Multi-modal fusion self-adaptive intelligent cold and hot touch comfort device

Abstract

The invention relates to the field of intelligent health, in particular to a multi-mode fusion self-adaptive intelligent cold and hot touch comfort device. Comprising a touch and comfort module, a control module and a multi-mode analysis module, the touch and comfort module adopts a TEC chip to achieve refrigeration or heating, and the temperature can be rapidly and accurately adjusted. And the multi-modal analysis module fuses and analyzes user feedback and physiological sensor data through an intelligent algorithm, and generates emotional scores and personalized adjustment suggestions in combination with a knowledge graph structure, so that closed-loop optimization of cold and hot touch experience is realized. The control unit is responsible for system coordination, temperature control and safety monitoring. The device has the advantages of being wearable, free of pollution, capable of being repeatedly used and the like, is suitable for the fields of sports relaxation, emotion relieving and personal health management, and provides scientific, safe and comfortable use experience.

Description

Technical Field

[0001] This invention relates to the field of smart health, specifically to a multimodal fusion adaptive smart hot and cold touch device. Background Technology

[0002] Currently, commonly used cold compress methods for rehabilitation and sports recovery mainly include self-cooling chemical ice packs, saline ice packs, ethanol ice packs, cuff-type cold compress devices, and cryotherapy pressure devices; the main hot compress methods include hot water bottles, hot compresses, hot compress patches, and microwave heat therapy. Therefore, existing cold and hot compress methods generally use disposable chemical materials or cold (heat) storage devices to achieve the desired effect. These methods may generate pollution and may require pre-storage, resulting in problems such as long application time, low efficiency, uneven effects, and potential pollution. Furthermore, existing methods struggle to achieve varied and controllable cold (heat) compress effects, making it impossible to analyze user experience and provide optimization suggestions.

[0003] Semiconductor cooling chips, also known as TEC chips, utilize the Peltier effect to directly convert input electrical energy into a temperature difference, achieving cooling. They generate current through the movement of electrons at different energy levels in two conductors. When electrons move from a higher energy level to a lower energy level, they release excess energy; conversely, they absorb energy. This energy is absorbed or released as heat at the interface between the two materials, resulting in a phenomenon where one end of the semiconductor cooling chip absorbs heat while the other releases it. Furthermore, changing the direction of the direct current reverses the flow, causing the cooling and heat-releasing ends to interchange. Semiconductor cooling chips are small, highly reliable, and easily adjustable. Using semiconductor cooling chips for cooling (heating) is fast, pollution-free, and reusable. However, existing medical and non-medical devices based on semiconductor cooling are still primarily based on "fixed temperature output" and "passive operation," lacking intelligent capabilities. Their main problems include: the inability to collect changes in the user's physiological indicators during use (such as heart rate, local temperature, blood flow, electromyography, etc.) and the inability to obtain the user's subjective experience (such as comfort, pain level, and temperature preference), thus hindering adaptive adjustments. Even those devices with simple sensor inputs lack intelligent algorithms such as pattern recognition, sentiment analysis, and personalized models. They also lack historical data accumulation and intelligent reasoning mechanisms, preventing them from generating differentiated strategies based on different individuals, body parts, and symptoms. This hinders the creation of precise, safe, and dynamically controllable temperature regulation effects. In particular, existing devices generally lack "multimodal fusion" capabilities, meaning they cannot unify and analyze user language feedback, emotional state, image data (local swelling, temperature distribution), and physiological sensor data, thus failing to establish user models with continuous learning capabilities. This lack of systematic analysis makes it difficult for these devices to provide truly personalized treatment and to form a feedback loop.

[0004] Therefore, existing technologies urgently need an intelligent hot and cold touch device with data acquisition capabilities, emotion analysis capabilities, multimodal modeling capabilities, and knowledge graph reasoning capabilities to solve problems such as lack of intelligent adjustment, lack of self-learning capabilities, and lack of user experience optimization mechanisms, thereby significantly improving the safety, comfort, accuracy, and consistency of hot and cold (hot) compresses. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multimodal fusion adaptive intelligent hot and cold touch therapy device, which solves the problems of existing touch therapy devices lacking intelligent adjustment, self-learning ability, and user experience optimization mechanism, thereby significantly improving the safety, comfort, accuracy, and consistency of hot and cold (hot) compresses.

[0006] To solve the aforementioned technical problem, the technical solution adopted by the present invention is: a multimodal fusion adaptive hot and cold touch device, comprising a touch module, a control module, and a multimodal analysis module; The Touch Comfort module is used to provide cooling or heating to the target area; The control module includes a PWM unit and a multimodal interface unit. The PWM unit is connected to the touch module and is used to control the working state of the touch module. The multimodal interface unit is used to receive environmental information, user information, and user interaction information collected by peripheral sensors and transmit them to the multimodal analysis module. The multimodal analysis module is connected to the control module and is used to receive data from the control module, analyze it, generate the adjustment strategy of the touch comfort module, and return it to the control module. The multimodal analysis module includes a sentiment analysis unit, a perception unit, an image analysis unit, and a knowledge graph unit. The sentiment analysis unit constructs an interrogation-based interactive scenario, identifies and evaluates the user's subjective emotional state and preferences in real time, and quantifies the subjective emotional state and preferences into an emotion score. The emotion score, along with the questions and user responses in the interactive scenario, is transmitted to the knowledge graph unit. The sensing unit collects multi-channel physiological data and, in conjunction with the application site and device settings, dynamically optimizes the sensor weights through a principal component analysis model, transmitting the sensor weight optimization results to the knowledge graph unit. The image analysis unit acquires image information of the target area, performs feature extraction and parameter quantization on the image information, generates image analysis results, and transmits them to the knowledge graph unit; After pre-training, the knowledge graph unit constructs a knowledge graph with various nodes and relationships. Nodes include question nodes, response nodes, sensor nodes, image nodes, emotion nodes, strategy nodes, execution result nodes, device mode nodes, treatment mode nodes, and historical treatment process nodes. Relationships include co-occurrence relationships, causal influence, parameter dependence, and temperature control feedback. The knowledge graph unit performs reasoning and retrieves similar states based on the graph structure, generates adjustment strategies, and supports incremental learning of subsequent models.

[0007] Furthermore, the sentiment analysis unit employs the BERT model. First, it poses a question and obtains user responses. Then, it performs analog-to-digital conversion and word segmentation on the user responses, dividing them into {vocabulary, position, sentence segmentation} and normalizing them. Useless information is discarded, and the remaining information is encoded. The BERT model has multiple encoding layers. Each layer receives vector encoding from the word segmentation stage or the previous encoding layer and performs sentiment analysis with reference to pre-trained knowledge graph parameters. After processing through multiple encoding layers, a sentiment score is generated.

[0008] Furthermore, the BERT model combines knowledge graphs with real-time sensor data to provide users with a feedback-based automatic optimization solution, which users then use to make their final decisions.

[0009] Furthermore, the image analysis unit uses a CNN model to process and analyze the image information acquired by the image sensor. First, it acquires the user's original image, then performs data enhancement and edge detection on the acquired image, extracts the ROI region and performs morphological filtering, and finally performs feature quantization on key information to form image acquisition information.

[0010] Furthermore, the knowledge graph employs a hierarchical association mechanism for reasoning, following the path from sensor node to image node to sentiment node to policy node, and outputs the current policy selection after a weighted path search.

[0011] Furthermore, after the strategy is executed, physiological changes, image changes, temperature difference feedback, and emotional fluctuation trends are written into the execution result node to update the strategy success rate and achieve dynamic learning.

[0012] Furthermore, the knowledge graph unit defines relationships as causal edges, temporal edges, weighted edges, semantic edges, and image feature edges.

[0013] Furthermore, the Touch Comfort module includes a TEC cooling plate and a heat sink. The TEC cooling plate includes a contact surface and a heat dissipation surface. In cooling mode, the contact surface cools, the heat dissipation surface heats, and the heat sink dissipates heat from the heat dissipation surface. In heating mode, the contact surface heats, the heat dissipation surface cools, and the heat sink does not work. The switching between cooling mode and heating mode is controlled by the direction of the current flowing through the TEC cooling plate.

[0014] Furthermore, the Touch Comfort module also includes the ADN8835 chip and the MAX6650 chip. The ADN8835 chip controls the operation of the TEC cold plate, and the MAX6650 chip controls the operation of the heat sink.

[0015] Furthermore, the TEC cold plate is equipped with a thermistor. The TEC cold plate is connected to the PWM unit through the ADC unit. The thermistor collects the temperature of the TEC cold plate and feeds it back to the PWM unit.

[0016] The beneficial effects of this invention are as follows: Using a semiconductor cold plate (TEC chip) as the cooling (heating) source enables rapid and pollution-free switching between cold and heat sources. By constructing a precise closed-loop feedback temperature control system, a fixed or dynamically variable cooling (heating) environment can be accurately generated based on user-set parameters and the real-time status of the cooling (heating) unit, achieving efficient and precise adjustment of specific parts. In terms of energy management, the device supports power supply from an internal lithium battery, AC power, and solar panels, realizing multi-dimensional energy input and storage, greatly enhancing the device's endurance and environmental adaptability. Furthermore, the device has end-to-end data acquisition and networking capabilities, uploading usage records, multi-dimensional sensor data, interaction content, and emotion scores to a server terminal for continuous optimization of a unified intelligent health model, ensuring that the device provides personalized adjustment solutions while possessing continuous learning and iteration capabilities.

[0017] Personalized and adaptive physiotherapy is achieved through a multimodal analysis module. The emotion analysis module (BERT) uses the BERT (Bidirectional Encoder Representations from Transformers) semantic model to innovatively construct an interrogation-based interactive scenario, identifying and evaluating the user's subjective emotional state and preferences in real time, quantifying the subjective experience of "comfort" into an emotion score. The perception module (PCA) collects multi-channel physiological data such as heart rate, body temperature, electromyography signals, and blood pressure, and, combined with the application site and device settings, dynamically optimizes sensor weights through principal component analysis (PCA) to ensure the accuracy and relevance of data collection in different physiotherapy scenarios. The image analysis module (CNN) uses a convolutional neural network (CNN) structure to process image information acquired by infrared and thermal imaging sensors, enabling feature extraction and parameter quantification of key physiological parameters such as temperature distribution, blood flow expansion, and swelling at the target site. The analysis unit uses {structured questions, user responses, quantified emotion scores, multimodal sensor information, and precise adjustment measures} as the core data structure to construct a high-dimensional knowledge graph, significantly improving the adjustment accuracy and adaptive capability of the intelligent hot and cold touch device. The optimization and training of the knowledge graph are completed on the server side, while the analysis unit generates real-time, high-confidence adjustment suggestions based on the local reasoning model.

[0018] Employing a multi-regional modular design, it includes various contact components such as a lumbar support, wristband, and knee brace to adapt to the therapeutic needs of different parts of the body. Each cooling (heating) unit consists of a TEC cooling plate and is precisely controlled by an ADN8835 chip to achieve forward and reverse current switching, thus quickly switching between cooling and heating modes. Power control of the ADN8835 chip is achieved through pulse width modulation (PWM), combined with a thermistor and analog-to-digital converter (ADC) module to implement a feedback-based temperature control mechanism, ensuring high precision and stability of the cooling and heating environment. In cooling mode, the cooling fan is independently managed by a MAX6650 chip, achieving dynamic and energy-saving heat dissipation control to ensure the TEC cooling plate operates continuously at high efficiency. In heating mode, the fan is silent, achieving energy saving through natural convection.

[0019] The intelligent hot and cold touch device uses an embedded Linux operating system as its control core, possessing powerful data processing, network transmission, and control capabilities. The control unit is responsible for receiving user input (including voice and human-computer interaction), data acquisition, network transmission, and independent, parallel control of the cooling (heating) units. Through precise matching of the PWM / ADC modules, the control unit can obtain the temperature status of the thermistor of each cooling unit in real time, achieving millisecond-level precise temperature control based on temperature feedback. This high-precision feedback mechanism not only ensures the safety and comfort of the user, but more importantly, it provides a highly reliable and timely underlying data foundation for the AI ​​models (BERT / PCA / CNN) of the analysis unit, which is a key guarantee for continuous model optimization and knowledge graph optimization.

[0020] This invention utilizes contact-type TEC cooling pads applied to specific areas such as muscles, joints, bones, and surrounding blood vessels designated by the user. It effectively relieves postoperative pain, restores bodily function, relieves muscle fatigue, and regulates local blood circulation according to settings. The intelligent hot and cold touch device has data collection and analysis capabilities, enabling it to gather user data during use, analyze treatment effects, provide feedback, and automatically adjust treatment parameters to optimize results. It can form a unified and efficient health model, which can be widely applied in the health and wellness field. This device is characterized by its intelligence, wearability, pollution-free operation, and reusability. It can generate both hot and cold modes according to user requirements. By collecting user biometric information, emotional information, and information about the surrounding environment, it analyzes bodily functions, establishes a feature model, and provides optimized usage suggestions. This device can provide a scientific, safe, and healthy user experience in the fields of medicine, sports, physiotherapy, and personal health, promoting the development of personal and public health. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the intelligent hot and cold touch device. Figure 2This is a schematic diagram illustrating the working principle of the Touch Comfort module; Figure 3 This is a schematic diagram of the control circuit for the TEC cooling plate; Figure 4 This is a schematic diagram of the control circuit for the radiator; Figure 5 This is a block diagram illustrating the principle of the multimodal analysis module; Figure 6 This is a schematic diagram of the knowledge graph structure; Figure 7 This is a flowchart of the usage process for the intelligent hot and cold touch device. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 This embodiment discloses a multimodal fusion adaptive intelligent hot and cold touch therapy device. This device aims to provide an intelligent, adaptive, and precise hot and cold therapy environment, such as... Figure 1 As shown, the device includes a touch-relaxation module, a control module, a multimodal analysis module, and a power supply module. The touch-relaxation module employs efficient temperature control technology, enabling rapid and precise temperature adjustment directly applied to designated areas such as muscles and joints. The multimodal analysis module is the core unit of the device; it uses intelligent algorithms to fuse and analyze user feedback and physiological sensor data, and combines this with a knowledge graph structure to generate an emotion score and personalized adjustment suggestions, achieving closed-loop optimization of the hot and cold touch experience. The control module is responsible for system coordination, temperature control, and safety monitoring. The power supply module provides power to the other modules. This device is wearable, pollution-free, and reusable, making it suitable for exercise relaxation, emotional relief, and personal health management, providing a scientific, safe, and comfortable user experience.

[0024] In this embodiment, the Touch Comfort module includes a TEC cold plate, a heat sink, a thermistor, and a control chip. The TEC cold plate is a cooling (heating) source and includes a contact surface and a heat dissipation surface, such as... Figure 2 As shown, in cooling mode, the contact surface cools (also known as the cold stage), and the heat dissipation surface heats (also known as the hot stage). The radiator dissipates heat from the heat dissipation surface, and the current flows from TEC+ to TEC-. In heating mode, the contact surface heats (also known as the hot stage), and the heat dissipation surface cools (also known as the cold stage). The radiator is not working, and the current flows from TEC- to TEC+. The contact surface and heat dissipation surface are the same in both cooling and heating modes. The direction of the current determines whether the contact surface cools or heats, and the cooling fan only starts in cooling mode.

[0025] In this embodiment, the control chips include the ADN8835 chip and the MAX6650 chip. (Including, for example...) Figure 3As shown, the TEC (Cooling Device) cooler is controlled by an ADN8835 chip. The ADN8835 chip, based on the PWM signal output from the control unit, controls the power supply time and direction of the 5V supply voltage to the TEC chip, thereby controlling the cooling (heating) mode and intensity of the TEC cooler. A thermistor is in close contact with the TEC cooler surface to provide feedback on the cooler temperature; the thermistor is also connected to a protection resistor for short-circuit protection and voltage division. The control module includes an ADC (Analog-to-Digital Converter) unit connected to the thermistor. The ADC unit obtains the TEC cooler contact surface temperature based on the voltage division change caused by the thermistor's resistance change. In this embodiment, an NTC thermistor is selected, operating in a wide temperature range of -45℃ to 85℃, with a temperature sensitivity of ±2% (1℃).

[0026] like Figure 4 As shown, the MAX6650 chip manages the cooling fan to achieve heat dissipation from the heat sink. The MAX6650 chip only operates in hot-drying mode. The parameters of the MAX6650 chip are determined by the low-speed I2C interface of the control unit. The fan speed is set according to the parameter control port TACH0, and the fan status is monitored according to the feedback port FB. The MAX6650 chip is selected by the control unit's GPIO interface, enabling independent control.

[0027] In this embodiment, the working units of the Touch Comfort module are designed in various forms, including contact modules such as lumbar supports, wristbands, knee braces, neck pillows, and patches, to adapt to the application needs of different parts of the body. Furthermore, the hot and cold compress parameters and heat dissipation of different Touch Comfort modules can be controlled independently. The ADN8835 chip is specifically designed for TEC cooling plate control. Based on the PWM output frequency provided by the control module, it performs precise power control on the TEC cooling plate, thereby achieving accurate temperature control of the target area. The heat dissipation section mainly uses a fan to quickly dissipate heat, maintaining the device's efficiency and safety. The cooling fan is controlled by the MAX6650 chip, which connects to the I2C interface of the intelligent hot and cold Touch Comfort device control unit and is responsible for receiving and processing control commands. In addition, the parameter reception of the MAX6650 chip is controlled by the control unit's I / O interface and can be configured independently.

[0028] The control module uses an embedded Linux system as its control platform, possessing rich peripheral interfaces and flexible software expansion capabilities. It can realize functions such as device-side learning and inference, cooling and heating control, sensor data processing, and security management. The control module includes a PWM unit, an ADC unit, and a multimodal interface unit. The PWM unit is connected to the touch comfort module and is used to send PWM signals to it. The ADC unit is connected to a thermistor and is used to receive the TEC cooling plate temperature fed back by the thermistor and feed it back to the PWM unit. The multimodal interface unit is used to receive environmental information, user information, and user interaction information collected by peripheral sensors and transmit them to the multimodal analysis module. In this embodiment, the multimodal interface unit includes a human-machine interface, a voice interaction interface, and a sensor interface. The human-machine interface is used to receive basic user settings, including selecting the cooling (heating) unit, setting the working mode, and adjusting temperature and treatment parameters. The voice interaction interface is used to capture real-time user feedback during use, assessing user comfort through emotion recognition and language parsing, thereby assisting in optimizing the cooling and heating stimulation strategy. The sensor interface collects environmental and user information, such as heart rate, blood oxygen status, blood pressure status, electromyography signals, bone temperature, vasodilation, swelling, cooling (heating) unit temperature, and cooling (heating) unit efficiency.

[0029] The control module can have multiple PWM units and ADC units. Each PWM unit and ADC unit are matched to independently control a specific cooling (heating) unit, forming a feedback-based cooling (heating) environment. GPIO interfaces and low-speed I2C interfaces select and control the cooling (heating) unit's cooling fan. Alternatively, the PWM units can be configured to independently drive multiple working units, creating a dynamic, zoned, and finely adjustable cooling and heating environment. The ADC unit samples the thermistors attached to each cooling (heating) unit, monitoring temperature changes in real time to achieve precise closed-loop control based on feedback.

[0030] The control module can connect to the server-side analysis unit via a network to enable model updates, knowledge graph expansion, and personalized parameter learning, giving the device continuous optimization capabilities and long-term adaptability.

[0031] The Touch Comfort module is the main power-consuming unit, and it has two power supply methods: external power and internal power. In external power supply mode, 220V AC power is converted to 5V to drive the cooling (heating) unit and simultaneously charge the internal lithium battery. In internal power supply mode, the intelligent cooling and heating Touch Comfort device's internal lithium battery can directly generate 5V to drive the cooling (heating) unit; the lithium battery can be charged by an external 220V source or a solar panel. External power supply is suitable for high-intensity, stable scenarios, such as gyms, ambulances, hospital wards, and physiotherapy centers. The internal lithium battery is suitable for low-intensity, temporary scenarios, such as outdoor sports, post-race recovery, home use, and in-vehicle use.

[0032] The multimodal analysis module is the core module for achieving personalized, adaptive physiotherapy. Deployed on a backend server, its input data is generated by the control unit's sensors, voice interface, and human-computer interaction interface, and provides analysis and decision-making services to the control unit via the network. For example... Figure 5 As shown, the multimodal analysis module consists of a sentiment analysis unit, a perception unit, an image analysis unit, and a knowledge graph, enabling multimodal deep perception of user status.

[0033] The sentiment analysis unit employs the BERT model for sentiment analysis, providing a user-interactive "question-and-answer dialogue" format to analyze user emotions during the operation of the intelligent hot and cold touch device and collect usage data. The BERT model first poses a question and obtains user responses, then performs analog-to-digital conversion on the responses. In the BERT preprocessing stage, user responses are categorized into {vocabulary, location, sentence segmentation} information, normalized, and useless information is discarded, while key information is encoded. BERT has multiple encoding layers, each receiving vector encodings from the preprocessing stage or the previous encoding layer, and referring to pre-trained knowledge graph parameters for sentiment analysis. The encoding layers consist of a multi-head self-attention layer, a normalization layer, a discarding layer, a reverse layer, and an output layer. After processing through multiple encoding layers, a sentiment score is generated. For example, BERT is used to parse and understand user responses to questions such as "Is the knee contact point too cold?", "Does the application frequency of patch A need to be increased?", "Should the wristband duration be increased?", and "Does the usage time need to be adjusted?". By filtering and extracting information such as keywords, tone, and pitch from the responses to these questions, user emotions and preferences are analyzed, generating a sentiment score.

[0034] In this embodiment, the BERT model not only analyzes the emotional tendencies in the user's language, but also combines knowledge graphs and real-time sensor data to provide the user with a feedback-based automatic optimization solution. Users can make final decisions based on these intelligent suggestions, thereby achieving a more personalized and precise treatment experience.

[0035] The sensing unit collects multi-channel physiological data such as heart rate, body temperature, electromyography signals, and blood pressure. Combined with the application site and device settings, it dynamically optimizes sensor weights using a principal component analysis model. The optimized sensor weights are then transmitted to the knowledge graph unit, ensuring the accuracy and relevance of data collection across different therapeutic scenarios. The sensing unit improves the accuracy and efficiency of data collection, ensuring more accurate capture of key data and enhancing the responsiveness and data processing capabilities of the intelligent hot and cold touch device.

[0036] like Figure 5As shown, the sensing unit uses a PCA model. It first acquires data from multiple environmental and biosensor data, assigning different sensor weights based on the applicable location. For example, for the knee location, the weights, from highest to lowest, are: heart rate, blood oxygen saturation, knee vascular blood pressure, knee bone temperature, knee electromyography, knee swelling, and other information. The PCA weights are {25%, 20%, 15%, 15%, 10%, 10%, 5%}. This location information is then collected and transmitted to a knowledge base.

[0037] The image analysis unit uses a CNN model to process sensor information acquired through infrared and thermal imaging technologies. These sensor technologies can acquire detailed data on key physiological parameters of the treatment site, such as swelling, bone temperature distribution, and vasodilation, to control and enhance knowledge graph decision-making. First, infrared and thermal imaging sensors acquire raw image data of the target area. This image data contains important physiological and pathological information about the treatment area, crucial for monitoring and adjusting the treatment process. Then, this raw data is input into a trained CNN model. The CNN model effectively processes this image data, extracting important features through its deep convolutional layers, performing necessary image enhancement, feature recognition, and quantification analysis. The data processed by the CNN model not only includes the intuitive features of the image but also quantifies these features, facilitating further analysis and application. The processed data is systematically stored in a knowledge base, becoming part of the knowledge graph. In this way, the device can more quickly identify similar symptoms and conditions in future use, enabling more precise adjustments to treatment strategies.

[0038] like Figure 5 As shown, the CNN model processes the image including: data augmentation and edge detection. Then, it extracts the Region of Interest (ROI) and performs morphological filtering using methods such as erosion, dilation, opening, and closing to enhance key information. Finally, it quantizes the key information to form the image acquisition information.

[0039] In this embodiment, the knowledge graph unit constructs a knowledge graph with various nodes and relationships after pre-training. The nodes include question nodes, response nodes, sensor nodes, image nodes, emotion nodes, strategy nodes, execution result nodes, device mode nodes, treatment mode nodes, and historical treatment process nodes. The relationships include co-occurrence relationships, causal influence, parameter dependence, and temperature control feedback. The knowledge graph unit performs reasoning and retrieves similar states based on the graph structure, generates adjustment strategies, and supports incremental learning of subsequent models.

[0040] like Figure 6As shown, the knowledge graph index consists of {questions, answers}, organizing user status, sensor information, and control strategies through question-and-answer interaction. Each time a question is posed, the system automatically associates historical data nodes and retrieves similar scenarios from the knowledge graph, enabling prediction of user status and strategy foresight.

[0041] When a user responds with "knee adjustment enhanced," the system acquires information such as knee blood pressure, bone temperature, local blood flow, muscle tension, heart rate, and historical emotional scores to form a "current perception node." The knowledge graph generates optimization strategies based on the similarity between these perception nodes and historical nodes, such as "adjust downwards by 1°C for 120 seconds." After user confirmation, this strategy is recorded as a "strategy execution node," forming a complete knowledge path.

[0042] When the user responds with "knee adjustment held," the system combines the current sensing node with historical "holding-type nodes" to infer and generate a strategy: "Cyclical adjustment every 10 seconds: decrease by 1°C → recover → increase by 1°C, lasting 120 seconds." This strategy is labeled as "microcirculation stimulation - constant regulation" in the knowledge graph and is connected to relevant nodes through flexible relational edges for subsequent adaptive optimization.

[0043] When a user responds with a decrease in knee adjustment, the system retrieves risk nodes related to "too cold" and "too strong stimulation," and generates a strategy based on causal chain inference: "Adjust upwards by 1°C for 120 seconds." This strategy is then confirmed by the user and recorded in the knowledge graph. These strategy nodes are tagged with "risk correction" to prevent repeated use in inappropriate situations, thus improving safety.

[0044] After the 120-second strategy execution period ends, the system will automatically initiate the next round of questions. The new question node is connected to the previous strategy execution node through a "time series edge," thereby establishing a continuous state path for the user at different treatment stages, supporting trend analysis, anomaly detection, and long-term optimization.

[0045] To enhance the system's reasoning capabilities, the knowledge graph comprises seven core node types: question nodes, response nodes, sensor nodes, image nodes, sentiment nodes, policy nodes, and execution result nodes. It also defines five types of logical relationships: causal edges, temporal edges, weighted edges, semantic edges, and image feature edges, constructing a multi-dimensional, multi-modal association structure to improve the reliability and accuracy of cold (hot) application control decisions.

[0046] The reasoning process of the knowledge graph adopts a hierarchical association mechanism, namely sensor node → image node → emotion node → policy node, and outputs the current policy selection after weighted path search. After the policy is executed, the system writes information such as physiological changes, image changes, temperature feedback, and emotional fluctuation trends into the "execution result node" to update the policy success rate and achieve dynamic learning.

[0047] The knowledge graph also stores temperature control logic mapping rules, such as "increased temperature difference → enhanced stimulation", "decreased blood flow → increased probability of discomfort", and "uneven temperature → risk node triggering", to constrain strategy selection and ensure the safety of medical treatment and physiotherapy.

[0048] Emotional scores are primarily used to train large models and build long-term user preferences, and are not used as a direct basis for strategy decisions. By continuously accumulating emotional data, the system can generate personalized comfort models, providing auxiliary weights in subsequent question-and-answer and strategy generation, making the long-term treatment experience more stable and adapted to user needs.

[0049] like Figure 7 As shown, the usage process of the intelligent hot and cold touch device described in this embodiment is as follows: The user wears the corresponding cooling (heating) unit and starts the intelligent hot and cold touch device. The control unit of the intelligent hot and cold touch device performs a self-test, loads the initial treatment program according to the user's settings, the analysis unit initializes, and starts the BERT, PCA, and CNN models and related resources. During the operation of the intelligent hot and cold touch device, the control unit monitors the cooling (heating) unit to ensure a continuous and stable hot and cold environment; the analysis unit continuously learns and optimizes to enrich the knowledge graph; the cooling (heating) unit controls the cooling pads according to the control unit and provides feedback on the status. Throughout the process, the analysis unit raises questions for the user to answer, and provides optimization suggestions based on the user's answers. When the treatment is completed, the control unit shuts down the intelligent hot and cold touch device, stops the cooling (heating) unit and the analysis unit, and completes the program.

[0050] The above description is merely the basic principle and preferred embodiment of the present invention. Improvements and substitutions made by those skilled in the art based on the present invention are within the scope of protection of the present invention.

Claims

1. A multimodal fusion adaptive intelligent hot and cold touch device, characterized in that: It includes a touch comfort module, a control module, and a multimodal analysis module; The Touch Comfort module is used to provide cooling or heating to the target area; The control module includes a PWM unit and a multimodal interface unit. The PWM unit is connected to the touch module and is used to control the working state of the touch module. The multimodal interface unit is used to receive environmental information, user information, and user interaction information collected by peripheral sensors and transmit them to the multimodal analysis module. The multimodal analysis module is connected to the control module and is used to receive data from the control module, analyze it, generate adjustment strategies, and return them to the control module. The multimodal analysis module includes a sentiment analysis unit, a perception unit, an image analysis unit, and a knowledge graph unit. The sentiment analysis unit constructs an interrogation-based interactive scenario, identifies and evaluates the user's subjective emotional state and preferences in real time, and quantifies the subjective emotional state and preferences into an emotion score. The emotion score, along with the questions and user responses in the interactive scenario, is transmitted to the knowledge graph unit. The sensing unit collects multi-channel physiological data and, in conjunction with the application site and device settings, dynamically optimizes the sensor weights through a principal component analysis model, transmitting the sensor weight optimization results to the knowledge graph unit. The image analysis unit acquires image information of the target area, performs feature extraction and parameter quantization on the image information, generates image analysis results, and transmits them to the knowledge graph unit; After pre-training, the knowledge graph unit constructs a knowledge graph with various nodes and relationships. Nodes include question nodes, response nodes, sensor nodes, image nodes, emotion nodes, strategy nodes, execution result nodes, device mode nodes, treatment mode nodes, and historical treatment process nodes. Relationships include co-occurrence relationships, causal influence, parameter dependence, and temperature control feedback. The knowledge graph unit performs reasoning and retrieves similar states based on the graph structure, generates adjustment strategies, and supports incremental learning of subsequent models.

2. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 1, characterized in that: The sentiment analysis unit uses the BERT model. First, it poses a question and obtains user responses. Then, it performs analog-to-digital conversion and word segmentation on the user responses, dividing them into {vocabulary, position, sentence segmentation} and normalizing them. Useless information is discarded, and the remaining information is encoded. The BERT model has multiple encoding layers. Each layer receives vector encoding from the word segmentation stage or the previous encoding layer and performs sentiment analysis with reference to the parameters of the pre-trained knowledge graph. After processing through multiple encoding layers, a sentiment score is generated.

3. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 2, characterized in that: The BERT model combines knowledge graphs with real-time sensor data to provide users with a feedback-based automatic optimization solution, which users then use to make their final decisions.

4. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 1, characterized in that: The image analysis unit uses a CNN model to process and analyze the image information acquired by the image sensor. First, it acquires the user's original image, then performs data enhancement and edge detection on the acquired image, extracts the ROI region and performs morphological filtering, and finally performs feature quantization on key information to form image acquisition information.

5. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 1, characterized in that: The knowledge graph uses a hierarchical association mechanism for reasoning, following the path from sensor node to image node to sentiment node to policy node, and outputs the current policy selection after a weighted path search.

6. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 5, characterized in that: After the strategy is executed, physiological changes, image changes, temperature difference feedback, and emotional fluctuation trends are written into the execution result node to update the strategy success rate and achieve dynamic learning.

7. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 1, characterized in that: The knowledge graph unit defines relationships as causal edges, temporal edges, weighted edges, semantic edges, and image feature edges.

8. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 1, characterized in that: The Touch Comfort module includes a TEC cooling plate and a heat sink. The TEC cooling plate includes a contact surface and a heat dissipation surface. In cooling mode, the contact surface cools, the heat dissipation surface heats, and the heat sink dissipates heat from the heat dissipation surface. In heating mode, the contact surface heats, the heat dissipation surface cools, and the heat sink does not work. The switching between cooling and heating modes is controlled by the direction of the current flowing through the TEC cooling plate.

9. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 8, characterized in that: The Touch Comfort module also includes the ADN8835 chip and the MAX6650 chip. The ADN8835 chip controls the operation of the TEC cold plate, and the MAX6650 chip controls the operation of the heat sink.

10. The multimodal fusion adaptive intelligent hot and cold touch device according to claim 8, characterized in that: The TEC cold plate is equipped with a thermistor. The TEC cold plate is connected to the PWM unit through the ADC unit. The thermistor collects the temperature of the TEC cold plate and feeds it back to the PWM unit.

Images