A Deep Learning-Based Real-Time Dynamic Compensation Measurement Method and System for Air Volume of Large-Scale Fans

Abstract

This invention discloses a method and system for real-time dynamic compensation measurement of air volume in large wind turbines based on deep learning, comprising: S1: real-time acquisition of initial data of the large wind turbine through preset multiple types of sensors; S2: optimization and synchronization alignment of the initial data of the wind turbine through industrial IoT edge nodes to obtain a multi-dimensional time-series data matrix with operating condition labels; S3: based on the multi-dimensional time-series data matrix, generating dynamic compensation features including eddy current suppression coefficients through an improved CNN-LSTM hybrid network; S4: decoding the optimal dynamic compensation air volume based on the dynamic compensation features through an adaptive reinforcement ensemble learning framework, and triggering a PID controller based on the optimal dynamic compensation air volume to realize dynamic compensation measurement of air volume. For wind turbine air volume, which is affected by multiple factors and has a certain degree of complexity, more accurate prediction can be achieved.

Description

Technical Field

[0001] This invention relates to the field of air volume measurement technology, and in particular to a method and system for real-time dynamic compensation measurement of air volume in large fans based on deep learning. Background Technology

[0002] In modern industrial production and energy supply systems, fans are key equipment widely used in many fields such as power, chemical, mining, and construction. For example, in thermal power plants, fans are responsible for providing the large amount of air needed for boiler combustion, and the precise control of their air volume is directly related to combustion efficiency and power generation costs; in mine ventilation systems, fans deliver fresh air to ensure the safety of underground workers, and a stable air supply is crucial.

[0003] Currently, traditional methods for measuring and controlling fan airflow have many limitations. In terms of airflow measurement, most methods use differential pressure or thermoelectric potential to measure airflow and then convert it into an airflow signal via a transmitter. However, this method is difficult to obtain accurate airflow data in complex duct environments (such as those with regulating dampers, support structures, elbows, baffles, and diameter changes). At the control level, traditional control strategies, represented by PID control, are difficult to adjust fan airflow flexibly and quickly in the face of frequent fluctuations in external environmental conditions (such as temperature, humidity, and duct pressure) due to their fixed control parameters. They often require manual intervention or excessive parameter presets, resulting in low adjustment efficiency. Therefore, this paper proposes a real-time dynamic compensation measurement method and system for large-scale fan airflow based on deep learning. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution:

[0005] A deep learning-based method for real-time dynamic compensation measurement of air volume in large-scale wind turbines includes:

[0006] S1: Real-time acquisition of initial data of large wind turbines through preset multiple types of sensors;

[0007] S2: Optimize, synchronize, and align the initial data of the wind turbine through industrial IoT edge nodes to obtain a multi-dimensional time-series data matrix with operating condition labels;

[0008] S3: Based on a multi-dimensional time-series data matrix, a dynamic compensation feature containing eddy current suppression coefficients is generated through an improved CNN-LSTM hybrid network;

[0009] The improved CNN-LSTM hybrid network is achieved by embedding a physical constraint layer based on the NS fluid dynamics equation into the original CNN-LSTM hybrid network, so that the output of the CNN-LSTM hybrid network can be dynamically compensated in accordance with the laws of fluid dynamics.

[0010] S4: Based on the dynamic compensation features, the optimal dynamic compensation air volume is decoded through an adaptive augmented ensemble learning framework, and the optimal dynamic compensation air volume triggers the PID controller to realize dynamic compensation measurement of air volume.

[0011] The various types of sensors include wind speed sensors, vortex sensors, motor speed sensors, and torque sensors.

[0012] The process of optimizing synchronization alignment is as follows:

[0013] Set a uniform data acquisition frequency f for each sensor, and at time intervals... Internal data acquisition, sensor setup At any moment The collected data is ,in Indicates the sensor number, Indicates the number of data collections;

[0014] Suppose that the sensor The received data sequence is After receiving data from n sensors, the industrial IoT edge node organizes this data into a two-dimensional matrix. Where m is the length of the time series, and optimized synchronization alignment is achieved.

[0015] The process for obtaining the multidimensional time-series data matrix with operating condition labels is as follows:

[0016] Suppose that the operating conditions of the fan are divided into k types, namely ;

[0017] Suppose there are l external environmental factors, namely ;

[0018] Based on the wind turbine's operating mode and external environmental conditions, operating condition labels are added to the collected data. Let's assume that at a certain time t, the wind turbine is in operating condition... External environmental factors are The data collected at that moment The corresponding operating condition label is ;

[0019] The final result is a multi-dimensional time-series data matrix with operating condition labels: m is the time series length, and n is the quantity index.

[0020] The improved CNN-LSTM hybrid network architecture includes a convolutional neural network, a long short-term memory network, and a physical constraint layer generated based on NS fluid dynamics.

[0021] The process of constructing the physical constraint layer based on NS fluid dynamics is as follows:

[0022] Based on the Navier-Stokes equations: ;

[0023] Where u is the velocity vector, t is the time, ρ is the fluid density, and p is the pressure. Kinematic viscosity, External force;

[0024] The NS equations are transformed to solve the vortex correlation equations, yielding the vortex suppression coefficient. Physical constraints generated based on NS fluid dynamics are embedded into the original CNN-LSTM hybrid network. The discretized form of the NS equations is used as part of the loss function to construct the physical constraint layer. Let the predicted vortex correlation features be... The actual physical characteristic is y, and the physical constraint loss function is... Represented as:

[0025] ,in, It is the number of features related to physical constraints;

[0026] After the physical constraint layer is constructed, the physical constraint loss function of the physical constraint layer is applied. The improved CNN-LSTM hybrid network is trained, and its output is a dynamically compensated feature Z containing vortex suppression coefficients.

[0027] The process of acquiring the adaptive reinforcement ensemble learning framework is as follows:

[0028] Set up a set of training samples and initialize the weights for each sample. After each iteration, train a weak learner. Obtain the error of the weak learner on the training set. According to the error Acquiring weak learners weight According to weak learners and corresponding weights Obtain a strong learner Where T is the number of iterations, For the nth preset dynamic compensation feature, complete the construction of the adaptive reinforcement ensemble learning framework.

[0029] The process for obtaining the optimal dynamic compensation air volume is as follows:

[0030] Input dynamic compensation features, and after T iterations during the training process of the adaptive reinforcement ensemble learning framework, obtain T weak learners h and their corresponding weights α. For each dynamic compensation feature Z, the optimal dynamic compensation air volume is obtained by decoding it through the strong learner of the adaptive reinforcement ensemble learning framework.

[0031] A deep learning-based real-time dynamic compensation measurement system for air volume of large-scale wind turbines includes:

[0032] Data acquisition module: Collects initial data of large wind turbines in real time through multiple preset sensors;

[0033] Time series acquisition module: Optimizes, synchronizes, and aligns the initial data of the wind turbine through industrial IoT edge nodes to obtain a multi-dimensional time series data matrix with operating condition labels;

[0034] Dynamic compensation module: Based on a multi-dimensional time-series data matrix, a dynamic compensation feature containing eddy current suppression coefficients is generated through an improved CNN-LSTM hybrid network;

[0035] Optimized measurement module: Based on dynamic compensation features, the optimal dynamic compensation air volume is decoded through an adaptive augmented ensemble learning framework, and the optimal dynamic compensation air volume triggers the PID controller to realize dynamic compensation measurement of air volume.

[0036] The present invention has the following beneficial effects:

[0037] In this invention, firstly, an improved CNN-LSTM hybrid network can extract deep dynamic compensation features from complex wind turbine operation data, effectively capturing the correlation of these data at different spatiotemporal scales. For data changes generated by wind turbines under different operating conditions (such as startup, stable operation, and variable speed operation), the network can adaptively learn and adjust. This avoids the limitations of traditional feature extraction methods when processing complex operating condition data.

[0038] Secondly, by embedding the Navier-Stokes equations (NS equations) as physical constraints into the network, the dynamic compensation features extracted by the network are ensured to conform to the laws of fluid dynamics. This makes the generated features more reasonable and reliable in a physical sense. For example, when predicting the air volume characteristics related to the flow field inside the fan, the intrinsic relationship between the fluid viscosity, pressure and velocity can be taken into account, thereby improving the accuracy of air volume prediction.

[0039] Finally, multiple weak learners (simple decision trees) are iteratively trained using an adaptive reinforcement ensemble learning framework, and the sample weights are dynamically adjusted based on their performance during training. Ultimately, these weak learners are combined into a strong learner. This mechanism can fully utilize the advantages of different weak learners and achieve more accurate predictions for data such as wind turbine air volume, which is affected by multiple factors and has a certain degree of complexity. Attached Figure Description

[0040] Figure 1 This diagram illustrates the steps of the real-time dynamic compensation measurement method and system for large-scale wind turbine air volume based on deep learning proposed in this invention.

[0041] Figure 2 This is a system block diagram of the real-time dynamic compensation measurement method and system for air volume of large wind turbines based on deep learning proposed in this invention. Detailed Implementation

[0042] 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.

[0043] Example 1: As Figure 1 As shown, the real-time dynamic compensation measurement method for air volume of large wind turbines based on deep learning proposed in this invention includes:

[0044] S1: Real-time acquisition of initial data of large wind turbines through preset multiple types of sensors;

[0045] Wind speed sensor data: Wind speed sensors are installed at the inlet and outlet of the fan. The inlet wind speed data can reflect the airflow entering the fan, while the outlet wind speed data can reflect the fan's acceleration effect on the airflow. Wind direction data helps to analyze the airflow path and the airflow interaction between the fan and the surrounding environment.

[0046] Let the inlet wind speed be The outlet wind speed is According to the law of conservation of mass, the air volume Q of the fan is related to the wind speed: ,in and These are the effective cross-sectional areas of the fan inlet and outlet, respectively.

[0047] Vortex sensor data: Vortex sensors are installed inside and around the fan to detect vortices in the airflow. The generation and development of vortices affect the fan's airflow and efficiency. The intensity and vorticity ω of the vortex are obtained by monitoring its intensity frequency.

[0048] Motor speed and torque data: Install motor speed and torque sensors. Let the motor speed be n. The relationship between the fan air volume Q and the motor speed is expressed as follows: ,That It is a proportionality coefficient related to the characteristics of the wind turbine, obtained through prior experience;

[0049] The relationship between motor torque T and motor power P is as follows: ,in, It is the angular velocity of the motor;

[0050] Specifically, the fan's air volume is directly related to the motor speed, and the fan's air volume is proportional to the motor speed. Torque data can reflect the motor's load and indirectly reflect the fan's working intensity. The motor power is related to the fan's air volume and pressure head. Through these relationships, the fan's operating status can be comprehensively analyzed.

[0051] S2: Optimize, synchronize, and align the initial data of the wind turbine through industrial IoT edge nodes to obtain a multi-dimensional time-series data matrix with operating condition labels;

[0052] The process of optimizing synchronization alignment is as follows:

[0053] Set a uniform data acquisition frequency f (f=100Hz) for each sensor, and set a time interval between them. Internal data collection;

[0054] Set sensor At any moment The collected data is ,in Indicates the sensor number, Indicates the number of data collections;

[0055] Industrial IoT edge nodes are responsible for receiving data from various sensors and performing preliminary processing on this data:

[0056] Suppose that the sensor The received data sequence is After receiving data from n sensors, the industrial IoT edge node organizes this data into a two-dimensional matrix. Where m is the length of the time series, and optimized synchronization alignment is achieved;

[0057] The process of obtaining a multidimensional time-series data matrix with operating condition labels is as follows:

[0058] First, the edge nodes sort the timestamped data to ensure that the data is arranged in chronological order. Then, the edge nodes perform an integrity check on the data to ensure that there are complete data records in each collection cycle.

[0059] If missing or abnormal data is detected, edge nodes will mark it and attempt to repair the data, such as by using interpolation or other data recovery algorithms to fill in the missing data points.

[0060] Suppose that the operating conditions of the fan are divided into k types, each represented by a different type. It is indicated that, at the same time, there are l external environmental factors, using express;

[0061] Based on the wind turbine's operating mode and external environmental conditions, operating condition labels are added to the collected data. Let's assume that at a certain time t, the wind turbine is in operating condition... External environmental factors are The data collected at that moment The corresponding operating condition label is ;

[0062] The final result is a multi-dimensional time-series data matrix with operating condition labels: It includes multi-dimensional data such as pressure, temperature, and vibration of the fan under different operating conditions.

[0063] S3: Based on a multi-dimensional time-series data matrix, a dynamic compensation feature containing eddy current suppression coefficients is generated through an improved CNN-LSTM hybrid network;

[0064] The obtained multidimensional time series data matrix is ​​normalized to the [0,1] interval;

[0065] Using a pure CNN-LSTM hybrid network, the normalized multidimensional time series data matrix is ​​segmented according to the time series. A sliding window method is used, with a window size of 100 data points and a step size of 10 data points, to divide the continuous time series data into multiple data subsets. Each subset serves as the input sample of the pure CNN-LSTM hybrid network.

[0066] Furthermore, by using sliding window processing, the correlation information between different time series data can be fully utilized to improve the network's ability to extract data features;

[0067] The improved CNN-LSTM hybrid network is achieved by embedding a physical constraint layer based on the NS fluid dynamics equation into the original CNN-LSTM hybrid network, so that the output of the CNN-LSTM hybrid network can be dynamically compensated in accordance with the laws of fluid dynamics.

[0068] Improvements to the CNN-LSTM hybrid network architecture include:

[0069] The improved CNN-LSTM hybrid network consists of a convolutional neural network (CNN), a long short-term memory network (LSTM), and a physical constraint layer generated based on NS fluid dynamics. The CNN part is mainly responsible for extracting the spatial features of the input data, while the LSTM part is used to process long-term dependencies in temporal data.

[0070] In the CNN part, multiple convolutional layers and pooling layers are set up, and the convolutional layers use small 3×3 convolutional kernels;

[0071] Convolution operations can be used to extract local features from data;

[0072] The pooling layer uses max pooling or average pooling to reduce the dimensionality of the convolutional features, reducing the amount of data while retaining the main features.

[0073] The LSTM part consists of multiple LSTM units, each containing an input gate, a forget gate, and an output gate. LSTM can remember information from long-term sequences, avoiding the gradient vanishing problem in traditional recurrent neural networks (RNNs). By performing temporal processing on features after CNN processing through LSTM units, dynamic changes in the data can be captured to obtain vortex-related features. ;

[0074] The process of constructing the physical constraint layer based on NS fluid dynamics is as follows:

[0075] Embed a physical constraint layer based on NS fluid dynamics generation into an improved CNN-LSTM hybrid network;

[0076] The Navier-Stokes equations are expressed as follows:

[0077] ;

[0078] Where u is the velocity vector, t is the time, ρ is the fluid density, and p is the pressure. Kinematic viscosity, External forces describe the laws governing fluid motion;

[0079] Specifically, by discretizing the NS equations and embedding them as physical constraints into the network, the dynamic compensation characteristics of the network output can conform to the laws of fluid dynamics.

[0080] By transforming the Navier-Stokes equations and solving the related equations of vorticity, the vortex suppression coefficient can be obtained. Specifically, by taking the curl of the Navier-Stokes equations, the vorticity equations are obtained:

[0081] ;

[0082] In this equation, by analyzing the generation, diffusion, and dissipation mechanisms of vorticity, the relationship between the vortex suppression coefficient and the terms in the equation can be defined, and the relationship between the vortex suppression coefficient and the diffusion term of vorticity can be established. The eddy current suppression coefficient is obtained by adjusting the coefficients of these terms.

[0083] The discretized form of the Navier-Stokes equations is used as part of the loss function to complete the construction of the physical constraint layer;

[0084] In practice, during network training, the network output is made to satisfy the physical relationship described by the Navier-Stokes equations as much as possible. For example, when calculating the loss function, in addition to traditional loss terms such as mean square error (MSE), physical constraint terms based on the Navier-Stokes equations are added. By adjusting the weights of the network, the vortex suppression coefficient output by the network is made to better match the actual fluid dynamics characteristics.

[0085] Network training and optimization:

[0086] The improved CNN-LSTM hybrid network is trained using a subset of multidimensional time-series data with working condition labels. The data subset is divided into training, validation, and test sets, for example, in proportions of 70%, 20%, and 10%.

[0087] During training, the Adam optimization algorithm is used. The Adam algorithm combines the advantages of the Adaptive Gradient Algorithm (Adagrad) and the Root Mean Square Propagation (RMSProp). It can adaptively adjust the learning rate, accelerate the convergence of the network, and continuously adjust the hyperparameters of the network, including the number of convolutional layers, the number of LSTM units, and the learning rate. The optimal combination of hyperparameters is selected by evaluating the network performance on the validation set.

[0088] Specifically, the network performance evaluation metric uses mean absolute error (MAE) to ensure that the network can accurately generate dynamic compensation features including vortex suppression coefficients.

[0089] Physical constraints generated based on NS fluid dynamics are embedded into the original CNN-LSTM hybrid network, and the discretized form of the NS equations is used as part of the loss function;

[0090] Let the predicted vortex correlation characteristics be: (Obtained from LSTM output), the actual physical characteristic (calculated based on the Navier-Stokes equations) is y, and the physical constraint loss function is... It can be represented as:

[0091] ;

[0092] in, It is the number of features related to physical constraints;

[0093] Loss function after physical constraints The improved CNN-LSTM hybrid network is trained, and its output is a dynamically compensated feature Z containing vortex suppression coefficients;

[0094] Specifically, assuming the final output layer of the network has d neurons, the dimension of the dynamic compensation feature Z is [d], and its element values ​​are the vortex suppression-related feature values ​​predicted by the network based on the input multidimensional time-series data. These feature values ​​can be used for subsequent dynamic airflow compensation calculations.

[0095] S4: Based on the dynamic compensation features, the optimal dynamic compensation air volume is decoded through an adaptive augmented ensemble learning framework, and the optimal dynamic compensation air volume triggers the PID controller to realize dynamic compensation measurement of air volume.

[0096] The dynamic compensation features output from the improved CNN-LSTM hybrid network have different numerical ranges and dimensions. These feature vectors are standardized so that the mean of each feature is 0 and the variance is 1.

[0097] We perform feature selection and dimensionality reduction on the standardized dynamic compensation features, and design an adaptive reinforcement ensemble learning framework to decode the dynamic compensation features.

[0098] The construction process of the adaptive reinforcement ensemble learning framework is as follows:

[0099] Let a set of training samples ,in, It is the nth preset dynamic compensation feature. It is the nth corresponding actual air volume value;

[0100] First, initialize the weights for each sample. After each iteration, a weak learner is trained. (e.g., a simple decision tree) to obtain the error of the weak learner on the training set. ,in It is an indicator function;

[0101] According to error Acquiring weak learners weight ;

[0102] Specifically, the process of updating sample weights is represented as: ,in, It is a normalization factor. Indicates the sample weights for the next iteration;

[0103] Based on weak learners and corresponding weights To obtain a strong learner, the final strong learner is a linear combination of multiple weak learners, expressed as: , where T is the number of iteration rounds, to complete the construction of the adaptive reinforcement ensemble learning framework;

[0104] Specifically, when classifying new samples, the strong learner H will comprehensively consider the prediction results of each weak learner and perform a weighted sum according to their weights to obtain the final classification result, thus completing the construction of the adaptive reinforcement ensemble learning framework.

[0105] In the completed adaptive reinforcement ensemble learning framework, the optimal dynamic compensation air volume is decoded through the strong learner of the adaptive reinforcement ensemble learning framework. The specific process is as follows:

[0106] The input data is the dynamically compensated feature Z obtained after previous processing. During the training process of the adaptive reinforcement ensemble learning framework, after T iterations, T weak learners h and their corresponding weights α are obtained. For each dynamically compensated feature Z, and according to the formula of the strong learner in the adaptive reinforcement ensemble learning framework... Decoding the optimal dynamic compensation airflow ;

[0107] Specifically, each weak learner h is trained by processing the feature vector using a specific algorithm (decision stump) to obtain an intermediate predicted value, which is then weighted and summed with its weight α to obtain the final optimal dynamic compensation air volume. ;

[0108] The process of achieving dynamic compensation measurement of air volume is as follows:

[0109] In the dynamic compensation of fan air volume, the set value is the desired air volume value, and the actual measured value is the optimal dynamic compensation air volume obtained through the above steps. The PID controller adjusts the control signal according to the magnitude and trend of the error so that the actual air volume of the fan approaches the desired air volume.

[0110] Based on the optimal dynamic compensation air volume, the PID control parameters are adjusted. The air volume error e(t) is calculated according to the selected optimal dynamic compensation air volume, and the air volume error is input into the PID controller. Dynamic compensation of air volume is achieved by adjusting the parameters of the PID controller.

[0111] When the airflow error is large, the proportional coefficient can be appropriately increased, for example:

[0112] Assume that when the air volume error e(t) is greater than or equal to When the air volume is measured in cubic meters per hour, the error is considered relatively large. In this case, the proportional gain of the PID controller can be appropriately increased. Let the initial... When the air volume error is greater than or equal to At that time, Increase to 1.0;

[0113] This is because a larger proportional gain allows the controller to respond more strongly to larger errors, thereby quickly reducing airflow errors;

[0114] Assuming the airflow error e(t) is less than If the airflow error is considered small, the proportional gain of the PID controller can be appropriately reduced. ,Will Reduced from 0.5 to 0.3;

[0115] A smaller proportional gain helps prevent the controller from over-adjusting when the error is small, thus preventing system oscillations.

[0116] By properly adjusting the parameters of the PID controller, precise dynamic compensation of the fan's air volume can be achieved, ensuring that the fan can stably output the desired air volume under different operating conditions.

[0117] Example 2: Figure 2 As shown, a large-scale wind turbine air volume real-time dynamic compensation measurement system based on deep learning includes:

[0118] Data acquisition module: Collects initial data of large wind turbines in real time through multiple preset sensors;

[0119] Time series acquisition module: Optimizes, synchronizes, and aligns the initial data of the wind turbine through industrial IoT edge nodes to obtain a multi-dimensional time series data matrix with operating condition labels;

[0120] Dynamic compensation module: Based on a multi-dimensional time-series data matrix, a dynamic compensation feature containing eddy current suppression coefficients is generated through an improved CNN-LSTM hybrid network;

[0121] Optimized measurement module: Based on dynamic compensation features, the optimal dynamic compensation air volume is decoded through an adaptive augmented ensemble learning framework, and the optimal dynamic compensation air volume triggers the PID controller to realize dynamic compensation measurement of air volume.

[0122] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time dynamic compensation measurement of air volume in large-scale fans based on deep learning, characterized in that, include: S1: Real-time acquisition of initial data of large wind turbines through preset multiple types of sensors; S2: Optimize, synchronize, and align the initial data of the wind turbine through industrial IoT edge nodes to obtain a multi-dimensional time-series data matrix with operating condition labels; S3: Based on a multi-dimensional time-series data matrix, a dynamic compensation feature containing eddy current suppression coefficients is generated through an improved CNN-LSTM hybrid network; The improved CNN-LSTM hybrid network is achieved by embedding a physical constraint layer based on the NS fluid dynamics equation into the original CNN-LSTM hybrid network, so that the output of the CNN-LSTM hybrid network can be dynamically compensated in accordance with the laws of fluid dynamics. S4: Based on the dynamic compensation features, the optimal dynamic compensation air volume is decoded through an adaptive augmented ensemble learning framework, and the optimal dynamic compensation air volume triggers the PID controller to realize dynamic compensation measurement of air volume.

2. The method for real-time dynamic compensation measurement of air volume in large-scale fans based on deep learning according to claim 1, characterized in that, The various types of sensors include wind speed sensors, vortex sensors, motor speed sensors, and torque sensors.

3. The method for real-time dynamic compensation measurement of air volume of large-scale fans based on deep learning according to claim 1, characterized in that, The process of optimizing synchronization alignment is as follows: Set a uniform data acquisition frequency f for each sensor, and at time intervals... Internal data acquisition, sensor setup At any moment The collected data is ,in Indicates the sensor number, Indicates the number of data collections; Suppose that the sensor The received data sequence is After receiving data from n sensors, the industrial IoT edge node organizes this data into a two-dimensional matrix. , where m is the length of the time series, to achieve optimized synchronization alignment.

4. The method for real-time dynamic compensation measurement of air volume of large-scale fans based on deep learning according to claim 1, characterized in that, The process for obtaining the multidimensional time-series data matrix with operating condition labels is as follows: Suppose that the operating conditions of the fan are divided into k types, namely ; Suppose there are l external environmental factors, namely ; Based on the wind turbine's operating mode and external environmental conditions, operating condition labels are added to the collected data. Let's assume that at a certain time t, the wind turbine is in operating condition... External environmental factors are The data collected at that moment The corresponding operating condition label is ; The final result is a multi-dimensional time-series data matrix with operating condition labels: m is the time series length, and n is the quantity index.

5. The method for real-time dynamic compensation measurement of air volume of large-scale fans based on deep learning according to claim 4, characterized in that, The improved CNN-LSTM hybrid network architecture includes a convolutional neural network, a long short-term memory network, and a physical constraint layer generated based on NS fluid dynamics.

6. The method for real-time dynamic compensation measurement of air volume of large-scale fans based on deep learning according to claim 5, characterized in that, The process of constructing the physical constraint layer based on NS fluid dynamics is as follows: Based on the Navier-Stokes equations: ; Where u is the velocity vector, t is the time, ρ is the fluid density, and p is the pressure. Kinematic viscosity, External force; The NS equations are transformed to solve the vortex correlation equations, yielding the vortex suppression coefficient. Physical constraints generated based on NS fluid dynamics are embedded into the original CNN-LSTM hybrid network. The discretized form of the NS equations is used as part of the loss function to construct the physical constraint layer. Let the predicted vortex correlation features be... The actual physical characteristic is y, and the physical constraint loss function is... Represented as: ,in, It is the number of features related to physical constraints; After the physical constraint layer is constructed, the physical constraint loss function of the physical constraint layer is applied. The improved CNN-LSTM hybrid network is trained, and its output is a dynamically compensated feature Z containing vortex suppression coefficients.

7. The method for real-time dynamic compensation measurement of air volume of large-scale fans based on deep learning according to claim 6, characterized in that, The process of acquiring the adaptive reinforcement ensemble learning framework is as follows: Set up a set of training samples and initialize the weights for each sample. After each iteration, train a weak learner. Obtain the error of the weak learner on the training set. According to the error Acquiring weak learners weight According to weak learners and corresponding weights Obtain a strong learner Where T is the number of iterations, For the nth preset dynamic compensation feature, complete the construction of the adaptive reinforcement ensemble learning framework.

8. The method for real-time dynamic compensation measurement of air volume of large fans based on deep learning according to claim 7, characterized in that, The process for obtaining the optimal dynamic compensation air volume is as follows: Input dynamic compensation features, and after T iterations during the training process of the adaptive reinforcement ensemble learning framework, obtain T weak learners h and their corresponding weights α. For each dynamic compensation feature Z, the optimal dynamic compensation air volume is obtained by decoding it through the strong learner of the adaptive reinforcement ensemble learning framework.

9. A real-time dynamic compensation measurement system for air volume of large-scale fans based on deep learning, implemented according to the method of any one of claims 1-8, characterized in that, include: Data acquisition module: Collects initial data of large wind turbines in real time through multiple preset sensors; Time series acquisition module: Optimizes, synchronizes, and aligns the initial data of the wind turbine through industrial IoT edge nodes to obtain a multi-dimensional time series data matrix with operating condition labels; Dynamic compensation module: Based on a multi-dimensional time-series data matrix, a dynamic compensation feature containing eddy current suppression coefficients is generated through an improved CNN-LSTM hybrid network; Optimized measurement module: Based on dynamic compensation features, the optimal dynamic compensation air volume is decoded through an adaptive augmented ensemble learning framework, and the optimal dynamic compensation air volume triggers the PID controller to realize dynamic compensation measurement of air volume.

Images