Rocket engine fault detection and diagnosis method based on time sequence adversarial network

Abstract

The invention discloses a rocket engine fault detection and diagnosis method based on a time sequence adversarial network. The method comprises the following steps: S1, constructing a time sequence adversarial network model; s2, training the time sequence adversarial network model by using normal data of an engine; s3, performing engine whole-process fault detection by using the trained time sequence adversarial network model; and S4, performing engine steady-state process fault diagnosis by using the trained time sequence adversarial network model. The whole-process fault detection of the engine is realized by using the time sequence adversarial network, the problem that the current whole-process fault detection of the engine needs various methods and models can be solved, and meanwhile, the method has the advantage of high diagnosis speed.

Description

Technical Field

[0001] The present invention relates to the field of rocket engine technology, and more specifically, to a rocket engine fault detection and diagnosis method based on a temporal adversarial network. Background Art

[0002] The launch failure of a launch vehicle is a cruel reality that major space powers have to face from time to time, and it often causes serious economic losses or casualties. Liquid rocket engines are the core component of the launch vehicle propulsion system. As the main power unit of the launch vehicle, their reliability is related to the success or failure of the space mission. In recent years, my country's aerospace industry has developed rapidly, and large-thrust reusable engines have successfully completed repeated flight tests. High-density space launches have increasingly higher requirements for engine safety and reliability. Therefore, it is necessary to further carry out research on engine fault detection and diagnosis.

[0003] Engine fault detection usually requires the use of multiple methods and models according to different engine operating conditions, which is highly complex and requires a lot of calculations. Existing engine steady-state fault diagnosis methods have the problem of low efficiency. Therefore, a low-complexity, high-computational-efficiency method is needed to realize full-process engine fault detection and rapid diagnosis of steady-state process faults. Summary of the invention

[0004] The purpose of the present invention is to provide a rocket engine fault detection and diagnosis method based on a temporal adversarial network to overcome the defects of the prior art.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] The rocket engine fault detection and diagnosis method based on the temporal adversarial network includes the following steps:

[0007] S1. Build a temporal adversarial network model;

[0008] S2. Using normal engine data to train the temporal adversarial network model;

[0009] S3, using the trained time series adversarial network model to perform engine fault detection throughout the entire process;

[0010] S4. Use the trained time series adversarial network model to perform engine steady-state process fault diagnosis.

[0011] Furthermore, the temporal adversarial network in step S1 is composed of a discriminator and a generator, wherein the discriminator is composed of a multi-layer perceptron MLP, and the generator is composed of a long short-term memory network LSTM.

[0012] The multilayer perceptron MLP is composed of multiple layers of neurons, including an input layer, a hidden layer, and an output layer. The input layer is composed of input signals, and the hidden layer and the output layer are fully connected layers. The calculation method is:

[0013]

[0014] Where, the subscript i represents the i-th neuron in this layer, j represents the j-th neuron in the previous layer, O is the output, x is the output of the previous layer, f is the activation function, W is the weight, and b is the bias vector;

[0015] The training process of the multi-layer perceptron MLP is realized by a back propagation algorithm. The learning process of the back propagation algorithm consists of forward propagation of signals and back propagation of errors. During forward propagation, the input signal is propagated to the output layer through the input layer and the hidden layer, and the error is calculated based on the actual output and the expected output. During the back propagation process, the error is propagated back along the original path, and the error is shared by each unit of each layer and the weight is corrected.

[0016] The long short-term memory network LSTM affects the output by using memory cells that store past inputs. The equation describing an LSTM basic unit at a certain time t is:

[0017] f t =σ(W hf h t-1 +W xf x t +b f )

[0018] i t =σ(W hi h t-1 +W xi x t +b i )

[0019] o t =σ(W ho h t-1 +W xo x t +b o )

[0020] c t ′=tanh(W hc h t-1 +W xc x t +b c )

[0021] c t =f t c t-1 +i t ct '

[0022] h t =o t tanh(c t )

[0023] In the formula, x is the input parameter, W is the weight, b is the bias, f represents the forget gate, which determines whether the value in the memory cell remains unchanged, decreased or reset, σ is the sigmoid function, tanh is the hyperbolic tangent function, i represents the input gate, which determines whether the current input is added to the storage cell and the value to be updated, o represents the output gate, which determines the output by the input and the value of the memory cell, c' controls the input to be updated, c represents the cell state, and h represents the hidden state;

[0024] The loss function of the discriminator selects the mean square error, which is calculated as follows:

[0025]

[0026] In the formula, x is the input sample, x r To reconstruct the sample;

[0027] The loss function of the generator is:

[0028] L g =(MSE1+MSE2) / 2

[0029] In the formula, MSE1 is calculated by the output parameters of the generator and the actual parameters, and MSE2 is calculated by the discrimination results of the discriminator and the actual labels.

[0030] Furthermore, the step S2 specifically includes:

[0031] Inputting random noise into the generator and outputting generated samples;

[0032] Inputting generated samples into the discriminator and outputting a discrimination result;

[0033] Calculate the error according to the loss function of the generator and update the network weights of the generator;

[0034] Inputting generated samples and real samples into the discriminator, outputting a discrimination result, calculating an error according to a loss function of the discriminator, and updating the network weights of the discriminator;

[0035] The training of the generator and the discriminator is performed alternately, and the training is completed when a set number of training rounds or a certain condition is reached.

[0036] Furthermore, the step S3 specifically includes:

[0037] The test sample is input into the discriminator, and the discrimination result is output. The whole process fault detection of the engine is realized through the discrimination result; if the discrimination result is greater than 0.5, it means that the test sample is relatively consistent with the training sample, which is a normal sample, and the output is set to 1; otherwise, the output is set to 0, indicating an abnormal sample, indicating that the engine has a fault.

[0038] Furthermore, the step S4 specifically includes:

[0039] The test sample is input into the generator, the predicted sample is output, and the absolute error between the predicted value and the true value of each parameter is calculated according to the following formula:

[0040] AE i =|x i -x ip |

[0041] Where: i represents the number of parameters, x i is the true value, x ip is the predicted value;

[0042] The absolute error between the predicted value and the true value of each parameter is standardized by the following formula:

[0043]

[0044] In the formula, μ i and σ i AE i The mean and standard deviation of

[0045] For each sample, the diagnostic indicator is a 24-dimensional vector SAE:

[0046] SAE=(SAE1,SAE2,...,SAE 24 )

[0047] The SAE of five consecutive samples after the engine failure is analyzed. According to the size and change trend of the SAE values ​​of different parameters and combined with the comprehensive analysis of the engine structure, the fault diagnosis result is obtained.

[0048] Furthermore, the generator is replaced with a recurrent neural network or a gated recurrent unit.

[0049] Compared with the prior art, the advantages of the present invention are: the present invention utilizes a temporal adversarial network to realize engine full-process fault detection, which can solve the problem that the current engine full-process fault detection requires multiple methods and models, and at the same time has the advantage of fast diagnosis speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] 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 or the description of the prior art 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 work.

[0051] Figure 1 It is a flow chart of the rocket engine fault detection and diagnosis method based on the timing adversarial network of the present invention.

[0052] Figure 2 It is a schematic diagram of the temporal adversarial network model in the present invention.

[0053] Figure 3 It is a structural diagram of the long short-term memory network LSTM unit in the present invention.

[0054] Figure 4 It is a schematic diagram of the temporal adversarial network training process in the present invention. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0056] See also Figure 1 As shown, this embodiment discloses a rocket engine fault detection and diagnosis method based on a temporal adversarial network, comprising the following steps:

[0057] Step S1: construct a temporal adversarial network model.

[0058] In this embodiment, the temporal adversarial network in step S1 is composed of a discriminator and a generator, the discriminator is composed of a multi-layer perceptron MLP, and the generator is composed of a long short-term memory network LSTM.

[0059] The multilayer perceptron MLP is composed of multiple layers of neurons, including an input layer, a hidden layer, and an output layer. The input layer is composed of input signals, and the hidden layer and the output layer are fully connected layers. The calculation method is:

[0060]

[0061] Where, the subscript i represents the i-th neuron in this layer, j represents the j-th neuron in the previous layer, O is the output, x is the output of the previous layer, f is the activation function, W is the weight, and b is the bias vector.

[0062] The training process of the multi-layer perceptron MLP is implemented by the back propagation algorithm (BP). The learning process of the back propagation algorithm consists of forward propagation of signals and back propagation of errors. During forward propagation, the input signal is propagated to the output layer through the input layer and the hidden layer, and the error is calculated based on the actual output and the expected output. The method for calculating the error is called the loss function. During the back propagation process, the error is transmitted back along the original path, and the error is shared by each unit of each layer and the weight is corrected.

[0063] The long short-term memory network LSTM affects the output by using memory cells that store past inputs. The equation describing an LSTM basic unit at a certain time t is:

[0064] f t =σ(W hf h t-1 +W xf x t +b f )

[0065] i t =σ(W hi h t-1 +W xi x t +b i )

[0066] o t =σ(W ho h t-1 +W xo x t +b o )

[0067] c t ′=tanh(W hc h t-1 +W xc x t +b c )

[0068] c t =f t c t-1 +i t c t '

[0069] h t =o t tanh(c t )

[0070] In the formula, x is the input parameter, W is the weight, b is the bias, f represents the forget gate, which determines whether the value in the memory cell remains unchanged, decreased or reset, σ is the sigmoid function, tanh is the hyperbolic tangent function, i represents the input gate, which determines whether the current input is added to the storage cell and the value to be updated, o represents the output gate, which determines the output by the input and the value of the memory cell, c' controls the input to be updated, c represents the cell state, and h represents the hidden state;

[0071] The loss function of the discriminator selects the mean square error, which is calculated as follows:

[0072]

[0073] In the formula, x is the input sample, x r To reconstruct the sample;

[0074] The loss function of the generator is:

[0075] L g =(MSE1+MSE2) / 2

[0076] In the formula, MSE1 is calculated by the output parameters of the generator and the actual parameters, and MSE2 is calculated by the discrimination results of the discriminator and the actual labels.

[0077] Step S2: Use normal engine data to train the temporal adversarial network model.

[0078] In this embodiment, step S2 specifically includes:

[0079] Input random noise to the generator and output generated samples; input generated samples to the discriminator and output discrimination results; calculate errors according to the loss function of the generator and update the network weights of the generator; input generated samples and real samples to the discriminator and output discrimination results, calculate errors according to the loss function of the discriminator and update the network weights of the discriminator; train the generator and the discriminator alternately, and complete the training when the set number of training rounds or certain conditions are reached.

[0080] Step S3: Use the trained sequential adversarial network model to perform engine full-process fault detection.

[0081] In this embodiment, step S3 specifically includes: inputting the test sample into the discriminator, outputting the discrimination result, and realizing the whole process fault detection of the engine through the discrimination result; if the discrimination result is greater than 0.5, it means that the test sample is relatively consistent with the training sample, and is a normal sample, and the output is set to 1, otherwise the output is set to 0, indicating an abnormal sample, indicating that the engine has a fault.

[0082] Step S4: Use the trained sequential adversarial network model to perform engine steady-state process fault diagnosis.

[0083] In this embodiment, step S4 specifically includes:

[0084] The test sample is input into the generator, the predicted sample is output, and the absolute error between the predicted value and the true value of each parameter is calculated according to the following formula:

[0085] AE i =|x i -x ip

[0086] Where: i represents the number of parameters, x i is the true value, x ip is the predicted value;

[0087] The absolute error between the predicted value and the true value of each parameter is standardized by the following formula:

[0088]

[0089] In the formula, μ i and σ i AE i The mean and standard deviation of

[0090] For each sample, the diagnostic indicator is a 24-dimensional vector SAE:

[0091] SAE=(SAE1,SAE2,...,SAE 24 )

[0092] The SAE of five consecutive samples after the engine failure is analyzed. According to the size and change trend of the SAE values ​​of different parameters and combined with the comprehensive analysis of the engine structure, the fault diagnosis result is obtained.

[0093] In this embodiment, all steps S1 to S4 can be completed through Python programming to achieve rocket engine fault detection and diagnosis based on the temporal adversarial network.

[0094] Preferably, the present invention may also replace the long short-term memory network LSTM with various networks such as a recurrent neural network (RNN), a gated recurrent unit (GRU), etc., which may realize fault diagnosis and achieve basically the same effect.

[0095] The present invention utilizes a temporal adversarial network to implement engine full-process fault detection, which can solve the problem that current engine full-process fault detection requires multiple methods and models, and at the same time has the advantage of fast diagnosis speed.

[0096] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, the patent owner may make various variations or modifications within the scope of the appended claims. As long as they do not exceed the protection scope described in the claims of the present invention, they should be within the protection scope of the present invention.

Claims

1. A rocket engine fault detection and diagnosis method based on a temporal adversarial network, characterized in that: The following steps are involved: S1. Build a temporal adversarial network model; S2. Using normal engine data to train the temporal adversarial network model; S3, using the trained time series adversarial network model to perform engine fault detection throughout the entire process; S4. Use the trained time series adversarial network model to perform engine steady-state process fault diagnosis.

2. The rocket engine fault detection and diagnosis method based on temporal adversarial network according to claim 1 is characterized in that: The temporal adversarial network in step S1 is composed of a discriminator and a generator, wherein the discriminator is composed of a multi-layer perceptron MLP, and the generator is composed of a long short-term memory network LSTM; The multilayer perceptron MLP is composed of multiple layers of neurons, including an input layer, a hidden layer, and an output layer. The input layer is composed of input signals, and the hidden layer and the output layer are fully connected layers. The calculation method is: Where, the subscript i represents the i-th neuron in this layer, j represents the j-th neuron in the previous layer, O is the output, x is the output of the previous layer, f is the activation function, W is the weight, and b is the bias vector; The training process of the multi-layer perceptron MLP is realized by a back propagation algorithm. The learning process of the back propagation algorithm consists of forward propagation of signals and back propagation of errors. During forward propagation, the input signal is propagated to the output layer through the input layer and the hidden layer, and the error is calculated based on the actual output and the expected output. During the back propagation process, the error is propagated back along the original path, and the error is shared by each unit of each layer and the weight is corrected. The long short-term memory network LSTM affects the output by using memory cells that store past inputs. The equation describing an LSTM basic unit at a certain time t is: f t =σ(W hf h t-1 +W xf x t +b f ) i t =σ(W hi h t-1 +W xi x t +b i ) o t =σ(W ho h t-1 +W xo x t +b o ) c t ′=tanh(W hc h t-1 +W xc x t +b c ) c t =f t c t-1 +i t c t ′ h t =o t fishy t ) In the formula, x is the input parameter, W is the weight, b is the bias, f represents the forget gate, which determines whether the value in the memory cell remains unchanged, decreased or reset, σ is the sigmoid function, tanh is the hyperbolic tangent function, i represents the input gate, which determines whether the current input is added to the storage cell and the value to be updated, o represents the output gate, which determines the output by the input and the value of the memory cell, c' controls the input to be updated, c represents the cell state, and h represents the hidden state; The loss function of the discriminator selects the mean square error, which is calculated as follows: In the formula, x is the input sample, x r To reconstruct the sample; The loss function of the generator is: L g =(MSE1+MSE2) / 2 In the formula, MSE1 is calculated by the output parameters of the generator and the actual parameters, and MSE2 is calculated by the discrimination results of the discriminator and the actual labels.

3. The rocket engine fault detection and diagnosis method based on temporal adversarial network according to claim 2 is characterized in that: The step S2 specifically includes: Inputting random noise into the generator and outputting generated samples; Inputting generated samples into the discriminator and outputting a discrimination result; Calculate the error according to the loss function of the generator and update the network weights of the generator; Inputting generated samples and real samples into the discriminator, outputting a discrimination result, calculating an error according to a loss function of the discriminator, and updating the network weights of the discriminator; The training of the generator and the discriminator is performed alternately, and the training is completed when a set number of training rounds or a certain condition is reached.

4. The rocket engine fault detection and diagnosis method based on temporal adversarial network according to claim 3 is characterized in that: The step S3 specifically includes: The test sample is input into the discriminator, and the discrimination result is output. The whole process fault detection of the engine is realized through the discrimination result; if the discrimination result is greater than 0.5, it means that the test sample is relatively consistent with the training sample, which is a normal sample, and the output is set to 1; otherwise, the output is set to 0, indicating an abnormal sample, indicating that the engine has a fault.

5. The rocket engine fault detection and diagnosis method based on temporal adversarial network according to claim 4 is characterized in that: The step S4 specifically includes: The test sample is input into the generator, the predicted sample is output, and the absolute error between the predicted value and the true value of each parameter is calculated according to the following formula: AE i =x i -x ip Where: i represents the number of parameters, x i is the true value, x ip is the predicted value; The absolute error between the predicted value and the true value of each parameter is standardized by the following formula: In the formula, μ i and σ i AE i The mean and standard deviation of For each sample, the diagnostic indicator is a 24-dimensional vector SAE: SAE=(SAE1,SAE2,...,SAE 24 ) The SAE of five consecutive samples after the engine failure is analyzed. According to the size and change trend of the SAE values ​​of different parameters and combined with the comprehensive analysis of the engine structure, the fault diagnosis result is obtained.

6. The rocket engine fault detection and diagnosis method based on temporal adversarial network according to claim 1 is characterized in that: The generator is replaced by a recurrent neural network or a gated recurrent unit.

Images