Intelligent body multi-task cooperative monitoring method and system for old people's family

Abstract

The present application relates to the field of intelligent old-age care and robot collaborative control, and particularly relates to an intelligent agent multi-task collaborative monitoring method and system for old people's families. The method embeds rehabilitation training actions into a nostalgic narrative scene associated with the personal experience of the old people to construct an emotional narrative background task flow, and uses real-time extracted implicit emotional tension to drive a medication reminder task to non-predictively disturb the background task flow in a haptic rhythm, so as to maintain the immersion of rehabilitation training while ensuring the timeliness of the medication reminder. The emotional immersion degree and the exercise fatigue degree of the old people are analyzed, the emotional state is updated accordingly, the rehabilitation narrative compliance coefficient is calculated, the hysteresis recovery characteristic correction is applied to the triggering threshold, and the narrative advancing rate is locally self-adaptively slowed down in weak compliance segments, so as to form a closed-loop collaborative regulation driven by the physical and mental state of the old people, thereby improving the compliance of rehabilitation training and the naturalness of emotional interaction under the premise of ensuring the safety of medication.

Description

Technical Field

[0001] This invention relates to the field of intelligent elderly care and robot collaborative control, and in particular to an intelligent agent multi-task collaborative monitoring method and system for elderly people's homes. Background Technology

[0002] As the population ages, intelligent agent multi-task collaborative monitoring technology for elderly people in their homes has gradually become a research hotspot in the field of smart elderly care. In existing technologies, intelligent robots used for home monitoring of the elderly typically employ a modular approach, with each module independently performing tasks such as medication reminders, rehabilitation training guidance, and emotional support. Medication reminders are often implemented using fixed-time audio-visual alarms, rehabilitation training guidance relies on pre-set action demonstration videos displayed on a screen, and emotional support services are mostly limited to simple voice dialogue. However, these methods have significant shortcomings in practical applications: Firstly, the tasks are independent of each other and lack a dynamic collaborative scheduling mechanism based on the elderly person's real-time physical and mental state. This can lead to the robot forcibly interrupting the elderly person's ongoing and well-compliant rehabilitation training when performing medication reminders, causing resistance and reducing training compliance. Secondly, the rehabilitation training guidance methods are too explicit and instructive, failing to establish an implicit connection with the elderly person's emotional state and personal life experiences, making it difficult to stimulate the elderly person's willingness to participate actively. Furthermore, it cannot simultaneously perceive the elderly person's emotional engagement and physical fatigue level and make adaptive adjustments while performing rehabilitation guidance. Therefore, how to implicitly embed rehabilitation training guidance into narrative scenarios related to the elderly's personal experiences, and how to construct a preemptible multi-task collaborative mechanism driven by the elderly's real-time emotional state, so as to maintain the immersion and compliance of rehabilitation training while ensuring the timeliness of medication reminders, has become an urgent technical problem to be solved in the field of intelligent agent multi-task collaborative monitoring for elderly families. Summary of the Invention

[0003] This invention overcomes the shortcomings of the prior art and provides an intelligent agent multi-task collaborative monitoring method and system for elderly families.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention discloses a multi-task collaborative monitoring method for intelligent agents in elderly households, comprising the following steps: Step 1: Collect the elderly’s home spatiotemporal behavior data through the multimodal sensor network of the robot agent, and extract the elderly’s life rhythm fingerprint and implicit emotional tension from it. Step 2: Based on the aforementioned life rhythm fingerprint and latent emotional tension, construct an emotional narrative background task flow that implicitly embeds rehabilitation training actions into nostalgic narrative scenes, and set medication reminders as preemptible prominent tasks triggered by the intensity of the aforementioned latent emotional tension. Step 3: Control the robot agent to execute the emotional narrative background task flow. When the preemptible salient task is not triggered, generate dynamic nostalgic image projection to guide rehabilitation actions. Once triggered, interrupt the background task with an unpredictable tactile rhythm and execute drug delivery. Step 4: Use the depth sensor and wearable device on the robot to acquire the elderly’s movement posture, facial micro-expression and skin electrical signals, and combine them with the life rhythm fingerprint to analyze the emotional immersion and exercise fatigue, so as to update the implicit emotional tension and calculate the rehabilitation narrative compliance coefficient. Step 5: Adjust the trigger condition threshold of the preemptible salient task according to the updated implicit emotional tension, and adjust the narrative progression rate of the nostalgic image projection in the emotional narrative background task flow according to the rehabilitation narrative compliance coefficient.

[0005] The second aspect of the present invention discloses an intelligent agent multi-task collaborative monitoring system for elderly households. The system includes a memory and a processor. The memory stores a program for an intelligent agent multi-task collaborative monitoring method for elderly households. When the program for the intelligent agent multi-task collaborative monitoring method for elderly households is executed by the processor, the steps of any one of the methods described above are implemented.

[0006] This invention addresses the technical deficiencies in the prior art and offers the following advantages: It implicitly embeds rehabilitation training movements into nostalgic narrative scenarios linked to the elderly person's personal experiences. Real-time emotional states drive asymmetric scheduling between medication reminders and rehabilitation guidance, allowing rehabilitation training to be integrated into the narrative immersion experience in a non-explicit manner. This avoids the emotional resistance and training interruptions caused by abrupt task switching in conventional approaches. Simultaneously, through joint analysis of multimodal signals such as movement following, facial emotional responses, postural tremors, and skin conductance, a real-time quantitative assessment of the elderly person's emotional immersion and physical fatigue is formed. Based on this, the trigger sensitivity of task scheduling and the pace of narrative progression are dynamically adjusted, achieving an adaptive balance between the timeliness of medication reminders and the continuity of rehabilitation immersion. This constructs a closed-loop collaborative mechanism driven by the elderly person's physical and mental state, improving adherence to rehabilitation training and the naturalness of emotional interaction while ensuring medication safety. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0008] Figure 1Flowchart of a multi-task collaborative monitoring method for intelligent agents in elderly households; Figure 2 This is an architecture diagram of an intelligent multi-task collaborative monitoring system for elderly households. Detailed Implementation

[0009] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0010] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0011] like Figure 1 As shown, the first aspect of this invention discloses a multi-task collaborative monitoring method for intelligent agents in elderly households, comprising the following steps: Step 1: Collect the elderly’s home spatiotemporal behavior data through the multimodal sensor network of the robot agent, and extract the elderly’s life rhythm fingerprint and implicit emotional tension from it. Step 2: Based on the aforementioned life rhythm fingerprint and latent emotional tension, construct an emotional narrative background task flow that implicitly embeds rehabilitation training actions into nostalgic narrative scenes, and set medication reminders as preemptible prominent tasks triggered by the intensity of the aforementioned latent emotional tension. Step 3: Control the robot agent to execute the emotional narrative background task flow. When the preemptible salient task is not triggered, generate dynamic nostalgic image projection to guide rehabilitation actions. Once triggered, interrupt the background task with an unpredictable tactile rhythm and execute drug delivery. Step 4: Use the depth sensor and wearable device on the robot to acquire the elderly’s movement posture, facial micro-expression and skin electrical signals, and combine them with the life rhythm fingerprint to analyze the emotional immersion and exercise fatigue, so as to update the implicit emotional tension and calculate the rehabilitation narrative compliance coefficient. Step 5: Adjust the trigger condition threshold of the preemptible salient task according to the updated implicit emotional tension, and adjust the narrative progression rate of the nostalgic image projection in the emotional narrative background task flow according to the rehabilitation narrative compliance coefficient.

[0012] Step one specifically involves: The robot intelligent agent moves in the home environment according to the preset daily inspection trajectory, and simultaneously activates the millimeter-wave radar, passive infrared pyroelectric sensor and environmental microphone array in the multimodal sensor network to continuously collect the spatiotemporal location sequence of the elderly in different rooms, changes in body activity heat radiation and environmental acoustic events, and generate raw spatiotemporal behavioral stream data. The original spatiotemporal behavioral stream data is segmented in the time domain to extract the behavioral patterns of the elderly in each fixed time window over multiple consecutive days. Spatiotemporal transition entropy is calculated for the behavioral patterns, and the spatiotemporal transition entropy is used to construct a life rhythm fingerprint that characterizes the degree of deviation of the elderly’s individualized daily activity rhythm. Based on the life rhythm fingerprint, abnormal behavior segments that deviate from the normal rhythm in the daily behavior sequence are identified. The environmental acoustic events at the time when the abnormal behavior segments occurred are traced back to separate the fundamental frequency micro-perturbation features and pause prosodic abnormal features of the speech signal, thereby constructing an acoustic marker of unexpressed emotions. The emotional acoustic markers are time-aligned with the changes in somatic activity thermal radiation captured during the abnormal behavior segment. The non-periodic fluctuation component of the thermal radiation change is extracted as a physiological stress thermal pattern. The physiological stress thermal pattern is then coupled with the emotional acoustic markers through cross-modal feature coupling to generate latent emotional tension.

[0013] Specifically, the robotic agent moves autonomously within the home environment according to a pre-set daily inspection trajectory. This trajectory can be generated in advance through human instruction or environmental mapping, covering key areas of the elderly person's daily activities, such as the bedroom, living room, kitchen, and bathroom. During the inspection, the robotic agent simultaneously activates its onboard multimodal sensor network, which includes at least three types of sensing devices: millimeter-wave radar, passive infrared pyroelectric sensors, and an environmental microphone array. The millimeter-wave radar acquires the elderly person's spatiotemporal location sequence in different rooms in real time. The passive infrared pyroelectric sensors capture changes in thermal radiation from the elderly person's body movements. Infrared radiation emitted from the elderly person's body surface is received by the sensor's sensitive elements and converted into electrical signals. When the elderly person walks, sits up, or moves their limbs, the amplitude and fluctuation pattern of the electrical signal change accordingly. The environmental microphone array simultaneously picks up various environmental acoustic events in the home environment, including the elderly person's voice, footsteps, and non-voice sounds such as furniture sounds.

[0014] After obtaining the raw spatiotemporal behavioral flow data, temporal domain segmentation processing is performed. Specifically, using fixed time windows as the basic unit, for example, dividing a day into several equal-length time periods, the distribution of the elderly person's stay time in each room, the frequency of activity heat radiation fluctuations, and the type and density of environmental acoustic events within each time window are statistically analyzed to form behavioral pattern vectors corresponding to each time window. Within a continuous multi-day monitoring period, the spatiotemporal transfer entropy of the behavioral pattern vectors for the same time window is calculated. Spatiotemporal transfer entropy is used to measure the consistency and predictability of the elderly person's activity patterns within the same time period. The lower the entropy value, the more stable the behavioral pattern during that period; an increase in entropy value reflects a trend of deviation from the normal activity rhythm during that period. Arranging the spatiotemporal transfer entropies of each time window in chronological order constitutes the life rhythm fingerprint, which characterizes the degree of deviation of the elderly person's individual daily activity rhythm. The life rhythm fingerprint is essentially a set of rhythm deviation curves that change with the passage of each day, reflecting the extent to which the elderly person's behavioral patterns deviate from their personal routines in different time periods.

[0015] Based on the generated circadian rhythm fingerprint, abnormal behavioral segments deviating from the normal rhythm in the daily behavioral sequence are further identified. The deviation from the normal rhythm refers to a difference between the actual behavioral pattern within a certain time window and the historical normal pattern for that period in the circadian rhythm fingerprint exceeding a preset deviation judgment standard. For example, an elderly person exhibiting continuous walking or frequent entry and exit from multiple rooms within their usual afternoon nap time window, or an unusually long period of stillness within their usual active time window, are typical abnormal behavioral segments. Once such abnormal behavioral segments are identified, the environmental acoustic event records corresponding to the time of occurrence of the abnormal behavioral segment are traced back. From the traced environmental acoustic events, the focus is on separating the speech signal and performing a refined analysis of its acoustic features. This invention employs the extraction of features from the speech signal that reflect the speaker's emotional state but are not easily expressed subjectively, namely, fundamental frequency micro-perturbation features and abnormal pause prosody features. Fundamental frequency micro-perturbation features refer to the involuntary, minute frequency jitters superimposed on the fundamental frequency curve of an elderly person's speech. The amplitude and frequency distribution of these jitters exhibit regular changes distinct from their calm state under emotional tension or depression, and these changes are usually imperceptible to the speaker. Pause prosodic abnormalities refer to the degree to which the distribution pattern of pause durations between sentences or phrases deviates from the speaker's usual speaking habits. Examples include abnormal elision pauses or excessively long pauses in non-grammatical positions within otherwise coherent expressions. Through quantitative analysis of these two types of features, emotional acoustic markers reflecting the elderly person's unexpressed emotional state are constructed. These emotional acoustic markers can be presented as a set of multidimensional feature vectors, with each dimension corresponding to the statistical parameters of the fundamental frequency micro-perturbation and the deviation measure of pause prosodic abnormalities, respectively.

[0016] Then, the aforementioned emotional acoustic markers were precisely aligned along the time axis with the changes in somatic activity thermal radiation captured by a passive infrared pyroelectric sensor during the same abnormal behavioral segment. After alignment, the aperiodic fluctuation component was extracted from the thermal radiation change signal. The aperiodic fluctuation component refers to the irregular fluctuation portion remaining after removing the periodic components generated by the elderly person's regular periodic movements (such as constant-speed walking and regular limb swinging) from the original thermal radiation change signal. This aperiodic fluctuation usually manifests as increased amplitude and irregular frequency when the elderly person is in a state of emotional stress, and therefore it is extracted and defined as physiological stress thermal patterns. Physiological stress thermal patterns essentially reflect the micro-change patterns of body surface temperature caused by the activity of the autonomic nervous system under the influence of emotional fluctuations. Finally, the physiological stress thermal patterns and emotional acoustic markers were coupled across modal features. Cross-modal feature coupling involves associating and combining emotional acoustic markers from the acoustic modality with physiological stress thermal patterns from the thermal radiation modality within a unified feature space. This can be achieved through feature splicing or feature fusion to generate a comprehensive multimodal joint feature that characterizes the latent emotional tension in the elderly. The magnitude of latent emotional tension reflects the degree of internal emotional stress that the elderly person is not currently expressing verbally but is revealed through both acoustic and physiological signals. Higher values ​​indicate greater emotional stress and a greater likelihood of potential negative emotional accumulation.

[0017] It should be noted that the aforementioned circadian rhythm fingerprint is not a fixed static benchmark, but rather continuously updated as the number of monitoring days increases. Whenever new daily behavioral data is entered, the system adds the new data to the historical database and recalculates the spatiotemporal transition entropy for each time window, gradually converging the circadian rhythm fingerprint to a stable form that most accurately reflects the individualized rhythmic characteristics of the elderly individual. This dynamic update mechanism ensures that the judgment of rhythm deviation is always based on the elderly individual's latest lifestyle habits, avoiding the problem of rhythm benchmark drift caused by factors such as seasonal changes and changes in health status.

[0018] Step two specifically involves: The life rhythm fingerprint is retrieved, and the time phase label of the elderly person's physiological rhythm on that day is matched from the preset personal historical behavioral rhythm benchmark library. At the same time, the implicit emotional tension is obtained, and the emotional-rhythm coupling deviation under the time phase label is calculated. It should be noted that the aforementioned personal historical behavioral rhythm benchmark database is a historical behavioral rhythm database established for individual elderly individuals during the long-term operation of the system. Its construction method involves: aligning the daily rhythm fingerprints of multiple consecutive days according to time, then accumulating and statistically analyzing them daily to extract the periodic changes in the elderly person's daily behavioral rhythm over time. The day is divided into several time phase intervals with different behavioral characteristics, such as the morning awakening and active period, the morning stable activity period, the midday rest period, the afternoon activity recovery period, and the nighttime rest period. Each time phase interval is assigned a time phase label. The specific method for matching the daily time phase label is to query the historical rhythm feature pattern of the corresponding time period from the benchmark database based on the current time position within the day, thereby determining which rhythm phase the elderly person is currently in.

[0019] While obtaining the time phase label, the implicit emotional tension output from step one is also acquired, and the emotion-rhythm coupling deviation under that time phase label is calculated. The calculation method for the emotion-rhythm coupling deviation is further explained: the emotion-rhythm coupling deviation is used to comprehensively measure the degree of mismatch between the elderly person's current emotional state and their physiological rhythm. Specifically, the rhythm deviation component corresponding to the time phase label and the current value of the implicit emotional tension within that time period are extracted from the circadian rhythm fingerprint, and the two are jointly evaluated. For example, if the current time phase label indicates that the elderly person is in their midday rest period, and the circadian rhythm fingerprint shows that the activity rhythm deviation during this period is significantly higher than the historical baseline, while the implicit emotional tension is also at a high level, it indicates that the elderly person has not only deviated from the usual midday rest pattern in behavior but also exhibits a relatively tense emotional state. In this case, the calculated emotion-rhythm coupling deviation is correspondingly high. Conversely, if the circadian rhythm fingerprint shows that although the activity rhythm deviation is increased, the implicit emotional tension remains at a normal level, the coupling deviation is relatively low, because this deviation may stem from some neutral external factors rather than emotional distress. Using the above methods, the deviation of the emotion-rhythm coupling can accurately reflect the overall degree of abnormality in the physical and mental state of the elderly.

[0020] Based on the aforementioned emotion-rhythm coupling deviation, narrative scene fragments associated with the elderly person's personal experiences are retrieved from the nostalgic narrative material library. The movement complexity level of the rehabilitation training movements is determined according to the degree of deviation of the activity rhythm in the life rhythm fingerprint. The rehabilitation training movements are then mapped to the character behavior sequence in the narrative scene fragments according to the movement complexity level, forming a temporal action-narrative binding mapping table. It should be noted that, based on the calculated emotion-rhythm coupling deviation, narrative scene fragments related to the elderly person's personal experiences are retrieved from the nostalgic narrative material library. The nostalgic narrative material library is a pre-built database of narrative scenes closely linked to the individual life experiences of the elderly person. The construction of this material library is typically completed with the assistance of family members or caregivers during the initial system deployment phase: information on the elderly person's important life experiences is collected through interviews or questionnaires, such as the historical context of their childhood, work scenarios during their youth, typical moments of family life, and past leisure activities. These experiences are then transformed into visualized narrative scene fragments and digitally stored. Each narrative scene fragment is labeled with emotional attribute tags, such as warmth, nostalgia, joy, and tranquility. During retrieval, the system prioritizes scene fragments from the material library whose emotional attributes match the current emotional regulation needs, based on the level of emotion-rhythm coupling deviation. For example, when the coupling deviation is high, scenes with a preference for warmth and tranquility are prioritized to facilitate emotional calming.

[0021] Simultaneously, the complexity level of rehabilitation training movements is determined based on the degree of deviation from the activity rhythm in the circadian rhythm fingerprint. A greater deviation indicates a more unstable daily behavioral pattern in the elderly person; in this case, rehabilitation training movements with lower complexity should be selected to avoid placing additional physical burden or causing frustration. After determining the complexity level, the system maps the corresponding rehabilitation training movements to the character behavior sequences within the selected narrative scene fragments. For example, if the retrieved narrative scene is a segment of an elderly person watering flowers in their youth, the character behavior sequence can be designed as movement nodes such as lifting the watering can, tilting the can, and moving along the flowerbed. For upper limb rehabilitation training needs, training movements such as grasping, raising the arm, and extending forward can be embedded according to their complexity level into the character behaviors of lifting the watering can, tilting the can, and moving along the flowerbed, respectively. After mapping, the rehabilitation training movements and the narrative character's behavior are bound on a timeline, forming a temporal action-narrative binding mapping table. The mapping table uses timestamps as indexes, recording the behavior that the character should exhibit in the narrative scene at each moment, as well as the rehabilitation training semantics carried by that behavior.

[0022] The temporal action-narrative binding mapping table is parsed, and key action nodes with rehabilitation guidance semantics are extracted from the character behavior sequence. Each key action node is arranged into a key action timeline according to its bound narrative timestamp, and a narrative advancement timeline with dynamic nostalgic image projection synchronized with the key action timeline is generated. The key action timeline and the narrative advancement timeline are aligned at the frame level, and a relaxed transition interval with a duration determined by the rhythm deviation of the life rhythm fingerprint is embedded between adjacent key action nodes. The key action timeline after embedding the relaxed transition interval is synthesized with the narrative advancement timeline to generate an emotional narrative background task flow. It should be noted that after obtaining the temporal action-narrative binding mapping table, it is parsed to extract key action nodes with rehabilitation guidance semantics. Key action nodes refer to those actions that correspond to a clear rehabilitation training intention and require the elderly to follow along, which are different from auxiliary actions in the narrative scene that only serve as transitions or atmosphere rendering. The extracted key action nodes are sorted according to their bound narrative timestamps to obtain the key action timeline. Corresponding to the key action timeline, the system generates a narrative progression timeline for a dynamic nostalgic video projection synchronized with the key action timeline. The narrative progression timeline defines the nostalgic video content that the projection device should output at each moment and its progression rhythm.

[0023] After generating the two timelines mentioned above, the key action timeline is aligned frame-by-frame with the narrative progression timeline to ensure that the presentation time of each key action node is precisely synchronized with the frame in the projected image where the character performs the corresponding action. After alignment, a relaxation transition interval is embedded between adjacent key action nodes. This relaxation transition interval is a relatively loose interval inserted between two key nodes where the elderly person needs to follow the rehabilitation actions. During this interval, the narrative progression timeline continues to play nostalgic video content to maintain scene immersion, but no new rehabilitation guidance actions are arranged in the key action timeline. The elderly person can freely adjust their posture or rest briefly during this interval without having to continuously follow the action guidance. The duration of the relaxation transition interval is determined by the degree of rhythm deviation in the circadian rhythm fingerprint: when the rhythm deviation is large, it indicates that the elderly person's physical and mental state is relatively unstable that day, and a longer relaxation transition interval should be allocated to give the elderly person more time for physical and emotional recovery; when the rhythm deviation is small, the relaxation transition interval can be shortened accordingly to improve the continuity and efficiency of rehabilitation training. By combining the key action timeline embedded in the relaxed transition interval with the narrative progression timeline, a complete emotional narrative background task flow is generated. This emotional narrative background task flow organically coordinates the visual guidance of nostalgic narrative, the implicit embedding of rehabilitation training actions, and the relaxed rhythm adapted by the rhythmic state.

[0024] Using the implicit emotional tension as input, the implicit emotional tension is extrapolated to obtain the predicted peak time of emotional tension within a preset time window. The predicted peak time of emotional tension is marked as the pre-trigger time node for drug delivery, and a preemptible prominent task instance that shares a common time axis with the emotional narrative background task flow is generated. The preemptible salient task instance is assigned an asymmetric task weight that takes priority over the emotional narrative background task flow. When the intensity of the latent emotional tension collected in real time exceeds the dynamic trigger threshold adjusted by the degree of rhythm deviation of the life rhythm fingerprint, the preemptible salient task instance immediately preempts the execution resources of the robot agent.

[0025] It should be noted that a preemptible highlight task instance also needs to be generated to allow the medication reminder function to interrupt the aforementioned background task flow when necessary. Specifically, the implicit emotional tension output from step one is used as input for trend extrapolation. The specific implementation of trend extrapolation is further explained as follows: The system maintains a sliding time window, continuously recording the numerical sequence of implicit emotional tension over a recent period. By analyzing the direction and rate of change of this sequence, the evolution trend of emotional tension in the short term is fitted. For example, if the recent sampling points show that implicit emotional tension is continuously increasing at an increasing rate, trend extrapolation will predict that this upward trend will continue in the next preset time window and may reach a local peak. The predicted peak position on the timeline is marked as the pre-trigger time node for medication delivery. Based on this, a preemptible highlight task instance sharing a common timeline with the aforementioned emotional narrative background task flow is generated. This instance, centered on the pre-trigger time node, contains all the execution instructions required for medication delivery, such as navigating to the medication storage location, grabbing the pillbox, moving to the elderly person's side, and performing the delivery action. Regarding task weight allocation, the system assigns asymmetric task weights to preemptible salient task instances, prioritizing them over the emotional narrative background task flow. This means that once the triggering condition is met, the preemptible salient task instance has the right to immediately interrupt the execution of the background task flow and occupy all execution resources of the robot agent. The triggering condition is set as follows: when the intensity of implicit emotional tension collected and continuously updated in step one exceeds the dynamic triggering threshold, the preemptible salient task instance immediately preempts execution. It should be noted that the dynamic triggering threshold is not a fixed value, but is adjusted by the degree of rhythm deviation of the circadian rhythm fingerprint. When the rhythm deviation is large, it indicates that the elderly person's physical and mental state is relatively fragile that day. In this case, the triggering threshold should be appropriately lowered, and medication reminders can be triggered at a lower level of emotional tension to intervene as early as possible and prevent further deterioration of the mood. Conversely, the triggering threshold should be appropriately increased to reduce unnecessary interruptions to rehabilitation training.

[0026] Step three specifically involves: The emotional narrative background task flow and the preemptible prominent task instance are obtained and loaded into the dual-channel task executor of the robot agent, and a parallel execution instruction stack is established with the emotional narrative background task flow as the main channel and the preemptible prominent task instance as the preemption channel. When no trigger signal of the preemptible salient task instance is detected, the robot agent is controlled to drive the projection module to project dynamic nostalgic images according to the emotional narrative background task flow, and simulates the guiding actions of the characters in the narrative scene segment through the torso posture of the robot agent in the relaxed transition interval, forming an emotional atmosphere field to guide the rehabilitation action. During the maintenance of the emotional atmosphere field, the intensity of implicit emotional tension is continuously monitored. When the intensity of implicit emotional tension exceeds the dynamic trigger threshold, the projection screen and torso posture simulation state of the current frame of the emotional narrative background task flow are immediately frozen, and a background task breakpoint scene containing the current narrative progress pointer and the completion degree of action guidance is generated. Activate the preemptible salient task instance, randomly select a tactile perturbation rhythm pattern from the pre-built non-repetitive tactile rhythm pattern library, control the robot agent to drive the contact actuator with the tactile perturbation rhythm pattern to generate an unpredictable tactile stimulation sequence on the elderly person's body surface, interrupt the elderly person's current immersion state and guide their attention to drug delivery. The non-repetitive tactile rhythm pattern library is a pre-built dataset containing various distinct tactile stimulus timing templates. Its construction involves the designers, during the initial system configuration phase, creating several sets of tactile rhythm patterns with different stimulus intervals, durations, and intensity variations, based on psychological research on tactile perception. Each rhythm pattern defines the start and end times and intensity of each contact stimulus in a complete tactile stimulus sequence. It is important to emphasize that the rhythm patterns in this library are designed to be non-repetitive in their temporal patterns, ensuring that elderly individuals cannot predict the next tactile stimulus pattern based on previous experience. This unpredictability aims to prevent the elderly from developing habitual adaptations to fixed tactile cues, thereby maintaining the effective capture of attention for tactile disturbances. During each extraction, the system randomly selects a rhythm pattern from the library that has not yet been used in the current task cycle, further enhancing the unpredictability of the stimulus pattern.

[0027] After completing drug delivery, the narrative progress pointer in the background task breakpoint is read, and the robot agent is controlled to restore the projection of the emotional narrative background task flow and the trunk posture simulation with the narrative progress pointer as the starting point. In the first relaxation transition interval after restoration, an emotional re-immersion lag decay interval with a duration determined by the degree of rhythm deviation of the life rhythm fingerprint is inserted.

[0028] It should be noted that the emotional narrative background task flow and preemptible highlight task instances are loaded into a dual-channel task executor. This executor supports two tasks running in parallel with different priorities. The main channel maintains the continuous operation of the background task flow, while the preemptive channel is in the background ready and listening for trigger conditions, forming a parallel execution instruction stack. Under normal operating conditions, the robot agent drives the projection module to project dynamic nostalgic images onto the wall according to the task flow timeline, updating frame by frame as the narrative progresses. During the relaxed transition interval, the robot uses movable parts such as its robotic arm and head joints to simulate guiding actions of characters in the narrative scene, such as lifting and tilting the watering can in a watering scene. This visual demonstration guides the elderly to follow and perform rehabilitation actions, avoiding explicit prompts such as voice commands. This, together with the projected images, creates an emotional atmosphere, allowing the elderly to focus their attention on the narrative content itself.

[0029] During the maintenance of the emotional atmosphere field, the intensity of latent emotional tension is continuously updated and compared with a dynamic trigger threshold. Once the intensity of latent emotional tension exceeds the threshold, preemption is immediately initiated: the current projection frame and trunk joint posture are frozen, the narrative progress pointer and the completion of the action guidance are recorded, and a background task breakpoint scene is generated. Subsequently, a tactile perturbation rhythm pattern is randomly selected from a non-repetitive tactile rhythm pattern library. The tactile perturbation rhythm pattern is a set of non-repetitive tactile stimulus sequence templates preset in the initial configuration phase of the system. Each set is different in terms of stimulus interval, duration, and intensity change pattern, to ensure that the elderly cannot predict the temporal pattern of the next tactile stimulus based on past experience, thus avoiding habituation. The robot drives the contact actuator with the extracted rhythm pattern to generate a non-predictable tactile stimulus sequence on the elderly's body surface, thereby interrupting their immersion state and guiding their attention to drug delivery.

[0030] After drug delivery, the narrative progress pointer at the breakpoint is retrieved, and the projection and torso posture simulation are restored from this starting point to ensure continuous scene transition. Considering the need for a buffer period for attention to re-immerse after being seized, an emotional re-immersion lag decay interval is inserted within the first relaxed transition interval after restoration. Its duration is determined by the degree of rhythm deviation from the life rhythm fingerprint; the greater the deviation, the longer the buffer. During this interval, the projection and posture simulation operate normally but no new key action nodes are introduced. After this interval, the regular guiding rhythm is restored, and subsequent task flows continue.

[0031] Step four specifically involves: During the execution of the emotional narrative background task flow, the depth sensor on the robot agent is simultaneously activated to acquire the skeletal joint point cloud data of the elderly, and the actual action posture sequence corresponding to the rehabilitation training action in the temporal action-narrative binding mapping table is extracted from the skeletal joint point cloud data. At the same time, the facial micro-expression electromyography signal and skin conduction signal of the elderly are collected through wearable devices. The circadian rhythm fingerprint is retrieved, and based on the time phase label of the elderly's daily physiological rhythm in the circadian rhythm fingerprint, the skin conductance baseline value and micro-expression baseline template that match the corresponding time phase label are retrieved from the preset individualized physiological baseline library to construct an immersion assessment baseline for the current task execution period. The actual action posture sequence is compared with the key action nodes of the character behavior sequence in the emotional narrative background task flow frame by frame to obtain the action follow deviation. Simultaneously, the facial micro-expression electromyography signals and the micro-expression baseline template are differentially quantified to obtain the facial emotional response attenuation degree. The movement following deviation degree and the facial emotional response attenuation degree are then weighted and synthesized. The weight of the movement following deviation degree is determined by mapping the physiological arousal level corresponding to the skin conductance benchmark value in the immersion assessment baseline, and the weight of the facial emotional response attenuation degree is determined by mapping the resting activity stability corresponding to the micro-expression baseline template. Based on this weighted synthesis, the emotional immersion degree is generated. Temporal analysis of the joint angular velocity between adjacent frames in the skeletal joint point cloud data is performed to extract the postural micro-tremor features that occur during the elderly's rehabilitation training movements; and the exercise fatigue level is generated based on the postural micro-tremor features and the attenuation slope of the skin conduction response in the skin conduction signal. Using the emotional immersion and exercise fatigue as joint correction factors, the implicit emotional tension generated in step one is updated, and the updated implicit emotional tension is mapped and correlated with the deviation of movement following, so as to determine the rehabilitation narrative compliance coefficient, which represents the elderly’s compliance with rehabilitation guidance in the emotional narrative background task flow.

[0032] It should be noted that facial micro-expression electromyography (EMG) signals are acquired using surface electrodes attached to the elderly person's face, reflecting changes in the weak electrical activity of facial muscle groups related to emotional expression; skin conductivity signals are measured using electrodes at the fingertips or wrists, measuring fluctuations in skin conductivity with sweat gland secretion, and are a commonly used indicator for assessing the level of autonomic nerve activation. Simultaneously with acquiring these signals, the system retrieves the circadian rhythm fingerprint generated in step one, and based on the time phase label of the elderly person's daily physiological rhythm, retrieves the corresponding baseline data from a pre-defined individualized physiological baseline database. This individualized physiological baseline database is a physiological state reference database established for each elderly person during long-term system operation. It is constructed by recording the skin conductivity levels of the elderly person in a calm and relaxed state at various time phase labels, as well as the resting activity patterns of facial EMG when there is no specific emotional task, over a continuous monitoring period of several days. After statistical smoothing, the corresponding skin conductivity baseline values ​​and micro-expression baseline templates for each time phase are formed. After determining the immersion assessment baseline, the system quantifies the elderly person's response from two dimensions. The first dimension is the movement following ability. The actual movement posture sequence is compared frame-by-frame with the key movement nodes of the character behavior sequence in the emotional narrative background task flow to obtain the movement follow-up deviation. This deviation reflects the degree of difference between the elderly person's actual rehabilitation movements and the demonstrated movements in terms of spatial trajectory and temporal coordination. The second dimension is facial emotional response, which involves quantifying the difference between facial micro-expression electromyographic signals and micro-expression baseline templates. The specific method of difference quantification is as follows: extract the temporal feature parameters of the electromyographic signals within the current time window, and calculate the difference with the same feature parameters in the baseline template under the corresponding time phase label. The larger the difference, the farther the current electromyographic activity deviates from the resting baseline. However, in the context of emotional immersion, a moderate deviation of electromyographic activity from the resting baseline precisely reflects that the elderly person is generating an emotional response to the narrative content rather than being in a state of indifference. Therefore, this difference is compared with the preset normal emotional response range. When the difference is lower than the lower limit of the range, it indicates that the facial emotional response tends to be weak, and the facial emotional response attenuation is quantified accordingly. The movement follow-up deviation and the facial emotional response attenuation are weighted and synthesized under the immersion assessment baseline to generate the emotional immersion level. When weighted, the weights of the two indicators can be adjusted according to the actual application scenario. For example, in the rehabilitation stage that focuses on action guidance, the deviation of action following can be given a higher weight, while in the narrative stage that focuses on emotional interaction, the opposite is true.

[0033] Then, temporal analysis was performed on the joint angular velocities between adjacent frames in the skeletal joint point cloud data to extract postural micro-tremors that appeared during the elderly's rehabilitation training movements. Postural micro-tremors refer to the high-frequency, low-amplitude irregular fluctuations superimposed on the joint angular velocity curve, which are usually related to the decline in motor control precision caused by muscle fatigue. Simultaneously, the attenuation slope of the skin conductance response signal was analyzed within the same time period. This slope reflects the rate at which the elderly's physiological arousal level gradually declines from an initially high state during continuous rehabilitation movements. The fatigue trend at the muscle level represented by the postural micro-tremors and the fatigue trend at the autonomic nervous level represented by the attenuation slope of the skin conductance response were jointly determined: when the intensity of the postural micro-tremors increased and the attenuation slope of the skin conductance response became steeper, it indicated that the elderly were experiencing fatigue accumulation in both motor control and peripheral physiological arousal dimensions. Based on this, the degree of change in both was comprehensively mapped to generate a motor fatigue level.

[0034] After obtaining the emotional immersion level and physical fatigue level, these are used as joint correction factors to update the implicit emotional tension generated in step one. The basic logic of the update is as follows: when the emotional immersion level is high, it indicates that the elderly person is currently in a good state of narrative participation, and the implicit emotional tension can be appropriately lowered to avoid over-intervention; when the physical fatigue level is high, the implicit emotional tension is appropriately increased to reflect the potential impact of physical load on the emotional state. The updated implicit emotional tension will be fed back to step five for adjusting the trigger threshold, and will also be used to calculate the rehabilitation narrative compliance coefficient.

[0035] Finally, the updated implicit emotional tension and the deviation from the action-following pattern were mapped and correlated to determine the rehabilitation narrative compliance coefficient. The specific mapping method was as follows: under the emotional state indicated by the updated implicit emotional tension, the magnitude of the action-following deviation was analyzed. Both factors combined reflect the extent to which the elderly person is currently willing and able to follow the rehabilitation guidance in the narrative. If the emotional tension is within an appropriate range and the action-following deviation is small, the rehabilitation narrative compliance coefficient is high, indicating that the elderly person currently has a good acceptance and execution of the rehabilitation guidance; conversely, the compliance coefficient is low, requiring corresponding adjustments to the narrative rate in subsequent steps.

[0036] Step five specifically involves: Obtain the updated implicit emotional tension, and retrieve the action guidance completion rate in the background task breakpoint scene recorded during the most recent preemptive execution of the preemptible salient task instance. Based on the action guidance completion rate and the updated implicit emotional tension, construct a preemption-emotional disturbance correlation pair that reflects the degree of disturbance of the emotional state caused by task preemption. Obtain the tactile perturbation rhythmic pattern sequence that has been extracted from the non-repetitive tactile rhythmic pattern library in the current task cycle, temporally couple the tactile perturbation rhythmic pattern sequence with the preemption-emotional perturbation correlation pair, calculate the stimulus-recovery correlation curve between tactile stimulation pattern and emotional recovery rate, and generate a task inhibition inertial field that acts on the next cycle based on the stimulus-recovery correlation curve. The task suppression inertial field is superimposed on the dynamic trigger threshold, and the dynamic trigger threshold is biased and corrected so that the trigger sensitivity of the preemptible salient task instance gradually recovers as the task suppression inertial field decays, generating a corrected trigger threshold with hysteresis recovery characteristics. Obtain the rehabilitation narrative compliance coefficient, trace back the frame-by-frame deviation between the actual action posture sequence recorded during the execution of the emotional narrative background task flow and the key action timeline, locate the weak compliance segment in the key action timeline where the action following deviation continuously exceeds the preset limit, and timestamp the weak compliance segment on the narrative advancement timeline. The original triggering conditions of the preemptible highlight task instance are replaced according to the modified trigger threshold, and non-uniform inter-frame interval stretching is applied to the weak compliance segments marked on the narrative advancement timeline, so that the narrative advancement rate of the emotional narrative background task flow at the weak compliance segments is lower than the narrative advancement rate of the other segments, forming a closed-loop adjustment of local adaptive slow motion.

[0037] The task inhibition inertial field includes an inhibition intensity attenuation coefficient determined by the rate of change of the stimulus interval between two adjacent tactile stimuli in the tactile perturbation rhythmic sequence, and an initial inhibition amplitude determined by the ratio of the action-guided completion degree to the updated implicit emotional tension.

[0038] It should be noted that, in terms of trigger threshold adjustment, the purpose of constructing the preemption-emotional perturbation correlation pair is to quantify the actual impact of each preemption event on the elderly person's emotions. By establishing a correspondence between the timing of preemption (measured by the completion of the action guidance) and the consequences of preemption (measured by the updated implicit emotional tension) through the correlation pair, a factual basis is provided for subsequent judgments on whether to suppress subsequent preemption. Based on this, the tactile perturbation rhythmic pattern sequence is temporally coupled with the preemption-emotional perturbation correlation pair. The specific method of temporal coupling is as follows: using the occurrence time of each preemption event as an alignment benchmark, the rate of change of the stimulus interval of the tactile perturbation rhythmic pattern used in each preemption is paired with the corresponding emotional recovery rate on the same time coordinate, forming a set of correlation data points reflecting "how tactile stimulation changes - how emotions recover." Curve fitting of these data points yields the stimulus-recovery correlation curve. The emotional recovery rate refers to the reciprocal of the time taken from the end of the preemption until the implicit emotional tension returns to the normal fluctuation range. If the stimulus-recovery correlation curve shows a negative correlation, it indicates that a more uneven tactile pattern in terms of stimulus intervals leads to a slower emotional recovery. Based on this, the system generates a task-inhibition inertia field. The physical meaning of this task-inhibition inertia field is to apply a time-decreasing inhibitory effect to the preemptible salient task trigger condition. Its initial inhibitory amplitude is determined by the ratio of the action-guided completion degree to the updated implicit emotional tension. The smaller this ratio, the more severe the emotional disturbance caused by the previous preemption, and the stronger the initial inhibitory intensity should be. The inhibition intensity decay coefficient is determined by the rate of change of the stimulus interval; the greater the rate of change, the slower the decay and the longer the inhibitory effect lasts. After this inertia field is superimposed on the dynamic trigger threshold, the threshold is temporarily raised and gradually falls back as the inertia field decays, thus giving the trigger condition a delayed recovery characteristic and preventing re-triggering of preemption during emotional recovery.

[0039] Regarding the adjustment of narrative progression rate, the system's logic for locating weak compliance segments is as follows: It traces the frame-by-frame deviation between the actual action posture sequence and the key action timeline. When the deviation exceeds a preset limit for several consecutive frames, the frame sequence is marked as a weak compliance segment. The preset limit can be pre-determined based on the individualized rehabilitation assessment results of the elderly. The non-uniform inter-frame interval stretching applied to the marked segments refers to increasing the time interval between adjacent frames within the marked segment without changing the inter-frame interval of unmarked segments, thus achieving local slowdown rather than overall deceleration. The degree of stretching is negatively correlated with the rehabilitation narrative compliance coefficient; the lower the compliance, the greater the stretching, and the more ample the follow-up time given to the elderly. After these two correction operations are written into the trigger conditions of the preemptible salient task instance and the inter-frame interval parameters of the narrative progression timeline, the closed-loop adjustment for this round is completed.

[0040] In a preferred embodiment of the present invention, it further includes: Obtain the circadian rhythm fingerprint sequence and the latent emotional tension sequence generated within a fixed time window over a consecutive preset number of days. Reconstruct the circadian rhythm fingerprint sequence and the latent emotional tension sequence into a joint trajectory in a multidimensional state space. The spatiotemporal transition entropy values ​​of the circadian rhythm fingerprint are used as the first-dimensional coordinate component of the trajectory, and the latent emotional tension values ​​are used as the second-dimensional coordinate component of the trajectory. Construct a set of state points for the joint trajectory under each time window. Calculate the distance between any two state points in the state point set and compare it with a preset neighborhood radius to generate a cross recursion matrix. In the cross recursion matrix, a matrix element with a value of one indicates that the joint state at the corresponding two time points is in a recursive state. Calculate the diagonal segment length distribution of the texture structure formed by the recursive points in the cross recursion matrix and extract the Shannon entropy of the diagonal segment length distribution as the system recursive complexity. Local extremum detection is performed on the evolution trajectory of the recursive complexity over time to obtain the turning point when the recursive complexity changes from steady fluctuation to monotonically decreasing. The state point corresponding to the turning point is marked as the laminar phase transition critical point to characterize the initial boundary of the rhythm-emotion coupling system from a non-periodic ordered state to a random disordered state. Randomly extract the action frame rate fine-tuning parameters and scene tone offset parameters of the guide character from the nostalgic narrative material library, and combine them to generate at least one narrative rhythm variant under the semantic invariance constraint of the rehabilitation guidance. When the system state is detected to have reached the critical point of the laminar phase transition, the generated narrative rhythm variant is injected into the relaxed transition interval of the original emotional narrative background task flow according to the timestamp, so as to disrupt the exponential divergence trend of the positive phase space trajectory, keep the recursive complexity in a stable fluctuation range, and thus delay the rigidification process of the rhythm-emotion coupling system.

[0041] It should be noted that during long-term system operation, although the aforementioned closed-loop regulation can effectively handle emotional fluctuations and compliance changes within a single task execution cycle, as the elderly become increasingly familiar with the narrative scenarios and rehabilitation guidance models, the response pattern of their rhythm-emotion coupling system may exhibit a monotonic trend. That is, the evolution trajectory of the system state in phase space gradually tends to repeat within a limited region, losing its original rich dynamic characteristics. If this trend is not addressed, it will lead to a gradual decline in the emotional arousal ability of emotional narratives on the elderly, thereby reducing long-term compliance with rehabilitation training. To address this, this embodiment reconstructs the life rhythm fingerprint sequence and the latent emotional tension sequence into a joint trajectory in a multi-dimensional state space. The reconstruction method is as follows: the spatiotemporal transition entropy values ​​of each life rhythm fingerprint are used as the first-dimensional coordinate component of the trajectory, and the latent emotional tension value is used as the second-dimensional coordinate component of the trajectory. This constructs a set of state points for the joint trajectory under each time window on a two-dimensional phase plane. Then, the distance between any two state points is calculated, and this distance is compared with a preset neighborhood radius to generate a cross-recursive matrix. The cross-recursion matrix is ​​a square matrix composed of zeros and ones, with rows and columns corresponding to time windows. When the distance between the joint states corresponding to two time windows in phase space is less than the neighborhood radius, the matrix element at the corresponding position takes a value of one, indicating that the joint state at these two times is in a recursive state, meaning the system exhibits similar dynamic behavior patterns at these two times; otherwise, the value is zero. Based on this, the diagonal segment length distribution of the texture structure formed by the recursive points in the cross-recursion matrix is ​​calculated. A diagonal segment refers to a segment in the matrix that continuously takes a value of one along the diagonal direction; its length reflects the duration for which the system states maintain similar trajectories in phase space. The Shannon entropy of this diagonal segment length distribution is extracted as the system's recursive complexity. Recursive complexity essentially quantifies the richness of the evolutionary patterns of the rhythm-emotion coupling system in state space: higher complexity indicates more diverse system state change patterns and better adaptive flexibility; continuously decreasing complexity indicates that the system is becoming rigid.

[0042] To promptly identify the starting point of a stagnant trend, local extrema detection is performed on the evolution trajectory of recursive complexity over time. This identifies the turning point where the recursive complexity transitions from steady fluctuation to monotonically decreasing, and the state point corresponding to this turning point is marked as the laminar phase transition critical point. The physical meaning of the laminar phase transition critical point is the initial boundary of the rhythm-emotion coupling system's evolution from an aperiodic ordered state to a random disordered state. Before this critical point, the system is still in a healthy dynamic fluctuation range; after crossing this critical point, the system will accelerate the loss of its original dynamic richness.

[0043] When the system state is detected to have reached the critical point of laminar phase transition, a micro-perturbation injection mechanism is triggered. Specifically, the frame rate adjustment parameters for the guide character's actions and the scene tone offset parameters are randomly extracted from the nostalgic narrative material library. The frame rate adjustment parameters are used to slightly increase or decrease the playback frame rate of the character's actions while maintaining the semantics of rehabilitation guidance, for example, slightly speeding up or slowing down the frame rate of the teapot-carrying action. The scene tone offset parameters are used to slightly shift the overall tone of the projected image, for example, adding a slight cool tone offset to a warm tone. The above parameters are combined to generate at least one narrative rhythm variant, which remains consistent with the original task flow at the level of rehabilitation guidance semantics, introducing only random micro-variations in the sensory presentation rhythm. The generated narrative rhythm variant is injected into the relaxed transition interval of the original emotional narrative background task flow according to the timestamp, so that the original relaxed transition interval is no longer a completely repetitive pattern, but a rhythmic perturbation with non-repetitive micro-variations. The role of this micro-perturbation is to disrupt the exponential divergence trend of the positive phase space trajectory, keeping the recursive complexity within a stable fluctuation range, thereby delaying the rigidification process of the rhythm-emotion coupling system. This embodiment introduces an active dynamic diversity preservation capability into the system at the long-term operation level through the above-described method, avoiding the problem of response dulling in the elderly caused by long-term repetitive interactions, and further improving the long-term operation performance of the system.

[0044] In a preferred embodiment of the present invention, it further includes: The spatiotemporal transfer entropy sequence, environmental acoustic event frequency sequence, and rehabilitation narrative compliance coefficient sequence in the circadian rhythm fingerprint are obtained. A three-channel multivariate time series tensor is constructed by aligning with the timestamp and divided into historical sequence segments and real-time sequence segments. A multivariate gated recurrent unit network is constructed. The update gate weight matrix of its hidden state update module imposes a sparsity constraint on the spatiotemporal transition entropy channel column vector to reduce its contribution weight. The reset gate weight matrix imposes a decay constraint on the rehabilitation narrative compliance coefficient channel column vector to extend the retention time of historical information. The network is trained using historical sequence segments to learn the mapping from the input sequence to the spatiotemporal transition entropy value of the next time step. The real-time sequence segments are fed into the trained network frame by frame. The hidden state vector of each frame is extracted and the principal components are sorted. The first two principal components are taken as the latent variable representation of the overall life state. The two-dimensional coordinate trajectory of the latent variable representation forms the latent state trajectory line. The cumulative angle change of the latent state trajectory line within a continuous sliding time window is calculated to obtain the short-term oscillation intensity, which is then smoothed by a first-order low-pass filter. When the intensity of smooth oscillation exceeds the system rigidity threshold determined by the historical fluctuation range of the life rhythm fingerprint for multiple consecutive windows, an oscillation trigger timestamp is generated. Using the oscillation trigger timestamp as the start time of the freeze, the bias correction of the dynamic trigger threshold of the preemptive salient task by the task suppression inertial field is prevented, so that the dynamic trigger threshold remains at the value before the freeze until the smooth oscillation intensity of subsequent consecutive windows falls below the system rigidity threshold and the life rhythm fingerprint returns to the normal fluctuation band, at which point the freeze is lifted.

[0045] It should be noted that in the aforementioned embodiments, the task-inhibiting inertial field generates a time-decreasing inhibitory effect by controlling the degree of emotional disturbance caused by each preemption, effectively preventing medication reminders from being frequently triggered during periods of emotional fluctuation in the elderly. However, in long-term operation, when the overall living conditions of the elderly tend to be highly regular and stable over a longer time scale, the continuously accumulated residual inhibitory amount of the inertial field may unnecessarily raise the dynamic trigger threshold, thereby reducing the response sensitivity of medication reminders. This apparent stability is actually a precursor to the rigidity of the rhythm-emotion coupling system, which needs to be identified and the inhibition of the inertial field paused in a timely manner. To this end, this embodiment starts from three dimensions: behavioral rhythm, environmental interaction, and rehabilitation response, aligning the three in time and dividing them into historical sequence segments for network training and real-time sequence segments for online inference, constructing a three-channel multivariate time series tensor as the input for subsequent analysis.

[0046] At the network design level, the multivariate gated recurrent unit network used in this embodiment applies channel-specific constraints to the weight matrices of the update and reset gates during initialization: In the update gate, a sparsity constraint is applied to the spatiotemporal transfer entropy channel to weaken the dominance of the behavioral rhythm dimension in short-term state updates, preventing temporary behavioral fluctuations in the elderly from masking the more fundamental state changes reflected by environmental interaction and rehabilitation response dimensions; in the reset gate, a decay constraint is applied to the rehabilitation narrative compliance coefficient channel to extend the retention time of historical information in this channel, allowing the long-term trend of compliance to be expressed more persistently in the hidden state. After training, the temporal dependencies between the three channels are solidified within the network. During the online inference phase, the system extracts the hidden state vector of each frame and takes its first two principal components as latent variable representations, thus forming a latent state trajectory line on a two-dimensional plane. The cumulative angle change is used to quantify the intensity of motion of the latent state trajectory line within the sliding time window. Its intuitive meaning can be compared to how many degrees a person turns in a specific time period while walking on a plane, and the smooth oscillation intensity is obtained after first-order low-pass filtering. The system stagnation threshold is determined by the historical fluctuation range of the circadian rhythm fingerprint over long-term operation. When the intensity of smooth oscillations falls below this threshold for multiple consecutive windows, it indicates that the latent state trajectory line tends to flatten, and the variability of the system state is significantly reduced, which is identified as a stagnation trend. Based on this, the system generates an oscillation trigger timestamp and triggers a freeze operation. The core of the freeze operation is to prevent the task suppression inertial field from continuing to apply bias correction to the dynamic trigger threshold, keeping the threshold unchanged at its current value, thereby restoring the medication reminder's normal response sensitivity to latent emotional tension. The conditions for unfreezing are that the intensity of smooth oscillations rises back above the threshold and the circadian rhythm fingerprint returns to its normal fluctuation range, thus realizing the adaptive start and stop of the trigger threshold protection mechanism during system stagnation.

[0047] like Figure 2 As shown, the second aspect of the present invention discloses an intelligent agent multi-task collaborative monitoring system for elderly households. The system includes a memory and a processor. The memory stores a program for an intelligent agent multi-task collaborative monitoring method for elderly households. When the program for the intelligent agent multi-task collaborative monitoring method for elderly households is executed by the processor, the steps of any one of the methods described above are implemented.

[0048] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent agent multi-task collaborative monitoring in elderly households, characterized in that, Includes the following steps: Step 1: Collect the elderly’s home spatiotemporal behavior data through the multimodal sensor network of the robot agent, and extract the elderly’s life rhythm fingerprint and latent emotional tension from it. Step 2: Based on the aforementioned life rhythm fingerprint and latent emotional tension, construct an emotional narrative background task flow that implicitly embeds rehabilitation training actions into nostalgic narrative scenes, and set medication reminders as preemptible prominent tasks triggered by the intensity of the aforementioned latent emotional tension. Step 3: Control the robot agent to execute the emotional narrative background task flow. When the preemptible salient task is not triggered, generate dynamic nostalgic image projection to guide rehabilitation actions. Once triggered, interrupt the background task with an unpredictable tactile rhythm and execute drug delivery. Step 4: Use the depth sensor and wearable device on the robot to acquire the elderly’s movement posture, facial micro-expression and skin electrical signals, and combine them with the life rhythm fingerprint to analyze the emotional immersion and exercise fatigue, so as to update the implicit emotional tension and calculate the rehabilitation narrative compliance coefficient. Step 5: Adjust the trigger condition threshold of the preemptible salient task according to the updated implicit emotional tension, and adjust the narrative progression rate of the nostalgic image projection in the emotional narrative background task flow according to the rehabilitation narrative compliance coefficient.

2. The method according to claim 1, characterized in that, Step one specifically involves: The robot intelligent agent moves in the home environment according to the preset daily inspection trajectory, and simultaneously activates the millimeter-wave radar, passive infrared pyroelectric sensor and environmental microphone array in the multimodal sensor network to continuously collect the spatiotemporal location sequence of the elderly in different rooms, changes in body activity heat radiation and environmental acoustic events, and generate raw spatiotemporal behavioral stream data. The original spatiotemporal behavioral stream data is segmented in the time domain to extract the behavioral patterns of the elderly in each fixed time window over multiple consecutive days. Spatiotemporal transition entropy is calculated for the behavioral patterns, and the spatiotemporal transition entropy is used to construct a life rhythm fingerprint that characterizes the degree of deviation of the elderly’s individualized daily activity rhythm. Based on the life rhythm fingerprint, abnormal behavior segments that deviate from the normal rhythm in the daily behavior sequence are identified. The environmental acoustic events at the time when the abnormal behavior segments occurred are traced back to separate the fundamental frequency micro-perturbation features and pause prosodic abnormal features of the speech signal, thereby constructing an acoustic marker of unexpressed emotions. The emotional acoustic markers are time-aligned with the changes in somatic activity thermal radiation captured during the abnormal behavior segment. The non-periodic fluctuation component of the thermal radiation change is extracted as a physiological stress thermal pattern. The physiological stress thermal pattern is then coupled with the emotional acoustic markers through cross-modal feature coupling to generate latent emotional tension.

3. The method according to claim 1, characterized in that, Step two specifically involves: The life rhythm fingerprint is retrieved, and the time phase label of the elderly person's physiological rhythm on that day is matched from the preset personal historical behavioral rhythm benchmark library. At the same time, the implicit emotional tension is obtained, and the emotional-rhythm coupling deviation under the time phase label is calculated. Based on the aforementioned emotion-rhythm coupling deviation, narrative scene fragments associated with the elderly person's personal experiences are retrieved from the nostalgic narrative material library. The movement complexity level of the rehabilitation training movements is determined according to the degree of deviation of the activity rhythm in the life rhythm fingerprint. The rehabilitation training movements are then mapped to the character behavior sequence in the narrative scene fragments according to the movement complexity level, forming a temporal action-narrative binding mapping table. The temporal action-narrative binding mapping table is parsed, key action nodes with rehabilitation guidance semantics are extracted from the character behavior sequence, and each key action node is arranged into a key action timeline according to its bound narrative timestamp, and a narrative advancement timeline with dynamic nostalgic image projection synchronized with the key action timeline is generated. The key action timeline is aligned with the narrative progression timeline at the frame level. A relaxed transition interval, the duration of which is determined by the degree of rhythm deviation of the life rhythm fingerprint, is embedded between adjacent key action nodes. The key action timeline after embedding the relaxed transition interval is then synthesized with the narrative progression timeline to generate an emotional narrative background task flow. Using the implicit emotional tension as input, the implicit emotional tension is extrapolated to obtain the predicted peak time of emotional tension within a preset time window. The predicted peak time of emotional tension is marked as the pre-trigger time node for drug delivery, and a preemptible prominent task instance that shares a common time axis with the emotional narrative background task flow is generated. The preemptible salient task instance is assigned an asymmetric task weight that takes priority over the emotional narrative background task flow. When the intensity of the latent emotional tension collected in real time exceeds the dynamic trigger threshold adjusted by the degree of rhythm deviation of the life rhythm fingerprint, the preemptible salient task instance immediately preempts the execution resources of the robot agent.

4. The method according to claim 1, characterized in that, Step three specifically involves: The emotional narrative background task flow and the preemptible prominent task instance are obtained and loaded into the dual-channel task executor of the robot agent, and a parallel execution instruction stack is established with the emotional narrative background task flow as the main channel and the preemptible prominent task instance as the preemption channel. When no trigger signal of the preemptible salient task instance is detected, the robot agent is controlled to drive the projection module to project dynamic nostalgic images according to the emotional narrative background task flow, and simulates the guiding actions of the characters in the narrative scene segment through the torso posture of the robot agent in the relaxed transition interval, forming an emotional atmosphere field to guide the rehabilitation action. During the maintenance of the emotional atmosphere field, the intensity of implicit emotional tension is continuously monitored. When the intensity of implicit emotional tension exceeds the dynamic trigger threshold, the projection screen and torso posture simulation state of the current frame of the emotional narrative background task flow are immediately frozen, and a background task breakpoint scene containing the current narrative progress pointer and the completion degree of action guidance is generated. Activate the preemptible salient task instance, randomly select a tactile perturbation rhythm pattern from the pre-built non-repetitive tactile rhythm pattern library, control the robot agent to drive the contact actuator with the tactile perturbation rhythm pattern to generate an unpredictable tactile stimulation sequence on the elderly person's body surface, interrupt the elderly person's current immersion state and guide their attention to drug delivery. After completing drug delivery, the narrative progress pointer in the background task breakpoint is read, and the robot agent is controlled to restore the projection of the emotional narrative background task flow and the trunk posture simulation with the narrative progress pointer as the starting point. In the first relaxation transition interval after restoration, an emotional re-immersion lag decay interval with a duration determined by the degree of rhythm deviation of the life rhythm fingerprint is inserted.

5. The method according to claim 1, characterized in that, Step four specifically involves: During the execution of the emotional narrative background task flow, the depth sensor on the robot agent is simultaneously activated to acquire the skeletal joint point cloud data of the elderly, and the actual action posture sequence corresponding to the rehabilitation training action in the temporal action-narrative binding mapping table is extracted from the skeletal joint point cloud data. At the same time, the facial micro-expression electromyography signal and skin conduction signal of the elderly are collected through wearable devices. The circadian rhythm fingerprint is retrieved, and based on the time phase label of the elderly's daily physiological rhythm in the circadian rhythm fingerprint, the skin conductance baseline value and micro-expression baseline template that match the corresponding time phase label are retrieved from the preset individualized physiological baseline library to construct an immersion assessment baseline for the current task execution period. The actual action posture sequence is compared with the key action nodes of the character behavior sequence in the emotional narrative background task flow frame by frame to obtain the action follow deviation. Simultaneously, the facial micro-expression electromyography signal and the micro-expression baseline template are differentially quantified to obtain the facial emotional response attenuation degree. The action following deviation degree and the facial emotional response attenuation degree are weighted and synthesized under the immersion evaluation baseline to generate the emotional immersion degree. Temporal analysis of the joint angular velocity between adjacent frames in the skeletal joint point cloud data was performed to extract the postural micro-tremor features that appeared during the elderly’s rehabilitation training movements. Exercise fatigue level is generated based on the posture micro-tremor characteristics and the attenuation slope of the skin conductance response in the skin conductance signal. Using the emotional immersion and exercise fatigue as joint correction factors, the implicit emotional tension generated in step one is updated, and the updated implicit emotional tension is mapped and correlated with the deviation of movement following, so as to determine the rehabilitation narrative compliance coefficient, which represents the elderly’s compliance with rehabilitation guidance in the emotional narrative background task flow.

6. The method according to claim 1, characterized in that, Step five specifically involves: Obtain the updated implicit emotional tension, and retrieve the action guidance completion rate in the background task breakpoint scene recorded during the most recent preemptive execution of the preemptible salient task instance. Based on the action guidance completion rate and the updated implicit emotional tension, construct a preemption-emotional disturbance correlation pair that reflects the degree of disturbance of the emotional state caused by task preemption. Obtain the tactile perturbation rhythmic pattern sequence that has been extracted from the non-repetitive tactile rhythmic pattern library in the current task cycle, temporally couple the tactile perturbation rhythmic pattern sequence with the preemption-emotional perturbation correlation pair, calculate the stimulus-recovery correlation curve between tactile stimulation pattern and emotional recovery rate, and generate a task inhibition inertial field that acts on the next cycle based on the stimulus-recovery correlation curve. The task suppression inertial field is superimposed on the dynamic trigger threshold, and the dynamic trigger threshold is biased and corrected so that the trigger sensitivity of the preemptible salient task instance gradually recovers as the task suppression inertial field decays, generating a corrected trigger threshold with hysteresis recovery characteristics. Obtain the rehabilitation narrative compliance coefficient, trace back the frame-by-frame deviation between the actual action posture sequence recorded during the execution of the emotional narrative background task flow and the key action timeline, locate the weak compliance segment in the key action timeline where the action following deviation continuously exceeds the preset limit, and timestamp the weak compliance segment on the narrative advancement timeline. The original triggering conditions of the preemptible highlight task instance are replaced according to the modified trigger threshold, and non-uniform inter-frame interval stretching is applied to the weak compliance segments marked on the narrative advancement timeline, so that the narrative advancement rate of the emotional narrative background task flow at the weak compliance segments is lower than the narrative advancement rate of the other segments, forming a closed-loop adjustment of local adaptive slow motion.

7. The method according to claim 6, characterized in that, The task inhibition inertial field includes an inhibition intensity attenuation coefficient determined by the rate of change of the stimulus interval between two adjacent tactile stimuli in the tactile perturbation rhythmic sequence, and an initial inhibition amplitude determined by the ratio of the action-guided completion degree to the updated implicit emotional tension.

8. A smart agent multi-task collaborative monitoring system for elderly households, characterized in that, The system includes a memory and a processor. The memory stores a program for a multi-task collaborative monitoring method for elderly households. When the processor executes the program for the multi-task collaborative monitoring method for elderly households, it implements the steps of the method as described in any one of claims 1 to 7.

Images