Unmanned retail terminal commodity recommendation method and system based on edge internet-of-things perception

Abstract

The invention relates to the technical field of unmanned retail recommendation, and discloses an unmanned retail terminal commodity recommendation method and system based on edge internet of things perception, and the method comprises the steps: carrying out the consumption event extraction of retail commodities and the consumption track extraction of a retail region from consumer behavior data; extracting associated consumption weights among different retail commodities and independent consumption weights of the retail commodities by utilizing a commodity interaction sequence mining mode; calculating visibility indexes and interaction probabilities of different display positions in the unmanned retail terminal; and a reward function for improving the purchase intention and the purchase conversion rate of the consumer is constructed and solved, and a retail commodity display layout scheme of the unmanned retail terminal is obtained. According to the method, edge internet-of-things perception and an interactive mining mode of retail commodities and display positions are combined, and the layout scheme of the unmanned retail terminal can be dynamically generated based on real interactive behaviors of consumers and display position values, so that the purchase intention expression effect and the actual purchase conversion rate are improved.

Description

Technical Field

[0001] This invention relates to the field of unmanned retail recommendation, and more particularly to a method and system for recommending goods in unmanned retail terminals using edge IoT sensing. Background Technology

[0002] With the continuous development of IoT, edge computing, and smart retail technologies, unmanned retail terminals are evolving from the early "unattended vending" model centered on automated vending to a "smart sensing and precise recommendation" model with behavioral perception, decision analysis, and dynamic optimization capabilities. Consumers' dwell time, movement paths, product pickup and return, and final purchase behavior in unmanned retail terminals constitute a complete and continuous consumption process. This behavioral data not only reflects consumers' immediate needs but also implies key information such as product attention, preference intensity, and related selection habits. Meanwhile, in practical applications, traditional retail terminal product layouts rely heavily on manual experience and static design, making it difficult to adapt to dynamic changes in consumer behavior. This leads to problems such as uneven product exposure, low utilization of display space, and limited sales conversion rates. Therefore, if the above behavioral data can be structured and comprehensively analyzed, a reliable data foundation can be provided for product recommendation strategy formulation and display layout optimization.

[0003] Existing research largely focuses on sales forecasting and inventory management. For example, patent CN118428998A proposes a sales forecasting method and system for unmanned retail smart cabinets. By collecting, cleaning, and extracting features from historical data, it uses an LSTM model to predict sales, thereby reducing the risk of stockouts and slow-moving inventory. This approach has certain advantages in terms of sales forecasting accuracy and system integration, but its core remains at the time-series level, primarily serving replenishment and inventory decisions. Furthermore, such methods typically rely on centralized cloud modeling, lacking sufficient support for real-time edge perception and instant recommendations, and making it difficult to directly map the forecast results to specific product recommendation and display layout adjustment strategies.

[0004] To address this issue, this invention proposes a product recommendation method and system for unmanned retail terminals based on edge IoT sensing. This method dynamically optimizes the display layout of retail products in unmanned retail terminals, enabling the placement and arrangement of retail products to enhance consumers' purchase intentions and thereby improve the purchase conversion rate of retail products. Summary of the Invention

[0005] This invention provides a product recommendation method and system for unmanned retail terminals based on edge IoT sensing. Step S1 collects infrared and RFID signal stream data at the edge using multimodal sensing devices, avoiding reliance on image data and enabling refined modeling of consumer dwell time, interaction, and purchasing behavior while protecting privacy. This solves the problems of high reliance on visual recognition and high deployment costs in traditional unmanned retail systems. Step S2 extracts consumer events and mines product interaction sequences, transforming discrete retail product picking and purchasing behaviors into calculable individual and associated consumer weights. This addresses the difficulty of directly extracting implicit product pairing relationships from raw sensing data. The technical challenges of acquisition are addressed in step S3, which introduces a visibility index and interaction probability to quantify the exposure and interaction intensity of different display positions in the consumption path, breaking through the limitations of relying solely on static locations or human experience to evaluate the value of display positions. Step S4 integrates retail product weights and display position characteristics into the reward function to construct an optimization objective that directly reflects purchase intent and conversion effects. It then introduces an adaptive layout optimization algorithm based on reinforcement learning, which can gradually approach the optimal retail product display layout scheme through continuous interaction and feedback, solving the problem that traditional rule-driven or one-time optimization methods cannot cope with the dynamic changes in consumer behavior.

[0006] To achieve the above objectives, the present invention provides a product recommendation method for unmanned retail terminals based on edge IoT sensing, comprising the following steps: S1: Collect consumer behavior data using multimodal sensing devices deployed in unmanned retail terminals, and extract consumption events of retail goods and consumption trajectories of retail areas from the consumer behavior data to obtain a set of consumption events of retail goods and a set of consumption trajectories of retail areas in the unmanned retail terminal. S2: Based on the set of consumption events for the retail goods, extract the associated consumption weights between different retail goods and the individual consumption weights of retail goods using the product interaction sequence mining method; S3: Based on the set of consumption trajectories in the retail area, calculate the visibility index and interaction probability of different display positions in the unmanned retail terminal; S4: Based on the associated consumption weights between different retail products, the individual consumption weights of retail products, the visibility index of different display positions, and the interaction probability, a reward function is constructed to improve consumers' purchase intention and purchase conversion rate. The reward function is solved using an adaptive layout optimization algorithm based on reinforcement learning to obtain the retail product display layout scheme of the unmanned retail terminal.

[0007] As a further improvement of the present invention: Furthermore, in step S1, consumer behavior data is collected using multimodal sensing devices deployed in unmanned retail terminals, including: S11: At the edge of the unmanned retail terminal, a multimodal sensing device consisting of infrared sensing sensors and RFID readers is deployed, with the infrared sensing sensors deployed in front of the unmanned retail terminal and in the merchandise retail channel inside the unmanned retail terminal. S12: The infrared sensing sensor is used to collect the dwell infrared signal stream representing the dwell state and the consumption infrared signal stream representing the consumption behavior. The infrared sensing sensors deployed in front of the unmanned retail terminal and in the merchandise retail channel inside the unmanned retail terminal collect the dwell infrared signal stream and the consumption infrared signal stream in sequence. S13: The RFID reader is used to collect the label change status of retail product labels in the unmanned retail terminal in real time, and the label change status of the retail product labels is constructed into an RFID signal stream; S14: The dwell infrared signal stream, consumption infrared signal stream and RFID signal stream are used as consumer behavior data.

[0008] Furthermore, step S1, which involves extracting consumption events for retail goods and consumption trajectories for retail areas from the consumer behavior data, also includes: S15: Based on the dwell infrared signal stream, calculate the signal energy change rate at the signal acquisition time in the dwell infrared signal stream, identify the initial dwell time when the consumer arrives at the unmanned retail terminal area and the end dwell time when the consumer leaves the unmanned retail terminal, and take the time period between the initial dwell time and the end dwell time as a consumption event and the time range corresponding to a consumption trajectory. S16: Based on the RFID signal stream, extract the set of retail goods whose tag status changed to disappear during the time period between the initial time of stay and the end time of stay, where the tag status changing to disappear indicates that the retail goods have been purchased; take the time period between the initial time of stay and the end time of stay as the event range of the consumption event, take the set of retail goods as the consumed goods of the consumption event, constitute a consumption event, construct all the currently extracted consumption events into a set of consumption events of retail goods in the unmanned retail terminal, and synchronously record the time when the tag status of the retail goods in the consumption event changes to disappear; S17: Based on the consumer infrared signal stream, calculate the signal energy change rate at the signal acquisition time in the consumer infrared signal stream, and count the number of signal acquisition times where the absolute value of the signal energy change rate is higher than the preset change rate threshold. Use the number of signal acquisition times as the number of operations performed by the consumer to take and put back retail goods. S18: Extract the set of retail goods in the consumption event corresponding to the time period between the initial time of stay and the end time of stay, sort the retail goods in the set of retail goods according to the order in which the tag status changes to disappear, and obtain the display positions of the sorted retail goods to form a display position sequence. S19: The number of signal acquisition times and the display bit sequence are taken as a consumption trajectory, and all the currently extracted consumption trajectories are constructed as a set of consumption trajectories in the retail area of ​​the unmanned retail terminal.

[0009] Furthermore, step S2, which uses product interaction sequence mining to extract the associated consumption weights between different retail products and the individual consumption weights of retail products, also includes: S21: Based on the set of consumption events for the retail goods, the time period between the initial time of stay and the end time of stay associated with the consumption event is taken as the consumption time period, and the length of the consumption time period is extracted as the consumption duration of the consumption event. The average consumption duration of all consumption events is calculated, and the weight factor of the consumption event is calculated using the average consumption duration. The retail goods in each consumption event are weighted to obtain the individual consumption weight of the retail goods. S22: Obtain the time when the label status of different retail products in the same consumption event changes to disappearance, and calculate the time correlation factor between any two retail products; Specifically, the formula for calculating the time correlation factor between any two retail products is as follows: ; ; in, Represents the uth type of retail goods With the vth type of retail goods The time-related factor in the m-th consumption event. M represents the total number of consumption events. Indicates the number of types of retail goods. These represent the u-th type of retail product, respectively. With the vth type of retail goods At the moment when the tag state changes to disappear in the m-th consumption event. This represents an exponential function with the natural constant as its base. Indicates the time decay scale; S23: Obtain the number of times different retail products co-occur in a consumption event, use an improved co-occurrence frequency algorithm to calculate the preliminary correlation between any two retail products, and use the preliminary correlation to weight the preliminary correlation to obtain the associated consumption weight between retail products.

[0010] Specifically, the formula for calculating the preliminary correlation degree is as follows: ; in, The u-th retail item in the set of retail goods representing a consumption event With the vth type of retail goods The number of simultaneous consumer events This indicates that the u-th retail item appears in the set of retail items. The number of consumer events This indicates that the v-th retail item appears in the set of retail items. The number of consumer events; The uth type of retail product With the vth type of retail goods The related consumption weight between them is .

[0011] Further, in step S21, the weighting factor of the consumption event is calculated using the average consumption duration, and the retail goods in each consumption event are weighted to obtain the individual consumption weight of the retail goods. The formula for calculating the individual consumption weight is as follows: ; in, Represents the uth type of retail product Individual consumption weight, This represents the weight factor for the m-th consumption event. This represents the duration of the m-th consumption event. Let M represent the average duration of all consumption events, and M represent the total number of consumption events. Indicates the number of types of retail goods. Represents a collection of retail goods Does the uth type of retail product exist? The discriminant function, Let m be the set of retail goods in the m-th consumption event. If the set of retail goods... There exists a type u of retail goods. ,but =1, otherwise It is 0.

[0012] Furthermore, step S3 calculates the visibility index and interaction probability of different display positions in the unmanned retail terminal, including: S31: Calculate the display weight of different display positions in the consumption trajectory according to the order in which the display positions appear in the consumption trajectory. S32: Calculate the visibility index of display positions in unmanned retail terminals by combining the indicator function value of whether the display position appears in the display position sequence and the display weight. S33: Based on the number of signal acquisition times in the consumption trajectory, the indicator function value of whether the display position appears in the display position sequence is weighted to generate the interaction probability of the display position in the unmanned retail terminal.

[0013] Furthermore, step S4 constructs a reward function to enhance consumer purchase intent and purchase conversion rate, the expression of which is: ; in, Represents the reward function, This represents the weighting factor of the reward function. Let represent the retail merchandise display layout scheme to be solved. This represents the reward function value corresponding to retail product display layout scheme A. Indicates the first One display space Visibility index Indicates the first One display space The probability of interaction, Indicates the number of display positions. This indicates the first step in retail merchandise display layout scheme A. One display space The retail products on display Indicates retail goods Individual consumption weight, Indicates retail goods The related consumption weight between them Indicates the u-th type of retail product. K represents the number of retail product types. This indicates the retail products in retail product display layout scheme A. The display space Indicates retail goods Display space distance constraints between them.

[0014] Furthermore, step S4, which employs a reinforcement learning-based adaptive layout optimization algorithm to solve the reward function and obtain a retail product display layout scheme for the unmanned retail terminal, also includes: S41: Initialize and generate a retail product display layout scheme as the current state; S42: Select the optimal layout adjustment action that may improve the reward function value from the pre-trained display layout adjustment action space, adjust the current state according to the selected layout adjustment action, generate the next time step state, and use the next time step state as the input value of the reward function to calculate the reward function value of the next time step state. S43: Use the Q-learning algorithm to update the Q value between the current state and the layout adjustment action. If the update magnitude is less than the preset threshold, or the number of Q value updates reaches the preset maximum number of iterations, it indicates that the algorithm has converged, and the state at the next moment is used as the retail product display layout scheme of the unmanned retail terminal obtained by solving. Otherwise, proceed to step S44. S44: Set the state of the next time step as the current state, record the number of times the Q value has been updated, and return to step S42.

[0015] This invention also proposes an edge IoT-sensing unmanned retail terminal product recommendation system, which includes a data processing module, an event extraction module, a consumer mining module, and a layout scheme recommendation module, to realize the edge IoT-sensing unmanned retail terminal product recommendation method described above.

[0016] Compared with existing technologies, this invention proposes a product recommendation method and system for unmanned retail terminals based on edge IoT sensing. This technology has the following beneficial effects: First, in calculating individual consumption weights, this invention introduces a weighting factor based on the average consumption duration to differentiate the weighting of consumption events. This effectively avoids the bias caused by measuring the importance of a product solely by its frequency of occurrence. The calculated individual consumption weights can comprehensively reflect the participation of retail products in different consumption events and their corresponding contribution to consumption duration. This allows consumption events with longer dwell times and higher levels of interaction to have a greater impact on individual consumption weights. This usually means that consumers have conducted more observation, comparison, and psychological evaluation of retail products, rather than just passing by or quickly browsing. This more realistically portrays consumers' attention to retail products and the intensity of their purchase intentions.

[0017] Meanwhile, this invention constructs a reward function by uniformly modeling the visibility index of display positions, interaction probability, individual consumption weight of retail products, and associated consumption weight between retail products, thereby achieving deep coupling between consumer spatial behavior and retail product consumption preferences. Specifically, the reward function constructs the consumption value of retail products by considering the individual consumption weight of retail products and the associated consumption weight between retail products. It uses the visibility index and interaction probability to weight display positions, guiding high-exposure and high-interaction display positions to prioritize high-consumption-value retail products, enhancing consumers' perceived attention and purchase intention. At the same time, the reward function introduces expected distance constraints based on associated consumption weights, ensuring that retail products with strong correlations remain spatially close in the display layout, reducing the movement and cognitive costs for consumers during multi-product joint purchases. This transforms the retail product display layout scheme from experience-based adjustments to data-driven optimization, effectively improving the probability of retail products being noticed, the occurrence rate of associated purchases, and the overall purchase conversion rate in unmanned retail scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a product recommendation method for an unmanned retail terminal based on edge IoT sensing, provided in an embodiment of the present invention. Figure 2 This is a deployment diagram of a multimodal sensing device for an unmanned retail terminal provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the system structure of a product recommendation system for an unmanned retail terminal provided in an embodiment of the present invention.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The realization of the objectives, functional characteristics, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] This invention provides a product recommendation method for unmanned retail terminals based on edge IoT sensing. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices: a server, a terminal, or any other electronic device configured to execute the method provided in this invention. In other words, the product recommendation method for unmanned retail terminals based on edge IoT sensing can be executed by software or hardware installed on a terminal device or a server device, where the software may be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0022] Reference Figure 1 as well as Figure 2 As shown, Embodiment 1 of the present invention is as follows: A product recommendation method for an edge IoT-sensing unmanned retail terminal, the method comprising: S1: Collect consumer behavior data using multimodal sensing devices deployed in unmanned retail terminals, and extract consumption events of retail goods and consumption trajectories of retail areas from the consumer behavior data to obtain a set of consumption events of retail goods and a set of consumption trajectories of retail areas in the unmanned retail terminal.

[0023] Step S1 involves collecting consumer behavior data using multimodal sensing devices deployed in unmanned retail terminals, including: S11: At the edge of the unmanned retail terminal, a multimodal sensing device consisting of infrared sensing sensors and RFID readers is deployed, with the infrared sensing sensors deployed in front of the unmanned retail terminal and in the merchandise retail channel inside the unmanned retail terminal. S12: The infrared sensing sensor is used to collect the dwell infrared signal stream representing the dwell state and the consumption infrared signal stream representing the consumption behavior. The infrared sensing sensors deployed in front of the unmanned retail terminal and in the merchandise retail channel inside the unmanned retail terminal collect the dwell infrared signal stream and the consumption infrared signal stream in sequence. S13: The RFID reader is used to collect the label change status of retail product labels in the unmanned retail terminal in real time, and the label change status of the retail product labels is constructed into an RFID signal stream; As an embodiment of the present invention, refer to as follows Figure 2 The diagram shows the deployment of multimodal sensing devices in an unmanned retail terminal. The infrared sensors are deployed in a zoned manner, installed in the front entrance area of ​​the unmanned retail terminal (as shown by infrared sensor 2 in the figure) and above the merchandise retail aisle or at the front edge of the shelves inside the unmanned retail terminal (as shown by infrared sensor 1 in the figure). The infrared sensor located in the front entrance area is used to sense the dwell status and duration of consumers after entering the unmanned retail terminal area, forming a dwell infrared signal stream. Its sensing range covers the key spatial areas where consumers enter, linger, and leave. The infrared sensors deployed in the merchandise retail aisle are arranged in an array along the direction of the merchandise retail aisle, used to capture consumer-related actions such as approaching, reaching out, and moving in front of the shelves, thereby forming a continuous consumer infrared signal stream.

[0024] The RFID reader is deployed inside or under the shelf, forming a near-field coverage relationship with the retail product tags. The retail product tags are RFID electronic tags bound to each retail product, used to store the unique identification information, product category code and basic attribute information of the retail product. The RFID electronic tags adopt a passive or low-power design, and are activated when they enter the effective sensing range of the RFID reader and return the corresponding tag status information, forming the tag change status, and realizing the automatic sensing of the tag change status. Each infrared sensing sensor and RFID reader is connected to a unified clock synchronization module through an edge computing node to achieve time alignment and centralized processing of multi-source sensing data.

[0025] Specifically, the RFID reader continuously transmits radio frequency excitation signals to the shelf area and receives the received signal strength returned by the retail product tags as the response status of the retail product tags at the corresponding time. By periodically scanning the response status of the retail product tags within its coverage area, when the retail product is picked up, moved, or put back, the spatial positional relationship between the retail product tag and the RFID reader changes, resulting in a change in the response status, forming a tag change status, thus realizing the real-time collection of the change status of retail product tags in unmanned retail terminals.

[0026] S14: The dwell infrared signal stream, consumption infrared signal stream and RFID signal stream are used as consumer behavior data.

[0027] Specifically, the representation of the dwell infrared signal stream, the consumption infrared signal stream, and the RFID signal stream is as follows: ; ; ; in, These represent the dwell infrared signal stream, the consumption infrared signal stream, and the RFID signal stream, respectively, where t represents the timing information of the signal stream. Indicates the time of signal acquisition. , , These represent the dwell infrared signal streams in sequence. Consumer infrared signal stream At the time of signal acquisition The signal value, Represents RFID signal flow At the time of signal acquisition The signal values, where the signal values ​​in the dwell infrared signal stream and the consumer infrared signal stream represent the infrared intensity values, indicate that the consumer has entered the sensing area of ​​the infrared sensor. A rapid increase in infrared intensity value from a low value indicates that the consumer has entered the sensing area of ​​the infrared sensor. , Indicates signal value The Middle The status of label changes for various retail products. Indicates the first Retail product labels for various retail goods This indicates the time when the RFID reader / writer acquires signals. The received signal strength returned by the retail product tag. Indicates the first Retail goods at the time of signal acquisition The tag status, The label status is "in place" to indicate that the retail product is still in the corresponding display position, and the label status is "disappeared" to indicate that the retail product has been taken from the display position and left the shelf area of ​​the unmanned retail terminal. Specifically, if ,but ,otherwise ,in The reference received signal strength, representing the label strength of retail goods, is measured in a laboratory environment.

[0028] It should be noted that the signal values ​​in the dwell infrared signal stream and the consumption infrared signal stream have been normalized.

[0029] Step S1, which involves extracting consumption events for retail goods and consumption trajectories for retail areas from the consumer behavior data, also includes: S15: Based on the dwell infrared signal stream, calculate the signal energy change rate at the signal acquisition time in the dwell infrared signal stream, identify the initial dwell time when the consumer arrives at the unmanned retail terminal area and the end dwell time when the consumer leaves the unmanned retail terminal, and take the time period between the initial dwell time and the end dwell time as a consumption event and the time range corresponding to a consumption trajectory. Specifically, the sliding window method is used to calculate the rate of change of signal energy at the signal acquisition time in the stationary infrared signal stream, where the signal acquisition time... The formula for calculating the rate of change of signal energy is: ; in, Indicates the dwell infrared signal stream At the time of signal acquisition The signal value, where L represents the length of the sliding window, is set to 4. Indicates the dwell infrared signal stream China and Israel The mean signal value within the central sliding window. Indicates the signal acquisition time The rate of change of signal energy; like If the value is higher than the preset start dwell threshold (e.g., 0.4), it indicates the signal acquisition time. When a consumer arrives at the unmanned retail terminal area, a signal collection time is set. At the initial moment of stay, if If the value is below the preset dwell end threshold (e.g., -0.4), it indicates the signal acquisition time. Signal collection time is set when consumers leave the unmanned retail terminal area. For any initial time of stay time1, iterate through the time to find the time of stay end time2 that is later than the initial time of stay time1 and closest to the initial time of stay time1. The time period between the initial time of stay time1 and the time of stay end time2 is taken as a consumption event and the time range corresponding to a consumption trajectory. S16: Based on the RFID signal stream, extract the set of retail goods whose tag status changed to disappear during the time period between the initial time of stay and the end time of stay, where the tag status changing to disappear indicates that the retail goods have been purchased; take the time period between the initial time of stay and the end time of stay as the event range of the consumption event, take the set of retail goods as the consumed goods of the consumption event, constitute a consumption event, construct all the currently extracted consumption events into a set of consumption events of retail goods in the unmanned retail terminal, and synchronously record the time when the tag status of the retail goods in the consumption event changes to disappear; S17: Based on the consumer infrared signal stream, calculate the signal energy change rate at the signal acquisition time in the consumer infrared signal stream, and count the number of signal acquisition times where the absolute value of the signal energy change rate is higher than a preset change rate threshold (e.g., 0.3). Use the number of signal acquisition times as the number of operations performed by consumers to take and put back retail goods. Specifically, the calculation method for the rate of change of signal energy at the moment of signal acquisition in the consumer infrared signal stream is the same as that in step S15; S18: Extract the set of retail goods in the consumption event corresponding to the time period between the initial time of stay and the end time of stay, sort the retail goods in the set of retail goods according to the order in which the tag status changes to disappear, and obtain the display positions of the sorted retail goods to form a display position sequence. S19: The number of signal acquisition times and the display bit sequence are taken as a consumption trajectory, and all the currently extracted consumption trajectories are constructed as a set of consumption trajectories in the retail area of ​​the unmanned retail terminal.

[0030] It should be noted that each consumption event corresponds to a consistent consumption trajectory over a certain period of time. This invention achieves a precise depiction of the entire consumption process—from consumer arrival to retail product interaction, purchase completion, and consumer departure—by collaboratively modeling the dwell infrared signal flow, consumption infrared signal flow, and RFID signal flow. Specifically, the invention adaptively identifies the initial and final dwell times based on the signal energy change rate, avoiding misjudgments caused by fixed time windows and improving the accuracy and robustness of consumption event segmentation. Furthermore, it accurately determines the purchased retail product through RFID tag state transitions and combines the consumption infrared signal flow to quantify the number of times the consumer takes and puts back items, thus giving the consumption trajectory a time dimension, a spatial display position dimension, and an interaction intensity dimension. Further, this invention incorporates the display position sequence and the number of operations into the consumption trajectory description, which helps to reveal the actual interaction paths and potential correlations between products and display positions.

[0031] S2: Based on the set of consumption events for the retail goods, extract the associated consumption weights between different retail goods and the individual consumption weights of retail goods using the product interaction sequence mining method.

[0032] Step S2, which uses product interaction sequence mining to extract the correlation consumption weights between different retail products and the individual consumption weights of retail products, also includes: S21: Based on the set of consumption events for the retail goods, the time period between the initial time of stay and the end time of stay associated with the consumption event is taken as the consumption time period, and the length of the consumption time period is extracted as the consumption duration of the consumption event. The average consumption duration of all consumption events is calculated, and the weight factor of the consumption event is calculated using the average consumption duration. The retail goods in each consumption event are weighted to obtain the individual consumption weight of the retail goods. S22: Obtain the time when the label status of different retail products in the same consumption event changes to disappearance, and calculate the time correlation factor between any two retail products; Specifically, the formula for calculating the time correlation factor between any two retail products is as follows: ; ; in, Represents the uth type of retail product With the vth type of retail goods The time-related factor in the m-th consumption event. M represents the total number of consumption events. Indicates the number of types of retail goods. These represent the u-th type of retail product, respectively. With the vth type of retail goods At the moment when the tag state changes to disappear in the m-th consumption event. This represents an exponential function with the natural constant as its base. Indicates the time decay scale, set It is 1.2; S23: Obtain the number of times different retail products co-occur in a consumption event, use an improved co-occurrence frequency algorithm to calculate the preliminary correlation between any two retail products, and use the preliminary correlation to weight the preliminary correlation to obtain the associated consumption weight between retail products.

[0033] Specifically, the formula for calculating the preliminary correlation degree is as follows: ; in, The u-th retail item in the set of retail goods representing a consumption event With the vth type of retail goods The number of simultaneous consumer events This indicates that the u-th retail item appears in the set of retail items. The number of consumer events This indicates that the v-th retail item appears in the set of retail items. The number of consumer events; The uth type of retail product With the vth type of retail goods The related consumption weight between them is .

[0034] It should be noted that this invention constructs a temporal association factor based on the time when the label state disappears, and combines it with an improved co-occurrence frequency. This can characterize the temporal proximity of retail goods in the same consumption event while ensuring statistical stability. This makes the associated consumption weight of two retail goods with closer purchase times greater, thereby avoiding false associations caused by relying solely on the number of co-occurrences and improving the accuracy of product association modeling and the rationality of layout optimization.

[0035] In step S21, the weighting factor of the consumption event is calculated using the average consumption duration, and the retail goods in each consumption event are weighted to obtain the individual consumption weight of the retail goods. The formula for calculating the individual consumption weight is as follows: ; in, Represents the uth type of retail goods Individual consumption weight, This represents the weight factor for the m-th consumption event. This represents the duration of the m-th consumption event. Let M represent the average duration of all consumption events, and M represent the total number of consumption events. Indicates the number of types of retail goods. Represents a collection of retail goods Does the uth type of retail product exist? The discriminant function, Let m be the set of retail goods in the m-th consumption event. If the set of retail goods... There exists a type u of retail goods. ,but =1, otherwise It is 0.

[0036] S3: Based on the set of consumption trajectories in the retail area, calculate the visibility index and interaction probability of different display positions in the unmanned retail terminal.

[0037] The S3 step calculates the visibility index and interaction probability of different display positions in the unmanned retail terminal, including: S31: Calculate the display weight of different display positions in the consumption trajectory according to the order in which the display positions appear in the consumption trajectory. Specifically, the formula for calculating the display weight of the display position in the consumption trajectory is as follows: ; in, Indicates the first One display space The display weight in the m-th segment of the consumption trajectory. Indicates the number of display positions. Indicates the first One display space The order in which the bit sequence appears is shown in the m-th segment of the consumption trajectory. Indicates the position attenuation coefficient, set It is 2. Represents an exponential function with the natural constant as its base; S32: Calculate the visibility index of display positions in unmanned retail terminals by combining the indicator function value of whether the display position appears in the display position sequence and the display weight. As an embodiment of the present invention, the first One display space The formula for calculating the visibility index is: ; in, Indicates the first One display space Visibility index Indicates display position In the m-th segment of the consumption trajectory, display the indicator function value indicating whether the bit sequence appears; if so, display the bit... If it appears in the bit sequence displayed in the m-th segment of the consumption trajectory, then =1, otherwise 0 S33: Based on the number of signal acquisition times in the consumption trajectory, the indicator function value of whether the display position appears in the display position sequence is weighted to generate the interaction probability of the display position in the unmanned retail terminal.

[0038] Specifically, the first One display space The formula for calculating the interaction probability is: ; in, Indicates the first One display space The probability of interaction, This represents the number of signal acquisition times for the m-th segment of the consumption trajectory.

[0039] It should be noted that this invention quantifies the degree of exposure of different display positions in the consumption trajectory by introducing an exponentially decaying display weight based on the display order. This gives higher weight to display positions that enter the consumer's field of vision earlier (i.e., priority purchase). Furthermore, by combining an indicator function of whether a display position appears, a visibility index is constructed to comprehensively characterize the frequency and intensity of attention paid to a display position. At the same time, the number of signal acquisition moments is used to weight the indicator function to obtain the interaction probability, so that display positions corresponding to high-frequency picking up and putting back strong interactive behaviors contribute more. These interactive behaviors objectively reflect that consumers have performed multiple close-range operations and attention lingering in front of the display position, indicating that the display position is not only seen, but also repeatedly evaluated and compared, thus more realistically reflecting the visibility and interactive value of the display position.

[0040] S4: Based on the associated consumption weights between different retail products, the individual consumption weights of retail products, the visibility index of different display positions, and the interaction probability, a reward function is constructed to improve consumers' purchase intention and purchase conversion rate. The reward function is solved using an adaptive layout optimization algorithm based on reinforcement learning to obtain the retail product display layout scheme of the unmanned retail terminal.

[0041] Step S4 constructs a reward function to enhance consumer purchase intent and purchase conversion rate. The expression for the reward function is as follows: ; ; ; in, Represents the reward function, This represents the weighting factor of the reward function, set... It is 0.2. Let represent the retail merchandise display layout scheme to be solved. This represents the reward function value corresponding to retail product display layout scheme A. Indicates the first One display space Visibility index Indicates the first One display space The probability of interaction, Indicates the number of display positions. This indicates the first step in retail merchandise display layout scheme A. One display space The retail products on display Indicates retail goods Individual consumption weight, Indicates retail goods The related consumption weight between them Indicates the u-th type of retail product. K represents the number of retail product types. This indicates the retail products in retail product display layout scheme A. The display space Indicates retail goods Display space distance constraints between them; Indicates display position The distance between them Indicates retail goods The expected distance between them This represents the minimum distance between any two display positions in an unmanned retail terminal. This represents the maximum distance between any two display positions in an unmanned retail terminal. Specifically, the display area is located in the shelf area of ​​the unmanned retail terminal, and each display area can hold several retail products, but the categories of retail products are the same.

[0042] Step S4 employs an adaptive layout optimization algorithm based on reinforcement learning to solve the reward function, thereby obtaining a retail product display layout scheme for the unmanned retail terminal. It also includes: S41: Initialize and generate a retail product display layout scheme as the current state; S42: Select the optimal layout adjustment action that may improve the reward function value from the pre-trained display layout adjustment action space, adjust the current state according to the selected layout adjustment action, generate the next time step state, and use the next time step state as the input value of the reward function to calculate the reward function value of the next time step state. S43: Use the Q-learning algorithm to update the Q value between the current state and the layout adjustment action. If the update magnitude is less than the preset threshold, or the number of Q value updates reaches the preset maximum number of iterations, it indicates that the algorithm has converged, and the state at the next moment is used as the retail product display layout scheme of the unmanned retail terminal obtained by solving. Otherwise, proceed to step S44. Specifically, the formula for updating the Q value is: ; in, This indicates that the current state P is performing a layout adjustment action. Q value, This indicates the update result of the Q value. Indicates the learning rate, set It is 0.85. Indicates the state at the next moment. The reward function value, Indicates the state at the next moment. Adjusting actions using arbitrary layout The maximum Q value, This indicates the space for layout adjustment actions. This represents the discount factor, set to 0.2; As an embodiment of the present invention, by acquiring the associated consumption weights among multiple groups of retail goods, the individual consumption weights of retail goods, the visibility index of different display positions, and the interaction probability, and manually generating (or simulating in real retail scenarios) the optimal retail goods display layout scheme for each group, the layout adjustment actions and the adjusted reward function improvement value are recorded during the generation process of the optimal retail goods display layout scheme. The recorded layout adjustment actions and the adjusted reward function improvement value are used to construct a pre-trained display layout adjustment action space, and the Q value between different retail goods display layout schemes and layout adjustment actions is initialized. During the adaptive layout process of the unmanned retail terminal, the Q value is updated online in real time. By periodically generating the retail goods display layout scheme of the unmanned retail terminal, and arranging the display positions of retail goods according to the retail goods display layout scheme, the retail goods layout strategy associated with the reinforcement learning-based adaptive layout optimization algorithm is continuously improved, thereby continuously adapting to the real purchasing behavior of consumers. Optionally, since the number of retail product types is higher than the number of display positions, before proceeding with the retail product display layout scheme generation process as described in steps S1 to S4, retail products with lower sales volumes can be excluded from the retail product display layout scheme generation process based on their sales volume.

[0043] S44: Set the state of the next time step as the current state, record the number of times the Q value has been updated, and return to step S42. Example

[0044] An edge IoT-sensing unmanned retail terminal product recommendation system 100 is provided to implement the edge IoT-sensing unmanned retail terminal product recommendation method as described in Embodiment 1. (Refer to...) Figure 3 The system structure diagram of the unmanned retail terminal product recommendation system 100 shown includes a data processing module 101, an event extraction module 102, a consumer mining module 103, and a layout scheme recommendation module 104. Data processing module 101 is used to collect consumer behavior data using multimodal sensing devices deployed in unmanned retail terminals; The event extraction module 102 is used to extract consumption events of retail goods and consumption trajectories of retail areas from consumer behavior data, so as to obtain a set of consumption events of retail goods and a set of consumption trajectories of retail areas in unmanned retail terminals. The consumer mining module 103 is used to extract the associated consumption weights between different retail products and the individual consumption weights of retail products by using the product interaction sequence mining method, and to calculate the visibility index and interaction probability of different display positions in the unmanned retail terminal. The layout scheme recommendation module 104 is used to construct a reward function to improve consumer purchase intention and purchase conversion rate. The reward function is solved by an adaptive layout optimization algorithm based on reinforcement learning to obtain the retail product display layout scheme of the unmanned retail terminal.

[0045] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in terms of the scope of the patent invention.

[0046] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0047] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0048] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A product recommendation method for unmanned retail terminals based on edge IoT sensing, characterized in that, The method includes: S1: Collect consumer behavior data using multimodal sensing devices deployed in unmanned retail terminals, and extract consumption events of retail goods and consumption trajectories of retail areas from the consumer behavior data to obtain a set of consumption events of retail goods and a set of consumption trajectories of retail areas in the unmanned retail terminal. S2: Based on the set of consumption events for the retail goods, extract the associated consumption weights between different retail goods and the individual consumption weights of retail goods using the product interaction sequence mining method; S3: Based on the set of consumption trajectories in the retail area, calculate the visibility index and interaction probability of different display positions in the unmanned retail terminal; S4: Based on the associated consumption weights between different retail products, the individual consumption weights of retail products, the visibility index of different display positions, and the interaction probability, a reward function is constructed to improve consumers' purchase intention and purchase conversion rate. The reward function is solved using an adaptive layout optimization algorithm based on reinforcement learning to obtain the retail product display layout scheme of the unmanned retail terminal.

2. The method for recommending goods in an unmanned retail terminal based on edge IoT sensing as described in claim 1, characterized in that, Step S1 involves collecting consumer behavior data using multimodal sensing devices deployed in unmanned retail terminals, including: S11: At the edge of the unmanned retail terminal, a multimodal sensing device consisting of infrared sensing sensors and RFID readers is deployed, with the infrared sensing sensors deployed in front of the unmanned retail terminal and in the merchandise retail channel inside the unmanned retail terminal. S12: The infrared sensing sensor is used to collect the dwell infrared signal stream representing the dwell state and the consumption infrared signal stream representing the consumption behavior. The infrared sensing sensors deployed in front of the unmanned retail terminal and in the merchandise retail channel inside the unmanned retail terminal collect the dwell infrared signal stream and the consumption infrared signal stream in sequence. S13: The RFID reader is used to collect the label change status of retail product labels in the unmanned retail terminal in real time, and the label change status of the retail product labels is constructed into an RFID signal stream; S14: The dwell infrared signal stream, consumption infrared signal stream and RFID signal stream are used as consumer behavior data.

3. The method for recommending goods in an unmanned retail terminal based on edge IoT sensing as described in claim 2, characterized in that, Step S1, which involves extracting consumption events for retail goods and consumption trajectories for retail areas from the consumer behavior data, also includes: S15: Based on the dwell infrared signal stream, calculate the signal energy change rate at the signal acquisition time in the dwell infrared signal stream, identify the initial dwell time when the consumer arrives at the unmanned retail terminal area and the end dwell time when the consumer leaves the unmanned retail terminal, and take the time period between the initial dwell time and the end dwell time as a consumption event and the time range corresponding to a consumption trajectory. S16: Based on the RFID signal stream, extract the set of retail goods whose tag status changed to disappear during the time period between the initial time of stay and the end time of stay, where the tag status changing to disappear indicates that the retail goods have been purchased; take the time period between the initial time of stay and the end time of stay as the event range of the consumption event, take the set of retail goods as the consumed goods of the consumption event, constitute a consumption event, construct all the currently extracted consumption events into a set of consumption events of retail goods in the unmanned retail terminal, and synchronously record the time when the tag status of the retail goods in the consumption event changes to disappear; S17: Based on the consumer infrared signal stream, calculate the signal energy change rate at the signal acquisition time in the consumer infrared signal stream, and count the number of signal acquisition times where the absolute value of the signal energy change rate is higher than the preset change rate threshold. Use the number of signal acquisition times as the number of operations performed by the consumer to take and put back retail goods. S18: Extract the set of retail goods in the consumption event corresponding to the time period between the initial time of stay and the end time of stay, sort the retail goods in the set of retail goods according to the order in which the tag status changes to disappear, and obtain the display positions of the sorted retail goods to form a display position sequence. S19: The number of signal acquisition times and the display bit sequence are taken as a consumption trajectory, and all the currently extracted consumption trajectories are constructed as a set of consumption trajectories in the retail area of ​​the unmanned retail terminal.

4. The method for recommending goods in an unmanned retail terminal based on edge IoT sensing as described in claim 1, characterized in that, Step S2, which uses product interaction sequence mining to extract the correlation consumption weights between different retail products and the individual consumption weights of retail products, also includes: S21: Based on the set of consumption events for the retail goods, the time period between the initial time of stay and the end time of stay associated with the consumption event is taken as the consumption time period, and the length of the consumption time period is extracted as the consumption duration of the consumption event. The average consumption duration of all consumption events is calculated, and the weight factor of the consumption event is calculated using the average consumption duration. The retail goods in each consumption event are weighted to obtain the individual consumption weight of the retail goods. S22: Obtain the time when the label status of different retail products in the same consumption event changes to disappearance, and calculate the time correlation factor between any two retail products; S23: Obtain the number of times different retail products co-occur in a consumption event, use an improved co-occurrence frequency algorithm to calculate the preliminary correlation between any two retail products, and use the preliminary correlation to weight the preliminary correlation to obtain the associated consumption weight between retail products.

5. The edge IoT sensing-based product recommendation method for unmanned retail terminals as described in claim 4, characterized in that, In step S21, the weighting factor of the consumption event is calculated using the average consumption duration, and the retail goods in each consumption event are weighted to obtain the individual consumption weight of the retail goods. The formula for calculating the individual consumption weight is as follows: ； in, Represents the uth type of retail goods Individual consumption weight, This represents the weight factor for the m-th consumption event. This represents the duration of the m-th consumption event. Let M represent the average duration of all consumption events, and M represent the total number of consumption events. Indicates the number of types of retail goods. Represents a collection of retail goods Does the uth type of retail product exist? The discriminant function, Let m be the set of retail goods in the m-th consumption event. If the set of retail goods... There exists a type u of retail goods. ,but =1, otherwise It is 0.

6. The method for recommending goods in an unmanned retail terminal based on edge IoT sensing as described in claim 1, characterized in that, The S3 step calculates the visibility index and interaction probability of different display positions in the unmanned retail terminal, including: S31: Calculate the display weight of different display positions in the consumption trajectory according to the order in which the display positions appear in the consumption trajectory. S32: Calculate the visibility index of display positions in unmanned retail terminals by combining the indicator function value of whether the display position appears in the display position sequence and the display weight. S33: Based on the number of signal acquisition times in the consumption trajectory, the indicator function value of whether the display position appears in the display position sequence is weighted to generate the interaction probability of the display position in the unmanned retail terminal.

7. The method for recommending goods in an unmanned retail terminal based on edge IoT sensing as described in claim 1, characterized in that, Step S4 constructs a reward function to enhance consumer purchase intent and purchase conversion rate. The expression for the reward function is as follows: ； in, Represents the reward function, This represents the weighting factor of the reward function. Let represent the retail merchandise display layout scheme to be solved. This represents the reward function value corresponding to retail product display layout scheme A. Indicates the first One display space Visibility index Indicates the first One display space The probability of interaction, Indicates the number of display positions. This indicates the first step in retail merchandise display layout scheme A. One display space The retail products on display Indicates retail goods Individual consumption weight, Indicates retail goods The related consumption weight between them Indicates the u-th type of retail product. K represents the number of retail product types. This indicates the retail products in retail product display layout scheme A. Display space Indicates retail goods Display space distance constraints between them.

8. The method for recommending goods in an unmanned retail terminal based on edge IoT sensing as described in claim 7, characterized in that, Step S4 employs an adaptive layout optimization algorithm based on reinforcement learning to solve the reward function, thereby obtaining a retail product display layout scheme for the unmanned retail terminal. It also includes: S41: Initialize and generate a retail product display layout scheme as the current state; S42: Select the optimal layout adjustment action that may improve the reward function value from the pre-trained display layout adjustment action space, adjust the current state according to the selected layout adjustment action, generate the next time step state, and use the next time step state as the input value of the reward function to calculate the reward function value of the next time step state. S43: Use the Q-learning algorithm to update the Q value between the current state and the layout adjustment action. If the update magnitude is less than the preset threshold, or the number of Q value updates reaches the preset maximum number of iterations, it indicates that the algorithm has converged, and the state at the next moment is used as the retail product display layout scheme of the unmanned retail terminal obtained by solving. Otherwise, proceed to step S44. S44: Set the state of the next time step as the current state, record the number of times the Q value has been updated, and return to step S42.

9. A product recommendation system for an unmanned retail terminal based on edge IoT sensing, characterized in that, The unmanned retail terminal product recommendation system includes a data processing module, an event extraction module, a consumer mining module, and a layout scheme recommendation module, to realize the edge IoT sensing unmanned retail terminal product recommendation method as described in any one of claims 1-8.

Images