Multi-parameter wind tunnel control method and system based on neural network decoupling

Abstract

The present application relates to the technical field of neural network model, solve the problem of mutual interference between parameters in the case of simultaneous adjustment of multiple environmental control parameters in the existing wind tunnel test, provide a kind of multiple parameter wind tunnel control method and system based on neural network decoupling.The method comprises: obtaining the measurement value and preset value of each environmental control parameter in the wind tunnel;The measurement value and the preset value are input into the pre-trained neural network decoupling model to obtain the decoupling compensation corresponding to each environmental control parameter;According to the measurement value, the preset value and decoupling compensation, PID control calculation is carried out to obtain each target control quantity;According to each target control quantity, the corresponding each environmental control parameter is adjusted to realize wind tunnel control.The present application can effectively inhibit the mutual interference between different environmental control parameters, and still maintain high control precision and stability in the case of simultaneous adjustment of multiple parameters, suitable for different test conditions.

Description

Technical Field

[0001] This invention relates to the field of neural network model technology, and in particular to a multi-parameter wind tunnel control method and system based on neural network decoupling. Background Technology

[0002] Wind tunnel testing is a common experimental method in aerospace, vehicle engineering, and materials and device reliability verification. Modern wind tunnel systems are no longer limited to simulating single wind speed or temperature conditions, but have gradually evolved into wind tunnel systems capable of simultaneously controlling multiple environmental parameters such as temperature, humidity, wind speed, and radiation conditions. However, in multi-parameter climate wind tunnels, there are often coupling relationships between the various environmental control parameters. For example, changes in wind speed will affect the distribution of temperature and humidity, and radiative heating or cooling will further affect the airflow state and temperature distribution. Adjustment of any parameter often causes changes in other parameters, and the system as a whole exhibits strong nonlinearity and time-varying characteristics.

[0003] Existing wind tunnel environmental control systems typically employ separate control loops for parameters such as temperature, wind speed, or radiation, using PID controllers for closed-loop regulation of each parameter. Some schemes utilize empirical compensation or simple coordination logic for compensation; however, these methods lack rapid response to changes and are not universally applicable. Research has attempted to introduce adaptive control or model-based control methods, but these are susceptible to model mismatch issues due to factors such as equipment response lag and environmental disturbances. Furthermore, while some existing technologies incorporate intelligent algorithms like neural networks for parameter adjustment or auxiliary control, they primarily focus on optimizing single environmental control parameters. Existing patent CN110702357B (A hot and humid climate wind tunnel and its multi-field coupling control system, 2020.06.26) discloses the combination and control coordination of temperature, humidity, wind speed, and other experimental module equipment within the wind tunnel. It mainly concerns system structural configuration, such as the combined hardware and control cabinet layout of the wind tunnel body, fans, temperature devices, and variable humidity devices, but rarely mentions specific control algorithm characteristics. It does not address parameter adjustment schemes through neural network decoupling, failing to fundamentally solve the problem of mutual interference between parameters.

[0004] In summary, effectively suppressing the mutual interference between multiple environmental control parameters when they are adjusted simultaneously has become an urgent technical problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a multi-parameter wind tunnel control method and system based on neural network decoupling, in order to solve the problem of mutual interference between parameters in the prior art when multiple environmental control parameters are adjusted simultaneously.

[0006] In a first aspect, embodiments of the present invention provide a multi-parameter wind tunnel control method and system based on neural network decoupling, the method comprising:

[0007] Obtain the measured and preset values ​​of various environmental control parameters preset in the wind tunnel;

[0008] The measured values ​​and the preset values ​​are input into a pre-trained neural network decoupling model to obtain the decoupling compensation amount corresponding to each environmental control parameter;

[0009] Based on the measured value, the preset value, and the decoupling compensation amount, PID control calculation is performed to obtain each target control quantity;

[0010] Based on the target control quantities, the corresponding environmental control parameters are adjusted to achieve wind tunnel control.

[0011] Preferably, the neural network decoupling model adopts a distributed structure, which includes a set of sub-networks. The set of sub-networks includes several independent sub-networks, each of which corresponds to an environmental control parameter. Each sub-network includes a multi-layer feedforward neural network, and each layer of the feedforward neural network uses the hyperbolic tangent activation function.

[0012] Preferably, before inputting the measured value and the preset value into the pre-trained neural network decoupling model to obtain the decoupling compensation amount corresponding to each environmental control parameter, the following steps are included:

[0013] Acquire several sets of raw data, each set of raw data including sample preset values ​​of environmental control parameters, sample measured values, and ideal control quantities when controlling the wind tunnel to a steady state;

[0014] The original data is normalized to obtain a preset training sample;

[0015] The weights of each subnetwork in the neural network are randomly initialized to obtain the initialized neural network.

[0016] Obtain the set of subnetworks and the target subnetwork within the set of subnetworks;

[0017] With minimizing the pre-constructed objective function as the training objective, the weights of each sub-network in the initialized neural network are adjusted to obtain the weights of the trained sub-networks.

[0018] According to the preset order, a non-target subnetwork in the subnetwork set is obtained as a new target subnetwork. Then, the steps of minimizing the pre-constructed objective function as the training objective and adjusting the weights of each subnetwork in the initialized neural network are returned to obtain the weights of the trained subnetworks. This process continues until the preset iterative training stopping condition is met, at which point the weights of all trained subnetworks are obtained.

[0019] Based on the weights of all trained subnetworks, a pre-trained neural network decoupling model is obtained.

[0020] Preferably, the step of obtaining a non-target subnetwork from the subnetwork set in a preset order as a new target subnetwork, returning to the step of adjusting the weights of each subnetwork in the initialized neural network with the training objective of minimizing the pre-constructed objective function, and obtaining the weights of the trained subnetworks, continues until a preset iterative training stopping condition is met, thus obtaining the weights of all trained subnetworks, including:

[0021] Obtain a non-target subnetwork from the subnetwork set as a new target subnetwork, and use minimizing the objective function as the training objective. Adjust the weights of each subnetwork in the initialized neural network to obtain the weights of the trained subnetwork.

[0022] When all sub-networks corresponding to each environmental control parameter have completed one round of training, determine whether the objective function value of each sub-network is less than the preset function threshold.

[0023] When there is an objective function value greater than or equal to a preset function threshold and the number of iterations is less than the preset number of training iterations, return to the step of obtaining a non-target subnetwork from the subnetwork set as a new target subnetwork, with minimizing the objective function as the training objective, adjusting the weights of each subnetwork in the initialized neural network, and obtaining the weights of the trained subnetworks, until the objective function values ​​corresponding to all subnetworks are less than the preset threshold or the number of iterations is equal to the preset number of training iterations, and then obtain the weights of all trained subnetworks.

[0024] Preferably, after completing the first round of training, the step of obtaining a non-target subnetwork from the subnetwork set as a new target subnetwork, with minimizing the objective function as the training objective, and adjusting the weights of each subnetwork in the initialized neural network to obtain the trained subnetwork weights includes:

[0025] Based on the objective function and the backpropagation algorithm, the gradient information of the environmental control parameters relative to each network weight is calculated to obtain the weight correction amount;

[0026] Obtain the weight change of the environmental control parameter weights in the previous iteration round, and weight the weight change according to the preset momentum factor to obtain the historical weight correction amount;

[0027] The weight update amount is obtained by weighting the weight correction amount with the historical weight correction amount;

[0028] Based on the weight update amount, the weights of the sub-networks corresponding to the environmental control parameters are adjusted to obtain the weights of the trained sub-networks.

[0029] Preferably, the multilayer feedforward neural network includes an input layer, a hidden layer, and an output layer. The step of inputting the measured value and the preset value into a pre-trained neural network decoupling model to obtain the decoupling compensation amount for each channel corresponding to the environmental control parameter includes:

[0030] The measured values ​​and the preset values ​​are grouped by channel to obtain channel input data;

[0031] The channel input data is input into the neural network decoupling model, and normalized through the input layer to obtain the normalized preset value and the measured value.

[0032] In the hidden layer, the normalized preset value and the measured value are weighted according to the preset bias parameters to obtain the weighted input value of the hidden layer.

[0033] Substitute the weighted input values ​​into a preset activation function to obtain the activation output result;

[0034] In the output layer, the activated output result is subjected to inverse normalization mapping processing according to the preset compensation value range to obtain the final decoupling compensation amount.

[0035] Preferably, the step of performing PID control calculations based on the measured value, the preset value, and the decoupling compensation amount to obtain each target control quantity includes:

[0036] The current decoupling error after compensation is calculated based on the preset value, the measured value, and the decoupling compensation amount.

[0037] Obtain historical decoupling errors and historical control variables;

[0038] Based on the current decoupling error and the historical decoupling error, PID control calculations are performed to obtain the control increment;

[0039] The target control quantity is calculated based on the control quantity increment and the historical control quantity.

[0040] Preferably, the method further includes:

[0041] Obtain the fluctuation range of the measured values ​​of each environmental control parameter;

[0042] When the fluctuation amplitude is less than the preset fluctuation threshold within a consecutive preset number of control cycles, the wind tunnel is determined to be in a stable state.

[0043] When the wind tunnel is in a stable state, the preset value vector, the measured value vector, and the ideal control vector when controlling the wind tunnel to a stable state are used as the current sample, and the Euclidean distance between the current sample and the adjacent training samples is calculated.

[0044] When the Euclidean distance is greater than a preset distance threshold, the current sample is output to a preset target sample set;

[0045] When the preset fine-tuning conditions are met, the sub-network is adjusted according to the target sample set to obtain the adjusted pre-trained neural network decoupling model. The preset fine-tuning conditions include the number of samples in the target sample set being equal to a preset sample number threshold or the time interval since the last fine-tuning being equal to a preset fine-tuning time threshold.

[0046] Preferably, the method further includes:

[0047] The training order of each sub-network is determined based on the coupling strength between various environmental control parameters and / or the preset control priority.

[0048] According to the training order, the sub-networks corresponding to each environmental control parameter are trained alternately to obtain a pre-trained neural network decoupling model.

[0049] In a second aspect, embodiments of the present invention provide a multi-parameter wind tunnel control system based on neural network decoupling, characterized in that it includes: at least one processor, at least one memory, and computer program instructions stored in the memory, wherein when the computer program instructions are executed by the processor, the method described in any of the first aspects is implemented.

[0050] In summary, the beneficial effects of the present invention are as follows:

[0051] The multi-parameter wind tunnel control method and system based on neural network decoupling provided in this invention obtains the measured values ​​and preset values ​​of various environmental control parameters within the wind tunnel; inputs the measured values ​​and preset values ​​into a pre-trained neural network decoupling model to obtain the decoupling compensation amount corresponding to each environmental control parameter; performs PID control calculation based on the measured values, preset values, and decoupling compensation amounts to obtain each target control amount; and adjusts the corresponding environmental control parameters according to each target control amount to achieve wind tunnel control. This invention compensates for the coupling effects of various environmental control parameters in a multi-parameter wind tunnel system through a distributed neural network decoupling structure, reducing overshoot and oscillations caused by mutual pulling between channels, while also reducing reliance on manual adjustment. It can effectively suppress mutual interference between different environmental control parameters and adapt to the nonlinear, time-varying, and frequently switching characteristics of wind tunnel systems. Therefore, it can maintain high control accuracy and stability even when multiple parameters are adjusted simultaneously, improving the reliability and applicability of the system under different experimental conditions. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, and these are all within the protection scope of the present invention.

[0053] Figure 1 This is a schematic diagram of the overall process of the multi-parameter wind tunnel control method based on neural network decoupling in Embodiment 1 of the present invention;

[0054] Figure 2 This is a schematic diagram of the process of performing weight updates on a sub-network in Embodiment 1 of the present invention;

[0055] Figure 3 This is a schematic diagram of the three-layer feedforward neural network in Embodiment 1 of the present invention;

[0056] Figure 4 This is a schematic diagram of the structure of the multi-parameter wind tunnel control system based on neural network decoupling in Embodiment 3 of the present invention. Detailed Implementation

[0057] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only configured to explain the present invention and are not configured to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely intended to provide a better understanding of the present invention by illustrating examples of the invention.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0059] Example 1

[0060] Please see Figure 1This invention provides a multi-parameter wind tunnel control method based on neural network decoupling, the method comprising:

[0061] S1: Obtain the measured and preset values ​​of various environmental control parameters preset in the wind tunnel;

[0062] Specifically, the preset environmental control parameters refer to the physical quantities that the wind tunnel needs to control, such as temperature, humidity, wind speed, irradiance, and background radiation. The measured values ​​are those acquired in real-time by sensors or measurement systems, while the preset values ​​are target values ​​issued by the experimenter or a host computer, such as setting the temperature to 40℃, humidity to 60%RH, and wind speed to 10 m / s. These measured values ​​can be acquired through temperature and humidity sensors, hot-wire anemometers, irradiance meters, infrared thermometers, or multi-sensor fusion modules.

[0063] S2: Input the measured value and the preset value into the pre-trained neural network decoupling model to obtain the decoupling compensation amount corresponding to each environmental control parameter;

[0064] Specifically, the pre-trained neural network decoupling model refers to a model obtained after training the neural network with samples. The neural network used is a BP neural network, RBF neural network (Radial Basis Function), CNN (Convolutional Neural Network), or LSTM (Long Short Term Memory), which possesses strong decoupling dynamic characteristics or time-series modeling capabilities. The decoupling compensation is the compensation term output by the model, i.e., the correction amount for the error of each environmental control parameter. For example, the decoupling compensation for temperature is used to offset the crosstalk between humidity regulation or wind speed changes and temperature. Within the controller, measured values ​​and preset values ​​can be grouped according to environmental control parameters and converted into input vectors, which are then input into the pre-trained neural network decoupling model, outputting the compensation amount corresponding to each environmental control parameter. Before entering the PID calculation, the error offset caused by channel coupling is estimated and processed, making the subsequent control of each environmental control parameter's corresponding channel closer to the ideal state of independent loop control, reducing overshoot and oscillation caused by the mutual pull of multiple loops.

[0065] S3: Based on the measured value, the preset value, and the decoupling compensation amount, perform PID control calculations to obtain each target control quantity;

[0066] Specifically, PID control calculations utilize proportional, integral, and derivative control algorithms to generate control outputs based on the current error. The target control quantity is the control command given to the actuator, such as controlling heating power, humidification rate, fan speed, or radiation source output power. The base error is obtained by subtracting the measured value from the preset value. This base error is then combined with decoupling compensation to form the compensated decoupled error. The target control quantities are calculated based on the rapid response of the proportional term, the elimination of steady-state deviation by the integral term, and the suppression of changing trends by the derivative term. Positional PID, incremental PID, PID with anti-saturation integral separation, or a filtered derivative term can be used to transform the decoupled error into an executable control quantity for the actuator, thereby achieving stable and adjustable closed-loop regulation. Decoupling compensation improves the accuracy of PID control in multi-channel coupled scenarios.

[0067] S4: Adjust the corresponding environmental control parameters according to the target control quantities to achieve wind tunnel control.

[0068] Specifically, the internal environment of the wind tunnel is altered through actuators. For example, the temperature of heaters or cooling components is adjusted, the humidity of humidifiers or dehumidifiers is adjusted, the wind speed of fans or airflow guides is adjusted, and the irradiance and radiation environment are adjusted by light sources or radiation panels. The target control quantity is converted into actuator drive signals, forming a complete closed-loop control.

[0069] In one embodiment, the neural network decoupling model adopts a distributed structure, which includes a set of sub-networks. The set of sub-networks includes several independent sub-networks, each of which corresponds to an environmental control parameter. Each of the sub-networks includes a multi-layer feedforward neural network, and each layer of the feedforward neural network uses the hyperbolic tangent activation function.

[0070] Specifically, a distributed structure refers to breaking down the overall decoupled model into multiple relatively independent model units rather than a single large model with unified output. Each channel of an environmental control parameter corresponds to a sub-network that outputs the compensation amount for that channel. Each sub-network is specifically designed to decouple one environmental control parameter. For example, five parameters—temperature, humidity, wind speed, irradiance, and background radiation—correspond to five different sub-networks. Each sub-network independently stores weights and performs independent inference. Multilayer feedforward neural networks are network structures where information propagates in one direction, such as... Figure 3 As shown, each subnetwork can employ a three-layer feedforward neural network, taking preset and measured values ​​as inputs and outputting the compensation amount. Different hidden layer sizes can be configured for different channels to balance computational complexity and fitting capability. The hyperbolic tangent activation function is a non-linear function whose output is between negative and positive values, suitable for expressing compensation in both positive and negative directions. The hyperbolic tangent activation function is as follows:

[0071]

[0072] The activation function expresses the channel coupling relationship through nonlinear mapping capability, while keeping the inference calculation controllable to adapt to real-time control, making the compensation quantity change more continuous and the symbol expression more natural. This helps to reduce the control shock caused by compensation abrupt changes and improve the smoothness and stability of multi-channel coordinated regulation.

[0073] In one embodiment, before S2, the following steps are included:

[0074] S01: Acquire several sets of raw data, wherein each set of raw data includes the sample preset value of environmental control parameters, the sample measured value, and the ideal control quantity when controlling the wind tunnel to a steady state;

[0075] Specifically, the raw data is the sample set used for subsequent training of the neural network. Each set of data contains at least preset values, measured values, and ideal control quantities. The ideal control quantity can be understood as the target drive quantity that each actuator should output when a steady-state condition is reached, such as the heating power should be maintained at a certain level under a certain temperature setting. Data is collected at multiple operating points, covering different seasons, different loads, and different external disturbances. Under several different operating conditions (preferably 50 or more), experienced operators manually and finely adjust the wind tunnel to a steady state, recording the measured value vector under each operating condition. and preset value vector and ideal control vector Multiple training samples were obtained, aiming to collect at least 500 sets. Orthogonal experimental design was used for operating condition design, covering more than 70% of the parameter range. Abundant training samples provide a correlation between input and expected output for decoupled model training, enabling the model to learn reasonable compensation and control tendencies under different settings and measurement conditions. The more comprehensive the sample coverage, the more robust the model will be to disturbances during actual wind tunnel operation, reducing the probability of compensation failure after deployment.

[0076] S02: Normalize the original data to obtain preset training samples;

[0077] Specifically, normalization is the process of scaling data with different dimensions and numerical ranges to a unified range. For example, different environmental control parameters such as temperature, humidity, and wind speed are mapped to 0 to 1 or -1 to 1 according to their upper and lower limits, avoiding the dominance of a single large numerical dimension in training. Min-max normalization, mean-variance standardization, or piecewise normalization can be used, and the normalization parameters can be saved.

[0078] S03: Randomly initialize the weights of each subnetwork in the neural network to obtain the initialized neural network;

[0079] Specifically, weights are the internal connection parameters of a neural network, used... express, ~U[-0.1,0.1], where weights determine how the input is mapped to the output. Random initialization is the process of assigning initial values ​​to the weights before training to break the symmetry. Initialization can be done using a uniform distribution, a normal distribution, or an initialization strategy better suited to the activation function.

[0080] S04: Obtain the set of subnetworks and the target subnetwork in the set of subnetworks;

[0081] Specifically, the subnetwork set is a collection of subnetworks corresponding to all preset environmental control parameters. The target subnetwork is the subnetwork currently selected as the training object; for example, the temperature subnetwork is trained first, followed by the humidity subnetwork. The target subnetwork can be specified in a predetermined order, or key channels can be trained preferentially based on coupling strength, or rotated according to control priority. A cyclic rotation approach can be used, or the training frequency of channels with stronger coupling can be increased. Establishing an alternating training organization method makes the training process controllable and reproducible, and clearly defines the optimization object for each training session, reducing the randomness and uncertainty of the training process.

[0082] S05: Using minimizing the pre-built objective function as the training objective, adjust the weights of each sub-network in the initialized neural network to obtain the weights of the trained sub-networks;

[0083] Specifically, the objective function is an evaluation metric for training optimization, used to measure the degree to which the output of the current target sub-network causes changes in the output of other channels, and the changes in the output of other environmental control parameter channels for non-target sub-networks. The changes in the output of each channel are calculated for the current sample, and the changes except for the target channel are summarized to form the objective function value. The objective function can be expressed by the following formula:

[0084]

[0085] in, Let be the output change of the q-th channel under the s-th sample; in subsequent processes, the backpropagation algorithm and the gradient descent method with the variable term are used to achieve the objective function. This guides the weight update of the p-th subnetwork. The decoupling training objective is transformed into an optimizable numerical metric, clearly defining the training direction as reducing the coupling effect between channels. Unlike existing techniques that only consider single-channel fitting, this approach suppresses interference with other channels, improving decoupling effectiveness. This enables the target subnetwork to gradually learn to compensate for its own channel without disturbing other channels, reducing coupling perturbations during multi-parameter collaborative adjustment.

[0086] S06: Obtain a non-target subnetwork from the subnetwork set in a preset order as a new target subnetwork, return to the step of minimizing the pre-constructed objective function as the training objective, adjust the weights of each subnetwork in the initialized neural network, and obtain the weights of the trained subnetworks, until the preset iterative training stopping condition is met, and obtain the weights of all trained subnetworks.

[0087] Specifically, the new target subnetwork is the channel subnetwork to be optimized in the next round. Returning to training means repeating the same decoupled training process for different environmental control parameter channels. After completing one round of training for all environmental control parameters, it begins to determine whether the preset objective function value requirement or the preset number of iterations is met, such as the objective function value being less than a preset threshold or the preset number of iterations reaching a preset number of training rounds. If all objective function values ​​are less than the preset threshold or the preset number of iterations reaches the preset number of training rounds, training stops, and the weights of all trained subnetworks are obtained. If there is an objective function value greater than or equal to the preset threshold but the preset number of training rounds has not been reached, the next round of training continues, again after completing one round of training for all environmental control parameters, and then it begins to determine whether the objective function threshold is less than the preset threshold or whether the preset number of training rounds has been reached. Through iterative training, the output, objective function and weights are repeatedly calculated and updated under the drive of training samples. The preset threshold is the stopping condition. For example, if the objective function is less than the preset threshold, the interference is considered to be small enough. The preset number of training rounds is the maximum upper limit of the number of iterations, which can be selected in turn according to the rotation order, such as temperature, then humidity and then wind speed. Channels that have not yet met the standard can also be dynamically selected according to the training convergence. Through alternating training, a decoupled model that can be used for all channels is formed, so that the system can still maintain stable control of multiple parameters at the same time under complex working conditions.

[0088] S07: Based on the weights of all trained sub-networks, obtain the pre-trained neural network decoupling model.

[0089] Specifically, the pre-trained neural network decoupling model is an overall model composed of all trained sub-networks. The neural network is configured according to the weights of all trained sub-networks to obtain a deployable decoupling computation model.

[0090] In one embodiment, S06 includes:

[0091] S061: Obtain a non-target subnetwork from the subnetwork set as a new target subnetwork, take minimizing the objective function as the training objective, adjust the weights of each subnetwork in the initialized neural network, and obtain the weights of the trained subnetwork.

[0092] Specifically, a previously unselected subnetwork is chosen from a set of multiple subnetworks as the new training target. This subnetwork is then trained using backpropagation algorithms to minimize the objective function, adjusting the network's weights and biases. Subsequently, each subnetwork is trained alternately, effectively optimizing the decoupling compensation for each control parameter. This alternating training method ensures that each subnetwork optimizes independently, unaffected by other channels, thus improving the decoupling effect.

[0093] S062: When all sub-networks corresponding to each environmental control parameter have completed one round of training, determine whether the objective function value of each sub-network is less than the preset function threshold.

[0094] Specifically, after all subnetworks have completed one round of training, the objective function value of each subnetwork is checked. If the objective function value of any subnetwork is still greater than a preset threshold, it indicates that further optimization is needed. If the objective function values ​​of all subnetworks are less than the preset threshold, then the training of all subnetworks can be considered complete, and the next step can proceed. According to the definition of the objective function, the objective function value is calculated based on the changes in other channels. This objective function value provides a quantitative basis for whether to continue training, allowing the training process to determine whether the decoupling objective has been achieved based on the value.

[0095] S063: When there is an objective function value greater than or equal to a preset function threshold and the number of iterations is less than the preset number of training iterations, return to the step of obtaining a non-target subnetwork from the subnetwork set as a new target subnetwork, with minimizing the objective function as the training objective, adjusting the weights of each subnetwork in the initialized neural network, and obtaining the weights of the trained subnetworks, until the objective function values ​​corresponding to all subnetworks are less than the preset threshold or the number of iterations is equal to the preset number of training iterations, and then obtain the weights of all trained subnetworks.

[0096] Specifically, the preset threshold is the standard for determining whether the training has met the target. For example, setting it to 0.001 or other empirical values ​​indicates that the interference to other channels is sufficiently small. After each round of training, the objective function can be re-evaluated under the new weights. The current objective function value is compared with the threshold to determine whether to continue updating. This process of weight updates and objective function evaluation iterates until the stopping condition is met, resulting in the weights of all trained sub-networks.

[0097] In one embodiment, such as Figure 2 As shown, after the first round of training is completed, S061 includes:

[0098] S0611: Based on the objective function and the error backpropagation algorithm, calculate the gradient information of the environmental control parameters relative to each network weight to obtain the weight correction amount;

[0099] Specifically, the backpropagation algorithm is used to calculate the gradient of the objective function with respect to the network weights. Gradient descent of the magnitude term is then used to overlay historical directions during updates to accelerate convergence and suppress oscillations. The updated weights are the new network connection parameters. In the training context, the gradient information is the partial derivative of the objective function with respect to the network weights. The weight adjustment is the update direction and magnitude calculated based on the gradient and learning rate. Backpropagation is performed on the target subnetwork, propagating the output layer error signal layer by layer back to the connection points between the hidden and input layers to obtain the gradient of each connection weight. By decoupling objective function The calculation shows that the hidden layer error is derived in reverse from the output layer error; It is the first The output of a layer neuron reflects the neuron's response to the first layer. The degree of influence of layer neurons, and then through the learning rate Scaling yields the weight adjustment for this iteration. Learning rate The step size coefficient for weight updates, ranging from 0.01 to 0.1, controls the magnitude of each weight adjustment. In one embodiment, 0.05 is preferred to avoid training oscillations or slow convergence. The gradient of the current round is calculated first, then combined with the weight changes of the previous round to form a momentum term. This yields the weight update amount for the current round, which is applied to the connection weights of the target subnetwork. The current learning direction is merged with the historical inertial direction to form a more reasonable update step size and direction, improving training convergence speed and stability, and reducing training inefficiency caused by repeated changes in gradient direction. Balancing response speed and stability makes training less prone to local oscillations and improves training success rate. The weight update direction is determined by the objective function. Subsequent updates to the weights of different subnetworks optimize the objective function, reducing interference between different parameters through training and improving decoupling capabilities from a mechanistic perspective.

[0100] S0612: Obtain the weight change of the environmental control parameter weights in the previous iteration round, and perform weighting processing on the weight change according to the preset momentum factor to obtain the historical weight correction amount;

[0101] Specifically, the weight change in the previous iteration is the result of the previous update, and the momentum factor α is used to preserve the historical update direction. To obtain the increment of each connection weight in the previous iteration: Multiply by momentum factor The historical weight correction amount was then obtained: Momentum factor The value can be set within the range of 0.5 to 0.9, and can also be dynamically adjusted with each training epoch to balance speed and stability. The historical weight correction utilizes historical update directions to smooth the training path, improve convergence speed, reduce fluctuations during training, make the objective function decrease more stably, and reduce the number of training epochs required.

[0102] S0613: The weight correction amount and the historical weight correction amount are weighted and calculated to obtain the weight update amount;

[0103] S0614: Based on the weight update amount, perform weight adjustment on the sub-network corresponding to the environmental control parameters to obtain the weights of the trained sub-network.

[0104] Specifically, the weight update is the final increment applied to the network weights, obtained by combining the current gradient term and the momentum term. The weight update is calculated by adding the current weight adjustment to the historical adjustment or by weighting them proportionally. For example, directly summing the current weight adjustment to the historical adjustment gives the weight update: Subsequently, based on the weight update amount, the weights of the sub-networks corresponding to the environmental control parameters are updated to obtain the weights of the trained sub-networks.

[0105] Specifically, the weights are updated item by item according to the formula: (The formula is:)

[0106]

[0107] Among them, the old weight This serves as the baseline for weight updates, with all adjustments based on the old weights. Only the weights of the target sub-network are updated, while the weights of the remaining sub-networks remain unchanged, in accordance with the strategy of alternating iterative training. By updating the parameters, the mapping relationship between the input and the compensation is changed, causing the objective function to tend to be smaller in the next round of evaluation. This continuously weakens the coupling perturbation of the target channel adjustment on other channels, thereby improving the control effectiveness of the overall decoupled model.

[0108] In one embodiment, such as Figure 3 As shown, the multilayer feedforward neural network includes an input layer, a hidden layer, and an output layer, and S2 includes:

[0109] S21: Group the measured values ​​and the preset values ​​according to channels to obtain channel input data;

[0110] Specifically, channel grouping pairs the preset value and measured value of each environmental control parameter to form a group of inputs. For example, the temperature channel input consists of the preset temperature value and the measured temperature value, and the humidity channel input consists of the preset humidity value and the measured humidity value. When there are five environmental control parameters—temperature, humidity, wind speed, solar radiation, and background radiation—the input is based on the current measured value. and preset values Construct sub-network inputs for modules corresponding to different environmental control parameters. , For the input of the i-th channel , All data are grouped according to environmental control parameters and converted into array or vector form as input to different sub-networks. This avoids data from different channels from being mixed and causing compensation errors, improves the determinism and consistency of inference input, and makes the compensation output of each sub-network more stable and reliable.

[0111] S22: Input the channel input data into the neural network decoupling model, and normalize it through the input layer to obtain the normalized preset value and the measured value;

[0112] Specifically, the normalization process in the input layer maps the channel input data to a uniform range according to the scaling rules used during training. The normalized preset values ​​and measured values ​​are the standardized inputs used for internal calculations within the decoupled neural network model. The same processing is performed on the online input using the normalization parameters saved during the training phase to ensure consistency between training and inference, avoid distortion of compensation due to changes in input scale, improve the model's generalization stability, and reduce compensation fluctuations caused by differences in units.

[0113] S23: In the hidden layer, the normalized preset value and the measured value are weighted according to the preset bias parameter to obtain the weighted input value of the hidden layer;

[0114] Specifically, the bias parameter is the baseline term of the neuron, the weighting calculation is a linear combination of the inputs using connection weights, and the weighted input values ​​of the hidden layers are intermediate results before activation function processing. The weighted input of each hidden layer neuron is obtained by summing the products of its input weights and adding the bias. For example, matrix multiplication can be used to accelerate the calculation, or fixed-point arithmetic can be adapted to embedded controllers. The following formula can be used for calculation:

[0115]

[0116] in, This represents the weighted sum of inputs to the j-th neuron in the hidden layer. This represents the weights from the first neuron in the input layer to the j-th neuron in the hidden layer. This represents the weights from the second neuron in the input layer to the j-th neuron in the hidden layer. This represents the bias term of the j-th neuron in the hidden layer. The processing in the hidden layer takes into account the nonlinear relationship caused by channel coupling, so that the compensation amount is no longer limited to linear correction, thus improving the decoupling accuracy.

[0117] S24: Substitute the weighted input value into the preset activation function to obtain the activation output result;

[0118] Specifically, the activation function is a non-linear mapping function, and the activation output is the output feature of the hidden layer neurons. The activation function is applied to the weighted input values ​​of each hidden layer to obtain the output, which is then used as the input for subsequent output layer calculations. Numerical stabilization processing can be performed on the activation output, such as pruning extreme inputs. Preset activation functions can include:

[0119]

[0120] After activation function processing, the network can fit complex coupling relationships rather than just linear superposition, improving the model's expressive power, making the decoupling compensation amount closer to the nonlinear coupling characteristics of the real system, and improving the control effect.

[0121] S25: In the output layer, the activated output result is subjected to inverse normalization mapping processing according to the preset compensation value range to obtain the final decoupling compensation amount.

[0122] Specifically, the denormalization mapping converts the normalized output within the network into the actual unit output. After obtaining the normalized output, it is then mapped according to a preset compensation range to obtain the final decoupling compensation amount.

[0123] The following formula can be used to perform inverse normalization mapping:

[0124]

[0125] in, For decoupling compensation amount, It refers to the maximum value within the target interval. This refers to the maximum value within the target interval. After inverse normalization mapping, a compensation quantity is generated that can directly participate in error correction, achieving decoupling and matching the actual situation. Furthermore, to match the response characteristics of the execution system, amplitude limiting or smoothing filtering can be applied after mapping.

[0126] In one embodiment, S3 includes:

[0127] S31: Calculate the compensated current decoupling error based on the preset value, the measured value, and the decoupling compensation amount;

[0128] Specifically, the decoupling error is the error after introducing decoupling compensation based on the difference between the preset value and the measured value. The decoupling error is calculated using the following formula. : ,in, As a preset value, For measured values, This is the decoupling compensation amount. The decoupling error, taking into account the coupling effect, can more accurately reflect the actual deviation that needs to be adjusted in this channel, reducing error offset caused by disturbances in other channels, and improving the targeting and stability of the control response.

[0129] S32: Obtain historical decoupling error and historical control quantity;

[0130] Specifically, the historical decoupling error refers to the error in the previous several cycles; the historical control quantity refers to the historical value of the control commands that have been output to the actuator in the previous several cycles. Number the channels (e.g., Corresponding temperature, Corresponding humidity, Corresponding to wind speed, Corresponding to solar radiation (corresponding to the sky background radiation), in calculating the first Decoupling error of the channel in the (k-1)th cycle At that time, the historical decoupling errors that need to be obtained include the first... Decoupling error of the channel in the (k-1)th cycle The decoupling error of the (k-2)th period Historical control quantities include the control quantities for the (k-1)th period. Introducing historical control values ​​into PID control calculations enables the system to eliminate steady-state deviations and predict trends, improving control continuity and reducing control jumps and steady-state drift.

[0131] S33: Based on the current decoupling error and the historical decoupling error, perform PID control calculations to obtain the control quantity increment;

[0132] Specifically, based on the historical decoupling error and historical control quantity obtained in the previous steps, PID control calculations are performed. The PID controllers corresponding to each environmental control parameter operate independently, using an incremental PID algorithm. The control cycle can be 1-5 seconds. The incremental control quantity is calculated using the following formula. :

[0133]

[0134] in, The scaling factor for the i-th channel is... Let be the integral coefficient of the i-th channel. Let be the differential coefficient of the i-th channel. The cumulative error of the integral term changes, and the trend of the differential term is estimated using error difference estimation. Further anti-saturation processing can be applied to the integral term, and low-pass filtering can be applied to the differential term to suppress sensor noise, reduce overshoot and oscillation caused by sudden changes in control output, and improve the efficiency of reaching a steady state.

[0135] S34: Calculate the target control quantity based on the control quantity increment and the historical control quantity.

[0136] Specifically, the target control quantity is the final control quantity output to the actuator in this cycle, obtained by adding the increment to the historical control quantity. The increment is then added to the control quantity of the previous cycle, according to the formula... Calculate the target control quantity for the i-th channel. Furthermore, the results can be limited, rate-limited, or subject to safety constraints. Different limiting strategies can be set for different actuators, such as limiting heating power or fan speed, to ensure system safety. The calculation results are converted into actual output commands to ensure closed-loop execution of the control link, making the control output continuously controllable and improving the long-term stable operation capability of the system.

[0137] In one embodiment, the method further includes:

[0138] S51: Obtain the fluctuation range of the measured values ​​of each environmental control parameter;

[0139] Specifically, fluctuation amplitude is an indicator that measures the range of change of a parameter over a period of time, such as the difference or standard deviation between the maximum and minimum values ​​within several control cycles. The fluctuation amplitude of environmental control parameters such as temperature, humidity, and wind speed can be statistically analyzed within a sliding window. The window length can be set according to the parameter inertia, for example, a longer window for temperature and a shorter window for wind speed. Furthermore, channels with high sensor noise can be filtered before statistical analysis.

[0140] S52: When the fluctuation amplitude is less than the preset fluctuation threshold within a consecutive preset number of control cycles, the wind tunnel is determined to be in a stable state.

[0141] Specifically, the fluctuation range of various environmental control parameters is used to determine whether the wind tunnel has entered a stable operating condition. An objective stability determination mechanism is established, reliable samples are collected, and the number of control cycles and the fluctuation threshold are used as stability determination conditions. For example, if the fluctuation range is less than 1% of the range for 10 consecutive cycles, the wind tunnel is determined to be in a stable state. This avoids accidental stability misjudgment, improves the accuracy of stable operating condition identification, and reduces the risk of model drift caused by accidental fine-tuning.

[0142] S53: When the wind tunnel is in a stable state, the preset value vector, the measured value vector, and the ideal control vector when controlling the wind tunnel to a stable state are used as the current sample, and the Euclidean distance between the current sample and the adjacent training samples is calculated.

[0143] Specifically, Euclidean distance is used to characterize the similarity between samples. For example, preset values, measured values, and ideal control variables can be transformed into vectors on the same scale to calculate the distance with the latest samples in the training sample library; or only preset values ​​and measured values ​​can be transformed into vectors on the same scale to calculate the distance with the latest samples in the training sample library. Euclidean distance is used to determine the difference between the current sample and existing samples. If there is a difference, samples under already covered conditions are avoided, improving online update efficiency.

[0144] S54: When the Euclidean distance is greater than a preset distance threshold, the current sample is output to a preset target sample set;

[0145] Specifically, the distance threshold is a criterion for determining the magnitude of the difference between samples. Exceeding the threshold indicates a significant difference between the current sample and existing training samples. When the Euclidean distance of the current sample is greater than the preset distance threshold, the sample is stored together with its timestamp, channel label, and stability determination result, and can be categorized and summarized by channel. The current sample is only saved when the Euclidean distance is greater than the preset distance threshold, thus filtering out the samples that are truly needed to update the model and avoiding a large number of similar samples occupying storage space.

[0146] S55: When the preset fine-tuning conditions are met, the sub-network is adjusted according to the target sample set to obtain the adjusted pre-trained neural network decoupling model. The preset fine-tuning conditions include the number of samples in the target sample set being equal to a preset sample number threshold or the time interval from the last fine-tuning being equal to a preset fine-tuning time threshold.

[0147] Specifically, when preset fine-tuning conditions are met, such as when the number of new samples in a certain channel reaches 20 or more than 24 hours have passed since the last fine-tuning, incremental training or parameter updates are performed on the sub-network. This allows the decoupling model to adaptively update as operating conditions drift and equipment ages, enabling the model to adapt to new operating conditions, improving the system's long-term robustness and operating condition coverage, and reducing decoupling failures and control quality degradation caused by environmental changes.

[0148] In one embodiment, the method further includes:

[0149] S07: Determine the training order of each sub-network based on the coupling strength between each environmental control parameter and / or the preset control priority;

[0150] Specifically, coupling strength is an indicator of the degree of mutual influence between channels. For example, a large crosstalk between humidity control and temperature indicates strong coupling. Control priority is the order determined by operating conditions or safety requirements, such as prioritizing temperature control. Coupling strength can be statistically analyzed using historical data, ranked based on engineering experience, or obtained through sensitivity analysis. Channels with the strongest coupling are trained first to suppress major interference sources. For example, when environmental control parameters include temperature, humidity, wind speed, solar radiation, and background radiation, determining the training order of each sub-network based on coupling strength involves weighted summation or averaging of the coupling strength between the current channel and other channels to obtain a coupling score. The training order is then determined by ranking the coupling scores from highest to lowest. If the temperature channel has high coupling strength with the humidity and wind speed channels, followed by the humidity channel with lower coupling strength, and the radiation-related channels have relatively weaker coupling with other channels, then the training order can be determined as: temperature channel, humidity channel, wind speed channel, solar radiation illuminance channel, and background radiation channel. The ordered sequence is: temperature channel, humidity channel, wind speed channel, solar radiation illuminance channel, and background radiation channel. When determining the training order of subnetworks in a neural network based on control priorities, for example, if the current experiment has the highest requirement for temperature stability and temperature exceeding the limit poses a safety risk, the preset control priorities can be set as follows: temperature channel highest, wind speed channel second, humidity channel next, solar irradiance channel and sky background radiation channel lower. Based on this, the training order is determined as: temperature channel, wind speed channel, humidity channel, solar irradiance channel, sky background radiation channel. By allowing the subnetworks corresponding to high-priority channels to enter the stable training phase earlier, the decoupling compensation of key parameters can be made available first, making the alternating training process more in line with actual control needs and improving overall convergence efficiency.

[0151] When determining the training order of subnetworks in a neural network based on the coupling strength between various environmental control parameter channels and the preset control priority, a comprehensive ranking index can be calculated first based on coupling strength scoring and then adjusted according to priority to obtain the final training order. Alternatively, a main ranking framework can be determined first based on control priority, and then specific rankings can be performed within the same priority level according to coupling strength. Taking temperature, humidity, wind speed, solar irradiance, and sky background radiation as examples, if the preset control priority requires temperature to be the highest, followed by wind speed, and the others to be lower, and the evaluation shows that the coupling strength between humidity and temperature and wind speed is significantly higher than that of radiation-related channels, then the training order can be comprehensively determined as temperature channel, wind speed channel, humidity channel, solar irradiance channel, and sky background radiation channel. Determining the training order of subnetworks makes alternating training more efficient, reduces the number of training rounds and training time, and improves the overall performance of the final decoupled model.

[0152] S08: According to the training order, the sub-networks corresponding to each environmental control parameter are trained alternately to obtain a pre-trained neural network decoupling model.

[0153] Specifically, alternating training refers to sequentially selecting the target sub-network for training, with each training iteration aiming to minimize changes in the output of other channels, gradually obtaining a fully trained model across all channels. During training, the weights of non-target sub-networks remain unchanged, while only the weights of the target sub-network are updated. After the update is complete, the system switches to the next sub-network to continue. This systematic approach to achieving full-channel decoupling under a distributed structure avoids global coupling residues caused by single-channel optimization, stabilizes output compensation, and improves the collaborative control performance of multi-parameter wind tunnels.

[0154] In one embodiment, S07 includes:

[0155] S071: Obtain preset training samples;

[0156] Specifically, after normalizing several sets of raw data, a preset training sample is obtained. Each set of raw data includes the preset values ​​of environmental control parameters, the measured values, and the ideal control quantities for controlling the wind tunnel to a steady state. Based on the raw data collected by the wind tunnel system under different operating conditions, the raw data undergoes scaling (such as min-max normalization and mean-variance standardization) to obtain a standardized training sample set. During the standardization process, the normalization coefficients of each parameter must be recorded to ensure the consistency and traceability of data processing. Normalization eliminates the differences in dimensions and numerical ranges of different environmental control parameters.

[0157] S072: Based on the preset training samples, construct data pairs of environmental control parameters, wherein each data pair includes data corresponding to two different environmental control parameters;

[0158] Specifically, a data pair refers to a pair of data units formed by combining complete data sequences of two different environmental control parameters selected from a pre-set training sample. Examples include temperature and humidity data pairs, and wind speed and irradiance data pairs. Each data pair contains all sample data for both parameters under the same operating conditions and at the same time stamp. For all environmental control parameters (such as temperature, humidity, wind speed, irradiance, and background radiation) included in the training sample, an unordered pairing rule is applied to combine all parameters pairwise, ensuring that all combination relationships between parameters are covered without repetition or omission. For example, when there are 5 parameters, a pairwise pairing rule needs to be constructed. There are 10 data pairs. These pairs are then further broken down to obtain the local coupling relationships between each pair of parameters, thus providing a reference for the training order. By constructing these data pairs, the complex coupling relationships between multiple parameters are structurally decomposed, providing a foundation for subsequent analysis of the coupling strength between environmental control parameters.

[0159] S073: Calculate the Pearson correlation coefficient between each pair of environmental control parameters based on the Pearson correlation coefficient calculation formula.

[0160] Specifically, the Pearson correlation coefficient is a statistical indicator used to measure the degree of linear correlation between two variables. Its value ranges from -1 to 1, where a value closer to 1 or -1 indicates a stronger linear correlation, and a value closer to 0 indicates a weaker linear correlation. For example, a Pearson correlation coefficient of 0.9 indicates a strong positive linear correlation, while a value of 0.1 indicates an extremely weak linear correlation. For each pair of environmental control parameter data, standardized sample sequences of the two parameters are obtained. Based on the Pearson correlation coefficient calculation formula, the covariance of the numerator and the standard deviation of the denominator are calculated by traversing the sample data to obtain the Pearson correlation coefficient for the two parameters. Furthermore, after calculation, the coefficient is tested for significance to eliminate spurious correlation results caused by accidental sample distribution. The Pearson correlation coefficient accurately reflects the linear coupling relationship between each pair of environmental control parameters, providing a core indicator of the linear dimension for subsequent comprehensive coupling strength calculations and providing a clear basis for identifying strongly linearly coupled parameter pairs (such as temperature and humidity).

[0161] S074: Calculate the physical influence coefficient between each pair of environmental control parameters based on the data and proportional relationship between the environmental control parameters.

[0162] Specifically, the physical influence coefficient refers to the nonlinear influence of a change in one parameter on another, quantified by a preset mathematical model based on the physical mechanism of wind tunnel system operation. Its value range is [0, +∞). In the selected environmental control parameter data pair, all other preset values ​​of environmental control parameters are fixed, and only the target parameter is changed. The preset values ​​are perturbed at preset steps (e.g., temperature ±5℃, wind speed ±2m / s), and other parameters before and after the perturbation are recorded. Change in measured value Data pairs, parameters calculated according to a preset mathematical model For parameters The physical influence coefficient is calculated using the following formula:

[0163]

[0164] The physical influence coefficient overcomes the limitation of the Pearson correlation coefficient, which can only reflect linear relationships, by quantifying the nonlinear coupling relationship between parameters based on physical mechanisms, and provides a key indicator of the nonlinear dimension for calculating the comprehensive coupling strength.

[0165] S075: Calculate the coupling strength value of each pair of environmental control parameters by weighted averaging the Pearson correlation coefficient and the physical influence coefficient;

[0166] Specifically, the coupling strength value is a quantitative indicator used to measure the overall coupling degree between two environmental control parameters, calculated by weighting the Pearson correlation coefficient (linear coupling index) and the physical influence coefficient (nonlinear coupling index). The value ranges from [0, +∞), with a larger value indicating more significant overall mutual interference between the two parameters. By integrating linear and nonlinear coupling characteristics, a quantitative indicator that comprehensively reflects the overall coupling degree between the two parameters is obtained, providing a core basis for determining the training order. The weighted averaging calculation method takes into account both linear and nonlinear coupling relationships between parameters, making the evaluation of the coupling strength value more comprehensive and avoiding misjudgments of coupling degree caused by a single indicator. Simultaneously, the differentiated setting of the weight coefficients can adapt to the coupling characteristics of different parameter combinations, further improving the reliability of the coupling strength value.

[0167] S076: Based on the coupling strength value, determine the training order of each sub-network corresponding to each environmental control parameter.

[0168] Specifically, the training order refers to the order in which the neural network subnetworks are trained, based on the coupling strength values ​​between each environmental control parameter and other parameters. Priority is given to training the subnetwork corresponding to the parameter with the highest total coupling strength with other parameters. The sum of the coupling strength values ​​between each environmental control parameter and all other parameters is calculated separately. For example, the coupling strength values ​​between temperature and humidity, wind speed, and irradiance are 0.8, 0.7, and 0.6 respectively, resulting in a total coupling strength of 2.1. All environmental control parameters are sorted from largest to smallest according to their total coupling strength values ​​to obtain the training order. Furthermore, the training order can be modified by considering control priorities. For instance, in aerospace experiments, wind speed has a higher control priority than background radiation. Based on the overall coupling degree between parameters and the control priorities, the training order of subnetworks can be rationally arranged to improve model training efficiency and decoupling effect.

[0169] In one embodiment, S076 includes:

[0170] S0761: According to the preset control priority, classify and sort each environmental control parameter by priority to obtain the preliminary training order of each environmental control parameter subset;

[0171] Specifically, environmental control parameters with the same control priority are divided into the same subset. Based on a preset control priority, for example, the subset of environmental control parameters with higher control priority is moved forward before the subset with lower control priority, resulting in the initial training order for each subset. This hierarchical approach ensures that high-priority channels are trained first, preventing the training of low-priority channels from interfering with their training. Placing high-priority channels first allows for the optimization of key control parameters in the early stages of training, prioritizing the decoupling and optimization of these parameters.

[0172] S0762: Sort the coupling strength values ​​to obtain the coupling strength order;

[0173] Specifically, the coupling strength order is a sorting of the magnitude of the coupling strength, with environmental control parameters appearing earlier in the coupling strength order as the coupling strength is greater.

[0174] S0763: According to the coupling strength order, sort the environmental control parameters in each of the environmental control parameter subsets to obtain the target training order within each of the environmental control parameter subsets;

[0175] Specifically, the initial training order is to sort the environmental control parameters according to the magnitude of their coupling strength within the subset. The higher the coupling strength of the environmental control parameters, the higher the target training order within the subset of environmental control parameters. Prioritizing the training of channels with higher coupling strength can solve the interference problem between key channels first and provide a decoupling basis for subsequent training.

[0176] S0764: Based on the preliminary training order and the target training order, obtain the training order of each sub-network corresponding to each environmental control parameter.

[0177] Specifically, the initial training order and the target training order jointly determine the training order of the subnetworks corresponding to each environmental control parameter. Although the coupling strength between some control parameters is high, based on control priority, some parameters may still need to be adjusted first. For environmental control parameters with higher control priority, their training order is prioritized; for channels with lower control priority, although their coupling strength may be high, their training order can be appropriately postponed. In this way, a training order that comprehensively considers coupling strength and control priority is finally obtained. For example, if the control priority of temperature is higher than that of humidity, then even if temperature is after humidity in the initial training order, the subnetwork of the temperature channel will be trained first.

[0178] In one embodiment, S07 includes:

[0179] S077: Obtain the preset control priority;

[0180] Specifically, control priority refers to determining the order of importance of various control parameters or control channels in a multi-environment control system based on different control needs or safety requirements. In some application scenarios, certain parameters may be more critical to the stability or safety of the system, thus requiring priority control. The setting of control priorities is usually based on system safety requirements, operating conditions, historical data, and experience. For example, temperature control may be crucial in some industrial processes; if the temperature runs out of control, it may lead to equipment damage or danger. Therefore, temperature control may be given the highest priority. The priority of each environmental control parameter may differ under different operating conditions. For instance, in an agricultural greenhouse control system, temperature and humidity may be the most prioritized parameters for adjustment, while light intensity is relatively secondary. By analyzing historical system operating data and combining it with engineering experience, it is determined which control parameter fluctuations have the greatest impact on system stability, and priorities are set accordingly. In some highly dynamic systems, priorities can also be adjusted based on real-time feedback. For example, in wind tunnel experiments, wind speed changes may be a key influencing factor, with a higher priority.

[0181] S078: Determine the training order of each sub-network according to the preset control priority;

[0182] Specifically, the subnetworks are trained according to their priority from high to low. First, high-priority subnetworks are trained to ensure effective decoupling and compensation of these critical control parameters. Then, the training gradually shifts to lower-priority subnetworks. This method ensures that the system converges first on the most critical control parameters, reducing interference with other important parameters.

[0183] Example 2

[0184] Furthermore, the multi-parameter wind tunnel control method based on neural network decoupling in Embodiment 2 of the present invention is essentially a general decoupling and collaborative control method for multi-parameter strongly coupled control objects. It can be applied to: other types of environmental simulation equipment, such as temperature and humidity chambers and vibration table composite environmental systems; multi-variable coupled systems in industrial process control, such as chemical reactors and boiler control; intelligent agricultural greenhouse environmental regulation; and multi-parameter coordinated control in building energy management systems.

[0185] In one embodiment, the method further includes:

[0186] Get the current test environment type;

[0187] Based on the current test environment type and the preset test environment type - environment control parameter template, obtain the preset environment control parameters;

[0188] Obtain the measured and preset values ​​of various environmental control parameters within the test environment.

[0189] Specifically, the preset environmental control parameters differ depending on the type of testing environment. For other types of environmental simulation equipment, such as temperature and humidity chambers and temperature-humidity-vibration composite environmental testing systems, the preset environmental control parameters for temperature and humidity chambers include temperature, relative humidity, and air pressure or gas composition (such as oxygen concentration and inert gas ratio). The preset environmental control parameters for temperature-humidity-vibration composite environmental testing systems include temperature, relative humidity, vibration frequency, vibration amplitude or acceleration, and air pressure or gas composition (such as oxygen concentration and inert gas ratio). Temperature and humidity are used to simulate high-temperature, low-temperature, and humid conditions, while vibration frequency and amplitude are used to simulate mechanical stress during transportation and operation. Temperature changes directly affect the accuracy of humidity control, while vibration conditions affect the stability of sensor measurements and the airflow distribution inside the chamber. There is a clear coupling relationship between these parameters, requiring multi-parameter coordinated control and decoupling compensation.

[0190] For multivariable coupled systems in industrial process control, such as chemical reactors, boiler systems, and multi-effect evaporation systems, preset environmental control parameters include temperature, pressure, material flow rate, material concentration or ratio, and liquid level. In chemical and energy systems, reaction temperature, pressure, and material flow rate directly determine the reaction rate and safety. Changes in material concentration affect the intensity of exothermic or endothermic reactions, thus inversely affecting temperature and pressure. These parameters exhibit strong nonlinear coupling characteristics, and single-parameter control can easily cause system oscillations or overshoot. Therefore, multi-parameter decoupling control is necessary to achieve stable, safe, and efficient operation.

[0191] For intelligent agricultural greenhouse environmental control systems, preset environmental control parameters include temperature, humidity, light intensity, carbon dioxide concentration, and soil moisture content. Crop growth is highly sensitive to the environment. Temperature and humidity directly affect crop transpiration and the probability of disease occurrence; light intensity determines photosynthetic efficiency; carbon dioxide concentration affects the photosynthetic rate; and soil moisture content determines the water absorption status of the roots. These multiple environmental parameters are coupled; for example, changing the light intensity will also change the temperature. Therefore, coordinated control and decoupling of these multiple environmental parameters are necessary.

[0192] For multi-parameter coordinated control in building energy management systems, such as those for large public buildings, data centers, and smart parks, preset environmental control parameters include temperature, humidity, ventilation rate, and electrical power. Building energy management needs to strike a balance between comfort and energy consumption. Ventilation rate affects temperature and humidity, and there are obvious coupling and constraint relationships between the parameters, making a general decoupling and coordinated control method suitable for strongly coupled multi-parameter control objects.

[0193] Example 3

[0194] Furthermore, the multi-parameter wind tunnel control method based on neural network decoupling in this embodiment of the invention can be implemented by a system. Figure 4A schematic diagram of the hardware structure of the system provided in an embodiment of the present invention is shown.

[0195] The system may include a processor and a memory storing computer program instructions.

[0196] Specifically, the processor may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0197] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0198] Computer-readable media include both permanent and non-permanent, removable and non-removable media, which can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient media, such as modulated communication signals and carrier waves.

[0199] The processor reads and executes computer program instructions stored in memory to implement any of the multi-parameter wind tunnel control methods based on neural network decoupling in the above embodiments.

[0200] In one example, the system may also include a communication interface and a bus. For example, Figure 4 As shown, the processor 401, memory 402, and communication interface 403 are connected through bus 410 and complete communication with each other.

[0201] The communication interface is mainly used to enable communication between various modules, devices, units and / or equipment in the embodiments of the present invention.

[0202] A bus, including hardware, software, or both, couples components of an electronic device together. For example, and not limitingly, a bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, a bus may include one or more buses. While specific buses are described and illustrated in embodiments of the invention, the invention contemplates any suitable bus or interconnect.

[0203] Example 4

[0204] Furthermore, in conjunction with the multi-parameter wind tunnel control method based on neural network decoupling in the above embodiments, this invention can be implemented using a computer-readable storage medium. This computer-readable storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the multi-parameter wind tunnel control methods based on neural network decoupling in the above embodiments.

[0205] In summary, the multi-parameter wind tunnel control method and system based on neural network decoupling provided in this invention obtains the measured values ​​and preset values ​​of various environmental control parameters within the wind tunnel; inputs the measured values ​​and preset values ​​into a pre-trained neural network decoupling model to obtain the decoupling compensation amount corresponding to each environmental control parameter; performs PID control calculation based on the measured values, preset values, and decoupling compensation amounts to obtain each target control amount; and adjusts the corresponding environmental control parameters according to each target control amount to achieve wind tunnel control. This invention compensates for the coupling effects of various environmental control parameters in a multi-parameter wind tunnel system through a distributed neural network decoupling structure, reducing overshoot and oscillations caused by mutual pulling between channels, while also reducing reliance on manual adjustment. It can effectively suppress mutual interference between different environmental control parameters and adapt to the nonlinear, time-varying, and frequent switching characteristics of wind tunnel systems. Therefore, it can maintain high control accuracy and stability even when multiple parameters are adjusted simultaneously, improving the reliability and applicability of the system under different experimental conditions.

[0206] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0207] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0208] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0209] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0210] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0211] It should also be noted that the exemplary embodiments mentioned in this invention describe methods or systems based on a series of steps or apparatus. However, this invention is not limited to the order of the steps described above; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0212] The above description is merely a specific embodiment of the present invention. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the protection scope of the present invention.

Claims

1. A multi-parameter wind tunnel control method based on neural network decoupling, characterized in that, include: Obtain the measured and preset values ​​of various environmental control parameters preset in the wind tunnel; The measured values ​​and the preset values ​​are input into a pre-trained neural network decoupling model to obtain the decoupling compensation amount corresponding to each environmental control parameter; Based on the measured value, the preset value, and the decoupling compensation amount, PID control calculation is performed to obtain each target control quantity; Based on the target control quantities, the corresponding environmental control parameters are adjusted to achieve wind tunnel control; The neural network decoupling model adopts a distributed structure. The neural network decoupling model includes a set of sub-networks, which includes several independent sub-networks. Each sub-network corresponds to an environmental control parameter. Each sub-network includes a multi-layer feedforward neural network, and each layer of the feedforward neural network uses the hyperbolic tangent activation function. The multilayer feedforward neural network includes an input layer, a hidden layer, and an output layer. The step of inputting the measured values ​​and the preset values ​​into a pre-trained neural network decoupling model to obtain the decoupling compensation amount for each channel corresponding to the environmental control parameters includes: The measured values ​​and the preset values ​​are grouped by channel to obtain channel input data; The channel input data is input into the neural network decoupling model, and normalized through the input layer to obtain the normalized preset value and the measured value. In the hidden layer, the normalized preset value and the measured value are weighted according to the preset bias parameters to obtain the weighted input value of the hidden layer. Substitute the weighted input values ​​into a preset activation function to obtain the activation output result; In the output layer, the activated output result is subjected to inverse normalization mapping processing according to the preset compensation value range to obtain the final decoupling compensation amount. The step of performing PID control calculations based on the measured values, the preset values, and the decoupling compensation amount to obtain each target control quantity includes: The current decoupling error after compensation is calculated based on the preset value, the measured value, and the decoupling compensation amount. Obtain historical decoupling errors and historical control variables; Based on the current decoupling error and the historical decoupling error, PID control calculations are performed to obtain the control increment; The target control quantity is calculated based on the control quantity increment and the historical control quantity.

2. The method according to claim 1, characterized in that, Before inputting the measured values ​​and the preset values ​​into the pre-trained neural network decoupling model to obtain the decoupling compensation amounts corresponding to each environmental control parameter, the process includes: Acquire several sets of raw data, each set of raw data including sample preset values ​​of environmental control parameters, sample measured values, and ideal control quantities when controlling the wind tunnel to a steady state; The original data is normalized to obtain a preset training sample; The weights of each subnetwork in the neural network are randomly initialized to obtain the initialized neural network. Obtain the set of subnetworks and the target subnetwork within the set of subnetworks; With minimizing the pre-constructed objective function as the training objective, the weights of each sub-network in the initialized neural network are adjusted to obtain the weights of the trained sub-networks. According to the preset order, a non-target subnetwork in the subnetwork set is obtained as a new target subnetwork. Then, the steps of minimizing the pre-constructed objective function as the training objective and adjusting the weights of each subnetwork in the initialized neural network are returned to obtain the weights of the trained subnetworks. This process continues until the preset iterative training stopping condition is met, at which point the weights of all trained subnetworks are obtained. Based on the weights of all trained subnetworks, a pre-trained neural network decoupling model is obtained.

3. The method according to claim 2, characterized in that, The steps of obtaining a non-target subnetwork from the subnetwork set according to a preset order as a new target subnetwork, returning to the step of adjusting the weights of each subnetwork in the initialized neural network with minimizing the pre-constructed objective function as the training objective, and obtaining the trained subnetwork weights, continue until a preset iterative training stopping condition is met, resulting in the weights of all trained subnetworks, including: Obtain a non-target subnetwork from the subnetwork set as a new target subnetwork, and use minimizing the objective function as the training objective. Adjust the weights of each subnetwork in the initialized neural network to obtain the weights of the trained subnetwork. When all sub-networks corresponding to each environmental control parameter have completed one round of training, determine whether the objective function value of each sub-network is less than the preset function threshold. When there is an objective function value greater than or equal to a preset function threshold and the number of iterations is less than the preset number of training iterations, return to the step of obtaining a non-target subnetwork from the subnetwork set as a new target subnetwork, with minimizing the objective function as the training objective, adjusting the weights of each subnetwork in the initialized neural network, and obtaining the weights of the trained subnetworks, until the objective function values ​​corresponding to all subnetworks are less than the preset threshold or the number of iterations is equal to the preset number of training iterations, and then obtain the weights of all trained subnetworks.

4. The method according to claim 3, characterized in that, After completing the first round of training, a non-target subnetwork from the subnetwork set is selected as the new target subnetwork. The training objective is to minimize the objective function. The weights of each subnetwork in the initialized neural network are adjusted to obtain the trained subnetwork weights, including: Based on the objective function and the backpropagation algorithm, the gradient information of the environmental control parameters relative to each network weight is calculated to obtain the weight correction amount; Obtain the weight change of the environmental control parameter weights in the previous iteration round, and weight the weight change according to the preset momentum factor to obtain the historical weight correction amount; The weight update amount is obtained by weighting the weight correction amount with the historical weight correction amount; Based on the weight update amount, the weights of the sub-networks corresponding to the environmental control parameters are adjusted to obtain the weights of the trained sub-networks.

5. The method according to claim 1, characterized in that, The method further includes: Obtain the fluctuation range of the measured values ​​of each environmental control parameter; When the fluctuation amplitude is less than the preset fluctuation threshold within a consecutive preset number of control cycles, the wind tunnel is determined to be in a stable state. When the wind tunnel is in a stable state, the preset value vector, the measured value vector, and the ideal control vector when controlling the wind tunnel to a stable state are used as the current sample, and the Euclidean distance between the current sample and the adjacent training samples is calculated. When the Euclidean distance is greater than a preset distance threshold, the current sample is output to a preset target sample set; When the preset fine-tuning conditions are met, the sub-network is adjusted according to the target sample set to obtain the adjusted pre-trained neural network decoupling model. The preset fine-tuning conditions include the number of samples in the target sample set being equal to a preset sample number threshold or the time interval since the last fine-tuning being equal to a preset fine-tuning time threshold.

6. The method according to any one of claims 1-5, characterized in that, The method further includes: The training order of each sub-network is determined based on the coupling strength between various environmental control parameters and / or the preset control priority. According to the training order, the sub-networks corresponding to each environmental control parameter are trained alternately to obtain a pre-trained neural network decoupling model.

7. A multi-parameter wind tunnel control system based on neural network decoupling, characterized in that, include: At least one processor, at least one memory, and computer program instructions stored in the memory, which, when executed by the processor, implement the method as described in any one of claims 1-6.

Images