Old people indoor monitoring system based on end-side artificial intelligence

Abstract

The invention discloses an indoor monitoring system for old people based on end-side artificial intelligence, and relates to the technical field of home monitoring. For the indoor risk faced by the current elderly living alone, the adopted scheme comprises a radar group, an AI terminal device, a local large model, a 5G communication module and an alarm module, and the radar group continuously monitors the indoor environment, collects the activity data of the elderly and transmits the activity data to the AI terminal device in real time; the AI terminal equipment is used as a core hub, interacts with a user through multiple modes, and processes and stores radar data; the local large model performs inference analysis on the data by means of an AI terminal, identifies an indoor activity mode of the old people, and judges whether the activity intensity is normal or not and whether abnormal behaviors occur or not; once abnormality is detected, the 5G communication module sends a short message containing abnormal information to a guardian, and the alarm module starts a loudspeaker to play preset alarm information. The system is used for monitoring indoor activities so as to quickly give an alarm when an abnormality occurs.

Description

Technical Field

[0001] The present invention relates to the field of home monitoring technology, and in particular to an indoor monitoring system for the elderly based on terminal-side artificial intelligence. Background Art

[0002] End-side AI refers to applications that run and process artificial intelligence algorithms directly on the terminal device. The terminal device processes data locally without sending the data to the cloud or server for processing. Running artificial intelligence applications on the end-side device helps protect user privacy and data security; reduces the terminal device's dependence on network connections; and reduces the cost of data transmission and cloud computing, thereby improving overall cost-effectiveness.

[0003] With the slowdown of population growth and the acceleration of population aging, the proportion of the elderly population in my country is increasing, especially in large cities. As the parents of the first batch of only children gradually enter old age, and the only children enter middle age, they are faced with the problem of supporting four elderly people and raising one child, and the family burden is getting heavier. In today's fast-paced lifestyle, the living habits and schedules of children or other family members are very different from those of the elderly, which has also made more elderly people live alone, facing many inconveniences or difficulties in home-based elderly care, and even safety risks. In order to resolve the safety risks of home-based elderly care for the elderly and better meet the basic elderly care service needs, it is particularly important to use intelligent systems to achieve real-time dynamic health monitoring of the elderly living alone. Summary of the invention

[0004] In view of the needs and shortcomings of current technological development, the present invention provides an indoor monitoring system for the elderly based on terminal-side artificial intelligence.

[0005] The present invention provides an indoor monitoring system for the elderly based on terminal-side artificial intelligence, and the technical solution adopted to solve the above technical problems is as follows:

[0006] An indoor monitoring system for the elderly based on end-side artificial intelligence, which includes five parts: radar group, AI terminal device, local large model, 5G communication module and alarm module, among which:

[0007] Radar group, including multiple radar sensors arranged indoors, is used to continuously monitor the indoor space, collect activity data of the elderly, and transmit the collected data to AI terminal devices in real time, providing basic data support for subsequent analysis and judgment;

[0008] The AI ​​terminal device, as the central hub of the system, is used to interact with users in a multimodal manner, collect and process the transmission data of the radar group, and store this data for subsequent analysis;

[0009] The local large model is used to achieve multimodal interaction with users through AI terminal devices, and use model algorithms to infer and analyze radar group data to identify the indoor activity patterns of the elderly, determine whether the activity intensity is normal, and whether abnormal behavior occurs;

[0010] The 5G communication module is integrated in the AI ​​terminal device. When the AI ​​terminal device combines the local large model to analyze and conclude that the elderly have abnormal behavior, it sends a text message containing abnormal information to the guardian's mobile phone according to the preset operation, ensuring that the guardian can be informed of the elderly's situation in a timely manner;

[0011] The alarm module is used to receive instructions from the AI ​​terminal device when the AI ​​terminal device combines the local large model to analyze and conclude that the elderly have abnormal behavior, and activate the speaker to play a pre-set alarm message to attract the attention of people around, prompting them to check the elderly’s condition in time and provide necessary help.

[0012] Optionally, a radar group including multiple radar sensors is arranged indoors by integrating the global search capability of the genetic algorithm and the local search capability of the gray wolf algorithm, wherein:

[0013] Utilize the global search capability of genetic algorithms: Starting from the initial population, through genetic operations such as selection, crossover, and mutation, conduct extensive searches in the entire solution space to find areas that may contain the optimal solution, so as to explore different radar group layout schemes within the preset range and obtain multiple optimal layouts;

[0014] Combined with the local search capability of the Gray Wolf Algorithm: In the area that may contain the optimal solution found by the genetic algorithm, the local search advantage of the Gray Wolf Algorithm is used to search in the local area, continuously approaching the local optimal solution, so that the radar group layout plan can be optimized locally and the accuracy of the layout can be improved.

[0015] Further optionally, a radar group including multiple radar sensors is arranged indoors by fusing the global search capability of the genetic algorithm and the local search capability of the gray wolf algorithm. The specific fusing process is as follows:

[0016] (1) Initialize the radar sensor layout, evenly arrange multiple radar sensors in the working area and introduce vectors to divide the working area. Measure the activity intensity of the elderly within the monitoring range of each radar sensor, and create an initial weight matrix that integrates multiple factors and covers the time dimension extension, providing a quantitative basis for subsequent layout optimization.

[0017] (2) The average impact value of the radar sensor is calculated to evaluate its sensitivity. A sparse matrix related to distance is constructed based on the evaluation index of comprehensive distance and attenuation effect. After setting the threshold and comparing the average impact value, a sparse weight matrix that retains high-impact weight elements is constructed to provide input data for the fusion of genetic and gray wolf algorithms.

[0018] (3) Using the global search capability of the genetic algorithm, a wide search is conducted in the entire solution space to find an area that may contain the optimal solution. Then, combined with the local search capability of the gray wolf algorithm, a search is conducted in the local area to continuously approach the local optimal solution, so that the radar group layout plan is locally optimized. The above two operations are continuously iterated to gradually find a better radar group layout plan until the preset termination condition is met.

[0019] (4) The algorithm is evaluated using a performance evaluation equation containing multiple factors. After setting the threshold, it is started and radar sensors with efficiency or coverage below the threshold are eliminated. The configuration with the largest weight in each row is selected as the recommended configuration. The threshold is fine-tuned according to the evaluation value to reduce the number of active radar sensors and ensure accuracy.

[0020] Optionally, step (1) specifically includes:

[0021] (1.1) h radar sensors are evenly placed in multiple working areas, where the working area refers to the indoor activity area of ​​the elderly. The working area is divided into a grid shape, and the working area is divided by introducing a vector P to detect and record the activity data of the elderly. The vector P is described as:

[0022]

[0023] In the formula, is a vector element, representing the h radar sensors in the t-th working area, each radar sensor covers at least one grid cell;

[0024] (1.2) Measure the activity intensity of the elderly within the monitoring range of each radar sensor, and create an initial weight matrix based on these data. The weight calculation method is:

[0025]

[0026] In the formula, Indicates that with distance d ij Increase the exponential decay term of signal attenuation; d ij represents the distance from the ith radar sensor to the target elder; α represents the attenuation coefficient that controls the influence of distance on the attenuation rate; h represents the total number of radar sensors; represents the normalization factor, which is the sum of the attenuation of the ith radar sensor relative to all targets and ensures that the sum of the weights is 1; the weight W ij It represents the influence weight of the j-th activity on the i-th sensor. The calculation takes into account the total number of activity frequencies of the elderly and expands it in the time dimension.

[0027] Optionally, step (2) specifically includes:

[0028] (2.1) Assuming that the changes in radar sensor values ​​caused by the elderly’s activities are recorded in s test cases, the average impact value V of the i-th radar sensor is i Calculated by the following formula:

[0029]

[0030] In the formula, ΔP i,k represents the value change of the i-th radar sensor in the k-th test case due to the activities of the elderly; d i,k represents the distance between the i-th radar sensor and the elderly's activity location; α is the attenuation coefficient, which is used to adjust the effect of distance on the response sensitivity of the radar sensor;

[0031] (2.2) Define an evaluation index that combines the distance between the elderly and the radar sensor and the attenuation effect α of the distance to measure the sensitivity of the radar sensor in monitoring the activities of the elderly, and further construct a sparse matrix W * , the sparse matrix W * Each element of It reflects the sensitivity of the i-th radar sensor to monitor the activity location of the j-th elderly person. The sparse matrix W * The formula is as follows:

[0032]

[0033] Where, d ij is the distance between the jth active position and the ith radar sensor; α is the attenuation coefficient, which is used to adjust the effect of distance on the response sensitivity of the radar sensor;

[0034] (2.3) Set the threshold L and the average impact value V of the radar sensor i When it is less than L, the sparse matrix W * The corresponding weights in are 0 or deleted from the matrix. When the average impact value Vi of the radar sensor is greater than L, the sparse matrix W is retained. * The corresponding weights in are:

[0035]

[0036] Sparse weight matrix W * The weight elements retained in represent the radar sensors with high influence on the system and serve as the input data for the fusion genetic algorithm and the grey wolf algorithm.

[0037] Optionally, step (3) specifically includes:

[0038] (3.1) Determine the initialization stage, including setting the population size, crossover rate, mutation rate, and generating the initial population;

[0039] (3.2) Entering the iterative calculation phase, each iteration includes genetic algorithm selection, crossover, mutation operations and gray wolf position update;

[0040] (3.3) The genetic algorithm updates the population by the following method:

[0041] x ′ =crossover(mutate(select(Pop),p m ),p c ),

[0042] Among them, select, crossover and mutate represent selection, crossover and mutation operations respectively, p m and p c represent mutation and crossover probabilities respectively;

[0043] (3.4) The gray wolf algorithm adjusts the positions of the remaining individuals according to the position of the best individual in the current population.

[0044] X new =X alpha -A*D+C*(X beta -X gamma ),

[0045] Where X alpha , X beta and X gamma are the locations of the best, second best, and third best sensors, respectively. A and C are the coefficients that control the search intensity and randomness. D is the vector of the distance between the radar sensor and the elderly’s motion.

[0046] (3.5) Through the iterative optimization process of steps (3.2)-(3.4), iterate until the termination condition is met, that is, the change of the objective function ΔF = |F(xnew)-F(xold)| between two consecutive generations is less than the threshold L, or the maximum number of iterations N is reached max , where the objective function is expressed as follows:

[0047] F(x)=α*Coverage(x)+β*Efficiency(x)+γ*Sensitivity(x),

[0048] Where x represents a radar sensor layout scheme, Coverage(x) represents the coverage of the layout scheme, Efficiency(x) represents the monitoring efficiency, Sensitivity(x) represents the sensitivity of the radar sensor, α, β, and γ are weighted coefficients used to adjust the relative importance of different performance indicators in the objective function. The specific values ​​need to be adjusted according to different application scenarios and priorities.

[0049] Optionally, step (4) specifically includes:

[0050] (4.1) Use the following performance evaluation equation to evaluate the effectiveness of the algorithm:

[0051]

[0052] Where N represents the total number of monitored areas, each area is covered by at least one radar sensor, x represents a specific radar sensor layout scheme, C(i,x) represents the coverage of the radar sensor in the i-th area, E(i,x) represents the efficiency of the radar sensor in the i-th area, S(i,x) represents the sensitivity of the radar sensor in the i-th area, λ1, λ2, and λ3 are weighted coefficients representing the relative importance of coverage, sensitivity, and efficiency, respectively; when the goals and requirements of different stages are different, the selected parameters will be adjusted according to the actual situation;

[0053] (4.2) After setting the threshold L = 1, the algorithm is started to remove radar sensors whose efficiency or coverage is lower than the threshold. If the weight matrix W has N rows, N radar sensors are finally selected. The radar sensors finally retained are the sum of the radar sensors with the largest weight in each row. The number and position of radar sensors obtained at this time are the recommended configuration;

[0054] (4.3) Calculate the performance evaluation value of the recommended configuration. If the evaluation value is not 0, continue to fine-tune the threshold until the performance requirements are met.

[0055] Optionally, the local large model involved realizes multimodal interaction with the user through the AI ​​terminal device, and uses the model algorithm to infer and analyze the radar group data, so as to identify the indoor activity pattern of the elderly, determine whether the activity intensity is normal, and whether abnormal behavior occurs. This process specifically includes:

[0056] Data reception and integration: AI terminal devices collect real-time data and multimodal data from the radar group, integrate the multimodal data, and prepare for subsequent unified analysis;

[0057] Data preprocessing: Use data cleaning algorithms to remove outliers and fill missing values, and extract key features from the integrated multimodal data;

[0058] Model reasoning: ① The local large model is pre-trained on a large amount of relevant data to learn the characteristics of various activity patterns, intensity ranges, and abnormal behaviors; when receiving the pre-processed data, the AI ​​terminal device loads the local large model and adapts the data to the input format of the local large model; ② The large model uses convolutional neural networks or recurrent neural networks and their variants in deep neural networks to perform feature analysis on radar data and identify any activity mode of walking, standing, sitting, and lying down; ③ Based on the extracted key features, the local large model predicts the activity intensity through a regression algorithm, and then compares it with the pre-set normal activity intensity range to determine whether the current activity intensity is normal; ④ The anomaly detection algorithm is used to compare the current activity pattern and activity intensity data with the normal pattern learned by the local large model. If the data point is outside the decision boundary of the local large model, or the deviation from the normal pattern exceeds the set threshold, it is determined to be an abnormal behavior;

[0059] Result output: The local large model generates a detailed report based on the results of the reasoning analysis, including the identified activity patterns, activity intensity assessment, and information on whether abnormal behavior exists. The report is presented in a structured form to facilitate further processing by the AI ​​terminal device.

[0060] Compared with the prior art, the indoor monitoring system for the elderly based on terminal-side artificial intelligence of the present invention has the following beneficial effects:

[0061] 1. Through the coordinated work of various parts, the present invention realizes all-round, real-time monitoring and efficient emergency treatment of the indoor activities of the elderly, providing a strong guarantee for the safety of the elderly;

[0062] 2. The present invention continuously monitors the indoor space through the radar group arranged indoors, collects the activity data of the elderly in real time, and transmits it to the AI ​​terminal device, which ensures that the system can obtain accurate and up-to-date information on the activities of the elderly, and provides a solid data foundation for subsequent analysis and judgment, such as accurately capturing the movement trajectory, activity frequency and other details of the elderly;

[0063] 3. The present invention uses AI terminal devices as the central hub. On the one hand, it interacts with users in a multimodal manner to facilitate the use of the elderly and guardians. On the other hand, it collects, processes and stores the data transmitted by the radar group. This not only improves the user experience, but also enables the orderly integration of data, providing convenience for the analysis of local large models. For example, the elderly can interact with the AI ​​terminal device through voice to query their own health information, and the AI ​​terminal device can quickly process and feedback;

[0064] 4. The local large model of the present invention uses AI terminal devices and model algorithms to perform reasoning and analysis on radar group data, accurately identify the activity patterns of the elderly, and determine whether the activity intensity is normal and whether abnormal behavior occurs. This intelligent analysis capability can detect potential risks in a timely manner, such as detecting abnormal behaviors of the elderly in advance, such as falling, being still for a long time, etc., to buy time for timely intervention;

[0065] 5. The 5G communication module of the present invention is integrated into the AI ​​terminal device. When the system determines that the elderly have abnormal behavior, it can send a text message containing abnormal information to the guardian's mobile phone according to the preset operation. By using the high-speed characteristics of the 5G network, the guardian can be informed of the elderly's situation in time, so as to take remote measures, such as contacting medical staff or notifying neighbors to check;

[0066] 6. When the alarm module of the present invention detects abnormal behavior of the elderly, it receives instructions from the AI ​​terminal device and starts the speaker to play the alarm information, which can quickly attract the attention of people around and prompt them to check the elderly's situation in time and provide help. This can provide timely on-site support for the elderly at critical moments and increase the safety of the elderly. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Attached Figure 1 is a system structure diagram of Embodiment 1 of the present invention;

[0068] Attached Figure 2 is a flow chart of arranging multiple radar sensors indoors in Embodiment 1 of the present invention;

[0069] Attached Figure 3 is an initial layout diagram of multiple radar sensors arranged indoors in the first embodiment of the present invention;

[0070] Attached Figure 4 This is a display diagram of the frequent activity areas of the elderly detected by multiple radar sensors in Example 1 of the present invention. DETAILED DESCRIPTION

[0071] In order to make the technical solution, the technical problem solved and the technical effect of the present invention more clearly understood, the technical solution of the present invention is clearly and completely described below in conjunction with specific embodiments.

[0072] Embodiment 1:

[0073] Combined with Figure 1 This embodiment proposes an indoor monitoring system for the elderly based on terminal-side artificial intelligence, which includes five parts: a radar group, an AI terminal device, a local large model, a 5G communication module, and an alarm module, wherein:

[0074] The radar group includes multiple radar sensors arranged indoors, which are used to continuously monitor the indoor space, collect activity data of the elderly, and transmit the collected data to the AI ​​terminal device in real time to provide basic data support for subsequent analysis and judgment. When the millimeter-wave radar signal of the radar sensor in this embodiment irradiates the human body, the reflected signal will produce Doppler frequency shift due to the movement of the human body, thereby causing the radar output data to fluctuate. Under the spatial layout of the radar sensors, according to the position relationship of the radar sensors and the radar sensor data, it can be accurately judged whether the elderly are exercising normally, lying down or falling down.

[0075] The AI ​​terminal device, as the central hub of the system, is used to interact with users in a multimodal manner, collect and process the transmission data of the radar group, and store this data for subsequent analysis. In this embodiment, the AI ​​terminal device uses RK3588 as the system master to realize data collection, storage, analysis and support the operation of local models.

[0076] The local large model is used to achieve multimodal interaction with users through AI terminal devices, and use model algorithms to infer and analyze radar group data to identify the indoor activity patterns of the elderly, determine whether the activity intensity is normal, and whether abnormal behavior occurs;

[0077] The 5G communication module is integrated in the AI ​​terminal device, and is used to send a text message containing abnormal information to the guardian's mobile phone according to the preset operation when the AI ​​terminal device combines the local large model to analyze that the elderly have abnormal behavior, so as to ensure that the guardian can know the elderly's situation in time; in this embodiment, the 5G communication module adopts the communication module of Quectel Electronics, which is connected to the RK3588 chip through the PCI interface;

[0078] The alarm module is used to receive instructions from the AI ​​terminal device when the AI ​​terminal device combines the local large model to analyze that the elderly have abnormal behavior, start the speaker to play the pre-set alarm information, attract the attention of people around, and prompt them to check the elderly in time to provide necessary help; in this embodiment, the alarm module uses the chip RK3588 with pre-stored alarm prompt voice. When the alarm is triggered, the voice data is output through the DAC and the voice is played through the speaker through the power amplifier to alert the caregiver that the elderly have an unexpected situation.

[0079] In this embodiment, multiple radar sensors are arranged indoors to form a radar group. The arrangement strategy of multiple radar sensors indoors is specifically implemented by integrating the global search capability of the genetic algorithm and the local search capability of the gray wolf algorithm, wherein:

[0080] Utilize the global search capability of genetic algorithms: Starting from the initial population, through genetic operations such as selection, crossover, and mutation, conduct extensive searches in the entire solution space to find areas that may contain the optimal solution, so as to explore different radar group layout schemes within the preset range and obtain multiple optimal layouts;

[0081] Combined with the local search capability of the Gray Wolf Algorithm: In the area that may contain the optimal solution found by the genetic algorithm, the local search advantage of the Gray Wolf Algorithm is used to search in the local area, continuously approaching the local optimal solution, so that the radar group layout plan can be optimized locally and the accuracy of the layout can be improved.

[0082] Specifically, combined with the Figure 2 , integrating the global search capability of the genetic algorithm and the local search capability of the gray wolf algorithm. The specific integration process is as follows:

[0083] (1) Initialize the radar sensor layout, evenly arrange multiple radar sensors in the working area and introduce vectors to divide the working area. Measure the activity intensity of the elderly within the monitoring range of each radar sensor, and create an initial weight matrix that integrates multiple factors and covers the time dimension extension, providing a quantitative basis for subsequent layout optimization. This process specifically includes:

[0084] (1.1) h radar sensors are evenly placed in multiple working areas, where the working area refers to the indoor activity area of ​​the elderly. The working area is divided into a grid shape, and the working area is divided by introducing a vector P to detect and record the activity data of the elderly. The vector P is described as:

[0085]

[0086] In the formula, is a vector element, representing the h radar sensors in the tth working area, each radar sensor covers at least one grid unit.

[0087] (1.2) Measure the activity intensity of the elderly within the monitoring range of each radar sensor, and create an initial weight matrix based on these data. The weight calculation method is:

[0088]

[0089] In the formula, Indicates that with distance d ij Increase the exponential decay term of signal attenuation; d ij represents the distance from the ith radar sensor to the target elder; α represents the attenuation coefficient that controls the influence of distance on the attenuation rate; h represents the total number of radar sensors; represents the normalization factor, which is the sum of the attenuation of the ith radar sensor relative to all targets and ensures that the sum of the weights is 1; the weight Wij It represents the influence weight of the j-th activity on the i-th sensor. The calculation takes into account the total number of activity frequencies of the elderly and expands it in the time dimension.

[0090] It is necessary to add that: Figure 3 For example, this is the initial layout of radar sensors indoors, not the final layout for detecting the daily activities of the elderly. At the beginning, a large number of sensors need to be evenly distributed throughout the room to comprehensively collect environmental data as algorithm input and optimize the sensor layout based on actual activity patterns and environmental feedback.

[0091] This embodiment collects the daily activity data of the elderly, evaluates the activity frequency of each area of ​​the room, and determines the weights of these areas, such as Figure 4 As shown, it shows the areas where the elderly frequently move around.

[0092] (2) Calculate the average impact value of the radar sensor to evaluate its sensitivity. Construct a sparse matrix related to distance based on the evaluation index of comprehensive distance and attenuation effect. After setting the threshold and comparing the average impact value, construct a sparse weight matrix that retains high-impact weight elements to provide input data for the fusion genetic and gray wolf algorithm. This process specifically includes:

[0093] (2.1) Assuming that the changes in radar sensor values ​​caused by the elderly’s activities are recorded in s test cases, the average impact value V of the i-th radar sensor is i Calculated by the following formula:

[0094]

[0095] In the formula, ΔP i,k represents the value change of the i-th radar sensor in the k-th test case due to the activities of the elderly; d i,k represents the distance between the i-th radar sensor and the elderly's activity location; α is the attenuation coefficient, which is used to adjust the effect of distance on the response sensitivity of the radar sensor;

[0096] (2.2) Define an evaluation index that combines the distance between the elderly and the radar sensor and the attenuation effect α of the distance to measure the sensitivity of the radar sensor in monitoring the activities of the elderly, and further construct a sparse matrix W * , the sparse matrix W * Each element of It reflects the sensitivity of the i-th radar sensor to monitor the activity location of the j-th elderly person. The sparse matrix W * The formula is as follows:

[0097]

[0098] Where, dij is the distance between the jth active position and the ith radar sensor; α is the attenuation coefficient, which is used to adjust the effect of distance on the response sensitivity of the radar sensor;

[0099] (2.3) Set the threshold L and the average impact value V of the radar sensor i When it is less than L, the sparse matrix W * The corresponding weights in are 0 or deleted from the matrix, and the average impact value V of the radar sensor i When it is greater than L, the sparse matrix W is retained * The corresponding weights in are:

[0100]

[0101] Sparse weight matrix W * The weight elements retained in represent the radar sensors with high influence on the system and serve as the input data for the fusion genetic algorithm and the grey wolf algorithm.

[0102] (3) Using the global search capability of the genetic algorithm, a wide search is conducted in the entire solution space to find an area that may contain the optimal solution. Then, combined with the local search capability of the gray wolf algorithm, a search is conducted in the local area to continuously approach the local optimal solution, so that the radar group layout plan is optimized locally. The above two operations are continuously iterated to gradually find a better radar group layout plan until the preset termination condition is met. This process specifically includes:

[0103] (3.1) Determine the initialization stage, including setting the population size, crossover rate, mutation rate, and generating the initial population;

[0104] (3.2) Entering the iterative calculation phase, each iteration includes genetic algorithm selection, crossover, mutation operations and gray wolf position update;

[0105] (3.3) The genetic algorithm updates the population by the following method:

[0106] x ′ =crossover(mutate(select(Pop),p m ),p c ),

[0107] Among them, select, crossover and mutate represent selection, crossover and mutation operations respectively, p m and p c Denote the mutation and crossover probabilities respectively. Some individuals are selected from the initial radar sensor layout, mutation operations are performed on the selected sensors, and finally crossover operations are performed on the mutated individuals to obtain a new sensor set. In this way, the sensor layout is continuously optimized.

[0108] (3.4) The gray wolf algorithm adjusts the positions of the remaining individuals according to the position of the best individual in the current population.

[0109] X new =X alpha -A*D+C*(X beta -X gamma ),

[0110] Where X alpha , X beta and X gamma are the locations of the best, second best, and third best sensors, respectively. A and C are the coefficients that control the search intensity and randomness. D is the vector of the distance between the radar sensor and the elderly’s motion.

[0111] (3.5) Through the iterative optimization process of steps (3.2)-(3.4), iterate until the termination condition is met, that is, the change of the objective function ΔF = |F(xnew)-F(xold)| between two consecutive generations is less than the threshold L, or the maximum number of iterations N is reached max ,According to the calculation process of the objective function, the performance of each radar sensor layout in the indoor behavior monitoring of the elderly is determined. First, a global search is performed through the genetic algorithm to determine the optimal radar sensor layout. Then, a local search and optimization is performed on the layout through the gray wolf algorithm to improve the accuracy and detection efficiency of the current radar sensor layout. The objective function is expressed as follows:

[0112] F(x)=α*Coverage(x)+β*Efficiency(x)+γ*Sensitivity(x),

[0113] In the formula, x represents a radar sensor layout scheme, Coverage(x) represents the coverage of the layout scheme, Efficiency(x) represents the monitoring efficiency, Sensitivity(x) represents the sensitivity of the radar sensor, α, β, and γ are weighted coefficients used to adjust the relative importance of different performance indicators in the objective function. The specific values ​​need to be adjusted according to different application scenarios and priorities. For example, in scenarios that require high coverage, a larger α value can be selected, while in applications that emphasize efficiency, the weight of β can be increased. In cases where sensitivity is more critical, the weight of γ can be increased. The specific selection can be adjusted based on experimental data to achieve the best monitoring effect.

[0114] (4) Use a performance evaluation equation containing multiple factors to evaluate the algorithm, set a threshold and start it, remove radar sensors with efficiency or coverage below the threshold, select the one with the largest weight in each row as the recommended configuration, and fine-tune the threshold based on the evaluation value to reduce the number of active radar sensors while ensuring accuracy; this process specifically includes:

[0115] (4.1) Use the following performance evaluation equation to evaluate the effectiveness of the algorithm:

[0116]

[0117] Where N represents the total number of monitored areas, each area is covered by at least one radar sensor, x represents a specific radar sensor layout scheme, C(i,x) represents the coverage of the radar sensor in the i-th area, E(i,x) represents the efficiency of the radar sensor in the i-th area, S(i,x) represents the sensitivity of the radar sensor in the i-th area, λ1, λ2, and λ3 are weighted coefficients representing the relative importance of coverage, sensitivity, and efficiency, respectively; when the goals and requirements of different stages are different, the selected parameters will be adjusted according to the actual situation;

[0118] (4.2) After setting the threshold L = 1, the algorithm is started to remove radar sensors whose efficiency or coverage is lower than the threshold. If the weight matrix W has N rows, N radar sensors are finally selected. The radar sensors finally retained are the sum of the radar sensors with the largest weight in each row. The number and position of radar sensors obtained at this time are the recommended configuration;

[0119] (4.3) Calculate the performance evaluation value of the recommended configuration. If the evaluation value is not 0, continue to fine-tune the threshold until the performance requirements are met, that is, to obtain the optimal recognition efficiency while using the minimum number of radar sensors.

[0120] In this embodiment, the local large model realizes multimodal interaction with the user through the AI ​​terminal device, and uses the model algorithm to infer and analyze the radar group data, so as to identify the indoor activity pattern of the elderly, determine whether the activity intensity is normal, and whether abnormal behavior occurs. This process specifically includes:

[0121] Data reception and integration: AI terminal devices collect real-time data and multimodal data from the radar group, integrate the multimodal data, and prepare for subsequent unified analysis;

[0122] Data preprocessing: Use data cleaning algorithms to remove outliers and fill in missing values, and extract key features from the integrated multimodal data. For example, for radar data, you may extract features such as movement speed, direction change frequency, and residence time in a specific area. Use signal processing techniques such as Fourier transform to convert time domain signals into frequency domain signals to obtain more dimensional information. For image data, use computer vision technology to extract features such as character outlines and gestures. For voice data, extract voice frequency, intonation, keywords, and other features. Through feature engineering, transform these raw data into a more representative form that is easier for models to process.

[0123] Model reasoning: ① The local large model is pre-trained on a large amount of relevant data to learn the characteristics of various activity patterns, intensity ranges, and abnormal behaviors; when receiving the pre-processed data, the AI ​​terminal device loads the local large model and adapts the data to the input format of the local large model; ② The large model uses convolutional neural networks or recurrent neural networks and their variants in deep neural networks to perform feature analysis on radar data and identify any activity mode of walking, standing, sitting, and lying down; ③ Based on the extracted key features, the local large model predicts the activity intensity through a regression algorithm, and then compares it with the pre-set normal activity intensity range to determine whether the current activity intensity is normal; ④ The anomaly detection algorithm is used to compare the current activity pattern and activity intensity data with the normal pattern learned by the local large model. If the data point is outside the decision boundary of the local large model, or the deviation from the normal pattern exceeds the set threshold, it is determined to be an abnormal behavior;

[0124] Result output: The local large model generates a detailed report based on the results of the reasoning analysis, including the identified activity patterns, activity intensity assessment, and information on whether abnormal behavior exists. The report is presented in a structured form to facilitate further processing by the AI ​​terminal device.

[0125] It should be added that the AI ​​terminal device feeds back the results to the user in a multimodal manner. If it is a direct interaction with the elderly, the current activity status of the elderly is informed in a clear and understandable voice through speech synthesis technology (for example, "You just walked normally in the living room with moderate activity intensity"). For guardians or caregivers, send them detailed activity reports via text messages. If abnormal behavior is detected, an alarm notification is immediately issued, such as an emergency reminder pushed via text message with an explanation of the abnormal situation, and the alarm module (such as an indoor sound alarm) is activated to attract the attention of people around. Users can interact with the local large model through AI terminal devices. For example, the guardian can query the detailed activity records of the elderly in a specific time period through the AI ​​terminal device, or provide feedback on the judgment results of the local large model (such as misjudgment correction). These feedback information can be used to further optimize the local large model and improve its accuracy in monitoring and analyzing the activities of the elderly.

[0126] In summary, the indoor monitoring system for the elderly based on terminal-side artificial intelligence of the present invention realizes all-round, real-time monitoring and efficient emergency response of the indoor activities of the elderly through the collaborative work of various parts, providing strong protection for the safety of the elderly.

[0127] The above specific examples are used to explain the principles and implementation methods of the present invention in detail. These examples are only used to help understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made by technicians in this technical field without departing from the principles of the present invention should fall within the scope of patent protection of the present invention.

Claims

1. An indoor monitoring system for the elderly based on terminal-side artificial intelligence, characterized in that: It consists of five parts: radar group, AI terminal equipment, local large model, 5G communication module and alarm module, among which: Radar group, including multiple radar sensors arranged indoors, is used to continuously monitor the indoor space, collect activity data of the elderly, and transmit the collected data to AI terminal devices in real time, providing basic data support for subsequent analysis and judgment; The AI ​​terminal device, as the central hub of the system, is used to interact with users in a multimodal manner, collect and process the transmission data of the radar group, and store this data for subsequent analysis; The local large model is used to achieve multimodal interaction with users through AI terminal devices, and use model algorithms to infer and analyze radar group data to identify the indoor activity patterns of the elderly, determine whether the activity intensity is normal, and whether abnormal behavior occurs; The 5G communication module is integrated in the AI ​​terminal device. When the AI ​​terminal device combines the local large model to analyze and conclude that the elderly have abnormal behavior, it sends a text message containing abnormal information to the guardian's mobile phone according to the preset operation, ensuring that the guardian can be informed of the elderly's situation in a timely manner; The alarm module is used to receive instructions from the AI ​​terminal device and start the speaker to play a pre-set alarm message to attract the attention of people around when the AI ​​terminal device combines the local large model to analyze and conclude that the elderly have abnormal behavior.

2. According to claim 1, the indoor monitoring system for the elderly based on terminal-side artificial intelligence is characterized in that: By integrating the global search capability of the genetic algorithm and the local search capability of the gray wolf algorithm, a radar group containing multiple radar sensors is deployed indoors, where: Utilize the global search capability of genetic algorithms: Starting from the initial population, through genetic operations such as selection, crossover, and mutation, conduct extensive searches in the entire solution space to find areas that may contain the optimal solution, so as to explore different radar group layout schemes within the preset range and obtain multiple optimal layouts; Combined with the local search capability of the Gray Wolf Algorithm: In the area that may contain the optimal solution found by the genetic algorithm, the local search advantage of the Gray Wolf Algorithm is used to search in the local area, continuously approaching the local optimal solution, so that the radar group layout plan can be optimized locally and the accuracy of the layout can be improved.

3. According to claim 2, the indoor monitoring system for the elderly based on terminal-side artificial intelligence is characterized in that: By integrating the global search capability of the genetic algorithm and the local search capability of the gray wolf algorithm, a radar group containing multiple radar sensors is arranged indoors. The specific fusion process is as follows: (1) Initialize the radar sensor layout, evenly arrange multiple radar sensors in the working area and introduce vectors to divide the working area. Measure the activity intensity of the elderly within the monitoring range of each radar sensor, and create an initial weight matrix that integrates multiple factors and covers the time dimension extension, providing a quantitative basis for subsequent layout optimization. (2) The average impact value of the radar sensor is calculated to evaluate its sensitivity. A sparse matrix related to distance is constructed based on the evaluation index of comprehensive distance and attenuation effect. After setting the threshold and comparing the average impact value, a sparse weight matrix that retains high-impact weight elements is constructed to provide input data for the fusion of genetic and gray wolf algorithms. (3) Using the global search capability of the genetic algorithm, a wide search is conducted in the entire solution space to find an area that may contain the optimal solution. Then, combined with the local search capability of the gray wolf algorithm, a search is conducted in the local area to continuously approach the local optimal solution, so that the radar group layout plan is locally optimized. The above two operations are continuously iterated to gradually find a better radar group layout plan until the preset termination condition is met. (4) The algorithm is evaluated using a performance evaluation equation containing multiple factors. After setting the threshold, it is started and radar sensors with efficiency or coverage below the threshold are eliminated. The configuration with the largest weight in each row is selected as the recommended configuration. The threshold is fine-tuned according to the evaluation value to reduce the number of active radar sensors and ensure accuracy.

4. According to claim 3, the indoor monitoring system for the elderly based on terminal-side artificial intelligence is characterized in that: The step (1) specifically comprises: (1.1) h radar sensors are evenly placed in multiple working areas, where the working area refers to the indoor activity area of ​​the elderly. The working area is divided into a grid shape, and the working area is divided by introducing a vector P to detect and record the activity data of the elderly. The vector P is described as: In the formula, is a vector element, representing the h radar sensors in the t-th working area, each radar sensor covers at least one grid cell; (1.2) Measure the activity intensity of the elderly within the monitoring range of each radar sensor, and create an initial weight matrix based on these data. The weight calculation method is: In the formula, Indicates that with distance d ij Increase the exponential decay term of signal attenuation; d ij represents the distance from the ith radar sensor to the target elder; α represents the attenuation coefficient that controls the influence of distance on the attenuation rate; h represents the total number of radar sensors; represents the normalization factor, which is the sum of the attenuation of the ith radar sensor relative to all targets and ensures that the sum of the weights is 1; the weight W ij It represents the influence weight of the j-th activity on the i-th sensor. The calculation takes into account the total number of activity frequencies of the elderly and expands it in the time dimension.

5. According to claim 4, the indoor monitoring system for the elderly based on terminal-side artificial intelligence is characterized in that: The step (2) specifically comprises: (2.1) Assuming that the changes in radar sensor values ​​caused by the elderly’s activities are recorded in s test cases, the average impact value V of the i-th radar sensor is i Calculated by the following formula: In the formula, ΔP i,k represents the value change of the i-th radar sensor in the k-th test case due to the activities of the elderly; d i,k represents the distance between the i-th radar sensor and the elderly's activity location; α is the attenuation coefficient, which is used to adjust the effect of distance on the response sensitivity of the radar sensor; (2.2) Define an evaluation index that combines the distance between the elderly and the radar sensor and the attenuation effect α of the distance to measure the sensitivity of the radar sensor in monitoring the activities of the elderly, and further construct a sparse matrix W * , the sparse matrix W * Each element of Reflects the sensitivity of the j-th radar sensor to monitor the i-th activity, and the sparse matrix W * The formula is as follows: Where, d ij is the distance between the jth active position and the ith radar sensor; α is the attenuation coefficient, which is used to adjust the effect of distance on the response sensitivity of the radar sensor; (2.3) Set the threshold L and the average impact value V of the radar sensor i When it is less than L, the sparse matrix W * The corresponding weights in are 0 or deleted from the matrix, and the average impact value V of the radar sensor i When it is greater than L, the sparse matrix W is retained * The corresponding weights in are: Sparse weight matrix W * The weight elements retained in represent the radar sensors with high influence on the system and serve as the input data for the fusion genetic algorithm and the grey wolf algorithm.

6. The indoor monitoring system for the elderly based on terminal-side artificial intelligence according to claim 5 is characterized in that: The step (3) specifically comprises: (3.1) Determine the initialization stage, including setting the population size, crossover rate, mutation rate, and generating the initial population; (3.2) Entering the iterative calculation phase, each iteration includes genetic algorithm selection, crossover, mutation operations and gray wolf position update; (3.3) The genetic algorithm updates the population by the following method: x ′ =crossover(mutate(select(Pop),p m ),p c ), Among them, select, crossover and mutate represent selection, crossover and mutation operations respectively, p m and p c represent mutation and crossover probabilities respectively; (3.4) The gray wolf algorithm adjusts the positions of the remaining individuals according to the position of the best individual in the current population. X new =X alpha -A*D+C*(X beta -X gamma ), Where X alpha , X beta and X gamma are the locations of the best, second best, and third best sensors, respectively. A and C are the coefficients that control the search intensity and randomness. D is the vector of the distance between the radar sensor and the elderly’s motion. (3.5) Through the iterative optimization process of steps (3.2)-(3.4), iterate until the termination condition is met, that is, the change of the objective function ΔF = |F(xnew)-F(xold)| between two consecutive generations is less than the threshold L, or the maximum number of iterations N is reached max , where the objective function is expressed as follows: F(x)=α*Coverage(x)+β*Efficiency(x)+γ*Sensitivity(x), Where x represents a radar sensor layout scheme, Coverage(x) represents the coverage of the layout scheme, Efficiency(x) represents the monitoring efficiency, Sensitivity(x) represents the sensitivity of the radar sensor, α, β, and γ are weighted coefficients used to adjust the relative importance of different performance indicators in the objective function. The specific values ​​need to be adjusted according to different application scenarios and priorities.

7. The indoor monitoring system for the elderly based on terminal-side artificial intelligence according to claim 6 is characterized in that: The step (4) specifically comprises: (4.1) Use the following performance evaluation equation to evaluate the effectiveness of the algorithm: Where N represents the total number of monitored areas, each area is covered by at least one radar sensor, x represents a specific radar sensor layout scheme, C(i,x) represents the coverage of the radar sensor in the i-th area, E(i,x) represents the efficiency of the radar sensor in the i-th area, S(i,x) represents the sensitivity of the radar sensor in the i-th area, λ1, λ2, and λ3 are weighted coefficients representing the relative importance of coverage, sensitivity, and efficiency, respectively; when the goals and requirements of different stages are different, the selected parameters will be adjusted according to the actual situation; (4.2) After setting the threshold L = 1, the algorithm is started to remove radar sensors whose efficiency or coverage is lower than the threshold. If the weight matrix W has N rows, N radar sensors are finally selected. The radar sensors finally retained are the sum of the radar sensors with the largest weight in each row. The number and position of radar sensors obtained at this time are the recommended configuration; (4.3) Calculate the performance evaluation value of the recommended configuration. If the evaluation value is not 0, continue to fine-tune the threshold until the performance requirements are met.

8. The indoor monitoring system for the elderly based on terminal-side artificial intelligence according to claim 1 is characterized in that: The local large model realizes multimodal interaction with users through AI terminal devices, and uses model algorithms to infer and analyze radar group data, thereby identifying the indoor activity patterns of the elderly, judging whether the activity intensity is normal, and whether abnormal behavior occurs. This process specifically includes: Data reception and integration: AI terminal devices collect real-time data and multimodal data from the radar group, integrate the multimodal data, and prepare for subsequent unified analysis; Data preprocessing: Use data cleaning algorithms to remove outliers and fill missing values, and extract key features from the integrated multimodal data; Model reasoning: ① The local large model is pre-trained on a large amount of relevant data to learn the characteristics of various activity patterns, intensity ranges, and abnormal behaviors; when receiving the pre-processed data, the AI ​​terminal device loads the local large model and adapts the data to the input format of the local large model; ② The large model uses convolutional neural networks or recurrent neural networks and their variants in deep neural networks to perform feature analysis on radar data and identify any activity mode of walking, standing, sitting, and lying down; ③ Based on the extracted key features, the local large model predicts the activity intensity through a regression algorithm, and then compares it with the pre-set normal activity intensity range to determine whether the current activity intensity is normal; ④ The anomaly detection algorithm is used to compare the current activity pattern and activity intensity data with the normal pattern learned by the local large model. If the data point is outside the decision boundary of the local large model, or the deviation from the normal pattern exceeds the set threshold, it is determined to be an abnormal behavior; Result output: The local large model generates a detailed report based on the results of the reasoning analysis, including the identified activity patterns, activity intensity assessment, and information on whether abnormal behavior exists. The report is presented in a structured form to facilitate further processing by the AI ​​terminal device.

Images