A gas turbine neural network fault-tolerant control method and system

Abstract

The present invention relates to a gas turbine neural network fault-tolerant control method and system. The present invention uses a neural network estimation model to simulate the gas turbine, determines the expected output in a fault-free situation, and determines based on a neural network compensation model that the actual output of the gas turbine in a fault situation is consistent with the expected output. The present invention realizes the fault-tolerant control of the gas turbine based on artificial intelligence, without the need to establish an accurate mathematical model of the gas turbine.

Description

Technical Field

[0001] The present invention relates to the technical field of fault-tolerant control, and in particular to a gas turbine neural network fault-tolerant control method and system. Background Art

[0002] Gas turbine systems are prone to problems such as actuator degradation and damage during long-term operation. In the event of an actuator failure, an effective fault-tolerant control method needs to be designed to solve some fault problems that occur during operation. There are two major design methods for fault-tolerant controllers: hardware redundancy and software redundancy. Hardware redundancy uses double or multiple backups of system components to improve fault tolerance. This method has high reliability and security, but as the number of redundant components increases, the cost is also high; software redundancy uses the functional redundancy of different components in the system, and has good flexibility and high cost performance.

[0003] At present, in the process of realizing fault-tolerant control of gas turbine control system actuators, there are mainly the following methods for fault-tolerant control:

[0004] 1. Fault-tolerant pole configuration based on parameter space design. This method uses a steady state or output feedback control to configure the closed-loop poles of the system within a predetermined region and ensures that the closed-loop poles remain within the region when any of the considered fault modes occurs.

[0005] 2. Design a robust fault-tolerant controller based on the Riccati equation and Lyapunov stability theorem.

[0006] 3. Fault-tolerant reorganization of control laws. This method uses the dynamic model of the system under normal conditions as the reference model, with the design goal of ensuring that the dynamic response of the system after a fault occurs is as close as possible to the response of the reference model, and realizes control law reorganization through an optimization algorithm.

[0007] Most of the fault-tolerant control methods listed above have two defects: one is that the precise model of the system must be understood; the other is that it is not easy to handle nonlinear objects. However, today's systems are becoming more and more complex, and the faulty systems are basically nonlinear. It is very difficult to obtain the precise mathematical model of the system. Traditional fault-tolerant control methods can no longer adapt to the current complex systems. Summary of the invention

[0008] In view of this, the present invention provides a gas turbine neural network fault-tolerant control method and system to achieve fault-tolerant control of the gas turbine without requiring an accurate mathematical model of the gas turbine.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A gas turbine neural network fault-tolerant control method, the method comprising the following steps:

[0011] Controlling the gas turbine based on the control instruction at time t, and collecting the actual output of the gas turbine at time t+1;

[0012] Inputting the control instruction at time t output by the controller of the gas turbine into the neural network estimation model, and obtaining the output of the neural network estimation model as the expected output of the gas turbine at time t;

[0013] Determine whether a fault occurs according to the expected output and the actual output, and obtain a determination result;

[0014] When the judgment result indicates yes, the difference between the expected output and the actual output is input into the neural network compensation model, the output of the neural network compensation model is obtained as the control compensation amount, and the gas turbine is controlled based on the control instruction at time t+1 and the control compensation amount, so that the actual output of the gas turbine at time t+1 is consistent with the expected output;

[0015] When the determination result indicates no, the gas turbine is controlled based on the control command at time t+1.

[0016] Optionally, the relationship between the input and output of the neural network estimation model is:

[0017]

[0018] Among them, y(t) is the output of the neural network estimation model, that is, the expected output at time t, is the input of the neural network estimation model, i.e., the control instruction at time t, θ is the parameter of the neural network estimation model, and f0(·) is the function expression of the neural network estimation model.

[0019] Optionally, the objective function for training the parameters of the neural network estimation model is:

[0020]

[0021] in, is the objective function for training the parameters of the neural network estimation model, D is the number of input data, Estimate the parameter thresholds of the model for the neural network, is the residual signal corresponding to the kth input data, represents the expected output of the neural network estimation model corresponding to the kth input data, y k represents the actual output of the gas turbine corresponding to the kth input data, where the input data is a historical control instruction.

[0022] Optionally, the corresponding relationship between the input and output of the neural network compensation model is:

[0023]

[0024] Among them, u f (t) is the output of the neural network compensation model, i.e., the control compensation amount, ε i is the residual value of the i-th neuron node in the input layer, w j is the weight of the jth neuron node in the hidden layer, v ij is the weight from the ith neuron node in the input layer to the jth neuron node in the hidden layer, n1 and n2 are the number of neuron nodes in the input layer and the hidden layer respectively, and f(·) is the activation function of the neural network compensation model.

[0025] Optionally, the objective function for training the neural network compensation model is:

[0026]

[0027] Among them, Δθ(t) is the objective function used to train the neural network compensation model, λ is a pre-set positive number, α is the momentum factor, η t represents the learning rate of the neural network compensation model training, e(t) represents the deviation between the output of the neural network compensation model and the actual output, y b (t) represents the actual output, μ(t) represents the control command received by the actuator, and μ f (t) represents the control instruction output by the neural network compensation model, and θ represents the parameter vector composed of the weights of the neuron nodes of the neural network compensation model.

[0028] Optionally, judging whether a fault occurs and obtaining a judgment result according to the expected output and the actual output, and judging whether a fault occurs and obtaining a judgment result specifically include:

[0029] According to the expected output and the actual output, the judgment formula Is it true? Among them, δ is the fault limit, y a and b are the expected output and the actual output respectively;

[0030] If so, the judgment result indicates no, and if not, the judgment result indicates yes.

[0031] A gas turbine neural network fault-tolerant control system, the system comprising:

[0032] A first control module, used to control the gas turbine based on the control instruction at time t, and collect the actual output of the gas turbine at time t+1;

[0033] A neural network estimator, used for inputting the control instruction outputted by the controller of the gas turbine at time t into the neural network estimation model, and obtaining the output of the neural network estimation model as the expected output of the gas turbine at time t+1;

[0034] A fault detector, used to determine whether a fault occurs according to the expected output and the actual output, and obtain a determination result;

[0035] A neural network compensator, configured to input the difference between the expected output and the actual output into a neural network compensation model when the judgment result indicates yes, obtain the output of the neural network compensation model as a control compensation amount, and control the gas turbine based on the control instruction at time t+1 and the control compensation amount, so that the actual output of the gas turbine at time t+1 is consistent with the expected output;

[0036] The second control module is used to control the gas turbine based on the control instruction at time t+1 when the judgment result indicates no.

[0037] Optionally, the relationship between the input and output of the neural network estimation model is:

[0038]

[0039] Among them, y(t+1) is the output of the neural network estimation model, that is, the expected output at time t+1, is the input of the neural network estimation model, i.e., the control instruction at time t, θ is the parameter of the neural network estimation model, and f0(·) is the function expression of the neural network estimation model.

[0040] Optionally, the objective function for training the parameters of the neural network estimation model is:

[0041]

[0042] in, is the objective function for training the parameters of the neural network estimation model, D is the number of input data, Estimate the parameter thresholds of the model for the neural network, is the residual signal corresponding to the kth input data, represents the expected output of the neural network estimation model corresponding to the kth input data, y k represents the actual output of the gas turbine corresponding to the kth input data, where the input data is a historical control instruction.

[0043] Optionally, the corresponding relationship between the input and output of the neural network compensation model is:

[0044]

[0045] Among them, u f (t) is the output of the neural network compensation model, i.e., the control compensation amount, ε i is the residual value of the i-th neuron node in the input layer, w j is the weight of the jth neuron node in the hidden layer, v ij is the weight from the ith neuron node in the input layer to the jth neuron node in the hidden layer, n1 and n2 are the number of neuron nodes in the input layer and the hidden layer respectively, and f(·) is the activation function of the neural network compensation model.

[0046] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0047] The present invention discloses a gas turbine neural network fault-tolerant control method and system, the method comprising the following steps: controlling the gas turbine based on a control instruction at time t, collecting the actual output of the gas turbine at time t+1; inputting the control instruction at time t output by the gas turbine controller into a neural network estimation model, obtaining the output of the neural network estimation model as the expected output of the gas turbine at time t+1; judging whether a fault occurs according to the expected output and the actual output, and obtaining a judgment result; when the judgment result indicates yes, inputting the difference between the expected output and the actual output into a neural network compensation model, obtaining the output of the neural network compensation model as a control compensation amount, and controlling the gas turbine based on the control instruction at time t+1 and the control compensation amount; when the judgment result indicates no, controlling the gas turbine based on the control instruction at time t+1. The present invention uses a neural network estimation model to simulate the gas turbine, determine the expected output in a fault-free situation, and determine based on the neural network compensation model that the actual output of the gas turbine in a fault situation is consistent with the expected output. The present invention realizes fault-tolerant control of the gas turbine based on artificial intelligence, without the need to establish an accurate mathematical model of the gas turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0049] Figure 1 A flow chart of a gas turbine neural network fault-tolerant control method provided by an embodiment of the present invention;

[0050] Figure 2 A schematic diagram of the structure of a gas turbine neural network fault-tolerant control system provided by an embodiment of the present invention;

[0051] Figure 3 A schematic diagram of nodes of a neural network estimation model provided by an embodiment of the present invention;

[0052] Figure 4 A comparison chart of simulation results provided by an embodiment of the present invention;

[0053] Figure 5 The embodiment of the present invention provides Figure 4 The simulation result comparison chart in Figure 2 is a local enlarged view of the fault location. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The object of the present invention is to provide a gas turbine neural network fault-tolerant control method and system, so as to realize the fault-tolerant control of the gas turbine without the need for an accurate mathematical model of the gas turbine.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] When a system fails, a large number of residuals are usually generated. These residuals play a key role in the fault diagnosis and fault-tolerant control of the system because they contain all the information about the system failure. According to the compensation control principle, these residuals are used to build a neural network compensation model. Using the residual between the expected output and the actual output of the faulty system, a neural network compensation model is designed. When a fault occurs, compensation control is introduced to weaken the impact of the fault, ensuring that it can still operate normally or close to normal conditions, and realizing its fault-tolerant control function.

[0059] When the system is operating normally, the neural network estimation model is set up in parallel with the controlled object as an approximate model of the controlled object (gas turbine). The neural network estimation model fully approximates the dynamic model of the controlled object, thereby establishing a neural network identification model for the entire system. When a fault occurs in the system, the expected output y of the neural network estimation model is used. a (t) and the actual output y of the controlled object b (t), and the difference ε(t) is sent to the fault detector and the neural network compensation model. The output u f (t) and the output u of the original controller c (t) jointly determine the output of the control quantity and compensate for the state deviation of the controlled object caused by the fault. Moreover, after training, the neural network parameters are no longer randomly selected, so the problem of difficult stability of the control system is effectively overcome, and the convergence speed is very fast. The performance of the neural network fault-tolerant control system depends on the quality of the original system controller design.

[0060] like Figure 1 As shown, embodiment 1 of the present invention provides a gas turbine neural network fault-tolerant control method, the method comprising the following steps:

[0061] Step 101, controlling the gas turbine based on the control instruction at time t, and collecting the actual output of the gas turbine at time t.

[0062] like Figure 2 As shown, the embodiment of the present invention makes the execution process of the gas turbine equivalent to Figure 2 The actual output is Figure 2 The y in b (t).

[0063] Step 102, inputting the control instruction at time t output by the controller of the gas turbine into the neural network estimation model, and obtaining the output of the neural network estimation model as the expected output of the gas turbine at time t.

[0064] The neural network estimation model is deployed in Figure 2 For the neural network compensator in , the expected output is Figure 2 The y in a (t).

[0065] Step 102 specifically includes the following steps:

[0066] Step 1: Construct a system state space model for the gas turbine speed and power control system;

[0067] Construct a system state space model for the gas turbine speed and power control system;

[0068] Assume the state equation of the system is:

[0069] x(t+1)=f(x(t),u(t))

[0070] The output equation is:

[0071] y(t)=g(x(t))

[0072] Where f(·) and g(·) are both nonlinear functions.

[0073] Step 2: Train the neural network estimation model

[0074] According to the input and output of the controlled object, the corresponding data is collected in the normal state to train the neural network, so as to determine the parameters of the neural network estimation model. That is:

[0075]

[0076] Among them, y(t) is the output of the neural network estimation model, that is, the expected output at time t, is the input of the neural network estimation model, i.e., the control instruction at time t, θ is the parameter of the neural network estimation model, and f0(·) is the function expression of the neural network estimation model.

[0077] The neural network estimation model is used to estimate the state of the gas turbine. The prediction error is the difference between the measured value (actual output) of the gas turbine output and the one-step prediction value (expected output):

[0078]

[0079] Through continuous training and learning of the neural network, and correction of the weights, the following is finally achieved:

[0080] e t (t)→0

[0081] Thus, a neural network estimation model is obtained, and the output of the neural network estimation model is equal to the system output under normal circumstances.

[0082] The following is a specific neural network estimation model modeling, using the commonly used three-layer BP neural network to establish the neural network estimation model of the system. The node diagram of the neural network estimation model is as follows Figure 3 shown.

[0083] The indicator function for modeling the neural network estimation model is defined as:

[0084]

[0085] Where ε(t) = y b (t)-ya (t) is the residual signal, then the objective function of the neural network estimation model learning is:

[0086]

[0087] Where D is the data length, Thresholds for estimating model parameters for a neural network.

[0088] The neuron activation function of the network hidden layer adopts the Sigmoid function, and the neuron learning rate is set to η.

[0089] Step 103: determine whether a fault occurs according to the expected output and the actual output, and obtain a determination result.

[0090] Step 103 specifically includes the following steps:

[0091] Step 3: Construct an actuator fault detection model for the constant gain fault of the actuator

[0092] By comparing the output of the neural network estimation model, that is, the expected output y a (t) and the actual output y of the controlled object b (t), establish the actuator constant gain fault detection model:

[0093]

[0094] In the formula, δ is the fault limit set by the system. When it is greater than the fault limit, the fault detection model detects the fault output and activates the neural network compensation model for control.

[0095] Step 104, when the judgment result indicates yes, the difference between the expected output and the actual output is input into the neural network compensation model, the output of the neural network compensation model is obtained as the control compensation amount, and the gas turbine is controlled based on the control instruction at time t+1 and the control compensation amount, so that the actual output of the gas turbine at time t+1 is consistent with the expected output.

[0096] The neural network compensation model is deployed in Figure 2 On the neural network compensator in .

[0097] Step 104 includes the following steps:

[0098] Step 4: Train the neural network compensation model

[0099] For the system that generates faults, a neural network compensation model is designed, and the good nonlinear approximation ability of the neural network is used to generate the control compensation amount corresponding to the fault. However, in the actual operation process, it is impossible to simulate each fault in a targeted manner, and then in the training process of the neural network compensation model, it is impossible to obtain the actual output of the gas turbine for different instructions under different faults. Based on this, the present invention first collects corresponding data under N potential fault states to train the neural network, so as to determine the parameters of the neural network estimation model, and then establish N state estimators. That is:

[0100] Applying these N state estimators to perform state estimation on the system, the state estimator of each model is the difference between the measured value of the system output and the one-step predicted value. Based on this, further training is performed on the basis of the neural network estimation model obtained in step 102 to obtain the state estimator corresponding to each potential fault. The state estimator can be used to simulate the actual output of the gas turbine for different instructions under different faults, thereby assisting the training of the neural network compensation model.

[0101] Similar to the structure of the neural network estimation model, the neural network compensation model uses a three-layer BP neural network. In order to maximize the use of the residual signal between the expected output of the neural network estimation model and the actual output of the controlled object during a fault, the output and input relationship of the neural network compensation model can be expressed by the following expression:

[0102]

[0103] Among them, ε i =ε(t-i+1), v and w are the weights of the neural network, n1 and n2 are the number of neuron nodes in the input layer and the hidden layer respectively, and f is the activation function of the neural network.

[0104] Let θ be a parameter vector consisting of v and w, e(t) = u(t) - y b (t), then the learning algorithm of θ based on the gradient descent method is expressed as follows:

[0105]

[0106] Δθ(t)=θ(t)-θ(t-1) (5)

[0107] Among them, η f is the learning rate of the neural network compensator. To ensure the convergence of the system, it is modified as follows:

[0108]

[0109] Among them, Δθ(t) is the objective function used to train the neural network compensation model, λ is a pre-set positive number, α is the momentum factor, η t represents the learning rate of the neural network compensation model training, e(t) represents the deviation between the output of the neural network compensation model and the actual output, y b (t) represents the actual output, μ(t) represents the control command received by the actuator, and μ f (t) represents the control instruction output by the neural network compensation model, and θ represents the parameter vector composed of the weights of the neuron nodes of the neural network compensation model. Exemplarily, in order to prevent the denominator from being zero, λ is a pre-set positive number close to zero.

[0110] Step 105: When the judgment result indicates no, the gas turbine is controlled based on the control instruction at time t+1.

[0111] Example 2

[0112] Embodiment 2 of the present invention provides a gas turbine neural network fault-tolerant control system, such as Figure 2 The system shown comprises:

[0113] The first control module is used to control the gas turbine based on the control instruction at time t and collect the actual output of the gas turbine at time t+1.

[0114] The neural network estimator is used to input the control instruction output by the gas turbine controller at time t into the neural network estimation model, and obtain the output of the neural network estimation model as the expected output of the gas turbine at time t+1.

[0115] The relationship between the input and output of the neural network estimation model is:

[0116]

[0117] Among them, y(t+1) is the output of the neural network estimation model, that is, the expected output at time t+1, is the input of the neural network estimation model, i.e., the control instruction at time t, θ is the parameter of the neural network estimation model, and f0(·) is the function expression of the neural network estimation model.

[0118] The objective function for training the parameters of the neural network estimation model is:

[0119]

[0120] in, is the objective function for training the parameters of the neural network estimation model, D is the number of input data, Estimate the parameter thresholds of the model for the neural network, is the residual signal corresponding to the kth input data, represents the expected output of the neural network estimation model corresponding to the kth input data, y k represents the actual output of the gas turbine corresponding to the kth input data, where the input data is a historical control instruction.

[0121] The fault detector is used to determine whether a fault occurs according to the expected output and the actual output, and obtain a determination result.

[0122] A neural network compensator is used to input the difference between the expected output and the actual output into a neural network compensation model when the judgment result indicates yes, obtain the output of the neural network compensation model as a control compensation amount, and control the gas turbine based on the control instruction at time t+1 and the control compensation amount, so that the actual output of the gas turbine at time t+1 is consistent with the expected output.

[0123] The corresponding relationship between the input and output of the neural network compensation model is:

[0124]

[0125] Among them, u f (t) is the output of the neural network compensation model, i.e., the control compensation amount, ε i is the residual value of the i-th neuron node in the input layer, w j is the weight of the jth neuron node in the hidden layer, v ij is the weight from the ith neuron node in the input layer to the jth neuron node in the hidden layer, n1 and n2 are the number of neuron nodes in the input layer and the hidden layer respectively, and f(·) is the activation function of the neural network compensation model.

[0126] The second control module is used to control the gas turbine based on the control instruction at time t+1 when the judgment result indicates no.

[0127] The first control module and the second control module correspond to Figure 2 The process of controlling the actuator and the controlled object.

[0128] Example 3

[0129] Example 3 of the present invention is used to verify the stability of controlling a gas turbine using the method in Example 1.

[0130] When the system operates normally, the input of the control quantity u=u c ; But when the system fails, the input of the control quantity is determined by the output u of the controllerc The compensation amount u of the compensator f Working together, that is:

[0131] u=u c +u f (7)

[0132] Under the control of the fault compensator, the impact of the system caused by the fault is determined by the control quantity u of the compensator. f To compensate, the faulty system is re-matched with the compensated model as much as possible. In this case, when the controller meets certain conditions, the stability of the closed-loop system can be guaranteed.

[0133] Take the Lyapunov function as Apply control algorithms (3), (6), and (7) if

[0134]

[0135] Where η1 is the learning rate of the neural network compensator

[0136]

[0137] Then the closed-loop system can ensure local asymptotic stability.

[0138] Proof: First, let Δe(t) = e(t+1)-e(t), then:

[0139]

[0140] Δe(t) is expanded using Lagrangian and its first-order partial derivative is obtained:

[0141]

[0142] and

[0143]

[0144] The controller and compensator have the same algorithm, then

[0145]

[0146] in

[0147]

[0148] make

[0149]

[0150] Substituting (4)(13) into formula (11), we can obtain:

[0151]

[0152] Among them, λ1, λ2>0 are pre-set positive numbers close to zero.

[0153]

[0154] From the above formula, we can know that to satisfy ΔV(t)<0, we only need to make

[0155]

[0156] Right now:

[0157]

[0158] Because there is

[0159]

[0160] And η1 is a pre-set learning rate, so formula (20) can be further transformed into:

[0161]

[0162]

[0163] Under the condition of satisfying formula (8), the stability of the closed-loop system can be guaranteed. The designed neural network fault-tolerant control system can be stable.

[0164] Learning rate η of the neural network fault compensator f Depends on S max When S max When θ is very small, Δθ also becomes very small, and the weights of the neural network used for compensation can be regarded as fixed, that is, the output control amount of the fault compensator is unchanged, which means that the system has achieved fault tolerance. In addition, the neural network is fault-tolerant. If the number of hidden layers and parameters of the neural network compensator are set appropriately, it can have better generalization ability and can also obtain better control effects for unlearned fault modes.

[0165] Example 4

[0166] Embodiment 4 of the present invention uses simulation to verify the effectiveness of the method in Embodiment 1 of the present invention.

[0167] During simulation, the neural network estimation model has 1 input node, 6 hidden nodes, and 1 output node; the neural network compensator model has 1 input node, 6 hidden nodes, and 1 output node. The initial value of the learning rate is 0.7, the error is 0.01, and the fault limit is set to 0.05. When t=10s, the actuator has a constant gain fault with an amplitude of 0.8. The fault is simulated, and the comparison results between the original controller and the neural network fault-tolerant compensation control are shown in the figure. Figure 4 and 5 As shown, it can be seen that the output of the controlled object is more stable when a fault occurs in the neural network fault-tolerant compensation control method of the present invention.

[0168] The present invention aims at the nonlinearity and uncertainty of the gas turbine system, and uses the neural network for the fault-tolerant control system to make up for the shortcomings of the traditional control method and improve the intelligent level of the control system. Since the neural network has good adaptive ability, learning ability and fault tolerance, the fault-tolerant control system designed by the neural network method can speed up the response speed of the system, enhance the adaptability, robustness and fault tolerance of the system, thereby improving the information processing ability of the entire system.

[0169] The present invention aims at the constant gain fault of the actuator in the speed power control system of the gas turbine, and designs a state estimator and a compensator based on the BP neural network to solve the influence of the fault on the system, so that the system can continue to operate stably, and at the same time greatly reduce the parameter fluctuation during the fault process, and reduce the influence of the fault on the stability of the system. The neural network estimator is trained by collecting the input and output data of the controlled object, and a complete neural network system state estimator is obtained after the training is completed. When there is no fault, the output of the neural network estimator is equal to the output of the controlled object. When a fault occurs in the system, the output of the neural network estimator and the output of the controlled object will have a certain deviation. The residual generated by them is used as the input of the neural network compensator, and the neural network compensator will generate a compensation control amount corresponding to the residual, and the compensation control amount and the output of the original controller of the system will work together to generate a control amount for control. Under the action of the fault compensator, the deviation caused by the fault of the system is compensated by the control amount of the compensator, so that the fault system and its model are re-matched as much as possible. In this case, the system can quickly reach stability again.

[0170] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0171] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A gas turbine neural network fault-tolerant control method, characterized in that: The method comprises the following steps: Controlling the gas turbine based on the control instruction at time t, and collecting the actual output of the gas turbine at time t; Inputting the control instruction at time t output by the controller of the gas turbine into the neural network estimation model, and obtaining the output of the neural network estimation model as the expected output of the gas turbine at time t; Determine whether a fault occurs according to the expected output and the actual output, and obtain a determination result; When the judgment result indicates yes, the difference between the expected output and the actual output is input into the neural network compensation model, the output of the neural network compensation model is obtained as the control compensation amount, and the gas turbine is controlled based on the control instruction at time t+1 and the control compensation amount, so that the actual output of the gas turbine at time t+1 is consistent with the expected output; When the judgment result indicates no, controlling the gas turbine based on the control instruction at time t+1; The objective function used to train the neural network compensation model is: Among them, Δθ(t) is the objective function used to train the neural network compensation model, λ is a pre-set positive number, α is the momentum factor, η t represents the learning rate of the neural network compensation model training, e(t) represents the deviation between the output of the neural network compensation model and the actual output, y b (t) represents the actual output, μ(t) represents the control command received by the actuator, and μ f (t) represents the control instruction output by the neural network compensation model, and θ represents the parameter vector composed of the weights of the neuron nodes of the neural network compensation model.

2. The gas turbine neural network fault-tolerant control method according to claim 1, characterized in that: The relationship between the input and output of the neural network estimation model is: Among them, y(t) is the output of the neural network estimation model, that is, the expected output at time t, is the input of the neural network estimation model, i.e., the control instruction at time t, θ is the parameter of the neural network estimation model, and f0(·) is the function expression of the neural network estimation model.

3. The gas turbine neural network fault-tolerant control method according to claim 1, characterized in that: The objective function for training the parameters of the neural network estimation model is: in, is the objective function for training the parameters of the neural network estimation model, D is the number of input data, Estimate the parameter thresholds of the model for the neural network, is the residual signal corresponding to the kth input data, represents the expected output of the neural network estimation model corresponding to the kth input data, y k represents the actual output of the gas turbine corresponding to the kth input data, where the input data is a historical control instruction.

4. The gas turbine neural network fault-tolerant control method according to claim 1, characterized in that: The corresponding relationship between the input and output of the neural network compensation model is: Among them, u f (t) is the output of the neural network compensation model, i.e., the control compensation amount, ε i is the residual value of the i-th neuron node in the input layer, w j is the weight of the jth neuron node in the hidden layer, v ij is the weight from the ith neuron node in the input layer to the jth neuron node in the hidden layer, n1 and n2 are the number of neuron nodes in the input layer and the hidden layer respectively, and f(·) is the activation function of the neural network compensation model.

5. The gas turbine neural network fault-tolerant control method according to claim 1, characterized in that: Determining whether a fault occurs and obtaining a determination result according to the expected output and the actual output, and determining whether a fault occurs and obtaining a determination result specifically include: According to the expected output and the actual output, the judgment formula Is it true? Among them, δ is the fault limit, y a and b are the expected output and the actual output respectively; If so, the judgment result indicates no, and if not, the judgment result indicates yes.

6. A gas turbine neural network fault-tolerant control system, characterized in that: The system applies the method described in any one of claims 1 to 5, and the system comprises: A first control module, used to control the gas turbine based on the control instruction at time t, and collect the actual output of the gas turbine at time t; A neural network estimator, used for inputting the control instruction at time t output by the controller of the gas turbine into the neural network estimation model, and obtaining the output of the neural network estimation model as the expected output of the gas turbine at time t; A fault detector, used to determine whether a fault occurs according to the expected output and the actual output, and obtain a determination result; A neural network compensator, configured to input the difference between the expected output and the actual output into a neural network compensation model when the judgment result indicates yes, obtain the output of the neural network compensation model as a control compensation amount, and control the gas turbine based on the control instruction at time t+1 and the control compensation amount, so that the actual output of the gas turbine at time t+1 is consistent with the expected output; The second control module is used to control the gas turbine based on the control instruction at time t+1 when the judgment result indicates no.

7. The gas turbine neural network fault-tolerant control system according to claim 6, characterized in that: The relationship between the input and output of the neural network estimation model is: Among them, y(t) is the output of the neural network estimation model, that is, the expected output at time t, is the input of the neural network estimation model, i.e., the control instruction at time t, θ is the parameter of the neural network estimation model, and f0(·) is the function expression of the neural network estimation model.

8. The gas turbine neural network fault-tolerant control system according to claim 6, characterized in that: The objective function for training the parameters of the neural network estimation model is: in, is the objective function for training the parameters of the neural network estimation model, D is the number of input data, Estimate the parameter thresholds of the model for the neural network, is the residual signal corresponding to the kth input data, represents the expected output of the neural network estimation model corresponding to the kth input data, y k represents the actual output of the gas turbine corresponding to the kth input data, where the input data is a historical control instruction.

9. The gas turbine neural network fault-tolerant control system according to claim 6, characterized in that: The corresponding relationship between the input and output of the neural network compensation model is: Among them, u f (t) is the output of the neural network compensation model, i.e., the control compensation amount, ε i is the residual value of the i-th neuron node in the input layer, w j is the weight of the jth neuron node in the hidden layer, v ij is the weight from the ith neuron node in the input layer to the jth neuron node in the hidden layer, n1 and n2 are the number of neuron nodes in the input layer and the hidden layer respectively, and f(·) is the activation function of the neural network compensation model.

Images