Central air conditioning system operation and maintenance method and system based on artificial intelligence diagnosis and regulation

Abstract

This invention discloses an operation and maintenance method and system for central air conditioning systems based on artificial intelligence diagnosis and control, belonging to the field of intelligent operation and maintenance technology for central air conditioning systems. The core steps of the method are: collecting and preprocessing multi-dimensional parameters; fusing data using an improved norm distance algorithm; identifying sub-health conditions of equipment based on the LSTM algorithm; diagnosing energy efficiency anomalies with a target COP ≥ 9.22 through a fusion model of Q-learning and gradient boosting decision trees; generating and executing the optimal control strategy using a digital twin model, implementing a three-level early warning system; iteratively optimizing parameters and strategies after verifying the control effect, supporting remote upgrades via BLRAT. The system includes corresponding functional modules. This invention achieves integrated sub-health identification, accurate diagnosis, and intelligent optimization, solving pain points such as low energy efficiency and delayed diagnosis in traditional operation and maintenance, improving operational efficiency, reducing energy consumption and maintenance costs, and balancing energy conservation and compliance.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for central air conditioning systems, and in particular to an operation and maintenance method and system for central air conditioning systems based on artificial intelligence diagnosis and control. Background Technology

[0002] Central air conditioning systems are the core unit of building energy consumption in industrial parks, accounting for 30%-50% of the total energy consumption. Their operational efficiency directly affects the park's carbon emission reduction targets and maintenance costs. Currently, central air conditioning operation and maintenance generally suffers from the following pain points: First, traditional operation and maintenance relies on manual periodic inspections and uses a fixed parameter control mode, which cannot respond to environmental changes and load fluctuations in real time, resulting in low energy efficiency and serious energy waste. Second, the diagnosis of energy efficiency anomalies is lagging, often relying on passive troubleshooting after equipment failure and shutdown, failing to identify sub-health conditions such as heat exchanger scaling, fan bearing wear, and minor refrigerant leaks, which can easily lead to major failures and increase maintenance costs. Third, monitoring data is fragmented, focusing only on single equipment operating parameters without integrating multi-dimensional data such as environment and load. Furthermore, traditional sensor monitoring suffers from poor anti-interference capabilities and insufficient accuracy, resulting in low diagnostic accuracy and a lack of scientific optimization strategies. Fourth, there is a lack of closed-loop control mechanisms, leading to a disconnect between diagnostic results and optimization execution, making it impossible to achieve full-link automation of "diagnosis-optimization-verification".

[0003] While some existing intelligent operation and maintenance solutions incorporate simple sensing and algorithm analysis, they have significant limitations: for example, the PID control or fixed-frequency regulation strategies used are difficult to adapt to the dynamic coupling characteristics of the system and are prone to getting trapped in local optima; the algorithms are simple, have weak generalization ability, and cannot adapt to complex operating conditions; they do not incorporate the latest industry standards, making it difficult to balance energy-saving effects and compliance; and the optimization range is limited, failing to achieve full-link collaborative optimization from cold source to terminal. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an operation and maintenance method and system for central air conditioning systems based on artificial intelligence diagnosis and control. It is mainly used for intelligent operation and maintenance of central air conditioning systems, so as to achieve the goal of integrated management of sub-health identification, accurate diagnosis, intelligent adjustment and continuous optimization by integrating cutting-edge algorithms and high-precision sensing technology.

[0005] This invention discloses an operation and maintenance method for a central air conditioning system based on artificial intelligence diagnosis and control, comprising the following steps:

[0006] S1. Multi-dimensional parameter dataset collection; acquire the device's operating parameters, environmental parameters, load parameters, and energy efficiency parameters to generate a multi-dimensional parameter dataset;

[0007] S2. Data fusion processing: An improved norm distance algorithm is used to associate and fuse multi-source heterogeneous data in a multi-dimensional parameter dataset to generate a standardized dataset;

[0008] S3. Sub-health status identification: The LSTM algorithm is used to build a time series prediction model. The model learns the parameter change pattern under normal operating conditions by using the historical parameter dataset of the equipment during normal operation, so as to generate a predicted value dataset for normal operation. By analyzing the deviation between the multi-dimensional parameter dataset and the predicted value dataset, the sub-health status of the equipment is identified.

[0009] S4. Energy efficiency anomaly diagnosis: An energy efficiency anomaly diagnosis model is constructed by integrating reinforcement learning Q-learning algorithm and gradient boosting decision tree. With COP≥9.22 as the target threshold, a standardized dataset is input, and the energy efficiency anomaly type is diagnosed and the degree of anomaly impact is calculated based on the preset energy efficiency anomaly dataset.

[0010] S5. Abnormal Handling; Based on the diagnostic results of steps S3 and S4, generate and automatically execute control strategies; including steps,

[0011] S501. Generate regulation strategy; Based on the sub-health state identification results and energy efficiency anomaly diagnosis results, combine digital twin model to simulate the operation effect of different regulation parameter combinations, compare the operation effect of different regulation parameter combinations, and generate the optimal regulation strategy.

[0012] S502. Automatically execute the control strategy; send the generated optimal control strategy to the central air conditioning control system, drive the execution unit to automatically adjust the operating parameters to the control parameters corresponding to the optimal control strategy;

[0013] S6. Validation of the regulation strategy: The multi-dimensional parameter data collected in S1, the sub-health identification results in S3, the energy efficiency anomaly diagnosis results in S4, the optimal regulation strategy in S5, and the energy efficiency changes after regulation are stored in the cloud database. The operation effect of the regulation strategy is validated by combining the digital twin model.

[0014] S7. Iteratively optimize parameters and control strategies; repeat steps S1-S6, and continuously iteratively optimize historical parameter datasets, energy efficiency anomaly datasets, control parameter combinations, and control strategies by analyzing historical data.

[0015] Furthermore, in step S1, the programmable logic controller built into the edge computing gateway is used to clean, reduce noise, standardize the format, and perform local basic control on the collected multi-dimensional parameter data, maintaining basic temperature control function in the offline state; wherein, the preprocessing is the preprocessing before the data fusion processing in step S2.

[0016] Furthermore, in the data fusion processing step S2, the existing central air conditioning control system is connected via BAC net IP and MQTT protocols.

[0017] Furthermore, in step S3, when the deviation is greater than a preset deviation threshold, the device is determined to be in a sub-healthy state; if the deviation is less than or equal to the preset deviation threshold, the device is determined not to be in a sub-healthy state.

[0018] Furthermore, in step S4, after diagnosing the type of energy efficiency anomaly and calculating the degree of its impact, the energy efficiency anomaly diagnosis model is continuously updated and optimized using a Q-table.

[0019] Furthermore, before step S501, a tiered early warning system is also included; three levels of early warning thresholds are preset, and corresponding early warning measures are implemented according to the risk level of the sub-health state and the severity of energy efficiency abnormalities.

[0020] The three-level early warning thresholds include attention level, early warning level, and alarm level;

[0021] When at the attention level, the parameters are automatically adjusted by the execution unit;

[0022] When the warning level is reached, a notification message is pushed to the maintenance personnel, and the control strategy is automatically executed through the execution unit. At the same time, the maintenance personnel go to the location to re-inspect the fault point.

[0023] When the alarm level is triggered, an audible and visual alarm is activated, and a maintenance work order is sent to the maintenance personnel for adjustment and maintenance.

[0024] Furthermore, in step S501, the control parameters include chilled water flow rate adjustment, fan speed adjustment, compressor operating frequency adjustment, dynamic adaptation of fresh air introduction ratio, and frequency conversion adjustment of water pump system.

[0025] Furthermore, in step S502, the execution unit includes an electric regulating valve and a frequency converter. The electric regulating valve is used to adjust the chilled water flow rate, chilled water flow rate, and fresh air introduction ratio; the frequency converter is used to adjust the compressor operating frequency; and the water pump system adopts an online adaptive frequency conversion control strategy based on Q-learning.

[0026] Furthermore, in step S7, during the iterative optimization of the historical parameter dataset, energy efficiency anomaly dataset, control parameter combination, and control strategy, remote configuration of control logic and firmware upgrades based on BLRAT technology are supported.

[0027] This invention also discloses an operation and maintenance system for a central air conditioning system based on artificial intelligence diagnosis and control, used to implement the operation and maintenance method for a central air conditioning system based on artificial intelligence diagnosis and control as described above, including:

[0028] The multi-dimensional data acquisition module is configured to execute step S1;

[0029] The data fusion processing module is configured to execute step S2;

[0030] The sub-health status identification module is configured to execute step S3;

[0031] The energy efficiency anomaly diagnosis module is configured to execute step S4;

[0032] The exception handling module is configured to execute step S5;

[0033] The control strategy verification module is configured to execute step S6;

[0034] The optimization iteration module is configured in execution step S7.

[0035] The beneficial effects of this invention are:

[0036] 1. Dynamic adaptive control addresses the root causes of low energy efficiency and energy waste. This invention abandons the traditional fixed-parameter operation mode. By sensing environmental and load changes in real time from multiple dimensions and combining AI algorithms, it achieves dynamic adaptive adjustment of operating parameters. It can quickly respond to load fluctuations, outdoor weather changes, and time-of-use electricity price signals, completely solving the problems of low energy efficiency and large waste caused by the inability of traditional operation and maintenance to adapt to operating conditions in real time, and significantly improving the overall operating efficiency of the central air conditioning system.

[0037] 2. Early identification of sub-health conditions enables a shift from reactive emergency repairs to proactive maintenance. Based on an LSTM time-series prediction model, sub-health conditions such as heat exchanger scaling, fan bearing wear, and refrigerant micro-leakage can be identified 7-15 days before equipment failure. Combined with high-precision sensing, the early warning accuracy rate is ≥95%, effectively preventing minor issues from escalating into major failures, significantly reducing the risk of unplanned downtime and equipment maintenance costs, and extending equipment lifespan.

[0038] 3. Deep fusion of multi-source data significantly improves the accuracy of anomaly diagnosis and the scientific nature of optimization. An improved norm distance algorithm is used to achieve unified fusion of multi-dimensional data on equipment, environment, load, and energy efficiency. Combined with high-precision anti-interference sensing technology, it solves the shortcomings of traditional monitoring data, such as fragmentation, low accuracy, and poor anti-interference. This provides high-quality data support for AI diagnosis, making energy efficiency anomaly judgment more accurate and optimization strategies more scientific.

[0039] 4. Construct a fully automated closed-loop autonomous system to automate the "diagnosis-optimization-verification" process. From data collection, AI diagnosis, intelligent decision-making, and command execution to effect verification and iterative optimization, a complete closed loop is formed. Diagnostic results and optimization commands are linked in real time, eliminating the need for manual intervention. This solves the problems of disconnect between diagnosis and execution and low operational efficiency in traditional solutions, enabling a highly autonomous system operation.

[0040] 5. Hybrid AI algorithms for collaborative optimization overcome the limitations of traditional control strategies and avoid local optima. By integrating LSTM time-series prediction, reinforcement learning Q-learning, and gradient boosting decision trees, compared to single PID, fixed-frequency control, or simple AI algorithms, it possesses stronger nonlinear fitting and dynamic adaptive capabilities. It can handle the strong coupling, large lag, and variable operating conditions of central air conditioning systems, achieving global optima rather than local optima, and significantly improving generalization ability.

[0041] 6. Adhering to the latest industry standards, energy-saving effects and compliance are achieved simultaneously. Diagnostic thresholds and energy efficiency targets are dynamically calibrated based on the latest industry standards, meeting compliance requirements while achieving high energy efficiency. This solves the problem of existing technologies struggling to balance energy saving and compliance, and can be directly integrated with zero-carbon industrial parks and smart building evaluation systems.

[0042] 7. Comprehensive optimization across the entire cooling source-terminal chain, resulting in wider coverage and higher energy savings. The optimization strategy covers the entire process from chiller units, pumps, fans, fresh air systems, to terminal valves, achieving coordinated optimization of cooling source generation, distribution, and terminal consumption. This breaks through the bottleneck of traditional solutions that only offer localized optimization and have limited energy-saving potential. With a COP ≥ 9.22 as the optimization target, energy consumption per unit of cooling capacity can be reduced by up to 22%.

[0043] 8. Remote iteration and continuous self-optimization extend system lifecycle and reduce long-term investment. Supports remote configuration of control logic, firmware and algorithm model upgrades via BLRAT, and continuously iterates parameter, exception and strategy libraries based on historical data, making the system more and more accurate with operation. No on-site maintenance or repeated modifications are required, and it adapts to equipment aging, scenario changes and standard upgrades. Its long-term economic efficiency and applicability are significantly better than traditional solutions. Attached Figure Description

[0044] Figure 1 The present invention discloses a flowchart of an operation and maintenance method for a central air conditioning system based on artificial intelligence diagnosis and control. Detailed Implementation

[0045] The present invention will be further described below.

[0046] This invention provides a central air conditioning system operation and maintenance method based on artificial intelligence diagnosis and control, mainly used for intelligent operation and maintenance of central air conditioning systems. By collecting multi-dimensional parameter datasets and utilizing data fusion processing, sub-health state identification, and energy efficiency anomaly diagnosis technologies, it generates and executes optimal control strategies. Finally, it verifies the control effect and iteratively optimizes parameters and control strategies to achieve integrated management of sub-health identification, accurate diagnosis, intelligent adjustment, and continuous optimization by integrating cutting-edge algorithms and high-precision sensing technologies. The method includes the following steps:

[0047] S1. Multi-dimensional parameter dataset collection: Acquire the device's operating parameters, environmental parameters, load parameters, and energy efficiency parameters to generate a multi-dimensional parameter dataset.

[0048] Furthermore, in step S1, the programmable logic controller built into the edge computing gateway is used to clean, reduce noise, standardize the format, and perform local basic control on the collected multi-dimensional parameter data, maintaining basic temperature control function in the offline state; wherein, the preprocessing is the preprocessing before the data fusion processing in step S2.

[0049] The above operating parameters include the inlet and outlet temperatures and pressures of the chiller unit, compressor current, fan speed, inverter frequency, and refrigerant level; the above environmental parameters include indoor and outdoor temperature and humidity, solar radiation intensity, atmospheric pressure, and CO2 concentration; the above load parameters include the building's occupancy density, cooling demand in each area, and time-of-use electricity price; and the above energy efficiency parameters include COP value, SCOP value, and energy consumption per unit of cooling capacity. Specifically, by deploying a distributed high-precision sensor array and an edge computing gateway, real-time acquisition and local preprocessing of multi-dimensional data are achieved. For temperature acquisition, a PT1000 platinum resistance sensor with integrated cold junction compensation is used, achieving a measurement accuracy of ±0.1℃. These sensors are deployed at key points such as chilled water supply and return, cooling water supply and return, indoor and outdoor air, and heat exchanger inlet and outlet. For pressure acquisition, a diffused silicon pressure sensor with an accuracy of ±0.5% is used, deployed at key points such as unit inlet and outlet, water pump inlet and outlet, and main pipeline. For power acquisition, a three-phase intelligent power acquisition module is used to collect current, voltage, active power, power factor, and cumulative energy value in real time. To suppress electromagnetic interference, differential input technology is used for all the above sensor acquisitions, and shielded cables with single-point grounding are used for sensor signals to adapt to the strong interference environment of industrial sites. The collection frequency and timing rules for the above parameters are as follows: operating parameters, such as water temperature, water pressure, and current, are collected once every 1 second; environmental parameters are collected once every 10 seconds; and energy efficiency statistics parameters are collected once every 1 minute. All data are accompanied by precise timestamps to ensure time alignment of multi-source data.

[0050] Before data is uploaded to the cloud, the edge computing gateway's built-in programmable logic controller (PLC) performs local preprocessing. Specifically, this includes: data cleaning: removing out-of-range, jump-type, and missing abnormal data, filling in short-term missing values, and ensuring time sequence continuity; noise reduction: using moving average filtering and median filtering to eliminate power frequency interference and mechanical vibration noise; format standardization: unifying timestamps, units, and data structures to output in a standard format; local basic control: in the event of a network outage or cloud offline state, the gateway independently maintains basic temperature control and safety protection to ensure that the equipment does not stop, does not exceed temperature, and does not exceed pressure. The edge gateway provides unified time synchronization, enabling snapshot-style data acquisition at the same time; it supports multiple protocols such as Modbus RTU / TCP, BAC net IP, MQTT, and OPCUA, directly connecting to central air conditioning controllers and building automation systems (BAS); and it adopts a distributed acquisition architecture, with independent data acquisition and uploading from the nearest location in the computer room, on each floor, and at the terminal, reducing cabling and transmission delays.

[0051] S2. Data fusion processing: An improved norm distance algorithm is used to associate and fuse multi-source heterogeneous data in a multi-dimensional parameter dataset to generate a standardized dataset.

[0052] Furthermore, it connects to the existing central air conditioning control system via BAC net IP and MQTT protocols.

[0053] An improved norm distance algorithm is employed to perform time alignment, dimensional normalization, correlation matching, and outlier removal on four types of multi-source heterogeneous data: equipment, environment, load, and energy efficiency. This results in a standardized dataset with unified timestamps, units, and structure, addressing the issues of inaccurate AI diagnostics caused by data asynchrony, inconsistent units, and weak correlations. During the integration process, BAC net IP and MQTT protocols are used to interface with the existing central air conditioning control system and building automation system (BAS), enabling data exchange and command issuance.

[0054] S3. Sub-health status identification: The LSTM algorithm is used to build a time series prediction model. The model learns the parameter change pattern under normal operating conditions by using the historical parameter dataset of the equipment during normal operation to generate a predicted value dataset for normal operation. By analyzing the deviation between the multi-dimensional parameter dataset and the predicted value dataset, the sub-health status of the equipment is identified.

[0055] Furthermore, in step S3, when the deviation is greater than a preset deviation threshold, the device is determined to be in a sub-healthy state; if the deviation is less than or equal to the preset deviation threshold, the device is determined not to be in a sub-healthy state.

[0056] Specifically, the LSTM algorithm is used to process time-series data, learn the long-term variation patterns of central air conditioning equipment under normal operating conditions, make accurate predictions of future parameters, and identify sub-health conditions by the deviation between predicted and measured values. It can identify early hidden dangers such as scaling, wear, and micro-leakage 7-15 days in advance, with an early warning accuracy rate of ≥95%. Sub-health conditions include heat exchanger scaling and bearing wear.

[0057] Specific implementation steps:

[0058] 1) Construct a normal operating condition dataset by collecting historical time-series data when the equipment is fault-free and energy efficiency meets the standards, and use it as the model training set.

[0059] 2) Build an LSTM prediction model. Input: time series of water temperature, pressure, current, load, etc. at past time T; output: predicted parameter values ​​for the next time step (such as heat exchanger inlet and outlet temperature difference, compressor current, etc.).

[0060] 3) Real-time prediction and deviation calculation, .

[0061] 4) Sub-health status determination: When e > preset threshold, it is determined to be a sub-healthy state; when e ≤ preset threshold, it is determined to be a normal state. The preset threshold is dynamically calibrated according to GB / T17981—2025 "Economic Operation of Air Conditioning Systems", for example, the heat exchanger temperature difference threshold is ≤0.3℃.

[0062] S4. Energy Efficiency Anomaly Diagnosis: An energy efficiency anomaly diagnosis model is constructed using a fusion reinforcement learning Q-learning algorithm and gradient boosting decision tree. With COP≥9.22 as the target threshold, a standardized dataset is input, and the energy efficiency anomaly type is diagnosed and the degree of anomaly impact is calculated based on the preset energy efficiency anomaly dataset.

[0063] Furthermore, in step S4, after diagnosing the type of energy efficiency anomaly and calculating the degree of its impact, the energy efficiency anomaly diagnosis model is continuously updated and optimized using a Q-table.

[0064] A hybrid model employing gradient boosting decision trees for anomaly classification and reinforcement learning Q-learning for policy optimization balances diagnostic accuracy and adaptability. This hybrid AI model exhibits stronger generalization than a single algorithm and maintains high diagnostic accuracy even under complex operating conditions. The energy efficiency anomaly diagnostic model is continuously updated and optimized using a Q-table to improve diagnostic accuracy; reinforcement learning mechanisms are utilized to continuously optimize model parameters based on historical data and feedback results, enhancing the system's adaptability and diagnostic accuracy, thus ensuring the effectiveness of energy efficiency optimization.

[0065] Implementation of Gradient Boosting Decision Tree; Input: Standardized multi-dimensional data; Output: Anomaly type, such as load imbalance, unreasonable parameters, equipment defects, heat exchanger blockage, etc., and anomaly severity score. Objective: To optimize COP ≥ 9.22 and achieve a classification accuracy ≥ 95%. Implementation of Reinforcement Learning Q-learning; State: Current operating condition, parameters, COP, deviation; Action: Parameter tuning direction, amplitude, actuator action; Reward: Reward for COP improvement, penalty for energy consumption increase; Automatically update the Q-table after each diagnosis to improve model accuracy with use.

[0066] S5. Abnormal Handling; Based on the diagnostic results of steps S3 and S4, generate and automatically execute control strategies; including steps,

[0067] Tiered early warning; three levels of early warning thresholds are preset, and corresponding early warning measures are implemented according to the risk level of sub-health status and the severity of energy efficiency abnormalities;

[0068] The three-level early warning thresholds include attention level, early warning level, and alarm level;

[0069] When at the attention level, the parameters are automatically adjusted by the execution unit;

[0070] When the warning level is reached, a notification message is pushed to the maintenance personnel, and the control strategy is automatically executed through the execution unit. At the same time, the maintenance personnel go to the location to re-inspect the fault point.

[0071] When the alarm level is triggered, an audible and visual alarm is activated, and a maintenance work order is sent to the maintenance personnel for adjustment and maintenance.

[0072] S501. Generate regulation strategy; Based on the sub-health state identification results and energy efficiency anomaly diagnosis results, combine digital twin model to simulate the operation effect of different regulation parameter combinations, compare the operation effect of different regulation parameter combinations, and generate the optimal regulation strategy.

[0073] In step S501, the control parameters include chilled water flow rate adjustment, fan speed adjustment, compressor operating frequency adjustment, dynamic adaptation of fresh air introduction ratio, and frequency conversion adjustment of water pump system.

[0074] S502. Automatically execute the control strategy; send the generated optimal control strategy to the central air conditioning control system, drive the execution unit to automatically adjust the operating parameters to the control parameters corresponding to the optimal control strategy.

[0075] In step S502, the execution unit includes an electric regulating valve and a frequency converter. The electric regulating valve is used to adjust the chilled water flow rate, chilled water flow rate, and fresh air introduction ratio. The frequency converter is used to adjust the compressor operating frequency. The water pump system adopts an online adaptive frequency conversion control strategy based on Q-learning.

[0076] The system presets three threshold levels: Attention Level, Warning Level, and Alarm Level. Attention Level: The system automatically adjusts parameters without manual intervention. Warning Level: A notification message is sent, automatic adjustment is implemented, and manual review is required. Alarm Level: Audible and visual alarms are triggered, maintenance work orders are sent, and emergency manual intervention is required. Digital twin model simulation and optimization are implemented; a 1:1 digital twin of the physical air conditioner is built in virtual space; multiple sets of control parameters are input to simulate operating effects, energy consumption, and COP; the scheme with the highest COP, lowest energy consumption, and best stability is selected as the optimal strategy. Control parameters and execution units: Control parameters include chilled water flow rate, fan speed, compressor frequency, fresh air ratio, and water pump frequency conversion; execution units include electric regulating valves (flow rate / fresh air) and frequency converters (compressor / fan); water pump control: Q-learning online adaptive frequency conversion is used for real-time optimization.

[0077] Differentiated handling based on risk level avoids indiscriminate alarms and operational redundancy, improving handling efficiency and response accuracy. Attention level: automatic parameter tuning without manual intervention reduces manpower input for maintenance and ensures continuous and stable system operation; Early warning level: automatic adjustment with manual verification balances automation efficiency and on-site safety redundancy, preventing misadjustments and omissions; Alarm level: mandatory audible and visual alarms and work order pushes quickly trigger manual handling, avoiding major risks such as equipment damage and production stoppages. Early warning rules are dynamically bound to the GB / T17981—2025 standard, ensuring compliance throughout optimization and maintenance, meeting energy conservation regulatory requirements. Virtual simulation is used for preliminary verification, avoiding energy waste, comfort fluctuations, and equipment impact caused by on-site trial and error; multi-parameter combination optimization across the entire chain achieves optimal synergy between cold source, distribution, and terminal, rather than local optimization of a single device, resulting in more significant overall energy savings; optimization strategy generation is fast, with an execution delay of ≤1 minute, allowing real-time adaptation to load and environmental changes, resulting in more timely energy efficiency improvements; the optimization process is visualized and traceable, facilitating maintenance review, management statistics, and quantitative display of energy-saving effects. It covers all adjustable parameters including chilled water flow rate, fan speed, compressor frequency, fresh air ratio, and water pump frequency conversion, adapting to complex operating conditions and different models. For the water pump, it employs Q-learning online adaptive frequency conversion to achieve continuous optimization under dynamic loads, further reducing energy consumption in transmission and distribution. The electric regulating valve and frequency converter have clearly defined functions and rapid response, ensuring accurate implementation of optimization commands without execution deviation. The actuator and AI strategy are deeply integrated, forming a closed loop of "diagnosis-decision-execution-feedback," enhancing the system's autonomy. It is compatible with existing equipment execution units, requiring no large-scale hardware modifications, reducing modification costs and shortening deployment cycles.

[0078] S6. Validation of the regulation strategy: The multi-dimensional parameter data collected in S1, the sub-health identification results in S3, the energy efficiency anomaly diagnosis results in S4, the optimal regulation strategy in S5, and the energy efficiency changes after regulation are stored in the cloud database. The operation effect of the regulation strategy is validated by combining the digital twin model.

[0079] The multi-dimensional parameter data collected in step S1, the sub-health identification results in step S3, the energy efficiency anomaly diagnosis results in step S4, the optimal control strategy generated in step S5, and the real-time energy efficiency change data after control are all stored in a cloud database to form a full-link traceable data archive. At the same time, the digital twin model is called to compare the actual operation data with the expected data of the twin simulation item by item to complete the verification of the control effect.

[0080] Data is stored in the cloud, including timestamps, device IDs, collected parameters, sub-health assessment results, anomaly types, anomaly levels, control commands, execution status, COP before / after control, energy consumption, temperature, pressure, flow rate, and other comprehensive data. The storage format is a structured time-series database, supporting rapid retrieval by device, time, work order, and anomaly type. The measured data from the cloud is written back to the digital twin model, and deviations are calculated between the measured data and simulation predictions. Verification indicators include: COP improvement rate, energy consumption reduction rate, parameter stability, indoor comfort, and equipment operational safety margin. Verification conclusions are output: if the measured effect reaches or exceeds the simulation effect, the control strategy is deemed effective, and the warning is automatically lifted; if the measured effect fails to meet the standard, the control strategy needs correction, and the system automatically triggers secondary diagnosis and re-optimization. From data collection, diagnosis, decision-making to execution and results, everything is traceable, meeting the needs of operation and maintenance auditing, energy-saving accounting, and fault tracing. A closed-loop comparison between virtual and real data is used, with digital twins used to verify the effect, avoiding "optimization without verification" and ensuring that every strategy is truly effective. Energy efficiency is quantifiable, automatically generating before-and-after comparison reports, making energy-saving effects intuitive and measurable, facilitating management and cost reduction reporting. Safety is guaranteed, with timely detection of execution deviations through verification, preventing energy efficiency decline or equipment risks caused by parameter drift or inadequate control.

[0081] S7. Iteratively optimize parameters and control strategies; repeat steps S1-S6, and continuously iteratively optimize historical parameter datasets, energy efficiency anomaly datasets, control parameter combinations, and control strategies by analyzing historical data.

[0082] Specifically, historical parameter dataset optimization involves continuously expanding samples of normal, sub-optimal, and faulty operating conditions, and updating the training set and judgment thresholds of the LSTM prediction model. Energy efficiency anomaly dataset optimization includes adding anomaly types, anomaly features, and fault causes to improve the anomaly database and enhance the diagnostic accuracy of reinforcement learning Q-learning algorithms and gradient boosting decision trees. Regulation parameter combination optimization utilizes historically successful strategies, continuously updating the reinforcement learning Q-table to create a better parameter combination library suitable for different seasons, loads, weather conditions, and electricity prices. Regulation strategy optimization solidifies and generalizes strategies with high success rates and high energy savings; inefficient strategies are marked and eliminated to achieve strategy self-evolution.

[0083] Furthermore, during the iteration and optimization process in step S7, the system supports remote configuration of control logic and upgrading of edge gateway firmware based on BLRAT technology. BLRAT remotely modifies control logic, early warning thresholds, optimization parameters, and AI model weights through an encrypted communication link; it can remotely complete firmware upgrades, algorithm model updates, and function iterations without the need for engineers to be present; the upgrade process supports breakpoint resume and rollback protection, without affecting the normal operation of the device.

[0084] By continuously learning from historical data, the accuracy of sub-health identification, abnormal diagnosis, and strategy generation is constantly improved, and the generalization ability is stronger. It has evolved from a "global general strategy" to a "scenario-specific optimal strategy", and can output the best solution under different loads, seasons, and weather. BLRAT remote configuration and remote upgrade significantly reduce on-site operation and maintenance costs, and cross-regional projects can be managed in a unified manner. Algorithms, strategies, and firmware can be continuously iterated to adapt to equipment aging, system transformation, and new standards and regulations. Upgrades can be performed without downtime, without on-site debugging, and without repeated modifications, resulting in lower overall investment and a longer benefit cycle.

[0085] This invention also provides a central air conditioning system operation and maintenance system based on artificial intelligence diagnosis and control, used to implement the aforementioned operation and maintenance method for central air conditioning systems based on artificial intelligence diagnosis and control. The system adopts a cloud-edge collaborative and modular full-link closed-loop architecture design, which can automatically complete the entire process of data acquisition, data fusion, AI diagnosis, intelligent optimization, effect verification, and continuous iterative optimization. The overall structure is clear, the functions are complete, and the adaptability is strong. Specifically, the system includes a multi-dimensional data acquisition module, a data fusion processing module, a sub-health state identification module, an energy efficiency anomaly diagnosis module, an anomaly handling module, a control strategy verification module, and an optimization iteration module. The specific configuration and operating logic of each module are as follows:

[0086] The multi-dimensional data acquisition module is configured to execute step S1. This module consists of a distributed high-precision sensor array, an edge computing gateway, a built-in programmable logic controller (PLC), and signal acquisition circuits. The sensors selected include a PT1000 temperature sensor with integrated cold junction compensation, a diffused silicon pressure sensor, a three-phase intelligent power acquisition module, and environmental sensors for temperature, humidity, CO2, and solar radiation intensity. All analog signal acquisition circuits adopt a differential input design to suppress electromagnetic interference on site. The edge computing gateway has multiple acquisition interfaces, data caching, and local real-time computing capabilities. It deploys acquisition points according to four dimensions: equipment operating parameters, environmental parameters, load parameters, and energy efficiency parameters, and completes data acquisition at a preset frequency. The PLC built into the gateway can automatically perform cleaning, noise reduction, interpolation, format standardization, unit unification, and timestamp alignment operations on the acquired data. It is also configured with network outage self-governance logic to maintain basic system temperature control and safety protection functions in the event of a network interruption, ensuring that the acquired data is stable, accurate, and reliable.

[0087] The data fusion processing module is configured to execute step S2. This module is deployed on a cloud server and runs on multi-core computing resources, a time-series database, and an algorithm container. The module has a built-in improved norm distance algorithm and a dynamic weight allocation mechanism, which can assign higher weights to key feature parameters such as water temperature, water pressure, load, and COP. It can automatically complete time alignment, dimension normalization, outlier removal, and multi-device association processing of multi-source heterogeneous data. At the same time, the module integrates multiple protocol communication stacks such as BAC net IP, MQTT, and Modbus TCP, which can establish stable two-way communication with central air conditioning field controllers, building automation systems, and park energy management platforms. Finally, it outputs a standardized dataset with unified time series, standard format, and complete features, providing high-quality input for subsequent AI models.

[0088] The sub-health state identification module is configured to execute step S3. This module deploys an LSTM long short-term memory network model in the cloud, completes the model building based on the preset network structure, time step size and activation function, and completes supervised training by loading historical time series data of normal equipment operation, so that the model can fully learn the changing patterns of various parameters under normal working conditions. When the module is running, it can receive standardized datasets in real time, output the predicted value of parameters at the next moment and calculate the deviation between the measured value and the predicted value. The module is configured with a dynamic deviation threshold according to the GB / T17981—2025 standard. When the calculated deviation exceeds the preset threshold, the equipment is determined to be in a sub-health state. When the deviation is within the threshold range, the equipment is determined to be operating normally, thereby realizing the accurate identification of early hidden dangers of the equipment.

[0089] The energy efficiency anomaly diagnosis module is configured to execute step S4. This module adopts a hybrid AI model that integrates gradient boosting decision trees and reinforcement learning Q-learning. The gradient boosting decision tree takes a standardized dataset as input and completes the determination of the energy efficiency anomaly type, anomaly degree, and influence weight calculation through multi-tree joint decision-making. The Q-learning algorithm is configured with a complete state space, action space, reward function, learning rate, and discount factor. It performs energy efficiency diagnosis and decision optimization with a target threshold of COP≥9.22. The module allocates a dedicated storage area for Q-table data storage. After each diagnosis, the Q-table is automatically updated according to preset rules to continuously optimize the judgment logic and decision-making ability of the diagnosis model and improve the diagnosis accuracy and generalization under complex working conditions.

[0090] The anomaly handling module is configured to execute step S5. This module includes a graded early warning unit, a strategy generation unit, and an instruction execution unit. The graded early warning unit presets three levels of early warning thresholds: attention level, early warning level, and alarm level. It automatically matches the handling methods such as prompts, alarms, and work order pushes according to the sub-health risk level and the severity of energy efficiency anomalies, and is configured with linkage interfaces with terminals and audible and visual alarms. The strategy generation unit is equipped with a 1:1 digital twin model of the central air conditioning system, which can load multiple sets of control parameter combinations for simulation operation. It comprehensively compares indicators such as COP, energy consumption, stability, and comfort, and outputs the optimal control strategy including chilled water flow, fan speed, compressor operating frequency, fresh air introduction ratio, and water pump frequency conversion adjustment. The instruction execution unit interfaces with actuators such as field electric regulating valves and frequency converters, and issues adjustment instructions through the control interface. The water pump system adopts online adaptive frequency conversion control logic based on Q-learning, and the instruction issuance response delay does not exceed 1 minute, ensuring fast and accurate execution of adjustments.

[0091] The regulation strategy verification module is configured to execute step S6. This module establishes data linkage with the cloud time series database and digital twin model. It is equipped with an automatic storage and long-term archiving mechanism, which can uniformly store multi-dimensional collected data, sub-health identification results, energy efficiency anomaly diagnosis results, optimal regulation strategies, and energy efficiency change data before and after regulation in the cloud and classify them for management. The module writes back the actual measured operation data to the digital twin model, automatically completes the comparison between the simulation expected value and the measured value, deviation calculation and effect scoring. Based on core indicators such as COP change rate, energy consumption reduction rate, parameter stability, temperature compliance rate, and equipment safety margin, it outputs the conclusion that the regulation strategy is effective or needs to be corrected, forming a traceable, quantifiable and closed-loop verification mechanism.

[0092] The optimization iteration module is configured to execute step S7. This module is equipped with a fully automatic cyclic execution mechanism, which can repeatedly call all the aforementioned modules according to a cycle or trigger conditions. By continuously accumulating historical operating data, it continuously expands and optimizes the historical parameter dataset and energy efficiency anomaly dataset, synchronously iterates the combination of control parameters and optimization strategies, and continuously strengthens the model's decision-making capabilities. The module integrates BLRAT remote bus technology, is configured with encrypted communication links, breakpoint resume and secure rollback mechanisms, and supports remote modification of control logic, adjustment of early warning thresholds, updating of algorithm models, and upgrading of edge gateway firmware. System function upgrades and parameter optimizations can be completed without on-site operation, ensuring long-term stable operation and continuous evolution of the system.

[0093] Example 1: Intelligent optimization of energy efficiency under high temperature and high load conditions in summer

[0094] Under high-temperature and high-load conditions in summer, the cooling demand of the park's central air conditioning system surges, leading to problems such as compressor load imbalance, insufficient fan matching, inadequate fresh air intake, and low energy efficiency. The system collects equipment operating parameters, environmental parameters, load parameters, and energy efficiency parameters through a distributed high-precision sensor array. The data undergoes preprocessing preprocessing, including cleaning, noise reduction, and format standardization, via an edge computing gateway. In the cloud, an improved norm distance algorithm is used to perform time alignment, dimensional normalization, and correlation fusion of multi-source heterogeneous data, generating a standardized dataset. An energy efficiency anomaly diagnosis model, employing a fusion of reinforcement learning Q-learning algorithm and gradient boosting decision tree, is used. With a target threshold of COP ≥ 9.22, the model diagnoses the anomaly as excessive load causing compressor load imbalance, fan speed not matching cooling demand, and insufficient fresh air intake, classifying the risk level as warning. The system pushes alerts to maintenance personnel and simultaneously simulates multiple sets of control parameter combinations based on a digital twin model to generate the optimal control strategy: increasing fan speed, dynamically adjusting compressor operating frequency, increasing the proportion of fresh air intake, and optimizing chilled water flow and pump frequency conversion output. The strategy is sent to the central air conditioning control system, where it is quickly executed by actuators such as frequency converters and electric regulating valves. The water pumps utilize online adaptive frequency conversion control based on Q-learning. Within 30 minutes of optimization, the chiller unit outlet temperature dropped to 9℃, the COP increased to 9.5, the energy consumption per unit cooling capacity decreased by 22%, and the indoor CO2 concentration dropped below 600ppm. The system verified that the control effect met the standards, automatically lifted the warning, and generated an energy efficiency optimization report, which was then pushed to the operation and maintenance platform.

[0095] Example 2: Early Warning and Preventive Maintenance of Scaling and Sub-health Status in Water Chiller Heat Exchangers

[0096] During routine system operation, an LSTM time-series prediction model is used to predict the inlet and outlet temperature difference of the chiller heat exchanger in real time. The measured and predicted values ​​are then compared to calculate the deviation and determine the system's status. Monitoring revealed that the inlet and outlet temperature difference deviation of the heat exchanger consistently exceeded 0.8℃, surpassing the ≤0.3℃ threshold specified in GB / T17981—2025 "Economic Operation of Air Conditioning Systems". Other operating parameters remained within safe limits. Combined with vibration sensor data analysis, the system determined the equipment was in a sub-healthy state due to heat exchanger scaling. This state did not trigger automatic optimization, so the system generated a preventative maintenance work order, prompting the heat exchanger to be cleaned and maintained within 10 days, and pushed this to the maintenance personnel's terminal. After the maintenance personnel completed the cleaning according to the work order, the inlet and outlet temperature difference deviation of the heat exchanger returned to 0.2℃, returning to the normal range. This intervention prevented compressor overload, energy efficiency degradation, and unplanned shutdowns caused by increased scaling, reducing subsequent maintenance costs by approximately 30%. The system stores the characteristics, treatment process, and effects of this sub-health condition in a cloud database, and iteratively optimizes the historical parameter dataset, abnormal feature library, and LSTM prediction model to improve the accuracy of sub-health identification.

Claims

1. An operation and maintenance method of a central air conditioning system based on artificial intelligence diagnosis and regulation, characterized in that, Including the following steps: S1. Multi-dimensional parameter dataset collection; acquire the device's operating parameters, environmental parameters, load parameters, and energy efficiency parameters to generate a multi-dimensional parameter dataset; S2. Data fusion processing: An improved norm distance algorithm is used to associate and fuse multi-source heterogeneous data in a multi-dimensional parameter dataset to generate a standardized dataset; S3. Sub-health status identification: The LSTM algorithm is used to build a time series prediction model. The model learns the parameter change pattern under normal operating conditions by using the historical parameter dataset of the equipment during normal operation, so as to generate a predicted value dataset for normal operation. By analyzing the deviation between the multi-dimensional parameter dataset and the predicted value dataset, the sub-health status of the equipment is identified. S4. Energy efficiency anomaly diagnosis; An energy efficiency anomaly diagnostic model is constructed by combining reinforcement learning Q-learning algorithm and gradient boosting decision tree. With COP≥9.22 as the target threshold, a standardized dataset is input, and the energy efficiency anomaly type is diagnosed and the degree of anomaly impact is calculated based on the preset energy efficiency anomaly dataset. S5. Abnormal Handling; Based on the diagnostic results of steps S3 and S4, generate and automatically execute control strategies; including steps, S501. Generate control strategy; Based on the sub-health state identification results and energy efficiency anomaly diagnosis results, the operation effect of different control parameter combinations is simulated by combining digital twin models. The operation effect of different control parameter combinations is compared and the optimal control strategy is generated. S502. Automatically execute the control strategy; send the generated optimal control strategy to the central air conditioning control system, drive the execution unit to automatically adjust the operating parameters to the control parameters corresponding to the optimal control strategy; S6. Validation of the regulation strategy: The multi-dimensional parameter data collected in S1, the sub-health identification results in S3, the energy efficiency anomaly diagnosis results in S4, the optimal regulation strategy in S5, and the energy efficiency changes after regulation are stored in the cloud database. The operation effect of the regulation strategy is validated by combining the digital twin model. S7. Iterative optimization of parameters and control strategies; Repeat steps S1-S6 to continuously iterate and optimize the historical parameter dataset, energy efficiency anomaly dataset, control parameter combination, and control strategy by analyzing historical data.

2. The operation and maintenance method of the central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, characterized in that: In step S1, the programmable logic controller built into the edge computing gateway is used to clean, reduce noise, standardize the format of the collected multi-dimensional parameter data, and perform local basic control to maintain basic temperature control function in the offline state; wherein, the preprocessing is the preprocessing before the data fusion processing in step S2.

3. The operation and maintenance method of the central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, characterized in that: In the data fusion processing step S2, the existing central air conditioning control system is connected via BACnet IP and MQTT protocols.

4. The operation and maintenance method of the central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, characterized in that: In step S3, when the deviation is greater than a preset deviation threshold, the device is determined to be in a sub-healthy state; if the deviation is less than or equal to the preset deviation threshold, the device is determined not to be in a sub-healthy state.

5. The operation and maintenance method of the central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, characterized in that: In step S4, after diagnosing the type of energy efficiency anomaly and calculating the degree of its impact, the energy efficiency anomaly diagnosis model is continuously updated and optimized using a Q-table.

6. The operation and maintenance method of the central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, characterized in that: The step S501 is preceded by a graded early warning system; a three-level early warning threshold is preset, and corresponding early warning measures are implemented according to the risk level of the sub-health state and the severity of the energy efficiency abnormality. The three-level early warning thresholds include attention level, early warning level, and alarm level; When at the attention level, the parameters are automatically adjusted by the execution unit; When the warning level is reached, a notification message is pushed to the maintenance personnel, and the control strategy is automatically executed through the execution unit. At the same time, the maintenance personnel go to the location to re-inspect the fault point. When the alarm level is triggered, an audible and visual alarm is activated, and a maintenance work order is sent to the maintenance personnel for adjustment and maintenance.

7. The operation and maintenance method of the central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, characterized in that: In step S501, the control parameters include chilled water flow rate adjustment, fan speed adjustment, compressor operating frequency adjustment, dynamic adaptation of fresh air introduction ratio, and frequency conversion adjustment of water pump system. 8.The operation and maintenance method of a central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, wherein: In step S502, the execution unit includes an electric regulating valve and a frequency converter. The electric regulating valve is used to adjust the chilled water flow rate, chilled water flow rate, and fresh air introduction ratio. The frequency converter is used to adjust the compressor operating frequency. The water pump system adopts an online adaptive frequency conversion control strategy based on Q-learning. 9.The operation and maintenance method of a central air conditioning system based on artificial intelligence diagnosis and regulation according to claim 1, wherein: In step S7, during the iterative optimization of historical parameter datasets, energy efficiency anomaly datasets, control parameter combinations, and control strategies, remote configuration of control logic and firmware upgrades based on BLRAT technology are supported.

10. An operation and maintenance system of a central air conditioning system based on artificial intelligence diagnosis and regulation, configured to implement the operation and maintenance method of the central air conditioning system based on artificial intelligence diagnosis and regulation according to any one of claims 1-9. include, The multi-dimensional data acquisition module is configured to execute step S1; The data fusion processing module is configured to execute step S2; The sub-health status identification module is configured to execute step S3; The energy efficiency anomaly diagnosis module is configured to execute step S4; The exception handling module is configured to execute step S5; The control strategy verification module is configured to execute step S6; The optimization iteration module is configured in execution step S7.

Images