A care robot home multi-source alarm traceability grading method

Abstract

The present application relates to a kind of care robot home multi-source alarm traceability grading method, specifically to the field of wisdom pension, by establishing unified timestamp and space coordinate system, the time-space alignment and fusion of multi-source sensor data are realized, effectively reduce single sensor false alarm;With the aid of preset cross-checking rule engine, identify candidate events and combine event confidence calculation formula Quantification degree of confidence, accurately distinguish different scenarios such as fall, fire coma, help-seeking;According to hard priority level and dynamic priority score function, conflict arbitration and dynamic sequencing are carried out on candidate events, to ensure that the most urgent event is responded first;By extracting space coordinates, structured alarm messages are generated and classified and pushed to different response parties, while the processing results are fed back to update weight parameters and historical recurrence coefficients, forming a closed-loop adaptive optimization, significantly improving the alarm accuracy, emergency response efficiency and long-term reliability of the home care system.

Description

Technical Field

[0001] This invention relates to the field of smart elderly care, and more specifically, to a method for tracing and classifying multi-source alarms in home care robots. Background Technology

[0002] With the accelerating aging of the population, the need for home safety monitoring for elderly people living alone is becoming increasingly urgent. Existing care robot systems generally integrate multiple sensors to achieve comprehensive perception of the elderly's living conditions. For example, millimeter-wave radar is used to detect human movement and posture, smoke sensors are used to monitor fire hazards, gas sensors are used to detect gas leaks, water immersion sensors are used to prevent water pipe bursts, one-button SOS is used for emergency assistance, voice call modules are used for voice interaction, and cameras are used for visual monitoring. In practical applications, the above sensors may capture different event signals within the same time window, and these signals often contradict each other. For example, the smoke sensor in the kitchen area may trigger a minor alarm due to a small amount of steam generated by cooking, while at the same time, the millimeter-wave radar in the living room detects the rapid descent of the elderly person as they bend over to pick up an item, while the camera's visual image does not show smoke or signs of a fall. In such situations, the system receives alarm signals from multiple sources simultaneously, but cannot determine which one is the real emergency, nor can it determine the causal or independent relationship between the events. This phenomenon of conflicting multi-source alarm signals occurs frequently in the home scenarios of elderly people living alone, seriously affecting the reliability and practicality of alarm systems.

[0003] Current monitoring equipment generally uses a single sensor-triggered alarm method, lacking effective coordination and cross-verification mechanisms between sensors. When multiple sensors alarm simultaneously or sequentially, the system typically pushes all alarm information independently to the monitoring center or family terminal, resulting in the recipient receiving multiple conflicting alarm messages and being unable to distinguish the authenticity and urgency of the event. Although some existing solutions integrate multiple alarm functions, their alarm processing methods remain at the stage of independent sensor triggering and SMS push, without introducing cross-verification rules and hierarchical logic for multi-source signals. Therefore, existing technologies cannot solve the problem of distinguishing conflicting alarm signals. The inability to distinguish between real-life scenarios such as falls, fires, and calls for help, and the inability to prioritize and dispatch tasks according to the urgency of events, directly leads to a high false alarm rate in alarm systems. This causes alarm fatigue among caregivers, leading them to ignore genuine emergency calls. Furthermore, incorrect prioritization can result in misallocation of emergency resources, delaying rescue of critically ill elderly individuals and even causing irreversible safety accidents. Therefore, there is an urgent need for an intelligent decision-making method capable of cross-validating multi-source alarm signals, prioritizing and arbitrarily dispatching tasks, and simultaneously tracing and locating the source, in order to reduce false alarm rates and improve the reliability and emergency response efficiency of home care systems. Summary of the Invention

[0004] The present application aims at the technical problems existing in the prior art, and provides a nursing robot home multi-source alarm traceability grading method to solve the problems in the background art.

[0005] The technical scheme for solving the above technical problems is as follows: specifically comprising the following steps: Step S1, a unified time stamp reference and a space coordinate system are established to perform space-time alignment and fusion on point cloud data from a millimeter wave radar, analog quantity values of a smoke sensor, a gas sensor and a water immersion sensor, a video frame sequence of a camera, and a Boolean signal of a one-key help button and a voice calling module, and to generate an extended multi-source fusion data frame; Step S2, based on the extended multi-source fusion data frame, a preset cross-checking rule engine is called to identify candidate events, and an event confidence calculation formula is used to calculate the confidence of each candidate event, wherein the event confidence calculation formula contains weight parameters of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module, so as to output an event list containing an event type label, a confidence and a trigger time stamp; Step S3, according to a preset hard priority level, a plurality of candidate events in the event list are subjected to conflict arbitration to obtain a candidate subset, and a dynamic priority scoring function is used to dynamically sort a plurality of candidate events in the candidate subset, wherein the dynamic priority scoring function contains the confidence, the inverse of the interval from the event to the current time and a historical recurrence coefficient, so as to determine a unique alarm event; Step S4, according to the unique alarm event, a corresponding space coordinate is extracted from the extended multi-source fusion data frame to generate a structured alarm message containing an event type label, a confidence, a trigger time stamp and a space coordinate, and the structured alarm message is pushed to different response parties according to the priority level, and the processing result after the push is fed back to update the weight parameters in the event confidence calculation formula and the historical recurrence coefficient in the dynamic priority scoring function; In one preferred embodiment, in the step S1, the main controller clock of the nursing robot is taken as a global time reference, and a global synchronous time stamp is stamped on each frame of data in the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signal of the one-key help button and the voice calling module. Then, the point cloud data of the non-uniformly sampled millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signals of the one-key help button and the voice calling module are unified to a fixed time grid by using an asynchronous alignment algorithm based on time offset mutual information maximization. For any one of the sensors, the reconstructed value at the alignment time is obtained by the following process: First, the output data values of any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module at all original sampling time points in the alignment time neighborhood are determined. Second, the time difference between each original sampling time point and the alignment time is calculated, and the adaptive kernel width is adaptively adjusted according to the sampling interval fluctuation variance of any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module in the nearest sliding window. Third, the exponentially weighted weight is calculated based on the time difference and the adaptive kernel width by using an exponential function. The form of the exponential function is e raised to the power of the ratio of the square of the negative time difference to twice the square of the adaptive kernel width. Finally, the calculated weights are normalized, and the normalized weights are multiplied by and summed with the output data values of the corresponding original sampling time points, thereby obtaining the reconstructed value of any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module at the alignment time, and generating the time-aligned point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signals of the one-key help button and the voice calling module.

[0006] In a preferred embodiment, the specific process of generating the extended multi-source fusion data frame is as follows: The time-aligned point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signals of the one-key help button and the voice calling module are mapped to a three-dimensional world coordinate system with the geometric center of the robot chassis as the origin and the north direction as the Y axis. For the millimeter wave radar, the polar coordinate form of the point cloud data is converted into the Cartesian coordinate form. For the camera, the pixel coordinates in the video frame sequence are back-projected to the three-dimensional space points by using the pre-calibrated intrinsic matrix and extrinsic rotation and translation matrix. The points in the local coordinate system of each sensor are converted to the points in the three-dimensional world coordinate system by using the space transformation matrix updated in real time by the factory calibration and the robot self-localization algorithm. The space transformation matrix is realized by a homogeneous transformation matrix composed of a rotation matrix and a translation vector. Then, all the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor after time alignment and space alignment, the video frame sequence of the camera, and the Boolean signal of the one-key help button and the voice calling module are organized into a structured multi-source fusion data frame according to time index, the multi-source fusion data frame is composed of a time stamp, the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the video frame sequence of the camera, the Boolean signal of the one-key help button, and the audio feature vector captured by the voice calling module in the time grid neighborhood; Finally, tensor decomposition fusion is performed on the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the pixel values of the video frame sequence of the camera, and the audio feature vector of the voice calling module in the multi-source fusion data frame, to form an extended multi-source fusion data frame, wherein the extended multi-source fusion data frame is composed of a time stamp, the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the video frame sequence of the camera, the Boolean signal of the one-key help button, the audio feature vector of the voice calling module, and a fusion feature vector, and the point cloud data of the millimeter wave radar contains multiple points, each point contains three-dimensional coordinates and Doppler velocity.

[0007] In a preferred embodiment, the process of the tensor decomposition fusion is as follows: The point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the pixel values of the video frame sequence of the camera, and the audio feature vector of the voice calling module in the multi-source fusion data frame are organized into a three-order tensor, an approximate decomposition of the tensor with a rank of three is iteratively solved by using an alternating least squares method, the weight coefficients of the first three principal components are extracted as a fusion feature vector, the fusion feature vector is attached to the multi-source fusion data frame, and an extended multi-source fusion data frame is formed.

[0008] In a preferred embodiment, in the step S2, the specific operation of calling the preset cross-checking rule engine to identify the candidate event is as follows: Three rules are preset, the first rule is used to determine a fall candidate event when the vertical velocity of the center of mass extracted from the point cloud data of the millimeter wave radar in the extended multi-source fusion data frame is greater than one point five meters per second, the analog quantity value of the smoke sensor in the extended multi-source fusion data frame is less than five hundred parts per million concentration, and the smoke probability output by the smoke detection model of the video frame sequence of the camera in the extended multi-source fusion data frame is less than zero point three; The second rule is used to determine a fire coma candidate event when the analog quantity value of the smoke sensor in the extended multi-source fusion data frame is greater than or equal to 500 parts per million and the human body center displacement extracted from the point cloud data of the millimeter wave radar in the extended multi-source fusion data frame is less than 0.1 meters in one second. The third rule is used to determine a help candidate event when the Boolean signal of the one-key help button in the extended multi-source fusion data frame is true or the audio feature vector of the voice calling module in the extended multi-source fusion data frame matches the preset help vocabulary through the keyword recognition model.

[0009] In a preferred embodiment, the specific process of calculating the confidence of each candidate event by the event confidence calculation formula is as follows: For each triggered candidate event, a membership function of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module in the extended multi-source fusion data frame participating in the determination of the candidate event is defined, and the membership function adopts a nonlinear transformation to map the sensor observation value to a membership value between zero and one. Then, based on the membership values of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module, and the corresponding weight parameters, a nonlinear confidence calculation formula with an evidence conflict penalty term is used to calculate the confidence of the candidate event. The evidence conflict penalty term is the product of the complements of the membership values of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module multiplied by a preset conflict penalty coefficient. The specific calculation process of the nonlinear confidence calculation formula is as follows: the products of the weight parameters and the membership values of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module are summed to obtain the numerator, the weight parameters of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module are summed and then the evidence conflict penalty term is added to obtain the denominator, and finally the numerator is divided by the denominator to obtain the confidence. For each triggered candidate event, the event type label, the confidence and the trigger timestamp of the candidate event are recorded, and the three together constitute an event entry. All candidate event entries are organized into an event list in chronological order, and if multiple rules are triggered at the same timestamp, all corresponding candidate events are added to the time list.

[0010] In a preferred embodiment, in the step S3, the specific operation of conflict arbitration of the multiple candidate events in the event list according to the preset hard priority hierarchy is as follows: The preset hard priority level is sequentially decreased in the order of fire coma, help-seeking, and falling, wherein the fire coma corresponds to the highest priority, the help-seeking corresponds to the second highest priority, and the falling corresponds to the lowest priority; the specific assignment process is as follows: If the event type label of the candidate event is fire coma, the value three is assigned; if the event type label of the candidate event is help-seeking, the value two is assigned; if the event type label of the candidate event is falling, the value one is assigned; each candidate event in the event list is traversed, the corresponding hard priority value is assigned according to the event type label of the candidate event, the candidate event with the maximum hard priority value is selected to constitute a candidate subset; if the candidate subset contains only one candidate event, the candidate event is directly determined as the unique alarm event, and the dynamic sorting is not performed; if the candidate subset contains multiple candidate events, the dynamic sorting is entered.

[0011] In a preferred embodiment, in the step S3, the specific operation of dynamically sorting the multiple candidate events in the candidate subset by the dynamic priority scoring function is as follows: For each candidate event in the candidate subset, the dynamic priority score is calculated, and the calculation process is as follows: Firstly, the first factor is calculated, and the confidence of the candidate event is squared; secondly, the second factor is calculated, and the second factor is the inverse of the time interval from the event; thirdly, the third factor is calculated, and the historical recurrence coefficient corresponding to the event type label of the candidate event is multiplied by zero point three; the result of the multiplication of the first factor and the second factor is multiplied by zero point seven, and then added to the third factor to obtain the dynamic priority score of the candidate event; Finally, the candidate event with the highest dynamic priority score is selected as the unique alarm event, and if there are multiple candidate events with equal dynamic priority scores and all the highest, the candidate event with the earliest trigger timestamp is selected as the unique alarm event.

[0012] In a preferred embodiment, in the step S4, the process of extracting the corresponding spatial coordinates from the extended multi-source fusion data frame according to the unique alarm event is as follows: Firstly, the most relevant sensor data of the unique alarm event is screened out from the extended multi-source fusion data frame according to the event type label of the unique alarm event; for the falling event, the three-dimensional coordinates of the human body center in the point cloud data of the millimeter wave radar in the extended multi-source fusion data frame are preferentially adopted; for the fire coma event, the center coordinates of the smoke area in the video frame sequence of the camera are combined with the installation position of the smoke sensor in the extended multi-source fusion data frame; for the help-seeking event, the sound source positioning result of the voice calling module and the preset position of the one-key help-seeking button in the extended multi-source fusion data frame are adopted, wherein the sound source positioning result is calculated by the time difference of arrival algorithm from the audio feature vector of the voice calling module. Then the multi-source weighted centroid positioning algorithm is used to fuse the position estimates provided by the millimeter wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-key help button and voice calling module participating in positioning; the sensor set participating in positioning is constructed, for each sensor in the sensor set, the corresponding weight is multiplied with the provided position coordinate component respectively and accumulated, then the sum of each component after accumulation is divided by the sum of all weights respectively, so as to obtain the final spatial coordinates; Next, the event type label, confidence, trigger timestamp and the final spatial coordinates calculated of the unique alarm event are packaged into a structured alarm message; Finally, according to the event type label of the unique alarm event, a hierarchical push strategy is executed: if the event type label is fire coma, it is pushed to the fire department, first-aid center, family terminal and community care center at the same time, and the push channel includes SMS, telephone voice call and application push; if the event type label is help, it is pushed to the family terminal and community care center, and the push channel is SMS and application push; if the event type label is fall, it is only pushed to the family terminal, and the push channel is application push, and all the pushes carry the generated structured alarm message.

[0013] In a preferred embodiment, the processing result after push is fed back to update the weight parameter in the event confidence calculation formula and the historical recurrence coefficient in the dynamic priority scoring function, specifically: The feedback signal returned by the responding party is listened to, and the feedback signal is a binary variable, where the value one represents a true event, and the value zero represents a false alarm; for the weight parameter in the event confidence calculation formula, the confidence propagation update rule is used, the new value of the weight parameter is set to the old value of the weight parameter plus the product of zero point one and the difference between the feedback signal and the historical average feedback rate, and then multiplied by the membership value of the millimeter wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-key help button and voice calling module in the unique alarm event; wherein the initial value of the historical average feedback rate is zero point five; For the historical recurrence coefficient in the dynamic priority scoring function, the exponential moving average method is used, the new value of the historical recurrence coefficient is set to the product of zero point three and the feedback signal plus the product of zero point seven and the old value of the historical recurrence coefficient.

[0014] The beneficial effects of the present application are: by establishing a unified time stamp and space coordinate system, the spatio-temporal alignment and fusion of multi-source sensor data are realized, and the false alarm of a single sensor is effectively reduced; by means of a preset cross-checking rule engine, candidate events are identified and the confidence level is quantified by combining with an event confidence calculation formula, so that different scenes such as falls, fire coma and help are accurately distinguished; according to the hard priority level and the dynamic priority scoring function, the candidate events are conflict arbitrated and dynamically sorted, so that the most urgent event is ensured to be responded first; by extracting the spatial coordinates to generate a structured alarm message and pushing it to different response parties in stages, the processing result is fed back to update the weight parameter and the historical recurrence coefficient, forming a closed loop adaptive optimization, and the alarm accuracy, emergency response efficiency and long-term reliability of the home care system are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a method flowchart of the present application; Figure 2 is a home environment sensor arrangement schematic diagram of the present application; Figure 3 is a raw signal waveform and an aligned waveform schematic diagram of the present application; Figure 4 is a fusion feature vector weight coefficient heat map of the present application; Figure 5 is a sensor membership function curve schematic diagram of the present application Figure 1 . Figure 6 is a dynamic priority scoring three-factor comparison schematic diagram of the present application; Figure 7 is a sensor weight parameter adaptive update curve schematic diagram of the present application; Figure 8 is a historical recurrence coefficient evolution curve schematic diagram of the present application; Figure 9 is a sensor membership function curve schematic diagram of the present application Figure 2 . DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1: This example provides the following... Figure 1-9 The method for tracing and classifying multi-source alarms in home care robots, as shown, specifically includes the following steps: Step S1: Establish a unified timestamp reference and spatial coordinate system to spatiotemporally align and fuse point cloud data from millimeter-wave radar, analog values ​​from smoke sensors, gas sensors and water immersion sensors, video frame sequences from cameras, and Boolean signals from one-click emergency button and voice call module to generate extended multi-source fusion data frames. Step S2: Based on the expanded multi-source fusion data frame, the preset cross-validation rule engine is called to identify candidate events, and the confidence of each candidate event is calculated using the event confidence calculation formula. The event confidence calculation formula includes weight parameters for millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-click emergency button and voice call module, thereby outputting an event list containing event type label, confidence and trigger timestamp. Step S3: Arbitrate conflicts among multiple candidate events in the event list according to the preset hard priority level to obtain a candidate subset. Then, dynamically sort the multiple candidate events in the candidate subset using a dynamic priority scoring function. The dynamic priority scoring function includes confidence, the reciprocal of the time interval between the event and the current time, and the historical recurrence coefficient, thereby determining the unique alarm event. Step S4, extracting the corresponding spatial coordinates from the extended multi-source fusion data frame according to the unique alarm event to generate a structured alarm message containing the event type label, confidence, trigger timestamp and spatial coordinates, and pushing the structured alarm message to different response parties according to the priority level, while feeding back the processing result after pushing to update the weight parameter in the event confidence calculation formula and the historical recurrence coefficient in the dynamic priority scoring function.

[0020] In this embodiment, it is specifically required to note that in step S1, the main controller clock of the care robot is taken as the global time reference, and a nanosecond-level global synchronization timestamp is stamped on each frame of data in the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signals of the one-key help button and the voice calling module. The arrangement of various sensors in the home environment is as shown in Figure 2 Then an asynchronous alignment algorithm based on time offset mutual information maximization is adopted, wherein the asynchronous alignment algorithm based on time offset mutual information maximization refers to: for each pair of sensors, the mutual information value of the observation sequence of the two sensors under different time offset is calculated, and the time offset that makes the mutual information value maximum is selected as the relative time delay estimation value of the sensor pair; then, taking the main controller clock as the reference, the time delay estimation values of all sensors are globally optimized, so that the alignment error sum of squares of the sensor data on the unified time grid is minimum; the input of the algorithm is the original sampling time sequence of each sensor, and the output is the best time offset of each sensor relative to the reference clock, which is used for subsequent reconstruction calculation. The point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signals of the one-key help button and the voice calling module are unified to a fixed time grid, and the grid spacing of the time grid is twenty milliseconds. For any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module, the output data value at any time is obtained by performing the following calculation process on the reconstruction value at the alignment time: ​Firstly, the output data values of any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module at all original sampling time points in the neighborhood of the alignment time point are determined; secondly, the time difference between each original sampling time point and the alignment time point is calculated, and the adaptive kernel width of any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module is adaptively adjusted according to the sampling interval fluctuation variance in the nearest sliding window, wherein the adjustment mode of the adaptive kernel width is: setting a sliding window length of one thousand sampling points, calculating the fluctuation variance of the sampling interval in the window; if the fluctuation variance is less than a preset threshold of zero point zero one millisecond square, the adaptive kernel width is set to five milliseconds; if the fluctuation variance is greater than or equal to zero point zero one millisecond square, the adaptive kernel width is set to the square root of the fluctuation variance multiplied by ten milliseconds; the preset threshold and the calculation mode ensure that a narrower kernel width is used for a stable sampling interval sensor to retain high-frequency details, and a wider kernel width is used for a non-stable sampling interval sensor to smooth noise; then, based on the time difference and the adaptive kernel width, the exponentially weighted weight is calculated through an exponential function, wherein the form of the exponential function is based on the natural constant, and the index is the ratio of the square of the negative time difference to twice the square of the adaptive kernel width, that is, the weight is equal to the exponential function value of the natural constant, the square of the negative time difference between the original sampling time point and the alignment time point divided by twice the square of the adaptive kernel width of any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module; finally, the calculated weights are normalized, wherein the normalization processing refers to: summing the exponential weight values corresponding to all original sampling time points to obtain a weight sum, then dividing each exponential weight value by the weight sum to obtain the normalized weight, so that the sum of all normalized weights is equal to one, and multiplying the normalized weight by the output data value of the corresponding original sampling time point and summing, thereby obtaining the reconstruction value of any one of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module at the alignment time point, thereby generating the time-aligned point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signals of the one-key help button and the voice calling module, as shown in Figure 3 ; The specific process of generating the extended multi-source fusion data frame is: Map the point cloud data of all completed time-aligned millimeter wave radars, analog quantity values of smoke sensors, gas sensors and water immersion sensors, video frame sequences of cameras and Boolean signals of one-key help-seeking buttons and voice calling modules to a three-dimensional world coordinate system with the geometric center of the care robot chassis as the origin and the north direction as the Y axis; wherein for the millimeter wave radar, convert the point cloud data in polar coordinate form to point cloud data in Cartesian coordinate form; for the camera, use the pre-calibrated intrinsic matrix and extrinsic rotation and translation matrix to back-project the pixel coordinates in the video frame sequences of the camera to three-dimensional space points; update the spatial transformation matrix in real time through factory calibration and care robot self-localization algorithm, wherein the care robot self-localization algorithm uses an iterative closest point matching algorithm based on laser radar point cloud and a pre-built map to calculate the current position and attitude of the robot at a frequency of ten hertz, and update the rotation matrix and translation vector in the spatial transformation matrix accordingly; the fixed installation positions and orientations of each sensor relative to the geometric center of the robot chassis have been measured and stored in advance during the factory calibration stage, and are used as the initial value of the spatial transformation matrix to convert the points in the local coordinate system of the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensors, the gas sensors and the water immersion sensors, the video frame sequences of the cameras and the Boolean signals of the one-key help-seeking buttons and the voice calling modules to points in the three-dimensional world coordinate system; the spatial transformation matrix is composed of a three-by-three rotation matrix and a three-by-one translation vector, by combining the rotation matrix and the translation vector into a four-by-four homogeneous transformation matrix, wherein the first three rows and the first three columns are the rotation matrix, the first three rows and the fourth column are the translation vector, the fourth row is a zero vector and a fixed value of one, and the homogeneous coordinates of the points in the local coordinate system of the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensors, the gas sensors and the water immersion sensors, the video frame sequences of the cameras and the Boolean signals of the one-key help-seeking buttons and the voice calling modules are multiplied by the homogeneous transformation matrix to calculate the homogeneous coordinates of the points in the three-dimensional world coordinate system, and then the homogeneous coordinates of the points in the three-dimensional world coordinate system are converted to Cartesian coordinates; Then, the point cloud data of all millimeter wave radars, the analog quantity values of smoke sensors, gas sensors and water immersion sensors, the video frame sequence of the camera and the Boolean signals of the one-key help-seeking button and the voice calling module after time alignment and space alignment are indexed by time to form a structured multi-source fusion data frame, which is composed of a time stamp, the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the video frame sequence of the camera, the Boolean signal of the one-key help-seeking button and the audio feature vector captured by the voice calling module in the time grid neighborhood, wherein the time grid neighborhood refers to a time interval of fifty milliseconds before and after the alignment time; in this interval, the audio signal collected by the voice calling module is processed by frame, each frame is twenty-five milliseconds long, the frame shift is ten milliseconds, one hundred and twenty-eight-dimensional mel frequency cepstral coefficients are extracted from each frame, and then the mel frequency cepstral coefficients of all frames are averaged to obtain the audio feature vector of the alignment time; Finally, the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the pixel values of the video frame sequence of the camera and the audio feature vector of the voice calling module in the multi-source fusion data frame are subjected to tensor decomposition fusion to form an expanded multi-source fusion data frame, as shown in Figure 4 The distribution of the first three principal component weight coefficients under different time windows is shown, wherein the expanded multi-source fusion data frame is composed of a time stamp, the point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the video frame sequence of the camera, the Boolean signal of the one-key help-seeking button, the audio feature vector of the voice calling module and the fusion feature vector, wherein the point cloud data of the millimeter wave radar contains multiple points, each point contains three-dimensional coordinates and Doppler velocity; the units of the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor are ppm or percentage lower explosive limit; the audio feature vector of the voice calling module is one hundred and twenty-eight-dimensional mel frequency cepstral coefficients; the fusion feature vector is composed of the first three principal component weight coefficients extracted by tensor decomposition fusion; The process of tensor decomposition fusion is as follows: The point cloud data of the millimeter wave radar, the analog quantity values of the smoke sensor, the analog quantity values of the gas sensor, the analog quantity values of the water immersion sensor, the pixel values of the video frame sequence of the camera, and the audio feature vector of the voice call module in the multi-source fusion data frame are organized into a three-order tensor, wherein three dimensions of the three-order tensor are defined as follows: the first dimension is the number of sensor types, which is six; the second dimension is the feature dimension of each sensor at a single alignment moment, wherein the point cloud density value (the number of points in a unit cubic meter) of the point cloud data of the millimeter wave radar, the original values of the analog quantity values of the smoke sensor, the gas sensor and the water immersion sensor, the gray mean value of the entire frame image of the pixel values of the video frame sequence of the camera, and the first component of the 128-dimensional Mel frequency cepstral coefficient of the audio feature vector of the voice call module are taken; the third dimension is the length of the time window, which is ten consecutive alignment moments; therefore, the size of the three-order tensor is six times one times ten, and the approximate decomposition of the tensor with a rank of three is solved by using the alternating least squares method, wherein the iteration termination condition of the alternating least squares method is that the relative residual error of the approximate tensors obtained by two adjacent iterations and the original tensor is less than a preset threshold of zero point zero zero one, or the iteration number reaches a preset maximum value of one hundred; the approximate decomposition with a rank of three means that the original three-order tensor is approximately represented as the sum of three rank-one tensors, and each rank-one tensor is composed of the outer product of three factor vectors; the alternating least squares method fixes two factor vectors in turn, solves the least squares solution of the third factor vector, and iterates cyclically until convergence, and the weight coefficients of the first three principal components are extracted as the fusion feature vector, wherein the weight coefficients of the first three principal components are the scaling factors corresponding to the three rank-one tensors obtained after the convergence of the alternating least squares method, that is, the weight values of each rank-one tensor; the three weight values are arranged in descending order to form a three-dimensional fusion feature vector; the reason for selecting the first three principal components is that experiments show that the cumulative variance contribution rate of the first three principal components exceeds ninety-five percent, which is sufficient to retain the main cross-correlation information of the multi-source sensor data, and the fusion feature vector is attached to the multi-source fusion data frame to form an expanded multi-source fusion data frame.

[0021] In the embodiment, the specific operation of identifying the candidate event by calling the preset cross-checking rule engine in step S2 is specifically as follows: The preset three rules are as follows: the first rule is used to determine a fall candidate event when the vertical velocity of the center of mass extracted from the point cloud data of the millimeter wave radar in the extended multi-source fusion data frame is greater than 1.5 m / s, the analog quantity value of the smoke sensor in the extended multi-source fusion data frame is less than 500 ppm, and the smoke probability output by the smoke detection model for the video frame sequence of the camera in the extended multi-source fusion data frame is less than 0.3; wherein the smoke detection model adopts an image classification model based on a deep convolutional neural network, the input of which is each frame of image in the video frame sequence of the camera in the extended multi-source fusion data frame, the image size is scaled to 224x224 pixels, and normalization processing is performed; the internal structure of the model includes five convolutional layers and three fully connected layers, the convolutional layer uses a 3x3 convolutional kernel, the activation function is a rectified linear unit, and the pooling layer adopts maximum pooling; the model output is a smoke probability value between 0 and 1, representing the confidence degree of the presence of smoke in the current image; the model is trained in advance using 10,000 labeled images containing kitchen steam, fire smoke and normal scenes, and the accuracy of the trained model on the validation set reaches 98%; the preset smoke probability threshold of 0.3 is determined by receiver operating characteristic curve analysis, which can control the false positive rate below 1% while ensuring the recall rate; The second rule is used to determine a fire coma candidate event when the analog quantity value of the smoke sensor in the extended multi-source fusion data frame is greater than or equal to 500 ppm, and the human body center displacement extracted from the point cloud data of the millimeter wave radar in the extended multi-source fusion data is less than 0.1 m in one second; The third rule is used to determine a help candidate event when the Boolean signal of the one-key help button in the extended multi-source fusion data frame is true or the audio feature vector of the voice calling module in the extended multi-source fusion data frame matches the preset help vocabulary through the keyword recognition model, and the help candidate event is not affected by the state of other sensors; wherein the keyword recognition model adopts a language model based on a recurrent neural network, the input of which is the audio feature vector of the voice calling module in the extended multi-source fusion data frame, i.e. 128-dimensional mel frequency cepstral coefficients; the internal structure of the model includes two layers of long short-term memory networks, each layer has 256 hidden units, followed by a fully connected layer and a Softmax classification layer; the model output is the probability distribution of each preset help vocabulary, which includes "save me" "help" "save me" "SOS" a total of four vocabularies; when the output probability of any one of the preset help vocabularies is greater than 0.8, it is determined that the matching is successful; the model is trained in advance using 100,000 audio samples containing help voice and daily conversation, and the recall rate of the trained model on the test set is 99.95%, and the false positive rate is 0.02%; the preset matching probability threshold of 0.8 is determined by balancing the precision and recall rate. The preset threshold value of 1.5 m / s of the vertical velocity of the center of mass is based on measured data of 100 old people in different activity states, and the threshold value can distinguish between normal bending (the speed is usually less than 0.8 m / s) and falling (the speed is usually greater than 1.5 m / s), and has a safety margin of 0.7 m / s; the preset threshold value of 500 ppm of the analog quantity value of the smoke sensor is set according to the lower limit of the alarm concentration of the smoke detector in the national fire alarm standard GB4715-2005, to ensure that the alarm can be triggered at the initial stage of the fire; the preset threshold value of 0.1 m of the displacement of the center of mass of the human body is based on statistical analysis of the slight shaking of the old people in the stationary state, and the displacement of the center of mass is usually within 0.02 m when standing or sitting normally, and 0.1 m can exclude false judgments caused by breathing or slight shaking; The specific process of calculating the confidence of each candidate event by the event confidence calculation formula is as follows: For each triggered candidate event, the membership function of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module in the extended multi-source fusion data frame participating in the candidate event judgment is defined, and the membership function adopts a nonlinear transformation to map the sensor observation value to a membership value between zero and one. For the fall candidate event, the membership function of the millimeter wave radar is obtained by nonlinear transformation based on the difference between the vertical velocity of the center of mass and 1.5 m / s. The specific operation of the nonlinear transformation is as follows: calculate the power function value with the natural constant e as the base and the negative five times the difference as the index, and then divide the power function value by one plus the power function value, so that the larger the difference, the closer the membership to one, and the smaller the difference, the closer the membership to zero. The membership function of the smoke sensor is obtained by taking the complement of the nonlinear transformation based on the difference between the analog quantity value of the smoke sensor and 500 ppm. The specific operation of the nonlinear transformation is as follows: calculate the power function value with the natural constant e as the base and the negative 0.01 times the difference as the index, and then subtract one from the power function value, so that when the analog quantity value of the smoke sensor is much less than 500 ppm, the membership is close to one, and when the analog quantity value of the smoke sensor is much greater than 500 ppm, the membership is close to zero. The membership function of the camera is based on one minus the smoke probability, that is, the closer the smoke probability output by the smoke detection model to zero, the closer the membership to one, and the closer the smoke probability to one, the closer the membership to zero, as shown in FIGS. 1, 2 and 3. Figure 5 and Figure 9 As shown, the trends of the memberships of the millimeter wave radar, the smoke sensor and the camera with respect to the observation values are shown. Subsequently, based on the membership values of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module, and the weight parameters of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module, the confidence of the candidate event is calculated by using a nonlinear confidence calculation formula with an evidence conflict penalty term, wherein the weight parameters of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module are dynamically adjusted by the linear mapping of the fusion feature vector in the extended multi-source fusion data frame. The specific operation of linear mapping is to multiply the three weight coefficients of the fusion feature vector by the regression coefficient between negative one and one, sum them up, and take the non-negative value of the sum result as the weight parameter. The regression coefficient is obtained by offline training: collect one thousand groups of extended multi-source fusion data frames in typical scenarios, each group of data frame corresponds to a known real event type, and use the gradient descent method to optimize to maximize the correlation coefficient between the weight parameters of each sensor and the event confidence. The range of the regression coefficient is limited between negative one and one to prevent the weight parameter from being too large or too small, resulting in imbalance of confidence calculation. The specific operation of taking the non-negative value of the sum result is: if the sum result is less than zero point one, it is forced to be assigned to zero point one to ensure that each sensor at least contributes ten percent of the weight in the confidence calculation. The evidence conflict penalty term is the product of the complements of the membership degrees of the millimeter wave radar, the smoke sensor, the gas sensor, the water immersion sensor, the camera, the one-key help button and the voice calling module multiplied by a preset conflict penalty coefficient, and the value of the preset conflict penalty coefficient is two. The value of two of the conflict penalty coefficient is determined by simulation experiment: on a test set containing five hundred groups of multi-source signal conflict scenes, different values of the conflict penalty coefficient from zero point five to five are tested respectively, and the weighted sum of the system false alarm rate and the miss rate is minimized as the target, and the optimal value is two. This value makes the confidence decrease by about forty percent when the sensor signals are completely contradictory (for example, one sensor membership is zero point nine and the other is zero point one), thereby effectively suppressing false events without excessively suppressing real events. The specific calculation process of the nonlinear confidence score formula is as follows: Summing the products of the weight parameters and membership values ​​of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call module, and voice call module yields the numerator. Summing the weight parameters of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call module, and voice call module again, and adding an evidence conflict penalty term yields the denominator. Dividing the numerator by the denominator gives the confidence score. When the signals from the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call module are consistent, the product of the membership complements approaches zero, making the denominator approach the sum of the weight parameters, and the confidence score approaches the weighted average. When the signals from the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call module are consistent, the confidence score approaches the weighted average. When signals from sensors, water immersion sensors, cameras, one-button emergency calls, and voice call modules conflict, the product of the complements of membership degrees increases significantly, leading to a significant increase in the denominator and a significant decrease in confidence. Specifically, signal consistency means that the membership values ​​of all participating sensors are greater than 0.7 or less than 0.3. In this case, the product of the complements of membership degrees is close to zero, the denominator is approximately equal to the sum of the weighting parameters, and the confidence is approximately equal to the weighted average. Signal conflict means that at least one sensor has a membership value greater than 0.7 while another sensor has a membership value less than 0.3. In this case, the product of the complements of membership degrees is significantly greater than zero, increasing the denominator and decreasing the confidence. This mechanism mimics the human decision-making logic of "seeing is believing, and listening to both sides leads to clarity," meaning that confidence is strengthened when multiple pieces of evidence point to the same conclusion, and weakened when evidence contradicts each other. For each triggered candidate event, record the event type label, confidence level, and trigger timestamp of the candidate event. These three elements together constitute an event entry. Organize all candidate event entries into an event list in chronological order. If multiple rules are triggered simultaneously under the same timestamp, all corresponding candidate events are added to the time list for conflict arbitration in step S3.

[0022] In this embodiment, it is specifically necessary to explain the following steps in step S3: the specific operation of arbitrating conflicts among multiple candidate events in the event list according to a preset hard priority level is as follows: The preset hard priority levels decrease sequentially in the order of fire-induced unconsciousness, calling for help, and falling, with fire-induced unconsciousness corresponding to the highest priority, calling for help to the second highest priority, and falling to the lowest priority; the specific assignment process is as follows: If the event type label of a candidate event is "fire and unconsciousness," it is assigned the value 3; if the event type label of a candidate event is "distress," it is assigned the value 2; if the event type label of a candidate event is "fall," it is assigned the value 1. Each candidate event in the event list is traversed, and a corresponding hard priority value is assigned based on its event type label. The candidate event with the highest hard priority value is selected to form a candidate subset. If the candidate subset contains only one candidate event, it is directly identified as the unique alarm event, and dynamic sorting is not performed. If the candidate subset contains multiple candidate events, dynamic sorting is performed. The hard priority values ​​of 3, 2, and 1 are based on a statistical analysis of the lethality of safety accidents involving elderly people living at home. The "fire and unconsciousness" event, involving both fire and unconsciousness, is assigned the highest priority value of 3. The "distress" event represents a distress signal actively issued by the elderly person, indicating that the elderly person is conscious but in danger, and is assigned the second highest priority value of 2. The "fall" event, although common, is assigned the lowest priority value of 1 because the elderly person usually has the ability to save themselves. This assignment system ensures that, with limited resources, events posing the greatest threat to life receive priority responses. In step S3, the specific operation of dynamically sorting multiple candidate events in the candidate subset using the dynamic priority scoring function is as follows: For each candidate event in the candidate subset, a dynamic priority score is calculated, as follows: First, the first factor is calculated by squaring the confidence scores of candidate events to amplify the advantage of high-confidence events. Squaring the confidence scores amplifies the gap between high-confidence and low-confidence events; for example, an event with a confidence score of 0.9 squares to 0.81, while an event with a confidence score of 0.5 squares to only 0.25, increasing the ratio from 1.8 to 3.24. This non-linear mapping makes the system more inclined to select high-confidence events, reducing the probability of low-confidence events being mistakenly selected due to random factors. The exponent for squaring is chosen to be two because experiments show that an exponent greater than two leads to low-confidence events being almost never selected, reducing the system's robustness; an exponent less than two results in insufficient discrimination. Next, the second factor is calculated. The second factor is the reciprocal of the time interval between the event and the current time (calculated by multiplying the exponent of the exponent function, with the exponent being negative zero-one, by the time interval between the trigger time of the candidate event), to give more weight to recent candidate events and to make the weight increase with time interval. The decay rate increases exponentially; the value of the exponential decay coefficient of 0.1 is determined as follows: when the time interval is 10 seconds, the value of the exponential decay term is approximately 0.37 (e to the power of negative 1), meaning that the weight of an event 10 seconds ago decays to about 37% of the current event; when the time interval is 30 seconds, the weight decays to approximately 0.05. This decay rate ensures that the system maintains high sensitivity to events within 30 seconds, while quickly reducing attention to earlier events, which meets the actual need for a response to home emergencies within tens of seconds. If the decay coefficient is too large (e.g., 0.5), the weight of an event 10 seconds ago drops to 0.067, which may cause earlier real emergencies to be ignored; if the decay coefficient is too small (e.g., 0.01), the weight of an event 1 minute ago is still 0.55, which cannot reflect the timeliness requirement. Then, the third factor is calculated by multiplying the historical recurrence coefficient corresponding to the event type label of the candidate event by 0.3; the result of multiplying the first and second factors is multiplied by 0.7, and then added to the third factor to obtain the dynamic priority score of the candidate event; for example... Figure 6The diagram illustrates the contribution ratios of the confidence squared term, time decay term, and historical recurrence term in different events. The values ​​of the balance factors 0.7 and 0.3 were determined through offline grid search optimization. On a validation set containing 300 multi-source alarm conflict scenarios, with the optimization objective of correctly selecting real emergency events, combinations of balance factors ranging from 0.5 to 0.9 with a step size of 0.05 were tested. The combination of 0.7 and 0.3 ultimately yielded the highest average success rate (97.5%). 0.7 was allocated to the confidence and timeliness terms, and 0.3 to the historical recurrence term, indicating that in most scenarios, the reliability and urgency of the current event are more important than historical experience. However, historical experience still plays a supplementary regulatory role, preventing the system from under-reacting to rare but dangerous events. The historical recurrence coefficient was updated using an exponential moving average. The formula is as follows: each time it is updated, 0.3 is multiplied by the flag value indicating whether the event type label of the most recent alarm event and the event type label of the current candidate event are consistent, and then 0.7 is added to the product of the old value of the historical recurrence coefficient. The result is then assigned to the historical recurrence coefficient. The learning rate of 0.3 is chosen based on the following: a larger learning rate (such as 0.5) will cause the historical recurrence coefficient to react too quickly to new events, making the system overly sensitive to accidental falls; a smaller learning rate (such as 0.1) will update slowly and will not be able to reflect changes in the elderly’s recent behavioral patterns in a timely manner; a learning rate of 0.3 allows the historical recurrence coefficient to rise from the initial value of 0.5 to about 0.8 after experiencing three similar events, which can quickly adapt to behavioral changes and filter out the influence of single accidental events; the initial value of the historical recurrence coefficient is set to 0.5, which represents an unbiased estimate of the prior probability of each event type. Finally, the candidate event with the highest dynamic priority score is selected as the sole alarm event. If multiple candidate events have the same and highest dynamic priority score, the candidate event with the earliest trigger timestamp is selected as the sole alarm event. The reason for selecting the candidate event with the earliest trigger timestamp as a tie-breaker rule is that an earlier event usually means that the problem has been ongoing for some time and needs to be dealt with first. If a later event is selected, earlier real emergencies may be ignored, which violates the basic principle of "first come, first served" in emergency response. This rule, together with the hard priority hierarchy, constitutes a complete conflict arbitration logic to ensure that an alarm event can be uniquely identified under any circumstances.

[0023] In this embodiment, it is specifically necessary to explain that in step S4, the process of extracting the corresponding spatial coordinates from the expanded multi-source fusion data frame based on the unique alarm event is as follows: First, based on the event type label of the unique alarm event, the sensor data most relevant to the unique alarm event is selected from the expanded multi-source fusion data frame. For fall events, the three-dimensional coordinates of the human centroid in the point cloud data of the millimeter-wave radar in the expanded multi-source fusion data frame are prioritized. For fire-related unconsciousness events, the installation location of the smoke sensor in the expanded multi-source fusion data frame is combined with the center coordinates of the smoke area in the video frame sequence of the camera. For distress calls, the sound source localization results of the voice call module in the expanded multi-source fusion data frame and the preset location of the one-button distress call are used. The sound source localization results are calculated by using the audio feature vector of the voice call module through a time difference of arrival algorithm. The algorithm calculates the time difference of arrival (TDOA) of the sound source relative to the microphone array by utilizing the time difference of arrival of the same sound signal received by multiple microphone arrays in the voice call module. Specifically, the voice call module is equipped with four microphones arranged in a square with a spacing of five centimeters. Cross-correlation analysis is performed on the cepstral coefficients of the Mel frequency in the audio feature vector to estimate the time difference of arrival between each microphone pair. Then, the azimuth and pitch angles of the sound source in the two-dimensional plane are solved by the triangulation method. Combined with the current position of the robot obtained by the self-localization algorithm of the care robot, the sound source orientation is converted into coordinates in the three-dimensional world coordinate system. The positioning error of this algorithm is less than 0.3 meters when the signal-to-noise ratio is greater than 10 decibels. Then, a multi-source weighted centroid localization algorithm is employed to fuse the position estimates provided by the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call, and voice call module involved in the localization process. The set of sensors involved in the localization is dynamically determined based on the event type label of the unique alarm event: for a fall event, the sensors involved include millimeter-wave radar and a camera; the millimeter-wave radar provides the three-dimensional coordinates of the human centroid, and the camera provides the coordinates of the foot contact point through key point detection. For a fire-related unconsciousness event, the sensors involved include a smoke sensor, a camera, and a millimeter-wave radar; the smoke sensor's installation location serves as the fire source reference point, the camera provides the coordinates of the center of the smoke area, and the millimeter-wave radar provides the static position of the human body. For an emergency call event, the sensors involved include a voice call module and a one-button emergency call; the voice call module provides the sound source localization coordinates, and the one-button emergency call provides the coordinates of a preset fixed position. The position estimates of each sensor are expressed in a three-dimensional world coordinate system, with the unit being meters. The weights of the sensor, camera, one-button emergency call, and voice call module are taken from the membership values ​​of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call, and voice call module under a unique alarm event in step S2. A set of sensors participating in the positioning is constructed. For each sensor in the set, its corresponding weight is multiplied by the provided location coordinate components and summed. Then, the sum of the summed components is divided by the sum of all weights to obtain the final spatial coordinates. The specific calculation process is as follows: First, obtain the sum of the products of each sensor weight participating in the positioning and its corresponding horizontal coordinate, and divide the sum by the sum of the weights of all sensors participating in the positioning to calculate the horizontal coordinate of the final spatial coordinates. Next, obtain the sum of the products of each sensor weight participating in the positioning and its corresponding vertical coordinate, and divide the sum by the sum of the weights of all sensors participating in the positioning to calculate the vertical coordinate of the final spatial coordinates. Finally, obtain the sum of the products of each sensor weight participating in the positioning and its corresponding vertical coordinate, and divide the sum by the sum of the weights of all sensors participating in the positioning to calculate the vertical coordinate of the final spatial coordinates. The weights reflect the reliability of millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-click emergency button and voice call module in the current event determination. The larger the weight, the higher the proportion of the sensor's position estimation in the final coordinate calculation. Next, the event type label, confidence level, trigger timestamp, and calculated final spatial coordinates of the unique alarm event are packaged into a structured alarm message. The structured alarm message is organized in key-value pair format. The specific fields of the structured alarm message include: an event type field, with values ​​of fire-induced unconsciousness, distress call, or fall; a confidence level field, with a value between zero and one calculated in step S2; a trigger time field, with a trigger timestamp recorded in step S2, in the format of year, month, day, hour, minute, second, and millisecond; a spatial coordinate field, containing the room number and three-dimensional coordinates. The room number is automatically obtained by the indoor positioning system of the care robot based on the spatial coordinates and matched with pre-stored map data. The three-dimensional coordinates are the final spatial coordinates calculated in step S4; and an event summary field, a natural language description automatically generated by the system based on the event type and sensor data, such as "Smoke was detected in the kitchen area and the person is stationary, indicating a risk of fire-induced unconsciousness." The total message length does not exceed one thousand bytes and is encoded in JSON format. Finally, a tiered push strategy is implemented based on the event type label of the unique alarm event: if the event type label is "fire and unconsciousness," it is simultaneously pushed to the fire department, emergency center, family terminal, and community monitoring center. Push channels include SMS, telephone voice calls, and application push notifications. SMS push notifications are sent through a third-party SMS gateway, containing the event type and room number, with a maximum of seventy Chinese characters. Telephone voice calls convert the event summary in the structured alarm message into speech using a text-to-speech engine, automatically dial a preset number via telephone line, and play the voice message. Each call is repeated a maximum of three times, with a thirty-second interval between each call. Application push notifications... The push notifications are sent via the messaging channel of the mobile application accompanying the care robot. The content includes all fields of a structured alarm message and a map location link. All push channels are equipped with a timeout retry mechanism. If no alarm confirmation signal is received from the responder within three minutes, the push level is automatically escalated. For example, a fall event is escalated to be pushed to the community monitoring center simultaneously. If the event type is labeled as a distress call, the notification is sent to both the family member's terminal and the community monitoring center via SMS and application push. If the event type is labeled as a fall, the notification is sent only to the family member's terminal via application push. All pushes carry the generated structured alarm message. The processing results after the push are fed back to update the weight parameters in the event confidence calculation formula and the historical recurrence coefficient in the dynamic priority scoring function, specifically: The system monitors feedback signals returned by the responders. These signals are binary variables, where a value of 1 indicates a confirmed event, and a value of 0 indicates a false alarm. For the weight parameters in the event confidence calculation formula, a confidence propagation update rule is used. The new value of the weight parameter is set to the old value plus 0.1 times the product of the feedback signal and the historical average feedback rate, then multiplied by the membership values ​​of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call, and voice call module in the unique alarm event. The learning rate of 0.1 times controls the single adjustment range of the weight parameter. If the learning rate is too large (e.g., 0.5), a single feedback may cause drastic fluctuations in the weight, reducing system stability. If the learning rate is too small (e.g., 0.01), the weight convergence is slow, failing to respond promptly to changes in sensor performance. A learning rate of 0.1 times allows the weight parameter to adjust from its initial value to near the target value after approximately ten similar feedbacks, balancing response speed and stability. The membership value acts as a multiplier, ensuring smooth event determination. Sensors contributing more to the signal receive larger weight adjustments, while those contributing less receive smaller adjustments. This accelerates the weight increase of effective sensors and suppresses the weight growth of ineffective sensors. The initial value of the historical average feedback rate is 0.5, used to smooth short-term fluctuations and ensure the stability of weight adjustments. The exponential moving average update method for the historical average feedback rate is as follows: after each feedback signal is received, the new value of the historical average feedback rate is set to 0.3 multiplied by the current feedback signal plus 0.7 multiplied by the old value of the historical average feedback rate. This update method allows the historical average feedback rate to smoothly track recent feedback trends. For example, if five consecutive real event feedbacks are received, the historical average feedback rate will rise from 0.5 to approximately 0.86; if five consecutive false alarm feedbacks are received, it will drop to approximately 0.14. The historical average feedback rate is used in the difference calculation in the weight parameter update formula. When the difference between the feedback signal and the historical average feedback rate is positive, the weight increases; conversely, the weight decreases, thus achieving adaptive weight adjustment. Figure 7 As shown, this illustrates the dynamic adjustment process of different sensor weights as the feedback signal changes; For the historical recurrence coefficient in the dynamic priority scoring function, an exponential moving average method is used. The exponential moving average update of the historical recurrence coefficient is only performed after a feedback signal is received. If the feedback signal is a value of one (confirmed as a real event), the new value of the historical recurrence coefficient increases, resulting in a higher third factor score for that event type in future dynamic priority scoring. If the feedback signal is a value of zero (false alarm), the new value of the historical recurrence coefficient decreases, causing a drop in the priority score for that event type. This mechanism ensures that only real event types that are repeatedly confirmed will receive a higher historical recurrence coefficient, thus gaining an advantage in dynamic priority scoring, while frequently false alarm event types will be gradually given less attention by the system. The historical recurrence coefficient is limited to a range of zero to one, with an initial value of 0.5. After each update, if the calculated result is less than zero, it is forcibly reset to zero; if it is greater than one, it is forcibly reset to one. The new value of the historical recurrence coefficient is set to the product of 0.3 and the feedback signal plus the product of 0.7 and the old value of the historical recurrence coefficient. This ensures that only event types confirmed as real events will see an increase in their corresponding historical recurrence coefficient; false alarm events will not see an increase in their historical recurrence coefficient. Figure 8 As shown, the historical recurrence coefficients of three event types—fire-related coma, distress calls, and falls—are displayed as a function of the number of responses.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for tracing and classifying multi-source alarms in home care robots, characterized in that, Specifically, the steps include the following: Step S1: Establish a unified timestamp reference and spatial coordinate system to spatiotemporally align and fuse point cloud data from millimeter-wave radar, analog values ​​from smoke sensors, gas sensors and water immersion sensors, video frame sequences from cameras, and Boolean signals from one-click emergency button and voice call module to generate extended multi-source fusion data frames. Step S2: Based on the expanded multi-source fusion data frame, the preset cross-validation rule engine is called to identify candidate events, and the confidence of each candidate event is calculated using the event confidence calculation formula. The event confidence calculation formula includes weight parameters for millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-click emergency button and voice call module, thereby outputting an event list containing event type label, confidence and trigger timestamp. Step S3: Arbitrate conflicts among multiple candidate events in the event list according to the preset hard priority level to obtain a candidate subset. Then, dynamically sort the multiple candidate events in the candidate subset using a dynamic priority scoring function. The dynamic priority scoring function includes confidence, the reciprocal of the time interval between the event and the current time, and the historical recurrence coefficient, thereby determining the unique alarm event. Step S4: Extract the corresponding spatial coordinates from the expanded multi-source fusion data frame based on the unique alarm event to generate a structured alarm message containing event type label, confidence level, trigger timestamp and spatial coordinates. Push the structured alarm message to different responders according to priority level. At the same time, feed back the processing results after the push to update the weight parameters in the event confidence calculation formula and the historical recurrence coefficient in the dynamic priority scoring function.

2. The method for tracing and classifying multi-source alarms in home care robots according to claim 1, characterized in that: In step S1, the main controller clock of the care robot is used as the global time reference to add a global synchronization timestamp to each frame of data in the millimeter-wave radar point cloud data, the analog values ​​of the smoke sensor, the gas sensor and the water immersion sensor, the video frame sequence of the camera, and the Boolean signal of the one-click emergency button and the voice call module. Then, an asynchronous alignment algorithm based on maximizing mutual information of time offset is used to unify the point cloud data of non-uniformly sampled millimeter-wave radar, analog values ​​of smoke sensors, gas sensors, and water immersion sensors, video frame sequences of cameras, and Boolean signals of one-button emergency call and voice call modules into a fixed time grid. For any sensor, the reconstructed value at the alignment time is obtained through the following process: First, determine the output data values ​​of any one of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call and voice call module at all original sampling times within the alignment time neighborhood; Secondly, the time difference between each original sampling moment and the alignment moment is calculated, and the corresponding adaptive kernel width is adaptively adjusted according to the sampling interval fluctuation variance of any one of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call button, and voice call module within the most recent sliding window. Next, based on the time difference and the adaptive kernel width, an exponentially weighted weight is calculated using an exponential function, where the exponential function is the ratio of the square of the time difference with a negative exponent (base of the natural constant) to the square of twice the adaptive kernel width. Finally, the calculated weights are normalized, and the normalized weights are multiplied by the output data value of the corresponding original sampling moment and summed to obtain the reconstructed value of any one of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call button, and voice call module at the alignment moment. This generates the time-aligned point cloud data of the millimeter-wave radar, the analog values ​​of the smoke sensor, gas sensor, and water immersion sensor, the video frame sequence of the camera, and the Boolean signals of the one-button emergency call button and the voice call module.

3. The method for tracing and classifying multi-source alarms in home care robots according to claim 2, characterized in that: The specific process for generating the extended multi-source fused data frame is as follows: All time-aligned millimeter-wave radar point cloud data, analog values ​​from smoke sensors, gas sensors, and water immersion sensors, video frame sequences from cameras, and Boolean signals from the one-button emergency call and voice call modules are mapped to a three-dimensional world coordinate system with the geometric center of the care robot chassis as the origin and due north as the Y-axis. For millimeter-wave radar, its polar coordinate point cloud data is converted to Cartesian coordinate form. For cameras, the pixel coordinates in the video frame sequence are back-projected to three-dimensional spatial points using pre-calibrated intrinsic and extrinsic rotation and translation matrices. Through factory calibration and real-time updates of the care robot's self-localization algorithm, the points in the local coordinate system of each sensor are transformed to points in the three-dimensional world coordinate system. The spatial transformation matrix is ​​implemented by a homogeneous transformation matrix composed of rotation matrices and translation vectors. Then, the point cloud data of all millimeter-wave radars, analog values ​​of smoke sensors, gas sensors and water immersion sensors, video frame sequences of cameras, and Boolean signals of the one-button emergency call button and voice call module, after time and spatial alignment, are organized into a structured multi-source fusion data frame by time index. This multi-source fusion data frame consists of timestamps, point cloud data of millimeter-wave radars, analog values ​​of smoke sensors, analog values ​​of gas sensors and water immersion sensors, video frame sequences of cameras, Boolean signals of the one-button emergency call button, and audio feature vectors captured by the voice call module in the time grid neighborhood. Finally, tensor decomposition and fusion are performed on the point cloud data of the millimeter-wave radar, the analog values ​​of the smoke sensor, the analog values ​​of the gas sensor, the analog values ​​of the water immersion sensor, the pixel values ​​of the video frame sequence of the camera, and the audio feature vector of the voice call module in the multi-source fusion data frame to form an expanded multi-source fusion data frame. The expanded multi-source fusion data frame consists of a timestamp, the point cloud data of the millimeter-wave radar, the analog values ​​of the smoke sensor, the analog values ​​of the gas sensor, the analog values ​​of the water immersion sensor, the video frame sequence of the camera, the Boolean signal of the one-button emergency call, the audio feature vector of the voice call module, and the fused feature vector. The point cloud data of the millimeter-wave radar contains multiple points, each containing three-dimensional coordinates and Doppler velocity.

4. The method for tracing and classifying multi-source alarms in home care robots according to claim 3, characterized in that: The tensor decomposition and fusion process is as follows: The point cloud data of millimeter-wave radar, the analog values ​​of smoke sensor, gas sensor, and water immersion sensor, the pixel values ​​of video frame sequences from the camera, and the audio feature vector of the voice call module in the multi-source fusion data frame are organized into a third-order tensor. The alternating least squares method is used to iteratively solve the approximate decomposition of the tensor with a rank of three. The weight coefficients of the first three principal components are extracted as the fusion feature vector, and the fusion feature vector is appended to the multi-source fusion data frame to form the expanded multi-source fusion data frame.

5. A method for tracing and classifying multi-source alarms in home care robots according to claim 4, characterized in that: In step S2, the specific operation of calling the preset cross-validation rule engine to identify candidate events is as follows: Three rules are preset. The first rule is used to determine a fall candidate event when the centroid vertical velocity extracted from the point cloud data of the millimeter-wave radar in the expanded multi-source fusion data frame is greater than 1.5 m / s, the analog value of the smoke sensor in the expanded multi-source fusion data frame is less than 500 million parts per second, and the smoke probability output by the smoke detection model of the video frame sequence of the camera in the expanded multi-source fusion data frame is less than 0.

3. The second rule is used to determine a fire coma candidate event when the analog value of the smoke sensor in the extended multi-source fusion data frame is greater than or equal to 500 million parts per second, and the displacement of the human centroid extracted from the point cloud data of the millimeter-wave radar in the extended multi-source fusion data frame is less than 0.1 meters for one consecutive second. The third rule is used to determine a candidate event for distress when the Boolean signal of the one-click distress button in the extended multi-source fusion data frame is true or when the audio feature vector of the voice call module in the extended multi-source fusion data frame is matched with a preset distress word by the keyword recognition model.

6. The method for tracing and classifying multi-source alarms in home care robots according to claim 5, characterized in that: The specific process of calculating the confidence score for each candidate event using the event confidence score calculation formula is as follows: For each triggered candidate event, the membership function of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-click emergency button and voice call module participating in the candidate event determination in the extended multi-source fusion data frame is defined. The membership function uses a nonlinear transformation to map the sensor observation values ​​to a membership value between zero and one. Subsequently, based on the membership values ​​of millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-click emergency button and voice call module, as well as the corresponding weight parameters, the confidence of candidate events is calculated using a nonlinear confidence calculation formula that introduces a conflict of evidence penalty term. The evidence conflict penalty item is the product of the membership complements of millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call and voice call module, multiplied by the preset conflict penalty coefficient; The specific calculation process of the nonlinear confidence calculation formula is as follows: sum the products of the weight parameters of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call and voice call module with the membership values ​​to obtain the numerator; sum the weight parameters of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call and voice call module and add the evidence conflict penalty term to obtain the denominator; then divide the numerator by the denominator to obtain the confidence score. For each triggered candidate event, record the event type label, confidence level, and trigger timestamp of the candidate event. These three elements together constitute an event entry. Organize all candidate event entries into an event list in chronological order. If multiple rules are triggered simultaneously under the same timestamp, all corresponding candidate events are added to the time list.

7. A method for tracing and classifying multi-source alarms for home care robots according to claim 6, characterized in that: In step S3, the specific operation of arbitrating conflicts among multiple candidate events in the event list according to a preset hard priority level is as follows: The preset hard priority levels decrease sequentially in the order of fire-induced unconsciousness, calling for help, and falling, with fire-induced unconsciousness corresponding to the highest priority, calling for help to the second highest priority, and falling to the lowest priority; the specific assignment process is as follows: If the event type label of a candidate event is fire and unconsciousness, assign a value of three; if the event type label of a candidate event is distress call, assign a value of two; if the event type label of a candidate event is fall, assign a value of one; iterate through each candidate event in the event list, assign a corresponding hard priority value according to the event type label of the candidate event, and select the candidate event with the highest hard priority value to form a candidate subset. If the candidate subset contains only one candidate event, the candidate event is directly identified as the unique alarm event, and dynamic sorting is not performed; if the candidate subset contains multiple candidate events, dynamic sorting is performed.

8. A method for tracing and classifying multi-source alarms in home care robots according to claim 7, characterized in that: In step S3, the specific operation of dynamically sorting multiple candidate events in the candidate subset using a dynamic priority scoring function is as follows: For each candidate event in the candidate subset, a dynamic priority score is calculated, as follows: First, calculate the first factor by squaring the confidence level of the candidate event; second, calculate the second factor, which is the reciprocal of the time interval between the event and the current time; then calculate the third factor by multiplying the historical recurrence coefficient corresponding to the event type label of the candidate event by 0.

3. Multiply the result of multiplying the first factor by the second factor by 0.7, and then add it to the third factor to obtain the dynamic priority score of the candidate event. Finally, the candidate event with the highest dynamic priority score is selected as the sole alarm event. If multiple candidate events have the same dynamic priority score and are all the highest, the candidate event with the earliest trigger timestamp is selected as the sole alarm event.

9. A method for tracing and classifying multi-source alarms in home care robots according to claim 8, characterized in that: In step S4, the process of extracting the corresponding spatial coordinates from the expanded multi-source fusion data frame based on the unique alarm event is as follows: First, based on the event type label of the unique alarm event, the sensor data most relevant to the unique alarm event is selected from the expanded multi-source fusion data frame. For fall events, the three-dimensional coordinates of the human body's center of mass in the point cloud data of the millimeter-wave radar in the expanded multi-source fusion data frame are used first. For fire-related unconsciousness events, the installation location of the smoke sensor in the expanded multi-source fusion data frame is combined with the center coordinates of the smoke area in the video frame sequence of the camera. For distress events, the sound source localization result of the voice call module in the expanded multi-source fusion data frame and the preset location of the one-button distress button are used. The sound source localization result is calculated by the audio feature vector of the voice call module using the time difference of arrival algorithm. Then, a multi-source weighted centroid localization algorithm is used to fuse the position estimates provided by the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call and voice call module involved in the localization. A set of sensors involved in the localization is constructed. For each sensor in the set, the corresponding weight is multiplied by the provided position coordinate component and accumulated. Then, the sum of the accumulated components is divided by the sum of all weights to obtain the final spatial coordinates. Next, the event type label, confidence level, trigger timestamp, and calculated spatial coordinates of the unique alarm event are packaged into a structured alarm message; Finally, a tiered push strategy is implemented based on the event type label of the unique alarm event: if the event type label is "fire and unconsciousness," it is pushed to the fire department, emergency center, family terminal, and community monitoring center simultaneously, with push channels including SMS, telephone voice call, and application push; if the event type label is "distress," it is pushed to the family terminal and community monitoring center, with push channels including SMS and application push; if the event type label is "fall," it is pushed only to the family terminal, with push channel including application push. All pushes carry the generated structured alarm message.

10. A method for tracing and classifying multi-source alarms in home care robots according to claim 9, characterized in that: The step of feeding back the processing results after the push to update the weight parameters in the event confidence calculation formula and the historical recurrence coefficient in the dynamic priority scoring function is as follows: The system monitors the feedback signal returned by the responder. The feedback signal is a binary variable, where a value of 1 indicates a confirmed real event, and a value of 0 indicates a false alarm. For the weight parameter in the event confidence calculation formula, a confidence propagation update rule is adopted. The new value of the weight parameter is set to the old value of the weight parameter plus 0.1 times the product of the difference between the feedback signal and the historical average feedback rate, and then multiplied by the membership value of the millimeter-wave radar, smoke sensor, gas sensor, water immersion sensor, camera, one-button emergency call, and voice call module in the unique alarm event. The initial value of the historical average feedback rate is 0.

5. For the historical recurrence coefficient in the dynamic priority scoring function, the exponential moving average method is used, and the new value of the historical recurrence coefficient is set to the product of 0.3 and the feedback signal plus the product of 0.7 and the old value of the historical recurrence coefficient.

Images