A temperature control method based on autoencoder fuzzy neural network

Abstract

The present invention discloses a temperature control method based on an autoencoding fuzzy neural network, which belongs to the field of temperature control technology. The temperature control method comprises: collecting temperature sensor data and performing preprocessing; using an autoencoding neural network to learn and reduce the dimension of the preprocessed temperature data, and generate a low-dimensional representation; using a fuzzy controller to control the low-dimensional representation and control error, and generate a corresponding control signal; and feeding the control signal back to the temperature control system for real-time control. The present invention utilizes an autoencoding neural network to reduce the dimension of the temperature data to generate a low-dimensional representation, and uses a fuzzy controller to control the low-dimensional representation. By controlling the temperature change, the efficiency and accuracy of the temperature control are achieved. At the same time, the method also has the advantages of strong adaptability and good robustness, and can adapt to various complex industrial production and manufacturing environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of temperature control, and in particular relates to a temperature control method based on an autoencoding fuzzy neural network. Background Art

[0002] Current temperature control systems typically use linear control methods like PID. However, these traditional methods have limitations when dealing with nonlinear systems. Traditional PID control methods often rely on precise models and linear assumptions, making them ineffective in addressing complex nonlinear temperature control problems. Furthermore, traditional control methods often require manual parameter adjustment, making real-time optimization difficult for dynamically changing systems. Therefore, in the field of temperature control, a more adaptable and self-adaptive control method is needed.

[0003] In recent years, autoencoder neural networks (ANNs) in the field of deep learning have been widely studied and applied for nonlinear modeling and control tasks. However, existing ANNF-based approaches still have some limitations in temperature control. For example, they may be restricted by issues such as the specific input data distribution, the choice of fuzzy set rules, and the design of the network architecture. To overcome these limitations of existing ANNF-based approaches and improve the accuracy and robustness of temperature control, this paper proposes a temperature control system based on ANNF-based approaches. Summary of the Invention

[0004] To address the shortcomings of the prior art, the present invention provides a temperature control method based on an autoencoder fuzzy neural network. This method combines an autoencoder neural network with fuzzy control methods, introducing fuzzy set rules and an improved network structure. This enables the network to better handle fuzziness and uncertainty, improving the performance of the temperature control system in complex environments.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a temperature control method based on an autoencoder fuzzy neural network, comprising the following steps:

[0006] (1) collecting temperature data to be controlled, preprocessing the temperature data, and obtaining preprocessed data;

[0007] (2) The preprocessed data is learned and dimensionality reduced using an autoencoder neural network to generate a low-dimensional representation;

[0008] (3) Using a fuzzy controller that takes the low-dimensional representation and the control error as input, and then inverts the output control increment to obtain the control signal;

[0009] (4) Feedback the control signal to the temperature control system for real-time control.

[0010] Compared to other neural networks, the autoencoder neural network can be trained without labeled data, making it suitable for large-scale datasets and unlabeled data scenarios. This gives the autoencoder neural network an advantage in tasks such as data preprocessing, feature learning, and data dimensionality reduction. By learning a low-dimensional representation of the input data, it can extract important features of the data, helping to reduce the data dimension and retain key information. Compared to traditional linear models, the autoencoder neural network has stronger nonlinear modeling capabilities. It can learn complex nonlinear relationships through a combination of multiple layers of neurons and activation functions, making it suitable for handling highly nonlinear temperature control problems.

[0011] The autoencoder neural network adopts a multi-layer perceptron architecture, which is expressed as:

[0012]

[0013] Where x is the input data; W1 and b1 are the weight and bias of the autoencoder neural network respectively; f is the Logsig activation function, and f(W1x+b1) expresses the transfer mode of the autoencoder neural network in the previous layer; The output expression of is expressed as the final output form of the neural network; W2 and b2 are the weight and bias of the fuzzy controller respectively;

[0014] The loss function of the autoencoder neural network is defined as:

[0015] L(θ)=L AE (θ)+L FC (θ)

[0016] Among them, L AE (θ) is the reconstruction error of the improved autoencoder neural network, L FC (θ) is the error of the fuzzy controller and θ is the parameter of the improved autoencoder neural network.

[0017] The low-dimensional representation described in the present invention is to represent data using fewer features or dimensions than the original data.

[0018] As a preferred embodiment of the present invention, the error L of the fuzzy controller FC The expression of (θ) is:

[0019]

[0020] Where N is the number of samples, i is the i-th sample, and y i is the fuzzy controller input, is the expected output.

[0021] As a preferred embodiment of the present invention, the hidden layer and the output layer of the autoencoding neural network both contain the Logsig activation function, which is defined as x is the input data.

[0022] In the forward propagation process of the autoencoder neural network, the output of each neuron can be calculated by multiplying the input vector x by the weight matrix W and adding the bias vector b, and then through the activation function f(x). The specific propagation process is shown in the following formula:

[0023]

[0024] Where x is the input of the neuron, y is the output of the neuron, W is the weight of the neuron, b is the bias, and b1 is the bias of the autoencoder neural network.

[0025] As a preferred embodiment of the present invention, in step (2), the pre-processed temperature data is input into the autoencoder neural network, and training is performed by the back propagation algorithm to learn the characteristics of the data and compress the pre-processed temperature data into a low-dimensional representation, which specifically includes the following steps:

[0026] S1: A temperature dataset containing N samples {x1,x2,...,x N}, where each sample x i ={x i1 ,x i2 ,...,x im There are m features. Import the dataset into the autoencoder neural network.

[0027] S2: Use the autoencoder in the autoencoder neural network to reduce the dimensionality of the data and iterate repeatedly to minimize the reconstruction error, where:

[0028] Assume that the number of nodes in the hidden layer z is k l , where both the input layer and the output layer have n l The goal of the autoencoder is to minimize the error between input and output, that is, to minimize the reconstruction error:

[0029]

[0030] Where N represents the number of samples, L(x i ,f(g(x i ))) is the input x i With the output f(g(x i )), g(x i ) represents the transformation from the input layer to the hidden layer, f(g(x i )) represents the transition from the hidden layer to the output layer; is a sparse regularization term used to control the number of activated neurons in the autoencoder neural network; θ is the autoencoder model parameter; L is the number of hidden layers, i is the i-th node in the hidden layer of the current calculation, and j is the j-th node in the input layer and output layer of the current calculation;

[0031] S3: When the convergence condition of minimizing the reconstruction error is reached, the training is completed and the output of the hidden layer of the autoencoder is used as the dimensionality reduction representation of the dataset:

[0032] x i →g(x i )→z i

[0033] where g(x i ) is the transformation from the input layer to the hidden layer, z i is the output of the hidden layer, i.e., the dimensionality reduction representation;

[0034] The convergence condition is that the reconstruction error is less than 1e-6.

[0035] As a preferred embodiment of the present invention, in step (3), the dimensionality reduction representation and control error are decomposed into multiple local models, a fuzzy controller is designed on each local model, and then the outputs of each fuzzy controller are synthesized using a weighted average method to obtain a control increment.

[0036] As a preferred embodiment of the present invention, the fuzzy controller is a TS fuzzy controller.

[0037] The input of the TS fuzzy controller of the present invention includes the dimension reduction representation and the control error, and the output is the control increment, which can be expressed as:

[0038]

[0039] where Δh a is the control increment on the ath local model, w a is the weight of the a-th local model.

[0040] The present invention can be inverted by a PID controller, which is expressed as:

[0041]

[0042] Where u(t) is the output of the PID controller and e(t) is the control error.

[0043] Therefore, the total control increment Δu of the PID controller is expressed as:

[0044]

[0045] Since the total control increment Δu of the PID controller is the result of the output of the TS fuzzy controller, the control error needs to be recalculated according to the actual situation. The specific calculation formula is as follows:

[0046] e(t)=y ref (t)-y(t)

[0047] In summary, the output control signal expression when the TS fuzzy controller is integrated with the PID controller is:

[0048]

[0049] As a preferred embodiment of the present invention, in step (3), a PID controller is used for inversion, and the output control signal expression when the TS fuzzy controller is integrated with the PID controller is:

[0050]

[0051] Where t is the current time, D h a is the control increment on the ath local model of the TS fuzzy controller, w a is the weight of the ath local model, K p , K i , K d are proportional, integral, and differential coefficients respectively; y ref (t) is the set reference output value, y(t) is the actual output value, N is the number of samples, and a is the ath local model.

[0052] As a preferred embodiment of the present invention, in step (1), the preprocessing is to normalize the collected temperature data and map the value range of all input parameter original data to [0, 1]. The normalization formula is:

[0053]

[0054] where x new is the parameter value after normalization, x is the parameter value before normalization, x min is the minimum parameter value in the sample, x max is the maximum parameter value within the sample.

[0055] As a preferred embodiment of the present invention, in step (1), before preprocessing, the collected data is divided into a test set and a training set.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] (1) The present invention utilizes an autoencoder neural network to reduce the dimensionality of temperature data to generate a low-dimensional representation, and uses a fuzzy controller to control the low-dimensional representation and control errors. By controlling temperature changes, the high efficiency and accuracy of temperature control are achieved, and it can adapt to various complex industrial production and manufacturing environments.

[0058] (2) The present invention combines the autoencoding neural network with fuzzy control, so that the autoencoding fuzzy neural network can better adapt to nonlinear temperature control systems, automatically extract the characteristics and laws of the system from the data using a data-driven modeling method, and can adapt to different working conditions and system changes. Ultimately, it can better predict system behavior and optimize control strategies, thereby achieving better control effects. It provides a technical reference for the modeling and control of nonlinear systems in industry, overcomes the limitations of existing technologies, and exhibits better performance in nonlinear temperature control systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a network structure diagram of the autoencoding neural network described in the present invention.

[0060] Figure 2 This is a control flow chart of the temperature control method based on the autoencoder fuzzy neural network of the present invention.

[0061] Figure 3 This is a comparison chart of the final predictive control result and the actual result described in Example 1 of the present invention.

[0062] Figure 4 This is a graph of the predicted mean square error of the temperature compared to the actual temperature of the temperature control method based on the autoencoder fuzzy neural network described in Example 1 of the present invention.

[0063] Figure 5 It is a response curve diagram of the temperature control method based on the autoencoder fuzzy neural network in the present invention and the traditional PID control method. DETAILED DESCRIPTION

[0064] In order to better illustrate the purpose, technical solutions and advantages of the present invention, the present invention will be further described below in conjunction with specific embodiments.

[0065] Example 1

[0066] A temperature control method based on an autoencoding fuzzy neural network comprises the following steps:

[0067] (1) The temperature data to be controlled is collected through the temperature acquisition module, and the temperature data is divided into a test set and a training set. The temperature data is normalized and the value range of all the input parameter original data is mapped to [0, 1]. The normalization formula is:

[0068]

[0069] where x new is the parameter value after normalization, x is the parameter value before normalization, x min is the minimum parameter value in the sample, x max is the maximum parameter value within the sample.

[0070] (2) Figure 1 As shown in the figure, the pre-processed data in the training set is trained using an autoencoder neural network through a back-propagation algorithm to learn the characteristics of the data and compress the pre-processed temperature data into a low-dimensional representation for the subsequent setting of the control rules of the fuzzy controller. The specific steps include:

[0071] S1: A dataset containing N samples {x1,x2,...,x N}, where each sample x i ={x i1 ,x i2 ,...,x im} has m features;

[0072] S2: Use autoencoder to reduce the dimension of data:

[0073] Assume that the number of nodes in the hidden layer z is k, and both the input layer and the output layer have n l The goal of the autoencoder is to minimize the error between input and output, that is, to minimize the reconstruction error:

[0074]

[0075] Where L(x i ,f(g(x i ))) is the input x i With the output f(g(x i )), g(x i ) represents the transformation from the input layer to the hidden layer, f(g(x i )) represents the transition from the hidden layer to the output layer; is a sparse regularization term used to control the number of activated neurons in the autoencoder neural network; θ is the autoencoder model parameter; L is the number of hidden layers, i is the i-th node in the hidden layer of the current calculation, and j is the j-th node in the input layer and output layer of the current calculation.

[0076] S3: When the convergence condition for minimizing the reconstruction error is reached, that is, the training is completed when the reconstruction error is less than 1e-6, and the output of the hidden layer of the autoencoder is used as the dimensionality reduction representation of the data set:

[0077] x i→g(x i )→z i

[0078] where g(x i ) is the transformation from the input layer to the hidden layer, z i is the output of the hidden layer, i.e., the dimensionality reduction representation.

[0079] The autoencoder neural network adopts a multi-layer perceptron architecture, which is expressed as:

[0080]

[0081] Where x is the input data; W1 and b1 are the weight and bias of the autoencoder neural network respectively; f is the Logsig activation function, and f(W1x+b1) expresses the transfer mode of the autoencoder neural network in the previous layer; The output expression of is expressed as the final output form of the neural network; W2 and b2 are the weight and bias of the fuzzy controller respectively;

[0082] The loss function of the autoencoder neural network is defined as:

[0083] L(θ)=L AE (θ)+L FC (θ)

[0084] Among them, L AE (θ) is the reconstruction error of the autoencoder neural network, L FC (θ) is the error of the fuzzy controller and θ is the parameter of the autoencoder neural network.

[0085] (3) First, design a fuzzy controller: Use the TS fuzzy model to construct a fuzzy controller, and use the low-dimensional representation and control error as input. Set the membership function of the fuzzy controller input to a Gaussian membership function, and train the fuzzy controller to construct the optimal fuzzy control rule; decompose the low-dimensional representation and control error into multiple local models, design a fuzzy controller for each local model, and then use the weighted average method to synthesize the outputs of each fuzzy controller to obtain the control increment;

[0086] Then, the output control increment is inverted through the PID controller to obtain the control signal. The output control signal expression when the TS fuzzy controller is integrated with the PID controller is:

[0087]

[0088] Where Δh a is the control increment on the ath local model of the TS fuzzy controller, w a is the weight of the ath local model, K p , Ki , K d They are proportional, integral and differential coefficients respectively. In this example, the original PID parameters are set as K p =0.001, K i =0.2, K d =0.0008;y ref (t) is the set reference output value, y(t) is the actual output value, N is the number of samples, and a is the ath local model.

[0089] (4) Feedback the control signal to the temperature control system for real-time control. Feedback the output signal of the fuzzy controller to the temperature control system for real-time control.

[0090] Depend on Figure 3 and Figure 4 It can be concluded that the temperature control method of the auto-encoding fuzzy neural network for nonlinear systems proposed in this invention can well predict the temperature of the nonlinear model, and Figure 5 From the image, we can see that the control method proposed in the present invention (i.e., the temperature control method of the improved autoencoder fuzzy neural network in the figure) has smaller fluctuations than the traditional PID control scheme (i.e., the PID in the figure) and can reach the expected temperature value more stably. At the same time, compared with the overshoot of 0.88% in the traditional PID, the overshoot of the improved autoencoder fuzzy neural network temperature control method for nonlinear systems proposed in the present invention is only 0.07%. Due to the limitation of the maximum temperature in high-temperature reactions, the smaller overshoot can effectively avoid the safety risks caused by uncontrollable high temperature.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A temperature control method based on autoencoder fuzzy neural network, characterized in that: The steps include: (1) collecting temperature data to be controlled, preprocessing the temperature data, and obtaining preprocessed data; (2) The preprocessed data is learned and dimensionality reduced using an autoencoder neural network to generate a low-dimensional representation; (3) First, a fuzzy controller is used to take the low-dimensional representation and control error as input, decompose the low-dimensional representation and control error into multiple local models, design a TS fuzzy controller on each local model, and then use the weighted average method to synthesize the outputs of each TS fuzzy controller to obtain the control increment; Then, the output control increment is inverted through the PID controller to obtain the control signal. The output control signal expression when the TS fuzzy controller is integrated with the PID controller is: Where Δh a is the control increment on the ath local model of the TS fuzzy controller, w a is the weight of the ath local model, K p , K i , K d are proportional, integral, and differential coefficients respectively; y ref (t) is the set reference output value, y(t) is the actual output value, N is the number of samples, and a is the ath local model; (4) Feedback the control signal to the temperature control system for real-time control; The autoencoder neural network adopts a multi-layer perceptron architecture, which is expressed as follows: Where x is the input data; W1 and b1 are the weight and bias of the autoencoder neural network respectively; f is the Logsig activation function, and f(W1x+b1) expresses the transfer mode of the autoencoder neural network in the previous layer; The output expression of is expressed as the final output form of the neural network; W2 and b2 are the weight and bias of the fuzzy controller respectively; The loss function of the autoencoder neural network is defined as: L(θ)=L AE (θ)+L FC (i) Among them, L AE (θ) is the reconstruction error of the autoencoder neural network, L FC (θ) is the error of the fuzzy controller and θ is the parameter of the autoencoder neural network.

2. The temperature control method based on the autoencoder fuzzy neural network according to claim 1, characterized in that: The specific calculation formula of the control error is as follows: e(t)=y ref (t)-y(t) Among them, y ref (t) is the set reference output value, y(t) is the actual output value, and e(t) is the control error.

3. The temperature control method based on the autoencoder fuzzy neural network according to claim 1, characterized in that: The error L of the fuzzy controller FC The expression of (θ) is: Where N is the number of samples, i is the i-th sample, and y i is the expected output, is the input of the autoencoder neural network.

4. The temperature control method based on the autoencoder fuzzy neural network according to claim 1, characterized in that: Each layer of the autoencoder neural network contains the Logsig activation function, which is defined as x is the input data.

5. The temperature control method based on the autoencoder fuzzy neural network according to claim 1, characterized in that: In the step (2), the pre-processed temperature data is input into the autoencoder neural network, and the training is performed by the back propagation algorithm to learn the characteristics of the data and compress the pre-processed temperature data into a low-dimensional representation, which specifically includes the following steps: S1: A dataset containing N samples {x1,x2,…,x N }, where each sample x i ={x i1 ,x i2 ,…,x im }There are m features; S2: Use autoencoder to reduce the dimension of data: Assume that the number of nodes in the hidden layer z is k l , where both the input layer and the output layer have n l The goal of the autoencoder is to minimize the error between input and output, that is, to minimize the reconstruction error: Where L(x i ,f(g(x i ))) is the input x i With the output f(g(x i )), g(x i ) represents the transformation from the input layer to the hidden layer, f(g(x i )) represents the transition from the hidden layer to the output layer; is a sparse regularization term used to control the number of activations of neurons in the autoencoder neural network; θ is the autoencoder model parameter; L is the number of hidden layers, i is the i-th node in the hidden layer of the current calculation, and j is the j-th node in the input layer and output layer of the current calculation; S3: When the convergence condition of minimizing the reconstruction error is reached, the training is completed and the output of the hidden layer of the autoencoder is used as the dimensionality reduction representation of the dataset: x i →g(x i )→z i where g(x i ) is the transformation from the input layer to the hidden layer, z i is the output of the hidden layer, i.e., the dimensionality reduction representation.

6. The temperature control method based on the autoencoder fuzzy neural network according to claim 1, characterized in that: In step (1), the preprocessing is to normalize the collected temperature data and map the value range of all input parameter original data to [0,1]. The normalization formula is: where x new is the parameter value after normalization, x is the parameter value before normalization, x min is the minimum parameter value in the sample, x max is the maximum parameter value within the sample.

7. The temperature control method based on the autoencoder fuzzy neural network according to claim 1, characterized in that: In the step (1), before preprocessing, the collected data is divided into a test set and a training set.

Images