CN-121980263-A - Machine learning-based two-stage indoor personnel occupation detection method

CN121980263ACN 121980263 ACN121980263 ACN 121980263ACN-121980263-A

Abstract

A two-stage indoor personnel occupation detection method based on machine learning comprises the steps of obtaining multi-mode time sequence data comprising a plurality of original data packets, enabling the original data packets to comprise sampling time stamps and physical sensor feature vectors, performing time removal on the original data packets, obtaining time sequence derived features and environment parameter interaction features, combining the time sequence derived features and the environment parameter interaction features with the physical sensor feature vectors to form a high-dimensional physical feature matrix, screening the high-dimensional physical feature matrix by utilizing a pearson correlation coefficient and a recursive feature elimination algorithm to remove redundant information to generate an optimal feature subset, generating feature vectors to be detected according to the optimal feature subset, selecting a presence detection model and a person number regression detection model in a candidate model set to construct a two-stage cascade model, performing binary judgment on whether a person exists or not by using the presence detection model, inputting the feature vectors to be detected into the person number regression detection model if the person exists, obtaining a person detection result, and outputting the number of persons to be zero if the person is judged to exist.

Inventors

XIA YUDONG
LI HEYU
LI YUN

Assignees

杭州电子科技大学

Dates

Publication Date: 20260505
Application Date: 20260119

Claims (10)

1. The two-stage indoor personnel occupation detection method based on machine learning is characterized by comprising the following steps of: S1, acquiring multi-mode time sequence data of an office comprehensive environment sensor, wherein the multi-mode time sequence data comprises a plurality of original data packets, and the original data packets comprise sampling time stamps and physical sensor feature vectors; S2, performing time elimination on the original data packet to obtain a time sequence derivative characteristic and an environment parameter interaction characteristic; s3, screening the high-dimensional physical feature matrix to remove redundant information, and generating an optimal feature subset; S4, generating a feature vector to be detected according to the optimal feature subset, selecting a presence detection model and a population regression detection model from the constructed candidate model set to construct a two-stage cascade model; S5, inputting the feature vector to be detected into the selected presence detection model to carry out binary judgment on whether people exist, if yes, inputting the feature vector to be detected into the selected people number regression detection model to obtain a person detection result, otherwise, outputting zero number of persons.
2. The method for detecting occupancy of two-stage indoor personnel based on machine learning of claim 1, wherein the physical sensor feature vector comprises the physical quantities of indoor carbon dioxide concentration, volatile organic compound concentration, noise decibels, indoor and outdoor relative humidity, indoor and outdoor temperature and illumination intensity.
3. The machine learning-based two-stage indoor personnel occupation detection method according to claim 1, wherein S1 comprises the steps of utilizing sampling time stamps to carry out integrity detection on multi-mode time sequence data, setting a fault judgment threshold value based on a sampling period of a physical sensor feature vector, traversing the multi-mode time sequence data, judging that a data fault exists at the position if the time stamp interval of adjacent original data packets is detected to be larger than the fault judgment threshold value, dividing data before and after the data fault into independent continuous subsequences, and carrying out time elimination on the continuous subsequences.
4. The machine learning based two-stage indoor personnel occupancy detection method of claim 1, wherein S2 comprises: S21, calculating dynamic changes of physical quantities through a relative time sequence relation determined by a sampling time stamp, so as to construct time sequence derivative characteristics comprising hysteresis characteristics, difference characteristics and sliding window statistical characteristics, wherein the hysteresis characteristics are sensor readings of a preset time step before the current moment; S22, constructing environmental parameter interaction characteristics for representing nonlinear coupling change generated by personnel existence among environmental physical quantities according to a multi-physical field coupling principle, wherein the environmental parameter interaction characteristics comprise acoustic metabolism coupling characteristics for capturing a sound-heat synchronization effect and a CO-occurrence rule of noise decibels and indoor CO 2 concentration, metabolism heat-humidity coupling characteristics for capturing a metabolism and heat-humidity linkage effect, indoor and outdoor environment difference value characteristics for representing relative change caused by internal personnel activities and multi-source heterogeneous comprehensive indexes for representing the overall environment state.
5. The machine learning-based two-stage indoor personnel occupancy detection method of claim 4, wherein the acoustic metabolic coupling features comprise acoustic temperature interaction features and acoustic carbon interaction features, the metabolic thermal-wet coupling features comprise carbon Wen Jiaohu features, carbon-wet interaction features and carbon-volatile interaction features, the indoor-outdoor difference features comprise indoor-outdoor difference features and indoor-outdoor difference features, the indoor-outdoor difference features are respectively an indoor air temperature difference and a relative humidity difference, and the multi-source heterogeneous comprehensive index comprises an indoor comfort index and an activity intensity index.
6. The machine learning based two-stage indoor personnel occupancy detection method of claim 1, wherein the step S3 of screening the high-dimensional physical feature matrix using a two-stage feature screening strategy comprises: And the second stage of screening, namely carrying out iterative elimination on the high-dimensional physical feature matrix subjected to the first stage of screening by using a recursive feature elimination algorithm, and screening out an optimal feature subset with the lowest feature dimension on the premise of maximizing the detection precision of the classification model.
7. The machine learning based two-stage indoor personnel occupancy detection method of claim 6, wherein in the second stage screening, the recursive feature elimination algorithm uses time series segmentation as a cross-validation strategy and the screening process targets the maximization of the accuracy of the classification model.
8. The machine learning based two-stage indoor personnel occupancy detection method of claim 1, wherein S4 comprises: S41, constructing a candidate model set; s42, classifying the personnel presence confidence through classification gating reasoning to generate a binary decision; S43, training is carried out on the full training data set with people and without people by using each candidate model, complete continuous mapping from environment perturbation to unmanned and unmanned is learned, and the candidate model with the lowest average absolute error is selected as a people number regression detection model.
9. The machine learning based two-stage indoor personnel occupancy detection method of claim 8, wherein the presence detection model and the person number regression detection model are trained and evaluated using 5-fold time series cross-validation.
10. The machine learning-based two-stage indoor personnel occupancy detection method of claim 8, wherein the person number regression detection model uses a post-processing module comprising non-negative cutoff and discretization rounding to make the person detection result a non-negative integer.

Description

Machine learning-based two-stage indoor personnel occupation detection method Technical Field The invention relates to the technical fields of computer data processing, pattern recognition and intelligent building control, in particular to a two-stage indoor personnel occupation detection method based on machine learning. Background The building department is used as an important energy consumption body, and the energy conservation and emission reduction of the building department face a great challenge. Heating Ventilation Air Conditioning (HVAC) systems typically account for 40% -60% of the total operating energy consumption of a building, and conventional building automation systems rely mostly on rigid control logic for preset schedules, resulting in energy waste phenomena. Thus, an Occupant-centric control (OCC) concept has developed, the core of which is to finely adjust the device operation according to real-time perceived indoor personnel information. The key to realizing OCC is accurate and real-time indoor personnel detection. The existing indoor personnel detection information acquisition technology mainly comprises a visual technology, a Passive Infrared (PIR) technology and a PIR (human input device) technology, wherein the visual technology is based on the highest precision of a visual identification technology of a camera, but has serious privacy invasion risks, and is difficult to popularize, the PIR technology is the most mature non-invasive technology, the PIR cost is low, but the physical principle determines that only a mobile heat source can be detected, false negative is easily generated for static personnel such as sitting office, reading and the like, and the specific number is difficult to detect. The environmental sensor fusion technology utilizes sensor data such as CO2, temperature and humidity, sound and light and the like to infer the existence and quantity of personnel. Such methods are non-invasive and privacy preserving, but the environmental data has a large hysteresis and the detection error is large. The existing personnel detection method based on the environment sensor has the following key technical bottlenecks in practical application: 1. The model is overly dependent on time characteristics-most existing algorithms directly take time stamps (e.g. hours, workdays) as input characteristics. Because the indoor personnel behaviors have certain regularity, the model is actually used for reciting the work and rest timetable of the building, and does not learn the physical law that the environmental parameters change with the personnel. Once the model is subjected to overtime, holiday rest or work and rest change, the model detection capability is greatly reduced, and meanwhile, the model generalization capability is poor. 2. Zero-valued dilation problems-in office or residential settings, there is often an unmanned period of considerable duration. When processing such highly skewed data distributions, it is difficult for a conventional single regression model to balance a large number of zero values with a small number of non-zero values, resulting in non-zero detection results being output when no one is present, or detection values being biased toward an average value when one is present, resulting in large detection errors. 3. Feature engineering lacks physical interactivity-raw sensor data tends to be disturbed by environmental background noise (e.g., HVAC operating noise, direct sunlight heating). The existing characteristic engineering construction method lacks of excavation of nonlinear coupling relation between physical quantities (such as interaction of sound pressure level and temperature and indoor and outdoor environment difference values), and is difficult to distinguish equipment operation and personnel activities in a complex dynamic environment. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a two-stage indoor personnel occupation detection method based on machine learning, which aims to solve the problems of poor generalization capability caused by feature redundancy and excessive dependence on time discipline and high zero value expansion and false alarm rate faced by a single regression model in the existing non-invasive indoor personnel detection technology, and can abandon dependence on time discipline and deep mining of environment physical interaction features and effectively solve zero value interference by constructing an interaction feature set with pure physical dimensions and adopting a two-stage modeling strategy, thereby improving the accuracy of indoor personnel detection on the premise of protecting privacy. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: A two-stage indoor personnel occupation detection method based on machine learning comprises the following steps: S1, acquiring multi-mode time sequence data of an office comprehensive environment se