An intelligent interaction system based on motion line triggering

Abstract

This invention relates to a movement-triggered intelligent interactive system. The system includes a positioning module, a voice acquisition module, an interactive agent, and a model optimization module. The positioning module collects location information and movement status in real time; the voice acquisition module collects voice information in real time; the interactive agent determines the risk level of the scene based on location and movement status, whether the location exceeds the recording area, and whether the voice information is valid, as well as whether vital signs are abnormal. If the triggering conditions are met, the interactive agent is activated and a confidence score is calculated. Then, a confidence score threshold is used to compare and match corresponding execution commands, which can directly execute operations, interact with the elderly person, or trigger safety responses to push warnings to caregivers. This invention aims to solve the problems of inconvenience, incomplete scene coverage, and delayed emergency response caused by the manual wake-up mode of existing elderly-friendly intelligent devices.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice interaction technology, specifically to an intelligent interaction system based on movement-triggered operation. Background Technology

[0002] Currently, the intelligent upgrading of elderly care services has become an industry trend, leading to the emergence of various smart devices specifically designed for the elderly, such as smart bracelets, smartwatches, and portable voice terminals. These devices typically integrate core functions such as location tracking, health monitoring, and voice interaction, aiming to improve the safety and convenience of the elderly's lives through technological means, and reduce the workload of caregivers. They have gradually become important auxiliary tools in home-based and institutional elderly care scenarios.

[0003] In the existing interactive design of smart devices, voice interaction is one of the core interaction methods. Its wake-up mechanism generally adopts a manual active wake-up mode, requiring elderly users to actively trigger the device's interactive function by using a specific wake-up word, such as "Little X, open XX," or by pressing a physical button or clicking a touchscreen. Only then can they express their needs, such as asking for help, seeking information, or calling functions. This type of wake-up design originates from the interaction logic of general smart devices. When targeting younger users, because younger users have strong learning abilities and high operational proficiency, they can quickly adapt to and accurately execute the wake-up operation, resulting in a better user experience and convenience.

[0004] However, directly applying the aforementioned human-activated interaction mode to smart devices serving the elderly presents several challenges. On one hand, as people age, elderly users generally experience memory decline and slower reaction times, making it difficult for them to accurately memorize and pronounce specific wake words. Even after repeated learning, they are prone to forgetting wake words or making pronunciation errors, leading to wake-up failure. On the other hand, some elderly users who are frail, disabled, or semi-disabled, such as those with limited mobility or Parkinson's disease, lack sufficient physical dexterity to reliably perform active operations such as pressing physical buttons or accurately clicking on the touchscreen, further hindering the realization of the wake-up function.

[0005] In summary, manual or physical wake-up, as a necessary prerequisite for voice interaction, increases the usage threshold and operational complexity for elderly users. For elderly users unfamiliar with the operating logic of smart devices, this can easily lead to wake-up failure or abandonment of use. Therefore, it is necessary to continue developing a functional system that can automatically wake up and initiate interaction without requiring active operation from the elderly user. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an intelligent interactive system based on movement line triggering, in order to solve the problems of inconvenience in use, incomplete scene coverage, and delayed emergency response caused by the manual wake-up mode of existing smart devices dedicated to serving the elderly.

[0007] This invention provides an intelligent interactive system based on movement path triggering, comprising:

[0008] The positioning module is used to collect location information and motion status in real time; The voice acquisition module is used to acquire voice information in real time; An interactive intelligent agent is configured with triggering conditions. When the triggering conditions are met, the interactive intelligent agent is triggered and a confidence score is calculated based on the location information, motion state, and voice information. Based on the confidence score, an execution instruction in a preset instruction library is matched. The execution instruction includes an execution response, an interactive query, and a security response. The model optimization module records the location information, motion state, voice information, and corresponding actual execution instructions as a set of historical interaction data. It selects samples that meet preset conditions from the historical interaction dataset as reinforcement training samples and generates a reinforcement training sample set to train the interactive agent.

[0009] By using multi-module collaboration, the system integrates and collects movement and voice data. Based on the integrated data, it determines whether the triggering conditions are met and whether there is no wake-up trigger. It also calculates the confidence level of the line and executes the corresponding instructions to meet the intelligent care needs of the elderly.

[0010] Furthermore, the interactive intelligent agent includes a confidence calculation submodule, specifically comprising: Multiple scoring items are obtained, including five preset basic scores across five dimensions: speech recognition clarity, intent recognition accuracy, scene adaptability, habit fit, and interactive feedback effectiveness. Specifically, speech recognition clarity reflects the clarity of the speech information converted into text; intent recognition accuracy reflects the degree of matching between the recognized intent and the actual need; scene adaptability reflects the degree of matching between the recognized intent and the current scene; habit fit reflects the degree of matching between the recognized intent and the elderly person's personal habits; and interactive feedback effectiveness reflects the consistency between the elderly person's feedback and the recognized intent. Each dimension is assigned a dynamic weight coefficient, and the confidence level is calculated based on the baseline scores of the five dimensions and the dynamic weight coefficient of each dimension.

[0011] The confidence score calculation covers a five-dimensional scoring system including voice, intent, scenario, habits, and feedback, making demand identification more comprehensive and accurate. Furthermore, the use of dynamic weighting coefficients allows the confidence score calculation to adapt to the personalized characteristics of different elderly people and different scenarios.

[0012] Furthermore, the dynamic weight coefficients are dynamically optimized based on the reinforcement training sample set, specifically as follows: For each set of reinforced training samples in the reinforced training sample set, the prediction confidence is calculated based on the initial weight coefficient. The prediction confidence is matched with the corresponding execution instruction and compared with the actual execution instruction in the reinforced training sample to see if they are consistent. If they are inconsistent, they are judged as errors. The initial weight coefficient is the dynamic weight coefficient obtained in the previous training. The corrected dynamic weight coefficients are obtained by calculating the error and based on the weight correction model.

[0013] The continuous optimization of dynamic weighting coefficients makes the confidence calculation more and more accurate, gradually reducing the system's recognition error of the elderly's needs and improving the reliability of instruction matching.

[0014] Furthermore, the interactive intelligent agent is configured with a confidence matching instruction strategy, specifically as follows: There are preset high confidence thresholds and low confidence thresholds, and the high confidence threshold is greater than the low confidence threshold. The confidence level is compared with both the high confidence threshold and the low confidence threshold. If the confidence level is higher than the high confidence threshold, a matching response is executed; if it is between the high confidence threshold and the low confidence threshold, an interactive query is matched; if it is lower than the low confidence threshold, a security response is executed.

[0015] Establish a one-to-one correspondence between confidence intervals and execution instructions, construct clear and practical quantitative rules, and achieve standardization and intelligentization of instruction matching.

[0016] Furthermore, the confidence matching instruction strategy includes a sub-strategy for dynamically adjusting the confidence threshold, specifically: The scenario is matched based on the location information, and the scenario risk level is obtained by matching the scenario with the scenario risk mapping table stored in the database; and the elderly’s vital signs information is obtained based on the external vital signs monitoring module of the system, and the elderly’s vital signs information is compared with the preset vital signs information to determine whether the vital signs information is abnormal. If the risk level of the scenario changes or the vital signs information becomes abnormal, the low confidence threshold will be dynamically adjusted. The mean confidence score is calculated based on the confidence scores in the historical interaction dataset. The mean confidence score is compared with a threshold. If it is lower than the threshold, the high confidence score threshold and the low confidence score threshold are dynamically adjusted.

[0017] Setting a threshold dynamic adjustment sub-strategy allows the confidence threshold to be dynamically adjusted, better adapting to the risks of the scenario and changes in the health status of the elderly, and enabling the instruction matching strategy to fit the real-time health status of the elderly and the risk level of the scenario.

[0018] Furthermore, the triggering conditions include: The scenario is a risk scenario with a risk level higher than the risk level threshold. By comparing the location information with the pre-stored location areas of elderly care institutions, it is determined that the location exceeds the location area of ​​elderly care institutions; The voice information is determined to be valid voice using a voice determination sub-strategy. The elderly person's vital signs information is compared with the preset threshold for elderly person's vital signs to determine if there are any abnormalities.

[0019] The design covers four dimensions: risk scenarios, abnormal location, proactive voice, and health abnormalities, with no wake-up trigger conditions. This allows the system to be activated automatically when the elderly have a need or face a risk, without requiring the elderly to cooperate in waking it up.

[0020] Furthermore, the voice determination sub-strategy is specifically as follows: Semantic text is obtained from the voice information, and the semantic text is matched with a preset keyword library. If the semantic text contains keywords, the voice information is valid. If the semantic text does not contain keywords, the scene is obtained through the location information, and the historical interaction count of the scene is matched from the historical interaction database. If it is a high-frequency interaction scene with a historical interaction count higher than the interaction threshold, then the voice information is valid voice.

[0021] The two-layer judgment logic of the voice judgment sub-strategy is clearly defined: judgment is made by matching keywords in the voice information; if no keywords are found, judgment is made by the scene interaction situation, so as to achieve accurate judgment of valid voice.

[0022] Furthermore, the interactive intelligent agent includes a general interactive intelligent agent for institutions and a personal interactive intelligent agent for the elderly; The general interactive intelligent agent of the institution is pre-trained through a pre-trained model. The data used by the pre-trained model includes speech recognition, elderly care knowledge base, medical and health common knowledge base, nursing home map layout, personnel information, and schedule data, in order to provide basic capability support for the elderly personal interactive intelligent agent. The elderly person's personal interactive intelligent agent is based on the pre-trained model of the institution's general interactive intelligent agent. It is trained in a personalized way by collecting the elderly person's voice information, movement data, and interaction feedback data, so as to learn the elderly person's voice habits, behavioral preferences and health taboos. The general interactive intelligent agent of the institution and the personal interactive intelligent agent of the elderly interact in real time. The personal interactive intelligent agent of the elderly calls the data generation instructions of the general interactive intelligent agent of the institution. The general interactive intelligent agent of the institution summarizes the common needs of multiple personal interactive intelligent agents of the elderly and iteratively optimizes the pre-trained model.

[0023] The design employs a two-tiered intelligent agent architecture, consisting of a general-purpose interactive intelligent agent for institutions and a personal interactive intelligent agent for elderly individuals. This architecture separates and integrates generality and personalization. The general-purpose intelligent agent uniformly carries the institution's basic capabilities, avoiding redundant training and reducing the model training cost for elderly care institutions. The personal intelligent agent is trained on the general model to achieve personalized interaction adaptation for each individual.

[0024] Furthermore, the preset conditions used in the model optimization module for selecting reinforcement training samples include: The same elderly person interacts with the same scenario more than the threshold number of times, and the subsequent interaction corrects the result of the previous interaction. If the confidence level is higher than the high confidence threshold and the execution response is completed, the judgment is determined to be incorrect based on feedback from the elderly person. If the confidence level is below the low confidence threshold and a safety response is completed, the risk is determined to be no based on feedback from the elderly person. Low-frequency interaction scenarios with a historical number of interactions below the interaction threshold, and after completing a safety response, are judged to have a risk based on the feedback results.

[0025] Targeted sample selection avoids indiscriminate collection of historical data, reduces the interference of redundant data on model training, improves the efficiency of model optimization, and reduces the consumption of computing resources.

[0026] Furthermore, the positioning module is configured with a location acquisition strategy, specifically: Fixed-position interactive units are evenly installed at room entrances, corridor corners, stairwells, and along railings on each floor of the elderly living facility, and the Bluetooth frequency bands of the fixed-position interactive units on different floors are differentiated. The fixed-position interaction unit interacts with the transceiver unit built into the positioning module, enabling the positioning module to obtain location information based on topological relationships and signal strength.

[0027] By deploying fixed-location interactive units across all scenarios, blind-spot-free coverage of location signals within elderly care facilities is achieved, solving the signal blind-spot problem of traditional positioning. Furthermore, the differentiated Bluetooth frequency band design for different floors avoids floor confusion in multi-story buildings from a hardware perspective, improving the accuracy of floor determination.

[0028] The beneficial effects of this invention are: This invention achieves accurate identification of the interaction needs of the elderly in nursing homes, rapid response, and proactive early warning of potential risks by using a positioning module for precise location acquisition and a voice acquisition module for effective contextual voice judgment. It combines a wake-up-free triggering mechanism with a confidence dynamic calculation and matching strategy to achieve intelligent protection. This not only conforms to the usage habits of the elderly but also meets the core needs of nursing homes for proactive care and safety, greatly improving the intelligence and safety of age-friendly interaction.

[0029] This invention achieves a deep integration of institutional universality and elderly personalization through a two-layer interactive intelligent agent. Relying on the targeted sample selection of the model optimization module and the dynamic optimization and adjustment of weights in the confidence calculation, as well as the dynamic adjustment of the confidence threshold, the system can continuously learn the elderly's voice habits, behavioral preferences and scene characteristics, achieving an adaptive effect that becomes more accurate and more suitable with use. At the same time, the modular architecture design and quantitative judgment and calculation rules take into account the system's feasibility, scalability and ease of operation and maintenance. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the logic of an intelligent interactive system based on motion-triggered interaction, showing the logic of triggering the interactive agent, calculating confidence, and matching and executing instructions. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0034] like Figure 1 As shown, this invention provides an intelligent interactive system based on movement path triggering, the system comprising: The positioning module is used to collect location information and movement status in real time. It is configured with a location acquisition strategy to accurately capture the real-time location of elderly residents in the multi-story building environment of senior living facilities and simultaneously record their movement status. Specifically: Fixed-position interaction units are evenly installed at room entrances, corridor corners, stairwells, and along railings on each floor of the elderly living facility. The Bluetooth frequency bands of the fixed-position interaction units on different floors are differentiated. The fixed-position interaction units interact with the transceiver unit built into the positioning module, enabling the positioning module to obtain location information based on topology and signal strength. Among them, the fixed location interaction unit is a fixed hardware device deployed in the elderly care facility to interact with the positioning module in the device worn by the elderly. Before installation and deployment, each fixed location interaction unit will store its own core information in advance, including a unique device identifier, the floor number, the specific installation location coordinates, and the type of functional area, such as the room door or the stairwell. This information is sent to the positioning module in sync during the signal interaction process, so that the positioning module can obtain the location information of the elderly. Furthermore, to avoid signal confusion and interference between different floors, different Bluetooth communication frequency bands are assigned to the fixed-position interaction units on different floors. For example, the fixed-position interaction unit on the first floor uses the 2.401GHz frequency band, the second floor uses the 2.403GHz frequency band, the third floor uses the 2.405GHz frequency band, and so on. The Bluetooth frequency band interval for each floor is set to 0.002GHz. This interval can effectively avoid frequency band overlap interference and ensure that the positioning module can accurately identify different frequency bands. The positioning module calculates its own precise position coordinates based on the position coordinates and corresponding signal strength of each fixed-position interaction unit received, and in combination with the topological relationship pre-stored in the system. The correspondence between signal strength and distance follows the free space propagation model, so the distance between the positioning module and the fixed-position interaction unit can be calculated by reverse calculation based on the signal strength. The calculation formula for this model is:

[0035] in, The signal strength received by the positioning module; λ is the transmit power of the fixed-position interaction unit; d is the straight-line distance between the positioning module and the fixed-position interaction unit; λ is the wavelength of the Bluetooth signal.

[0036] In practical applications, when the positioning module simultaneously receives signals from at least three adjacent fixed-position interaction units, triangulation can be used to calculate precise position coordinates. For example, if the positioning module receives signals from three fixed-position interaction units A, B, and C, and the pre-stored coordinates of these three interaction units are A(…),… B ( ), C( Based on the signal strength, the distances between the positioning module and the three fixed-position interaction units were calculated as follows: , , Then, using the coordinates of the three units as vertices and the corresponding distances as radii, draw circles. The intersection of the three circles is the precise coordinates (x, y) of the positioning module. When the positioning module only receives signals from one or two fixed-position interaction units, it combines the topological relationship and signal strength to determine the relative position. For example, if the positioning module only receives signals from the fixed-position interaction unit at the entrance of the second-floor restaurant, then based on the signal strength and the model, the distance d is calculated to be 1.5 meters, and it is determined that the elderly person is currently in the area at the entrance of the second-floor restaurant. In addition, the positioning module is also equipped with GPS to obtain the elderly's location information. It prioritizes using the interaction with the fixed location interaction unit to obtain the elderly's location information. When the elderly's location cannot be obtained through the fixed location interaction unit, GPS is used to collect location information, such as when the elderly walk out of the nursing home alone and cannot obtain location information through the fixed location interaction unit. In addition to acquiring the elderly person's location information in real time, the positioning module also uses built-in motion sensors, such as a three-axis accelerometer and a gyroscope, to collect the elderly person's motion status data. By analyzing changes in acceleration and angular velocity, it can determine whether the elderly person is walking, stationary, or slowly wandering.

[0037] The intelligent interaction system also includes a voice acquisition module, which is used to collect voice information in real time. The hardware configuration of the voice acquisition module must include a high-sensitivity microphone array. The microphones are evenly distributed and installed in the wristband worn by the elderly, which can capture the elderly’s voice signals from all directions. The system also performs preprocessing operations such as noise reduction and filtering on the collected voice information to remove environmental noise, unintentional sounds such as coughing and sighing, and unrecognizable unclear speech.

[0038] The system also includes an interactive agent, which is configured with trigger conditions. When the trigger conditions are met, the interactive agent is triggered and confidence calculation is started. The confidence is calculated based on location information, motion state and voice information. Then, based on the confidence, the system matches the execution instructions in the preset instruction library. The execution instructions include execution response, interactive inquiry and security response. The triggering conditions mentioned above specifically include: The first method involves determining the elderly person's location based on their location information. A scene refers to a specific spatial area within the elderly care facility, categorized by function and risk characteristics. These areas include high-frequency interaction scenes, high-risk scenes, routine scenes, and low-frequency interaction scenes. The risk level of the scene is then determined by matching the scene with a risk level mapping table stored in the database. This mapping table is a structured data table pre-stored in the system database to define the corresponding risk level for each scene. The table contains fields such as scene identifier, scene name, risk level, and risk description. The scene risk level is categorized into three levels: low risk, medium risk, and high risk, based on the degree of safety hazards in the scene. For example, scenes such as bedrooms, living rooms, and gardens are considered low-risk; scenes such as restaurants and corridors are considered medium-risk; and scenes such as stairwells, railings, restrooms, or equipment rooms are considered high-risk. When the risk level of the scene where the elderly are located is higher than the risk level threshold, the interactive intelligent agent is activated without wake-up. The risk level threshold is a critical value for determining whether to activate the interactive intelligent agent. For example, the system can preset medium risk as the judgment threshold. If the risk level of the scene is high risk, the triggering condition is met. The second type uses GPS positioning in the positioning module to obtain the location information of the elderly, determines whether the elderly's location exceeds the pre-stored location area of ​​the elderly care institution, sets the location area of ​​the elderly care institution as a safe area, and if the current location of the elderly exceeds the safe area, it is determined that the elderly has entered an unfamiliar area, triggering the interactive intelligent agent to start without wake-up. For example, if an elderly person is wandering alone and climbs over the fence on the east side of the nursing home to enter the street, the location information of the elderly person collected by the GPS in the positioning module is not within the safe area of ​​the nursing home pre-stored in the system. The system determines that the elderly person is in an unfamiliar area and immediately triggers the wake-up-free activation of the interactive intelligent agent. The third type uses a vital sign monitoring module set in the wristband to collect the elderly’s vital sign information in real time, such as heart rate, blood pressure, and blood oxygen saturation. The collected data is compared with the preset vital sign thresholds for the elderly, which are the normal range values ​​for the elderly’s heart rate, blood pressure, and blood oxygen saturation. When the collected vital sign data exceeds the normal range, it is determined that the vital sign information is abnormal and the interactive smart agent will not be woken up. The fourth category involves using a voice determination sub-strategy to determine the validity of preprocessed voice information. If the voice is valid, the interactive agent is activated without wake-up. This voice determination sub-strategy employs a two-layer determination logic. One layer is: after converting the collected voice information into semantic text, the system matches the semantic text with the preset keyword library one by one. If the semantic text contains at least one keyword, the voice information is directly determined to be valid voice. The keyword library stores a database of words related to the high-frequency needs of the elderly in the elderly care scenario, such as core keywords related to life needs, health assistance, and equipment control. For example, when an elderly person stops at the stairwell and says, "I feel a little dizzy," the voice acquisition module converts the voice information into semantic text. The semantic text "I feel a little dizzy" contains the keyword "dizzy," which meets the first-level keyword matching of the fourth category in the triggering conditions. Therefore, it is determined to be valid voice, and the interactive intelligent agent is activated without wake-up. Another layer is: if the semantic text does not contain any keywords, the system matches the scene with the elderly person's current location information obtained by the positioning module, and matches the number of historical interactions in that scene from the historical interaction database. The historical interaction database stores all the elderly person's interaction records in various scenes over a period of time. Therefore, the number of historical interactions in the current scene can be known through the historical interaction records. If the number of historical interactions in that scene is higher than the interaction threshold, then the scene is determined to be a high-frequency interaction scene. The voice information collected in the high-frequency interaction scene is valid voice. For example, in the living room, a high-frequency activity area for the elderly, if an elderly person softly says "Turn on the power" in the living room, the voice acquisition module will capture this voice signal in real time. After preprocessing, it will be converted into semantic text "Turn on the power". The system will match this semantic text with a preset keyword library. If no keywords are found, the system will activate the second layer of the voice judgment sub-strategy, which will match the current scene as the living room. The system will retrieve the historical interaction count of this scene from the historical interaction database and confirm that it is a high-frequency interaction scene. Therefore, even if there are no keywords in the semantic text, the voice information will still be judged as valid voice, which will meet the fourth type of triggering condition of the interactive agent. The interactive agent will then start without wake-up.

[0039] The triggering conditions in the interactive agent are met, enabling it to start without wake-up, and the confidence score is calculated through the confidence score calculation submodule set in the interactive agent, specifically: Multiple scoring items are obtained, including five preset basic scores for speech recognition clarity, intent recognition accuracy, scene adaptability, habit fit, and interaction feedback effectiveness. The preset basic scores refer to the quantitative scoring results set for each of these five dimensions. The scoring range for each dimension is from 0 to 1. The higher the score, the better the performance of that dimension and the more accurate it is in supporting the demand intent. Among them, speech recognition clarity refers to the quantitative indicator of the clarity and completeness of the speech acquisition module in converting the elderly’s speech information into semantic text, reflecting the clarity of semantic acquisition; Intent recognition accuracy refers to a quantitative indicator of the degree of matching between the demand intent extracted from semantic text and the actual needs of the elderly, reflecting the accuracy of understanding the needs of the elderly. Scene adaptability refers to a quantitative indicator of the degree to which the extracted demand intention matches the functional attributes of the scene in which the elderly are currently located. For example, the adaptability of dining needs in a restaurant scene is higher than that of watching TV. Habit fit refers to a quantitative indicator of the degree of matching between the extracted demand intention and the personal behavioral habit preferences formed in the elderly’s historical interaction records. For example, if an elderly person is accustomed to taking a walk in the garden at 10 a.m., the demand fit for walking in that time and scenario is higher. Interactive feedback effectiveness refers to a quantitative indicator of the consistency between the elderly person's feedback information on the system's initial response and the needs and intentions extracted by the system. If the elderly person clearly agrees or agrees, the effectiveness is high; if they clearly disagree, the effectiveness is low. Furthermore, a dynamic weight coefficient is assigned to each dimension, with the weight of each dimension ranging from 0.05 to 0.4, and the sum of the weights of the five dimensions is 1. The higher the weight value, the greater the contribution of that dimension to the confidence calculation. The confidence score is calculated using a weighted summation method based on the baseline scores of the five dimensions and the dynamic weight coefficient of each dimension. The confidence level is calculated using the following formula: = , Where Conf is the confidence level, which ranges from 0 to 1; For speech recognition clarity; For the accuracy of intent recognition; For scene adaptation; For habit compatibility; To ensure the effectiveness of interactive feedback; , , , and The dynamic weight coefficients for the five dimensions are respectively and satisfy the following conditions: + + + + =1.

[0040] The basic scoring rules for the five dimensions in the above formula are as follows: Speech recognition clarity The calculation uses a signal-to-noise ratio weighted scoring method, and the formula is as follows: = , in the formula: The accuracy rate of the text output by the speech recognition module is determined by the proportion of matching characters between the recognized semantic text and the manually annotated standard text, relative to the total number of characters in the standard text. For example, if the standard semantic text of an elderly person's speech is "I want to drink water," and the recognized text is also "I want to drink water," then the text accuracy rate is 100%. =1; It is the signal-to-noise ratio correction factor, which is determined by the signal-to-noise ratio of the speech signal. When the signal-to-noise ratio is ≥20dB, =1; When 10dB≤Signal-to-noise ratio<20dB =0.8; when the signal-to-noise ratio is <10dB, =0.5; Intent recognition accuracy The calculation uses an intent matching scoring method. The system has a pre-built intent library that includes major categories such as life needs, health assistance, and device control. The intents are further subdivided into major categories and each sub-intent is associated with a set of keywords. The accuracy of intent recognition is determined by calculating the ratio of the number of keywords that match the semantic text to the total number of keywords for that intent. For example, if an elderly person says in the living room, "I want to drink some water," the original semantic text of the speech conversion is "I want to drink some water." The keyword group for the sub-intent of drinking water in the life needs category of the system's intent library consists of five words: drinking water, water, thirst, drinking water, and hydration. If the semantic text matches the keyword "drink water," the matching ratio is 0.2, and the accuracy of intent recognition is 0.2. Scene adaptability The calculation adopts the scene intent correlation score method. The scene intent correlation table is pre-stored to clarify the correlation weight between each scene and each sub-intent. The correlation weight is set based on the usage logic of the elderly care institution scene. For example, the correlation weight between the restaurant scene and the dining intention is 0.9, while the correlation weight between the restaurant scene and the dining intention is 0.1. The habitual compatibility score S4 is calculated using the historical match rate scoring method, and the formula is as follows: , Where n represents the number of interactions with the same intent in the current scenario and time period within the past 30 days, and m represents the total number of interactions in the current scenario and time period within the past 30 days; if there are no interaction records in the current scenario and time period within the past 30 days... It is 0.5; For example, if an elderly person has 8 interactions in the living room at 10 am in the past 30 days, 3 of which are with the intention of turning on the TV, then the habit fit S4 is 0.375; Effectiveness of interactive feedback Using a feedback result quantification method, if the elderly person clearly acknowledges the system's initial response via voice or button presses, then... Set to 1; if there is no feedback by default, then... It is 0.8; if explicitly denied, then The value is 0.2; if no feedback from the elderly is required, then the operation can be performed directly. The average feedback effectiveness for this intention in history is taken. For example, if the elderly person clearly answers "yes" after the system gives an initial response to the intention to turn on the TV, then the interaction feedback effectiveness is considered. =1; The dynamic weight coefficients in the confidence calculation formula are obtained by retrieving the latest optimized dynamic weight coefficients from the model optimization module. These dynamic weight coefficients are obtained through iterative optimization using a reinforced training sample set, and initial weights are preset based on experience with age-friendly scenarios. For example, the initial weights can be set to 0.1 for speech recognition clarity, 0.3 for intent recognition accuracy, 0.2 for scenario adaptability, 0.25 for habit fit, and 0.15 for interaction feedback effectiveness. Each subsequent optimization round uses the dynamic weight coefficients from the previous round as initial values ​​and corrects for errors to obtain new weights, ensuring that the weights continuously align with the interaction characteristics of the elderly. For example, for elderly people with cognitive impairment, the system will increase the weights of habit fit and interaction feedback effectiveness through training optimization, while decreasing the weight of speech recognition clarity. Note that the weights must be normalized after dynamic adjustment to ensure... + + + + =1.

[0041] After obtaining the confidence level, the confidence level is matched with the execution instructions in the preset instruction library using confidence level matching instruction measurement, and the corresponding instructions are executed. Specifically: The system has preset high-confidence thresholds and low-confidence thresholds. The high-confidence threshold is the confidence threshold used to determine whether to directly execute a response, representing a high reliability standard for recognizing the elderly person's needs and intentions. The default initial value is 0.8, and the value range is 0.7 to 0.9. The low-confidence threshold is the confidence threshold used to distinguish between interactive inquiries and safety responses, representing the lowest reliability standard for recognizing the elderly person's needs and intentions. The system's default initial value is 0.5, and the value range is 0.4 to 0.6. At the same time, the high-confidence threshold must always be greater than the low-confidence threshold. The obtained confidence level is compared with the high confidence threshold and the low confidence threshold respectively. If it is higher than the high confidence threshold, the execution response is matched. The execution response is the type of instruction that is directly executed when the elderly person's needs and intentions are clear and reliable based on the high confidence level. No additional feedback is required from the elderly person. For example, directly turn on the TV or adjust the air conditioner temperature. If the confidence level is between the high confidence threshold and the low confidence threshold, an interactive query is matched. The interactive query is based on the medium confidence level. When there is some uncertainty in the elderly's intention to seek help, the elderly are given clear feedback through voice confirmation or option guidance before the corresponding operation is executed. For example, if the intention is to drink water, the voice will ask if the elderly want to drink water. If the elderly answer yes, the water delivery service will be dispatched. If the confidence level is below the low confidence threshold, a safety response is executed. If the low confidence level indicates that the elderly person's needs and intentions cannot be accurately identified, or if there is a potential safety risk, risk warning and protection instructions are activated. These may include sending warnings to caregivers or providing audio and visual prompts through the wristband. For example, if the confidence level is below the low threshold, the wristband will directly send a warning to the caregiver. The caregiver's terminal will receive information including the elderly person's location, current scene, voice information, vital signs data, and confidence level.

[0042] Furthermore, the confidence matching instruction strategy also includes a sub-strategy for dynamically adjusting the confidence threshold. This sub-strategy adjusts the low confidence threshold in real time based on changes in the risk level of the elderly person's current environment and abnormal vital signs, ensuring that the threshold matches the real-time risk and health status. This avoids missed warnings in high-risk scenarios or when the elderly person is unwell due to a fixed threshold. Additionally, it calculates the average confidence level in historical interaction data and compares it with a preset average threshold. When the average confidence level falls below this threshold, the high and low confidence thresholds are dynamically lowered. This ensures that when the overall reliability of demand identification is low, reducing excessive interaction by lowering the threshold threshold improves execution response efficiency, thereby ensuring that instruction matching is both accurate and meets the needs of safe care.

[0043] The specific sub-strategy for dynamically adjusting the confidence threshold is as follows: The system collects the elderly’s location information in real time based on the positioning module, matches the elderly’s current scene based on the location information, and then retrieves the scene risk mapping table from the database. The risk level of the scene is obtained by associating it with the scene identifier. For example, if the elderly’s location information is matched to the scene at the stairwell, the risk level corresponding to the stairwell in the scene risk mapping table is high risk. And based on the judgment of abnormal vital signs information, the vital signs information of the elderly are collected through the vital signs monitoring module, including vital signs data such as heart rate, blood pressure, and blood oxygen saturation. Each vital sign data collected is compared with the corresponding preset vital signs information one by one. If any vital signs data exceeds the preset normal range, it is judged as abnormal vital signs information. For example, if the collected heart rate is 115 beats per minute, which exceeds the preset normal range of 60 to 100 beats per minute, it is judged as abnormal vital signs information. When the risk level of the scene changes or the vital signs information becomes abnormal, the low confidence threshold is dynamically adjusted, while still maintaining the constraint that the high threshold is greater than the low threshold.

[0044] The formula for dynamically adjusting the low confidence threshold is: = ;in, This serves as the initial low-confidence threshold. Threshold adjustment coefficients are configured for different risk levels in different scenarios, under low-risk scenarios. =1.0, medium-risk scenario =1.1, High-risk scenarios =1.2; Adjustment coefficients are configured for abnormal vital signs information; when vital signs information is normal... =1.0, when vital signs information is abnormal =1.1; It should be noted that if the adjusted value of the low confidence threshold exceeds its maximum boundary value, then the maximum boundary value is taken as the adjusted low confidence threshold.

[0045] For example, if the elderly person is semi-independent and has a history of hypertension, a high confidence threshold should be initially set. =0.8, low confidence threshold =0.5. One day, an elderly person walks alone to the stairwell. The location module collects the elderly person's location information and determines the current scenario to be a high-risk scenario. =1.2. The elderly person suddenly felt dizzy and mumbled that they felt unwell. The vital signs monitoring module in the device recorded that the elderly person's heart rate at this time was 108 beats per minute, which is outside the normal range and is judged as abnormal vital signs information. =1.1, the interaction agent starts and calculates the confidence level, which is Conf=0.65, because the current scenario of the elderly is a high-risk scenario and the elderly's vital signs are abnormal. If the interaction agent is triggered without wake-up, the interaction agent starts and calculates the confidence level. If the confidence level is Conf=0.65, the interaction query instruction will be executed if the initial threshold is compared. Based on the abnormal vital signs of the elderly, a dynamic adjustment of the threshold is immediately triggered, which can be calculated to obtain =0.5×1.2×1.1=0.66, then the adjusted confidence thresholds are respectively =0.66, then perform confidence matching command judgment; Based on a confidence level of Conf=0.65, which is below the low confidence threshold, a safety response command is triggered. The system pushes the elderly person's location, vital signs, and confidence level to the caregiver's terminal. Upon receiving the notification and arriving at the scene, the caregiver will find that the elderly person has elevated blood pressure, administer antihypertensive medication promptly, and assist them to a safe area to rest, preventing an accident. After a period of time, the system detects that the elderly person's scene has switched to the living room, at which point the risk level is low, and the elderly person's vital signs have returned to normal. The system then automatically adjusts the threshold back to the previous level. =0.5; Based on the model optimization module's storage of a collection of all complete historical interaction data from the past 30 days, each historical interaction data point contains location information, motion status, voice information, calculated confidence level, and corresponding actual execution instructions. All confidence level values ​​can be extracted from the historical interaction dataset, and the calculated mean confidence level is used to reflect the overall reliability level of recognizing the elderly person's needs and intentions in recent times. The mean confidence level ranges from 0 to 1. The mean confidence level is compared with a threshold; if it falls below the threshold, the high and low confidence thresholds are dynamically adjusted. The mean threshold is the critical value used to determine whether the confidence threshold needs adjustment, representing the minimum acceptable overall recognition reliability level of the system. The system's default initial value is 0.65, and its range is from 0.6 to 0.7. The high-confidence and low-confidence thresholds are dynamically adjusted using a linear downward adjustment method. When the mean confidence level is compared with the mean threshold, if it is lower than the mean threshold, the overall recognition reliability of the system is deemed insufficient, and dynamic threshold adjustment is initiated. The adjustment formula is as follows: = - , = - ; in For high confidence threshold, The low confidence threshold This serves as the initial high-confidence threshold. The initial low confidence threshold is ΔH, the reduction magnitude of the high threshold is ΔH, and the reduction magnitude of the low threshold is ΔL. The formula for calculating the reduction magnitude is: = = ( ), The mean threshold, The confidence level is the mean. The adjustment coefficient is set to 0.5 to ensure that the threshold reduction is proportional to the mean deviation and to avoid over-adjustment.

[0046] For example, an elderly person with mild cognitive impairment and weak expressive ability may have a lower initial confidence threshold in the system. =0.8, low confidence threshold =0.5, mean threshold =0.65. In the past 30 days, the system recorded 80 valid interaction data points from the elderly. Due to the vagueness of the elderly's expressions, the confidence scores were generally low, resulting in a total confidence score of ΣConf=49.6, from which the mean confidence score was obtained. Since 0.62 < 0.65, the system startup threshold is adjusted. =0.785, =0.485. When the elderly person is detected saying "Turn it on" in the living room, it is determined to be valid speech. The interactive agent is activated and performs confidence calculation. The confidence value is 0.79. If the judgment is based on the threshold before adjustment, an interactive query will be executed. However, if the judgment is based on the threshold after adjustment, a direct match and response can be executed. This avoids frequent queries caused by the original threshold being too high and improves the interaction efficiency.

[0047] The system also includes a model optimization module, which is designed to continuously train the interactive agent by strengthening the training sample set, and continuously optimize the confidence calculation accuracy of the interactive agent. This allows the system to gradually adapt to the personalized needs of the elderly and the care scenarios of elderly care institutions as it is used, achieving an adaptive effect that becomes more and more accurate with use. The model optimization module records historical interaction data. Each set of historical interaction data includes location information, motion state, voice information, and corresponding actual execution instructions. It also selects samples that meet preset conditions from the historical interaction dataset as reinforcement training samples. The reinforcement training samples are key data that can expose the current shortcomings of the system, such as misjudgment, missed judgment, or unreasonable thresholds. This type of data has direct guiding significance for correcting the algorithm parameters, weight coefficients, and threshold settings of the interactive agent. The selected reinforcement training samples are then combined into a reinforcement training sample set to train the interactive agent. The preset conditions for screening the above samples specifically include: First, count the number of consecutive interactions of the same elderly person in the same scene. Continuous interaction refers to the interaction events of the same elderly person in the same scene. If the time interval is less than the interval threshold and the elderly person has not left the scene, it is counted as continuous interaction. Then, determine whether the number of consecutive interactions exceeds the number threshold. The number threshold can be set to 2 times. When the number of interactions exceeds 2 times and the actual execution command of the later interaction is inconsistent with the initial matching execution command of the previous interaction, that is, the later interaction corrects the result of the previous interaction, all the continuous interaction data in this group are included in the reinforcement training sample. For example, after an elderly person sits down on the sofa in the living room, they softly say "Turn it on." The voice acquisition module captures this voice information and converts it into semantic text "Turn it on." Since the living room is a high-frequency interaction scenario, it is determined to be valid voice, triggering the interactive agent to start without wake-up. At the same time, the confidence level is calculated, and the Conf is 0.772. Since the system presets a high confidence threshold H=0.8 and a low confidence threshold L=0.5, the confidence level is determined to be between 0.5 and 0.8. The interactive query command is executed, and the wristband voice asks if you want to turn on the TV. The elderly person immediately responds, "No, I want to turn on the air conditioner." After determining that the elderly person has not left the living room scenario and the interaction time interval has not exceeded the continuous interaction time interval threshold, it is determined to be a continuous interaction. The number of continuous interactions reaches 2, and the actual execution command of the second interaction is to turn on the air conditioner, which is inconsistent with the initial matching execution command of the previous interaction, which asks to turn on the TV. This belongs to the case where the second interaction corrects the result of the previous interaction. This set of historical interaction data is included in the reinforcement training sample set. Secondly, after the interactive agent is triggered and confidence is calculated, if the confidence is higher than the high confidence threshold, then after the response instruction is executed, if the elderly person's feedback is negative or the caregiver verifies that the executed instruction is inconsistent with the actual needs, it is judged as data with incorrect judgment and included in the reinforcement training sample set. For example, when an elderly person says "turn it on" in the bedroom, the interactive agent is triggered after the speech is deemed valid. The confidence score is calculated to be Conf=0.87, which is higher than the high confidence threshold of 0.8. The interactive agent matches the execution response as turning on the bedside lamp in the bedroom. However, after hearing the lamp being turned on, the elderly person responds, "I don't want to turn on the bedside lamp, I want to turn on the radio," clearly denying the execution response. This is judged as an error, and this historical interaction data is included in the reinforcement training sample set. Third, after the interactive agent is triggered and confidence is calculated, if the confidence is lower than the low confidence threshold, a safety response is executed. If the elderly person's feedback indicates that there is no risk or the caregiver verifies that there is no safety hazard after the safety response is completed, it is also considered a judgment error and is included in the reinforcement training sample set. For example, when an elderly person enters the bathroom, the scenario and risk level are determined to be a medium-risk scenario, which also meets the triggering conditions, causing the interactive agent to be activated. At this time, the elderly person mumbles a vague syllable. The confidence level is calculated to be 0.335, which is lower than the low confidence threshold. The safety response command is executed. After receiving the warning, the caregiver rushes to the bathroom and finds that the elderly person is washing up normally and is just humming casually without any safety hazards. Alternatively, if the elderly person reports that everything is normal through the wristband button, it is considered an incorrect judgment. In this case, the historical interaction data is included in the reinforcement training sample set. Fourth, first calculate the number of historical interactions for each scenario, then filter out low-frequency interaction scenarios where the number of historical interactions is lower than the interaction threshold, then extract the data where the initial matching execution command in the scenario is a safety response, and finally filter out the records where the elderly report that they have an emergency need or the caregiver verifies that there is a safety risk, i.e., the data that is judged to be at risk, and include them in the enhanced training sample set. For example, when an elderly person enters the equipment room of a nursing home, historical data shows that the number of interactions the elderly person has had in the current scenario is below the interaction threshold, classifying it as a low-frequency interaction scenario. Furthermore, if the scenario risk mapping table indicates that the elderly person is currently in a high-risk scenario, the interactive agent is activated. If the elderly person, while searching for old items stored away years ago, accidentally knocks over a toolbox in the corner, making a clattering sound, the confidence level is calculated to be 0.3, below the low confidence threshold. Therefore, a safety response command is matched, and the elderly person's current information is immediately sent to the caregiver's terminal. The caregiver arrives and finds that the elderly person has lost their balance due to the knocked-over toolbox, has a slight sprained right leg, and is unable to move while holding onto the wall. This indicates a low-frequency interaction scenario with a confirmed safety risk. This interaction data is included in the reinforcement training sample set for subsequent optimization of the interactive agent's confidence calculation logic and safety response sensitivity in low-frequency, high-risk scenarios, reducing the probability of missed risk detection.

[0048] In one embodiment of the present invention, the interactive intelligent agent includes a general interactive intelligent agent for institutions and a personal interactive intelligent agent for the elderly. The general interactive intelligent agent for institutions refers to a public intelligent agent module for all elderly people in elderly care institutions. It is equipped with basic interactive capabilities in elderly care scenarios through a pre-trained model, and provides unified data support and functional foundation for all personal interactive intelligent agents for the elderly. The personal interactive intelligent agent for the elderly refers to a personalized intelligent agent module built for each elderly person. It is based on the pre-trained model of the general interactive intelligent agent for institutions and learns the elderly person's personal data to form exclusive interactive logic. First, the institutional general interactive intelligent agent is pre-trained using a pre-trained model, which can employ a Transformer deep learning model. The data used for model training includes: speech recognition data, which includes Mandarin and common dialects such as Cantonese, Sichuanese, and Henanese, with each sample containing audio and corresponding standard text; a knowledge base for elderly care, containing entries on dietary care, daily living care, rehabilitation training, and psychological counseling; a medical and health knowledge base, containing entries on common disease prevention, symptom recognition, emergency treatment, and medication contraindications; a map layout of the nursing home, specifically including the spatial coordinates of each floor, room identification and functions, and equipment identification and deployment coordinates of fixed-location interactive units; personnel information, specifically including the number of caregivers and doctors, their job responsibilities, and contact information; and schedule data, which includes information such as daily meal times, group rehabilitation activity times, and weekly physical examination schedules. The above data is divided into training and validation sets according to the proportion, and then input into the pre-trained model for training to obtain a mature pre-trained model. When an elderly person uses the system for the first time, the system automatically creates a personalized interactive intelligent agent for them, using the pre-trained model parameters of the institution's general interactive intelligent agent as initial parameters. During subsequent use, the elderly person's personalized interactive intelligent agent collects three types of personalized data in real time: voice information, including all of the elderly person's voice input and corresponding semantic text; movement data, including the elderly person's location trajectory, scene dwell time, and changes in movement state; and interactive feedback data, including the elderly person's approval, disapproval, and supplementary explanations of the system's executed instructions. The collected personalized data is used for personalized training to learn the elderly person's voice habits, behavioral preferences, and health restrictions. For example, if an elderly person is used to taking a walk in the garden at 3 pm every day, they should avoid sweets if they have diabetes, or they should avoid pollen environments if they have asthma. The organization's general interactive intelligent agent and the elderly's personal interactive intelligent agent work together in real time. When the elderly's personal interactive intelligent agent encounters a need that it cannot handle on its own, such as querying information about other elderly people, accessing the organization's group schedule, or obtaining professional medical knowledge, it can call the data and functions of the organization's general interactive intelligent agent. For example, if an elderly person asks "What is tomorrow's group activity?", the elderly person's personal interactive intelligent agent can call the schedule data of the organization's general interactive intelligent agent to generate the corresponding answer instruction. The organization's general interactive agent regularly summarizes the usage data of all elderly people's personal interactive agents. For example, it conducts a common needs analysis once a week, and filters out common needs that occur more than 5 times during the period. For example, multiple elderly people misinterpreted the intention of "adding water" in a restaurant setting, or multiple elderly people responded to a certain type of safety response as risk-free. The interaction data corresponding to these common needs are used as supplementary training data to iteratively optimize the pre-trained model. The optimized pre-trained model parameters are synchronously updated to all elderly people's personal interactive agents, achieving a collaborative improvement effect where one person's optimization benefits everyone.

[0049] For example, if an elderly person says "I want something sweet" in a restaurant, the elderly person's personal interactive intelligent agent can access the medical and health knowledge base of the institution's general interactive intelligent agent and the health taboo data stored in their own account to generate an interactive question: "You have diabetes and should not eat sweets. Do you need me to recommend some low-sugar snacks for you?" In a nursing home, the personal interactive agent for elderly people misjudged eight elderly people who expressed "add water" as "add water" in the restaurant setting. After these common needs were summarized, the institution's general interactive agent used "add water" as a related keyword for the intention of "add water" and updated the pre-trained model. After all the personal interactive agents for elderly people synchronized this parameter, they were able to accurately identify the intention of "add water".

[0050] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A smart interactive system based on movement path triggering, characterized in that, include: The positioning module is used to collect location information and motion status in real time; The voice acquisition module is used to acquire voice information in real time; An interactive intelligent agent is configured with triggering conditions. When the triggering conditions are met, the interactive intelligent agent is triggered and a confidence score is calculated based on the location information, motion state, and voice information. Based on the confidence score, an execution instruction in a preset instruction library is matched. The execution instruction includes an execution response, an interactive query, and a security response. The model optimization module records the location information, motion state, voice information, and corresponding actual execution instructions as a set of historical interaction data. It selects samples that meet preset conditions from the historical interaction dataset as reinforcement training samples and generates a reinforcement training sample set to train the interactive agent.

2. The intelligent interactive system based on movement-line triggering according to claim 1, characterized in that, The interactive intelligent agent includes a confidence calculation submodule, specifically comprising: Multiple scoring items are obtained, including five preset basic scores across five dimensions: speech recognition clarity, intent recognition accuracy, scene adaptability, habit fit, and interactive feedback effectiveness. Specifically, speech recognition clarity reflects the clarity of the speech information converted into text; intent recognition accuracy reflects the degree of matching between the recognized intent and the actual need; scene adaptability reflects the degree of matching between the recognized intent and the current scene; habit fit reflects the degree of matching between the recognized intent and the elderly person's personal habits; and interactive feedback effectiveness reflects the consistency between the elderly person's feedback and the recognized intent. Each dimension is assigned a dynamic weight coefficient, and the confidence level is calculated based on the baseline scores of the five dimensions and the dynamic weight coefficient of each dimension.

3. The intelligent interactive system based on movement-line triggering according to claim 2, characterized in that, The dynamic weight coefficients are dynamically optimized based on the reinforcement training sample set, specifically as follows: For each set of reinforced training samples in the reinforced training sample set, the prediction confidence is calculated based on the initial weight coefficient. The prediction confidence is matched with the corresponding execution instruction and compared with the actual execution instruction in the reinforced training sample to see if they are consistent. If they are inconsistent, they are judged as errors. The initial weight coefficient is the dynamic weight coefficient obtained in the previous training. The corrected dynamic weight coefficients are obtained by calculating the error based on the weight correction model.

4. The intelligent interactive system based on movement-line triggering according to claim 2, characterized in that, The interactive agent is configured with a confidence matching instruction strategy, specifically: There are preset high confidence thresholds and low confidence thresholds, and the high confidence threshold is greater than the low confidence threshold. The confidence level is compared with both the high confidence threshold and the low confidence threshold. If the confidence level is higher than the high confidence threshold, the response is executed; if the confidence level is between the high confidence threshold and the low confidence threshold, the interactive query is executed. If the confidence level is below the low confidence threshold, a safety response is executed.

5. The intelligent interactive system based on movement-line triggering according to claim 3, characterized in that, The confidence matching instruction strategy includes a sub-strategy for dynamically adjusting the confidence threshold, specifically: The scenario is matched based on the location information, and the scenario risk level is obtained by matching the scenario with the scenario risk mapping table stored in the database; and the elderly’s vital signs information is obtained based on the external vital signs monitoring module of the system, and the elderly’s vital signs information is compared with the preset vital signs information to determine whether the vital signs information is abnormal. If the risk level of the scenario changes or the vital signs information becomes abnormal, the low confidence threshold will be dynamically adjusted. The mean confidence score is calculated based on the confidence scores in the historical interaction dataset. The mean confidence score is compared with a threshold. If it is lower than the threshold, the high confidence score threshold and the low confidence score threshold are dynamically adjusted.

6. The intelligent interactive system based on movement line triggering according to claim 5, characterized in that, The triggering conditions include: The scenario is a risk scenario with a risk level higher than the risk level threshold. By comparing the location information with the pre-stored location areas of elderly care institutions, it is determined that the location exceeds the location area of ​​elderly care institutions; The voice information is determined to be valid voice using a voice determination sub-strategy. The elderly person's vital signs information is compared with the preset threshold for elderly person's vital signs to determine if there are any abnormalities.

7. The intelligent interactive system based on movement-line triggering according to claim 6, characterized in that, The specific speech determination sub-strategy is as follows: Semantic text is obtained from the voice information, and the semantic text is matched with a preset keyword library. If the semantic text contains keywords, the voice information is valid. If the semantic text does not contain keywords, the scene is obtained through the location information, and the historical interaction count of the scene is matched from the historical interaction database. If it is a high-frequency interaction scene with a historical interaction count higher than the interaction threshold, then the voice information is valid voice.

8. The intelligent interactive system based on movement-line triggering according to claim 1, characterized in that, The interactive intelligent agents include general interactive intelligent agents for institutions and personal interactive intelligent agents for the elderly; The general interactive intelligent agent of the institution is pre-trained through a pre-trained model. The data used by the pre-trained model includes speech recognition, elderly care knowledge base, medical and health common knowledge base, nursing home map layout, personnel information, and schedule data, in order to provide basic capability support for the elderly personal interactive intelligent agent. The elderly person's personal interactive intelligent agent is based on the pre-trained model of the institution's general interactive intelligent agent. It is trained in a personalized way by collecting the elderly person's voice information, movement data, and interaction feedback data, so as to learn the elderly person's voice habits, behavioral preferences and health taboos. The general interactive intelligent agent of the institution and the personal interactive intelligent agent of the elderly interact in real time. The personal interactive intelligent agent of the elderly calls the data generation instructions of the general interactive intelligent agent of the institution. The general interactive intelligent agent of the institution summarizes the common needs of multiple personal interactive intelligent agents of the elderly and iteratively optimizes the pre-trained model.

9. The intelligent interactive system based on movement-line triggering according to claim 1, characterized in that, The preset conditions used in the model optimization module to select reinforcement training samples include: The same elderly person interacts with the same scenario more than the threshold number of times, and the subsequent interaction corrects the result of the previous interaction. If the confidence level is higher than the high confidence threshold and the execution response is completed, the judgment is determined to be incorrect based on feedback from the elderly person. If the confidence level is below the low confidence threshold and a safety response is completed, the risk is determined to be no based on feedback from the elderly person. Low-frequency interaction scenarios with a historical number of interactions below the interaction threshold, and after completing a safety response, are judged to have a risk based on the feedback results.

10. The intelligent interactive system based on movement line triggering according to claim 1, characterized in that, The positioning module is configured with a location acquisition strategy, specifically: Fixed-position interactive units are evenly installed at room entrances, corridor corners, stairwells, and along railings on each floor of the elderly living facility, and the Bluetooth frequency bands of the fixed-position interactive units on different floors are differentiated. The fixed-position interaction unit interacts with the transceiver unit built into the positioning module, enabling the positioning module to obtain location information based on topological relationships and signal strength.

Images