Intelligent self-adaptive scheduling method for multi-protocol industrial fan group control

Abstract

The invention discloses an intelligent self-adaptive scheduling method for multi-protocol industrial fan group control. The method comprises the following steps: S1, time-space self-consistency real-time sensing and semantic-level data generation of a multi-protocol heterogeneous fan group; s2, performing multi-protocol analysis based on dynamic semantic mapping; s3, realizing accurate integration of data of a plurality of sensors by using Kalman filtering and a data fusion technology; s4, an operation strategy is adjusted in real time according to the equipment operation state, the environment change and the fan load, and dynamic optimization of fan group control is achieved; and S5, performing fault prediction and setting an early warning mechanism. By introducing technologies such as multi-protocol data analysis and mapping, an adaptive scheduling algorithm, intelligent energy efficiency optimization control, machine learning fault prediction and the like, a fan group control management platform with high compatibility, intelligence, energy conservation and reliability is constructed, the system adapts to a complex and dynamic operation environment of an industrial field, and the system is suitable for large-scale popularization and application. And the overall operation efficiency and energy efficiency of the fan group control system are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial Internet of Things and intelligent control technology, and in particular relates to an intelligent adaptive scheduling method for multi-protocol industrial fan group control. Background Technology

[0002] With the deepening of industrialization, wind turbine systems, as indispensable equipment in industrial production, building air conditioning, data centers, and other scenarios, have become a crucial factor determining energy efficiency. Traditional wind turbine group control systems typically employ fixed control strategies, such as frequency conversion regulation based on timing or load. These methods struggle to achieve efficient and flexible energy efficiency optimization under complex operating conditions, leading to significant energy waste and high operating costs.

[0003] Furthermore, current industrial field equipment typically comes from different manufacturers and uses different protocols (such as BACnet, M-Bus, EtherCAT, etc.). Insufficient cross-protocol interoperability between these devices leads to system integration complexity and increases commissioning and maintenance costs. Multi-protocol data exchange, status monitoring, and collaborative control of wind turbines require a highly efficient system capable of automatically adapting to different protocols and operating environments to optimize their operating strategies, reduce energy consumption, and improve operational efficiency. Summary of the Invention

[0004] The purpose of this invention is to propose an intelligent adaptive scheduling method for multi-protocol industrial fan group control in order to solve the problems of poor protocol compatibility, insufficient energy efficiency optimization, and lagging real-time scheduling in current industrial fan group control systems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent adaptive scheduling method for multi-protocol industrial fan group control specifically includes the following steps: S1. Spatiotemporal self-consistency real-time perception and semantic-level data generation of multi-protocol heterogeneous wind turbine groups; S2. Multi-protocol parsing based on dynamic semantic mapping; S3. Use Kalman filtering and data fusion technology to achieve accurate integration of data from multiple sensors; S4. Adjust the operating strategy in real time according to the equipment operating status, environmental changes and fan load to achieve dynamic optimization of fan group control; S5. Perform fault prediction and set up an early warning mechanism.

[0006] As a further description of the above technical solution: In step S1, the system collects wind turbine operating parameters, including multi-dimensional data such as current, voltage, speed, power, temperature, air volume, and pressure, through a real-time sensing mechanism for a multi-protocol heterogeneous wind turbine group. The system first uses multi-protocol adaptive parsing technology to uniformly parse device data from different protocols such as BACnet, Modbus, EtherCAT, and M-Bus. Then, it converts the raw data of different formats into a unified device semantic model through protocol semantic mapping rules to ensure data consistency among heterogeneous devices. At the same time, it uses a time-series synchronization mechanism and Kalman filtering noise suppression method to perform spatiotemporal alignment and dynamic fusion of multi-source sensor data to improve data stability and predictability. The high-quality feature data after the above processing is transmitted to the scheduling decision module in real time. The scheduling algorithm evaluates the operating status and identifies the demand based on standardized multi-dimensional environmental information and wind turbine load status, providing accurate, timely, and consistent input data for subsequent group control optimization.

[0007] As a further description of the above technical solution: In S2, protocol identification and parsing: For different industrial protocols, such as BACnet, M-Bus, EtherCAT, etc., the protocol type is first automatically identified, and a unified protocol parsing engine is used to classify the devices. During the protocol parsing process, information such as data fields, data types and function codes are abstracted into a unified semantic model to form a general data processing framework. Semantic mapping and dynamic conversion: By establishing a protocol semantic mapping library, the parsing engine dynamically maps data according to the semantic information of the protocol used by the device and performs corresponding format conversion. The system can automatically adjust the data format when the device is connected, avoiding data misreading or loss due to protocol differences. Cross-device data fusion: For different connected devices, the system can fuse the data of each device in real time according to the protocol mapping library and device feature library, form a unified data stream, and transmit it to the group control and scheduling system to realize cross-protocol collaborative control between wind turbines.

[0008] As a further description of the above technical solution: In S3, Kalman filtering is a recursive optimal estimation algorithm that is widely used in state estimation problems in dynamic systems. It is particularly suitable for multi-source signal fusion and noise suppression. Kalman filtering technology can achieve the following functions: noise suppression and data cleaning, spatiotemporal data fusion, and improved decision accuracy and response speed.

[0009] As a further description of the above technical solution: The noise suppression and data cleaning process includes: Sensors involved in the wind turbine group control system, such as temperature, humidity, wind speed, current, and voltage, are subject to varying degrees of noise interference. This noise may affect the accuracy of the data and lead to instability in the wind turbine control strategy. The Kalman filter uses a weighted average of the predicted value of the system state and the sensor measurement value to suppress sensor measurement errors and noise, and outputs an optimal estimate. The recursive nature of the Kalman filter allows it to process each new data input in real time, effectively remove outliers, and improve data accuracy.

[0010] As a further description of the above technical solution: The spatiotemporal data fusion includes: In a wind turbine group control system, the temporal and spatial synchronization of data from multiple sensors, such as temperature and humidity sensors, wind speed sensors, and power sensors, is crucial. Kalman filtering is used to fuse the temporal data of multiple sensors to ensure their spatiotemporal consistency. During wind turbine operation, the effects of changes in ambient temperature and wind speed need to be considered simultaneously. The Kalman filter can update the prediction model in real time and combine it with actual measurement data to make optimal estimates, obtaining a comprehensive value of multiple sensor data, thus avoiding the influence of errors from a single sensor data on control decisions.

[0011] As a further description of the above technical solution: The improvement in decision-making accuracy and response speed includes: through the real-time data update and prediction function of Kalman filtering, the system can accurately estimate the current state of the wind turbine group control system, such as load and operating efficiency, providing high-quality input data for the intelligent scheduling algorithm, significantly improving the accuracy and response speed of control decisions, and ensuring that the system is always in the optimal operating state under changing working conditions; The prediction steps and core formula of Kalman filtering are expressed as follows: S101, State Prediction:

[0012] in, It is a predicted system state. It is the state transition matrix. It is a control input matrix. It is a control input (such as adjusting the operation of the fan); S102, Predicting Covariance:

[0013] in: It is the predicted covariance matrix. It is the process noise covariance; S103, Kalman gain calculation:

[0014] in: It is Kalman gain. It is the observation matrix. It measures the noise covariance; S104, Status Update:

[0015] in: This is the updated status. These are actual measured values; S105, Covariance Update:

[0016] in: It is the updated covariance matrix; By fusing data from different sensors, such as temperature, wind speed, and current, and removing noise, accurate real-time state estimates are obtained, providing precise input for subsequent scheduling decisions.

[0017] As a further description of the above technical solution: In S5, fault prediction includes a machine learning fault prediction model, which uses convolutional neural networks (CNN) and support vector machines (SVM) to extract and predict fault features. By training a large amount of wind turbine operation data, it can effectively identify abnormal patterns of the equipment and issue fault warning signals in advance.

[0018] As a further description of the above technical solution: In S5, the early warning mechanism includes fault warning and automatic response. When it is predicted that a fault is about to occur in the equipment, the system will issue an alarm in a timely manner and automatically adjust the operating strategy according to the fault type, such as activating the backup fan and adjusting the workload, to reduce the impact of the fault on the overall system operation.

[0019] An application system for an intelligent adaptive scheduling method for multi-protocol industrial fan group control is characterized by comprising five core layers: a multi-protocol unified access layer, a multi-source feature fusion layer, a state assessment and prediction layer, a digital twin modeling layer, an AI scheduling and optimization decision layer, and an edge execution and system feedback layer. Unified access layer for multiple protocols: It realizes unified parsing and model abstraction of BACnet, M-Bus, EtherCAT, MODBUS, private protocols, etc., and adopts protocol DSL, automatic parser generation (ParserGen), buffer queue mechanism, etc. Multi-source feature fusion layer: Integrates multi-source sensor data (temperature, humidity, pressure, wind speed, current, voltage, etc.) from the wind turbine equipment, and obtains a high-dimensional feature vector of wind turbine operation through Kalman filtering, state estimation and feature extraction, including algorithms such as EKF, UKF, Attention-Based Fusion and noise suppression; State assessment and prediction layer: performs health assessment (HI), load prediction, power consumption prediction, etc. on wind turbines, including LSTM, CNN, Transformer, and physical constraint-based models; Digital twin modeling layer: Constructs a simulation model that couples the fan, duct, and environment to achieve real-time mapping, inversion inference, and virtual-real comparison, and coordinates with aerodynamic models, thermal management models, system simulation models, etc. AI scheduling and optimization decision layer: Based on reinforcement learning, game theory mechanism, multi-agent collaboration, formulate the global optimal scheduling strategy, DQN, DDPG, PPO, MADDPG, multi-objective optimization, and multi-constraint solution; Edge execution and system feedback layer: executes scheduling strategies, monitors execution results, handles exceptions and rollbacks, performs feedback learning, integrates with edge controllers, provides millisecond-level response, and enables autonomous control during network outages.

[0020] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention introduces technologies such as multi-protocol data parsing and mapping, adaptive scheduling algorithms, intelligent energy efficiency optimization control, and machine learning fault prediction to construct a highly compatible, intelligent, energy-saving, and reliable wind turbine group control management platform. It adapts to the complex and dynamic operating environment of industrial sites and effectively improves the overall operating efficiency and energy efficiency of the wind turbine group control system.

[0021] 2. This invention proposes an adaptive scheduling algorithm that can adjust the operating strategy in real time according to the equipment operating status, environmental changes and wind turbine load, realize the dynamic optimization of wind turbine group control, and, combined with multi-protocol data parsing and compatibility technology, enable seamless access to devices with different protocols, reduce the difficulty of equipment integration, and improve the interoperability and stability of the system. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the system architecture of an intelligent adaptive scheduling method for multi-protocol industrial fan group control proposed in this invention; Figure 2 This is a flowchart illustrating the Kalman filtering and data fusion process of an intelligent adaptive scheduling method for multi-protocol industrial fan group control proposed in this invention. Figure 3 This is an overall flowchart of the entire system workflow of the intelligent adaptive scheduling method for multi-protocol industrial fan group control proposed in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1-2 This invention provides a technical solution: an intelligent adaptive scheduling method for multi-protocol industrial fan group control, characterized by the following steps: S1. The system collects wind turbine operating parameters, including multi-dimensional data such as current, voltage, speed, power, temperature, air volume, and pressure, through a real-time sensing mechanism for a multi-protocol heterogeneous wind turbine group. First, the system utilizes multi-protocol adaptive parsing technology to uniformly parse device data from different protocols such as BACnet, Modbus, EtherCAT, and M-Bus. Then, it converts the raw data of different formats into a unified device semantic model through protocol semantic mapping rules, ensuring data consistency among heterogeneous devices. Simultaneously, a time-series synchronization mechanism and Kalman filtering noise suppression method are used to perform spatiotemporal alignment and dynamic fusion of multi-source sensor data, improving data stability and predictability. The high-quality feature data processed above is transmitted in real-time to the scheduling decision module. The scheduling algorithm evaluates the operating status and identifies demands based on standardized multi-dimensional environmental information and wind turbine load status, providing accurate, timely, and consistent input data for subsequent group control optimization. S2. Multi-protocol parsing based on dynamic semantic mapping; Kalman filtering is a recursive optimal estimation algorithm that is widely used in state estimation problems in dynamic systems, especially suitable for multi-source signal fusion and noise suppression. The following functions can be achieved through Kalman filtering: noise suppression and data cleaning, spatiotemporal data fusion, and improved decision accuracy and response speed. S3. Use Kalman filtering and data fusion technology to achieve accurate integration of data from multiple sensors; S4. Adjust the operating strategy in real time according to the equipment operating status, environmental changes and fan load to achieve dynamic optimization of fan group control; S5. Perform fault prediction and set up an early warning mechanism.

[0025] Protocol identification and parsing: For different industrial protocols, such as BACnet, M-Bus, EtherCAT, etc., the protocol type is first automatically identified. A unified protocol parsing engine is used to classify the devices. During the protocol parsing process, information such as data fields, data types and function codes are abstracted into a unified semantic model to form a general data processing framework. Semantic mapping and dynamic conversion: By establishing a protocol semantic mapping library, the parsing engine dynamically maps data according to the semantic information of the protocol used by the device and performs corresponding format conversion. The system can automatically adjust the data format when the device is connected, avoiding data misreading or loss due to protocol differences. Cross-device data fusion: For different connected devices, the system can fuse the data of each device in real time according to the protocol mapping library and device feature library, form a unified data stream, and transmit it to the group control and scheduling system to realize cross-protocol collaborative control between wind turbines.

[0026] The noise suppression and data cleaning include: Sensors involved in the wind turbine group control system, such as temperature, humidity, wind speed, current, and voltage, are subject to varying degrees of noise interference. This noise may affect the accuracy of the data and lead to instability in the wind turbine control strategy. The Kalman filter uses a weighted average of the predicted value of the system state and the sensor measurement value to suppress sensor measurement errors and noise, and outputs an optimal estimate. The recursive nature of the Kalman filter allows it to process each new data input in real time, effectively removing outliers and improving data accuracy. The spatiotemporal data fusion includes: In the wind turbine group control system, the temporal and spatial synchronization of data from multiple sensors, such as temperature and humidity sensors, wind speed sensors, and power sensors, is crucial. The Kalman filter fuses the temporal data from multiple sensors to ensure their spatiotemporal consistency. During wind turbine operation, the effects of changes in ambient temperature and wind speed need to be considered simultaneously. The Kalman filter can update the prediction model in real time and combine it with actual measurement data to make an optimal estimate, obtaining a comprehensive value from multiple sensor data, avoiding the influence of errors from a single sensor data on control decisions. The improvement in decision-making accuracy and response speed includes: through the real-time data update and prediction function of Kalman filtering, the system can accurately estimate the current state of the wind turbine group control system, such as load and operating efficiency, providing high-quality input data for the intelligent scheduling algorithm, significantly improving the accuracy and response speed of control decisions, and ensuring that the system is always in the optimal operating state under changing working conditions.

[0027] Kalman filtering is used to fuse data from multiple sensors and suppress measurement noise. The basic principle of Kalman filtering is to optimally estimate the system state through a prediction and update process. The core formula for Kalman filtering is as follows: Prediction Steps State prediction:

[0028] in, It is a predicted system state. It is the state transition matrix. It is a control input matrix. It is a control input (such as the operation adjustment of the fan).

[0029] Predicting covariance:

[0030] in: It is the predicted covariance matrix. It is the process noise covariance.

[0031] Kalman gain calculation:

[0032] in: It is Kalman gain. It is the observation matrix. It measures noise covariance. Status Update:

[0033] in: This is the updated status. This is the actual measured value. Covariance update:

[0034] in: It is the updated covariance matrix By using Kalman filtering, the system can fuse data from different sensors (such as temperature, wind speed, current, etc.) and remove noise, thereby obtaining accurate real-time state estimates and providing precise input for subsequent scheduling decisions.

[0035] In S5, fault prediction includes a machine learning fault prediction model, which uses convolutional neural networks (CNN) and support vector machines (SVM) to extract and predict fault features. By training a large amount of wind turbine operation data, it can effectively identify abnormal patterns of the equipment and issue fault warning signals in advance.

[0036] In S5, the early warning mechanism includes fault warning and automatic response. When it is predicted that a fault is about to occur in the equipment, the system will issue an alarm in a timely manner and automatically adjust the operating strategy according to the fault type, such as activating the backup fan and adjusting the workload, to reduce the impact of the fault on the overall system operation.

[0037] An application system for an intelligent adaptive scheduling method for multi-protocol industrial fan group control includes five core layers: multi-protocol unified access layer, multi-source feature fusion layer, state assessment and prediction layer, digital twin modeling layer, AI scheduling and optimization decision layer, and edge execution and system feedback layer. The Unified Protocol Access Layer (UPAL) is a fundamental module of this invention. It enables unified parsing and model abstraction of BACnet, M-Bus, EtherCAT, MODBUS, and proprietary protocols, employing a protocol DSL, automatic parser generation (ParserGen), and a buffer queue mechanism. Its mission is to address long-standing issues in industrial fan group control scenarios, such as inconsistent device protocols, inconsistent data formats, unstable communication behavior, and difficulty in interconnecting devices from different vendors. UPAL's core objective is to build a unified data semantic expression, a unified access interface, a unified timing model, and a unified error handling system, providing stable and standardized data input for subsequent AI fusion, prediction, and scheduling. UPAL's design essentially follows the concept of "three abstractions and one generation": protocol abstraction, communication abstraction, semantic abstraction, and automatic parser generation. Therefore, this invention proposes: (1) Protocol Syntax Abstraction (PSA), in which all protocols are abstracted into a unified five-tuple: P=(H,A,D,C,E), where: H: Header, protocol header structure; A: Addressing, addressing information (device ID, object ID, register, etc.); D: Data Field, data content; C: Control Field, control instructions, such as read / write type, function code; E: Error Model, error semantic description. The significance of PSA is to eliminate protocol differences and enable the parser to be generated using templates. The concept of multi-protocol unified modeling (Protocol Abstraction Model) is used in the wind turbine control system in industrial fields, which involves a variety of mainstream protocols, including: BACnet / IP, BACnet MS / TP, M-Bus (Meter-Bus), EtherCAT (based on communication objects such as CoE and SoE), MODBUS RTU / TCP, and vendor proprietary protocols. These different protocols vary greatly in frame structure, semantic expression, data type, communication latency, and error model. Protocol Semantic Abstraction (PSeA) addresses the issue of different protocols providing varying semantic descriptions of wind turbine objects. Examples include BACnet (object / attribute model), M-Bus (data record), EtherCAT (PDO + SDO), and MODBUS (register model). This invention proposes a unified semantic model, UOM (Unified Object Model): Obj = {id, type, value, timestamp, quality}. All protocols output this unified format, facilitating direct use by subsequent AI applications. To reduce the high cost of manually writing parsers for each protocol, this invention designs a Protocol Domain-Specific Language (P-DSL). This P-DSL describes the protocol format using simple rules, and the system automatically generates a parser (ParserGen). The system automatically generates: a frame decoder, a verification module, and a field extraction and semantic mapping module. Advantages include: new protocols can be quickly added to the system; no significant code modifications are required; and the parser has a consistent error handling mechanism. High-performance protocol parser auto-generation (ParserGen) tool automatically generates: state machine parser (FSM-based Parser), stream parser (StreamParser), and zero-copy memory-mapped parser (ZeroCopyParser) through DSL. Its core algorithm is based on finite state machine (FSM): S_(t+1)=δ(S_t,input_t). FSM can handle complex protocol state transitions, such as BACnet segmented messages, M-Bus long messages, and EtherCATDC synchronization frames. Adaptive Quality of Communication (AQE) and dynamic handover are crucial in complex industrial field networks where different protocols exhibit varying communication quality. This invention proposes a communication quality assessment model:

[0038] Where: L: network latency, J: jitter, E: error rate, R: retransmission count, and Q value are used to determine: multi-protocol interface switching, multi-link load balancing, and device access priority adjustment; the system maintains a dynamic communication QoS table. This enables adaptive access across different devices and protocols. The time alignment model addresses the issue of different data reporting cycles across various protocols, such as BACnet COV (several seconds), EtherCAT (1 ms), M-Bus (2-10 seconds), and proprietary protocols (uncertain). To enable data from different time scales to participate in the AI ​​model simultaneously, this invention employs interpolation and resampling.

[0039] Methods employed: linear interpolation, cubic spline interpolation, and Kalman time alignment, ensuring that all data at time step t satisfy the following:

[0040] Ensure that the data is correctly input into the LSTM and Transformer models; The Unified Error Model (UEM) defines different error semantics for different protocols. Examples include: BACnet: Timeout, Offline, Unknown Property; M-Bus: Parity Error, Checksum Error; EtherCAT: Sync Error, PDO Mismatch; MODBUS: CRC Error, Illegal Address. Standardized error format:

[0041] The system can provide unified alarms for all protocol devices; Multi-source feature fusion layer: Integrates multi-source sensor data (temperature, humidity, pressure, wind speed, current, voltage, etc.) from the wind turbine equipment, and obtains a high-dimensional feature vector of wind turbine operation through Kalman filtering, state estimation and feature extraction, including algorithms such as EKF, UKF, Attention-Based Fusion and noise suppression; The Feature Fusion Layer (FDL) is a crucial link in this invention, which is responsible for converting data from different protocols, interfaces, sampling frequencies, units, and physical meanings into multi-dimensional feature vectors with consistent time scale, high confidence, and high correlation, for use in subsequent state assessment and intelligent scheduling. The core tasks of FDL include: calibration and denoising of multi-source sensor data, spatiotemporal synchronization of data, data fusion (Kalman filtering, Bayesian estimation, attention fusion), feature extraction (temporal features, frequency domain features, statistical features), construction of wind turbine equipment profiles, and generation of high-dimensional feature vectors for AI models. The entire FDL is a complex "data processing plant" responsible for transforming messy raw data into "learnable, predictable, and controllable" model inputs. Analysis of the characteristics and fusion requirements of multi-source data: Typical sensors involved in a wind turbine group control system include: temperature (ambient temperature, fan temperature), humidity, wind speed (fan outlet and inlet air), air pressure (static pressure, dynamic pressure), current, voltage, power, power factor, motor speed (RPM), air volume, environmental load (heat load, CO2 concentration, air quality, etc.), vibration, and noise (in some scenarios). These data have the following characteristics: Features illustrate Different sampling frequencies Current 1kHz, wind speed 1Hz, temperature 0.2Hz Inconsistent data quality Different sensors have different noise characteristics Time delay asynchrony LoRa data latency can reach several seconds. Different dimensions Wind speed (m / s), current (A), air volume (m³ / h) Different equipment models Centrifugal fans, axial fans, mixed-flow fans Data missing and numerous anomalies Packet loss, jitter, communication errors The primary task of FDL is to transform heterogeneous data into homogeneous features.

[0042] Data cleaning and denoising: Raw data often contains noise. This invention employs multiple filtering methods: (1) Sliding window filtering (Moving Average) Suitable for smoothing trends such as wind speed and current.

[0043] (2) Median Filter to remove abrupt transitions:

[0044] Suitable for high-noise sensors (such as pressure sensors).

[0045] (3) First-order Kalman filter (KF), assuming the system is:

[0046]

[0047] Kalman Update: predict:

[0048] renew:

[0049] Used for continuous physical quantities such as motor speed, current, and pressure; Spatio-temporal alignment requires a unified time scale due to the different sampling frequencies of different sensors. This invention proposes a "unified time grid model".

[0050] All data were resampled to a 1-second time interval.

[0051] The following methods were used: (1) Linear Interpolation

[0052] (2) Cubic Spline It provides smooth, continuous curves for wind speed and power.

[0053] (3) Kalman Time Alignment (KTA) For scenarios with variable LoRa / WiFi latency.

[0054] Make:

[0055] It becomes a unified input vector.

[0056] Multi-source data fusion (core contribution): Based on the processed multi-source sensor data, this invention proposes a three-level fusion method: Low-level fusion (sensor-level fusion, EKF / UKF), such as wind speed + wind pressure to estimate air volume:

[0057] Using EKF: Suitable for nonlinear scenarios (such as air volume estimation); Mid-level fusion (model-level fusion, Bayesian fusion), such as inferring thermal demand from temperature and humidity plus environmental load: Used for air heat load estimation; High-level fusion (semantic fusion, attention mechanism), attention weights:

[0058] Features after fusion:

[0059] This method can automatically learn the importance of different sensors under different operating conditions. For example, temperature has a greater weight when the temperature is high, current has a greater weight when the fan load changes suddenly, and pressure has a greater weight when the air duct is blocked. Compared with the traditional Kalman fusion, the multi-layer fusion of this invention can better adapt to dynamic operating conditions.

[0060] The core achievement of FDL is generating a "wind turbine profile":

[0061] Where: F_t: high-dimensional fusion feature vector, HI_t: health index, Load_t: load level inference, Pattern_t: typical operating condition characteristics (thermal type, high air volume type, etc.); Wind turbine profiling is used for: energy consumption analysis, environmental matching, and wind turbine formation / group scheduling. The accuracy of wind turbine profiling directly determines the effectiveness of the entire system's AI control. Feature vector construction (AI input), ultimately constructing a unified AI input vector:

[0062] Dimensions are generally in the following areas:

[0063] This is the key input for entering LSTM, Transformer, and RL.

[0064] State assessment and prediction layer: performs health assessment (HI), load prediction, power consumption prediction, etc. on wind turbines, including LSTM, CNN, Transformer, and physical constraint-based models; The Intelligent State & Evaluation Prediction Layer (ISEP) is the core intelligent component of this invention, designed to provide real-time understanding, accurate prediction, and highly reliable early warning of the operating status of wind turbine equipment. Its design goal is to achieve: Real-time assessment of the current operating status of the fan (temperature, air pressure, air speed, motor load, etc.) Short-term forecasts (seconds to minutes): air volume demand, fan load, power consumption Medium-term forecast (minutes to hours): operational risks and trend changes. Long-term prediction (days): Health (HI) ISEP integrates machine learning and deep learning, including models such as LSTM, GRU, CNN, Transformer, Autoencoder, and GAN, to achieve intelligent analysis of multi-dimensional time-series data of large wind turbine groups. The overall architecture of the status assessment system, ISEP, comprises five core modules: Health Assessment Module (HI Estimation) Load / Power Forecast Module Operating Condition Recognition Module Anomaly Detection Module The overall process is as follows: data fusion F_t -> temporal modeling (LSTM / GRU) -> CNN local feature extraction -> Transformer global attention association -> HI / Load / Power / Anomaly output. ISEP is the "intelligent understanding" part of the whole system, providing the necessary wind turbine state estimation and future prediction for reinforcement learning scheduling strategies.

[0065] The Health Index (HI) model is used to characterize whether a wind turbine is currently in a normal, sub-healthy, overloaded, or fault-prone state. The HI metric was constructed with the device's historical health baseline set at 1.0:

[0066] in: : The i-th feature (temperature, current, wind pressure, wind speed). Baseline health values Feature weights, learned through an attention mechanism. Normalization factor; When: HI ≥ 0.8: healthy, 0.5 ≤ HI<0.8: sub-healthy, HI<0.5: increased risk of abnormality.

[0067] Deep learning estimation of HI Using an LSTM + Attention network:

[0068] in This is the data from the most recent L seconds.

[0069] The network can automatically learn "which characteristics are more critical to health".

[0070] Load & Power Forecast Model The load changes of the fan are strongly correlated with changes in ambient temperature and humidity, duct resistance, and spatial heat load. ISEP uses an LSTM-based time series forecasting model:

[0071]

[0072] Used to predict load and power consumption in the next k steps.

[0073] Typical model structure: 2-layer LSTM + 1-layer fully connected; input dimension d=64~128; output: load and power consumption in the next 5s / 10s / 30s.

[0074] Prediction accuracy: Load prediction error ≤ 5%, power consumption prediction error ≤ 4%, providing forward-looking data for AI scheduling; The pattern recognition model identifies the current operating mode of the wind turbine, including: Low load / low airflow mode, high airflow mode, blockage mode (increased airflow resistance), thermal load impact mode, and early-stage "sub-healthy" mode. The architecture employs a CNN + Transformer structure: CNN extracts local variations (such as wind speed / current perturbations), while Transformer models long-distance dependencies (such as long-term temperature upward trends).

[0075] Anomaly Detection ISEP uses a multi-model joint detection method for anomalies: (1) Based on probability statistics

[0076] An anomaly is determined when |z| is greater than the threshold.

[0077] (2) Reconstruction error based on Autoencoder Autoencoder (AE) is trained on normal data and used to predict abnormal data:

[0078] when:

[0079] This method is used to identify abnormalities and is suitable for detecting subtle, early-stage faults.

[0080] (3) Isolation Forest Used for detecting rare events, such as sudden power consumption jumps.

[0081] ISEP output unified state vector ISEP outputs a unified state vector:

[0082] It is the most critical input for reinforcement learning scheduling strategies.

[0083] This vector summarizes the wind turbine's "past-present-future" state in a mathematical form.

[0084] Digital twin modeling layer: Constructs a simulation model that couples the fan, duct, and environment to achieve real-time mapping, inversion inference, and virtual-real comparison, and coordinates with aerodynamic models, thermal management models, system simulation models, etc. AI scheduling and optimization decision layer: Based on reinforcement learning, game theory mechanism, multi-agent collaboration, formulate the global optimal scheduling strategy, DQN, DDPG, PPO, MADDPG, multi-objective optimization, and multi-constraint solution; AI Decision & Optimization Layer (AIDO) The AI ​​Scheduling and Optimization Decision Layer (AIDO) is the decision-making "brain" of this invention, and its core objective is: While meeting ventilation and operating requirements, the goal is to achieve the lowest overall energy consumption, the most stable operation, the fastest response, and the longest lifespan for the entire fan group.

[0085] AIDO takes the "wind turbine profile" and "future prediction" provided by the State Assessment Layer (ISEP) as input, and uses a fusion of reinforcement learning, multi-agent collaborative optimization, and engineering control methods to output the optimal speed, optimal air volume, and optimal operating strategy for each wind turbine. Its innovation lies in the fact that it no longer treats each wind turbine as an independent entity, but rather allows the entire wind turbine group to operate as an intelligent collaborative system, enabling it to achieve "system optimization" rather than "single-device local optimization" as in traditional control methods.

[0086] The overall structure of AIDO consists of five core modules: Multi-objective Optimization Model, Deep Reinforcement Learning Policy Network, Multi-Agent RL, and Security Constraint and Interpretability Control Model. The data flow is as follows: State vector S_t (from ISEP) -> Reinforcement learning policy generation π(a|s) -> Multi-agent cooperative scheduling -> Feedback results written to the environment -> Reward update -> Policy self-learning optimization. AIDO forms a complete closed-loop self-learning system.

[0087] Multi-objective optimization formula (the mathematical foundation of intelligent system control) The core objectives of fan group control are: to meet ventilation volume requirements (environmental constraints), to meet temperature / pressure requirements (comfort and safety constraints), to minimize overall power consumption (energy consumption constraints), to reduce equipment wear (lifespan constraints), and to maintain steady-state operation (stability constraints). Its mathematical expression is:

[0088] satisfy:

[0089]

[0090]

[0091] in: The power consumption of the i-th wind turbine, Air volume contribution : Required air volume, Let be the rotational speed of the i-th wind turbine. This is a nonlinear, multi-constraint, multi-objective, and multi-variable optimization problem, which is difficult to solve using traditional PID and fuzzy control methods. Therefore, reinforcement learning is introduced.

[0092] Deep reinforcement learning model (core innovation) AIDO uses a reinforcement learning-based policy π(a|s) that maps input states (health, load forecast, operating conditions, demand) to output actions (new speed and fresh air volume for each fan). The RL models used are: PPO (Proximal Policy Optimization) – stable and reliable, suitable for continuous operation (speed); DDPG – continuous control, suitable for precise speed regulation; MADDPG (Multi-Agent RL) – for multi-wind turbine collaboration; and V-MARL (Value-based Multi-agent RL) – for sparse interaction in clusters of thousands of wind turbines. The decision-making dimension for wind turbines is continuous; therefore, an Actor-Critic structure is chosen.

[0093]

[0094] State space and action space (advanced definition), state space S; high-dimensional states provided by SEP:

[0095] Includes: Health Status (HI), Current Load, Future Load Prediction, Power Consumption Prediction, Operating Mode, Remaining Lifespan, Ambient Temperature and Humidity, Airflow Resistance Changes, Dimensions: 64-256.

[0096] Action Space A: Actions of each fan:

[0097] These are the actual control variables: rotational speed (RPM), airflow setpoint, air pressure target, and coordinated airflow strategy. The actions are continuous values ​​(not discrete), so DDPG / PPO processing is required.

[0098] The reward function (core innovation) is crucial to the entire scheduling strategy. The reward in this invention is a weighted sum:

[0099] in: (1) Energy consumption bonus: The lower the energy consumption, the higher the reward.

[0100] (2) Airflow meets the reward: Ensure that ventilation requirements are met.

[0101] (3) Safety Rewards (Health Penalties): Encourage wind turbines to operate in a healthy manner.

[0102] (4) Life protection bonus: Avoid overload operation, which can lead to accelerated degradation.

[0103] These rewards enable the system to achieve a multi-objective balance between energy saving, performance, safety, and lifespan.

[0104] Multi-Agent Reinforcement Learning (MARL) addresses the coupling between multiple wind turbine groups, such as: one turbine accelerating increases the air intake of another, wind pressure changes are not linear, and duct blockage affects local turbines. Traditional single-agent RL cannot handle these coupling relationships.

[0105] Therefore, this invention adopts a multi-agent framework: one agent per wind turbine, global rewards, shared state coding, and local control execution.

[0106] Key Model:

[0107] By employing MADDPG, each wind turbine learns its role within the system. After group coordination: energy consumption is reduced by 20-35%, duct fluctuations are reduced by 30-50%, and collaborative response speed is improved by 40-60%.

[0108] Safety constraints and interpretability mechanisms prevent dangerous actions from occurring in RL. This invention incorporates strong constraints:

[0109] It also employs safety filters: Lindberg perturbation suppression (LQR-like Filter), motion smoothing constraints, and a threshold for the difference between actions at adjacent times; Ensure that control signals are smooth and do not damage the equipment.

[0110] In addition, an interpretability mechanism (SHAP) is introduced to explain which features affect action decisions and which operating conditions affect wind turbine speed regulation, making it easier for maintenance personnel to understand AI behavior.

[0111] Edge execution and system feedback layer: executes scheduling strategies, monitors execution results, performs exception rollback, performs feedback learning, integrates with edge controllers, provides millisecond-level response, and enables autonomous control when the network is out of service; The Edge Control & Execution (ECE) layer is a key engineering module for the implementation of this invention, serving as the final link in "real-world execution of AI decision-making results on the wind turbine equipment." Located on the wind turbine side (Edge side), the ECE features strong real-time performance, high compatibility, and strong anti-interference capabilities. In the overall architecture, the ECE directly receives action commands from the AIDO (AI Scheduling and Optimization Decision Layer) and completes actual control, closed-loop feedback, parameter correction, anomaly protection, and multi-protocol adaptation.

[0112] ECE mainly includes: multi-protocol compatible controller (BACnet / modbus / CanOpen / M-Bus / EtherCAT), edge real-time controller (lower-level control of wind turbine frequency converter and control board), edge model inference module (simplified AI), real-time monitoring and protection strategy, and equipment-side data feedback and self-correction mechanism. Multi-protocol compatibility and edge execution gateway (EPG) To address the issue of protocol inconsistencies between different wind turbine manufacturers and controllers, the ECE of this invention introduces a multi-protocol edge execution gateway (EPG) that supports: BACnet / IP, BACnet MS / TP, Modbus RTU / Modbus TCP, CANopen, RS485 custom protocol, EtherCAT (Slave), and MQTT (cloud control backup). EPG uses a “protocol abstraction layer + device description file (DDF)” approach to map different protocols to a unified instruction set, such as: CTRL_SET_SPEED=0x01,GET_STATUS=0x02,GET_SENSOR=0x03; Regardless of whether the upper-level AIDO sends out BACnet objects, Modbus register addresses, or EtherCATCoE parameters, they are all uniformly mapped to the standard control interface of EPG, making the wind turbine group control change from "protocol-dependent" to "protocol-independent". At the edge, each fan is monitored by a real-time control unit (RTU), whose core tasks include: executing speed / airflow control commands (millisecond level), ensuring smooth control actions, preventing damage to the motor, collecting sensor data (temperature, pressure, vibration, current, power), handling fan faults, self-testing, abnormal temperature rise, overcurrent protection, and transmitting data back to the AI ​​layer in real time. RTU converts high-dimensional AI action commands into control curves that the wind turbine can execute:

[0113] Vector acceleration and deceleration control is implemented to prevent mechanical shock from the fan due to rapid changes in speed.

[0114] Edge-AI Lite (Edge AI Inference Module) To improve system stability, this invention introduces a small, lightweight AI model (such as TFLite or ONNX Runtime versions) into the ECE as a "local shadow model" for cloud / server AIDO. Its functions include: providing alternative AI decision-making during network outages (offline operation capability) and performing simple load forecasting.

[0115] It enables local airflow demand estimation and automatic compensation, executes local optimization strategies in latency-sensitive scenarios, and reduces reliance on the cloud. Edge-AI Lite is a simplified version of AIDO, but it is sufficient to support independent control of the wind turbine in a short period of time, giving the system extremely high robustness and reliability.

[0116] Edge-side anomaly detection and self-protection strategy (Safety Layer) ECE has a built-in industrial-grade security module that employs a dual mechanism of rules and prediction: (1) Hard rule protection: Includes: overcurrent protection, overtemperature protection, impeller stall protection, overspeed protection, sensor failure protection, and communication interruption protection (reducing to safe speed). Example of anomaly detection formula:

[0117]

[0118] (2) Predictive protection: Predicting future states using Edge-AI:

[0119] When the temperature is predicted to exceed the threshold in the next 5 minutes: reduce the speed in advance → avoid failure. This is an intelligent prevention mechanism that traditional control cannot achieve.

[0120] Closed-loop feedback mechanism of control loop The wind turbine's operating status is transmitted back to the ECE in real time, forming a closed loop through sensor data: AIDO command -> RTU execution -> sensor feedback -> ECE evaluation and correction -> adjustment and transmission back to AIDO as necessary, using the following adaptive correction formula:

[0121] in: For adaptive gain, : Required air volume, Actual air volume.

[0122] This enables automatic compensation for changes in duct resistance, temperature and humidity, and efficiency decline caused by equipment aging, thus keeping the control system in optimal condition over the long term.

[0123] Summary of the entire system's workflow The complete system operation process is as follows: ① Multi-protocol wind turbine access → EPG performs protocol adaptation, connecting devices such as BACnet, Modbus, and EtherCAT to a unified control framework; ② Real-time data perception → ISEP establishes a wind turbine profile and status prediction: current operating conditions (air volume / speed / pressure / power consumption), health status, failure probability, and load prediction for the next 10~30 minutes; ③ AIDO executes AI decisions and calculates the optimal scheduling strategy: RL outputs the speed / air volume of each fan, multi-agent coordination, and energy-minimizing scheduling; ④ ECE converts AI actions into actual control signals: smooth acceleration and deceleration, safety constraints, and local real-time adjustment; ⑤ Feedback data → Adaptive correction → Formation of a loop; Through closed-loop learning, the system can maintain its optimal state over a long period of time.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent adaptive scheduling method for multi-protocol industrial fan group control, characterized in that, Specifically, the following steps are included: S1. Spatiotemporal self-consistency real-time perception and semantic-level data generation of multi-protocol heterogeneous wind turbine groups; S2. Multi-protocol parsing based on dynamic semantic mapping; S3. Use Kalman filtering and data fusion technology to achieve accurate integration of data from multiple sensors; S4. Adjust the operating strategy in real time according to the equipment operating status, environmental changes and fan load to achieve dynamic optimization of fan group control; S5. Perform fault prediction and set up an early warning mechanism.

2. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 1, characterized in that, In S1, the core challenges of "multi-protocol devices not recognizing each other, multi-sensor data not synchronizing, high data noise, and lack of unified features for AI use" in wind turbine group control systems are addressed. This serves as the prerequisite for subsequent fusion, prediction, and scheduling. Wind turbine operating data, such as current, voltage, and temperature, are collected and transmitted to the scheduling decision module. Data acquisition employs high-precision sensors and multi-protocol data fusion technology to ensure data accuracy, timeliness, and spatiotemporal consistency. The scheduling algorithm automatically assesses the equipment's operating status and environmental requirements based on multi-dimensional environmental data and wind turbine load data.

3. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 1, characterized in that, In S2, protocol identification and parsing: For different industrial protocols, such as BACnet, M-Bus, EtherCAT, etc., the protocol type is first automatically identified, and a unified protocol parsing engine is used to classify the devices. During the protocol parsing process, information such as data fields, data types and function codes are abstracted into a unified semantic model to form a general data processing framework. Semantic mapping and dynamic conversion: By establishing a protocol semantic mapping library, the parsing engine dynamically maps data according to the semantic information of the protocol used by the device and performs corresponding format conversion. The system can automatically adjust the data format when the device is connected, avoiding data misreading or loss due to protocol differences. Cross-device data fusion: For different connected devices, the system can fuse the data of each device in real time according to the protocol mapping library and device feature library, form a unified data stream, and transmit it to the group control and scheduling system to realize cross-protocol collaborative control between wind turbines.

4. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 1, characterized in that, In S3, Kalman filtering is a recursive optimal estimation algorithm that is widely used in state estimation problems in dynamic systems. It is particularly suitable for multi-source signal fusion and noise suppression. Kalman filtering technology can achieve the following functions: noise suppression and data cleaning, spatiotemporal data fusion, and improved decision accuracy and response speed.

5. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 4, characterized in that, The noise suppression and data cleaning process includes: Sensors involved in the wind turbine group control system, such as temperature, humidity, wind speed, current, and voltage, are subject to varying degrees of noise interference. This noise may affect the accuracy of the data and lead to instability in the wind turbine control strategy. The Kalman filter uses a weighted average of the predicted value of the system state and the sensor measurement value to suppress sensor measurement errors and noise, and outputs an optimal estimate. The recursive nature of the Kalman filter allows it to process each new data input in real time, effectively remove outliers, and improve data accuracy.

6. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 4, characterized in that, The spatiotemporal data fusion includes: In a wind turbine group control system, the temporal and spatial synchronization of data from multiple sensors, such as temperature and humidity sensors, wind speed sensors, and power sensors, is crucial. Kalman filtering is used to fuse the temporal data of multiple sensors to ensure their spatiotemporal consistency. During wind turbine operation, the effects of changes in ambient temperature and wind speed need to be considered simultaneously. The Kalman filter can update the prediction model in real time and combine it with actual measurement data to make optimal estimates, obtaining a comprehensive value of multiple sensor data, thus avoiding the influence of errors from a single sensor data on control decisions.

7. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 4, characterized in that, The improvement in decision-making accuracy and response speed includes: through the real-time data update and prediction function of Kalman filtering, the system can accurately estimate the current state of the wind turbine group control system, such as load and operating efficiency, providing high-quality input data for the intelligent scheduling algorithm, significantly improving the accuracy and response speed of control decisions, and ensuring that the system is always in the optimal operating state under changing working conditions; The prediction steps and core formula of Kalman filtering are expressed as follows: S101, State Prediction: in, It represents the predicted system state, A is the state transition matrix, B is the control input matrix, and u... k-1 It is a control input (such as adjusting the operation of the fan); S102, Predicting Covariance: in: It is the predicted covariance matrix, and Q is the process noise covariance; S103, Kalman gain calculation: Where: K k Here, H is the Kalman gain, H is the observation matrix, and R is the measurement noise covariance. S104, Status Update: in: This is the updated state, z k These are actual measured values; S105, Covariance Update: Where: P k It is the updated covariance matrix; By fusing data from different sensors, such as temperature, wind speed, and current, and removing noise, accurate real-time state estimates are obtained, providing precise input for subsequent scheduling decisions.

8. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 1, characterized in that, In S2, fault prediction includes a machine learning fault prediction model, which uses convolutional neural networks (CNN) and support vector machines (SVM) to extract and predict fault features. By training a large amount of wind turbine operation data, it can effectively identify abnormal patterns of the equipment and issue fault warning signals in advance.

9. The intelligent adaptive scheduling method for multi-protocol industrial fan group control according to claim 1, characterized in that, In S2, the early warning mechanism includes fault warning and automatic response. When it is predicted that a fault is about to occur in the equipment, the system will issue an alarm in a timely manner and automatically adjust the operating strategy according to the fault type, such as activating the backup fan and adjusting the workload, to reduce the impact of the fault on the overall system operation.

10. An application system for an intelligent adaptive scheduling method for multi-protocol industrial fan group control according to any one of claims 1-9, characterized in that, Specifically, it includes five core layers: multi-protocol unified access layer, multi-source feature fusion layer, state assessment and prediction layer, digital twin modeling layer, AI scheduling and optimization decision layer, and edge execution and system feedback layer; Unified access layer for multiple protocols: It realizes unified parsing and model abstraction of BACnet, M-Bus, EtherCAT, MODBUS, private protocols, etc., and adopts protocol DSL, automatic parser generation (ParserGen), buffer queue mechanism, etc. Multi-source feature fusion layer: Integrates multi-source sensor data (temperature, humidity, pressure, wind speed, current, voltage, etc.) from the wind turbine equipment, and obtains a high-dimensional feature vector of wind turbine operation through Kalman filtering, state estimation and feature extraction, including algorithms such as EKF, UKF, Attention-Based Fusion and noise suppression; State assessment and prediction layer: performs health assessment (HI), load prediction, power consumption prediction, etc. on wind turbines, including LSTM, CNN, Transformer, and physical constraint-based models; Digital twin modeling layer: Constructs a simulation model that couples the fan, duct, and environment to achieve real-time mapping, inversion inference, and virtual-real comparison, and coordinates with aerodynamic models, thermal management models, system simulation models, etc. AI scheduling and optimization decision layer: Based on reinforcement learning, game theory mechanism, and multi-agent collaboration, formulate the global optimal scheduling strategy, using DQN, DDPG, PPO, MADDPG, multi-objective optimization, and multi-constraint solution. Edge execution and system feedback layer: executes scheduling strategies, monitors execution results, handles exceptions and rollbacks, performs feedback learning, integrates with edge controllers, provides millisecond-level response, and enables autonomous control during network outages.

Images