A multi-dimensional data fusion indoor environment quality intelligent evaluation method

Abstract

The present application relates to the technical field of environmental parameter measurement, and more particularly to a multi-dimensional data fusion indoor environmental quality intelligent evaluation method, comprising: step 1: multi-source data space-time synchronous acquisition; step 2: dynamic correlation model construction, generating a parameter coupling matrix based on time delay cross-correlation analysis between parameters, the time delay cross-correlation analysis is realized by sequence translation and correlation coefficient calculation in a sliding time window, and an environment state class is constructed according to a parameter mean vector, fluctuation intensity and abnormal event marker, and each environment state class is bound to an independent parameter contribution weight set; step 3: two-stage environmental quality evaluation; step 4: adaptive incremental optimization, when the deviation between the comprehensive evaluation index and the subjective evaluation exceeds the threshold value, the parameter contribution weight set of the environment state class is adjusted, and the parameter coupling matrix is periodically reconstructed. By considering the time delay coupling between parameters, the prediction ability is improved; it has the ability of self-learning and optimization, and the stability of long-term operation of the system is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental parameter measurement, and in particular to an indoor environmental quality intelligent evaluation method based on multi-dimensional data fusion. BACKGROUND

[0002] Indoor environmental quality (IEQ) evaluation is a core technology in the fields of intelligent buildings, healthy living, etc. The existing evaluation methods mainly have the following limitations:

[0003] The current mainstream technology relies on the independent threshold judgment of parameters such as temperature, humidity, PM2.5, etc. However:

[0004] The dynamic coupling effect between parameters is ignored: for example, temperature and humidity jointly affect the reproduction rate of microorganisms (when the relative humidity is greater than 60%, the bacterial proliferation rate increases by 30% for every 5℃ increase in temperature), but the existing method does not establish such a cross-parameter correlation model;

[0005] Consequences: In complex scenarios such as hospital operating rooms, only controlling a single parameter to meet the standard may still lead to the risk of exceeding the microorganism standard, and the evaluation result deviates significantly from the true environmental state.

[0006] Reasonableness of the problem: Indoor environment is a complex system of multi-parameter interaction, and independent evaluation violates the laws of environmental physics, which is particularly prominent in health-sensitive places.

[0007] In addition, the existing technology also uses a fixed weight fusion scheme, which still has inherent defects:

[0008] Unable to respond to dynamic events: when sudden gathering of people causes a sharp rise in CO2 concentration, the CO2 weight does not increase dynamically, making the evaluation value lag behind the actual pollution level;

[0009] For example, when the CO2 concentration in a conference room rises from 600ppm to 1500ppm during use, the evaluation value of the fixed weight model changes by only 15%, while the actual subjective dissatisfaction increases by 70%. Since unexpected events occur frequently in building environments, static models are difficult to meet the real-time health protection needs.

[0010] Therefore, there is an urgent need for an indoor environmental quality intelligent evaluation method based on multi-dimensional data fusion to solve the above problems. SUMMARY

[0011] Based on the above purpose, the present application provides an indoor environmental quality intelligent evaluation method based on multi-dimensional data fusion, comprising the following steps:

[0012] Step 1: Time and space synchronization of multi-source data, temperature value, humidity value, PM2.5 concentration value, CO2 concentration value, TVOC concentration value, and light intensity value are synchronously collected by a distributed sensor network, and spatial position coordinates and synchronous time stamps are attached;

[0013] Step 2: Dynamic correlation model construction, based on the time delay cross-correlation analysis between parameters, the parameter coupling matrix is generated, the time delay cross-correlation analysis is realized by sequence translation and correlation coefficient calculation in a sliding time window, and the environment state class is constructed according to the parameter mean vector, fluctuation intensity and abnormal event marker, and each environment state class is bound with an independent parameter contribution weight set;

[0014] Step 3: Two-stage environmental quality assessment, real-time state assessment: match the current data to the environment state class, calculate the real-time quality index based on the parameter contribution weight set, trend evolution assessment: predict the future trend of change using the parameter coupling matrix, calculate the trend risk index combined with the parameter sensitivity coefficient, generate the comprehensive evaluation index by fusing the real-time quality index and the trend risk index, and the fusion coefficient is dynamically adjusted according to the change of the environment state;

[0015] Step 4: Adaptive incremental optimization, when the deviation between the comprehensive evaluation index and the subjective evaluation exceeds the threshold, adjust the parameter contribution weight set of the environment state class, and periodically reconstruct the parameter coupling matrix.

[0016] The beneficial effects of the present application are:

[0017] 1. By introducing time delay cross-correlation analysis based on sliding time window, the coupling relationship matrix between multiple parameters is established, effectively revealing the synergistic or inverse effect between parameters such as temperature, humidity, CO2 and TVOC.

[0018] 2. By constructing environment state categories, combining historical fluctuation characteristics and abnormal event markers, the fine classification of typical environment states is realized, and independent parameter contribution weights are configured for different states. This can avoid the one-size-fits-all evaluation model and achieve targeted evaluation. The mechanism of binding environment state and evaluation weight significantly improves the judgment accuracy in complex scenarios, making the evaluation value closer to the actual perception and health risk degree of human body.

[0019] 3. The present application adopts a dynamic weight mechanism, which adaptively selects a parameter weight set according to the currently matched environment state class. When a sudden event such as CO2 surge occurs, the system can immediately match to a high CO2 sensitive state class and automatically increase the evaluation weight of the parameter, so that the comprehensive evaluation value quickly responds to the pollution change, overcoming the problem of response lag of traditional fixed weight model.

[0020] 4. By introducing an adaptive incremental optimization mechanism, when there is a deviation between user's subjective perception and comprehensive evaluation value, the parameter weight set is automatically corrected, and the parameter coupling structure is periodically reconstructed, so that the model continuously approaches the real feeling of individuals or groups. This optimization strategy not only improves the individualized adaptability of evaluation, but also ensures the stability and user satisfaction in the long-term use. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the present application or the prior art, the drawings needed to be used in the following embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without any creative effort.

[0022] Fig. 1 Flow chart for the step of calculating spatial influence weight for each sensor in the method of the present application;

[0023] Fig. 2 Flow chart for the step of calculating spatial influence weight for each sensor in the method of the present application;

[0024] Fig. 3 Flow chart for the step of calculating the trend risk index in step 3 of the method of the present application, including the calibration of the parameter sensitivity coefficient. DETAILED DESCRIPTION

[0025] The present application will be described in detail below with reference to the drawings and specific embodiments. It should be noted here that, in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement some known technologies; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.

[0026] Please refer to Figs. 1-3 The embodiment of the present application provides a multi-dimensional data fusion indoor environment quality intelligent evaluation method. First, in the data collection stage, through the deployment of multiple types of sensors at different positions in the room, environmental parameters such as temperature, humidity, PM2.5, CO2, TVOC, and light intensity are collected in real time and synchronously. Each piece of data is attached with specific collection time and spatial coordinates to realize unified time and space identification. By setting a unified synchronization period and a network clock calibration mechanism, the consistency of all data sources on the time axis is ensured, effectively supporting subsequent correlation analysis.

[0027] Subsequently, in the dynamic correlation modeling stage, based on the time dependence between multiple parameters, sequence analysis is performed through sliding time window technology. For any two parameters, by sliding the time series of one parameter, the correlation coefficient of the other parameter at each time translation position is calculated, and the maximum correlation coefficient and the corresponding time offset are extracted as the coupling strength and dominant time lag value of the parameter pair. The analysis results of all parameter pairs are summarized as a parameter coupling matrix. Combined with the historical parameter mean value, fluctuation intensity and abnormal marker, an environment state category is constructed. Each state category is bound to an independent parameter contribution weight set to reflect the relative importance of different parameters to the environment quality in that environment state.

[0028] In the quality assessment link, a two-stage evaluation mechanism is designed. First, the current collected data is matched to the most similar environmental state category in real time, and the weight set under this category is used to weight the parameter values, and the real-time environmental quality index is output. Subsequently, according to the parameter coupling matrix, the future evolution trend of each parameter is predicted, and combined with the parameter sensitivity index, the trend risk index that may cause the decline of environmental quality is calculated. Finally, the real-time index and the risk index are weighted and fused according to the state adaptive fusion coefficient to form the comprehensive evaluation value. The fusion coefficient is dynamically adjusted according to the environmental stability, and the real-time index is focused on in the stable state, and the trend risk is paid attention to in the fluctuation state.

[0029] Finally, when the comprehensive evaluation value and the user's subjective feeling have a significant difference (such as the deviation exceeds the set threshold), the parameter weight adjustment mechanism will be triggered to further refine the subjective experience. At the same time, the environmental state category change and the parameter coupling structure stability are periodically checked, and the coupling matrix is reconstructed if necessary, to improve the overall evaluation accuracy and adaptability.

[0030] The present application realizes the dynamic perception and intelligent evaluation of the indoor complex environment state, can accurately reflect the current state and warn the potential risk; by considering the time delay coupling between parameters, the prediction ability is improved; at the same time, it has the ability of self-learning and optimization, which improves the stability and user satisfaction of long-term operation of the system.

[0031] In one possible implementation, first, the spherical neighborhood radius of each sensor needs to be calculated according to the volume of the indoor space and the distribution density of the sensor. The larger the volume, the more sparse the distribution of the sensor, and the neighborhood radius should be increased accordingly; on the contrary, the higher the density of the sensor, the neighborhood radius is correspondingly reduced. This calculation process ensures that the influence of each sensor is reasonably considered under different spatial layouts, avoiding inaccurate evaluation caused by too concentrated or too dispersed sensor settings.

[0032] Then, according to the Euclidean distance principle, the influence weight of other sensors in the neighborhood on the target point is calculated. The weight decreases exponentially with the increase of distance, which means that the farther the sensor is from the target point, the smaller the influence of the data, and vice versa. This decay law conforms to the signal transmission characteristics in the actual physical environment, and can truly reflect the influence range of the sensor in space.

[0033] Since sensors may have abnormal data due to faults, environmental factors or other reasons during operation, the introduction of the signal quality factor is used to correct the initial weight component. The factor is determined by calculating the abnormal rate of the historical data of the sensor. The higher the abnormal rate, the smaller the correction coefficient, indicating that the data of the sensor is less reliable, and its influence on the target point will also be weakened accordingly. This correction mechanism ensures that the model can more accurately reflect the real environment state when there are unreliable sensors, avoiding misleading evaluation results caused by unreliable data.

[0034] Finally, the calculated spatial influence weight will be used for subsequent data repair and spatial interpolation. In the data repair process, the spatial influence weight of the sensor helps to select appropriate sensor data for fusion and correction, ensuring the accuracy and consistency of the data. In the spatial interpolation process, the weight setting helps to accurately estimate the environmental data at positions not directly measured, improving the comprehensiveness and accuracy of the evaluation method.

[0035] The implementation of this technical feature effectively solves the evaluation error caused by uneven distribution of sensors or signal quality differences. By introducing the spatial influence weight, not only the data repair accuracy is improved, but also the adaptability and reliability of the evaluation method in complex spatial environments are enhanced. For dynamic changing environments, this weight calculation method based on physical distance and sensor quality significantly improves the authenticity and accuracy of the evaluation, avoiding the "data distortion far from the sensor position" problem that may exist in traditional evaluation methods, while effectively dealing with the uncertainty caused by sensor failure or poor signal, thereby improving the stability and accuracy of the entire evaluation process.

[0036] In one possible implementation, first, two groups of parameter sequences A and B are processed. In this process, sequence A will be translated in a sliding time window with a fixed step size. The choice of the length of the sliding time window and the step size is critical, as it will directly affect the accuracy and computational efficiency of the analysis results. By translating sequence A, the correlation between sequence A and sequence B at different time points can be evaluated step by step, ensuring a comprehensive analysis of the potential time lag between the two.

[0037] After each translation, the correlation coefficient between the translated sequence A and sequence B is calculated. The correlation coefficient is a measure of the linear correlation between two sequences, with a value range of -1 to 1. The closer the value is to 1, the closer the relationship between the two sequences. By calculating the correlation coefficients at all translation steps, the maximum value is selected as the coupling strength value of the parameter pair, reflecting the strongest coupling relationship between sequence A and B at a certain time. The selection of this maximum value means that the coupling strength at this time is the strongest, which helps to further identify the interaction pattern between the sequences.

[0038] After finding the maximum correlation coefficient, the step number of the translation step at this time is recorded and converted into a time unit, and this time value is the dominant time lag value. The dominant time lag value represents the time deviation between the two sequences when the maximum correlation occurs, reflecting the time delay of the influence of one parameter change on another parameter. The time lag value is an important parameter for describing the timing relationship between the two, which helps to better understand the timing coupling characteristics between them.

[0039] Finally, the trigger condition for matrix update is defined. Specifically, if the rate of change of the number of environmental state classes exceeds the set proportion threshold, or the fluctuation amplitude of the dominant time lag value within the same class exceeds the stability threshold, the reconstruction of the analysis results is triggered. This triggering mechanism ensures that the analysis model can be adjusted in time when the environmental state changes significantly, thereby improving the adaptability and accuracy of the model.

[0040] By sliding the time window and fixed step translation, the timing correlation between sequences at multiple time points can be captured, making the time lag analysis more detailed and comprehensive. Secondly, the selection of the maximum correlation coefficient effectively reflects the strongest coupling relationship between parameters, ensuring the accuracy of the time lag analysis. By recording the dominant time lag value, the timing relationship between parameters can be effectively revealed, providing a reliable basis for subsequent dynamic adjustment and prediction. In addition, the trigger condition for matrix update ensures the flexibility and dynamic response capability of the analysis when the environment changes, avoiding data distortion caused by environmental fluctuations, making the overall analysis results more stable and reliable. These steps work together to greatly improve the practicality of the time lag analysis method, effectively improving the modeling and prediction ability of complex environmental systems in practical applications.

[0041] In one possible implementation, first, the fluctuation intensity of each parameter in the historical data is calculated, which usually refers to the degree of change of the parameter over time. By calculating the fluctuation intensity of each parameter in the historical data, the next step is to perform quantile analysis, that is, to sort and segment the fluctuation intensities into several intervals. Through quantile analysis, the distribution of parameter fluctuation intensity can be determined, providing data basis for subsequent cluster radius determination.

[0042] In the density clustering algorithm, the clustering radius is an important parameter that determines whether the data points belong to the same class. To determine the clustering radius, this method uses the quantile method to select the initial radius. The specific steps are to select the fluctuation intensity value of the N quantile and multiply it by a spatial topology complexity coefficient as the initial radius of clustering. The N quantile value can be adjusted according to different needs to adapt to different data distribution characteristics. The spatial topology complexity coefficient is a compensation for the complexity of the data space structure, which can help adjust the adaptability of the initial radius, so that the density clustering can better reflect the true structure of the data.

[0043] In the density clustering process, it is necessary to dynamically evaluate whether a new environment state class should be added. The basis for adding a class is that the weighted Euclidean distance between the current data vector and all existing class centers is greater than a certain set multiple of the maximum distance within the class. The weighted Euclidean distance is a measure of the distance between data points, in which the weight of the parameter is weighted according to the inverse of its historical fluctuation amplitude. That is, the parameter with larger historical fluctuation amplitude is given smaller weight in distance calculation, which can effectively reduce its influence on the clustering result, so as to ensure that the clustering result is more stable and accurate.

[0044] Through the above steps, the environment state class can be dynamically updated every time new data arrives. When the weighted Euclidean distance between the current data vector and the existing class center is large, it means that the data is significantly different from the existing environment state class, so it can be considered as a new environment state class for division. This can ensure that the clustering result adapts to the change of data, avoids excessive fusion between classes, and ensures that each environment state class can accurately reflect certain environmental characteristics.

[0045] By introducing the density clustering algorithm and determining the clustering radius based on the fluctuation intensity quantile, the dependence on parameter selection and radius setting in traditional clustering methods is avoided. The combination of fluctuation intensity value calculated by quantile and topological complexity coefficient makes the clustering radius have stronger adaptability, which can automatically adjust according to the actual distribution of data, thereby effectively improving the clustering precision and stability. At the same time, the strategy of weighted Euclidean distance minimizes the influence of parameters with large historical fluctuation amplitude on the clustering result, ensuring that the clustering process is not disturbed by individual abnormal data, and the clustering effect is more in line with actual needs. In addition, the mechanism of dynamically updating the environment state class ensures that the method can adjust the clustering result in time when facing changing data, and maintains the accuracy and timeliness of data classification.

[0046] In one possible implementation, first, the data set of extreme events needs to be extracted from historical data. Extreme events refer to the largest fluctuation or abnormal change events that have occurred in the past historical data. These events usually cause a dramatic change in the state of the system, so detailed analysis of them can help identify which parameters play a key role in the system. In this step, the maximum fluctuation amplitude of each parameter is recorded, and the subjective evaluation drop corresponding to each event is also recorded. The subjective evaluation drop is usually given by experts or experienced evaluators to describe the overall impact of the event on the system.

[0047] After obtaining the historical extreme event dataset, the next step is to fit the mapping relationship between parameter volatility amplitude and subjective evaluation reduction through machine learning algorithms. The purpose of this step is to discover the regularity between each parameter volatility intensity and its actual impact on the system (i.e., reduction). Typically, regression analysis or other supervised learning algorithms are used to establish a relationship model between volatility amplitude and reduction, generating a sensitivity base value for each parameter. This base value reflects the sensitivity of each parameter to changes in volatility, which is the basis for calculating the trend risk index.

[0048] After generating the sensitivity base value, it needs to be weighted and corrected according to the coupling relationship between parameters. The coupling matrix reflects the strength and direction of the mutual influence between different parameters. In this process, first, calculate the average correlation strength of each parameter in the coupling matrix, i.e., the correlation degree between a certain parameter and other parameters. Then, the sensitivity base value is weighted and corrected with the average correlation strength of the parameter, thus obtaining a more accurate sensitivity value. This weighted correction can take into account the interaction between different parameters, further improving the accuracy of the prediction.

[0049] Finally, the chain propagation model is used to predict the change rate of trend risk. In this process, the dominant time lag value is used to capture the time delay effect of parameter changes in the system. The chain propagation model simulates the influence of parameter changes in the system on the final trend risk index, taking into account the influence, sensitivity, and time lag effect between parameters, thus being able to predict the future trend risk of the system. Through this method, the complex time dependence and mutual influence in the system can be better captured, providing more accurate risk prediction.

[0050] This method accurately calibrates the sensitivity of each parameter, making the prediction of trend risk more accurate. By extracting historical extreme events and using machine learning techniques, the deep connection between parameter volatility and actual system impact can be revealed, optimizing the calculation and correction process of sensitivity. The weighted correction of the sensitivity base value takes into account the interaction between parameters, making the risk prediction more consistent with the actual operation of the system. The use of the chain propagation model and the dominant time lag value further optimizes the time accuracy of risk prediction and the response speed of change trend. In summary, this method can provide an efficient, dynamic, and accurate solution for trend risk prediction, which is of great significance for identifying potential risks in advance and adjusting decisions in a timely manner.

[0051] In one possible implementation, first, one of the conditions for increasing the real-time evaluation item coefficient is that the duration of the environmental state class exceeds the stability threshold determined by the historical average duration of the class. The environmental state class refers to the state characteristics exhibited by the environment under certain conditions, while the duration refers to the length of time the environment remains in that state. To determine whether an environmental state has stabilized, the historical average duration can be calculated and a stability threshold can be set. If the current environmental state duration exceeds this threshold, it is considered that the current environmental state has stabilized, and the coefficient can be further adjusted so that the evaluation item is more in line with the changing trend of the current environment.

[0052] Second, the growth rate of the trend risk index needs to be lower than the sensitivity threshold set by the characteristics of the environmental state class. This condition determines the speed of change of the trend risk. If the growth rate of the trend risk index is low, it means that the risk of the system changes slowly and there are no sharp fluctuations. In this case, the real-time evaluation item coefficient can be increased moderately to strengthen the attention to the current state, so as to more accurately predict the future risk changes.

[0053] In actual operation, the increase condition of the trend evaluation item coefficient is opposite to that of the real-time evaluation item coefficient. That is, when the real-time evaluation item coefficient increases, the trend evaluation item coefficient decreases accordingly, and vice versa. Moreover, the sum of the two is always 1. This means that during the adjustment process, the weight of the evaluation item is always in balance, thereby avoiding excessive bias to a certain coefficient. In this way, it can be ensured that the weight of the evaluation item is dynamically adjusted under different environmental conditions to adapt to different risk assessment needs.

[0054] By setting the increase condition of the real-time evaluation item coefficient, the method can dynamically adjust the evaluation coefficient to more accurately reflect the changes in the environmental state and the evolution of the trend risk. Specifically, when the environmental state is stable and the risk changes slowly, increasing the real-time evaluation item coefficient helps to maintain high sensitivity to existing trends, thereby improving the prediction accuracy of future changes. In the case of rapid changes in trend risk, the increase of the trend evaluation item coefficient helps to quickly capture environmental changes, thereby effectively responding to sudden risks. Since the sum of the real-time evaluation item coefficient and the trend evaluation item coefficient is always 1, the problem of unbalanced evaluation item weight is avoided, making the entire evaluation process more scientific and stable.

[0055] In one possible implementation, first, in the same environment state class, the data is divided into multiple sub-intervals according to the time sequence. The division of these sub-intervals is usually based on the continuity of time and its influence on the change of the dominant time lag value. These sub-intervals represent different periods of time, and the time lag value in each sub-interval has a certain stability. In each sub-interval, the environment state remains basically unchanged, and the changes of various indicators are relatively regular. Therefore, by dividing the time into multiple sub-intervals, the fluctuation range of the dominant time lag value in different time periods can be independently evaluated, thereby avoiding the interference of the fluctuation of the global data on the local evaluation.

[0056] Next, the standard deviation of the dominant time lag value in each sub-interval is calculated. The standard deviation is a statistical measure of the fluctuation or dispersion of data. In this step, the standard deviation is used to quantify the fluctuation range of the dominant time lag value in each sub-interval. If the standard deviation of the dominant time lag value in a sub-interval is large, it indicates that there is a large fluctuation of the dominant time lag value in that interval, which may reflect some unstable factors; otherwise, it indicates that the time lag value in that interval changes less, and the system is relatively stable.

[0057] Finally, when the standard deviation of the dominant time lag value in consecutive multiple sub-intervals exceeds the preset threshold value for a certain number of times, it is determined that the fluctuation range is abnormal. This determination condition can help to identify the abnormal situation of the fluctuation of the dominant time lag value in time. The abnormal fluctuation range may indicate that the environment state has changed unusually, or the system has some abnormal response. If the fluctuation range of some sub-intervals is continuously abnormal, it indicates that there are factors in the environment or system that affect the change of the dominant time lag value, which needs to be paid enough attention.

[0058] Through this fluctuation range evaluation method, the change of the dominant time lag value can be accurately monitored, and the fluctuation that may affect the stability of the system can be found in time. In practical applications, the dominant time lag value often plays an important role in risk assessment and environment prediction. By quantifying the fluctuation range and setting a threshold to judge abnormality, it can ensure that the system can respond quickly when abnormal fluctuation occurs, and avoid making wrong judgments or decisions due to excessive fluctuation. Therefore, this technical feature can effectively improve the accuracy of environment state evaluation and risk prediction, improve the ability to respond to emergencies, reduce the risk in system decision-making, and ensure the stable operation of the system in complex environments.

[0059] In one possible implementation, to solve the problem of missing parameters due to lack of sensors at a certain point P in space, the method proposes the following multi-level estimation process:

[0060] Firstly, a spherical neighborhood with adaptive radius is dynamically constructed with the point P as the center. The radius is determined based on the distribution density of regional sensors and data availability. If the sensor distribution is sparse, the radius is appropriately increased to ensure that enough valid sensor data sources are covered in the neighborhood. This approach avoids the problem of fixed radius failure in sparse areas and achieves flexibility in spatial range.

[0061] After determining the neighborhood, all sensors within it are retrieved, and their corresponding measurement values are extracted. Subsequently, based on factors such as spatial distance between each sensor and point P, environmental similarity, and historical stability, the spatial influence weight of each sensor on point P is calculated. For example, the inverse distance principle or exponential decay model. Finally, by weighting and summing the measurement values of each sensor with their corresponding weights, the estimated value of point P is obtained.

[0062] If there are no valid sensors within the set neighborhood, even if the neighborhood radius is expanded, no valid data sources are obtained. At this time, the cross-regional compensation mechanism is enabled. This mechanism first analyzes the movement trajectory data of the personnel in the region and extracts the proportion of their stay time in the target region. Subsequently, from other regions with high behavioral relevance to the target region, the corresponding parameter data is extracted, and the measurement values are fused with the stay time of each region as the weight to estimate the parameters of the target region. This compensation method not only fills the spatial data gap but also introduces crowd behavior as an important reference dimension, enhancing the semantic reasonableness of data estimation.

[0063] In one possible implementation, first, the similarity retrieval technique is used to analyze similar categories in historical environmental data to the current environmental state. By comparing the attributes of the current environment and the historical state categories, the similarity between the two is calculated, and the state category closest to the current environment is selected. For these historical state categories, the weight values of their related parameters are extracted as the initial weights. This way, the experience and rules in historical data can be used to provide reasonable weight initialization for the current environment, avoiding the instability caused by random initialization.

[0064] In some cases, it may not be possible to find a similar historical category to the current environmental state. At this time, the method uses a global regression model of parameters and subjective evaluation to generate initial weights. Through global analysis and regression modeling of historical data, combined with the feedback of subjective evaluation, the initial parameter weights suitable for the current environment are generated. This strategy ensures that even if there is no direct similar class for reference, appropriate weights can be obtained through global modeling, avoiding the initialization problem caused by the lack of historical data.

[0065] For new categories of environment, the initial sample can be less, so the weight initialization can face the problem of inaccuracy. To solve this problem, in the initial stage, the step of weight adjustment will be dynamically reduced according to the number of samples. Specifically, when the number of samples is small, the adjustment range of the weight is large, so as to quickly adapt to new data; with the increase of the number of samples, the step is gradually reduced, so as to avoid the instability caused by excessive adjustment. Such a method ensures that the new class has enough flexibility in the initial stage, and gradually stabilizes with the accumulation of data, and finally realizes more accurate weight allocation.

[0066] In one possible implementation, at each round of weight update, the weight change trend in the previous iteration or iterations is retained and introduced as a momentum term into the current weight calculation. The introduction of the momentum term can effectively alleviate the sharp fluctuations caused by sample discreteness or sudden data, thereby realizing "inertia control" in the update process. The initial value of the momentum coefficient is set to be high, giving the historical change trend sufficient influence, but as the number of samples of this category increases, the momentum coefficient decreases according to a predetermined strategy, which is usually negatively correlated with the total number of samples. This mechanism can realize the natural transition from relying on historical trends in the initial stage to relying more on current data in the later stage.

[0067] To prevent the newly introduced data from causing the deviation of the internal structure of the weight vector, the method imposes an intra-class consistency constraint on the weight gradient corresponding to the new data. This constraint term constructs a local consistency regularization term by analyzing the distribution of intra-class samples in the feature space, forcing the new data to maintain consistency with the existing data in direction and scale when adjusting its parameter contribution. The constraint strength is designed to be inversely proportional to the square root of the number of samples in the class, and the more the number of samples, the smaller the constraint strength, so as to give higher degree of freedom when the data is rich, and to strengthen the local consistency when the data is scarce.

[0068] The present application encompasses any alternatives, modifications, equivalent methods and schemes made on the essence and scope of the present application. In order for the public to have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details by those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0069] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can also be made, which should be considered as the protection scope of the present application.

Claims

1. A multi-dimensional data fusion indoor environment quality intelligent evaluation method, characterized in that, Comprise the following steps: Step 1: Multi-source data space-time synchronous acquisition, through distributed sensor network synchronous acquisition temperature value, humidity value, PM2.5 concentration value, CO2 concentration value, TVOC concentration value, light intensity value, and additional spatial position coordinates and synchronous time stamp; Step 2: Dynamic correlation model construction, based on the time delay cross-correlation analysis of parameters to generate parameter coupling matrix, the time delay cross-correlation analysis is realized by sequence translation and correlation coefficient calculation in sliding time window, according to parameter mean vector, fluctuation intensity and abnormal event marking to construct environment state class, each environment state class binds independent parameter contribution weight set; The specific process of time delay cross-correlation analysis in step 2 includes: a. To two parameter sequences A and B, sequence A is translated in a sliding time window with fixed step; b. Calculate the correlation coefficient of sequence A and sequence B after translation, take the maximum value of correlation coefficient as the coupling strength value of the parameter pair; c. Record the translation step number when the maximum correlation coefficient is reached, convert it into time unit as the dominant time delay value; Matrix update trigger condition: when the change rate of environment state class number exceeds the set proportion threshold, or the fluctuation amplitude of dominant time delay value in the same class exceeds the stability threshold, start reconstruction; The fluctuation amplitude of dominant time delay value is evaluated by the following way: In the same environment state class, divide multiple subintervals in time sequence; Calculate the standard deviation of dominant time delay value in each subinterval; When the standard deviation exceeds the set threshold continuously for a certain number of times, it is judged as abnormal fluctuation amplitude; Step 3: Two-stage environmental quality evaluation, real-time state evaluation: match the current data to the environment state class, calculate the real-time quality index based on the parameter contribution weight set, trend evolution evaluation: predict the future trend by using parameter coupling matrix, calculate trend risk index combined with parameter sensitivity coefficient, generate comprehensive evaluation index by fusing real-time quality index and trend risk index, and adjust the fusion coefficient dynamically according to the change of environment state; Step 4: Self-adaptive incremental optimization, when the deviation between comprehensive evaluation index and subjective evaluation exceeds the threshold, adjust the parameter contribution weight set of environment state class, and periodically reconstruct parameter coupling matrix. 2.The method of claim 1, wherein, Step 1 also includes calculating the spatial influence weight for each sensor, the calculation process is: a. Determine the spherical neighborhood radius according to the indoor space volume and sensor distribution density; b. Calculate the initial weight component of each sensor to the target point in the neighborhood, the initial weight component decays according to negative exponential law with the increase of Euclidean distance; c. Introduce signal quality factor to modify the initial weight component, the signal quality factor is calculated based on the historical data anomaly rate of the sensor, the higher the anomaly rate, the smaller the correction coefficient, the spatial influence weight is used for subsequent data repair and spatial interpolation. 3.The multi-dimensional data fusion-based indoor environment quality intelligent evaluation method according to claim 1, characterized in that, The clustering process in step 2 of building environment state class adopts density clustering algorithm, the determination method of clustering radius in density clustering algorithm: a. Calculate the quantile of parameter fluctuation intensity in historical data; b. Take the N quantile value multiplied by the spatial topology complexity coefficient as the initial radius; The condition for adding a new environmental state class is that the weighted Euclidean distance between the current data vector and all existing class centers is greater than a set multiple of the maximum intra-class distance, and the parameter weight in the weighted Euclidean distance is the inverse of the historical fluctuation amplitude. 4.The method of claim 1, wherein, The calculation of the trend risk index in step 3 includes the calibration of the parameter sensitivity coefficient, and the specific process includes: a. Extract the historical extreme event data set, record the maximum fluctuation amplitude of each parameter and the corresponding subjective evaluation reduction; b. Through machine learning, the mapping relationship between parameter fluctuation amplitude and subjective evaluation reduction is fitted to generate the sensitivity basis value of each parameter; c. According to the average correlation strength of the parameter in the coupling matrix, the sensitivity basis value is weighted and corrected, and the calculation of the change rate uses a chain propagation model based on the dominant time lag value. 5.The multi-dimensional data fusion-based indoor environment quality intelligent evaluation method according to claim 1, characterized in that, The increase condition of the real-time evaluation item coefficient in the dynamic adjustment of the fusion coefficient with the change of the environmental state also includes: a. The duration of the current environmental state class exceeds the stability threshold determined by the historical average duration of the class; b. The growth rate of the trend risk index is lower than the sensitivity threshold set by the characteristics of the environmental state class; The increase condition of the trend evaluation item coefficient is opposite to that of the real-time evaluation item coefficient, and the sum of the two is always 1. 6.The multi-dimensional data fusion-based indoor environment quality intelligent evaluation method according to claim 2, characterized in that, The specific process of using spatial influence weight for spatial interpolation includes: Parameter value estimation for sensorless position point P: a. Retrieve all sensors within the spherical neighborhood centered at P with an adaptively determined radius; b. Calculate the spatial influence weighted sum of each sensor measurement, and the weight is calculated according to claim 2; c. When there is no valid sensor in the neighborhood, start cross-region compensation, which includes: Obtain the proportion of the residence time of the personnel movement trajectory in the region; Fuse the data of the associated region with the residence time as the weight. 7.The multi-dimensional data fusion-based indoor environment quality intelligent evaluation method according to claim 3, characterized in that, The parameter weight initialization in the weighted Euclidean distance includes: Based on the similarity retrieval of the historical similar environmental state class, the weight is used as the initial value; If there is no similar class, use the global regression model of the parameter and the subjective evaluation to generate the initial weight; During the period of insufficient initial samples of the new class, the weight adjustment step is reduced by the proportion of the sample quantity. 8.The multi-dimensional data fusion-based indoor environment quality intelligent evaluation method according to claim 1, characterized in that, The parameter contribution weight set of the environmental state class in step 4 also includes a momentum constraint mechanism: Retain the momentum term of the historical weight set, and the momentum coefficient decreases with the increase of the data quantity of the class; Apply intra-class consistency constraint to the weight gradient of new data, and the constraint strength is inversely proportional to the square root of the sample quantity of the class.

Images