Dynamic air purification cruising method based on real-time health perception

Abstract

The invention discloses a dynamic air purification cruising method based on real-time health perception, which comprises the following steps of: constructing an initial characteristic variable set based on acquired multi-dimensional data; performing a round of feature screening by using an adaptive weight distribution algorithm to obtain a first feature variable set; performing two rounds of feature screening through information entropy and mutual information analysis to obtain a second feature variable set; based on the deep reinforcement learning model and the second feature variable set, constructing a dynamic path optimization strategy, and obtaining an optimal cruise path; and based on the optimal cruise path, an air purification task is executed, and environment evaluation is conducted. According to the method, a traditional fixed mode is broken through, two types of core data-user real-time physiological indexes and indoor environment parameters are innovatively combined, precise positioning of a pollution source can be achieved, meanwhile, the equipment cruise path is dynamically adjusted according to data changes, and waste of purification resources is avoided.

Description

Technical Field

[0001] This application belongs to the field of air purification technology, specifically relating to a dynamic air purification cruise method based on real-time health perception. Background Technology

[0002] With the rapid development of smart home and health environment monitoring technologies, air purifiers are gradually evolving from simple purification devices into intelligent devices with real-time sensing and dynamic adjustment capabilities, playing a significant role in improving indoor air quality. Especially in the field of automatic cruise air purifiers, these devices effectively purify the air in various indoor areas by moving and planning their own paths, further enhancing the user experience. However, current technology still has significant shortcomings in real-time health sensing and dynamic adjustment capabilities, failing to meet users' personalized and precise expectations for indoor air quality.

[0003] Patent CN113757829B discloses a method for controlling an air purifier based on weather information. It acquires external weather data and pushes appropriate purification scenarios to achieve intelligent monitoring and purification. However, this technology largely relies on weather information as a trigger, failing to adequately consider the impact of real-time indoor health data (human breathing patterns, changes in harmful gas concentrations, etc.) on the air purification strategy. Therefore, it may exhibit slow or inaccurate responses when facing complex indoor environments. Furthermore, the control logic of this solution is relatively rigid, lacking the ability to personalize adjustments based on the health needs of different user groups and failing to improve purification effectiveness according to the user's real-time condition.

[0004] The patent with publication number CN110686335B utilizes 3D image generation and the Floyd algorithm to plan optimal paths, enabling air purifiers to efficiently inspect and purify different areas indoors. While this technology offers advantages in path optimization and efficiency improvement, it lacks a real-time health sensing mechanism, meaning it doesn't dynamically monitor human physiological indicators or specific pollutants. Therefore, it cannot dynamically adjust the purification strategy based on the user's current health status. Furthermore, this solution focuses on improving path efficiency but neglects the differences in health priorities across areas, potentially leading to insufficient purification in critical areas to meet the user's needs in a timely manner.

[0005] The aforementioned issues demonstrate that current air purification control methods have significant shortcomings in terms of real-time health perception, dynamic adjustment capabilities, and personalized services. In particular, in the application scenario of automatic cruise air purifiers, how to integrate real-time health data monitoring and analysis mechanisms, and meet users' personalized needs by dynamically adjusting the cruise path and purification strategy, remains an urgent technical challenge to be solved. Summary of the Invention

[0006] This application provides a dynamic air purification cruise method based on real-time health perception to solve the above-mentioned technical problems.

[0007] To address the aforementioned technical problems, this application adopts the following technical solution: a dynamic air purification cruise method based on real-time health perception, comprising:

[0008] S1. Construct an initial set of feature variables based on the collected multidimensional data;

[0009] S2. Based on the initial feature variable set, perform a round of feature selection using an adaptive weight allocation algorithm to obtain the first feature variable set;

[0010] S3. Based on the first set of feature variables, a second set of feature variables is obtained by performing a second round of feature selection through information entropy and mutual information analysis;

[0011] S4. Based on the deep reinforcement learning model and the second feature variable set, construct a dynamic path optimization strategy to obtain the optimal cruising path;

[0012] S5. Based on the optimal cruise path, perform air purification tasks and conduct environmental assessments.

[0013] Furthermore, the method in step S1 includes:

[0014] S11. Real-time collection of user physiological data, including respiratory rate and heart rate data, through the smart bracelet worn by the user, and collection of environmental data, including indoor PM2.5 and CO concentration, through the sensors on the air purifier;

[0015] S12. Based on user physiological data and environmental data, extract real-time monitoring data of all sampling points within a cycle to obtain the cycle-extracted data;

[0016] S13. Based on periodic data extraction, obtain an initial set of feature variables; wherein, the initial set of feature variables includes respiratory rate fluctuation index, harmful gas concentration gradient, particulate matter distribution density, and heart rate variability index.

[0017] Furthermore, the method in step S2 includes:

[0018] S21. Based on the initial feature variables in the initial feature variable set, obtain the correlation coefficient between each initial feature variable and the target variable; where the target variable is the air quality health index;

[0019] S22. Based on the correlation coefficient, obtain the adaptive weights for each initial feature variable;

[0020] S23. Analyze the contribution of features to the target variable using a random forest regressor, select the critical importance score when the cumulative contribution rate reaches 85% as the threshold, retain the initial feature variables with adaptive weights greater than the threshold, and obtain the first feature variable set.

[0021] Furthermore, the method in step S3 includes:

[0022] S31. Based on the first feature variables in the first feature variable set, obtain the information entropy of the first feature variables;

[0023] S32. Based on information entropy, obtain the mutual information of two adjacent first feature variables;

[0024] S33. Find the threshold of mutual information through cross-validation; if the mutual information of the two first feature variables is greater than the threshold and their correlation coefficient is greater than 0.8, then retain the first feature variable with the higher mutual information value to the target variable and remove the other; if the mutual information of the two first feature variables is less than the threshold, then remove the corresponding feature variable.

[0025] S34. Based on the retained first feature variables, obtain the second feature variable set.

[0026] Furthermore, the method in step S4 includes:

[0027] S41. By cross-validating the second feature variable set through time series analysis, the samples in the second feature variable set are divided into training set, validation set and test set in a ratio of 7:2:1.

[0028] S42. Construct a learning model based on a deep Q-network, and adopt an architecture that combines competitive deep Q-networks with dual Q-learning to obtain the model loss function;

[0029] S43. Based on the validation set, the hyperparameters in the model loss function are adjusted using the Bayesian optimization method to obtain the reinforcement learning model;

[0030] S44. Based on the training set, train the reinforcement learning model to obtain the dynamic path optimization strategy containing the trained optimization learning model;

[0031] S45. Calculate the optimal cruise path based on the dynamic path optimization strategy.

[0032] Furthermore, the method in step S5 includes:

[0033] S51. Establish an environmental change response mechanism: When the purifier detects that the rate of change of any environmental parameter exceeds the preset physical reasonable threshold within 30 seconds, it is determined to be an environmental change, the current optimization path is immediately interrupted, and an emergency replanning process is initiated to prioritize the change area, with the response time controlled within 1 minute.

[0034] S52. Based on the air quality improvement rate, health indicator satisfaction, and energy consumption efficiency, evaluate the purification effect and establish a model update trigger mechanism;

[0035] S53. Based on the evaluation results, dynamically adjust the parameters and feature weights of the dynamic path optimization strategy, and continuously optimize the reinforcement learning model using an online learning algorithm.

[0036] The beneficial effects of this application are as follows: Breaking away from traditional fixed models, this application innovatively combines two types of core data—real-time user physiological indicators (such as respiratory rate and heart rate) and indoor environmental parameters (such as harmful gas concentration and particulate matter distribution)—to achieve precise location of pollution sources. Simultaneously, it dynamically adjusts the equipment's patrol path based on data changes, avoiding waste of purification resources. This application not only collects various data related to user health but also performs in-depth data analysis, providing a core basis for adjusting purification strategies, making the purification process more aligned with user health needs, rather than relying solely on single environmental data. By combining the analysis results collected by the health sensing system with an improved dynamic path planning algorithm, this application effectively solves the purification challenges in complex indoor environments (such as multi-room, multi-obstacle layouts), overcoming the limitations of traditional algorithms' low efficiency in complex scenarios and achieving highly efficient purification coverage. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating an embodiment of the dynamic air purification cruise method based on real-time health perception according to this application.

[0038] Figure 2 yes Figure 1 A flowchart illustrating step S1 of an embodiment;

[0039] Figure 3 yes Figure 1 A flowchart illustrating step S2 of an embodiment;

[0040] Figure 4 yes Figure 1 A flowchart illustrating step S3 of an embodiment;

[0041] Figure 5 yes Figure 1 A flowchart illustrating step S4 in one embodiment;

[0042] Figure 6 yes Figure 1 A flowchart illustrating step S5 of an embodiment;

[0043] Figure 7 The curves showing the changes in user physiological indicators and environmental parameters in the living room scene of Embodiment 1 in this application;

[0044] Figure 8The curves showing the changes in user physiological indicators and environmental parameters in the kitchen scene of Embodiment 1 in this application;

[0045] Figure 9 The curves showing the changes in user physiological indicators and environmental parameters in the master bedroom sleep scenario of Embodiment 1 in this application;

[0046] Figure 10 The specific process of determining the threshold for grid search optimization in Embodiment 1 of this application;

[0047] Figure 11 The curve showing the change of the reward function for path optimization in the living room scene of Embodiment 1 in this application;

[0048] Figure 12 The curve showing the change of the reward function for the path optimization in the kitchen scene of Embodiment 1 in this application;

[0049] Figure 13 The reward function change curve for the master bedroom sleep scene path optimization in Embodiment 1 of this application;

[0050] Figure 14 Line graph showing the CO purification effect of the real-time health sensing air purification adjustment method in the kitchen cooking scene of Embodiment 1 of this application;

[0051] Figure 15 This is a line graph showing the PM2.5 purification effect of the real-time health sensing air purification adjustment method in the kitchen cooking scene of Embodiment 1 of this application.

[0052] Figure 16 This is a line graph showing the CO purification effect in a living room activity scene of the real-time health sensing air purification adjustment method of Embodiment 1 in this application.

[0053] Figure 17 This is a line graph showing the PM2.5 purification effect in a living room activity scene of the real-time health sensing air purification adjustment method of Embodiment 1 in this application.

[0054] Figure 18 This is a line graph showing the CO purification effect in the master bedroom sleeping scene of the real-time health sensing air purification adjustment method of Embodiment 1 in this application.

[0055] Figure 19 This is a line graph showing the PM2.5 purification effect in the master bedroom sleeping scene of the real-time health sensing air purification adjustment method of Embodiment 1 in this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0057] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0058] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the dynamic air purification cruise method based on real-time health perception according to this application. The method includes:

[0059] S1. Based on the collected multidimensional data, construct an initial set of feature variables.

[0060] For details, please refer to Figure 2 The method in step S1 includes:

[0061] S11. Real-time collection of user physiological data, including respiratory rate and heart rate, through the smart bracelet worn by the user, and collection of environmental data, including indoor PM2.5 and CO concentration, through the sensors on the air purifier.

[0062] Specifically, the system monitors the user's respiratory rate and heart rate variability in real time through a smart bracelet worn by the user, while high-precision sensors mounted on the autonomous cruise vehicle collect indoor PM2.5 and CO concentrations in real time. The collected multi-source heterogeneous data undergoes unified signal preprocessing, including filtering and denoising physiological signals, aligning data from devices with different sampling frequencies using a time synchronization method, and then performing spatiotemporal alignment and data fusion on the synchronized data to establish the correlation between the user's physiological indicators and environmental parameters.

[0063] S12. Based on user physiological data and environmental data, extract real-time monitoring data from all sampling points within a cycle to obtain the cycle-extracted data. .

[0064] Specifically, user physiological indicators (respiratory rate, heart rate variability) and environmental parameters (PM2.5 concentration, CO concentration) are used as data samples to extract real-time monitoring data from all sampling points within a period:

[0065] (1);

[0066] in, Indicates window size. This represents the sampled data at time t.

[0067] S13. Based on periodic data extraction, obtain an initial set of feature variables; wherein, the initial set of feature variables includes respiratory rate fluctuation index, harmful gas concentration gradient, particulate matter distribution density, and heart rate variability index.

[0068] Specifically, based on the extracted data, multiple feature variables are set, wherein the feature variables The respiratory rate fluctuation index is equal to the root mean square of the amplitude of respiratory rate change over a period of time, and is expressed as follows:

[0069] (2);

[0070] in, This represents the respiratory rate at time i. The average respiratory rate during the period is represented by N, which represents the total number of sampling points.

[0071] Feature variables This represents the concentration gradient of harmful gases, specifically as follows:

[0072] (3);

[0073] in, This indicates the change in the concentration of harmful gases. Indicates a time interval;

[0074] Feature variables The distribution density of particulate matter is represented as follows:

[0075] (4);

[0076] in, Let A represent the particulate matter concentration in the j-th region, A represent the total area of ​​the region, and M represent the number of regions.

[0077] Feature variables Represents the heart rate variability index, specifically as follows:

[0078] (5);

[0079] Wherein, SDNN (Standard Deviation of Normal-to-Normal intervals) represents the standard deviation of heart rate, and RMSSD (Root Mean Square of Successive Differences) represents the root mean square of the difference between adjacent heartbeats; the corresponding formulas are as follows:

[0080] (6);

[0081] (7);

[0082] Where N represents the total number of sampling points, Indicates the first A normal sinus interval, This represents the (i+1)th normal sinus interval. The arithmetic mean over the period represented.

[0083] S2. Based on the initial feature variable set, use an adaptive weight allocation algorithm to perform a round of feature selection to obtain the first feature variable set.

[0084] For details, please refer to Figure 3 The method of step S2 includes:

[0085] S21. Based on the initial feature variables in the initial feature variable set, obtain the correlation coefficient between each initial feature variable and the target variable; the target variable is the air quality health index.

[0086] Specifically, an adaptive weighting assumption is established, assuming that the contribution of each feature variable to the purification effect is proportional to its weight. The correlation coefficient between each initial feature variable and the target variable (Air Quality Index, AQI) is calculated. The specific formula is as follows:

[0087] (8);

[0088] Where cov represents covariance. The standard deviation is represented by the formula for calculating the AQI:

[0089] (9);

[0090] The corresponding calculation of the air quality index for each pollutant. (This sub-index represents the air quality index of a single pollutant, here referring to the air quality index of PM2.5 and CO.) The formula is:

[0091] (10);

[0092] in, This represents the real-time concentration measurement of pollutant p. and These are respectively defined according to the national "Ambient Air Quality Standards" and... The upper and lower limits of the concentration limits corresponding to the relevant concentration range; and They are and The corresponding air quality sub-index limit.

[0093] The entire system is based on a pollutant concentration-sub-index lookup table constructed according to the aforementioned national standards. The real-time concentration is obtained by searching this table. Matching , , as well as Find the value and substitute it into the formula to complete the calculation.

[0094] S22. Based on the correlation coefficient, obtain the adaptive weights for each initial feature variable.

[0095] Specifically, based on the correlation coefficient matrix, adaptive weights for each feature variable are calculated. Specifically, it is expressed as follows:

[0096] (11);

[0097] in, Let K represent the correlation coefficient between the k-th feature variable and the target variable, where K represents the total number of feature variables.

[0098] S23. Analyze the contribution of features to the target variable using a random forest regressor, select the critical importance score when the cumulative contribution rate reaches 85% as the threshold, retain the initial feature variables with adaptive weights greater than the threshold, and obtain the first feature variable set.

[0099] Specifically, the contribution of features to the target variable AQI is analyzed using a random forest regressor, and the critical importance score at which the cumulative contribution rate reaches 85% is selected as the threshold. and retain The characteristics constitute the first set of characteristic variables. .

[0100] S3. Based on the first set of feature variables, a second set of feature variables is obtained by performing a second round of feature selection through information entropy and mutual information analysis.

[0101] For details, please refer to Figure 4 The method of step S3 includes:

[0102] S31. Based on the first feature variable in the first feature variable set, obtain the information entropy of the first feature variable.

[0103] Specifically, based on the first feature variable set For each first feature variable in the dataset, its information entropy is calculated, as shown below:

[0104] (12);

[0105] in, Let represent the probability of the i-th eigenvalue appearing, and n represent the total number of eigenvalues.

[0106] S32. Based on information entropy, obtain the mutual information of two adjacent first feature variables.

[0107] Specifically, based on information entropy, the mutual information between the first feature variables is calculated, as follows:

[0108] (13);

[0109] in, Describes the joint probability distribution. and These represent the marginal probability distributions.

[0110] S33. Finding the threshold of mutual information through cross-validation If the first characteristic variable With the first characteristic variable The mutual information is greater than the threshold. If the correlation coefficient between the two first feature variables is greater than 0.8, then the first feature variable with the higher mutual information value to the target variable AQI is retained, and the other is removed; if the mutual information of both first feature variables is less than the threshold, i.e. If so, then the corresponding feature variable is removed.

[0111] S34. Based on the retained first feature variables, obtain the second feature variable set. .

[0112] S4. Based on the deep reinforcement learning model and the second feature variable set, a dynamic path optimization strategy is constructed to obtain the optimal cruising path.

[0113] Specifically, define the hyperparameters of the reinforcement learning model:

[0114] (14);

[0115] in, Let represent the value of the k-th second characteristic variable at time t. E represents the degree of air quality improvement, T represents the time cost (seconds), and H represents the degree of health improvement. The weighting coefficients are automatically optimized through value iteration during the reinforcement learning process.

[0116] See Figure 5 The method of step S4 includes:

[0117] S41. By cross-validating the second feature variable set through time series analysis, the samples in the second feature variable set are divided into training set, validation set and test set in a ratio of 7:2:1.

[0118] S42. A learning model is constructed based on a deep Q-network, employing an architecture combining competitive deep Q-networks and dual Q-learning to improve the accuracy of value estimation and training stability, and to obtain the model loss function. The model loss function is as follows:

[0119] (15);

[0120] in, For online network parameters, For the target network parameters, Here, E represents the discount factor and the expected value.

[0121] S43. Based on the validation set, the hyperparameters in the model loss function are adjusted using the Bayesian optimization method to obtain the reinforcement learning model;

[0122] S44. Based on the training set, train the reinforcement learning model to obtain the dynamic path optimization strategy containing the trained optimization learning model;

[0123] S45. Calculate the optimal cruise path based on the dynamic path optimization strategy.

[0124] Specifically, based on the dynamic path optimization strategy, the optimal cruising path is calculated, and its objective function is:

[0125] (16);

[0126] in, This represents the distance of the i-th path segment. This represents the purification cost of the i-th path segment; , These are weighting coefficients, dynamically adjusted through online learning. The maximum path length. This is the minimum health benefit threshold.

[0127] S5. Based on the optimal cruise path, perform air purification tasks and conduct environmental assessments.

[0128] Specifically, based on the optimal cruising path, an air purification task is performed; the cruising vehicle moves along the planned path at a speed of 0.5 m / s, and the purification intensity is dynamically adjusted according to the needs of the scenario (where CADR refers to "Clean Air Delivery Rate", and the specific levels are: low 180 m³ / h, medium 320 m³ / h, and high 480 m³ / h).

[0129] See Figure 6 The method in step S5 includes:

[0130] S51. Establish an environmental change response mechanism: When the purifier detects that the rate of change of any environmental parameter exceeds the preset physical reasonable threshold within 30 seconds, it is determined to be an environmental change, the current optimization path is immediately interrupted, and an emergency replanning process is initiated to prioritize the change area, with the response time controlled within 1 minute.

[0131] S52. Based on air quality improvement rate Health indicators and satisfaction and energy efficiency The purification effect is evaluated, and a model update triggering mechanism is established.

[0132] Specifically, multidimensional evaluation indicators are used to assess the purification effect:

[0133] (17);

[0134] in, The air purification number before purification. The air purification number is the number of purified air. The representative was normalized. This represents the weight of the k-th health characteristic variable.

[0135] S53. Based on the evaluation results, dynamically adjust the parameters and feature weights of the dynamic path optimization strategy, and continuously optimize the reinforcement learning model using an online learning algorithm.

[0136] Specifically, establish a model update trigger mechanism:

[0137] If this triggers model retraining (18);

[0138] in, The threshold for air quality improvement rate. The minimum satisfaction threshold is determined using statistical process control methods. Based on the evaluation results, strategy parameters and feature weights are dynamically adjusted, and an online learning algorithm is used to continuously optimize system performance.

[0139] To verify the effectiveness of the present invention, numerical simulation was performed using Matlab.

[0140] Example 1

[0141] To verify the effectiveness of this invention, a 120㎡ three-bedroom family scenario (living room 35㎡, master bedroom 25㎡, second bedroom 18㎡, kitchen 12㎡, bathroom 8㎡, corridor and other areas 14㎡) was selected as the test environment, and the experiment was conducted according to the following steps:

[0142] Step 1: Test Environment Setup and Data Acquisition

[0143] The system simulates a family member's daily activity pattern (8:00-18:00 living room activities, 18:00-20:00 cooking in the kitchen, 22:00-7:00 the next day resting in the master bedroom). In this scenario, the system monitors the user's respiratory rate and heart rate variability in real time via a smart bracelet, while simultaneously collecting indoor PM2.5 and CO concentrations using high-precision sensors on an automatic cruise air purifier. Real-time data from 60 sampling points (sampling frequency 1 time / minute) is collected over one cycle (1 hour), and the multi-source heterogeneous data undergoes signal preprocessing and spatiotemporal alignment fusion.

[0144] Step 2: Feature Variable Extraction

[0145] See Figure 7-9 Based on the collected data, the system extracts four feature variables: respiratory rate fluctuation index, harmful gas concentration gradient, particulate matter distribution density, and heart rate variability index. Figure 7 , Figure 8 , Figure 9 The changing trends of each feature parameter in these three scenarios are shown respectively, providing a data foundation for subsequent feature selection.

[0146] Step 3: First round of feature selection (adaptive weight allocation)

[0147] Based on 3000 historical data points from multiple scenarios, a random forest regression model was used to calculate the importance scores of each feature to the Air Quality Index (AQI). Importance analysis results showed that the particulate matter distribution density was 0.81, the harmful gas concentration gradient was 0.68, the respiratory rate fluctuation index was 0.49, and the heart rate variability index was 0.26. An initial threshold of 0.15 was determined through cumulative contribution rate analysis. To further improve the performance of the final purification strategy, the average cumulative reward obtained on the validation set after inputting the aforementioned selected feature subset into the deep reinforcement learning model was used as the optimization objective. A grid search was performed to optimize this initial threshold. Parameter optimization was conducted within the search range of (0.1, 0.5) with a step size of 0.01, ultimately determining the optimal threshold that maximizes the average cumulative reward. =0.23. The specific process of determining the threshold using grid search optimization is as follows: Figure 10 As shown.

[0148] Table 1 shows the correlation coefficients and corresponding weights between the characteristic variables and the target variable in the four adaptive scenario categories.

[0149]

[0150] Analysis of the data in Table 1 shows that the weights of two feature variables in the living room activity scenario are greater than the threshold. =0.23 (F2=0.40, F3=0.44); in the kitchen cooking scenario, the weights of the three feature variables are greater than 0.23 (F1=0.38, F2=0.40, F3=0.36); in the master bedroom sleeping scenario, the weights of the two feature variables are greater than 0.23 (F2=0.28, F3=0.48). The first round of feature selection is complete, and the feature subsets corresponding to each scenario enter the second round of selection.

[0151] Step 4: Second round of feature selection (information entropy and mutual information analysis)

[0152] A two-round feature selection process using information entropy and mutual information analysis was employed. This was to determine the optimal mutual information threshold. The average cumulative reward of the aforementioned deep reinforcement learning model on the validation set was used as the evaluation metric, and optimization was performed using 5-fold hierarchical cross-validation. The 3000 data points were divided into 5 groups of 600 data points each (200 data points per scenario). A search was conducted within the candidate threshold range [0.3, 0.7] with a step size of 0.05, and the threshold that maximized the average cumulative reward was ultimately selected. =0.45.

[0153] Table 2 shows the original information entropy and mutual information values ​​of the feature variables for the three types of scenarios.

[0154]

[0155] As shown in Table 2, the mutual information values ​​between variables within each feature subset of each scene are all less than the threshold. =0.45, with no significant information redundancy, all features were retained to form the final optimized feature sets for each scenario: Living room activity scenario optimized feature set: {F2, F3}, kitchen cooking scenario optimized feature set: {F1, F2}, master bedroom sleeping scenario optimized feature set: {F3}. Among them, the information entropy of F2 in the kitchen cooking scenario is the highest (1.25), indicating that the CO concentration fluctuation is the most significant in this scenario, which needs special attention in subsequent path optimization.

[0156] Step 5: Deep Reinforcement Learning Model Training and Path Optimization

[0157] See Figure 11-13 ,in Figure 11-13The convergence curves of the reward functions for three scenarios are shown. The retained feature variables from the three scenarios are input into a deep reinforcement learning model for dynamic path optimization strategy training. The model adopts an architecture combining a competitive deep Q-network and dual Q-learning as described in step S43, with the input state vector containing feature variables and position coordinates. The training process employs an experience replay mechanism, using the Bayesian optimization method described in step S44 to fine-tune key hyperparameters, with the optimization objective being to maximize the cumulative reward on the validation set. The training convergence criterion is an average reward change of <1% over 10 consecutive cycles, and training for a single scenario takes approximately 2 hours.

[0158] The optimal cruising paths and purification strategies for each scenario after training are shown in Table 3.

[0159] Table 3 shows the optimized paths and purification strategies for the three scenarios.

[0160]

[0161] The training set contains 3,000 data samples (1,200 for living room activities, 800 for kitchen cooking, and 1,000 for master bedroom sleeping), while the test set contains 60 samples (20 for each of the three scenarios). During training, the system automatically determines the purification intensity level based on the pollution fluctuation characteristics and health risk levels of different scenarios: when the PM2.5 concentration change rate exceeds 15 μg / m³ / min or the CO concentration change rate exceeds 120 ppm / min, it is determined to be a sudden high pollution state, and the high purification intensity (CADR 480 m³ / h) is immediately implemented; when the pollution change is within a moderate range, a medium purification intensity (CADR 320 m³ / h) is used; and in scenarios with stable air quality and small pollution fluctuations, a low purification intensity (CADR 180 m³ / h) is implemented to balance purification effect and energy consumption. The kitchen cooking scenario is a typical environment of sudden pollution. When the system detects a sharp increase in CO and PM2.5 concentrations in a short period of time, it triggers a high-level purification mode to quickly reduce the pollution concentration. In the living room activity scenario, pollution fluctuations are moderate, and the system maintains a medium-level purification mode. In the master bedroom sleeping scenario, the air is stable, and the system maintains a low-level purification mode to reduce noise and energy consumption. The verification results of the environmental sudden change response mechanism show that when the detected PM2.5 concentration change rate is >15μg / m³ / min or CO concentration change rate is >120 ppm / min, the system can complete emergency path replanning and restore the optimal purification strategy within 1 minute, ensuring the continuity and stability of air quality control under sudden environments.

[0162] Step Six: Performance Verification and Comparative Analysis

[0163] The optimized dynamic path strategy was compared with the traditional fixed path strategy in a comparative test.

[0164] Table 4 compares the purification effects of the two methods.

[0165]

[0166] As shown in Table 4, the method of this application is significantly better than the traditional method in terms of PM2.5 purification efficiency, CO purification efficiency, daily energy consumption and response time to environmental changes. Figures 14 to 19 The system showcases a comparative analysis of purification effects in three typical scenarios (kitchen cooking, living room activities, and master bedroom sleeping). Specifically: in the kitchen cooking scenario ( Figure 14 , 15 The path strategy optimized by the selected feature variables can reduce CO concentration from 45 ppm to 12 ppm and PM2.5 concentration from 85 μg / m³ to 32 μg / m³ within 15 minutes; in a living room activity scenario ( Figure 16 , 17 The CO concentration decreased from 20 ppm to 8 ppm within 15 minutes, and the PM2.5 concentration decreased from 75 μg / m³ to 35 μg / m³; in the master bedroom sleeping scenario ( Figure 18 , 19 The CO concentration decreased from 12 ppm to 5 ppm and the PM2.5 concentration decreased from 60 μg / m³ to 28 μg / m³ within 15 minutes. In contrast, traditional fixed-path methods require approximately 30 minutes to achieve similar purification effects in the same scenarios. The purification effect curves for all scenarios consistently demonstrate that the method of this invention has significant advantages in response speed, purification efficiency, and adaptability to different environmental characteristics, fully verifying the universal effectiveness and robustness of this method in multi-scenario applications.

[0167] Step 7: Continuous System Optimization and Verification

[0168] During the continuous operation phase, based on the aforementioned model update mechanism, the system can autonomously relearn parameters and fine-tune strategies over long periods, achieving continuous performance optimization. After 30 days of continuous operation testing, the system triggered automatic optimization three times, resulting in an overall purification efficiency increase of 15.2% and a 9.7% reduction in energy consumption, verifying the effectiveness and stability of the model update mechanism in a real-world environment.

[0169] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0170] (1) Precise pollution treatment capability: The new air purification adjustment method breaks through the traditional fixed mode and innovatively combines two types of core data - real-time physiological indicators of users (such as respiratory rate, heart rate, etc.) and indoor environmental parameters (such as concentration of harmful gases, distribution of particulate matter, etc.). It can achieve precise location of pollution sources and dynamically adjust the equipment cruise path according to data changes to avoid waste of purification resources.

[0171] (2) Highly sensitive health perception support: The designed health perception system is highly sensitive. It can not only collect a variety of data related to user health, but also perform in-depth analysis of the data, providing core basis for adjusting the purification strategy, so that the purification process is more in line with the user's health needs, rather than relying solely on a single environmental data.

[0172] (3) High-efficiency purification in complex scenarios: By combining the analysis results collected by the health sensing system with the improved dynamic path planning algorithm, the purification problem in complex indoor environments (such as multi-room and multi-obstacle layouts) can be effectively solved, breaking through the limitations of traditional algorithms in low efficiency in complex scenarios and achieving high-efficiency purification coverage.

[0173] (4) Dual optimization of purification efficiency and energy consumption: Experimental verification shows that the purification efficiency of this method is more than 30% higher than that of traditional methods in different user groups (such as the elderly, adults, children, etc.) and diverse scenarios (such as family bedrooms, offices, living rooms, etc.); at the same time, on the basis of high-efficiency purification, the energy consumption of the equipment is significantly reduced, achieving the dual advantages of "high efficiency + low consumption".

[0174] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A dynamic air purification cruise method based on real-time health perception, characterized in that, include: S1. Construct an initial set of feature variables based on the collected multidimensional data; S2. Based on the initial feature variable set, perform a round of feature filtering using an adaptive weight allocation algorithm to obtain the first feature variable set; S3. Based on the first set of feature variables, a second set of feature variables is obtained by performing a second round of feature screening through information entropy and mutual information analysis; S4. Based on the deep reinforcement learning model and the second feature variable set, construct a dynamic path optimization strategy to obtain the optimal cruising path; S5. Based on the optimal cruise path, perform air purification tasks and conduct environmental assessments.

2. The method according to claim 1, characterized in that, The method of step S1 includes: S11. Real-time collection of user physiological data, including respiratory rate and heart rate data, through the smart bracelet worn by the user, and collection of environmental data, including indoor PM2.5 and CO concentration, through the sensors on the air purifier; S12. Based on the user's physiological data and the environmental data, extract real-time monitoring data of all sampling points within a cycle to obtain the cycle-extracted data; S13. Based on the periodic data extraction, obtain an initial set of feature variables; wherein, the initial set of feature variables includes respiratory rate fluctuation index, harmful gas concentration gradient, particulate matter distribution density, and heart rate variability index.

3. The method according to claim 1, characterized in that, The method of step S2 includes: S21. Based on the initial feature variables in the initial feature variable set, obtain the correlation coefficient between each initial feature variable and the target variable; the target variable is the air quality health index; S22. Based on the correlation coefficient, obtain the adaptive weight of each of the initial feature variables; S23. Analyze the contribution of features to the target variable using a random forest regressor, select the critical importance score when the cumulative contribution rate reaches 85% as a threshold, retain the initial feature variables whose adaptive weights are greater than the threshold, and obtain the first feature variable set.

4. The method according to claim 1, characterized in that, The method of step S3 includes: S31. Based on the first feature variables in the first feature variable set, obtain the information entropy of the first feature variables; S32. Based on the information entropy, obtain the mutual information of two adjacent first feature variables; S33. Find the threshold of mutual information through cross-validation; if the mutual information of the two first feature variables is greater than the threshold and their correlation coefficient is greater than 0.8, then retain the first feature variable with a higher mutual information value to the target variable and remove the other; if the mutual information of the two first feature variables is less than the threshold, then remove the corresponding feature variable. S34. Based on the retained first feature variables, obtain the second feature variable set.

5. The method according to claim 1, characterized in that, The method of step S4 includes: S41. By cross-validating the second feature variable set through time series analysis, the samples in the second feature variable set are divided into a training set, a validation set, and a test set in a ratio of 7:2:

1. S42. Construct a learning model based on a deep Q-network, and adopt an architecture that combines competitive deep Q-networks with dual Q-learning to obtain the model loss function; S43. Based on the validation set, the hyperparameters in the model loss function are adjusted using the Bayesian optimization method to obtain the reinforcement learning model; S44. Based on the training set, train the reinforcement learning model to obtain the dynamic path optimization strategy containing the trained reinforcement learning model; S45. Calculate the optimal cruise path based on the dynamic path optimization strategy.

6. The method according to claim 1, characterized in that, The method of step S5 includes: S51. Establish an environmental change response mechanism: When the purifier detects that the rate of change of any environmental parameter exceeds the preset physical reasonable threshold within 30 seconds, it is determined to be an environmental change, the current optimization path is immediately interrupted, and an emergency replanning process is initiated to prioritize the change area, with the response time controlled within 1 minute. S52. Based on the air quality improvement rate, health indicator satisfaction, and energy consumption efficiency, evaluate the purification effect and establish a model update trigger mechanism; S53. Based on the evaluation results, dynamically adjust the parameters and feature weights of the dynamic path optimization strategy, and continuously optimize the reinforcement learning model using an online learning algorithm.

Images