducing the propagation depth factor, shallow propagation and deep propagation are distinguished.

[0102] e θt represents the search popularity time enhancement factor, θ>0 represents the time enhancement coefficient, t represents the time variable, indicating that the recent search volume shows exponential growth, which is used to distinguish short-term bursts and long-term stable popularity; represents the S-shaped curve attenuation factor of the discussion heat (T is the time of the heat peak), indicating that the discussion heat first rises and then decays over time (such as the cooling period of a hot event), and is used to fit the actual discussion heat decay curve. k' represents the steepness parameter of the S-shaped function, which is used to control the transition speed from low to high weights (for example, when k' = 0.1, the transition is smooth; when k' = 1, the switch is rapid); lg(1 + τt) represents the logarithmic growth factor of the propagation heat (τ > 0), indicating that the propagation volume slowly grows over time (such as long-tail content), and is used to avoid the distortion of the heat value caused by too large a propagation volume. τ represents the logarithmic growth scaling coefficient, which is used to adjust the amplitude of the potential value growing over time; Q j represents the interaction quality factor (range: 0 - 1). The interaction types include collection, like, comment, and share, and it can be obtained by weighting the interaction types. For example, collection = 0.4, like = 0.3, comment = 0.2, share = 0.1. Of course, the values of different interaction types can also be defined according to the actual situation, and it is used to finely distinguish the value of interaction behaviors; U j represents the user profile factor (range: 0 - 1), which can be understood as the stratification of user value (such as VIP users having a higher interaction weight), and is used to improve the accuracy of heat calculation by combining the user profile; γ' and η respectively represent the interaction and profile weights, which are determined through A / B testing (such as in the e-commerce scenario, γ' and η are set to 0.6 and 0.4 respectively), and are used to dynamically balance the influence of user behavior and profile; ξ represents the outlier filtering coefficient (range: 0 - 1), which is used to filter robot behaviors based on the chi-square test (such as a large number of repeated interactions in a short period of time), and excludes the interference of abnormal data.

[0103] The independent heat formula of the goods j has the following advantages:

[0104] 1. Multi-dimensional dynamic weight allocation. By using LSTM to predict the user behavior trend, the weights of search, discussion, and propagation heat are automatically adjusted. For example, the search weight is increased during the new product launch period, and the propagation weight is increased during the long-tail content period.

[0105] 2. Enhanced time sensitivity. The search heat uses an exponential growth factor to capture short-term bursts; the discussion heat simulates the heat life cycle through an S-shaped curve; the propagation heat uses a logarithmic function to smooth the long-tail effect.

[0106] 3. Fine-grained modeling of user behavior. The interaction quality factor distinguishes the value of behaviors such as collection and like, and the user profile factor enhances the contribution of high-value users. For example, in the case of mother and baby products, the interaction weight of the mother group is higher.

[0107] 4. Anti-noise ability. The outlier filtering coefficient ξ excludes robot brushing based on the chi-square test to ensure the authenticity of the heat value.

[0108] The comprehensive heat acquisition module is used to obtain the comprehensive heat of goods according to the comprehensive access frequency, recent access frequency, and associated goods heat. Specifically, the comprehensive heat acquisition module calculates the comprehensive heat of goods according to the formula H i = α1·F Com_i + α2·F Recent_i + α3·R i where F Com_i and F Recent_i and R i represent the comprehensive access frequency, recent access frequency, and associated goods heat respectively, and α1, α2, and α3 represent the weight coefficients of the comprehensive access frequency, recent access frequency, and associated goods heat respectively.

[0109] The associated goods heat strengthens the influence of strongly associated goods by integrating co-occurrence frequency, time decay, and semantic similarity, avoiding the equalization of the influence of high-heat associated goods. By combining behavior data C ij , time decay e -λΔt and semantic information Sim(i,j), the accuracy of the associated heat is improved. Based on the associated goods heat, the finally obtained comprehensive heat of goods is more in line with the actual situation and has higher accuracy.

[0110] Regarding the allocation of the optimal storage location according to the access heat map and the comprehensive heat of goods, the specific method includes: pre-constructing the mapping relationship between the comprehensive heat of goods and the access heat, and allocating the corresponding storage location for the goods to be stored according to the constructed mapping relationship.

[0111] Specifically, the storage locations can be divided into regions according to the access heat map. The storage locations in different regions correspond to different comprehensive heats of goods. According to the comprehensive heat of the goods to be stored, the corresponding storage location is allocated.

[0112] By improving the accuracy of the comprehensive heat of goods and the heat of storage locations, and integrating the heat of storage locations with the comprehensive heat of the goods to be stored to consider the goods allocation, it is beneficial to optimize the storage location allocation of goods by the intelligent storage rack, and further optimize the goods access path.

[0113] That is to say, the intelligent storage rack can optimize the goods access path, shorten the operation time, improve the space utilization rate and reduce the comprehensive cost by constructing a heat map and obtaining the comprehensive heat of goods, and allocating the optimal storage location according to the heat map and the comprehensive heat of goods, which is beneficial to improving the intelligence level of the storage rack.

[0114] As a preferred technical solution, for the outlier filtering coefficient ξ, its core objective is to identify and exclude abnormal behaviors such as robot traffic brushing through statistical tests, and specifically includes the following steps:

[0115] Step 1: Data collection. Since robot traffic usually manifests as high-frequency, repetitive, or abnormal interaction patterns within a short period, the interaction behavior data (such as like, comment, share timestamps, user IDs, IP addresses, device fingerprints, etc.) related to goods i and j can be collected for further analysis to determine whether the interaction data is abnormal.

[0116] Step 2: Feature construction. Extract anomaly detection features, including but not limited to the interaction frequency per unit time (such as a user liking multiple times within 1 minute), the deviation of the active period of the device / account from the user profile (such as high-frequency interaction late at night), and the contradiction between geographical location and device information (such as multiple devices interacting simultaneously with the same IP).

[0117] Step 3: Chi-square test implementation. First, make a hypothesis. Since the chi-square test detects anomalies by comparing the differences between observed frequencies and expected frequencies, normal interaction behaviors conform to the Poisson distribution or the normal distribution, while abnormal behaviors deviate from this distribution. Immediately afterwards, perform binning. After discretizing continuous data to apply the chi-square test, group the interaction behaviors by time or user dimension (such as the number of interactions per hour). Finally, calculate the chi-square value and perform significance judgment. If the chi-square value is greater than the critical value, it is judged that there is an anomaly, where O i and E i represent the observed frequency and the expected frequency respectively.

[0118] Step 4: Dynamically adjust the value of the outlier filtering coefficient according to the anomaly ratio. If the detected proportion of abnormal interactions ≤ 5%, the outlier filtering coefficient ξ = 1; if the proportion of abnormal interactions is 5% - 15%, the outlier filtering coefficient ξ = 1 - 2 × the proportion of abnormal interactions (for example, if the proportion of abnormal interactions is 10%, then the outlier filtering coefficient ξ = 1 - 2 * 10% = 0.8); if the proportion of abnormal interactions > 15%, the outlier filtering coefficient ξ = 0.5 to forcibly filter high-risk data. Of course, according to the actual situation and needs, the outlier filtering coefficients corresponding to different proportions of abnormal interactions can be dynamically adjusted.

[0119] Based on the outlier filtering coefficient, the credibility weight of the independent popularity of item j can be dynamically adjusted. Its effects on the independent popularity formula include: 1. Enhanced anti-noise ability. By excluding robot brushing (such as a large number of repeated likes in a short period), it avoids the false interaction from artificially inflating the popularity value. 2. Enhanced authenticity of the popularity value. The introduction of the outlier filtering coefficient makes the independent popularity formula of item j closer to real user behavior. For example, in the category of mother and baby products, if the proportion of interactions of VIP users (high-value users) is diluted by abnormal behavior, ξ will reduce the overall weight to ensure that the behavior of high-value users dominates the popularity calculation. 3. Dynamically adaptive optimization. The value of ξ changes with the abnormal proportion, enabling the independent popularity formula to have an adaptive ability. For example, during major e-commerce promotions, the normal interaction frequency may increase briefly. At this time, the chi-square test can adjust the expected frequency in combination with historical peaks to avoid misjudgment. 4. Multi-dimensional collaborative filtering. ξ forms a complement with the interaction quality factor and the user portrait factor. The interaction quality factor is used to distinguish the value of behavior (such as collection > like), the user portrait factor is used to weight the user value (such as higher VIP weight), and ξ excludes systematic anomalies from a statistical level. The three jointly improve the accuracy of the independent popularity of item j.

[0120] The outlier filtering coefficient dynamically identifies and suppresses robot brushing through the chi-square test, significantly enhancing the anti-interference ability and authenticity of the independent popularity value formula. Its synergistic effect with the interaction quality factor and the user portrait factor makes the independent popularity value of item j more accurately reflect real user behavior and is applicable to high-noise scenarios such as e-commerce recommendations.

[0121] To improve the intelligence level of item access in the intelligent storage rack, as a preferred technical solution, the intelligent storage rack further includes a storage and retrieval robotic arm signal-connected to the controller. The storage and retrieval robotic arm is used to respond to control instructions for order sorting. The storage and retrieval robotic arm can be installed on an AGV navigation cart, and the AGV navigation cart is signal-connected to the controller and is used to respond to control instructions and move along the planned path. After moving to the designated position, the storage and retrieval robotic arm picks up items according to the picking sequence.

[0122] Specifically, the controller further includes a path planning module.

[0123] The path planning module is used to, after receiving the sorting order, obtain the picking sequence and the planned path according to the Q function and generate control instructions according to the picking sequence and the planned path.

[0124] Among them, s represents the current system state, s t represents the operating state of the system at time t, a represents the currently selected action, a t represents the specific action of the storage and retrieval robotic arm at time t, γ represents the discount factor, r t+k$R_{t + k}$ represents the immediate reward of the system at time $t + k$, $E$ represents the expectation operator, $k$ represents the future time step index starting from the current time $t$, and its value range is $k = 0, 1, 2, \cdots$, corresponding to the reward sequence $r_{t + 1}$, $r_{t + 2}$, $r_{t + 3}$, $\cdots$, $r_{t + k}$. t , $r_{t + 2}$ t+1 , $r_{t + 3}$ t+2 …, $r_{t + k}$ t $R$ represents the reward function, including time efficiency reward, energy consumption penalty, and collision risk penalty. $\gamma$ k represents the decay weight of the reward at the $k$-th future step. In the Q function, by accumulating the discounted rewards of all time steps, it can comprehensively evaluate the long-term benefits of action $a$ in state $s$.

[0125] Specifically, the operating state of the system at time $t$ includes the warehouse layout (such as the position of the storage rack, the width of the aisle, etc.), the real-time position and remaining power of the access manipulator / AGV navigation vehicle, the priority of the orders to be sorted, the position of the storage space on the target storage rack, and other dynamic factors (such as the paths of other access manipulators, the distribution of obstacles). The specific actions of the access manipulator at time $t$ include the moving direction (forward, turning, avoidance path planning) and the grasping / releasing operation of the goods. The discount factor $\gamma\in[0, 1]$ is used to balance the importance of the current reward and the future reward. When the discount factor $\gamma$ is 0, it means only focusing on the immediate reward and ignoring all subsequent rewards. When the discount factor $\gamma\rightarrow1$, more importance is attached to the long-term cumulative benefits (such as global path optimization), and the Q function will assign almost the same weight to the future reward and the current reward.

[0126] The time efficiency represents the number of orders completed per unit time or the sorting action efficiency. The time efficiency reward includes positive rewards and negative rewards. For positive rewards, it can be understood that the reward value is increased (e.g., +1 or +0.5) when each order sorting is completed or the sorting cycle is shortened. For negative rewards, it can be understood that the reward is deducted (e.g., -0.5 or -1) when the sorting task is not completed on time. The time efficiency reward is used to reduce the ineffective moving time by optimizing the path and improve the order throughput.

[0127] The energy consumption mentioned above includes the energy consumption of actions such as the movement of the AGV navigation vehicle and the grasping of the access manipulator. The energy consumption penalty includes negative penalties, which can be understood as being based on the moving distance (e.g., per meter, the penalty value is defined as -1 or -2), the acceleration frequency (e.g., each rapid acceleration, the penalty value is defined as -0.5 or -1). The energy consumption penalty is used to reduce the repeated paths and lower the overall energy consumption through a hierarchical strategy.

[0128] The collision risk represents the probability of the access manipulator colliding with other devices, personnel, or obstacles. The collision risk penalty includes a negative penalty (such as real-time detection of surrounding dynamic obstacles, calculating the penalty value according to distance and speed, and when the risk level is high, the negative penalty value is defined as -1.5 or -2) and a safety reward (gradually increasing the positive reward during continuous collision-free operation, such as every 10 minutes, the safety reward +0.3 or 0.6).

[0129] Obtaining the picking sequence and planning path according to the Q function specifically includes: maximizing the expected cumulative reward during the sorting process, that is, finding the optimal policy π * such that π * = argmax π Q(s,a); by adjusting the weight coefficients of each index in the reward function, the policy bias under different business requirements is realized. In the scenario of high-priority orders, the weight of the time efficiency reward is increased. In the energy-saving mode, the weight of the energy consumption penalty is increased. When facing a complex environment, the weight of the collision risk penalty is strengthened.

[0130] In this embodiment, obtaining the picking sequence and planning path according to the Q function, and balancing the sorting efficiency (time), energy consumption, and collision risk through the reward function can reduce the collision risk. The shelf body, the feature access module, the controller, and the access robot are combined, and the optimal storage location is allocated according to the collocation access heat map and the comprehensive heat of the goods, which is not only beneficial to optimizing the storage location allocation of the intelligent shelf for goods, and then optimizing the goods access path. On this basis, the space utilization rate can be further improved and the comprehensive cost can be reduced, and the intelligence level and warehousing efficiency of the shelf can be further improved.

[0131] An embodiment of the present invention also provides a method for using an intelligent shelf, as Figure 2 shown, which includes:

[0132] S1, obtaining the feature information of the goods to be stored and the historical access data of the storage location.

[0133] The feature information includes but is not limited to the image information, weight information, and label information of the goods to be stored. The type of the goods to be stored is identified according to the feature information. The access data of the storage location can be obtained according to the feature information first and processed and stored by a controller such as a client, a processing center, or a server, etc., to generate the historical access data. Specifically, each historical access event corresponds to a piece of historical access data.

[0134] S2, obtaining the heat value of each storage location according to the historical access data, and constructing an access heat map according to the heat values of multiple storage locations.

[0135] Preferably, in step S2, as Figure 3As shown, the specific method for constructing the access heat map includes:

[0136] S21, obtaining the three-dimensional Euclidean distance d between the storage location and each access location in the historical access events i .

[0137] S22, obtaining the contribution degree of each historical access event to the storage location according to the three-dimensional Euclidean distance

[0138] S23, performing weighted superposition on the contribution degrees to obtain the heat value, and constructing the access heat map according to the heat value.

[0139] Among them, represents the heat value of the three-dimensional coordinate point (x, y, z) of the storage location, N represents the total number of historical access events, σ represents the diffusion coefficient of the Gaussian kernel, μ represents the distance offset, and e represents the natural constant.

[0140] Specifically, the three-dimensional Euclidean distance characterizes spatial proximity, which represents the three-dimensional Euclidean distance between the three-dimensional coordinate point (x, y, z) of the storage location and the access location of the i-th time, and can be marked as (xi, yi, zi) represents the three-dimensional coordinate point of the access location of the i-th time.

[0141] For the heat map, by obtaining the heat value of each storage location and using professional data analysis tools (such as Matplotlib and Seaborn libraries in Python), the heat map is drawn based on the heat value of each storage location. In the heat map, different colors can represent different heat levels, and the darker the color, the higher the heat. According to the heat map, a heat threshold is set, and the area with a heat value higher than the heat threshold is determined as the high-frequency area (usually the area where goods are most frequently in and out of storage)

[0142] S3, obtaining the comprehensive heat of the goods to be stored according to the historical access data and characteristic information.

[0143] Preferably, the comprehensive heat of the goods i to be stored Among them, C i ' represents the total number of accesses of the goods i to be stored within the statistical period D', can be understood as the access density or heat value of the goods i to be stored within the statistical period D', C j ' represents the total number of accesses of the associated goods j within the corresponding statistical period D j ', can be understood as the access density or heat value of the associated goods j within the statistical period D j ', M represents the number of associated goods j, It represents the average association heat of the associated goods j of the goods i to be stored, and β1 and β2 respectively represent the weight coefficients of the goods i to be stored and the average association heat, which can be set by technicians.

[0144] Specifically, the other goods j that have access associations with the goods i to be stored, that is, the associated goods, can be mined based on user behavior data (such as co-existing access, semantic association, etc.).

[0145] S4. Allocate the optimal storage location according to the access heat map and the comprehensive heat of the goods. Specifically, the storage locations can be divided into regions according to the access heat map. The storage locations in different regions correspond to different comprehensive heats of the goods, and the corresponding storage locations are allocated according to the comprehensive heat of the goods to be stored.

[0146] To sum up, the method for using the intelligent storage rack can optimize the access path of the goods, shorten the operation time, improve the space utilization rate and reduce the comprehensive cost by constructing a heat map and obtaining the comprehensive heat of the goods, and allocating the optimal storage location according to the heat map and the comprehensive heat of the goods, which is beneficial to improving the intelligence level of the storage rack.

[0147] In the comprehensive heat of the goods i to be stored In the formula, there are two deficiencies in the average association heat of the associated goods j of the goods i to be stored. First, the influence of high-heat associated goods is evenly distributed, and the heat value of strong associated goods is easily weakened by taking the average value. Second, the direct co-existing access and weak semantic association are not distinguished, ignoring the difference in association strength.

[0148] In order to optimize the comprehensive heat of the goods formula, as a preferred technical solution, as Figure 4 shown, in step S3, the specific method for obtaining the comprehensive heat of the goods to be stored includes:

[0149] S31. Obtain the comprehensive access frequency F Com_i of the goods to be stored, the recent access frequency F Recent_i and the associated goods heat R i ;

[0150] S32. Obtain the comprehensive heat H i of the goods according to the comprehensive access frequency, the recent access frequency and the associated goods heat Com_i = α1·F Recent_i + α2·F i + α3·R

[0151] Among them, α1, α2, and α3 respectively represent the weight coefficients of the comprehensive access frequency, the recent access frequency, and the associated goods heat, and can all be set by technicians.

[0152] The comprehensive storage frequency The recent access frequency where C i ' represents the total number of accesses of the goods i to be stored within the statistical period D'. It can be counted through the RFID scanning module or the RGB-D camera image recognition log. The statistical period D' can be understood as the total number of days of system operation or the time span of goods shelving. For example, if the goods are shelved for 30 days, then the statistical period D' = 30. C i ” represents the number of accesses of the goods i to be stored in the recent D”. The recent D” is a preset time window (such as the most recent 7 days). By calculating the recent access frequency, the problem of lagging heat of long-tail goods can be solved.

[0153] Associated goods heat

[0154] where M represents the number of associated goods j, C ij represents the co-occurrence access times of goods i and j, C total represents the total number of accesses of goods i, e -λΔt represents the time decay factor, λ represents the decay coefficient (which can be set by technicians according to experience), Δt represents the time difference between the current time and the most recent co-occurrence (calculated in days), Sim(i,j) represents the semantic similarity weight, H j represents the independent heat of goods j. When the heat of the associated goods themselves is higher, the contribution to R i is greater. For example, if the associated goods j are popular goods, even if the co-occurrence times are average, their high heat will increase the R i value. For the independent heat of goods j, it can be calculated according to the formula where C j ' represents the total number of accesses of goods j within the statistical period D'.

[0155] The improved comprehensive heat of goods can be realized as follows: through the triple weights of co-occurrence frequency, time decay, and semantic similarity, the influence of strongly associated goods is strengthened, and the equal-sharing effect is avoided; combined with the behavior data C ij , time decay e -λΔt and semantic information Sim(i,j), the accuracy of the associated heat is improved.

[0156] Based on the associated goods heat, the finally obtained comprehensive heat of goods is more in line with the actual situation and has higher accuracy.

[0157] That is to say, by improving the accuracy of the comprehensive heat of goods and the heat of storage positions, and integrating the heat of storage positions with the comprehensive heat of goods to be stored to consider the goods allocation, it is beneficial to optimize the storage position allocation of goods by the intelligent storage rack, and further optimize the goods access path.

[0158] As a preferred technical solution, as Figure 5 shown, the method for using the intelligent storage rack further includes:

[0159] S5. After receiving the sorting order, obtain the picking sequence and plan the path according to the Q function and generate a control instruction according to the picking sequence and the planned path;

[0160] S6. The access manipulator responds to the control instruction and performs order sorting;

[0161] where s represents the current system state, s t represents the operating state of the system at time t, s represents the currently selected action, s t represents the specific action of the access manipulator at time t, γ represents the discount factor, r t+k represents the immediate reward of the system at time t + k, E represents the expectation operator, k represents the future time step index starting from the current time t, r t represents the reward function, including time efficiency reward, energy consumption penalty, and collision risk penalty.

[0162] Obtaining the picking sequence and planning the path according to the Q function specifically includes: maximizing the expected cumulative reward during the sorting process, that is, finding the optimal policy π * such that π * = argmax π Q(s,a); by adjusting the weight coefficients of each index in the reward function, the policy bias under different business requirements is realized. In the high-priority order scenario, the weight of the time efficiency reward is increased. In the energy-saving mode, the weight of the energy consumption penalty is increased. When facing a complex environment, the weight of the collision risk penalty is strengthened.

[0163] In this embodiment, obtaining the picking sequence and planning the path according to the Q function, and balancing the sorting efficiency (time), energy consumption, and collision risk through the reward function can reduce the collision risk. The storage rack body, the feature access module, the controller, and the access robot are combined, and the optimal storage position is allocated according to the matching access heat map and the comprehensive heat of the goods, which is not only beneficial to optimizing the storage location allocation of the intelligent storage rack for goods, and then optimizing the goods access path. On this basis, the space utilization rate can be further improved and the comprehensive cost can be reduced, and the intelligence level and warehousing efficiency of the storage rack are further improved.

[0164] As a preferred technical solution, as Figure 6 shown, the method for using the intelligent storage rack further includes the following steps:

[0165] S7. Obtain the basic path cost item; wherein, the basic path cost item includes the actual movement cost g(n) from the starting point to the current node n and the estimated cost h(n) from the current node n to the end point.

[0166] S8. Construct the spatio-temporal conflict prediction item; wherein, the spatio-temporal conflict prediction item is composed of the conflict penalty factor γ”, the prediction time window length T, and the spatio-temporal conflict indicator function Collision. The spatio-temporal conflict indicator function is expressed as outputting 1 when there is a collision risk at node n at time t, and 0 otherwise.

[0167] S9. Construct a cost function based on the basic path cost item and the spatio-temporal conflict prediction item, and obtain the planned path based on the cost function.

[0168] Specifically, the cost function The actual movement cost includes but is not limited to distance, time, and energy consumption, and can be obtained by weighted calculation of the path length, movement time, and movement energy consumption.

[0169] Preferably, the estimated cost (x e , y e ) represents the end point coordinates, (x n , y n ) represents the current node coordinates, that is, the real-time position of the end effector of the access manipulator, which is obtained through a joint encoder or a vision positioning system. λ” represents the obstacle density influence coefficient, which is used to adjust the influence intensity of the obstacle density on the path valuation. The value range is usually 0.2 - 0.5. For low-risk scenarios (such as an empty passage), λ” can be set to 0.2 to slightly correct the path (the path valuation is increased by 20%). For high-risk scenarios (such as a temporary stacking area), λ” can be set to 0.5 to significantly correct the path (the path valuation is increased by 50%). For the obstacle density influence coefficient, it can be dynamically adapted and adjusted according to different scenarios.

[0170] obstacle_density represents the obstacle density, and can be calculated by the formula R represents the preset radius; the effective obstacle can be understood as a dynamic / static object within a circular area with a radius R around node n and with a height within the manipulator movement envelope (such as 0.5 - 1.5 m).

[0171] The estimated cost function of the traditional A* algorithm only considers the geometric distance, while the formula Here, the distance estimation is dynamically weighted by the correction term (1 + λ”·obstacle_density): in the sparse obstacle area (obstacle_density≈0), it can maintain the original distance estimation and prefer the straight-line path; while in the dense obstacle area (obstacle_density>0.3 / m 2 ), it can significantly improve the path estimation in this area and guide the algorithm to detour. Therefore, based on the improved predicted cost function, it can not only reduce the collision probability by avoiding high-risk areas in advance, but also reduce the expansion of invalid nodes by dynamically adjusting the path estimation by quantifying the environmental risk, making the robotic arm efficient and safe in complex scenarios.

[0172] The conflict penalty factor can be appropriately adjusted according to the actual scenario. For example, when in a safety-sensitive environment (such as human-robot collaboration), the value of the conflict penalty factor can be appropriately increased, such as setting γ” between 2.0 - 5.0. When in an efficiency-priority environment (such as a pure machine environment), its value can be appropriately decreased, such as setting γ” between 0.5 - 1.5.

[0173] The unit of the prediction time window length T is seconds, which determines the depth of the algorithm's foresight of future risks. For example, a short cycle (T = 3 seconds) is suitable for high-speed sorting scenarios (robotic arm speed ≥ 2m / s), and a long cycle (T = 5 - 8 seconds) is used for large-scale AGV cluster scheduling.

[0174] For the spatio-temporal conflict indication function Collision, its collision condition is: Among them, R safe respectively represent the predicted position of the current robotic arm at time t, the predicted positions of other robotic arms / obstacles, and the safety radius, and t current represents the starting time of the path planning decision. For the collision condition of the spatio-temporal conflict indication function, every time a path planning iteration is completed, t current is updated to the latest time, and the time window [t current ,t current ,+T] rolls forward, and ||·|| represents the three-dimensional space distance, which is used to characterize the actual distance between two objects in space.

[0175] The collision condition can be understood as: there exists a certain time t within the window from the current time to the current time plus T, such that the distance between the predicted position of the current robotic arm and the predicted positions of other robotic arms or obstacles is less than the safety radius, then the output is 1, otherwise the output is 0.

[0176] The specific method for obtaining a planned path based on a cost function includes: algorithm initialization and parameter setting (such as environmental modeling, configuration of input features / output results / update frequency of the LSTM prediction module, setting of parameters including a conflict penalty factor and the length of the prediction time window in the cost function), conflict probability evaluation, node expansion and cost calculation, priority queue management, and incremental path update. Among them, conflict probability evaluation includes predicting the trajectories of all adjacent robotic arms in the next 30 steps through LSTM, generating a spatio-temporal occupancy heatmap, and determining the collision risk; node expansion and cost calculation include actual movement cost calculation, prediction cost calculation, and conflict penalty superposition calculation, etc.; priority queue management includes using a binary heap to store nodes to be expanded, arranging them in ascending order according to the cost function, and evaluating the head node of the queue every 100 ms. If the value of the cost function increases by more than a preset threshold due to environmental changes, path backtracking is triggered, etc.; incremental path update includes when a new obstacle is detected or the prediction error > 5 cm, retaining the first 50% of the planned path and restarting the A* search from the intermediate node, etc.

[0177] Here, converting the LSTM prediction result into a probabilistic occupancy heatmap to replace traditional binary collision detection can improve the tolerance to uncertainty.

[0178] In summary, for obtaining a planned path based on a cost function, it predicts the trajectories of other robotic arms / AGVs within the next 3 seconds through an LSTM network, generates a spatio-temporal conflict heatmap, and adjusts the path cost function in combination with a conflict penalty factor, which can generate a detour path in advance, avoid sudden stops or path replanning, and further improve the intelligence level of the described intelligent storage rack usage method. Constructing a cost function based on the basic path cost item and the spatio-temporal conflict prediction item, comprehensively considering multi-dimensional optimization objectives including distance, energy consumption, efficiency, and safety, can better perform path optimization and resource conservation.

[0179] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0180] The above embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it cannot be understood as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the invention patent should be subject to the appended claims.

Claims

1. An intelligent storage rack, comprising a storage rack body, a controller, and at least one feature acquisition module for acquiring feature information of goods to be stored. A plurality of storage positions are arranged on the storage rack body. The controller is used for storing historical access data of the storage positions. The feature acquisition module is connected to the controller and feeds back the feature information to the controller, characterized in that, The controller is further configured to: Obtain the heat value of each storage location according to the historical access data, and construct an access heat map based on the heat values of multiple storage locations; Obtain the comprehensive heat of the goods to be stored according to the historical access data and the feature information; And allocate the optimal storage location according to the access heat map and the comprehensive heat of the goods.

2. The intelligent storage rack according to claim 1, characterized in that, The controller includes: A distance acquisition module, configured to acquire the three-dimensional Euclidean distance between the storage location and each access location in the historical access events; A contribution degree acquisition module, configured to acquire the contribution degree of each historical access event to the storage location according to the three-dimensional Euclidean distance; A heat map construction module, configured to perform weighted superposition on the contribution degrees to obtain the heat value, and construct an access heat map according to the heat value.

3. The intelligent storage rack according to claim 2, characterized in that, The controller further includes: An access frequency acquisition module, configured to acquire the comprehensive access frequency and the recent access frequency of the goods to be stored; An associated heat acquisition module, configured to acquire the heat of associated goods; A comprehensive heat acquisition module, configured to acquire the comprehensive heat of the goods according to the comprehensive access frequency, the recent access frequency, and the heat of associated goods.

4. The intelligent storage rack according to claim 3, wherein, The heat map construction module constructs an access heat map according to the formula to construct an access heat map; Among them, H(x, y, z) represents the heat value of the three-dimensional coordinate point (x, y, z) of the storage location, N represents the total number of historical access events, and d i represents the three-dimensional Euclidean distance between the three-dimensional coordinate point (x, y, z) of the storage location and the three-dimensional access position of the i-th time, σ represents the diffusion coefficient of the Gaussian kernel, μ represents the distance offset, e represents the natural constant, represents the contribution degree of the i-th historical access event to the storage location.

5. The intelligent storage rack according to claim 4, wherein, The comprehensive heat acquisition module obtains the comprehensive heat of the goods according to the formula H i = α1·F Com_i + α2·F Recent_i + α3·R i ; Among them, F Com_i , F Recent_i , R i respectively represent the comprehensive access frequency, the recent access frequency, and the associated goods popularity, and α1, α2, and α3 respectively represent the weight coefficients of the comprehensive access frequency, the recent access frequency, and the associated goods popularity.

6. An intelligent storage rack according to claim 5, wherein, The intelligent storage rack further includes an access robotic arm that is signal-connected to the controller. The access robotic arm is configured to respond to a control instruction and perform order sorting. The controller further includes: A path planning module, which is configured to obtain a picking sequence and plan a path according to the Q function after receiving a sorting order, and generate a control instruction according to the picking sequence and the planned path; ​ Among them, s represents the current system state, s t represents the operating state of the system at time t, s represents the currently selected action, s t represents the specific action of the access manipulator at time t, γ represents the discount factor, r t+k represents the immediate reward of the system at time t + k, E represents the expectation operator, k represents the future time step index starting from the current time t, r t represents the reward function, including time efficiency reward, energy consumption penalty, and collision risk penalty.

7. A method for using an intelligent storage rack, characterized in that, The method for using the intelligent storage rack includes: Obtain the feature information of the goods to be stored and the historical access data of the storage location; Obtain the heat value of each storage location according to the historical access data, and construct an access heat map based on the heat values of multiple storage locations; Obtain the comprehensive heat of the goods to be stored according to the historical access data and the feature information; Allocate the optimal storage location according to the access heat map and the comprehensive heat of the goods.

8. The method for using an intelligent storage rack according to claim 7, wherein, The specific method for constructing the access heat map includes: Obtain the three-dimensional Euclidean distance d between the storage location and each access location in the historical access events i ; Obtain the contribution degree of each historical access event to the storage location according to the three-dimensional Euclidean distance Perform weighted superposition on the contribution degrees to obtain the heat value, and construct an access heat map according to the heat value; Among them, represents the heat value of the three-dimensional coordinate point (x, y, z) of the storage location, N represents the total number of historical access events, σ represents the diffusion coefficient of the Gaussian kernel, μ represents the distance offset, and e represents the natural constant.

9. The method for using an intelligent storage rack according to claim 8, characterized in that, The specific method for obtaining the comprehensive heat of the goods to be stored includes: Obtain the comprehensive access frequency F of the goods to be stored Com_i , the recent access frequency F Recent_i and the popularity R of related goods i ; Obtain the comprehensive popularity H of goods based on the comprehensive access frequency, recent access frequency, and the popularity of associated goods i = α1·F Com_i + α2·F Recent_i + α3·R i . Wherein, α1, α2, and α3 respectively represent the weight coefficients of the comprehensive access frequency, the recent access frequency, and the heat of associated goods.

10. A method for using an intelligent storage rack according to claim 9, characterized in that, The method for using the intelligent storage rack further includes: After receiving the sorting order, according to the Q function Obtain the picking sequence and plan the path, and generate a control instruction according to the picking sequence and the planned path; The access robotic arm responds to the control instruction and performs order sorting; Among them, s represents the current system state, s t represents the operating state of the system at time t, s represents the currently selected action, s t represents the specific action of the access manipulator at time t, γ represents the discount factor, r t+k represents the immediate reward of the system at time t + k, E represents the expectation operator, k represents the future time step index starting from the current time t, r t represents the reward function, including time efficiency reward, energy consumption penalty, and collision risk penalty.

Images