Product Purchase System

Abstract

We provide a product purchasing system device with cross-language voice interaction capabilities, secure transaction processing, and high system stability. [Solution] The product purchase system device comprises a user terminal, a store server, and a cloud system. The user terminal is equipped with a microphone device 101 that captures the user's voice input, and the microphone device has a built-in RNNoise noise reduction module and a Vosk voice recognition engine. The voice processing core 102 has a built-in Whisper engine. The store server is equipped with an LED status display and a built-in control board module 202, which controls an electromagnetic lock mechanism 203 to lock / unlock the product storage compartment, and controls a communication module 201 and a payment module 204. The cloud payment ASIC chip 302 in the cloud system enables both the payment user and the store to use the system easily and securely.

Description

Technical Field

[0001] The present invention relates to a product purchase support system for convenience stores, and particularly to a system integrating multilingual voice guidance, a physical inventory locking mechanism, and an international settlement function. This system mainly belongs to the International Patent Classification G07G 1 / 00 (retail store equipment), and is also related to technical fields such as G10L 15 / 22 (voice amplification), G06Q 20 / 38 (settlement risk management), etc. Specifically, it focuses on high-speed dialect processing based on FPGA chips, multi-currency settlement processing by ASIC chips using the TSMC28nm process, and high-speed communication between hardware based on the Bluetooth 6.0 beacon + Wi-Fi 7 standard. (Figure 1) The system of the present invention

[0002] Voice recognition terminal The RNNoise deep noise reduction module integrated into the microphone (101) of the user terminal effectively suppresses environmental noise of 70 decibels or more and improves the signal-to-noise ratio of foreign language voice signals to 35 decibels or more. By using a multi-microphone beamforming algorithm, voice signals in a specific direction are emphasized, ambient noise is reduced, and the accuracy of voice recognition is improved. (Figure 2)

[0003] To improve the continuity of voice conversations in the network interruption state of the user terminal (101), the present invention provides a local offline voice cache module to realize the processing and buffer storage of voice data even when cloud connection is unavailable. The aforementioned offline audio cache module works in conjunction with an offline engine based on the local speech recognition unit Vosk (open-source speech recognition engine), supporting audio data processing time up to approximately 25 minutes. Based on a 16kHz sampling rate and 16-bit single-channel PCM format, the audio data rate is approximately 32KB / s, and the total cache data size does not exceed 50MB. When employing a highly compressed audio encoding format (such as Opus), the system can support audio processing and caching time up to approximately 80 minutes (Figure 3).

[0004] Recognition results are cached in a structured format, which includes fields such as speech-to-text conversion text, word-level timestamps, and recognition confidence, with an average generation rate of 1-3 KB / s. The recognized data is stored synchronously with the corresponding original or compressed audio data and serves as the basis for subsequent upload, comparison, or error correction processes.

[0005] When the local cache approaches the system-configured capacity limit (e.g., 50MB), the system triggers the cache management mechanism, prioritizing a "first-in, first-out" strategy to delete the earliest cached data; optionally, data compression and summary storage strategies are activated, and tiered processing is performed on non-key data to avoid the impact on system operation due to cache overflow.

[0006] After cloud connectivity is restored, the module completes cache state detection within 0.8 seconds and initiates a data synchronization flow via a physically isolated RDMA dedicated channel. Cached audio and recognition result data are transmitted using a highest-priority scheduling strategy, and the transmission process employs the AES-256 encryption algorithm for end-to-end encryption, ensuring the integrity and security of data transmission.

[0007] This system further incorporates a hardware-level QoS (Quality of Service) mechanism to prioritize and schedule data flows within the RDMA channel, ensuring that voice data is processed preferentially even under bandwidth-limited or high-load conditions, thereby guaranteeing low-latency response capabilities for core voice services.

[0008] In the embodiment of this invention, the voice processing unit (102) connected to the user terminal (101) employs a Xilinx Zynq-7000 series FPGA chip and integrates a multilingual / dialect speech conversion engine optimized based on a whisper large model structure (this is an example of an open-source implementation, and the invention is not limited thereto). The engine performs frame-level processing operations on the user's voice input, specifically including a speech segmentation strategy with a frame length of 20ms and a frame shift of 10ms, and extracts multimodal feature information such as 16-dimensional Mel frequency cepstrum coefficients (MFCCs), fundamental frequencies, and speaker embedding vectors using a window function and a fused coding module.

[0009] The aforementioned multimodal features are processed by a "feature coupling unit" and then input into a built-in multilingual, multidialect acoustic model. Similarity matching with the target language model is performed, and then feature level alignment and standardization are completed via a transformer coding network. Finally, the features are mapped to the speech feature output in the standard language space. This transformation process has low latency characteristics, and the overall processing delay from speech input to standard feature output is controlled to within 0.5 seconds (Figure 4).

[0010] This system further incorporates a language selection module and an intonation adaptation module, allowing it to adaptively adjust the language path based on speaker characteristics and speech rhythm, and dynamically call a speech recognition model path that matches the speaker. The final output standard language speech features support the native languages ​​and dialects (including multiple minor languages) of the vast majority of foreign visitors to Japan, achieving highly versatile speech dialogue capabilities.

[0011] The system supports continuous processing of online voice input streams for up to 180 minutes, and under stable RDMA channel and network conditions, can continuously transmit multimodal voice feature data streams of up to approximately 512 KB / s, including information such as voice frame features, voiceprint vectors, intonation parameters, and language model indices. If the continuously input voice data exceeds the processing capacity limit, the system automatically activates an emergency cache mechanism or switches to offline compensation mode to avoid task blocking or degradation of recognition performance due to feature accumulation.

[0012] In addition, (102) performs real-time speech recognition and semantic alignment on the user's expressed purchase request, extracts product information contained therein, and transmits the converted standard language representation to the downstream processing unit (202) and the store clerk terminal via a dedicated RDMA channel for real-time dialogue response and task execution.

[0013] The FPGA chip (102) further includes an interface module that supports dynamic model hot updates, enabling it to receive new dialect acoustic models or language model parameter packages from a cloud system via a wireless connection method without interrupting local voice processing, and to complete local replacement or extended updates of the model. This update process employs an incremental loading method and supports model hot switching, ensuring system operation stability and model adaptability.

[0014] Store Server To achieve physical control of a product storage cabinet, an embodiment of the present invention provides an execution mechanism including an LED control board (202) and an electromagnetic lock component (203) (Figure 5). The working process and design are as follows: After the physical operation of storing purchased goods is completed, the solenoid valve (203) is driven by a control signal emitted from an ASIC chip (302) connected to an FPGA, and has a locking force of 30N or more. This locking force acts on the safe lock core via a mechanical transmission structure, achieving physical closing control of the product storage area (Figure 6). To improve performance stability, the solenoid valve (203) integrates a thermal compensation circuit and an EEPROM temperature calibration module. The EEPROM supports more than 1 million data writes, and the calibration error is controlled to within ±0.5℃. The temperature working range of the locking mechanism is -40℃ to +85℃, referring to the automotive electronics standard AEC-Q100 Grade 3, and the system controls the locking force error at each temperature point to within ±1N. The maximum continuous energizing time of the solenoid valve does not exceed 60 seconds, the steady-state operating current is less than 500 mA, and thermal stability is controlled so that the temperature rise after operation does not exceed 15°C.

[0015] The LED control board (202) is equipped with an inventory status display matrix and an ambient light adaptive adjustment module to provide visual indications of product status. The inventory status display matrix consists of RGBW four-color LEDs and supports at least four types of status indications: green flashing (50% duty cycle) indicates sufficient stock and available for purchase; red constant illumination indicates insufficient stock or the product is locked; yellow breathing light (1Hz frequency) indicates the product is being processed for payment; and blue flashing (2Hz frequency) indicates the product is being restocked and placed on the shelves. The ambient light adaptive module detects ambient illuminance in real time using an integrated light sensor, refreshes the brightness setting at a 1Hz frequency, and achieves automatic adjustment of LED brightness within the range of 50cd / m² to 300cd / m². The response time is controlled to be within 500ms, ensuring that the displayed content can be clearly identified within a +45° viewing angle and a range of 2 meters.

[0016] The control board (202) further incorporates a product attribute analysis module that automatically identifies and matches the control status of the corresponding storage cabinet based on the temperature storage requirements marked in the product information, and indicates this using an LED visualization method. The temperature control matching logic employs a control tolerance of ±2℃ and supports locking of three types of storage cabinet status categories: frozen, refrigerated, and ambient temperature. In addition, the control board supports user guidance via a Bluetooth / Wi-Fi positioning module, linking the positioning process with the storage cabinet status, and visualizing and displaying the guidance, target cabinet position, and progress.

[0017] The control board (202) is further equipped with transaction interaction capabilities and can display the user's completed operations and retrieval status based on the transaction code generated by the settlement module (204). The transaction code is displayed on the control interface in text or graphic code format (such as a QR code) and is updated in conjunction with the locked cabinet information.

[0018] To ensure the reliable operation of the system, the store server is equipped with a dual watchdog circuit, including an independent hardware watchdog and a kernel-level software watchdog (Figure 7). The hardware watchdog is driven by an independent clock source and polls to detect the system reset state every 1.5 seconds; the software watchdog is integrated into the operating system kernel and detects the operational status of key processes every 3 seconds. If either watchdog detects a system anomaly, it immediately triggers an automatic recovery mechanism, controlling the total automatic recovery time to within 10 seconds (Figure 8). After three consecutive automatic recovery failures, the system enters fault protection mode, stops the key execution logic, and notifies maintenance personnel via a control terminal.

[0019] The system's communication module (201) supports the estimation of the user terminal's location and spatial matching of the locked product storage cabinet based on Bluetooth 6.0 channel detection technology or the RSSI (Received Signal Strength Indicator) value of a Wi-Fi beacon. In a typical commercial environment, the positioning accuracy reaches ±1.5 meters, and can be further improved to ±0.5 meters when multi-beacon fusion positioning is employed. The payment module (204) immediately generates a unique transaction code after successful payment and transmits it synchronously to the user terminal and the store clerk terminal. This code is used to bind the dialogue path between the user and the target product cabinet.

[0020] The payment module (204) further includes an AR navigation processing unit that generates and displays a route map on the user terminal interface based on the user terminal's real-time Bluetooth positioning results and the location of the product cabinet on the map. The route map is built on the latest store map data and has a default validity period of 24 hours; if the map data version is mismatched or invalid, the system will offer the terminal a synchronized update to ensure the accuracy and stability of the navigation guidance (Figure 9).

[0021] As described above, the product dialogue system in this invention ensures stable operation in complex commercial environments by setting clear physical control parameters, recognition state boundaries, communication error correction mechanisms, and environmental adaptability. Each control module is provided with an operating range, tolerance thresholds, and recovery strategies to ensure consistency in the user experience, timeliness of dialogue responses, and security of data processing.

[0022] Cloud system To achieve highly efficient and secure payment and risk control processing, this invention provides a custom ASIC chip (302) manufactured using the TSMC 28nm process, integrating dedicated hardware logic modules for multi-currency payment processing, nationality identification, and federal learning risk control strategy modeling (Figure 10). The ASIC chip has a working voltage range of 0.8V to 1.2V, a typical operating frequency of 400MHz, a rated power consumption of no more than 2.5W, and possesses reliable thermal stability and power consumption control capabilities, making it suitable for commercial deployment in ambient to high-temperature environments.

[0023] The aforementioned ASIC chip (302) incorporates a nationality-based payment risk assessment circuit and can automatically identify the user's nationality information based on nationality identifiers (including, but not limited to, bank card BIN codes, mobile phone number location prefixes, and internationalized email domains) included in the payment data submitted by the user. If nationality information is missing, conflicting, or cannot be determined, the system defaults to adopting a "limit transaction" strategy and prompts the store clerk to manually review and confirm the transaction.

[0024] The Federal Learning Risk Control Module is integrated into the system-on-a-chip and is used for periodic aggregation and model updates of anonymized, deidentified payment data uploaded from multiple store servers. The Federal Learning Module does not re-transmit the original data, ensuring user privacy and data security. The model training cycle is set to once daily by default and can be dynamically adjusted based on system load. The generated risk control model is used in real time to assist in transaction decisions, risk control scoring, and limit management strategies.

[0025] For example, when the nationality identification circuit detects that the user is of Chinese nationality, the system automatically calls the built-in pre-authorization processing circuit and executes a pre-authorization operation for RMB transactions that complies with the "Bank Card Pre-Authorization Business Specification" issued by the People's Bank of China. This process includes multiple steps such as freezing the pre-charged amount, verifying the fund validity period, and matching the merchant authentication identifier, ensuring the compliance, controllability, and traceability of the entire RMB settlement process.

[0026] The ASIC chip (302) further integrates a multi-currency settlement parallel processing unit and includes four independently operating clearing cores. Each core supports real-time clearing calculations for major circulating currencies such as the Japanese yen (JPY), US dollar (USD), RMB (CNY), and euro (EUR). The operations between the clearing cores are isolated, supporting parallel processing of settlement transactions in four different currencies, ensuring the stable operation of the system under high concurrency conditions. The signal transmission method is accelerated through LVDS. (Figure 12).

[0027] The standard processing time for each transaction is controlled within 1.5 seconds, and the entire system supports an average transaction throughput capacity of 240 transactions per minute. After continuous high-load operation exceeds 10 minutes, the chip automatically enables the dynamic voltage and frequency scaling (DVFS) mechanism to perform frequency reduction management, maintaining the power consumption below the safety threshold and preventing overheating. All settlement instructions and key data paths are processed in the chip's on-chip trusted execution environment (TEE), ensuring the encrypted and isolated transmission of the user's identity, fund data, and nationality information, and ensuring financial security and compliance with cross-border settlement regulations. The external monitoring system refers to supervision agencies such as the Japanese Financial Information Center (JAFIC) established based on the Financial Services Agency and the Law on Prevention of Transfer of Criminal Proceeds. The reported data shall be statistical data that complies with the Personal Information Protection Law and has undergone irreversible anonymization processing, or report data that includes personal identification information only within the scope mandated by law. This function is for fulfilling the obligation to report suspicious transactions based on Article 8 of the Act on Prevention of Transfer of Criminal Proceeds. The data interface module automatically converts data into the electronic reporting format (e.g., XML Schema) specified by JAFIC and transmits it via a dedicated secure communication channel. (Figure 11)

Background Art

[0028] According to the prediction of the Japan National Tourism Organization, the number of foreign tourists visiting Japan in 2025 is expected to exceed 40 million (Note 1). However, in the current technical environment of convenience stores in a chain, when foreign tourists shop at retail stores such as convenience stores in Japan, the language barrier often causes confusion for both tourists and store clerks. Some operators have introduced multilingual call centers and support the communication between store clerks and foreign tourists through interpreters at the centers (Note 2). At the same time, due to the rapid increase in the number of foreign tourists and the diversity of their nationalities, many inconveniences have occurred in the existing product purchase technology system environment. For example, when multiple tourists who speak different languages and use their respective local dialects shop and check out in the store at the same time, the parallel processing ability of multi-currency settlement in the existing system is limited, resulting in processing delays during peak hours. In addition, the existing technical system may not operate properly under extreme weather conditions.

[0029] There is room for improvement in the existing product purchase system environment. For example, in terms of privacy and security, the encryption strength and key update mechanism are insufficient, and there is a risk of data leakage. There are also issues in terms of processing efficiency. At the same time, the unauthorized use of credit cards has reached a record high. According to reports, the damage amount due to unauthorized use of Apple Pay has increased by 17% and reached approximately 14 billion yen (Note 3). The loss due to unauthorized use of credit cards in 2024 reached 55.5 billion yen, setting a record high (Note 4).

[0030] The present invention was developed to solve the above and related problems, comprehensively strengthening safety, processing efficiency, and environmental adaptability with new technologies, and significantly improving the purchase experience and settlement success rate of foreign tourists. [Overview of the Initiative] [Problems that the invention aims to solve]

[0031] The problem of insufficient speech recognition accuracy in the current technological environment: Under current technological conditions, the system microphone equipment lacks sufficient environmental noise suppression capabilities, resulting in a low signal-to-noise ratio (SNR) for foreign language and dialect voice signals. This can cause speech recognition accuracy to drop to below 70%, especially during peak hours in convenience stores. Furthermore, the system lacks the ability to recognize foreign languages, their new dialects, and slang, impacting its practicality. This system improves accuracy through noise reduction and dialect recognition, effectively preventing misidentification of inventory.

[0032] Problems with insufficient system stability in the current technological environment: In the current technological environment, the system can hang up during prolonged operation due to software bugs or hardware malfunctions, requiring significant time for recovery and impacting store operations.

[0033] The problem of insufficient offline operation capacity in the current technological environment: In the current technological environment, if the cloud connection is lost, the speech recognition and dialect conversion functions become almost completely unusable, preventing users from purchasing products normally.

[0034] The problem of insufficient safety in the current technological environment: The current technological environment's systems suffer from insufficient encryption algorithm strength, long key update cycles, and a lack of mutual authentication mechanisms between devices, resulting in a high risk of data leakage and tampering during transmission, and potentially making it impossible to guarantee the security of confidential data such as international payment information.

[0035] The problem of insufficient processing efficiency in the current technological environment: The current technology environment has limited parallel processing capabilities for multi-currency payments, resulting in processing delays when multiple foreign customers attempt to make payments simultaneously during peak times. This impacts customer experience and store operational efficiency, especially during tourist seasons (e.g., cherry blossom season, year-end shopping festivals).

[0036] The problem of insufficient environmental adaptability in the current technological environment: Systems in the current technological environment are typically designed for temperatures ranging from -20°C to 40°C. Under extreme temperature conditions, the locking force of the solenoid valve may become unstable, leading to problems such as a decrease in the processing speed of the FPGA chip, and potentially preventing normal operation. [Effects of the Invention]

[0037] Strong cross-language voice interaction capabilities: Supports real-time recognition and standard voice alignment of multiple languages ​​and dialects, achieving a recognition rate of over 95% with the combination of Vosk+RNNoise; effectively prevents duplicate sales by utilizing inventory locking and payment integration, improving the shopping experience for foreign users in their local environment.

[0038] Secure and efficient transaction processing: Integrated ASIC chips enable nationality identification, risk control decisions, and multi-currency parallel processing, shortening transaction times and improving the economic profitability of retailers.

[0039] Strong compliance capabilities: The system automatically adapts payment rules based on the user's nationality, and is particularly adaptable to the pre-authorized payment mode in countries with strict financial supervision.

[0040] Strong system stability: Dual-core hot standby and an automatic recovery mechanism significantly reduce the risk of trading interruptions due to system failures.

[0041] Flexible risk control: By combining federal learning and dynamic scoring strategies, risk control achieves both real-time capabilities and local privacy protection.

[0042] Data uploads are consistent and traceable: Transaction activities and risk control results can be transmitted in a directional manner to the supervisory system, meeting the transparency requirements for cross-border settlement supervision.

[0043] Compared to the apps currently widely used in chain convenience stores, this system enables users to lock onto and purchase items from the nearest store's inventory. On the other hand, existing app technologies are designed for inventory management in a broader environment and cannot meet the needs of users.

[0044] We provide AR navigation services from shopping needs to product pickup, significantly enhancing the experience for foreign tourists. [Brief explanation of the drawing]

[0045] [Figure 1] Overall structural block diagram of the cross-border retail payment and risk control system according to the present invention. [Figure 2] Flowchart for noise reduction work on microphone equipment. [Figure 3] The Vosk workflow diagram is deployed on the internal edge of the user terminal (101). [Figure 4] This exhibit (102) presents a schematic diagram of the FPGA speech recognition processing flow, illustrating the processing path from frame-level feature extraction of user input speech to transformer alignment conversion and standard speech generation. [Figure 5] The product control cabinet control logic diagram includes the LED state matrix, temperature control electromagnetic lock, and positioning induction module interaction relationships. [Figure 6] A locking force of 30N or more is applied to the locking core, enabling physical control of the product storage area. [Figure 7] A dual watchdog circuit is installed. [Figure 8] If the watchdog detects a system anomaly, it immediately triggers an automatic recovery mechanism. [Figure 9]The system prompts the terminal for synchronized updates, ensuring the accuracy and stability of navigation guidance. [Figure 10] Overall structural block diagram of a custom ASIC chip (302) manufactured using the TSMC 28nm process. [Figure 11] This diagram shows the internal module structure of an ASIC chip, which includes five modules: nationality assessment, risk control modeling, multi-currency settlement, cryptographic coprocessing, and supervisory interaction. [Figure 12] Supports parallel processing of settlement transactions using four different currencies. [Figure 13] The schematic diagram of dynamic risk control score threshold adjustment demonstrates the adjustment process by which the score engine tightens or loosens the threshold in response to changes in the trading environment. [Figure 14] Dual ASIC chip switching flow diagram in a hot standby mechanism. Switching logic flow diagram for dual ASIC hot standby. [Figure 15] The system and supervisory interface interaction flowchart demonstrates the logic for uploading transaction data, confirming receipt, and resending failed transmissions. [Modes for carrying out the invention]

[0046] When a user issues a voice command to purchase a product in a foreign language to the microphone device (101), the microphone device first processes the original voice signal using an RNNoise depth noise reduction module and a beamforming algorithm for a multi-microphone array, and then transmits the denoised signal to the FPGA chip (102). Based on its built-in multilingual transformer model, the FPGA chip completes foreign language voice recognition and standard voice feature conversion within 0.5 seconds, forming a standardized control signal used for product identification and inventory identification.

[0047] The control signal is transmitted to the LED control board (202) via a communication module (201) configured based on Bluetooth 6.0 channel detection technology and a Wi-Fi 7 protocol stack. To improve the real-time and stability of communication, the Bluetooth beacon broadcast interval is set to 20ms, a 2Mbps encoding scheme is employed, and the broadcast data packet includes 37 bytes of conversion command and device status data. The system supports an extended broadcast packet format to transmit encrypted inventory warning information and effectively avoids radio interference through dynamic frequency hopping channel technology.

[0048] After receiving the above-mentioned control signals, the LED control board (202) performs an RGBW matrix display based on the product's inventory status. Before communication, mutual identity authentication is performed by both devices, and the data packets are signed with a 256-bit AES encryption algorithm to verify their integrity and prevent tampering or forgery of the signals during transmission.

[0049] All owned nodes within the system, including user terminals, store servers, and cloud systems, are connected via end-to-end encrypted communication links. These links employ the AES-256-GCM encryption algorithm to ensure the confidentiality and integrity of data transmission. Each device possesses a unique digital certificate and dynamically updates its session key every 15 minutes via frequency hopping channels under the Bluetooth 6.0 and Wi-Fi 7 protocols to counter man-in-the-middle and replay attacks.

[0050] Once the user confirms the purchase and submits the payment request, the cloud system invokes a nationality-based payment risk assessment module located within the ASIC chip (302), performs a compliance check based on the nationality identifier in the user's payment information, and completes the multi-currency payment operation. After successful payment, the ASIC chip sends an unlock signal to the solenoid valve (203) of the product storage cabinet via the GPIO interface, and the solenoid valve completes a physical unlock operation within 10 seconds, allowing the user to retrieve the product.

[0051] To ensure system stability, the store server integrates a dual watchdog monitoring circuit, including hardware and software watchdog modules, which detect system status every 1.5 seconds and 3 seconds, respectively. If a failure such as system operation abnormality, communication timeout, or task freeze is detected, an automatic restart and failure recovery flow are triggered, and the system completes recovery within 10 seconds, ensuring no interruption to key transactions.

[0052] Specific implementation methods (supplementary explanation) In actual applications, users perform product selection operations via a terminal device that supports voice input. The system first processes the voice command using an FPGA chip, performing frame processing with a 20ms frame length and a 10ms frame shift. It extracts multimodal features including 16-dimensional MFCCs, fundamental frequencies, and speaker features, which are then input into a speech alignment network based on a transformer architecture to achieve standard speech spatial conversion. The conversion results are input to the store clerk terminal or product processing node in the form of intermediate semantic features, avoiding misunderstandings due to language barriers.

[0053] If the system detects a network disconnection during the speech processing process, it will switch to offline recognition using the local "Vosk" (open-source speech recognition engine) engine within 0.5 seconds, control the delay to within 0.8 seconds, and automatically upload the cached content after the connection is restored.

[0054] In the payment flow, the ASIC chip classifies users by nationality based on nationality identifiers such as bank card BIN code and mobile phone number. For example, if identified as a Chinese user, the system automatically triggers a RMB pre-authorization flow that conforms to the "Bank Card Pre-Authorization Business Standards." Throughout the entire payment flow, four independent payment cores can process transactions in different currencies in parallel, and the delay for single transactions is controlled to within 1.5 seconds.

[0055] In terms of risk control, the federal learning module within the chip performs model aggregation on de-identified payment data from multiple stores daily to generate new risk control score parameters. The scoring engine updates thresholds in real time based on factors such as historical behavior, geographic location, and supervisory alerts, adapting to risk control needs under different scenarios (Figure 13).

[0056] If the system detects that the chip temperature has risen above 90°C or that an abnormality has occurred in the key module during operation, the backup chip will take over the task within 200ms. The primary and backup chips synchronize their operating status, transaction context, and intermediate data via a dedicated channel to ensure the continuity of transactions (Figure 14). All confidential data acquired during the transaction process undergoes signature verification and channel encryption via an encryption coprocessor within the chip, meeting security standards such as FIPS 140-2. The supervisory interface module communicates with government or clearing agency servers in an asynchronous queue format to ensure coordinated uploads (Figure 15). Annotation Explanation 1. A navigation module placed at the user's end point to guide the user to the store is protected by a single claim (divisional application). 2. A temperature-controlled storage cabinet is protected by a single claim (divisional application). 3. Storing / retrieving products is a physical operation performed by the user and store staff (not covered by patent). 4. "Vosk" (Open Source Speech Recognition Engine) English original: Vosk. 5. Whisper Large. 6. Transformers. (Original English: Transformer) Cited materials (Note 1): The total number of tourists visiting Japan in 2025 is expected to exceed 40 million. >Source: NHK News, July 16, 2025 (Note 2): Japanese convenience stores are taking many measures to better accommodate foreign tourists. 7-Eleven convenience stores have introduced multilingual call centers to support communication between staff and foreign tourists through interpreters at the centers. >Source: Issued by JNTO on February 19, 2025 (Note 3): Title: Frequent Apple Pay Fraud in Japan Attracts Attention to Credit Card Security. Content: This article reports on the issue of Apple Pay being used for credit card fraud due to vulnerabilities in identity verification, and states that the amount of damage caused by credit card fraud in 2016 increased by 17%, reaching approximately 14 billion yen, according to statistics from the Japan Credit Card Industry Association. >Source: Nikkei, Publication Date: August 15, 2017 (Note 4): Credit card fraud occurred in Japan, with losses reaching 55.5 billion yen, a record high. According to data from the Japan Credit Card Association, losses due to credit card fraud in 2024 reached 55.5 billion yen, a new record high, with "card number theft" accounting for 92.5% of the total. >Source: NHK News (Reprinted from Asia Pacific Express) Publication Date: March 30, 2025

Claims

1. A product purchasing system device for use in multi-level retail businesses, including the following: The user terminal's microphone device (101) captures the user's voice input and executes voice command dialogue; the FPGA chip module (102) integrates a speech recognition and Whisper Large conversion engine, which employs a multilingual speech feature alignment network based on a Transformer structure, performs frame-level processing on the user's voice, and converts it into standard speech features; The user terminal is equipped with a microphone device (101) for language voice input, which integrates an RNN depth noise reduction module, a multi-microphone beamforming algorithm, and a Vosk (open-source speech recognition engine) local offline real-time speech recognition module. The voice is transmitted within 0.5 seconds to an FPGA chip (102) capable of converting the native language and dialects (including multiple minor languages) of more than 99% of foreign visitors to Japan. The main control board module (202) includes an LED status indicator and an electromagnetic locking mechanism (203) to control the physical unlocking and inventory status display of the product storage cabinet, and also includes a communication module (201) and a payment module (204); This is an ASIC chip module (302), manufactured using the TSMC 28nm process, and integrates the following functional modules: a nationality-based risk assessment circuit that identifies the user's nationality based on the nationality identifier in the user's payment information; It is a federal learning risk control model module that performs aggregate learning on anonymized transaction data from multiple store servers and updates the risk control score model in real time; it is a multi-currency settlement parallel processing unit that includes multiple independent clearing cores and supports parallel processing of transactions of multiple currencies; and it is an encryption algorithm coprocessor that supports AES and NIST ECC encryption and decryption operations. It is a highly available backup mechanism, including a hot standby switching structure for primary and backup chips; and as a data interface module with an external supervision system, it performs reporting of transaction summaries and risk management data based on a specified data format. The modules transmit user transaction data via trusted channels, and the ASIC chip provides risk control assessment, settlement calculation, and data encryption throughout the entire transaction processing process.

2. The system according to claim 1, wherein the user terminal is equipped with a microphone device (101) for voice input, the microphone device has an RNNoise depth noise reduction module at the edge, employs a multi-microphone beamforming algorithm, and transmits voice immediately to an FPGA chip (102) under normal network conditions.

3. The system according to claim 1, wherein the user terminal integrates a module that supports offline speech recognition, employs a local speech recognition engine based on the Vosk (open source speech recognition engine) architecture, automatically switches to offline mode when the cloud connection is interrupted, and the switching delay does not exceed 0.5 seconds.

4. The system according to claim 1, wherein the (102) FPGA chip module includes a model update interface and supports receiving and loading new foreign language dialect acoustic models wirelessly.

5. The system according to claim 1, wherein the main control board module includes an RGBW LED display matrix for displaying inventory status, and supports four types of status indication methods: green blinking, red constant illumination, yellow breathing, and blue blinking.

6. The system according to claim 1, wherein the locking force of the electromagnetic locking mechanism is 30N or more, and it comprises a thermal compensation circuit and an EEPROM temperature calibration module, is capable of stable operation in a temperature range of -40°C to +85°C, and controls the locking error to within ±1N.

7. The system according to claim 1, wherein the store server's dual watchdog circuit includes an independent hardware watchdog and a software watchdog, the hardware watchdog is driven by an external clock source and detects a reset signal every 1.5 seconds, the software watchdog is built into the OS kernel and monitors the process state every 3 seconds, and if either watchdog detects an abnormality, the system automatically recovers after triggering a melting mechanism, and the automatic recovery time is controlled to be within 10 seconds.

8. The system according to claim 1, wherein the multi-currency settlement parallel processing unit of the ASIC chip module includes four clearing cores, supports simultaneous processing of currency types such as yen, dollar, euro, and yuan, and dynamically adjusts power consumption according to the load using DVFS technology.

9. The system according to claim 1, characterized in that the encryption algorithm coprocessor supports completing 500 or more AES-256 encryption and decryption operations and 100 or more ECC P-256 digital signature verifications per second.

10. The system according to claim 1, wherein the primary and backup chips of the ASIC chip are equipped with a hot standby mechanism, and in the event of a failure in the primary chip, failure switching and task continuation are completed within 200ms.

11. The system according to claim 1, wherein the risk control score model automatically adjusts the score threshold based on transaction history, supervisory notices and federal learning results, recalculates the weight coefficients of risk factors based on federal learning every 10 minutes, and dynamically updates the matrix.

12. The system according to claim 1, wherein the external supervision data interface module supports reporting of risk control data in RESTful API, MQTT, and ISO 20022 protocol formats, and comprises an asynchronous queue and an acknowledgment mechanism.

Images