A design method for linear active disturbance rejection controller based on neural network prediction

Abstract

The invention discloses a design method of a linear active disturbance rejection controller based on neural network prediction, comprising: step 1, obtaining and preprocessing sample data, constructing a sample training set and a sample test set; step 2, constructing an ELMAN neural network prediction model using the sample training set, and verifying it with the sample test set; step 3, inputting operating condition data into the ELMAN neural network prediction model to obtain output values ​​to transform an extended state observer. The control effect of the controlled object under various disturbances is improved by designing a linear active disturbance rejection controller.

Description

Technical Field

[0001] The invention belongs to the field of automatic control, and in particular relates to a design method of a linear active disturbance rejection controller based on neural network prediction. Background Art

[0002] Active disturbance rejection control (ADRC) takes the unmodeled dynamics and unknown disturbances of the system as total disturbances, estimates them through the extended state observer, and eliminates them in feedback control. ADRC does not rely on the system model, has strong robustness, is suitable for complex field control environments, and can be relatively easily built in the configuration. It has been widely used in industry. In complex industrial environments, there are generally control objects with large inertia and large delay. Classical ADRC technology is difficult to directly handle large delays. Therefore, the application of classic ADRC technology in large delay systems faces the situation of poor regulation effect.

[0003] The existing Chinese patent application number is 201410495004.1, which discloses a design and tuning method of an ADRC control system for a time-delay system. Based on the ADRC technology, the method first fits the complex controlled object into a first-order inertia link plus pure time-delay mathematical model, and at the same time, the time delay is attributed to the disturbance quantity. The time-delay-reduced linear extended state observer is used to estimate the unknown total disturbance containing time delay, and actively compensate for the influence of the total disturbance on the system, thereby reducing the time-delay system to the "integrator series type" of the ADRC standard, and realizing the compensation of the time-delay system. Finally, the closed-loop transfer function of the system is derived, the pure time-delay link in the characteristic equation is eliminated, and the universal ADRC single-parameter tuning formula and the numerical relationship between the adjustable parameters are given accordingly. The simulation results verify that the designed practical ADRC has good stability, rapidity, accuracy and anti-disturbance.

[0004] Another improved version of the linear ADRC was proposed. It is based on the structure of the ordinary linear ADRC and uses a high-order inertia link in the feedback loop instead of pure delay to compensate for the delay time.

[0005] Both of the above controllers utilize partial information of the controlled object and can improve the control effect of large delay objects to a certain extent. However, actual engineering objects often face a variety of external interferences in complex environments, and their characteristics change. In particular, time-delay systems often show time-varying phenomena, and delay characteristics are prone to fluctuations and changes. The existing controlled objects are affected by external disturbances, resulting in the existing improved anti-disturbance controller having a poor or even unstable regulation effect in actual engineering, and it is difficult for the anti-disturbance control to achieve satisfactory control effects. Summary of the invention

[0006] In view of the deficiencies in the prior art, an object of the present invention is to provide a design method for a linear active disturbance rejection controller based on neural network prediction.

[0007] The design method of the linear active disturbance rejection controller of the present invention comprises:

[0008] Step 1: Obtain and preprocess sample data to construct sample training sets and sample test sets;

[0009] Step 2: Use the sample training set to construct the ELMAN neural network prediction model, and verify it with the sample test set;

[0010] Step 3: Input the operating condition data into the ELMAN neural network prediction model to obtain the output value and transform the extended state observer.

[0011] Furthermore, the acquisition and preprocessing of sample data in step 1 and the construction of sample training sets and sample test sets include:

[0012] Step 101, dividing the collected historical sample data into training samples and test samples, wherein the training samples and the test samples both include interference factors of the controlled object, the controlled amount of the controlled object and the output value of the controlled object under different working conditions;

[0013] Step 102: Substitute the training sample and the test sample into the control object model to calculate the training sample delay time and the test sample delay time, which are defined as labeled data and unlabeled data respectively;

[0014] Step 103, according to the interference factor of the controlled object under the rated working condition, the controlled variable of the controlled object and the output value of the controlled object, the controlled object model is substituted to calculate and obtain the rated working condition delay time;

[0015] Step 104, according to the label output value calculation model, use the rated operating delay time and the labeled data to calculate the labeled output vector; use the rated operating time and the unlabeled data to calculate the unlabeled output vector, as a test to see whether the constructed model meets the preset requirements.

[0016] Furthermore, the control object model in step 103 is calculated as follows:

[0017]

[0018] In the above formula (1), y is the output value of the controlled object, s is the Laplace operator, K is the static gain of the process, u is the input value of the controlled object, τ is the delay time of the controlled object, and T is the inertia time of the process.

[0019] Furthermore, the label output value calculation model in step 104 is as follows:

[0020] f(d,u) k =τ k -τ0.......(2),

[0021] In the above formula (2), f(d,u) k is the delay time deviation caused by the kth disturbance, d is the interference factor of the controlled object, τ k is the delay time estimate obtained through parameter identification in the kth process, k is the number, k∈(1..n), and n represents the number of parameter identifications.

[0022] Furthermore, in step 2, the sample training set is used to construct the ELMAN neural network prediction model, as shown in the following formula (3):

[0023]

[0024] In the above formula (3), W U (k), W C (k) and W O (k) is the weight value of the input layer, the weight matrix of the receiving layer and the output layer, u in (k) is the kth input vector containing the main disturbance factors and control quantities, v in (k) is the input vector of the hidden layer, y out (k) is the kth output variable, x H (k) is the output vector of the hidden layer, x c (k) is the output vector of the receiving layer, f1 and f2 are the functions of the input layer and the output layer, and α is the gain of the receiving layer;

[0025] The ELMAN neural network prediction model is verified using a sample test set, including:

[0026] Step 201: Construct a training data set X = (u in ,y out ), the input vector is composed of the interference factor of the controlled object in the training sample as the interference factor of the controlled object d and the input value u of the controlled object, which is The output vector is composed of the label output vector, which is

[0027] Step 202: Set the training data set X = (u in ,y out ) is substituted into the ELMAN neural network prediction model to solve the values ​​of each parameter and determine the ELMAN neural network prediction model;

[0028] Step 203: Model verification, constructing a test data set X'=(u in test ,y out test), using unlabeled data to construct the input vector for testing, which is u in test , using the unlabeled output vector to construct the output vector for testing, y out test , the input vector u of the test in test Substitute into the ELMAN neural network prediction model to solve the predicted output vector y' out test ;

[0029] Step 204: predict the output vector y' out test and the test output vector y out test Substitute the test model into the test model to determine whether the ELMAN neural network prediction analysis model meets the pre-set requirements. If it does not meet the pre-set requirements, re-select sample data to train the ELMAN neural network prediction model to determine the optimal parameter values. If it meets the pre-set requirements, the ELMAN neural network prediction model is constructed.

[0030] Furthermore, the ELMAN neural network prediction model test in step 3 is calculated according to the following formula (4):

[0031]

[0032] In the above formula (4), ε is a preset value.

[0033] Furthermore, in step 3, the inputting of the operating condition data into the ELMAN neural network prediction model to obtain the output value and transform the extended state observer includes:

[0034] Step 301, obtaining interference factors of the control object and input values ​​of the control object in the current state;

[0035] Step 302: Substitute the interference factors of the control object and the input value of the control object in the current state into the ELMAN neural network prediction and analysis model to solve the delay time estimation value of the current working condition;

[0036] Step 303: Input the estimated delay time value of the current working condition into the extended state observer for transformation, and the transformed value is the following formula (5):

[0037]

[0038] In the above formula (5), L0 = [β1 β2 L β n β n+1 ] T represents the state observer gain, C0=[1 0 … 0 0] T , z(t) is the state variable in the extended state observer, is the derivative of the state variable in the extended state observer, is the estimated value of the state extended observer for the object output y(t), u(t-(τ0+y out (k))) is the input quantity u past (τ0+y out (k)) value at the moment.

[0039] The beneficial effects of the present invention are:

[0040] The present invention proposes a design method for an improved active disturbance rejection controller based on neural network prediction. The controller proposed by this method utilizes a neural network to realize the prediction function of the delay characteristics of complex objects under multiple interferences, thereby realizing precise control of large-delay objects under multiple interferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a schematic diagram of the controller structure of the present invention;

[0042] Figure 2 It is a specific flow chart of the design method of the improved active disturbance rejection controller based on neural network prediction described in the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings.

[0044] The invention proposes a design method of an improved active disturbance rejection controller based on neural network prediction.

[0045] like Figure 1 , Figure 2 As shown, the design method of the linear active disturbance rejection controller includes:

[0046] Step 1: Obtain and preprocess sample data to construct sample training sets and sample test sets;

[0047] Step 2: Use the sample training set to construct the ELMAN neural network prediction model, and verify it with the sample test set;

[0048] Step 3: Input the operating condition data into the ELMAN neural network prediction model to obtain the output value, transform the extended state observer, and construct ANN-LADRC.

[0049] Furthermore, the acquisition and preprocessing of sample data in step 1 and the construction of sample training sets and sample test sets include:

[0050] Step 101, dividing the collected historical sample data into training samples and test samples, wherein the training samples and the test samples both include interference factors of the controlled object, the controlled amount of the controlled object and the output value of the controlled object under different working conditions;

[0051] Step 102: Substitute the training sample and the test sample into the control object model to calculate the training sample delay time and the test sample delay time, which are defined as labeled data and unlabeled data respectively;

[0052] Step 103, according to the interference factor of the controlled object under the rated working condition, the controlled variable of the controlled object and the output value of the controlled object, the controlled object model is substituted to calculate and obtain the rated working condition delay time;

[0053] Step 104, according to the label output value calculation model, use the rated operating delay time and the labeled data to calculate the labeled output vector; use the rated operating time and the unlabeled data to calculate the unlabeled output vector, as a test to see whether the constructed model meets the preset requirements.

[0054] Furthermore, the control object model in step 103 is calculated as follows:

[0055]

[0056] In the above formula (1), y is the output value of the controlled object, s is the Laplace operator, K is the static gain of the process, u is the input value of the controlled object, τ is the delay time of the controlled object, and T is the inertia time of the process.

[0057] Furthermore, the label output value calculation model in step 104 is as follows:

[0058] f(d,u) k =τ k -τ0.......(2),

[0059] In the above formula (2), f(d,u) k is the delay time deviation caused by the kth disturbance, d is the interference factor of the controlled object, τ k is the delay time estimate obtained through parameter identification in the kth process, k is the number, k∈(1..n), and n represents the number of parameter identifications.

[0060] Furthermore, in step 2, the sample training set is used to construct the ELMAN neural network prediction model, as shown in the following formula (3):

[0061]

[0062] In the above formula (3), W U (k), W C (k) and W O (k) is the weight value of the input layer, the weight matrix of the receiving layer and the output layer, u in (k) is the kth input vector containing the main disturbance factors and control quantities, v in(k) is the input vector of the hidden layer, y out (k) is the kth output variable, x H (k) is the output vector of the hidden layer, x c (k) is the output vector of the receiving layer, f1 and f2 are the functions of the input layer and the output layer, and α is the gain of the receiving layer;

[0063] The sample test set is used to verify the ELMAN neural network prediction model, including:

[0064] Step 201: Construct a training data set X = (u in ,y out ), the input vector is composed of the interference factor of the controlled object in the training sample as the interference factor of the controlled object d and the input value u of the controlled object, which is The output vector is composed of the label output vector, which is

[0065] Step 202: Set the training data set X = (u in ,y out ) is substituted into the ELMAN neural network prediction model to solve the values ​​of each parameter and determine the ELMAN neural network prediction analysis model;

[0066] Step 203: Model verification, constructing a test data set X'=(u in test ,y out test ), using unlabeled data to construct the input vector for testing, which is u in test , using the unlabeled output vector to construct the output vector for testing, y out test , the input vector u of the test in test Substitute into the ELMAN neural network prediction model to solve the predicted output vector y' out test ;

[0067] Step 204: predict the output vector y' out test and the test output vector y out test Substitute the test model into the test model to determine whether the ELMAN neural network prediction model meets the pre-set requirements. If it does not meet the pre-set requirements, re-select sample data to train the ELMAN neural network prediction model to determine the optimal parameter values. If it meets the pre-set requirements, the ELMAN neural network prediction analysis model is constructed.

[0068] Furthermore, the ELMAN neural network prediction model test in step 3 is calculated according to the following formula (4):

[0069]

[0070] In the above formula (4), ε is a preset value.

[0071] Furthermore, in step 3, the operating condition data is input into the ELMAN neural network prediction model to obtain the output value, and the extended state observer is transformed, including:

[0072] Step 301, obtaining interference factors of the control object and input values ​​of the control object in the current state;

[0073] Step 302: Substitute the interference factors of the control object and the input value of the control object in the current state into the ELMAN neural network prediction and analysis model to solve the delay time estimation value of the current working condition;

[0074] Step 303: Input the estimated delay time value of the current working condition into the extended state observer for transformation, and the transformed value is the following formula (5):

[0075]

[0076] In the above formula (5), L0 = [β1 β2 L β n β n+1 ] T represents the state observer gain, C0=[1 0 … 0 0] T , z(t) is the state variable in the extended state observer, is the derivative of the state variable in the extended state observer, is the estimated value of the state extended observer for the object output y(t), u(t-(τ0+y out (k))) is the input quantity u past (τ0+y out (k)) value at the moment.

[0077] The specific step 3 is, assuming that the model of the process object is:

[0078]

[0079] Among them, y is the system output; u is the system input; w is the external disturbance of the system;

[0080] is the generalized disturbance of the system, is the unknown function of the unknown disturbance inside the system and the external disturbance of the system; x(t) is the state variable of the system; b0 is the gain coefficient;

[0081] Expand g to a new state, let x1 = y, x n =y (n-1) , x n+1 =g;

[0082] Assume that g is differentiable and The system is represented as:

[0083]

[0084] in,

[0085] C0=[1 0 … 0 0] T ,

[0086] Combined with the output of the ELMAN neural network model, LSEO is designed as:

[0087]

[0088] Where, L0=[β1 β2 L β n β n+1 ] T ——State observer gain.

[0089] Based on LESO's estimation of system state and disturbance, the system control rate is designed as:

[0090]

[0091] Among them, the state feedback control rate u o (t) Designed to:

[0092]

[0093] Then we get:

[0094]

[0095] Among them, r(t) is the input signal to be tracked, ——state feedback gain;

[0096] The remaining unknown parameters in the controller are b0, L0, and K0, which are adjusted using an optimization algorithm.

[0097] The present invention is not limited to the above-mentioned embodiments. Without departing from the essential content of the present invention, any deformation, improvement and substitution that can be thought of by those skilled in the art shall fall into the protection scope of the present invention.

Claims

1. A design method for a linear active disturbance rejection controller based on neural network prediction, characterized in that: include: Step 1: Obtain and preprocess sample data to construct sample training sets and sample test sets: Step 101, dividing the collected historical sample data into training samples and test samples, wherein the training samples and the test samples both include interference factors of the controlled object, the controlled amount of the controlled object and the output value of the controlled object under different working conditions; Step 102: Substitute the training sample and the test sample into the control object model to calculate the training sample delay time and the test sample delay time, which are defined as labeled data and unlabeled data respectively. The control object model is as follows: In the above formula (1), y is the output value of the controlled object, s is the Laplace operator, K is the static gain of the process, u is the input value of the controlled object, τ is the delay time of the controlled object, and T is the inertia time of the process; Step 103, according to the interference factor of the controlled object under the rated working condition, the controlled variable of the controlled object and the output value of the controlled object, substitute into the control object model calculation formula (1) to obtain the rated working condition delay time; Step 104: According to the label output value calculation model, the labeled output vector is calculated using the rated operating condition delay time and the labeled data; the unlabeled output vector is calculated using the rated operating condition time and the unlabeled data, as a test to see whether the constructed model meets the preset requirements. The label output value calculation model is as follows: f(d, u) k =t k -τ0.......(2), In the above formula (2), f(d,u) k is the delay time deviation caused by the disturbance d at the kth time, τ k is the estimated value of the delay time obtained through parameter identification in the kth process, k is the number, k∈(1..n), and n represents the number of parameter identifications; Step 2: Use the sample training set to construct the ELMAN neural network prediction model, as shown in formula (3): In the above formula (3), W U (k), W C (k) and W O (k) is the weight value of the input layer, the weight matrix of the receiving layer and the output layer, u in (k) is the kth input vector containing the main disturbance factors and control quantities, v in (k) is the input vector of the hidden layer, y out (k) is the kth output variable, x H (k) is the output vector of the hidden layer, x c (k) is the output vector of the receiving layer, f1 and f2 are the functions of the input layer and the output layer, and α is the gain of the receiving layer; The ELMAN neural network prediction model is verified using a sample test set, including: Step 201: Construct a training data set X = (u in ,y out ), the input vector is composed of the interference factor d of the controlled object and the input value u of the controlled object in the training sample, The output vector is composed of the label output vector, Step 202: Set the training data set X = (u in ,y out ) is substituted into the ELMAN neural network prediction model to solve the values ​​of each parameter and determine the ELMAN neural network prediction model; Step 203: Model verification, constructing a test data set X'=(u in test ,y out test ), using unlabeled data to construct the input vector for testing, which is u in test , using the unlabeled output vector to construct the output vector for testing as y out test , the input vector u of the test in test Substitute into the ELMAN neural network prediction model to solve the predicted output vector y' out test ; Step 204: predict the output vector y' out test and the test output vector y out test Substitute the test model to determine whether the ELMAN neural network prediction analysis model meets the pre-set requirements. If it does not meet the pre-set requirements, re-select sample data to train the ELMAN neural network prediction model to determine the optimal parameter values. If it meets the pre-set requirements, the ELMAN neural network prediction analysis model is constructed; Step 3: Input the operating condition data into the ELMAN neural network prediction model, obtain the output value and transform the extended state observer, as shown in the following formula (4): In the above formula (4), ε is a preset value; Input the operating condition data to the ELMAN neural network prediction model to obtain the output value. The modified extended state observer includes: Step 301, obtaining interference factors of the control object and input values ​​of the control object in the current state; Step 302: Substitute the interference factors of the control object and the input value of the control object in the current state into the ELMAN neural network prediction and analysis model to solve the delay time estimation value of the current working condition; Step 303: Input the estimated delay time value of the current working condition into the extended state observer for transformation. The transformed extended state observer is the following formula (5): In the above formula (5), L0 = [β1β2…β n β n+1 ] T represents the state observer gain, C0=[1 0 … 0 0] T , z(t) is the state variable in the extended state observer, is the derivative of the state variable in the extended state observer, is the estimated value of the state extended observer for the object output y(t), u(t-(τ0+y out (k)) is the input quantity u past (τ0+y out (k)) value at the moment.

Images