Air conditioner terminal energy-saving control method and system based on millimeter wave radar perception

Abstract

The present application relates to the technical field of radio waves, in particular to an air conditioner terminal energy-saving control method and system based on millimeter wave radar sensing, which extracts vital sign features by analyzing the micro-Doppler effect of echo signals, realizes accurate differentiation of operating personnel and non-living moving targets, after confirming the human target, combines joint angle-Doppler analysis and time-of-flight ranging to solve its three-dimensional coordinates, analyzes macro-Doppler frequency shift and micro-motion characteristics, obtains its motion vector and activity state, and forms a multi-dimensional personnel state data stream. Then, the state data stream is fused with the operation work order, the long short-term memory network is used to predict the future walking track and key operation point of personnel, a deep Q network is constructed for global optimization calculation, and a control instruction sequence is output. The instruction is used to dynamically define a local cooling area following the movement of personnel and pre-cooling, realizes high-precision, forward-looking and personalized environmental protection for personnel, and greatly reduces the total energy consumption of the system.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of radio waves, in particular to an air conditioner terminal energy-saving control method and system based on millimeter wave radar sensing. BACKGROUND

[0002] In constructing the technical difficulties of high global refrigeration energy consumption and inability to accurately supply air to personnel on demand for air conditioning systems in existing large spaces such as data centers, especially in complex indoor environments such as data centers, accurate sensing and tracking of personnel are the technical premise for intelligent management, but the existing technology has significant defects; low-dimensional sensing technologies such as passive infrared can only provide a fuzzy existence judgment and cannot locate or classify targets; although the optical vision scheme has rich information, it faces serious privacy compliance, light dependence and physical obstruction problems in scenarios such as machine rooms;

[0003] And when the traditional wireless positioning technology is applied to such environments, not only will it cause serious multipath effects and electromagnetic interference due to the dense metal cabinets, but more importantly, it lacks the ability to analyze the fine details of the target echo signal, cannot analyze the biological properties of the target, and is difficult to reliably distinguish between human bodies and non-living mobile targets such as automatic inspection vehicles, and cannot sense the specific activity state, so it is difficult to meet the needs of intelligent applications for high-precision, multi-dimensional sensing information;

[0004] Therefore, the application provides an air conditioner terminal energy-saving control method and system based on millimeter wave radar sensing. SUMMARY

[0005] The application aims to provide an air conditioner terminal energy-saving control method and system based on millimeter wave radar sensing, which forms an intelligent energy-saving control method for air conditioner terminals by integrating millimeter wave radar high-precision sensing, deep learning trajectory prediction, and reinforcement learning optimal decision-making. To achieve the above purpose, the application provides the following technical solutions:

[0006] The air conditioner terminal energy-saving control method based on millimeter wave radar sensing comprises:

[0007] The millimeter wave radar sensor array is deployed in the machine room, emits a frequency-modulated continuous wave, receives the radar echo signal generated by the reflection of the moving target, analyzes the micro-Doppler effect of the radar echo signal, extracts vital sign features, and distinguishes between maintenance personnel and non-living mobile targets;

[0008] After confirming the moving target as a human body, a joint angle-Doppler analysis is used for angle of arrival estimation, combined with time-of-flight ranging of the radar echo signal, to calculate real-time three-dimensional spatial coordinates of the moving target, and to synchronously extract temperature readings of the corresponding area; an activity state classification is obtained according to micro-motion characteristics, and real-time motion speed and direction are calculated by analyzing macro-Doppler frequency shift; the activity state classification, motion speed and direction are integrated into a personnel state data stream;

[0009] The personnel state data stream and operation and maintenance work order are input into a trajectory prediction model to obtain key operation points and three-dimensional space walking trajectories, a deep Q network is applied for global optimization calculation, a local cooling area range and refrigeration intensity are dynamically defined with the key operation points as the center, and a control instruction sequence of the air conditioner is output.

[0010] Preferably, the echo signal of the moving target is subjected to time-frequency analysis, and the non-periodic micro-Doppler spectrum characteristics generated by the non-rigid motion of the human body walking and working, and the like are compared with the periodic spectrum of rigid and / or quasi-rigid targets such as robots and inspection vehicles, to realize preliminary identification of the human body target;

[0011] Under the condition that the preliminary identification is a human body target and it is judged to be in a static or micro-motion state, further micro-Doppler modulation characteristics with unique period and amplitude caused by human respiration and heartbeat are extracted from the time-frequency spectrum as vital sign characteristics, and high-confidence confirmation of the human body target is completed by comparison with a preset model, to effectively distinguish non-living moving targets.

[0012] Preferably, the step of calculating the real-time three-dimensional spatial coordinates of the moving target specifically includes: static clutter filtering processing is performed on the target echo signal confirmed as a human body to enhance the signal signal-to-noise ratio; the clarity of the micro-Doppler spectrum characteristics matched with the human motion pattern is used as a confidence score for evaluating the quality of each signal path, by using a joint angle-Doppler analysis method with multiple receiving channels of the millimeter wave radar sensor array; the angle with the highest confidence score is identified as the main path of the signal to obtain the angle of arrival estimation containing the azimuth angle and the pitch angle; the target distance is determined based on the signal characteristics of the frequency-modulated continuous wave to calculate the round-trip time of flight of the signal; the coordinates of the operation and maintenance personnel in the three-dimensional space are calculated by fusing the angle of arrival estimation and the target distance, and the temperature readings of the corresponding area are synchronously extracted.

[0013] Preferably, the step of classifying the activity state according to the micro-motion characteristics after confirming the target as a human body specifically comprises: performing time-frequency analysis on the target echo signal confirmed as a human body to extract micro-Doppler spectrum characteristics generated by limb swinging; comparing and matching the extracted spectrum characteristics with a preset activity state feature library, the feature library pre-labeling activity states including low metabolic rate level of sitting operation, medium metabolic rate level of walking at constant speed, and high metabolic rate level of high-intensity operation; and classifying the current activity state of the operation and maintenance personnel into discrete values associated with metabolic rate levels according to the matching result.

[0014] Preferably, the step of calculating the real-time motion speed and direction specifically comprises: performing two-dimensional fast Fourier transform on the radar target echo signal data confirmed as a human body to generate a range-Doppler heat map; locating a signal peak point with the largest radar scattering cross section by a peak value detection algorithm, the signal peak point representing a strong reflection point of the human body trunk; and extracting a macroscopic Doppler frequency shift of the signal peak point in the Doppler dimension.

[0015] According to the macroscopic Doppler frequency shift and the radar carrier frequency, the radial velocity of the operation and maintenance personnel along the radar line-of-sight direction is calculated; the three-dimensional spatial coordinates obtained at continuous multiple time points are time-series tracked by using a Kalman filter, and the calculated radial velocity is input into the filter as an observation value, the historical trajectory information is fused by the Kalman filter to estimate and output the real-time motion speed and direction vector of the personnel in the machine room three-dimensional coordinate system, and the activity state classification together constitutes a personnel state data stream.

[0016] Preferably, the step of obtaining the key operation points and the three-dimensional space walking trajectory specifically comprises: performing natural language processing and structured information extraction on the operation and maintenance work order extracted from the operation and maintenance system to parse the task target containing the target device ID, cabinet number and corresponding physical location coordinates, and defining the parsed physical location coordinates as a key operation point sequence; inputting the defined key operation point sequence, the real-time obtained starting position of the operation and maintenance personnel, the historical movement trajectory data, and the generated personnel state data stream into a trajectory prediction model; the trajectory prediction model is a long short-term memory network, and based on the learned personnel movement mode and space constraints, the three-dimensional space walking trajectory connecting all the key operation points is predicted, and the estimated stay time at each operation point is estimated to generate a future position prediction data stream.

[0017] Preferably, the step of applying the deep Q network to perform global optimization calculation comprises: defining the multi-dimensional perception data acquired in real time as the state space input of the deep Q network, the state space at least including: real-time three-dimensional coordinates of the operation and maintenance personnel, movement speed and direction, activity state metabolic rate level, current temperature and air supply of the local area where the personnel are located, and current time and external data of time-of-use electricity price; defining a set of discretized instructions for adjusting the refrigeration intensity of the local cooling area as the action space of the deep Q network; learning the optimal mapping strategy from the state space to the action space through the deep Q network, and the optimal mapping strategy learning is guided by maximizing a long-term cumulative reward which integrates the implicit feedback on personnel comfort and the punishment on high energy consumption of the air conditioner.

[0018] Preferably, the step of outputting the control instruction sequence of the air conditioner specifically comprises: based on the generated future position prediction data stream, extracting the key operation point to be reached by the operation and maintenance personnel and the predicted arrival time, and defining a dynamic local cooling area centered on the predicted key operation point in advance; feeding the predicted state associated with the key operation point, including the predicted metabolic rate level and the initial temperature of the area, to the trained deep Q network; determining the target refrigeration intensity according to the optimal action output by the deep Q network, and generating specific air conditioner terminal control instructions, which accurately open the air outlets of the target area and close the air outlets of the non-target area before the personnel arrive, and adjust the refrigeration power of the air conditioner host as needed.

[0019] The air conditioner terminal energy-saving control system based on millimeter wave radar perception comprises:

[0020] The target differentiation module: millimeter wave radar sensor arrays are deployed in the machine room, transmitting frequency-modulated continuous waves, receiving radar echo signals generated by the reflection of moving targets, extracting vital sign features by analyzing the micro-Doppler effect of the radar echo signals, and distinguishing between maintenance personnel and non-living moving targets;

[0021] The behavior recognition module: after confirming that the moving target is a human body, joint angle-Doppler analysis is used for angle of arrival estimation, combined with the time-of-flight ranging of the radar echo signals, to calculate the real-time three-dimensional spatial coordinates of the moving target and synchronously extract the temperature readings of the corresponding area; the activity state classification is obtained according to the micro-motion characteristics, and the real-time movement speed and direction are calculated by analyzing the macro-Doppler frequency shift; the activity state classification, movement speed and direction are integrated into the personnel state data stream;

[0022] The instruction output module: inputs the personnel state data stream and the operation and maintenance work order into the trajectory prediction model to obtain the key operation point and the three-dimensional space walking trajectory, applies the deep Q network to perform global optimization calculation, dynamically defines the local cooling area range and the refrigeration intensity centered on the key operation point, and outputs the control instruction sequence of the air conditioner.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] 1. By analyzing the micro-Doppler effect of radar echoes, the vital sign features such as respiration and heartbeat are extracted, which can accurately distinguish between maintenance personnel and non-living moving targets such as inspection robots, and fundamentally avoid the wrong identification of control targets; combined with joint angle-Doppler analysis and Kalman filter tracking, the position, speed and direction of personnel are stably and accurately calculated in complex electromagnetic environments, providing a reliable data foundation for subsequent accurate control.

[0025] 2. The present application not only realizes real-time tracking of personnel, but also predicts the future moving track and key operation point of personnel by fusing operation order information and long short-term memory network; this makes the air conditioning system change from passive response to "pre-cooling" active service, and provides a comfortable environment for the operation area that will arrive in advance; at the same time, through reinforcement learning, the user's implicit feedback, i.e. manual adjustment behavior, is continuously learned to build a dynamic personalized comfort model, realizing fine air supply according to different people and different activity states.

[0026] 3. Taking deep Q network as the decision core, the generation of control strategy is constructed as a multi-objective optimization problem. This method not only considers the real-time position and activity state of personnel, but also dynamically fuses external economic factors such as time-of-use electricity price, and through the design of reward function, continuously searches for the control strategy with the lowest energy consumption under the premise of ensuring personnel comfort; its essence is to replace the traditional "global space cooling mode" with a "dynamic local cooling area" that follows the movement of personnel, so as to accurately put cold energy into the effective area, greatly reduce the power utilization efficiency of the data center, and the energy-saving effect is remarkable. BRIEF DESCRIPTION OF DRAWINGS

[0027] Fig. 1 The step flow chart of the air conditioning terminal energy-saving control method and system based on millimeter wave radar perception of the present application;

[0028] Fig. 2 The flowchart of the air conditioning terminal energy-saving control method and system based on millimeter wave radar perception of the present application;

[0029] Fig. 3 The structural diagram of the dynamic local cooling area based on trajectory prediction of the present application. DETAILED DESCRIPTION

[0030] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0031] Please refer to Figs. 1 to 3 The present application provides an air conditioner terminal energy-saving control method and system based on millimeter wave radar perception, and the technical solutions are as follows: with reference to Fig. 1 The step flow chart is Fig. 2 The flow chart is a schematic diagram, and specifically includes:

[0032] A millimeter wave radar sensor array is deployed in a machine room, a frequency-modulated continuous wave is transmitted, a radar echo signal generated by a moving target reflection is received, vital sign features are extracted by analyzing the micro-Doppler effect of the radar echo signal, and living personnel and non-living moving targets are distinguished.

[0033] After confirming that the moving target is a human body, joint angle-Doppler analysis is used for angle of arrival estimation, time-of-flight ranging of the radar echo signal is combined, real-time three-dimensional spatial coordinates of the moving target are calculated, and temperature readings of the corresponding area are synchronously extracted; activity state classification is obtained according to the micro-motion features, real-time motion speed and direction are calculated by analyzing macro-Doppler frequency shift; the activity state classification, motion speed and direction are integrated into a personnel state data stream.

[0034] The personnel state data stream and operation and maintenance work orders are input into a trajectory prediction model to obtain key operation points and walking trajectories, a deep Q network is applied for global optimization calculation, a local cooling area range and refrigeration intensity are dynamically defined around the key operation points, and a control instruction sequence of the air conditioner is output.

[0035] In a large data center machine room, there are multiple rows of server cabinets, and temperature control in the machine room is crucial for stable operation of equipment; at the same time, operation and maintenance personnel will periodically or irregularly enter the machine room for inspection, equipment maintenance or fault handling; traditional air conditioning systems usually uniformly cool the entire machine room, resulting in continuous consumption of a large amount of energy in areas where personnel are not active; the method of the present application aims to precisely cool only the personnel activity area, thereby achieving significant energy saving.

[0036] Embodiment 1: First, a millimeter wave radar sensor array is deployed in the machine room, a frequency-modulated continuous wave is transmitted, a radar echo signal generated by a moving target reflection is received, vital sign features are extracted by analyzing the micro-Doppler effect of the radar echo signal, and living personnel and non-living moving targets are distinguished.

[0037] Eight millimeter-wave radar sensors are uniformly deployed on the top of the data center room to form an array; these radar sensors continuously emit frequency-modulated continuous wave signals with a frequency of 77 GHz; when the maintenance personnel A enters the room, the maintenance personnel A maintains the No. 3 cabinet, and the radar sensor receives the radar echo signal reflected by the body; at the same time, the automatic inspection robot existing in the room will also produce an echo.

[0038] Further, the distinction between the maintenance personnel and the non-living moving target is specifically implemented in the embodiment as follows:

[0039] Step one: preliminary identification, first, the received moving target echo signal is subjected to time-frequency analysis, and the complexity and non-periodicity features of the micro-Doppler spectrum are extracted to distinguish between human bodies and robots and other rigid / quasi-rigid targets; this specifically includes calculating the spectral entropy of the time-frequency spectrum in the Doppler dimension and an energy concentration index, wherein a signal with high spectral entropy and low energy concentration is considered to have high complexity; and performing autocorrelation analysis on the signal to calculate the ratio of the maximum autocorrelation peak value to the secondary peak value, and the ratio is lower than the preset threshold (for example, 1.5) and is considered to have significant non-periodicity; subsequently, the system inputs the extracted feature vectors into a pre-trained support vector machine (SVM) classifier, and the classifier outputs the preliminary identification result that the target is a "human body" or a "non-human body"; the SVM classifier is trained using a dataset containing at least 1000 labeled samples.

[0040] Step two: high-confidence confirmation, when the target is preliminarily identified as a "human body" and the system judges that it has been in a stationary or micro-motion state, a high-confidence confirmation process is started. This process further extracts the micro-Doppler modulation features (for example, a periodic modulation with a period of about 1.2 Hz and an amplitude fluctuating within a range of 20 Hz) with unique periodicity and amplitude caused by human respiration and heartbeat from the time-frequency spectrum, and compares them with the pre-set human vital sign model; when the matching degree of the vital sign feature reaches 95% (higher than the pre-set threshold of 80%), the system finally confirms that the current tracking target is the maintenance personnel A; in contrast, the echo signal time-frequency spectrum of the inspection robot does not contain such periodic micro-Doppler features, thereby achieving effective and highly reliable distinction.

[0041] Using the unique respiration, heartbeat and other vital sign micro-Doppler features of the human body as the basis for identification provides the system with a biological level "target verification" capability; this ensures that the control system can accurately lock the human body target from the source, effectively avoiding false identification and invalid refrigeration of non-living targets such as inspection robots and mobile device vehicles in the room, and is a reliable prerequisite for the implementation of all subsequent precise control and energy-saving strategies.

[0042] After confirming the moving target as a human body, a joint angle-Doppler analysis is used for angle of arrival estimation, combined with time-of-flight ranging of the radar echo signal, to calculate the real-time three-dimensional spatial coordinates of the moving target, and to synchronously extract temperature readings of the corresponding area; according to the micro-motion characteristics, an activity state classification is obtained, and by analyzing the macro-Doppler shift, the real-time motion speed and direction are calculated; the activity state classification, motion speed and direction are integrated into a personnel state data stream; specifically, the target echo signal confirmed as a human body is subjected to static clutter filtering processing to enhance the signal-to-noise ratio, for example, by setting a high-pass filter with a cutoff frequency of 0.1 Hz; further, a joint angle-Doppler analysis method is used by using multiple receiving channels of the millimeter wave radar sensor array; in the presence of multipath interference, the system analyzes multiple echo signals from different angles. The echo signal with the clearest and stable human micro-Doppler feature (such as gait cycle when walking) is considered to have the highest probability of coming from the direct path or high-quality reflection path of the target. The angle measurement value of this path is given a higher weight in the subsequent Kalman filter tracking algorithm, and the weight of other feature-fuzzy path signals is reduced.

[0043] (1) Target distance measurement: based on the signal characteristics of frequency-modulated continuous wave, the round-trip time of flight of the signal is calculated; for example, the measured time of flight corresponds to a target distance of 12.5 meters;

[0044] (2) Three-dimensional coordinate fusion: fusing the angle of arrival, including azimuth angle-5° and elevation angle 10°, and the target distance of 12.5 meters, the real-time coordinates of the operation and maintenance personnel A are calculated in three-dimensional space, for example, (X=12.3m, Y=-1.1m, Z=2.1m);

[0045] (3) Synchronous extraction of temperature readings: after calculating the real-time three-dimensional coordinates of personnel A, the temperature readings of the area where the coordinates are located are synchronously extracted, for example, the current temperature of 28°C is obtained through the distributed temperature sensor located in the area.

[0046] The step of calculating the micro-Doppler spectrum feature clarity and generating a confidence score is specifically: first, identify the periodic main frequency peak value corresponding to the pedestrian gait in the time-frequency spectrum (usually 1-2 Hz), and calculate the signal-to-noise ratio of the peak value, which is taken as a quantitative indicator of clarity. Then, through an S-shaped function, the SNR value is mapped to a confidence score ranging from 0 to 1; in the subsequent Kalman filter update step, the confidence score is used to adjust the measurement noise covariance of the corresponding angle measurement; when the clarity is high and the confidence score is close to 1, it means that the measurement value is more reliable, and its noise covariance is reduced, so that it obtains a higher weight in state estimation.

[0047] By adopting the joint angle-Doppler analysis method, the angle estimation is bound with the unique micro-Doppler characteristics of the human body, which can effectively suppress the serious multipath reflection interference generated by the metal cabinet in the data center; compared with the traditional direction finding method, it can more accurately identify the signals generated by the human body direct or main path reflection, and improve the accuracy of personnel three-dimensional space positioning in complex electromagnetic environment.

[0048] Further, real-time motion speed and direction calculation, performing two-dimensional fast Fourier transform on the radar target echo signal data confirmed as human body, generating a range-Doppler heat map;

[0049] By peak detection algorithm to locate the signal peak point with the largest radar scattering cross section, the peak point represents the strong reflection point of the human body trunk; for example, the center frequency offset of the peak point in the Doppler dimension is +500Hz, which is the macroscopic Doppler shift; the radial velocity calculation is calculated according to the macroscopic Doppler shift (+500Hz) and the radar carrier frequency (77GHz), and the radial velocity of the operation and maintenance personnel A along the radar line of sight direction is +0.97m / s, indicating that the personnel A is approaching the radar at a speed of 0.97m / s.

[0050] Three-dimensional motion speed and direction vector estimation is to use Kalman filter to track the three-dimensional space coordinates obtained at continuous multiple times; the calculated radial velocity is input into the filter as an observation value; the Kalman filter fuses historical trajectory information, estimates and outputs the real-time motion speed vector (for example, X direction 0.5m / s, Y direction-0.2m / s, Z direction 0.1m / s) and direction vector of personnel A in the three-dimensional coordinate system of the machine room;

[0051] Personnel state data stream integration is to integrate the activity state classification (fine operation, discrete value 2), real-time motion speed (for example, X direction 0.5m / s) and direction vector into personnel state data stream.

[0052] By introducing Kalman filter, the inherent limitation of millimeter wave radar that can only directly measure radial velocity is solved; it can fuse the three-dimensional coordinate information and the current radial velocity measurement value, and accurately estimate the complete motion vector of personnel; this provides high-quality, non-jittering dynamic motion parameter input for subsequent trajectory prediction and activity state analysis, and improves the analysis accuracy of the whole system.

[0053] Further, the activity state classification is performed, the target echo signal confirmed as a human body is subjected to time-frequency analysis, and micro-Doppler spectrum characteristics generated by limb swing are extracted; for example, when personnel A performs maintenance operation in front of cabinet No. 3, the fine movement of the hands and upper limbs of personnel A will generate micro-Doppler modulation with a frequency in the range of ±0.5 kHz and a periodicity of about 0.8 seconds on the time-frequency spectrum; the extracted spectrum characteristics are compared and matched with a preset activity state characteristic library; the characteristic library pre-labels the low metabolic rate level of sitting operation, the spectrum characteristics are concentrated in ±0.1 kHz, the medium metabolic rate level of walking at a constant speed, the spectrum characteristics are concentrated in ±1 kHz, the high metabolic rate level of high-intensity operation, the spectrum characteristics are extended to more than ±2 kHz, and the periodicity is irregular, and the activity state; according to the matching result, the current activity state of the operation and maintenance personnel A is classified as “fine operation”, corresponding to the medium metabolic rate level, and is assigned a discrete value of 2.

[0054] The micro-Doppler spectrum of the abnormal behavior period is extracted, compared with a preset operation fingerprint library, and the unplanned operation intention is confirmed and classified; specifically, the micro-Doppler spectrum of the radar echo signal in the long stationary period is intercepted and analyzed; and the spectrum characteristics are quickly compared with a preset micro-Doppler operation fingerprint library; this qualitative analysis capability enables the cooling intensity of the air conditioner to be matched with the actual heat load of personnel on demand, providing truly personalized and resource-optimized environmental control, and avoiding excessive cooling for low-intensity operations; the fingerprint library pre-enters and labels the unique signal characteristics of various typical fine operations, for example:

[0055] Fingerprint library A-keyboard operation: the spectrum is characterized by periodic signals with high frequency, small amplitude, and concentrated in a specific frequency band;

[0056] Fingerprint library B-cable plugging: the spectrum is characterized by a wideband signal of large arm movement, followed by a weak and complex narrowband signal of fine positioning of the hands;

[0057] Fingerprint library C-tool use: the spectrum is characterized by unique, non-human motion harmonic frequencies generated by the rotation or vibration of tools such as electric screwdrivers;

[0058] Only when the detected micro-Doppler characteristics match a certain item in the fingerprint library above a preset threshold, the system finally confirms that the behavior anomaly is an unplanned operation event, and classifies the operation type according to the matched fingerprint.

[0059] The quantification of the intensity of personnel activities is achieved, and the abstract "activity state" is associated with the specific "metabolic rate level". By accurately identifying sitting, walking or high-intensity work, the system can assess the heat production load of personnel in real time, so as to dynamically match the refrigeration intensity with the actual physiological demand. Compared with the control strategy based on position only, this method realizes more personalized comfort management and more refined energy distribution.

[0060] Reference Fig. 3 , wherein A, B, C represent the main functional areas of the machine room, the light gray square (legend: Off) represents the outlet in the closed state; O is the generated dynamic local cooling area, the black circle (legend: On) represents the open state; the black dotted line 601 is the historical trajectory, and the black short dash line 602 is the predicted trajectory; the personnel state data stream and the operation and maintenance work order are input into the trajectory prediction model to obtain the key operation points and the walking trajectory, a deep Q network is applied for global optimization calculation, the local cooling area range and the refrigeration intensity are dynamically defined with the key operation points as the center, and the control instruction sequence of the air conditioner is output. Specifically, the operation and maintenance work order is extracted from the operation and maintenance system, for example, "performing disk replacement operation on No. 3 cabinet (ID: DC-CAB-003)"; the work order is subjected to natural language processing and structured information extraction, and the task target containing the target equipment ID "DC-CAB-003", the cabinet number "No. 3" and the corresponding physical position coordinates (X=10m, Y=5m, Z=2m) are analyzed out; the analyzed physical position coordinates are defined as the first point in the key operation point sequence.

[0061] Further, the defined key operation point sequence (No. 3 cabinet coordinates) and the starting position, current coordinates (X=12.3m, Y=-1.1m, Z=2.1m) of the operation and maintenance personnel A, historical movement trajectory data such as the movement path in the past 30 seconds and the generated personnel state data stream are jointly input into a long short-term memory network (LSTM) trajectory prediction model.

[0062] The trajectory and stay time prediction is based on the personnel movement mode and space constraint learned by the LSTM model, and a three-dimensional space walking trajectory connecting all the key operation points is predicted; for example, the predicted trajectory is the path from the current position to No. 3 cabinet along the aisle of the machine room; at the same time, the predicted stay time at each operation point is estimated, for example, it is predicted that No. 3 cabinet will be stayed for 15 minutes; the prediction model generates a future position prediction data stream, indicating that personnel A will arrive at No. 3 cabinet in the next 2 minutes.

[0063] By analyzing historical trajectory data through machine learning, the system automatically discovers spatial hotspots where personnel frequently reside, mines behavior correlations, and predicts the next key work point. Specifically, by analyzing radar trajectory data over the past few months or even a year, the system can automatically discover the areas and equipment where the operation and maintenance personnel most frequently stay using machine learning algorithms such as DBSCAN clustering. Moreover, the system can discover correlations between behaviors. For example, data shows that after operating the C12 cabinet, the operation and maintenance personnel have a 70% probability of operating the F05 cabinet. The hotspots discovered through historical data are themselves considered potential key work points by the system, and the probabilistic correlations between tasks can further assist the system in accurately predicting the next target point from the current location, which is particularly effective when there is no explicit work order.

[0064] By analyzing operation and maintenance work orders and applying a long short-term memory network, the system is given the ability to "predict the future". It transforms aimless random wandering into a path prediction with a clear task goal and a high probability, enabling air conditioning control to upgrade from passive lagging response to proactive service with foresight. This foresight can provide personnel with "cold air" and greatly improve the immediate comfort when entering the work area suddenly.

[0065] Furthermore, a deep Q network global optimization calculation is performed:

[0066] The state space input is to define the multi-dimensional perception data obtained in real time as the state space input of the deep Q network. This state space includes: the real-time three-dimensional coordinates of the operation and maintenance personnel A, such as X=10.1m, Y=4.9m, Z=2.0m, which has approached cabinet No. 3, the movement speed is 0.1m / s, and the activity state metabolic rate level corresponds to fine operation, discrete value 2, the local area where the personnel are, that is, the current temperature reading of 28°C in front of cabinet No. 3 and the air supply amount of 0 CFM, because it has not been turned on, and the current time is 2:30 pm corresponding to the off-peak electricity price external data is the peak electricity price, 0.8 yuan / d;

[0067] The action space definition is to define a set of discrete instructions for adjusting the cooling intensity of the local cooling area as the action space of the deep Q network. For example, action 0: close the air outlet, i.e. cooling intensity 0; action 1: open the air outlet, low cooling intensity, air supply amount 50 CFM, target temperature 26°C; action 2: open the air outlet, medium cooling intensity, air supply amount 100 CFM, target temperature 24°C; action 3: open the air outlet, high cooling intensity, air supply amount 150 CFM, target temperature 22°C.

[0068] The objective function includes comfort maintenance cost, external data source constraint, energy efficiency, and equipment thermal risk. The comfort penalty is proportional to the square of the difference between the actual temperature and the target temperature, and the energy consumption penalty is proportional to the air conditioning power, with a weight of β,

[0069] The economic penalty is proportional to the time-of-use electricity cost; the device thermal risk penalty is a very high penalty weight when the temperature of the critical IT device has a trend to exceed its optimal health interval;

[0070] Specifically, the predicted comfort maintenance cost is the total cooling capacity required to meet the comfort requirement of all points on the future trajectory according to the model prediction; the external data source constraint is to actively incorporate external data such as time-of-use electricity data (increase the weight of energy consumption during peak electricity prices), weather forecast data (affect the base thermal load of the machine room), as a constraint condition or cost item of optimization; the energy efficiency optimization is to take the minimum energy consumption required to achieve the above goals as the core optimization direction; in the objective function, a "virtual hot spot" is set for the critical IT device (such as the core server, switch) or the known weak cooling point in the machine room as the device thermal risk; even if there is no personnel in the current area, once the temperature of the point has a trend to exceed its optimal health interval, the objective function will impose a very high penalty weight; ensuring that any energy-saving strategy cannot absolutely sacrifice the stable operation of the IT device at the cost of increasing the risk control and redundancy protection of the entire system.

[0071] The optimal mapping strategy learning maximizes a long-term cumulative reward that combines personnel comfort, such as a comfort zone of a target temperature of 24℃, ±1℃ implicit feedback, and air conditioning energy consumption penalty, such as a penalty when the cooling intensity is high and the electricity price is high; the optimal mapping strategy from the state space to the action space is learned through a deep Q network; for example, when the personnel are in the "fine operation" state and the local area temperature is high at 28℃, the DQN will learn to select the "medium cooling intensity" action 2 to obtain a higher cumulative reward.

[0072] An adaptive and self-optimizing intelligent decision-making core is constructed; the deep Q network can autonomously find the control strategy with the highest long-term return in a complex state space containing personnel, environment, energy price, etc. through continuous learning and trial and error; this method eliminates the dependence on static thresholds and fixed rules, enabling the system to dynamically adapt to various changes and ensuring that the control decisions at any time tend to be globally optimal.

[0073] Further, the air conditioning control instruction sequence output is output, and specifically, the local cooling area predefinition is based on the generated future position prediction data stream, extracts the key operation point to which the operation and maintenance personnel A will arrive and the predicted arrival time, for example, the time to arrive at the No. 3 cabinet is 15 seconds, and a dynamic local cooling area is predefined around the predicted key operation point, for example, a square area with the No. 3 cabinet as the center and a side length of 2 meters.

[0074] Predictive state feed involves feeding the predicted state associated with critical operation points, including the predicted metabolic rate level (fine operation) and the initial temperature of the region (28°C), into the trained deep Q network; target cooling intensity determination is based on the output of the deep Q network in the current state to determine the optimal action, for example, "Action 2", which indicates that the target cooling intensity is medium.

[0075] The air conditioning terminal control command is generated based on a defined target cooling intensity. Specific control commands are generated: 10 seconds before person A arrives, the command control system precisely opens the air conditioning vent above cabinet 3, sets the air volume to 100 CFM, and sets the target outlet temperature to 18℃; simultaneously, it closes the vents in non-target areas and adjusts the cooling power of the air conditioning unit as needed to ensure that the temperature in cabinet 3 can quickly drop to a comfortable range of 24℃ when person A arrives, and gradually adjusts the cooling power after person A leaves to achieve energy saving.

[0076] Accurately implementing the results of forecasts and decisions constitutes a closed loop for proactive service execution; it ensures that cooling resources are deployed to key work points before personnel arrive and that the correct cooling intensity is applied; through precise switching and power adjustment of terminal air vents, dynamic localized cooling that "follows the person" is achieved, which is the ultimate physical guarantee for maximizing the system's energy-saving goals.

[0077] By constructing an integrated computer system, the three modules are linked together to form a highly efficient data processing closed loop. The system first uses a target discrimination module to accurately distinguish between people and static objects; then, it uses a behavior recognition module to analyze the movement trajectory and micro-motion characteristics of the tracked target to accurately determine its activity status; finally, the command output module dynamically generates and sends control commands to the air conditioning terminal equipment based on the behavior recognition results, thereby realizing on-demand cooling and heating and achieving intelligent energy saving.

[0078] Example 2: In a large multi-story data center, there are multiple interconnected server room areas (Area A, Area B, Area C), each of which is equipped with a millimeter-wave radar sensor array; the data center requires multi-area and multi-person collaborative maintenance and emergency response, with two maintenance personnel (Personnel E and Personnel F) performing different maintenance tasks at the same time, and sudden equipment failures may occur.

[0079] The millimeter wave radar system continuously scans all areas. When personnel E enters area A and personnel F enters area B, the system extracts vital signs such as respiration and heartbeat by analyzing the micro-Doppler effect of the respective echo signals, accurately distinguishes personnel E and personnel F, and distinguishes them from the AGV inspection robots, handling equipment and other non-living targets in the machine room; The system calculates the three-dimensional coordinates, motion speed, direction and activity state of personnel E and personnel F in real time (such as personnel E sitting in front of cabinet No. 2 in area A, and personnel F walking at a constant speed between cabinet Nos. 5 and 6 in area B), and synchronously obtains the temperature readings of the respective areas; These form two independent and real-time "personnel state data streams".

[0080] Further, trajectory prediction and task fusion, the system simultaneously receives two operation and maintenance work orders. Specifically, work order 1 corresponds to personnel E, and maintains cabinet No. 2 in area A (coordinates X1, Y1, Z1). Work order 2 corresponds to personnel F, and inspects all cabinets in area B, and restarts the equipment at cabinet No. 7 in area C (coordinates X2, Y2, Z2); The system extracts structured information from the two work orders to obtain the respective key operation point sequences.

[0081] The state data stream of personnel E and work order 1, and the state data stream of personnel F and work order 2, are respectively input into the respective long short-term memory network (LSTM) trajectory prediction model; The model predicts that personnel E will stay in front of cabinet No. 2 for 30 minutes, and generates a local cooling area around the point; The model predicts the inspection path of personnel F, estimates the time to arrive at cabinet No. 7 in area C, and predicts to stay at the point for 10 minutes, and plans the follow-up cooling path in area B and the pre-cooling area of cabinet No. 7 in area C in advance.

[0082] Further, sudden failure and emergency response, while personnel E and personnel F are respectively performing tasks, cabinet No. 1 in area C suddenly issues a high-temperature warning and triggers a high-level fault work order, requiring immediate handling; The system immediately promotes the priority of this high-level fault work order (including the coordinates X3, Y3, Z3 of cabinet No. 1 in area C) and inputs it into the trajectory prediction model for all personnel;

[0083] The model quickly evaluates that personnel F is closest to cabinet No. 1 in area C, and that his current task (inspection) can be interrupted; The system predicts that personnel F will immediately change the path to cabinet No. 1 in area C, and estimates the arrival time.

[0084] Further, deep Q network global optimization and multi-air-conditioner end collaborative control, the system inputs all multi-dimensional perception data such as real-time coordinates, activity state, predicted trajectory, current area temperature, time-of-use electricity price of the two personnel into the state space of the deep Q network (DQN); DQN aims to maximize the comprehensive reward (personnel comfort, energy consumption penalty, equipment thermal risk penalty, emergency response priority) for global optimization calculation.

[0085] Further, output control instructions: personnel E sat in front of No. 2 cabinet operation, DQN output "low cooling intensity" instruction, accurate opening A area No. 2 cabinet above the local air outlet, air supply 80 CFM, target temperature 25℃; personnel F changed path to C area, DQN output "close" instruction, B area follow cooling area air outlet gradually closed, energy saving; due to C area No. 1 cabinet failure, DQN output "emergency high cooling intensity" instruction, immediately open C area No. 1 cabinet above the air outlet, air supply 150 CFM, target temperature 20℃, while dispatch adjacent air conditioning host refrigeration power; in personnel F will arrive at C area No. 1 cabinet, C area No. 7 cabinet precooling plan is temporarily shelved or adjusted.

[0086] Through the implementation of the above multi-personnel scene, the method can simultaneously track multiple personnel, process multiple task flows, and quickly respond and optimize resource allocation when an emergency occurs, realize fine, forward-looking, personalized and emergency environmental control for different personnel and different areas, maximize energy saving effect, while ensuring personnel comfort and equipment safety.

[0087] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, the scope of the present application being defined by the appended claims and their equivalents.

Claims

1. An energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing, characterized in that, Includes the following steps: A millimeter-wave radar sensor array is deployed in the equipment room to transmit frequency-modulated continuous waves and receive radar echo signals generated by reflections from moving targets. By analyzing the micro-Doppler effect of the radar echo signals, vital signs are extracted to distinguish maintenance personnel from non-living moving targets. After confirming that the moving target is a human body, joint angle-Doppler analysis is used to estimate the angle of arrival. Combined with the time-of-flight ranging of the radar echo signal, the real-time three-dimensional spatial coordinates of the moving target are calculated, and the temperature readings of the corresponding area are extracted simultaneously. The activity state is classified according to the micro-motion characteristics, and the real-time motion speed and direction are calculated by analyzing the macroscopic Doppler frequency shift. The activity state classification, motion speed, and direction are integrated into a personnel status data stream. The personnel status data stream and maintenance work orders are input into the trajectory prediction model to obtain key work points and three-dimensional spatial walking trajectories. A deep Q-network is applied for global optimization calculation. With the key work points as the center, the local cooling area range and cooling intensity are dynamically defined, and the control command sequence of the air conditioner is output.

2. The energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing according to claim 1, characterized in that, The specific steps for distinguishing maintenance personnel from inanimate moving targets are as follows: Time-frequency analysis is performed on the echo signal of the moving target to obtain a time-frequency spectrum. Based on the non-periodic micro-Doppler spectrum characteristics generated by non-rigid motions such as human walking and limb swinging during work, the spectrum is compared with the periodic spectrum of rigid and / or quasi-rigid targets such as robots and inspection vehicles to achieve preliminary identification of human targets. Under the condition that the target is initially identified as a human body and the human body is judged to be in a static and / or slightly moving state, the micro-Doppler modulation features with unique periodicity and amplitude caused by human breathing and heartbeat are further extracted from the time-frequency spectrum as vital signs. By comparing with the preset model, the confidence level of the human body target is confirmed, and the non-living moving target is effectively distinguished.

3. The energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing according to claim 1, characterized in that, The step of calculating the real-time three-dimensional spatial coordinates of the moving target specifically includes: Static clutter filtering is performed on the echo signals of targets identified as human bodies to enhance the signal-to-noise ratio. Using multiple receiving channels of the millimeter-wave radar sensor array, a joint angle-Doppler analysis method is employed to use the sharpness of the micro-Doppler spectral features matching the human motion pattern as a confidence score for evaluating the quality of each signal path. The angle with the highest confidence score is identified as the main signal path, yielding an estimated angle of arrival including azimuth and elevation angles. The round-trip time of flight is calculated based on the signal characteristics of the frequency-modulated continuous wave to determine the target distance. The estimated angle of arrival and the target distance are fused to calculate the coordinates of the maintenance personnel in three-dimensional space, and temperature readings for the corresponding area are extracted simultaneously.

4. The energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing according to claim 1, characterized in that, The specific steps for classifying activity states based on micro-motion features are as follows: Time-frequency analysis is performed on the target echo signal identified as a human body to extract the micro-Doppler spectral features generated by limb movement. The extracted spectral features are then compared and matched with a pre-defined activity state feature library, which pre-labels activity states including low metabolic rate level sedentary operation, medium metabolic rate level uniform walking, and high metabolic rate level high-intensity work. Based on the matching results, the current activity state of the maintenance personnel is classified into discrete values ​​associated with the metabolic rate level.

5. The energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing according to claim 1, characterized in that, The specific steps for calculating the real-time motion speed and direction are as follows: A two-dimensional fast Fourier transform is performed on the radar target echo signal data that has been identified as a human body to generate a range-Doppler heatmap; the signal peak point with the largest radar cross section is located by a peak detection algorithm, and the signal peak point represents the strong reflection point of the human torso; the center frequency shift of the signal peak point in the Doppler dimension is extracted to obtain the macroscopic Doppler frequency shift; Based on the macroscopic Doppler frequency shift and radar carrier frequency, the radial velocity of the maintenance personnel along the radar line of sight is calculated. A Kalman filter is used to perform time-series tracking of the three-dimensional spatial coordinates acquired at multiple consecutive moments, and the calculated radial velocity is input into the filter as an observation value. The Kalman filter fuses historical trajectory information to estimate and output the real-time movement velocity and direction vector of the personnel in the three-dimensional coordinate system of the computer room. Together with the activity status classification, they constitute the personnel status data stream.

6. The energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing according to claim 1, characterized in that, The steps for obtaining the key operation points and the three-dimensional spatial walking trajectory specifically include: The maintenance work orders extracted from the maintenance system are processed using natural language processing and structured information extraction to parse out the task objectives, which include the target device ID, rack number, and corresponding physical location coordinates. The parsed physical location coordinates are defined as a sequence of critical work points. The defined sequence of critical work points, along with the real-time acquired starting positions of maintenance personnel, historical movement trajectory data, and the generated personnel status data stream, are input into a trajectory prediction model. The trajectory prediction model is a long short-term memory network that predicts the three-dimensional spatial walking trajectory connecting all critical work points based on learned personnel movement patterns and spatial constraints. It also estimates the expected dwell time at each work point and generates a future location prediction data stream.

7. The energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing according to claim 1, characterized in that, The steps for performing global optimization calculations using the applied deep Q-network include: The real-time acquired multi-dimensional sensing data is defined as the state space input of the deep Q-network. The state space includes at least: the real-time three-dimensional coordinates of the maintenance personnel, their movement speed and direction, the metabolic rate level of their activity state, the current temperature and air volume of the local area where the personnel are located, and external data such as the current time and time-of-use electricity price. A set of discretized instructions used to adjust the cooling intensity of the local cooling area is defined as the action space of the deep Q-network. The deep Q-network learns the optimal mapping strategy from the state space to the action space. The optimal mapping strategy learning is guided by maximizing a long-term cumulative reward, which integrates implicit feedback on personnel comfort and penalties for high energy consumption of the air conditioner.

8. The energy-saving control method for air conditioning terminals based on millimeter-wave radar sensing according to claim 1, characterized in that, The steps of the output air conditioner control command sequence specifically include: Based on the generated future location prediction data stream, the system extracts the key work points that maintenance personnel are about to arrive at and their estimated arrival time, and predefines dynamic local cooling zones centered on the predicted key work points. The predicted states associated with the key work points, including the predicted metabolic rate level and the initial temperature of the area, are fed into the trained deep Q network. The target cooling intensity is determined based on the optimal action output by the deep Q network, and specific air conditioning terminal control instructions are generated to accurately open the air vents in the target area and close the air vents in non-target areas before personnel arrive, and adjust the cooling power of the air conditioning unit as needed.

9. An energy-saving control system for air conditioning terminals based on millimeter-wave radar sensing, characterized in that, include: Target differentiation module: A millimeter-wave radar sensor array is deployed in the equipment room to transmit frequency-modulated continuous waves and receive radar echo signals generated by the reflection of moving targets. By analyzing the micro-Doppler effect of the radar echo signals, vital signs are extracted to distinguish maintenance personnel from non-living moving targets. Behavior recognition module: After confirming that the moving target is a human body, it uses joint angle-Doppler analysis to estimate the angle of arrival, combines the time-of-flight ranging of the radar echo signal to calculate the real-time three-dimensional spatial coordinates of the moving target, and simultaneously extracts the temperature readings of the corresponding area; it obtains activity state classification based on micro-motion characteristics, and calculates real-time motion speed and direction by analyzing macroscopic Doppler frequency shift; it integrates the activity state classification, motion speed, and direction into a personnel status data stream; Command output module: Inputs the personnel status data stream and maintenance work order into the trajectory prediction model to obtain key operation points and three-dimensional spatial walking trajectories. It applies a deep Q-network to perform global optimization calculations, dynamically defines the local cooling area range and cooling intensity with the key operation points as the center, and outputs the air conditioner control command sequence.

Images