Edge-optimized retail analytics system to support store-level decision-making.

Abstract

An edge-optimized retail analytics system for autonomous in-store decision support, consisting of: an edge computing unit housed in a robust, thermally conductive enclosure and configured for use in a retail store environment; System-on-Module (SoM) mounted on a multilayer printed circuit board (PCB), wherein the SoM comprises a multi-core central processing unit (CPU) configured to manage sensor data orchestration and rule-based inference processing, and a neural processing unit (NPU) configured to perform deep neural network inference operations in real time; a high-bandwidth volatile memory module electrically connected to the SoM to buffer time-aligned multimodal sensor streams; a non-volatile solid-state drive (SSD) configured to persistently store AI model weights, inference outputs, and business-specific event logs; a power management circuit integrated on the printed circuit board, configured to regulate the input voltage of a Power-over-Ethernet (PoE) line; and a sensor interface bus that is connected to a variety of sensor modules, including visible spectrum cameras, thermal imaging sensors, passive infrared motion sensors, RFID readers, and load cell-based shelf weight sensors.

Description

Technical field

[0001] The present invention relates to the field of retail data analysis and edge computing, and in particular to a hardware-integrated, edge-optimized system configured for in-store data acquisition, local processing, real-time analysis, and actionable decision support for retail operations at the store level. BACKGROUND

[0002] Traditional retail analytics systems rely heavily on a centralized cloud infrastructure. This results in latency, connectivity dependencies, and high bandwidth costs when processing massive data streams from stores, such as customer movement, shelf levels, transaction logs, and environmental metrics. The growing need for store-level decision autonomy—especially in geographically distributed retail environments or those with limited connectivity—demands a new class of systems that can operate locally, process data at the network edge, and deliver real-time insights without relying on a continuous cloud connection. Furthermore, retailers face the challenge of integrating disparate sensors, real-time demand signals, and multimodal input sources into a unified, edge-ready analytics framework.Therefore, a system is urgently needed that combines sensor-level data acquisition, edge inference engines, and localized decision support modules in a hardware-integrated analytics structure for the retail sector.

[0003] In the evolving retail landscape, the application of data-driven decision-making is crucial for improving customer experience, optimizing inventory management, and increasing operational efficiency. Traditional retail analytics systems predominantly utilize centralized cloud infrastructures for data aggregation, processing, and decision-making. These systems operate by transferring large volumes of data collected by in-store sensors, point-of-sale terminals, loyalty programs, and CCTV units to remote cloud servers that host data lakes and analytics engines. After processing, the insights are sent back to the stores for operational action.While these solutions have driven the growth of enterprise-wide business intelligence systems, they are increasingly proving inadequate to meet the differentiated real-time requirements of branch operations, particularly in geographically distributed environments or environments with limited infrastructure.

[0004] A fundamental limitation of traditional cloud-based retail analytics architectures lies in their reliance on a high-bandwidth, low-latency internet connection. For every event—be it a customer entering a store, an item being taken from a shelf, or a shelf nearly empty—raw or partially processed data must be transmitted to central servers. This incurs significant data transfer overhead and introduces latency that can impact the timeliness of operational decisions. For example, by the time a stockout alert is generated in the cloud and sent back to the store, the opportunity for just-in-time restocking or proactive sales recovery may already be lost.The latency problem is exacerbated in environments with unreliable internet connections or when data privacy regulations restrict the transmission of customer-identifying information outside the store. Furthermore, round-trip communication in the data center often exceeds acceptable thresholds for applications requiring a sub-second response, such as targeted, customer-behavior-driven in-store promotions.

[0005] Furthermore, centralized analytics systems are designed for scalability rather than contextual specificity. While they can effectively identify macro trends across a chain, they often fail to capture the unique characteristics of individual store environments, such as local customer preferences, store layout constraints, real-time shelf conditions, and fluctuations in foot traffic. Store managers are therefore frequently disconnected from the analytics cycle or reliant on delayed reports that do not immediately reflect operational realities. Even when stores are provided with dashboards and KPIs from centralized analytics, these results are often too general to enable granular interventions at the store level. Consequently, the value of data-driven insights is diluted, and the potential of retail analytics remains underutilized precisely where customer interaction takes place.

[0006] Another challenge with existing solutions is the lack of interoperability between the various hardware and software systems in the retail sector. Retailers often use a mix of legacy systems—such as standalone inventory management software, analog monitoring systems, and isolated environmental sensors—and newer IoT-enabled platforms. Integrating these disparate sources into a unified analytics pipeline requires significant IT effort, middleware customization, and ongoing maintenance. Cloud-based platforms struggle to normalize and contextualize multimodal data streams from disparate sources with inconsistent formats, sampling rates, and reliability.The result is that valuable signals – such as the relationship between temperature fluctuations and spoilage of perishable goods or between customer dwell time and sales completion – are either lost or recognized too late to initiate preventive measures.

[0007] Furthermore, the high costs of cloud computing services and the associated data transfer fees represent a financial burden for large retailers, especially those with hundreds or thousands of stores. Continuously streaming high-resolution video feeds, RFID telemetry, and sensor data to the cloud not only consumes significant bandwidth but also incurs ongoing costs for storage, compute cycles, and analytics services. In many cases, the cost-benefit ratio of cloud-based analytics becomes unfavorable, particularly for low-margin retailers who require cost-effective technology solutions to remain competitive. While some vendors have attempted to introduce hybrid solutions with partial edge computing capabilities, their functionality is often limited and requires significant manual configuration to adapt them to specific store operations.

[0008] Efforts to address these issues through localized compute nodes or gateways have been only partially successful. Retail edge gateways, in their current form, are typically designed as simple protocol converters or local data buffers that perform basic preprocessing tasks such as data filtering or format conversion before forwarding the data to the cloud. While this reduces bandwidth consumption to some extent, it does not eliminate the architectural bottlenecks caused by remote processing. These devices rarely possess the computing power or AI inference engines required for real-time analytics or in-store decision-making.Furthermore, they lack standardized support for model deployment, lifecycle management, and dynamic orchestration – features that are essential for implementing advanced AI workflows such as anomaly detection, time series forecasting, or real-time object detection.

[0009] Data privacy and data security concerns further complicate the use of centralized analytics solutions for retailers. Regulations such as the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA), which mandate strict controls on personal data, increasingly restrict retailers in the collection, storage, and analysis of customer data. Centralized systems that transfer raw or enriched behavioral data to remote servers expose retailers to compliance risks, especially if encryption, access control, and data anonymization mechanisms are not implemented consistently. Edge-based alternatives, which process data locally and transfer only non-identifiable summaries, are considered more compliant but do not yet offer the full range of analytics capabilities that modern retailers require.

[0010] Finally, existing retail analytics platforms are largely reactive. They are designed to analyze historical data rather than provide predictive or prescriptive decision support. This limits their usefulness in dynamic store environments where conditions can change rapidly due to increased foot traffic, product demand spikes, or competitor promotions. The inability to adapt quickly to such changes undermines the strategic value of analytics, leading to missed sales opportunities, suboptimal resource allocation, and reduced customer satisfaction. Even when predictive capabilities are available, they are typically cloud-based and require extensive batch processing pipelines that do not allow for real-time inference.

[0011] The current generation of retail analytics systems—dominated by centralized cloud architectures—has significant drawbacks in terms of latency, cost, lack of contextual granularity, integration complexity, and regulatory risks. While these systems enable macro-level insights and long-term planning, they fail to provide store-level decision-makers with real-time, context-specific information. The lack of edge-native processing capabilities and limited support for AI-driven reasoning in data generation represent a fundamental limitation of the current paradigm.There is an urgent need for a new class of solutions that decentralize analytics at the branch level, integrate seamlessly with heterogeneous data sources, operate autonomously during connection interruptions, and provide retailers with precise, timely, and contextual decision support. SUMMARY

[0012] The present invention describes an edge-optimized retail analytics system implemented as a compact, modular, and robust hardware device for in-store use. The system comprises integrated sensor units, edge computing modules, AI inference accelerators, decision support controllers, and a multi-layered data bus—all housed in a temperature-controlled, dustproof enclosure suitable for use in diverse retail environments. Using pre-trained AI models, the system performs real-time analysis of customer behavior, shelf utilization, inventory movements, sales data, and environmental factors directly at the store level. Decision support results are generated locally and include restock alerts, heat maps, planogram compliance feedback, predictive reordering suggestions, and personalized advertising triggers.The system supports asynchronous synchronization with a central retail dashboard, provided connectivity allows, while ensuring continuous autonomous operation during offline periods.

[0013] The main objective of the present invention is to provide an edge-optimized retail analytics system that enables real-time decision support at the store level without relying on continuous cloud connectivity. The invention aims to decentralize data processing by integrating a high-performance, hardware-based edge computing unit directly into the retail environment. This eliminates the latency associated with remote analytics and enables immediate feedback loops for operational actions such as inventory replenishment, planogram computing, customer traffic heatmap creation, and customer retention optimization.A further objective of the invention is to create a unified, hardware-based platform that seamlessly interacts with various sensors, video cameras, RFID readers, and point-of-sale systems in the branches, and to process the resulting multimodal data streams locally using embedded AI inference engines. This allows the system to gain actionable insights with minimal data transmission overhead while ensuring high responsiveness and operational autonomy.

[0014] Another objective of the invention is to provide context-related, branch-specific information. The device can adapt and optimize its analytical models based on local trends. This supports hyper-personalized advertising triggers and predictive inventory analytics that reflect the sales velocity and individual customer behavior of each branch in real time. Furthermore, the invention aims to address data privacy and compliance challenges by ensuring that personal data is processed and anonymized at the network edge, and that only aggregated insights are selectively synchronized with cloud platforms when needed. This supports compliance with legal regulations such as GDPR and CCPA while simultaneously reducing the security risks associated with centralized data processing.

[0015] Another objective of the invention is to create a modular and robust device architecture that can be used in a variety of retail environments – from small individual stores to large supermarkets. It supports scalable, mesh-based networks between multiple devices operating within a single store or across multiple stores. The invention also aims to enable simplified installation, configuration, and model deployment via a local user interface and a remote over-the-air update mechanism, thereby minimizing the technical effort required by store staff and IT personnel. Furthermore, the invention provides for the integration of self-diagnostic and failover mechanisms that maintain system availability, manage thermal and computational load balancing, and ensure consistent operation even under adverse power or network conditions.Through these integrated functions, the invention ultimately aims to transform traditional retail operations into intelligent, self-optimizing environments that enable data-driven decisions at the physical edge, thereby increasing operational efficiency, customer satisfaction, and profitability. BRIEF DESCRIPTION OF THE FIGURE

[0016] These and other features, aspects, and advantages of the present invention will be better understood if the following detailed description is read with reference to the accompanying drawing, in which the same symbols consistently represent the same parts. The following applies: Fig. Figure 1 shows a block diagram of an edge-optimized retail analytics system for store-level decision support.

[0017] Experts will also recognize that the elements in the drawing are shown for the sake of simplicity and are not necessarily to scale. For example, the flowcharts illustrate the process by highlighting the main steps to enhance understanding of the aspects of this disclosure. Furthermore, with regard to the design of the device, one or more components of the device may be represented in the drawing by conventional symbols, and the drawing may show only the specific details relevant to understanding the embodiments of this disclosure, so as not to clutter the drawing with details that are readily apparent to those skilled in the art after reading this description. Detailed description of the invention

[0018] For a better understanding of the inventive principles, reference is made below to the embodiment shown in the drawing, which is described in specific terminology. However, this does not limit the scope of the invention. Changes and further modifications of the illustrated system, as well as further applications of the inventive principles, are possible, as would normally occur to a person skilled in the art in the field of invention.

[0019] It is clear to the person skilled in the art that the preceding general description and the following detailed description are exemplary and explanatory of the invention and are not intended as a limitation of it.

[0020] References in this specification to “an aspect”, “another aspect”, or similar expressions mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, occurrences of the expressions “in one embodiment”, “in another embodiment”, and similar expressions in this specification may all refer to the same embodiment, but need not.

[0021] The terms "includes," "include," or other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that includes a list of steps may not only contain those steps but may also include other steps not expressly listed or inherent in such process or method. Likewise, the statement "includes..." in the case of one or more devices, subsystems, elements, structures, or components does not, without further limitations, preclude the existence of other devices, subsystems, elements, structures, components, or additional devices, subsystems, elements, structures, or components.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by a person skilled in the art in the field of the invention. The system, methods, and examples provided here serve only for illustration and are not to be construed as a limitation.

[0023] Embodiments of the present disclosure are described in detail below with reference to the attached drawing.

[0024] In Fig.Figure 1 shows a block diagram of an edge-optimized retail analytics system for store-level decision support. The system 100 comprises: an edge computing unit (102) housed in a rugged, thermally conductive enclosure and configured for use in a retail store; a system-on-module (SoM) (104) mounted on a multilayer printed circuit board (PCB), the SoM comprising a multi-core central processing unit (CPU) configured for orchestrating sensor data and rule-based inference processing, and a neural processing unit (NPU) (104a) configured for real-time deep neural network inference operations; a high-bandwidth volatile memory module (106) electrically connected to the SoM to buffer time-aligned multimodal sensor streams;a non-volatile solid-state drive (SSD) (108) configured for persistent storage of AI model weights, inference outputs, and store-specific event logs; an on-PCB power management circuit (110) configured to regulate the input voltage from a Power-over-Ethernet (PoE) line; and a sensor interface bus (112) connected to a variety of sensor modules, including visible spectrum cameras, thermal imaging sensors, passive infrared motion sensors, RFID readers, and load cell-based shelf weight sensors, the edge computing unit further configured to execute a multi-threaded orchestration layer (114) to perform parallel data acquisition, feature extraction, object tracking, and local event-driven actuation without relying on a cloud-based infrastructure.

[0025] In one embodiment, the edge computing unit (102) comprises a hardware-isolated coprocessor configured to perform secure model parameter updates. The coprocessor includes a Trusted Execution Environment (TEE) implemented on a dedicated ARM TrustZone or secure enclave, wherein the TEE is programmed to receive encrypted gradient updates via a federated learning protocol and to decrypt, validate, and integrate these updates into the locally stored AI model, while preventing direct memory access by insecure processes.

[0026] In one embodiment, the sensor interface bus (112) comprises a modular I 2C and SPI hybrid architecture that supports time-synchronized polling of heterogeneous sensor types, with the interface managed by a programmable real-time clock controller that aligns the input signals across all sensor channels to a synchronization threshold of 3 milliseconds to ensure a time-coherent input for downstream inference.

[0027] In one embodiment, the NPU (104a) is configured to execute quantized Convolutional Neural Networks (Q-CNNs) for low-latency inference in object detection tasks, wherein the Q-CNNs are compiled via a vendor-specific optimization pipeline and stored in intermediate representation formats of TensorRT or OpenVINO within the SSD, and wherein the NPU dynamically loads and removes model segments based on a usage priority matrix calculated by the CPU based on the current customer frequency density in the store.

[0028] In one embodiment, the edge computing unit (102) further comprises a passive thermal management subsystem comprising a heat spreader plate made of anodized aluminum, which is thermally bonded to the SoM and the memory modules by means of a thermally conductive adhesive, wherein the heat spreader is mechanically coupled to a finned heat sink structure integrated into the housing, and wherein thermal profiles are monitored by integrated thermistors and used to throttle the CPU and NPU clock frequencies during high-inference workloads to prevent temperature overshoot above 80 °C.

[0029] In one embodiment, the solid-state drive (108) comprises a separate logical partition that is encrypted with AES-256 and provided as a read-only file system during the inference process. The partition is configured to store pre-trained model parameters and AI rule sets and further includes an update daemon that performs an authenticated model exchange only after verification of a cryptographic signature by a central controller via a secure MQTT protocol.

[0030] In one embodiment, the multithreaded orchestration layer (102a) comprises an event-driven pipeline manager implemented using an asynchronous event loop. The pipeline manager is configured to allocate computational resources to image inference tasks, product identification tasks, and inventory outage detection tasks based on a token-bucket priority scheduler, with the scheduler dynamically rebalancing the processing threads based on real-time metrics such as frame processing rate, memory usage, and event queue length.

[0031] In one embodiment, the edge computing unit (102) is also connected via a GPIO expansion module to an array of LED-based shelf displays, and the displays are triggered in response to localized stockout detection events identified by a comparison of the current load cell weight and the predicted planogram baseline, with the triggering being controlled by a low-latency rule set encoded in a local inference engine implemented using a lightweight ONNX runtime library.

[0032] In one embodiment, the edge-optimized retail analytics system is based on a modular edge computing unit housed in a ruggedized enclosure to withstand the environmental fluctuations of a typical retail store. The central processing platform consists of a system-on-module (SoM) mounted on a custom multilayer printed circuit board (PCB) and includes a multi-core CPU and a dedicated neural processing unit (NPU). The CPU is responsible for orchestrating sensor data acquisition, executing low-level control logic, and coordinating the inference pipeline, while the NPU accelerates deep learning inference for computer vision models implemented locally on the device.

[0033] The circuit board contains a high-bandwidth DDR4 memory module for buffering multimodal data from various sensors. These include visible-light cameras for monitoring product shelves, passive infrared (PIR) motion sensors for detecting human movement, load cell-based shelf sensors for real-time product weight measurement, RFID antennas for product identity verification, and thermal imaging sensors for anomaly detection in low-light conditions. These sensors are electrically connected via a sensor interface bus, which functions as a hybrid interface. 2 The system is implemented using C and SPI architecture. To ensure coherent timing alignment between different sensors, it includes a programmable real-time clock (RTC) module that provides a timestamp granularity with a tolerance of 3 ms, which is essential for effective data fusion and inference consistency.

[0034] The software architecture is based on a multi-threaded orchestration layer that operates with an asynchronous, event-driven model. This layer is responsible for instantiating and managing multiple concurrent data pipelines corresponding to the main functional modules: object detection and tracking, planogram compliance checks, inventory forecasting, customer traffic estimation, and product recognition. Each module is assigned a token-bucket-style resource priority queue, where real-time metrics such as frame throughput, memory consumption, and event latency are used to dynamically adjust thread scheduling. For example, if the number of motion events per unit of time exceeds a threshold, the orchestration layer allocates more NPU cycles to the customer traffic estimation pipeline by evicting lower-priority visual inference tasks from memory.

[0035] The models were precompiled into optimized intermediate representations using TensorRT or Intel OpenVINO. These compiled models are stored in a secure, read-only logical partition of the solid-state drive (SSD), which is encrypted with AES-256. The SSD also contains pre-computed embeddings, feature maps, and inference rule sets. A lightweight inference runtime, such as ONNX Runtime, is provided within the edge node to perform model inference on small input stacks with minimal latency.

[0036] Model updates are supported by a secure federated learning mechanism, where each edge node independently performs local training or fine-tuning with business data without transmitting raw data to the cloud. A hardware-isolated coprocessor—implemented with ARM TrustZone or an equivalent secure enclave—decrypts and integrates authenticated gradient updates received via secure MQTT channels. This ensures that model parameters can evolve based on local environmental conditions while maintaining data privacy and security.

[0037] To minimize heat buildup, the SoM and memory modules are thermally bonded via a conductive adhesive to an anodized aluminum heatsink, which in turn is connected to a finned passive heatsink integrated into the device housing. Embedded thermistors monitor the temperature in real time and relay these readings to the CPU. If temperature thresholds are exceeded, clock throttling is triggered at both the CPU and NPU levels to prevent hardware degradation or inference errors.

[0038] The decision logic for detecting and triggering events in the store is implemented via a low-latency rule engine at the CPU level. For example, if the load cell value of a shelf for a specific item falls below a predefined threshold and the vision module does not detect the product within the expected bounding box, a stockout anomaly is reported. This event is then forwarded via the GPIO expander module to trigger LED indicators on the corresponding shelf, thus notifying store staff to restock.

[0039] Finally, all logs—including event triggers, inference metadata, sensor timestamps, and system health diagnostics—are asynchronously written to a local file storage on the SSD and periodically aggregated for uploading to a central analytics dashboard using bandwidth-saving delta encoding protocols. In this way, the system provides a self-contained, privacy-friendly, and real-time edge intelligence solution for brick-and-mortar retail.

[0040] In one embodiment, the disclosed invention comprises a hardware-implemented Edge Retail Analytics Device (ERAD) designed as a compact edge processing unit and housed in a tamper-proof, industrial-grade aluminum enclosure with embedded heat dissipation fins and vibration-damping mounts. The ERAD is dimensionally optimized for mounting at shelf or ceiling height in standard shelving or kiosks and features an integrated power control subsystem for compatibility with various retail voltage standards (110 V / 220 V AC with internal DC conversion).

[0041] The device is internally divided into functional areas. A Sensor Interface Layer (SIL) comprises several I / O ports configured to receive inputs from shelf load sensors, passive infrared (PIR) motion detectors, thermal imaging cameras, RFID readers, barcode scanners, and environmental sensors for measuring temperature, humidity, and ambient light. These signals are digitized using a high-resolution analog-to-digital converter module and transmitted to the Edge Analytics Core Module (EACM) via a dedicated sensor data bus.

[0042] The EACM consists of an ARM-based multi-core SoC (System-on-Chip) processor, complemented by a GPU or a dedicated TPU (Tensor Processing Unit) for accelerated AI inference. The EACM runs on an embedded Linux kernel and a custom, containerized runtime environment optimized for real-time analytics. Pre-trained deep learning models—including convolutional neural networks (CNNs) for object detection, recurrent neural networks (RNNs) for time-series demand forecasting, and clustering models for customer segmentation—are deployed locally and analyze incoming data streams without requiring external server calls.

[0043] The Decision Support Engine (DSE) acts as a policy-based inference module that overlays the analytics core. It uses rule sets, thresholds, and adaptive learning mechanisms to generate alerts and insights. For example, if a shelf weight sensor indicates that inventory has fallen below a dynamic reorder threshold predicted by sales velocity patterns, the DSE triggers a restock alert and optionally notifies store staff via a local display or audio alert. Additionally, the system stores an on-device temporary buffer of data logs in encrypted solid-state storage for regulatory audits, anomaly tracking, and delayed cloud synchronization.

[0044] The device also includes an Edge Communication Gateway (ECG) that enables secure, policy-compliant data transmission to the central retail cloud management system. This is achieved through adaptive transmission planning based on bandwidth availability and the relevance of the insights. The ECG supports dual-mode communication via Wi-Fi 6 and LTE / 5G modules and integrates an embedded firewall, a VPN stack, and a device identity keyring for compliance with zero-trust security regulations.

[0045] A visual feedback and configuration interface (VFCI) is integrated as a power-saving E-Ink touchscreen display or, optionally, as an HDMI output for larger screens. This interface allows store personnel to view alerts, analytics dashboards, and device diagnostics without cloud access. Configurations such as AI model updates, rule optimization, sensor calibration, and network settings can be managed via an administrator-protected local GUI or remotely via secure over-the-air updates.

[0046] In one embodiment, the ERAD includes a self-diagnostic and failover management unit that monitors the internal temperature, processing load, sensor status, memory integrity, and power supply stability. Upon detecting anomalies or potential failure conditions, the unit executes recovery routines, including thermal throttling, service restarts, or safe shutdowns, to prevent system damage and ensure uptime.

[0047] The modular architecture allows multiple ERAD units to operate as a mesh network, enabling store-wide data aggregation and the exchange of inferences. In such configurations, a distributed consensus protocol governs shared decision triggers, such as aisle-level customer traffic analysis or coordinated advertising displays, without requiring constant cloud arbitration.

[0048] The invention also includes an integrated model adaptation trainer that uses current business data to selectively train parts of the analytical models in order to refine predictions and reduce domain drift. Training takes place during off-peak hours and utilizes principles of federated learning. Anonymized model updates are regularly transmitted to the central server to improve the model globally while simultaneously protecting user privacy.

[0049] The drawing and the preceding description show examples of embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another embodiment. For example, the sequence of the processes described here can be changed and is not limited to the manner described here. Furthermore, the actions of a flowchart need not be implemented in the sequence shown; nor does it necessarily have to be performed by all actions. Actions that are not dependent on other actions can also be performed in parallel with the other actions. The scope of the embodiments is in no way limited by these specific examples.Numerous variations are possible, whether explicitly stated in the specification or not, such as differences in structure, dimensions, and material use. The range of embodiments is at least as broad as specified in the following claims.

[0050] Advantages, further benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and all components that can lead to an advantage, benefit, or solution occurring or becoming more apparent are not to be construed as critical, necessary, or essential features or components of individual or all claims. References 100 A Edge-Optimized Retail Analysis System To Support Decisions at the Store Level. 102 Edge Computing Units 104 System-On-Modules (Som) 104a Neural Processing Unit (Npu) 106 High Bandwidth Volatile Memory Module 108 Non-volatile solid-state drive (SSD) 110 Energy Management Circuit 112 Sensor interface bus 114 Multithreaded Orchestration Layer

Claims

[1] An edge-optimized retail analytics system for autonomous in-store decision support, consisting of: an edge computing unit housed in a robust, thermally conductive enclosure and configured for use in a retail store environment; System-on-Module (SoM) mounted on a multilayer printed circuit board (PCB), wherein the SoM comprises a multi-core central processing unit (CPU) configured to manage sensor data orchestration and rule-based inference processing, and a neural processing unit (NPU) configured to perform deep neural network inference operations in real time; a high-bandwidth volatile memory module electrically connected to the SoM to buffer time-aligned multimodal sensor streams; a non-volatile solid-state drive (SSD) configured to persistently store AI model weights, inference outputs, and business-specific event logs; a power management circuit integrated on the printed circuit board, configured to regulate the input voltage of a Power-over-Ethernet (PoE) line; and a sensor interface bus that is connected to a variety of sensor modules, including visible spectrum cameras, thermal imaging sensors, passive infrared motion sensors, RFID readers, and load cell-based shelf weight sensors. [2] System according to claim 1, wherein the edge computing unit comprises a hardware-isolated coprocessor configured to perform secure model parameter updates, the coprocessor comprising a Trusted Execution Environment (TEE) implemented on a dedicated ARM TrustZone or secure enclave, the TEE being programmed to receive encrypted gradient updates via a federated learning protocol and decrypt, validate, and integrate the updates into the locally stored AI model, while preventing direct memory access by insecure processes. [3] System according to claim 1, wherein the sensor interface bus is a modular I 2It incorporates a C and SPI hybrid architecture that supports time-synchronized querying of heterogeneous sensor types, and where the interface is managed by a programmable real-time clock controller that aligns the input signals across all sensor channels to a synchronization threshold of 3 milliseconds to ensure a temporally coherent input for downstream inference. [4] System according to claim 1, wherein the NPU is configured to execute quantized Convolutional Neural Networks (Q-CNNs) for low-latency inference in object recognition tasks, wherein the Q-CNNs are compiled via a vendor-specific optimization pipeline and stored in the intermediate representation formats TensorRT or OpenVINO within the SSD, and wherein the NPU dynamically loads and removes model segments based on a usage priority matrix calculated by the CPU based on the current customer frequency density in the store. [5] System according to claim 1, wherein the edge computing unit further comprises a passive thermal management subsystem comprising a heat spreader plate made of anodized aluminium which is thermally bonded to the SoM and the memory modules by means of a thermally conductive adhesive, wherein the heat spreader is mechanically connected to a finned heat sink structure integrated into the housing and wherein thermal profiles are monitored by integrated thermistors and used to throttle the CPU and NPU clock frequencies during high inference workloads to prevent temperature overshoot above 80 °C. [6] System according to claim 1, wherein the multi-threaded orchestration layer comprises an event-driven pipeline manager implemented using an asynchronous event loop, the pipeline manager being configured to allocate computational resources to image inference tasks, product identification tasks and inventory outage detection tasks based on a token bucket priority planner, the planner dynamically rebalancing processing threads based on real-time metrics such as frame processing rate, memory usage and event queue length. [7] System according to claim 1, wherein the edge computing unit is further connected via a GPIO expansion module to an array of LED-based shelf displays and wherein the displays are triggered in response to localized stockout detection events identified by a comparison of the current load cell weight and the predicted planogram baseline, wherein the triggering is controlled by a low-latency rule set encoded in a local inference engine implemented using a lightweight ONNX runtime library.

Images