Construction engineering operation and maintenance big data visual management and control system and method

Abstract

The invention relates to the technical field of constructional engineering informatization management, and particularly discloses a visual management and control system and method for constructional engineering operation and maintenance big data. According to the system, full-dimension operation and maintenance data is acquired through a multi-source heterogeneous data acquisition module, after the full-dimension operation and maintenance data is processed by a data fusion and feature engineering module, a dynamic three-dimensional thermodynamic diagram and an efficiency instrument panel are generated by a space-comfort dynamic visualization engine to be presented, and an intelligent self-adaptive management and control decision module based on deep reinforcement learning is utilized to realize real-time management and control. And generating an optimization regulation and control instruction for the environment and energy system. According to the invention, transparent insight and self-adaptive precise management and control of the operation state of the building are realized, and the space resource allocation efficiency and the energy utilization efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information management of construction engineering, and particularly relates to a big data visualization management and control system and method for construction engineering operation and maintenance. BACKGROUND

[0002] In the field of building and facility management, with the deep integration of Internet of Things, big data and artificial intelligence technology, smart building operation and maintenance is developing towards fine, intelligent and data-driven direction. The goal is to realize efficient, energy-saving and humanized management and control of building energy consumption, environment, equipment and space elements through the collection, analysis and application of massive heterogeneous data in buildings, so as to improve the overall operation efficiency and user experience of buildings.

[0003] Big data visualization management and control of construction engineering operation and maintenance is an important technical direction of smart building operation and maintenance. This technology aims to integrate multi-source, multi-dimensional data generated during building operation and maintenance through data fusion, model calculation and visualization, to provide intuitive and comprehensive decision support for managers, so as to realize real-time monitoring, abnormal warning and optimization control of building operation status.

[0004] The existing technology usually relies on deploying various sensors to collect environmental and equipment data, and building independent visualization panels for display. The existing system has the following problems: the collected data is single-dimensional, mainly focusing on basic physical quantities such as temperature and humidity, energy consumption, and lacks effective integration and correlation analysis of multi-modal data such as actual occupancy rate of space, personnel flow pattern and user subjective comfort feedback; the visualization presentation is mainly static charts or simple dashboards, which cannot dynamically and intuitively reflect the spatio-temporal distribution and trend of building space utilization efficiency and personnel comfort; the control logic is based on pre-set fixed thresholds or simple rules, which is difficult to adapt to the dynamic changes of different space functions, use periods and personalized needs, resulting in rigid allocation of space resources, and inability to realize on-demand, precise environmental service supply and energy efficiency optimization. Therefore, how to build a system and method that can deeply integrate multi-source operation and maintenance data, realize dynamic visualization of space and comfort, and support intelligent adaptive management and control, has become a technical problem to be solved to improve the intelligent level of modern construction engineering operation and maintenance. SUMMARY

[0005] The purpose of the present application is to provide a big data visualization management and control system and method for construction engineering operation and maintenance, to solve the technical problems of single-dimensional data, static and rigid visualization presentation, and control logic that cannot adapt to dynamic changes in the prior art, resulting in low efficiency of building space resource allocation and insufficient personalized environmental service supply.

[0006] To achieve the above purpose, the present application provides a big data visualization management and control system for construction engineering operation and maintenance, comprising: A multi-source heterogeneous data acquisition module is configured to acquire full-dimensional data streams in real time during the operation and maintenance of a building, and comprises a physical environment sensing unit, a space occupancy sensing unit, a personnel behavior analysis unit, and a subjective feedback collection unit. A data fusion and feature engineering module is connected to the multi-source heterogeneous data acquisition module, configured to perform time synchronization, space alignment, cleaning, and fusion processing on the received full-dimensional heterogeneous data, and extract high-order features for representing building space efficiency and personnel comfort state. A space-comfort dynamic visualization engine is connected to the data fusion and feature engineering module, configured to generate and render a three-dimensional visualization scene of building space efficiency and comfort based on the processed multi-modal data cube and extracted high-order features. An intelligent adaptive management and control decision module is connected to the space-comfort dynamic visualization engine and the data fusion and feature engineering module, configured to generate optimization control instructions for building environment control systems and energy systems based on dynamic visualization insights and real-time data features.

[0007] Preferably, the physical environment sensing unit is deployed in each functional subarea of the building to acquire temperature, humidity, light intensity, carbon dioxide concentration, particulate matter concentration, and sub-energy consumption data of each energy-using subsystem. The space occupancy sensing unit integrates a millimeter wave radar array and an infrared thermal imaging sensor network to obtain the precise three-dimensional coordinates and micro-motion features of targets in the space through the millimeter wave radar, and obtain thermal distribution images in the space through the infrared thermal imaging sensor network, and fuse the data to calculate the actual number of people, static and moving state distribution, and space density thermal map of each functional subarea. The personnel behavior analysis unit analyzes the personnel flow path, gathering hot spot area, and stay duration through computer vision algorithms based on anonymized video analysis terminals deployed in public areas and terminal connection data collected by wireless access points, to construct a personnel spatio-temporal behavior pattern map. The subjective feedback collection unit periodically or event-triggered collects subjective evaluation data of building users on the comfort of the local space environment through a lightweight questionnaire and instant scoring interface deployed in a mobile terminal application or a fixed interactive interface, and the evaluation dimensions include thermal sensation, air quality, lighting satisfaction, and overall comfort.

[0008] Preferably, the data fusion and feature engineering module establishes a unified spatio-temporal coordinate system based on the building information model, synchronizes the time stamps of all collected data to a unified time server, and maps the spatial positions of sensors to the spatial subarea numbers corresponding to the building information model. The data fusion and feature engineering module performs data cleaning, eliminating abnormal values and null values caused by sensor failure or communication interruption, and using a linear interpolation method based on a sliding window to fill in the data; At the data fusion level, the data fusion and feature engineering module associates and splices the physical environment data, space occupancy data, and personnel behavior data according to a unified space-time coordinate system, forming a multi-modal data cube with time as the primary key and space partition as the dimension. The feature engineering process extracts two types of feature sets from the multi-modal data cube: space efficiency features and comprehensive comfort index. The space efficiency feature set includes unit area energy consumption intensity, space occupancy rate, personnel flow coefficient, and space function matching degree deviation. The comprehensive comfort index is generated by a multi-layer perceptron model, which takes the physical environment parameter vector, space personnel density value, and average of recent historical subjective scores as input, and outputs a scalar value between 0 and 100.

[0009] Preferably, the space-comfort dynamic visualization engine uses the three-dimensional geometric model and semantic information of the building information model as the basis, and superimposes and renders the space efficiency feature set and the comprehensive comfort index to the corresponding three-dimensional space entity in the form of dynamic texture and particle system. For each building partition, the space-comfort dynamic visualization engine drives a group of dynamic particles representing personnel distribution and flow direction to move within the three-dimensional model of the building partition according to its real-time calculated space occupancy rate and personnel flow coefficient. The particle density and velocity vector reflect the personnel density and flow trend. The color and transparency of the surface of the three-dimensional model of the building partition are dynamically and gradually mapped according to the real-time comprehensive comfort index. The higher the index, the more the color tends to green and the transparency decreases, and the lower the index, the more the color tends to red and the transparency increases, forming a three-dimensional comfort heat map covering the entire building. The space-comfort dynamic visualization engine also generates a series of two-dimensional overlay views, including a space efficiency dashboard with floor as the plane, which associates and displays the unit area energy consumption intensity and function matching degree deviation in the form of a combination chart, and supports time axis playback and spatiotemporal range data drilling analysis.

[0010] Preferably, the intelligent adaptive control decision module has an adaptive decision agent based on deep reinforcement learning. The state space of the adaptive decision-making agent is defined as the environmental parameter vector of all spatial partitions, the spatial performance feature vector, the comprehensive comfort index, and the space reservation schedule information within the next 2 hours; the action space is defined as the adjustment instruction combination of the set value of the air conditioning terminal device of each partition, the air volume of the fresh air unit, the brightness of the lighting circuit, and the angle of the sunshade louver. The reward function is designed as a multi-objective weighted sum, including a comfort reward item, an energy consumption penalty item, and a device action smoothness penalty item. The comfort reward item is positively correlated with the comprehensive comfort index of each partition, and a positive reward is obtained when the index is higher than the preset comfort baseline, and a negative reward is obtained when the index is lower than the comfort baseline. The energy consumption penalty item is positively correlated with the total energy consumption of the system. The device action smoothness penalty item is positively correlated with the change amplitude of the action vector at adjacent time steps. The adaptive decision-making agent optimizes the strategy through a combination of offline historical data training and online incremental learning, with a decision-making period of 15 minutes. In each decision-making period, the agent observes the current state, outputs the optimal action instruction based on its strategy network, and issues it to the corresponding building automation system executor for execution. The intelligent adaptive control decision-making module is provided with a rule verification layer to ensure that the action instruction output by the agent does not violate the safe operation boundary of the device and the highest priority manual forced intervention instruction.

[0011] Preferably, the data fusion process of the millimeter wave radar array and the infrared thermal imaging sensor in the space occupancy perception unit is specifically as follows: Cluster the millimeter wave radar point cloud data, distinguish independent targets, and calculate the three-dimensional coordinates of the centroid of each target. Segment the human body contour in the infrared thermal imaging image, extract the human body heat source area, then perform nearest neighbor matching of the radar target centroid and the heat source area center in a unified coordinate system, assign a personnel label to the targets that match successfully, and fuse the accurate position and micro-motion information provided by the radar and the body surface temperature distribution information provided by the thermal imaging, finally count the number and static distribution of personnel in each partition.

[0012] Preferably, the trigger logic of the lightweight questionnaire in the subjective feedback collection unit is based on context awareness; when the data fusion and feature engineering module detects that the comprehensive comfort index of a certain spatial partition has been continuously decreasing for 3 consecutive decision-making periods and is lower than the threshold, or the personnel density in the spatial partition exceeds 80% of the design density, a short comfort feedback questionnaire is automatically pushed to the authorized mobile terminal application located in the spatial partition, and the questionnaire results are fed back to the data fusion and feature engineering module in real time as key evidence for updating the subjective score input in the comprehensive comfort index calculation model.

[0013] Preferably, the deep reinforcement learning agent in the intelligent adaptive management and control decision module adopts a proximal policy optimization algorithm for training. Both the policy network and the value network are fully connected neural networks containing 3 hidden layers, and the activation function adopts a rectified linear unit. During the training process, the agent pre-trains in a simulation environment constructed using historical operation and maintenance data, and learns a basic regulation and control strategy. After the system goes online, the agent switches to an online learning mode, and the next period state data collected after the actual execution of actions and the calculated reward value form an experience tuple, which is stored in an experience replay buffer, and data is periodically sampled from the buffer for parameter updating of the policy network and the value network, so as to realize continuous adaptive optimization of the policy.

[0014] Preferably, the system further comprises a digital twin simulation deduction module. The digital twin simulation deduction module receives the to-be-executed regulation and control instructions generated by the intelligent adaptive management and control decision module, simulates the execution of the regulation and control instructions in advance in a high-fidelity digital twin model constructed based on a building information model and physical laws, predicts the environmental parameter and energy consumption change trend of each subzone of the building within the next 30 minutes, and superimposes the prediction results in the form of a comparison curve on the interface of the space-comfort dynamic visualization engine for evaluation and confirmation by operation and maintenance personnel, and after confirmation, the results are issued to the real equipment for execution.

[0015] Preferably, the specific calculation process of the reward function is as follows: The comfort reward item is calculated as the sum of the parts of all subzone comprehensive comfort indexes exceeding the comfort baseline; The energy consumption penalty item is positively correlated with the total energy consumption of the system; The device action smoothness penalty item is calculated as the square of the Euclidean distance between the current action vector and the last period action vector; The final reward value is the weighted sum of the comfort reward item, the energy consumption penalty item and the device action smoothness penalty item.

[0016] Compared with the prior art, the present application has the following advantages: 1. The present application integrates millimeter wave radar, infrared thermal imaging, video analysis, wireless terminal detection and subjective feedback interface to construct a full-dimensional, multi-modal data acquisition system covering physical environment, space occupation, personnel behavior and subjective feeling, solving the problem of single data dimension in traditional systems. The unified spatio-temporal coordinate system and multi-modal data cube established by the data fusion and feature engineering module realize the internal correlation and deep fusion of heterogeneous data, providing a data basis for understanding the building operation state.

[0017] 2、The space-comfort dynamic visualization engine provided by the application innovatively combines building information model, dynamic particle system and color gradient mapping technology, generates three-dimensional dynamic heat maps and performance instrument panels that can reflect the distribution of people inside the building, flow trend and spatial difference of comprehensive comfort in real time and intuitively. This visualization method surpasses the expression ability of static charts, enables the operation and maintenance manager to grasp the space-time evolution law of the utilization efficiency and comfort condition of the building space at a glance, and realizes the transparency and immersive insight into the operation state of the building.

[0018] 3、The application adopts an intelligent adaptive control decision module based on deep reinforcement learning, the decision mechanism is no longer a simple rule depending on a fixed threshold, but an adaptive optimization process that can continuously perceive the dynamic state of the building, weigh multiple targets such as comfort and energy consumption, and consider future schedule information. The agent continuously optimizes the strategy through online learning, so that the system can actively adapt to different spatial function differences, use time period changes and personalized demand fluctuations, realize the transformation from homogeneous supply to on-demand precise regulation, and improve the flexibility of building energy utilization efficiency and spatial resource allocation under the premise of ensuring user comfort. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is the overall technical scheme architecture schematic diagram of the application; Figure 2 is the core principle framework schematic diagram of the adaptive decision agent based on deep reinforcement learning in the application; Figure 3 is the logic flow framework diagram of multi-source heterogeneous data acquisition and fusion in the application; Figure 4 is the multi-level interaction relationship and data flow schematic diagram of the space-comfort dynamic visualization engine in the application; Figure 5 is the collaborative work flow schematic diagram of the intelligent adaptive control decision module and the digital twin simulation deduction module in the application. DETAILED DESCRIPTION

[0020] The building engineering operation and maintenance big data visualization control system described in the application has the overall technical architecture as shown in Figure 1 , which includes a multi-source heterogeneous data acquisition module, a data fusion and feature engineering module, a space-comfort dynamic visualization engine and an intelligent adaptive control decision module. The units are tightly coupled through a high-throughput, low-latency data bus, forming a closed-loop control system from perception, fusion, presentation to decision execution. The specific implementation of each component of the system will be described in detail below in combination with Figure 1 to Figure 5 .

[0021] The multi-source heterogeneous data collection module, as the perception front end of the system, is deployed in each functional partition and public area inside the building, and is responsible for real-time collection of full data streams covering four dimensions of physical environment, space occupancy, personnel behavior and subjective feeling. The multi-source heterogeneous data collection module specifically includes a physical environment sensing unit, a space occupancy sensing unit, a personnel behavior analysis unit and a subjective feedback collection unit. The physical environment sensing unit is composed of distributed deployment of temperature and humidity sensors, illumination intensity sensors, carbon dioxide concentration detectors, particulate matter concentration monitors and sub-item electric energy metering devices. All sensors are connected to the building local area network through wired or wireless means, and report the current readings to the data fusion and feature engineering module at a period of every 5 seconds.

[0022] The sensor distribution follows the spatial partition topology in the building information model, and each spatial partition is equipped with at least one complete environmental sensing suite to ensure that the spatial granularity of data collection corresponds to the building functional unit. All sensing data carry accurate time stamps and unique spatial partition identifiers. The time stamp is uniformly time-synchronized by a system-level time server with an error of less than ±10 milliseconds.

[0023] The space occupancy sensing unit is composed of a millimeter wave radar array and an infrared thermal imaging sensor network. The millimeter wave radar array uses a 77GHz frequency band and is deployed at the four corners of the ceiling of each spatial partition, forming an overlapping coverage, which can achieve a three-dimensional positioning accuracy of better than ±5 centimeters for targets in the space, and a micro-motion detection sensitivity of 0.1 millimeter / second. The infrared thermal imaging sensor network is composed of non-cooled thermal imagers with a resolution of not less than 320x240, and is installed in a staggered manner with the millimeter wave radar to avoid mutual interference. The data collection frequency of both types of sensors is 1 Hz. At the data processing level, the millimeter wave radar outputs raw point cloud data, which is separated into independent targets by clustering algorithms such as DBSCAN, and the three-dimensional coordinates of the centroid of each target are calculated; the infrared thermal imaging image is processed by a semantic segmentation model based on the U-Net architecture to extract the human heat source area contour, and the geometric center is calculated.

[0024] In the unified world coordinate system based on the building information model, the radar target centroid and the heat source area center are nearest neighbor matched, and the matching threshold is set to be less than 0.8 meters in Euclidean distance. The matched target is labeled as "personnel", and the accurate position, velocity vector and micro-motion characteristics provided by the radar, as well as the body surface temperature distribution information provided by the thermal imaging are fused, to finally generate a structured data package containing the number of personnel, static / moving state, spatial density thermal map, and upload it to the data fusion and feature engineering module at a frequency of 1 Hz.

[0025] The personnel behavior analysis unit is supported by anonymized video analysis terminals deployed in public areas such as corridors, halls, elevator halls, and the log data of wireless access points in the building. The video analysis terminal adopts an edge computing architecture, with a lightweight YOLOv5s model built-in, only outputting personnel detection box coordinates and tracking ID, and the original video stream is not stored or uploaded, ensuring privacy compliance. The system constructs a cross-regional personnel flow path atlas through a multi-camera cooperative tracking algorithm, records the entry time, exit time and stay duration of each tracking ID in each spatial partition.

[0026] The wireless access point collects the media access control address, signal strength and connection duration of all connected terminal devices, which are anonymized and hashed to assist in verifying the existence and activity range of personnel. After aligning the two types of data in time, the personnel spatio-temporal behavior pattern atlas is generated, including hot area identification, average stay duration distribution, peak period traffic statistics and other indicators, updated once every 15 minutes and pushed to the data fusion and feature engineering module.

[0027] The subjective feedback collection unit is implemented through a lightweight interface integrated into the building user's mobile terminal application or a wall-mounted interactive screen. The lightweight interface provides four-dimensional ratings: thermal sensation, air quality, lighting satisfaction and overall comfort. The questionnaire trigger logic is driven by a context awareness mechanism: when the data fusion and feature engineering module detects that the comprehensive comfort index of a spatial partition has continuously decreased for 3 consecutive decision cycles and is below the threshold of 75, or the real-time personnel density of the spatial partition exceeds 80% of the designed density, the system automatically pushes a pop-up questionnaire to the authorized mobile terminals located in the spatial partition. The questionnaire response data is transmitted immediately after encryption, and then injected into the input buffer of the feature engineering module to dynamically correct the calculation weights of the comprehensive comfort index, ensuring that subjective feelings can affect the system evaluation in real time.

[0028] The data fusion and feature engineering module receives all raw data streams from the multi-source heterogeneous data acquisition module, and performs a five-stage processing flow of time synchronization, spatial alignment, cleaning, fusion and feature extraction. The data fusion and feature engineering module establishes a unified space-time coordinate system based on the building information model, with the building main entrance as the origin, the X-axis along the east-west direction, the Y-axis along the south-north direction, and the Z-axis vertically upward. All sensor positions are mapped to three-dimensional coordinates in the space-time coordinate system through the component ID in the building information model. Time synchronization is guaranteed by a network time protocol server, and all data packets are calibrated to a unified time reference before entering the data fusion and feature engineering module, with a maximum time delay jitter of not more than 50 milliseconds.

[0029] The data cleaning stage adopts a double-layer filtering mechanism. The first layer is hard threshold filtering, which removes values that are obviously beyond the physical reasonable range. The second layer is sliding window anomaly detection, which uses a sliding window of length 12 sampling points (i.e. 60 seconds) to calculate the standard deviation of the data in the window. If the current point deviates from the window mean by more than 3 times the standard deviation, it is determined to be abnormal and is removed. For null values or data gaps after removal, a linear interpolation method based on adjacent valid points is used to fill in the gaps, with a maximum interpolation window span of 3 sampling periods.

[0030] In the data fusion stage, the module data fusion and feature engineering module associates and splices the physical environment data, space occupancy data, and personnel behavior data according to a unified space-time coordinate system to form a multi-modal data cube with time as the primary key and space partition as the dimension. Each data unit of the multi-modal data cube contains the following fields: timestamp, space partition ID, temperature, humidity, light intensity, carbon dioxide concentration, particulate matter concentration, sub-energy consumption, number of people, number of stationary people, number of moving people, space density, personnel flow path sequence, stay duration distribution, and recent subjective score average. The multi-modal data cube is stored in a time series database in columnar storage format, supporting millisecond-level query response.

[0031] The feature engineering process extracts two types of high-order feature sets from the multi-modal data cube: space efficiency features and comprehensive comfort index. The space efficiency feature set includes four indicators: unit area energy consumption intensity, calculated as the current sub-energy consumption of the space partition divided by its building area; space occupancy rate, defined as the ratio of the current number of people to the designed maximum capacity; personnel flow coefficient, defined as the net flow of people in and out of the space partition per unit time divided by the average number of people present; and space function matching degree deviation, which calculates the mode difference by comparing the current actual use mode with the design function marked in the building information model, with a larger deviation indicating a higher degree of function mismatch.

[0032] The comprehensive comfort index is generated by a multi-layer perceptron model. The input of the multi-layer perceptron model is the physical environment parameter vector at the current time, the space personnel density value, and the average of the recent historical subjective scores, with a total of 7-dimensional input. The output is a scalar value between 0 and 100. The model structure is a 3-layer fully connected network with hidden layer neuron counts of 16, 8, and 4 respectively, using a rectified linear unit as the activation function, and a Sigmoid function for scaling to the target interval in the output layer. The model parameters are obtained by offline training of historical data, and are fine-tuned online during system operation based on newly collected subjective feedback data, with a learning rate of 0.001.

[0033] The space-comfort dynamic visualization engine takes the three-dimensional geometric model and semantic information of the building information model as the foundation, receives the multi-modal data cubes and high-order features from the data fusion and feature engineering module, and generates dynamic and interactive three-dimensional visualization scenes. Please refer to the attached Figure 4 The rendering pipeline of the space-comfort dynamic visualization engine includes five sub-stages: geometry loading, feature mapping, particle driving, color gradient, and two-dimensional overlay. The building information model is imported in IFC format, and after lightweight processing, it is loaded into the WebGL rendering space-comfort dynamic visualization engine, retaining all the semantic properties of the spatial partitions.

[0034] For each building partition, the space-comfort dynamic visualization engine drives a set of dynamic particles to move within its three-dimensional model based on its real-time calculated space occupancy rate and personnel flow coefficient. When the particle system is initialized, the total number of particles is proportional to the current number of personnel, with a proportionality coefficient of 1.2. The initial position of the particles is randomly sampled based on the personnel distribution heat map provided by the space occupancy perception unit. The particle velocity vector is determined by the personnel flow coefficient and the historical flow direction: if the flow coefficient is greater than 0.3, the particle moves at a constant speed along the main flow direction; otherwise, the particle does Brownian motion in a small local range, simulating the static state. The particle color uses a blue-white-red gradient color band, with blue representing static, red representing high-speed movement, and white representing the intermediate state.

[0035] The color and transparency of the three-dimensional model surface of the building partition are dynamically mapped based on the real-time comprehensive comfort index. The mapping rules are as follows: when the index is ≥90, the model surface is pure green (RGB:0,255,0) and the transparency is 0%; when the index is between 75 and 90, the color gradually changes from yellow-green to green and the transparency linearly decreases from 20% to 0%; when the index is between 60 and 75, the color is yellow (RGB:255,255,0) and the transparency is 30%; when the index is <60, the color gradually changes from orange to red (RGB:255,0,0) and the transparency increases from 40% to 60%. This dynamic mapping strategy makes the low comfort area appear "semi-transparent red" in vision, facilitating quick identification of problem spaces.

[0036] The space-comfort dynamic visualization engine also generates a series of two-dimensional overlay views, including a space performance dashboard based on the floor plane. The space performance dashboard uses a combination chart form, with a column chart on the left showing the energy intensity per unit area of each partition, and a line chart on the right showing the functional matching degree deviation, both sharing the same horizontal axis (space partition list). Users can replay data from any past time period through the time axis control, or perform spatiotemporal range data drilling through the box selection operation, and the drilling results will be displayed in the form of a pop-up window, showing detailed statistical indicators and trend curves for the subset.

[0037] The intelligent adaptive control decision module is as followsFigure 2 and Figure 5 An adaptive decision-making agent based on deep reinforcement learning is shown. The state space of the adaptive decision-making agent is defined as the environmental parameter vector of all N spatial partitions at the current time, the spatial performance feature vector, the comprehensive comfort index, and the space reservation schedule information within the next 2 hours, with a total dimension of 11N+1. The action space is defined as the adjustment instruction combination of the set temperature of the air conditioning terminal device of each partition, the air volume of the fresh air handling unit, the brightness of the lighting circuit, and the angle of the sunshade louver, with a total dimension of N+M+P+Q. The reward function is designed as a multi-objective weighted sum:

[0038] The comfort reward term is calculated as the sum of the parts of all partition comprehensive comfort indexes exceeding the comfort baseline of 85; is the total energy consumption of the system; is the action space of the last control cycle; is the device action smoothness penalty term, which measures the squared Euclidean distance between the current action and the action of the last cycle; the weight coefficient is determined through multiple rounds of simulation experiments.

[0039] The adaptive decision-making agent uses the proximal policy optimization algorithm for training. The policy network and the value network are both fully connected neural networks containing 3 hidden layers, with 256, 128, and 64 neurons per layer, respectively, and the activation function uses a rectified linear unit. The training is divided into two stages: in the first stage, the historical operation and maintenance data of the past year are used to build a Markov decision process simulation environment for offline pre-training, with a cumulative interaction step of 1 million steps; in the second stage, after the system goes online, it switches to an online learning mode, and the agent executes a decision-making cycle every 15 minutes. In each cycle, the agent observes the current state, samples actions through the policy network, and after checking by the rule verification layer, it issues the instructions to the building automation system. The rule verification layer ensures that all action instructions meet the safety boundaries of the devices and do not override the highest priority manual forced intervention instructions (such as fire linkage signals).

[0040] After executing the actions, the system collects the next cycle state, calculates the reward, forms the experience tuples, and stores them in an experience replay buffer with a capacity of 10,000. Every 10 decision-making cycles, the agent randomly samples 512 tuples from the buffer, updates the policy network and value network parameters using the Adam optimizer, and sets the learning rate to 3×10 -4 This double filtering mechanism enables the agent to continuously adapt to seasonal changes and unexpected events in building usage patterns.

[0041] As an important enhancement component of the system, the digital twin simulation deduction module is used to Figure 5As shown, the received intelligent adaptive control decision module generates a to-be-executed control instruction, and performs advanced simulation in a high-fidelity digital twin model. The digital twin model is constructed based on a building information model, integrates physical laws such as building envelope thermal performance parameters, heating, ventilation and air conditioning system equipment characteristic curves, lighting system power model, and solves the energy balance equation using the finite difference method. The simulation step is 1 minute, and the prediction is for the temperature, humidity, carbon dioxide concentration and energy consumption trend of each partition in the next 30 minutes. The prediction result is displayed in the form of a comparison curve on the interface of the space-comfort dynamic visualization engine, with solid lines representing predicted values and dashed lines representing current actual values. The operation and maintenance personnel can confirm or reject the control instruction on this interface, and the instruction is only issued to the real equipment for execution after confirmation, thereby forming a safe closed loop of "prediction-evaluation-confirmation-execution".

[0042] In summary, the embodiment builds a closed-loop system of full-dimensional perception, deep fusion, dynamic visualization and intelligent decision-making, realizes efficient allocation of building space resources and precise supply of personalized environmental services, and improves the intelligent level of building operation and maintenance and user experience.

Claims

1. A big data visualization management and control system for building engineering operation and maintenance, characterized in that: include: A multi-source heterogeneous data acquisition module is used to collect full-dimensional data streams in the building operation and maintenance process in real time. The multi-source heterogeneous data acquisition module includes a physical environment sensing unit, a space occupancy sensing unit, a personnel behavior analysis unit, and a subjective feedback collection unit. The data fusion and feature engineering module is connected to the multi-source heterogeneous data acquisition module. It is used to perform time synchronization, spatial alignment, cleaning and fusion processing on the received multi-dimensional heterogeneous data, and extract high-order features to characterize the building space efficiency and human comfort status. The space-comfort dynamic visualization engine, connected to the data fusion and feature engineering module, is used to generate and render a three-dimensional visualization scene of building space performance and comfort based on the processed multimodal data cube and extracted high-order features. The intelligent adaptive control and management decision module is connected to the space-comfort dynamic visualization engine and the data fusion and feature engineering module. It is used to generate optimized control instructions for the building environment control system and energy system based on dynamic visualization insights and real-time data features.

2. The building engineering operation and maintenance big data visualization management and control system according to claim 1, characterized in that, The physical environment sensing units are deployed in various functional areas of the building to collect data on temperature, humidity, light intensity, carbon dioxide concentration, particulate matter concentration, and energy consumption data of each energy-consuming subsystem. The space occupancy sensing unit integrates a millimeter-wave radar array and an infrared thermal imaging sensor network. It is used to obtain the precise three-dimensional coordinates and micro-motion characteristics of targets in space through millimeter-wave radar, and to obtain the thermal distribution image in space through the infrared thermal imaging sensor network. The two data are fused to calculate the actual number of people in each functional area, the distribution of stationary and moving states, and the spatial density heat map. The personnel behavior analysis unit uses computer vision algorithms to analyze personnel flow paths, gathering hotspots, and dwell time based on terminal connection data collected by anonymized video analysis terminals deployed in public areas and wireless access points, in order to construct a spatiotemporal behavior pattern map of personnel. The subjective feedback collection unit collects subjective evaluation data of building users on the comfort of their current local spatial environment periodically or in an event-triggered manner through lightweight questionnaires and instant scoring interfaces deployed on mobile terminal applications or fixed interactive interfaces. The evaluation dimensions cover thermal sensation, air quality, lighting satisfaction and overall comfort.

3. The building engineering operation and maintenance big data visualization management and control system according to claim 2, characterized in that, The data fusion and feature engineering module establishes a unified spatiotemporal coordinate system based on the building information model, synchronizes the timestamps of all collected data to a unified time server, and maps the spatial location of the sensor to the spatial partition number corresponding to the building information model. The data fusion and feature engineering module performs data cleaning, removes outliers and null values ​​caused by sensor failures or communication interruptions, and uses a sliding window-based linear interpolation method to complete the data. At the data fusion level, the data fusion and feature engineering module associates and splices physical environment data, space occupancy data, and personnel behavior data according to a unified spatiotemporal coordinate system to form a multimodal data cube with time as the primary key and spatial partitioning as the dimension. The feature engineering process extracts two types of feature sets from the multimodal data cube: spatial performance feature set and comprehensive comfort index; The spatial performance feature set includes energy consumption intensity per unit area, space occupancy rate, personnel flow coefficient, and spatial function matching deviation. The comprehensive comfort index is calculated and generated by a multilayer perceptron model. The input of the multilayer perceptron model is the physical environment parameter vector at the current moment, the spatial personnel density value, and the mean of recent historical subjective scores. The output is a scalar value between 0 and 100.

4. The building engineering operation and maintenance big data visualization management and control system according to claim 3, characterized in that, The space-comfort dynamic visualization engine uses the three-dimensional geometric model and semantic information of the building information model as a base, and overlays and renders the space performance feature set and comprehensive comfort index onto the corresponding three-dimensional space entity in the form of dynamic textures and particle systems. For each building zone, the space-comfort dynamic visualization engine drives a set of dynamic particles representing the distribution and flow direction of people to move within the three-dimensional model of the building zone based on its real-time calculated space occupancy rate and personnel flow coefficient. The particle density and velocity vector reflect the personnel density and flow trend. The color and transparency of the surface of the 3D model of the building's partitions are dynamically mapped according to the real-time comprehensive comfort index. The higher the index, the more the color leans towards green and the lower the transparency. The lower the index, the more the color leans towards red and the higher the transparency, thus forming a 3D comfort heat map covering the entire building. The space-comfort dynamic visualization engine also generates a series of two-dimensional overlay views, including a space efficiency dashboard with the floor as the plane. This dashboard displays the energy consumption intensity per unit area and the functional matching deviation in a combined chart format, and supports playback by time axis and data drill-down analysis across spatiotemporal ranges.

5. The building engineering operation and maintenance big data visualization management and control system according to claim 4, characterized in that, The intelligent adaptive control and management decision-making module has a built-in adaptive decision-making intelligent agent based on deep reinforcement learning; The state space of the adaptive decision-making agent is defined as the environmental parameter vector, spatial efficiency feature vector, comprehensive comfort index, and spatial reservation schedule information for all spatial partitions at the current moment, as well as the spatial reservation schedule information for the next 2 hours. The action space is defined as a combination of adjustment commands for the set values ​​of the air conditioning terminal equipment in each zone, the air volume of the fresh air unit, the brightness of the lighting circuit, and the angle of the shading louvers. The reward function is designed as a multi-objective weighted sum, including a comfort reward term, an energy consumption penalty term, and a device motion smoothness penalty term; The comfort reward item is positively correlated with the comprehensive comfort index of each zone. When the index is higher than the preset comfort baseline, a positive reward is obtained, and when it is lower than the comfort baseline, a negative reward is obtained. The energy consumption penalty term is positively correlated with the total energy consumption of the system. The device motion smoothness penalty term is positively correlated with the change amplitude of the motion vector in adjacent time steps; The adaptive decision-making agent optimizes its strategy by combining offline historical data training with online incremental learning, with a decision cycle of 15 minutes. In each decision cycle, the agent observes the current state and outputs the optimal action command based on its policy network, which is then sent to the corresponding building automation system actuator. The intelligent adaptive control and management decision-making module is equipped with a rule verification layer to ensure that the action commands output by the intelligent agent do not violate the equipment's safe operation boundaries and the highest priority manual intervention commands.

6. The building engineering operation and maintenance big data visualization management and control system according to claim 5, characterized in that, The data fusion process between the millimeter-wave radar array and the infrared thermal imaging sensor in the space occupancy sensing unit is as follows: Clustering is performed on millimeter-wave radar point cloud data to distinguish independent targets, and the three-dimensional coordinates of the centroid of each target are calculated. Human body contour segmentation is performed on infrared thermal imaging images to extract the heat source region of the human body; Then, in a unified coordinate system, the centroid of the radar target is matched with the center of the heat source area for nearest neighbor matching. The successfully matched targets are assigned personnel tags, and the precise location and micro-motion information provided by the radar and the body surface temperature distribution information provided by thermal imaging are fused together to finally calculate the number and static distribution of personnel in each zone.

7. The building engineering operation and maintenance big data visualization management and control system according to claim 6, characterized in that, The triggering logic of the lightweight questionnaire in the subjective feedback collection unit is based on context awareness. When the data fusion and feature engineering module detects that the comprehensive comfort index of a certain spatial partition has been continuously decreasing and falling below the threshold for three consecutive decision cycles, or that the personnel density of the spatial partition exceeds 80% of the design density, it automatically pushes a short comfort feedback questionnaire to the authorized mobile terminal application located in the spatial partition. The questionnaire results are fed back to the data fusion and feature engineering module in real time as key evidence to update the subjective rating input in the comprehensive comfort index calculation model.

8. The building engineering operation and maintenance big data visualization management and control system according to claim 7, characterized in that, The deep reinforcement learning agent in the intelligent adaptive control and management decision-making module is trained using a proximal policy optimization algorithm. Both the policy network and the value network are fully connected neural networks with three hidden layers, and the activation function is a rectified linear unit. During training, the agent is pre-trained using a simulation environment built from historical operation and maintenance data to learn basic control strategies. After the system goes online, the agent enters online learning mode. It uses the state data of the next cycle collected after the actual action is performed and the calculated reward value to form an experience tuple, which is stored in the experience replay buffer. The agent also periodically samples data from the buffer to update the parameters of the policy network and the value network, so as to achieve continuous adaptive optimization of the policy.

9. The building engineering operation and maintenance big data visualization management and control system according to claim 8, characterized in that, The system also includes a digital twin simulation and deduction module; The digital twin simulation and deduction module receives the control instructions to be executed generated by the intelligent adaptive control and management decision module. It simulates the execution of the control instructions in advance in a high-fidelity digital twin model built based on building information model and physical laws. It predicts the environmental parameters and energy consumption change trends of each zone of the building in the next 30 minutes and displays the prediction results as a comparison curve on the interface of the space-comfort dynamic visualization engine for operation and maintenance personnel to evaluate and confirm. After confirmation, it is sent to the real equipment for execution.

10. The building engineering operation and maintenance big data visualization management and control system according to claim 9, characterized in that, The specific calculation process of the reward function is as follows: The comfort bonus is calculated as the sum of the portions of the overall comfort index of all zones that exceed the comfort baseline; The energy consumption penalty term is positively correlated with the total energy consumption of the system. The device motion smoothness penalty term is calculated as the square of the Euclidean distance between the current motion vector and the motion vector of the previous cycle. The final reward value is the weighted sum of the comfort reward value minus the energy consumption penalty value and the device motion smoothness penalty value.

Images