Method of inventory management for unmanned store based on regional consumer data

Abstract

A method for managing inventory in an unmanned store based on local consumer data involves the multiple consumer terminals located within a commercial area each transmitting location data of consumers obtained from the multiple consumer terminals to a central server. The multiple seller terminals each transmit product data purchased by the consumers, obtained from the multiple seller terminals provided within the store within the commercial area, to the central server. Based on the location data, the central server calculates the expected movement paths of the consumers. Based on the product data, the central server calculates the expected products likely to be purchased by the consumers along the expected movement paths. The central server transmits the expected products to at least one of the multiple seller terminals located along the expected movement path.

Description

Technology Field

[0001] The present invention relates to a method for managing inventory in an unmanned store based on local consumer data, and more specifically, to a method for managing inventory in an unmanned store based on local consumer data that enables the operation of a localized unmanned store by utilizing consumer data such as the location, movement paths, and purchase history of local consumers to efficiently manage the inventory of the unmanned store and predicting and recommending products with high sales potential. Background Technology

[0002] Conventional unmanned store operation technologies are limited to managing inventory based solely on sales data or manually replenishing stock at fixed intervals, failing to reflect regional consumer preferences or commercial district characteristics. Since this approach fails to account for consumption patterns and demand shifts that vary by store, low-selling items often remain as excess inventory, while popular products sell out early. Furthermore, unmanned stores lack on-site staff, making it difficult to monitor inventory in real-time, encourage purchases through customer interaction, or manage loss rates. Additionally, there is a lack of integrated management systems or forecast-based inventory management technologies to address these issues. The problem to be solved

[0003] One objective of the present invention is to provide a local consumer data-based unmanned store inventory management method that can simultaneously solve the problems of excess inventory and stockouts by predicting products with high sales potential based on the location, movement path, and purchase data of local consumers and reflecting this in the inventory management and operation of the unmanned store. means of solving the problem

[0004] A method for managing inventory of an unmanned store based on local consumer data according to exemplary embodiments for achieving one objective of the present invention comprises: location data of consumers obtained from a plurality of consumer terminals located within a commercial area, which are each transmitted by the plurality of consumer terminals to a central server. Product data purchased by the consumers, obtained from a plurality of seller terminals provided within a store within the commercial area, which are each transmitted by the plurality of seller terminals to the central server. Based on the location data, the central server calculates the expected movement paths of the consumers. Based on the product data, the central server calculates the expected purchase products that are highly likely to be purchased by the consumers along the expected movement paths. The central server transmits the expected purchase products to at least one of the plurality of seller terminals located along the expected movement path.

[0005] In exemplary embodiments, the seller terminal further includes transmitting a voice guidance when at least one of the plurality of consumer terminals approaches within a preset distance of the seller terminal located on the expected travel path.

[0006] In exemplary embodiments, the expected purchase products are calculated based on product characteristics, product sales cycle, product price range, product category, shelf life, delivery feasibility, and delivery difficulty.

[0007] In exemplary embodiments, the plurality of seller terminals are provided within the store and receive motion data tracked through real-time motion positioning correction technology from product samples having a location sensor and a gyroscope sensor, and transmit this data to the central server, and the central server calculates the positional movement amount and rotation amount of each of the product samples from the motion data, and the central server calculates the expected purchase items such that the greater the positional movement amount and the rotation amount, the higher the probability of purchasing the product corresponding to the product samples.

[0008] In exemplary embodiments, the central server calculates the predicted movement paths through a first artificial intelligence algorithm having a conditional variational autoencoder-based path prediction algorithm, and calculates the predicted purchase products through a second artificial intelligence algorithm having a session-based transformer recommendation algorithm. Effects of the invention

[0009] According to exemplary embodiments, a local consumer data-based unmanned store inventory management method comprises: transmitting location data of consumers obtained from multiple consumer terminals located within a commercial area to a central server, wherein multiple consumer terminals each transmit to a central server the product data purchased by the consumers obtained from multiple seller terminals provided within a store within the commercial area, wherein the central server calculates the expected movement paths of the consumers based on the location data, wherein the central server each calculates the expected purchase products that are highly likely to be purchased by the consumers on the expected movement paths based on the product data, and wherein the central server transmits the expected purchase products to at least one of the multiple seller terminals located on the expected movement path.

[0010] Accordingly, the above-described regional consumer data-based unmanned store inventory management method can predict the movement paths of consumers by integrally collecting and analyzing the location data and purchase data of consumers in the region on the central server, and can calculate products with a high probability of purchase along those paths and automatically transmit them to the seller terminals. The sellers operating the stores can reflect the actual purchasing patterns and changes in demand of consumers in real time, and can simultaneously reduce excess inventory and stockouts by efficiently adjusting inventory. By prioritizing the placement of products with a high probability of sale instead of products with low sales efficiency, the inventory turnover rate can be improved, and the unmanned store operator or headquarters can automatically manage customized product configurations for each region without separate human intervention.

[0011] Furthermore, the aforementioned regional consumer data-based unmanned store inventory management method can quantitatively analyze product interest and customer dwell patterns by utilizing motion data and position correction data obtained from sensors equipped on product samples. Through this, product placement within the store can be optimized by predicting consumers' purchasing intentions, and promotion or event strategies for low-sales products can be automatically generated. Centralized management through an integrated control room enables real-time remote monitoring of the store's security status, inventory fluctuations, and customer responses; this improved management efficiency simultaneously enhances the operational stability and profitability of the unmanned store.

[0012] However, the effects of the present invention are not limited to the effects mentioned above and may be extended in various ways without departing from the spirit and scope of the present invention. Brief explanation of the drawing

[0013] FIG. 1 is a schematic diagram illustrating a local consumer data-based unmanned store inventory management method according to exemplary embodiments. FIG. 2 is a flowchart illustrating a local consumer data-based unmanned store inventory management method according to exemplary embodiments. Specific details for implementing the invention

[0014] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

[0015] In each drawing of the present invention, the dimensions of the structures are depicted enlarged compared to the actual dimensions for the clarity of the invention.

[0016] In the present invention, terms such as "first," "second," etc., may be used to describe various components, but said components should not be limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0017] The terms used in this invention are used merely to describe specific embodiments and are not intended to limit the invention. The singular expression includes the plural expression unless the context clearly indicates otherwise. In this application, terms such as "comprising" or "having" are intended to indicate the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0018] With respect to the embodiments of the present invention disclosed in the text, specific structural or functional descriptions are provided merely for the purpose of explaining the embodiments of the present invention, and the embodiments of the present invention may be implemented in various forms and should not be interpreted as being limited to the embodiments described in the text.

[0019] That is, the present invention is capable of various modifications and may take various forms, and specific embodiments are illustrated in the drawings and described in detail in the text. However, this is not intended to limit the present invention to the specific disclosed forms, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention.

[0020] FIG. 1 is a schematic diagram illustrating a method for managing inventory of an unmanned store based on local consumer data according to exemplary embodiments. FIG. 2 is a flowchart illustrating a method for managing inventory of an unmanned store based on local consumer data according to exemplary embodiments.

[0021] Referring to FIGS. 1 and 2, first, the local consumer data-based unmanned store inventory management method (20) can be performed by the local consumer data-based unmanned store inventory management system (10). The local consumer data-based unmanned store inventory management method (20) can transmit location data of consumers obtained from multiple consumer terminals (200) located within a commercial area to the central server (100) respectively (S110).

[0022] In exemplary embodiments, the local consumer data-based unmanned store inventory management method (20) can intelligently manage the inventory of the unmanned store based on various consumer data such as the location, movement path, and purchase history of local consumers, and predict products with a high probability of being purchased by consumers, thereby enabling efficient inventory management and sales strategies. Through this, the store operator can automatically adjust the product composition to reflect regional consumption trends, prevent excess inventory of products with low sales rates, and minimize stockouts of popular products. The local consumer data-based unmanned store inventory management method (20) can establish a data-based predictive management system by analyzing data in real time through a network connecting multiple consumer terminals (200) and multiple seller terminals (300) via a central control system, thereby moving away from the existing manual and experience-dependent inventory management method.

[0023] In exemplary embodiments, a plurality of consumer terminals (200) operating within the commercial area may periodically acquire location data with time information attached through a location measurement function installed in each terminal or by an event trigger. The acquisition period of the location data may be variably set considering battery consumption and network costs, for example, the sampling period may be shortened only when the movement speed exceeds a certain threshold, or conversely, the sampling period may be extended when a stationary state persists. The plurality of consumer terminals (200) may load the acquired location data into a transmission queue and then transmit it to a central server (100) using a predefined communication protocol, wherein the transmission may be performed in a real-time streaming method, a batch upload method, or a combination of both methods.

[0024] To prepare for cases where packet loss or network disconnection occurs during transmission, each location data may include a sequence identifier and a checksum, and a retransmission policy (Retry with exponential backoff) may be applied. Pre- and post-filtering based on quality indicators (e.g., satellite visibility, signal strength, consistency with inertial sensors) may be applied to remove duplicate location data or low-quality samples from the same consumer terminal (200). For privacy protection, the terminal's unique identifier may be unilaterally hashed or replaced with a pseudo-random number-based pseudonym identifier, and excessive precision may be mitigated through a snapping technique for a predetermined spatial and temporal grid. Multiple consumer terminals (200) may undergo a user consent procedure before transmitting location data, and the scope of data sharing may be selected according to the terminal settings. The central server (100) may construct a commercial area-level movement trace by performing timestamp normalization, coordinate system transformation, outlier removal, and terminal-specific session merging on the received location data, which can be utilized for path prediction and purchase probability estimation in subsequent stages.

[0025] The aforementioned commercial area refers to a geographical range established around a point where an unmanned store is located or a group of multiple stores, and can be defined using administrative districts, radius-based buffers, or network distance indicators such as walking and vehicle access times. The commercial area can be dynamically expanded or contracted based on changes in pedestrian traffic by time of day, traffic conditions, and event hosting information, and can be classified into multiple layers based on store types and the purchasing zone characteristics of the products handled. The boundaries of the commercial area can be represented as a map tile grid or polygons, and the inclusion of location data can be determined by point-polygon inclusion operations or nearest neighbor network matching.

[0026] Multiple consumer terminals (200) may include smartphones, tablets, wearables, vehicle terminals, or other portable electronic devices capable of location measurement carried by consumers present within the commercial area. Each terminal may include at least one GNSS receiver, cellular / Wi-Fi based positioning, BLE beacon scanner, and inertial sensor (IMU), and may generate location data either independently or through a fusion method. Multiple consumer terminals (200) may collect location data through an application-level SDK or an operating system-level location service interface and transmit it to a central server (100) through a network interface. The sampling period, batch size, and transmission timing may be adaptively controlled according to the power status, network status, and user settings of each terminal.

[0027] The above location data includes latitude and longitude coordinates and timestamps as basic elements, and may optionally include elevation, velocity, acceleration, azimuth, horizontal precision (HDOP), data source, and quality indicators. The location data may be normalized to a standard geographic coordinate system for coordinate system consistency and snapped to a road / pedestrian network for map matching. For privacy considerations, coordinates may be quantized into a predetermined grid unit or replaced with cluster centroids; to prevent re-identification of a single user, only data exceeding a certain time interval may be retained, or the data may be converted and stored at an aggregation level that meets anonymity criteria. To enhance data reliability, surrounding AP / beacon scan information, inertial sensor fusion results, or map constraints may be included as auxiliary metadata.

[0028] The central server (100) may represent a computation and storage infrastructure that receives, stores, and preprocesses location data from multiple consumer terminals (200) and provides a data pipeline necessary for commercial area analysis and route prediction. The central server (100) may store and manage data in a database (110). The central server (100) may include a collection API gateway, a stream / batch collection module, a data consistency verification module, a geographic information processing engine, and a metadata management store. The central server (100) may perform deduplication, time synchronization, coordinate system unification, and outlier detection on the received location data, configure terminal-unit sessions, and then apply spatial indexing and tiling for commercial areas to ensure efficiency in subsequent analysis. The central server (100) may apply encryption during transmission and storage for security and privacy protection, and may maintain data governance through access control, logging, and audit trail functions. The central server (100) is linked with the consumer movement path calculation and expected purchase product calculation modules in subsequent stages, and may adopt a stream processing framework or a message queue-based asynchronous pipeline for real-time processing.

[0029] The central server (100) can transmit data between multiple consumer terminals (200) and multiple seller terminals (300). The central server (100) can exchange data with multiple consumer terminals (200) and multiple seller terminals (300) via a wireless communication network. The wireless communication network may utilize a global system for mobile communications.

[0030] For example, wireless communication networks can utilize wireless communication technologies such as WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi (Wireless Fidelity) Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed ​​Downlink Packet Access), HSUPA (High Speed ​​Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), and 5G (5th generation).

[0031] Next, the product data purchased by the consumers obtained from the multiple seller terminals (300) provided in the store within the above commercial area can be transmitted by each of the multiple seller terminals (300) to the central server (100) (S120).

[0032] In exemplary embodiments, a plurality of seller terminals (300) provided in the store located within the commercial area can capture the product data generated during the process of consumers selecting and paying for products in event units and transmit them to a central server (100). The transmission method may be implemented as at least one of a real-time streaming method immediately after transaction approval, a batch upload method that bundles and transmits at predetermined time intervals, and a delayed synchronization method that temporarily stores in a local buffer and retransmits in the event of network instability. During the process of transmitting the product data, a transaction identifier, a terminal identifier, and a timestamp may be included to prevent duplicate reception and order disruption, and a retransmission policy and a checksum or signature for integrity verification may be applied to prepare for packet loss.

[0033] Multiple seller terminals (300) can perform item-unit aggregation, error detection, and removal of sensitive information immediately before transmission, and the central server (100) can safely input the data into the inventory calculation pipeline by performing schema validation, timestamp normalization, terminal-specific offset correction, and duplicate removal immediately upon receipt. Correction events such as sales cancellation, returns, and price adjustments are backpropagated through correlation referencing with the original transaction identifier to ensure consistency of real-time inventory and sales indicators.

[0034] Multiple seller terminals (300) can update a local inventory snapshot based on the amount of inventory fluctuation included in the product data, and then transmit the snapshot or fluctuation log together to a central server (100). In this case, the multiple seller terminals (300) can maintain a correlation between the physical layout in the store and inventory changes by attaching a shelf location or storage zone identifier (e.g., shelf ID, bin ID, refrigerated zone ID), and in the case of expiration date-based products, can contribute to establishing a first-in, first-out strategy by including metadata such as the manufacturing date, lot, and expiration date. During a promotion period, promotional information such as event codes, discount types, and bundle configurations can be enclosed so that the central server (100) can estimate sales elasticity and display optimization. In a network and power-constrained environment, data transmission efficiency can be increased through low-power protocols or short-distance communication with a local gateway, and when multiple terminals transmit simultaneously, the store gateway can act as an edge aggregation that performs placement, compression, and deduplication.

[0035] Multiple seller terminals (300) may be devices that detect, process, and transmit product sales or inventory change events within an unmanned store. Multiple seller terminals (300) may include a kiosk-type payment terminal, a touch display including POS functions, a vending machine controller, a smart shelf, an RFID / barcode scanner module, an entrance reader, a local store gate, and an edge computing device linked thereto. Multiple seller terminals (300) may be equipped with a network interface, a security module, a local storage, and a time synchronization function, and may support multiple input modes such as single item scanning, multi-item batch scanning, and shopping cart-based scanning. In the case of refrigerated and frozen products, multiple seller terminals (300) may add a temperature and humidity sensor to generate an alarm event when storage conditions are deviated from, and may link with an entrance gate to generate an event to prevent unpaid outflow.

[0036] The above product data may be structured information generated in response to the consumer's purchasing behavior or inventory fluctuations. The above product data may include product identifiers, product names, categories, brands, specifications / capacities, unit prices, sales prices, quantities, discount / promotion information, transaction identifiers, terminal identifiers, timestamps, etc. The above product data may include lot, expiration date, manufacturing date, storage conditions, country of origin, package composition, bundle / set relationships, display location identifiers, inventory fluctuation types, inventory figures before and after fluctuation, promotion codes, coupon application information, tax codes, etc. For the protection of personal information, payment method information may be tokenized or restricted to a level of anonymization, and in the case of multiple items within the same transaction, referential integrity may be maintained through an item line structure. For data quality management, the terminal creation time and server reception time may be recorded together to enable delayed analysis, and rules for quantity / price range verification, catalog reference verification, and event conflict detection may be recorded as metadata.

[0037] Next, a plurality of seller terminals (300) are provided within the store and can receive movement data tracked through real-time movement positioning correction technology from product samples (400) having a location sensor (410) and a gyroscope sensor (420) and transmit it to a central server (100) (S130).

[0038] In exemplary embodiments, a plurality of seller terminals (300) installed in the store within the commercial area can receive motion data generated from a plurality of product samples (400) placed in front of a product displayed in the store in real time and transmit it to a central server (100).

[0039] Each of the multiple seller terminals (300) is connected via a wireless communication module to a position sensor (410) and a gyroscope sensor (420) mounted on the product samples (400), and can detect in real time the actions of a consumer holding, lifting, or rotating the product samples (400). The multiple seller terminals (300) collect raw sensor signals transmitted periodically or on an event basis from each product sample (400) and apply Real-Time Kinematic (RTK) technology to track minute changes in the sample's movement with high precision. According to this configuration, the multiple seller terminals (300) can transmit the raw data received from the sensors, including time information, sample identifiers, and signal quality indicators, to a central server (100). To ensure data accuracy, the seller terminals (300) can apply a channel separation technique to prevent interference between multiple samples, and if the reception of the Real-Time Kinematic signal is unstable, they can perform interpolation or predictive filtering to maintain data continuity. The central server (100) can synchronize the received data spatially and temporally to form basic data for analyzing the intensity and duration of interaction between consumers and samples within the store.

[0040] Product samples (400) may refer to models or actual specimens placed on the front of a display stand so that consumers can check the shape, texture, size, etc. of the actual product without directly unpacking it in an unmanned store. Each product sample (400) can detect detailed motion data, such as the number of times a consumer lifts the sample, the holding time, the direction of rotation, and the shaking pattern, through an embedded position sensor (410) and gyroscope sensor (420), and the sensor module can transmit data to a nearby seller terminal (300a) via low-power BLE communication or ZigBee protocol. The product sample (400) is equipped with a rechargeable battery or a wireless power transfer (WPT) module to enable continuous operation, and the sensor module may be embedded inside a plastic, metal, or composite housing. To prevent data quality degradation due to changes in the position of the product samples (400), a reference tag may be installed at the fixed position of the sample, and relative coordinates can be calculated during real-time motion positioning correction using this reference point.

[0041] The above-mentioned mobile positioning correction technology may refer to a technology that secures centimeter-level precision by comparing and correcting the carrier phase in real time based on satellite signals received from the position sensor (410). The mobile positioning correction technology may be applied to precisely measure minute displacements and rotational changes that occur when the product sample (400) is moved by the consumer. The seller terminals (300) act as base stations to correct phase data received from the sensors of each product sample (400), and can reduce the position error to less than a few centimeters using differential signals. By applying the above-mentioned mobile positioning correction technology, even minute movements such as the consumer holding the sample in their hand, rotating it, or lifting and putting it down can be accurately tracked, and data collection with much higher precision than traditional GPS or BLE location tracking methods may be possible. To minimize position errors caused by radio wave reflection or multipath in a store environment, the mobile positioning correction signal may be distributed and received through multiple correction antennas installed inside the store.

[0042] The above motion data refers to time-continuous motion information collected from a position sensor (410) and a gyroscope sensor (420) mounted on a product sample (400), and may include 3D position coordinates, velocity, acceleration, rotation angle, rotation speed, holding time, and event occurrence time information. Each data set may be packetized by a seller terminal (300) along with a unique identifier of the product sample and transmitted to a central server (100), and the central server (100) may reconstruct the movement trajectory and motion profile of the sample based on this. The motion data may undergo filtering and normalization processes to remove noise or discontinuous sections, and detailed motion patterns may be distinguished through temporal rate of change analysis. To increase data reliability, the motion data of each sample may include a signal-to-noise ratio, sensor calibration information, and environmental parameters at the time of data collection.

[0043] Next, based on the above location data, the central server (100) can calculate the expected movement paths of the consumers (S140).

[0044] In exemplary embodiments, the central server (100) may perform time synchronization, coordinate system unification, outlier removal, and quality indicator correction on the location data collected from a plurality of consumer terminals (200), and then construct a movement trajectory segment by matching it with the geography and store map information of the commercial area. The movement trajectory segment may be expanded with derived features such as movement start and end times, average and instantaneous speeds, rate of change of direction, duration of stay, and repeated visit patterns, and may extract representative movement templates according to time of day, day of the week, and event situation through cluster-based similar trajectory grouping. The central server (100) may receive preprocessed and standardized spatiotemporal features as input and probabilistically calculate future movement paths per consumer or cohort through a prediction model that considers space-time interaction.

[0045] The central server (100) can generate multiple sets of alternative paths in a multimodal form rather than a single path during prediction, and can calculate the probability of arrival, the distribution of expected dwell time, the order of transit zones, the path length, and the reliability of the transition point for each path. In this case, by simultaneously reflecting the local motion of the recent trajectory and cluster-level congestion and flow direction signals, a path distribution combining an individual's immediate intention and environmental constraints can be estimated. By performing sliding window-based incremental estimation on new location data flowing in as a real-time stream, the path distribution can be continuously re-evaluated, and existing predictions can be re-weighted and updated as necessary. Prediction reliability can be estimated by data quality indicators and model uncertainty, and the central server (100) can record the confidence interval, entropy, calibration error, etc. for each path to support their use in subsequent stages.

[0046] The representation format of the prediction result can be coded as a polyline sequence or a zone index sequence on a commercial area map, and a timestamp offset, expected stay, and transition probability can be added to each sequence element. If mutual exclusivity or partial overlap relationships exist between paths, the central server (100) can ensure normalization of the entire distribution by applying probability preservation constraints on the path graph.

[0047] The above predicted movement paths refer to a set of one or more candidate paths that a consumer is likely to move along in the future on a commercial area and store map, and each path may include an ordered sequence of locations, the probability of proceeding along that sequence, the expected value and variance of the time spent per segment, and the transition probability of intermediate points. Multiple predicted movement paths may coexist at the same time, and the sum of the probabilities among them may be normalized to 1. Each path is encoded in the form of a map-matched polyline or zone index for consistency of representation, and the path probability distribution may be recalculated upon the inflow of new location data when the prediction is updated.

[0048] In exemplary embodiments, the central server (100) may calculate the predicted movement paths through a first artificial intelligence algorithm. The first artificial intelligence algorithm may include a conditional variational autoencoder-based path prediction algorithm. The central server (100) may perform preprocessing on the location data flowing in from a plurality of consumer terminals (200), then combine it with map contexts such as a map of a commercial area, movement path, zone, shelf area, etc. to construct an observation trajectory in the form of a sequence, and receive this as input to calculate the predicted movement paths of consumers. In this case, the central server (100) may calculate multiple candidate paths, the probability of reaching each candidate, the expected value of the time of stay per section, the transition probability, and the prediction reliability together through the first artificial intelligence algorithm.

[0049] The first artificial intelligence algorithm may include a conditional variational autoencoder-based path prediction algorithm. The first artificial intelligence algorithm may be composed of a conditional encoder that encodes contextual information such as past observation trajectories, maps, time zones, and congestion levels; a probabilistic latent space that expresses multiple modalities (multimodality) of future movement through latent variables; and a decoder that generates a future location sequence by combining latent variables and condition information. The conditional encoder may extract time-series features (velocity, acceleration, rate of change of direction, dwell flag) of observed locations and form a condition vector by combining map-matched zone embeddings, POI embeddings of major movement paths, time zone one-hot / sine-cosine period features, and nearby population flow indicators. The decoder may receive the corresponding condition vector and sampled latent variables as input and output a coordinate distribution for future multi-stage time points in stages.

[0050] Training can be performed using training samples generated from historical data, with incremental lower bound maximization or KL divergence minimization combined with an equivalent reconstruction loss. The reconstruction loss uses the log-likelihood between the predicted coordinate distribution and the actual coordinates, and the KL term can be weighted to suppress divergence between the prior and posterior distributions of the latent variable. To stabilize training, KL annealing, teacher forcing, scheduled sampling, gradient clipping, label smoothing, calibration-assisted loss (temperature scaling), etc., can be applied.

[0051] During inference, multiple candidate paths can be generated by multi-sampling latent variables, and the likelihood or posterior probability of each path can be calculated to select and normalize the top K paths. Each candidate path is encoded as a polyline or zone index sequence and may include meta-information such as the probability of arrival at each time point, the expected stay at each segment, the path length, the reliability of the turning point, and the prediction entropy. The central server (100) can recalculate path probabilities by performing sliding window-based incremental updates whenever new location data is received, and, if necessary, apply exponential decay that weights recent data. If the data distribution changes structurally, the distribution drift detection metric can trigger model retraining or fine-tuning of the adapter layer.

[0052] The first artificial intelligence algorithm described above can explicitly calculate prediction uncertainty. Specifically, it can express random uncertainty (aletoric) using the covariance of the distribution output by the decoder, and provide confidence intervals for each path by approximating model uncertainty (epistemic) through dropout sampling or ensemble. The uncertainty information can be utilized for ranking between paths, reweighting in subsequent steps, and risk management during path selection. For computational efficiency, the server supports batch inference and microbatching, and can apply model lightweighting to ensure inference time within a latency threshold.

[0053] Next, the central server (100) can calculate the positional displacement and rotational amount of each of the product samples from the above movement data (S150).

[0054] In exemplary embodiments, the central server (100) can calculate the positional movement amount and the rotation amount by aligning the time series for each product sample (400) after time synchronization and missing data correction for the movement data collected and transmitted through multiple seller terminals (300). The positional coordinates obtained from the position sensor (410) are normalized according to coordinate system unification and map matching (display stand reference coordinate system), and a low-pass filter or Kalman / extended Kalman filter may be applied to remove sensor noise and spikes. The central server (100) calculates the total distance traveled by accumulating the positional vector difference between consecutive time points, and can also calculate the Euclidean distance between the start and end positions from the display stand reference point, the distance traveled per segment, and the ratio of movement to the dwelling segment. When an RTK correction signal is used, positional accuracy of several centimeters can be secured by applying a carrier phase-based correction value, and continuity can be maintained through inertial walking correction and interpolation in the event of temporary loss of the correction signal.

[0055] The central server (100) can calculate the change in angle between time points by time-integrating the angular velocity collected from the gyroscope sensor (420), and calculate the accumulated rotation amount while avoiding gimbal lock through quaternion-based attitude estimation. The raw angular velocity can undergo static interval calibration and Allan variance-based noise parameter estimation for bias removal and scale correction, and if gravity vectors and magnetic sensors are present, attitude alignment to the storage reference coordinate system can be performed using the corrected gravity and magnetic references. The central server (100) can calculate the total rotation amount through the integration of the rotation angle (absolute value) for each time interval, and can calculate the rotation amount per axis, maximum rotation speed, rotation-stop transition point, and significant rotation threshold relative to micro-vibration. The central server (100) can jointly analyze linear displacement and rotational displacement of the same interval to record reliability indicators for position displacement and rotation amount, outlier flags, and summary statistics per segment so that motion patterns such as lifting, tilting, and rotation observation can be reliably distinguished.

[0056] Next, based on the product data, the central server (100) can calculate the expected purchase products that are highly likely to be purchased by the consumers on the expected movement paths (S160).

[0057] In exemplary embodiments, the central server (100) can calculate expected purchase products with a high probability of purchase for each consumer or cohort unit by reflecting the exposure context on the expected movement paths of the consumers based on the product data. Specifically, the central server (100) can estimate the purchase probability for each product and calculate the ranking of candidate products based thereon by combining path context features, such as the probability of reaching each expected movement path, zones on the path, shelf exposure status, and the expected time of stay in the corresponding section, with attribute features of the product data. In this case, the prediction results for each product may include not only the purchase probability but also a reliability indicator and a prediction validity window, and if conflicts occur between multiple candidates within the same path, the priority can be readjusted by applying weights for the probability of reaching and the time of stay. If the path is calculated as multiple alternatives, the central server (100) can calculate the final purchase probability by weighting the probability of reaching each path and apply conservative weights to sections with large prediction deviations to mitigate overfitting.

[0058] In exemplary embodiments, the expected purchase products may be calculated based on product characteristics, product sales cycle, product price range, product category, shelf life, delivery feasibility, and delivery difficulty.

[0059] The above product characteristics can be subdivided into capacity, packaging type, taste, scent, material, seasonality, storage conditions, etc., and each characteristic can be weighted through interaction terms with past sales responses and exposure contexts along the channel. The above product sales cycle may include repetition in daily, weekly, or monthly units, temporary elasticity of promotion periods, and repurchase interval distribution, and the central server (100) can adjust the possibility of purchase at specific points in time by combining the said periodicity and the time-of-the-week and day-of-the-week patterns of the channel. The above product price range may include price elasticity, relative price compared to competing products, and whether bundles / sets are included, and can be reflected by increasing the weight of low-involvement and low-price impulse purchase items in sections with short expected dwell times along the channel, and increasing the weight of comparative and deliberative items in sections with long dwell times. The above product categories can model mutual substitution and complementary relationships using hierarchical classifications such as snacks, beverages, stationery, and household goods, and can form a set of candidates by considering internal competition within the same category and cross-selling possibilities together.

[0060] The above expiration date may include remaining days, turnover rate, history of deviation from storage conditions, etc., and the central server (100) may increase the priority recommendation weight in the section with a high probability of exposure on the path for products where the remaining days are close to a threshold. The above delivery probability may be calculated based on supplier inventory, lead time, delivery reliability, etc., and the above delivery difficulty may include minimum order quantity, logistics cost variability, work process requirements, etc. The central server (100) may lower the priority of items with low delivery probability or high difficulty, even if the purchase probability is high, and provide candidates with high feasibility first. The central server (100) may normalize the scale of each of the above features, interpolate and mitigate omissions and outliers, and then construct a score function including an interaction term with the path context to calculate the above expected purchase product for each path, each zone, and each time window. The calculated results can be expressed as a set of feasible candidates that simultaneously satisfy product-specific purchase probabilities, expected sales and margins, and expiration date and delivery constraints, and can be immediately recalculated through sliding window-based incremental updates when new product data or path updates are introduced.

[0061] In exemplary embodiments, the central server (100) may calculate the expected purchase items such that the greater the positional movement amount and the rotation amount, the higher the probability of purchasing the product corresponding to the product samples. The central server (100) may calculate the expected purchase items by integrally analyzing the positional movement amount and the rotation amount calculated from the movement data to estimate how actively the consumer observed and manipulated the product sample (400), and by utilizing this as a key indicator related to the consumer's purchase intention. Specifically, unlike when the product sample (400) is simply picked up and placed back down while on display in a store, the positional movement amount and the rotation amount may be measured as relatively large when the consumer picks it up with their hand, rotates it from various angles, or examines it closely. The central server (100) determines that the higher the intensity of such interaction, the higher the likelihood that the consumer will take an interest in the product and form a purchase intention, and may adjust the probability of purchasing the product upward in proportion to that ratio.

[0062] The central server (100) can normalize interest based on the relative magnitude of interactions occurring within the same product group, and can ensure fairness in comparison by correcting deviations in movement amount that may occur depending on the physical size, weight, display height, etc. of each sample. For example, since larger volume products may generate a greater amount of movement with the same hand movement than smaller volume products, the central server (100) can weight the influence of the positional movement amount and the rotation amount differently according to the physical attributes of the sample. The central server (100) can analyze interaction patterns over time to distinguish between short observations driven by simple curiosity and continuous observations that are highly likely to lead to actual purchases, and can estimate the final purchase probability for each product by considering the continuous observation time and the intensity of the movement together.

[0063] The central server (100) interprets the positional movement amount and rotation amount of the product sample (400) as behavioral feedback of the consumer and calculates the probability of purchase in real time, thereby enabling more precise and immediate prediction of purchase intent than existing prediction methods based on simple sales data or click logs. The local consumer data-based unmanned store inventory management method (20) implements a human behavior-based correlation in a data analysis structure, which states that the probability of actual purchase increases as consumers observe a product for longer, more frequently, and from more diverse angles. This provides a technical effect that allows for the real-time calculation of the probability of purchase by utilizing actual consumer behavioral data in an unmanned store.

[0064] In exemplary embodiments, the central server (100) can calculate the likelihood of purchase in real time by interpreting the amount of rotation as behavioral feedback of the consumer in priority over the amount of positional movement. This reflects the fact that when a consumer shows actual interest in an object, there is a strong tendency to hold it in their hand and rotate it from various angles rather than simply moving it up and down or making short positional changes. Accordingly, the central server (100) prioritizes processing the angular velocity-based rotation signal collected from the gyroscope sensor (420), and noise caused by short micro-vibrations and shaking of the display stand can be excluded using debouncing rules based on critical time and critical angular displacement. Subsequently, normalization and bias correction using reference values ​​at the sample / category level are performed to offset differences in ease of rotation according to the physical characteristics (weight, volume, mounting height) of the sample, and drift can be suppressed by periodic attitude initialization using a reference stand complete contact event. The central server (100) uses the refined rotation amount as a core feedback, while the position movement amount calculated from the position sensor (410) is referenced in parallel as an auxiliary indicator, and can apply an adaptive weighting policy that temporarily increases the weight of the position movement amount in sections where the reliability of the rotation amount is low (degraded sensor quality, excessive congestion). In addition, repeated observation events within the same session are merged into a window aggregation, and an upper limit and a re-trigger deferral are set to prevent excessive accumulation and thus prevent over-interpretation of the signal. Finally, by combining the rotation-centered interaction intensity and observation duration to calculate the real-time purchase probability for each product and recording the reliability together, it is possible to make recommendation and exposure decisions that precisely reflect the psychological progression stages leading from interest to observation to purchase.

[0065] In exemplary embodiments, the expected purchase products may be calculated through a second artificial intelligence algorithm. The second artificial intelligence algorithm may include a session-based transformer recommendation algorithm.

[0066] Since the second artificial intelligence algorithm includes the session-based transformer recommendation algorithm, the central server (100) can encode a series of interaction histories constituting the current visit session for a specific consumer or cohort into a chronological or masked sequence, interpret the context of the session bidirectionally, and calculate product candidates with a high probability of purchase at the next point in time. In this case, the second artificial intelligence algorithm can receive static and dynamic features such as product attributes, store context, and operational constraints together as input, learn the interaction between the session context and product attributes, and calculate the priority of the candidate products. The central server (100) can add purchase probability, prediction confidence, and valid time window together to the calculated candidates to finally determine the expected purchase product that is suitable for the path context and feasible.

[0067] The above session-based transformer recommendation algorithm can effectively capture patterns of consumers re-entering a specific zone or cross-exploring different categories by modeling not only adjacent events within a session but also long-range dependencies across the session through an attention mechanism. The central server (100) can recalculate recommendation results by incrementally updating the session sequence whenever a new event is received, and can apply correction techniques such as calibration, dropout-based uncertainty estimation, and time decay weighting to prevent excessive bias. Candidates that are difficult to provide immediately due to inventory limits or delivery constraints can have their priority automatically lowered, and diversity constraints considering substitute and complementary relationships can be applied to mitigate excessive competition within the same category.

[0068] Since the first AI algorithm calculates the consumer's predicted movement paths, reach probabilities, expected stays per segment, and prediction reliability using the conditional variational autoencoder-based path prediction algorithm, additional combination effects may occur when used simultaneously with the second AI algorithm. By injecting the path-based exposure context provided by the first algorithm as an input feature of the second algorithm, recommendations can be reordered around candidates that the consumer can actually access. By weighting the reach probabilities and expected stay information from the first AI algorithm onto the output results of the second AI algorithm, a set of actionable recommendations can be constructed that considers both purchase potential and reachability. By propagating the prediction uncertainty information provided by the first AI algorithm to the decision-making of the second AI algorithm, a robust operational policy can be implemented, such as adopting conservative recommendations in segments where path prediction is unstable and applying aggressive recommendations in segments with high reliability. When the first and second AI algorithms are used in competition, it is possible to move beyond static recommendations based on simple sales history and enable predictive inventory management, display, and promotion decision-making that reflects real-time movement context.

[0069] Next, the central server (100) can transmit the expected purchase products to at least one (300a) located on the expected travel path among the plurality of seller terminals (300) (S170).

[0070] In exemplary embodiments, the central server (100) may select a path among the consumer’s expected movement paths calculated through the first artificial intelligence algorithm that has a reach probability greater than a predetermined threshold, map a plurality of seller terminals (300) arranged in correspondence with zones, shelves, and entry / exit paths on the path, and then designate at least one of them (300a) as a transmission target to transmit the expected purchase products. In this case, the central server (100) may calculate the transmission priority by comprehensively considering the reach probability per path, the expected time of stay per section, the physical location of the terminal, the terminal status, the availability of screen display slots, inventory availability, and the valid time window, and may apply a rate limit and a round-robin policy at the terminal / zone level to prevent excessive duplicate exposure of the same product.

[0071] The payload transmitted by the central server (100) may include product identifiers, categories, expected purchase probability, exposure priority, display duration, recommendation reasons, application constraints, and experiment identifiers, and may also include template or rendering instruction information to match the screen configuration, resolution, and language settings of each terminal. Transmission is performed through a secure channel, and the central server (100) receives acknowledgment (ACK), display start / end, user interaction events, etc. from the terminal to form a real-time feedback loop, and in the event of failure, may perform exponential backoff-based retries or alternative terminal swapping, and if the consumer's location is updated and the path probability changes, the transmission target terminal and the exposure time may be immediately recalculated to ensure exposure synchronized with the latest path context.

[0072] The local consumer data-based unmanned store inventory management method (20) can implement an advanced data-based inventory management system that precisely analyzes the actual behavioral patterns of consumers and predicts purchase intentions by integrally applying the conditional variational autoencoder-based path prediction algorithm, the session-based transformer recommendation algorithm, and the motion positioning correction technology. Specifically, the conditional variational autoencoder-based path prediction algorithm can probabilistically predict the movement path of each consumer based on location data collected from consumer terminals (200), calculate the expected duration of stay and the probability of arrival, and identify the flow of consumers in real time. The session-based transformer recommendation algorithm can utilize the results of this movement path prediction as input features to analyze products that consumers are likely to be interested in on the path on a session basis, and predict products that are likely to be purchased by considering product characteristics, price, expiration date, delivery availability, etc. Furthermore, the amount of position movement and the amount of rotation measured using the movement position correction technology from the product sample (400) installed in the store can quantitatively reflect the extent to which the consumer actually holds the sample in their hand and observes it, thereby allowing the consumer's actual interest to be evaluated based on behavior.

[0073] Because the above conditional variational autoencoder-based path prediction algorithm, the above session-based transformer recommendation algorithm, and the above motion positioning correction technology are combined, the central server (100) can simultaneously predict the purchase intention and the possibility of depleting inventory by comprehensively analyzing the consumer's location, path, and behavior data, and based on this, adjust the inventory in the store in real time and provide purchase inducement guidance at the optimal time and location through the seller terminal (300). Accordingly, the local consumer data-based unmanned store inventory management method (20) can learn the consumer's behavior data to maximize the operational efficiency of the unmanned store, improve sales prediction accuracy and inventory turnover rate, and at the same time personalize the consumer's purchasing experience.

[0074] Next, when at least one (200a) among the plurality of consumer terminals (200) approaches the seller terminal (300a) located on the expected movement path within a preset distance, the seller terminal (300a) can transmit voice guidance (S180).

[0075] In exemplary embodiments, a central server (100) or a seller terminal (300) sets an access zone (geofence) of a predetermined radius centered on the location of a specific seller terminal (300a) on an expected movement path calculated through the first artificial intelligence algorithm, and when at least one of a plurality of consumer terminals (200) (200a) enters within the preset distance threshold of the access zone, it is determined that the voice guidance trigger condition is satisfied, and the seller terminal (300a) can transmit voice guidance. The access determination can be performed by comprehensively considering the quality indicator of location data collected from the consumer terminal (200a), the in-store received signal strength (RSSI), beacon identification, and inertial estimation trajectory, and hysteresis, dwell time threshold, re-trigger grace time, and congestion-based weighting can be applied to prevent excessive triggers due to frequent entry or exit from the boundary surface. To suppress excessive repeated transmission to the same consumer terminal (200), a session unit upper limit, a zone unit cooldown, and a campaign unit exposure frequency limit may be set, and the volume, guidance length, and number of plays may be dynamically adjusted according to the store environment noise level, time of day, and customer distance and direction.

[0076] The above voice guidance is intended to induce purchases by consumers and may be configured to include messages such as recommended product summaries, promotion information, shelf location guidance, notifications of approaching expiration dates or limited quantities, and cross-selling proposals. The seller terminal (300) may be linked with a built-in speaker, a directional speaker, or a store PA system to ensure the quality of the guidance. The guidance language may be automatically selected in multiple languages ​​according to store settings or time of day and customer flow characteristics, and may be extended to multimodal notifications linked with screen banners, vibrations, or light signals to enhance visual and auditory accessibility. To protect personal information, the voice guidance may be designed to be transmitted using generalized phrases that do not specify individual consumers, and event logs from the guidance transmission process may be collected by a central server (100) and used for conversion rate analysis, retrigger policy optimization, and A / B experiment evaluation.

[0077] As described above, the local consumer data-based unmanned store inventory management method (20) can predict the movement paths of the local consumers by integrally collecting and analyzing the location data and purchase data of the local consumers at a central server (100), and can calculate products with a high probability of purchase along the path and automatically transmit them to seller terminals (300). The sellers operating the stores can reflect the actual purchase patterns and demand changes of the consumers in real time, and can efficiently adjust inventory to simultaneously reduce excess inventory and stockouts. By prioritizing the placement of products with a high probability of sale instead of products with low sales efficiency, the inventory turnover rate can be improved, and the unmanned store operator or headquarters can automatically manage customized product compositions for each region without separate personnel intervention.

[0078] In addition, the local consumer data-based unmanned store inventory management method (20) can quantitatively analyze the level of interest and customer dwell patterns for each product by utilizing motion data and position correction data obtained from sensors equipped on product samples. Through this, the purchasing intentions of the consumers can be predicted to optimize product placement within the store, and promotion or event strategies for products with low sales rates can be automatically generated. Centralized management through an integrated control room is possible, allowing for real-time remote monitoring of the store's security status, inventory fluctuations, and customer reactions. Through this improvement in management efficiency, the operational stability and profitability of the unmanned store can be increased simultaneously.

[0079] Although it has been described above that all components constituting an embodiment of the present invention are combined or operate as a single unit, the present invention is not necessarily limited to such an embodiment. That is, within the scope of the purpose of the present invention, all components may be selectively combined in one or more ways to operate.

[0080] The foregoing description is merely an illustrative explanation of the technical concept of the present invention, and those skilled in the art to which the present invention pertains will be able to make various modifications and variations within the scope of the essential characteristics of the present invention. Accordingly, the embodiments disclosed in the present invention are intended to explain, not limit, the technical concept of the present invention, and the scope of the technical concept of the present invention is not limited by these embodiments. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present invention. Explanation of the symbols

[0081] 10: Unmanned Store Inventory Management System 20: Unmanned Store Inventory Management Methods 100: Central server 200: Consumer terminals 300: Seller terminals 400: Product samples 410: Position sensor 420: Gyroscope sensor

Claims

Claim 1 Location data of consumers obtained from multiple consumer terminals located within a commercial area is transmitted by each of the multiple consumer terminals to a central server; product data purchased by the consumers obtained from multiple seller terminals provided within a store in the commercial area is transmitted by each of the multiple seller terminals to the central server; the central server calculates the expected movement paths of the consumers based on the location data; the central server calculates the expected purchase products that are highly likely to be purchased by the consumers on the expected movement paths based on the product data; the central server transmits the expected purchase products to at least one of the multiple seller terminals located on the expected movement path; and the multiple seller terminals receive motion data tracked through real-time motion positioning correction technology from product samples provided within the store and equipped with a location sensor and a gyroscope sensor, and transmit it to the central server; A method for managing inventory in an unmanned store based on local consumer data, comprising: a central server calculating the positional movement amount and rotation amount of each of the product samples from the above movement data; the central server calculating the expected purchase items such that the greater the positional movement amount and rotation amount, the higher the probability of purchasing the product corresponding to the product samples; the central server calculating the expected movement paths through a first artificial intelligence algorithm having a conditional variational autoencoder-based path prediction algorithm; and calculating the expected purchase items through a second artificial intelligence algorithm having a session-based transformer recommendation algorithm. Claim 2 A method for managing inventory of an unmanned store based on local consumer data according to claim 1, further comprising the seller terminal transmitting a voice guidance when at least one of the plurality of consumer terminals approaches within a preset distance of the seller terminal located on the expected movement path. Claim 3 In claim 1, the method for managing inventory of an unmanned store based on local consumer data, wherein the calculation of the expected purchase products is based on product characteristics, product sales cycle, product price range, product category, expiration date, delivery feasibility, and delivery difficulty.

Images