Fault diagnosis method for aero-engine combustion chamber

Abstract

The invention provides a fault diagnosis method for an aero-engine combustion chamber, which comprises the following steps of: processing a signal by using a wavelet packet analysis technology, selecting a db3 wavelet function, reducing the noise of the signal by using a hard threshold wavelet packet noise reduction method, decomposing and reconstructing a wavelet packet to extract an energy feature vector of the signal, and performing fault diagnosis on the energy feature vector of the signal. Taking the obtained feature vector as the input of a neural network; designing a GA + BP algorithm, and combining a genetic algorithm with a neural network; the GA performs early-stage optimization on the BP neural network, and determines an optimal network structure and an initial weight, a threshold and a learning rate of the network corresponding to the structure; and then, constructing a neural network with the optimal structure and parameters to carry out fault diagnosis. The workload of the GA + BP neural network is less than that of a BP neural network, the defect of local minimum is overcome, and better training performance is achieved; the fault diagnosis accuracy of the GA + BP neural network is higher than that of the BP neural network.

Description

Technical Field

[0001] The invention belongs to the technical field of engine control, and in particular relates to a method for diagnosing faults in a combustion chamber of an aero-engine. Background Art

[0002] With the rapid advancement of science and technology, modern production equipment is becoming increasingly large-scale, high-speed, efficient, automated, and continuous. Aircraft engine combustion chambers are not only expected to perform well and be highly efficient, but also to operate with minimal or no failures. The operational quality of these high-tech aircraft engine combustion chambers directly impacts both social and economic benefits. However, these high-tech aircraft engine combustion chambers are subject to varying degrees of failure, ranging from minor performance degradation to complete failure and, in even more serious cases, immeasurable damage to property and life. Therefore, research on aircraft engine combustion chamber condition monitoring and fault diagnosis technologies has become a critical topic.

[0003] Fault diagnosis is a comprehensive discipline that uses information about the operating status of aircraft engine combustion chambers to identify fault sources and determine appropriate responses. The ultimate goal of fault diagnosis technology is to prevent failures (especially major accidents), ensure personal safety and the safety of aircraft engine combustion chambers, promote reforms in aircraft engine combustion chamber maintenance systems, extend maintenance cycles, improve maintenance accuracy and speed, reduce maintenance costs, increase production efficiency, and achieve optimal economic benefits.

[0004] Acquiring fault signals from aircraft engine combustion chambers is a crucial step in combustor fault diagnosis and lifespan assessment. As a core component of aircraft engines, the operating status of the combustion chamber is directly related to engine performance and safety. Therefore, collecting and analyzing fault signals from the combustion chamber, as well as assessing its lifespan, are crucial for ensuring proper engine operation and flight safety. Summary of the Invention

[0005] In view of this, the present invention aims to overcome the deficiencies of the above-mentioned problems in the prior art and proposes a method for diagnosing combustion chamber faults of an aircraft engine.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0007] A first aspect of the present invention provides a method for diagnosing a combustion chamber fault in an aircraft engine, comprising the following steps:

[0008] Step 1: Collect the fault signal of the aircraft engine combustion chamber, perform wavelet packet decomposition and reconstruction on the signal, and extract the energy feature vector of the signal;

[0009] Step 2: Construct a BP neural network, use genetic algorithm to optimize the structure and parameters of the BP neural network, obtain a GA-BP neural network, and train the GA-BP neural network;

[0010] Step 3: Input the energy feature vector of the extracted signal into the trained GA-BP neural network to obtain the fault type.

[0011] Furthermore, the aircraft engine combustion chamber fault signal in step 1 includes a pressure signal, a temperature signal, a vibration signal, and a flame signal, which are used to reflect the pressure fluctuation, temperature distribution, vibration condition, and flame state information inside the combustion chamber.

[0012] Furthermore, the step 1 also includes processing the collected signal, including filtering, denoising, amplifying, discretizing and performing wavelet packet noise reduction on the signal.

[0013] Furthermore, in step 1, performing wavelet packet decomposition and reconstruction on the signal to extract the energy feature vector of the signal includes:

[0014] (1) Normalization processing of fault signals:

[0015] The data normalization method is:

[0016] x i (t)′=2(x i (t)-minx i (t)) / (maxx i (t)-minx i (t))-1

[0017] Where x i (t) represents the original data, x i (t)′ represents the normalized data, maxx i (t) and minx i (t) represents the maximum and minimum values ​​of all data of the same input node;

[0018] (2) Decompose the signal into N layers of wavelet packets, using X Nj Represents the Nth layer low frequency to high frequency 2 N The decomposition coefficient vector of the frequency band is obtained from the Nth layer from low frequency to high frequency 2 N The characteristic signal S of the frequency band Nj , j=1,2,…,2 N ;

[0019] (3) Calculate the energy of each frequency band signal:

[0020]

[0021] (4) Construct the eigenvector:

[0022] To E Nj Perform normalization:

[0023]

[0024] The energy eigenvector is thus determined to be

[0025] Furthermore, in step 2, the structure and parameters of the BP neural network are optimized using a genetic algorithm to obtain a GA-BP neural network including:

[0026] Initialize the structure, weights, thresholds and learning rates of the BP neural network;

[0027] The structure, weights, thresholds and learning rates of BP neural network are optimized by using the optimization capability of genetic algorithm.

[0028] Furthermore, the optimization of the structure, weights, thresholds and learning rates of the BP neural network using the genetic algorithm optimization capability includes:

[0029] Genetic encoding and decoding:

[0030] Decompose each chromosome into linker genes and parameter genes;

[0031] The connection gene corresponds to the number of hidden layer nodes of the neural network. It uses binary coding. The coding string consists of binary "0" and "1". Each binary number in the coding string represents a hidden layer neuron. "1" means that the neuron exists, and "0" means that it does not exist. The binary coding length is equal to the number of hidden layer nodes.

[0032] The parameter gene adopts real number encoding. Each connection weight, threshold and learning rate is directly represented by a real number. The parameter gene is divided into two parts: weight threshold gene and rate gene. The encoding length of weight threshold gene is equal to the total number of all weights and thresholds in each layer of the neural network. The encoding length of rate gene is one. The encoding length of parameter gene is the sum of the length of weight threshold gene and rate gene. The total encoding length of a chromosome is the sum of the encoding length of connection gene and the encoding length of parameter gene.

[0033] Design fitness function:

[0034] The fitness function of each chromosome is taken as

[0035] f=1(1+E)

[0036] In the formula is the total error in the neural network, represents the ideal output, represents the true output, K is the number of sample sets;

[0037] Design genetic operators, including selection, crossover, and mutation:

[0038] The selection operation adopts a method combining the best individual preservation and fitness ratio;

[0039] The connection gene adopts a one-point crossover method, and the weight threshold gene and rate gene in the parameter gene adopt an arithmetic crossover method. The two parts cross each other without interfering with each other.

[0040] The connection gene adopts the basic mutation mode, and the weight threshold gene and rate gene in the parameter gene adopt the non-uniform mutation mode, but the two parts mutate independently without interfering with each other;

[0041] Select the control parameters.

[0042] A second aspect of the present invention provides an aircraft engine combustion chamber fault diagnosis device, comprising:

[0043] A signal acquisition unit is used to collect the fault signal of the combustion chamber of the aircraft engine, perform wavelet packet decomposition and reconstruction on the signal, and extract the energy characteristic vector of the signal;

[0044] A neural network construction unit is used to construct a BP neural network, optimize the structure and parameters of the BP neural network using a genetic algorithm to obtain a GA-BP neural network, and train the GA-BP neural network;

[0045] The signal processing unit is used to input the energy feature vector of the extracted signal into the trained GA-BP neural network to obtain the fault type.

[0046] A third aspect of the present invention provides an electronic device comprising a processor and a memory communicatively connected to the processor and used to store instructions executable by the processor, wherein the processor is used to execute the above-mentioned method for diagnosing a combustion chamber fault of an aircraft engine.

[0047] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned method for diagnosing a fault in a combustion chamber of an aircraft engine.

[0048] Compared with the prior art, the method for diagnosing a combustion chamber fault of an aircraft engine according to the present invention has the following advantages:

[0049] The GA+BP neural network of the present invention has less workload, overcomes the disadvantage of being trapped in local minimum, and has better training performance;

[0050] The GA+BP neural network of the present invention has a high fault diagnosis accuracy rate and can better perform fault diagnosis work on aircraft engines. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings, which constitute part of the present invention, are provided to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are provided to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0052] Figure 1 is a flow chart of the method of the present invention;

[0053] Figure 2 The energy characteristic vector histograms of the four working states of the present invention are as follows;

[0054] Figure 3 This is a schematic diagram of the genetically encoded three-layer BP neural network structure of the present invention. DETAILED DESCRIPTION

[0055] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0056] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0057] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0058] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0059] Example 1:

[0060] like Figure 1 As shown, the present invention provides a method for diagnosing a fault in an aircraft engine combustion chamber, comprising the following steps:

[0061] Step 1: Collect the fault signal of the aircraft engine combustion chamber, perform wavelet packet decomposition and reconstruction on the signal, and extract the energy feature vector of the signal;

[0062] Fault signal acquisition primarily involves four aspects: signal selection, sensor placement, signal preprocessing, and signal sampling. First, signal selection involves selecting signals that reflect the combustion chamber fault based on its characteristics and diagnostic requirements. Commonly used signals include pressure, temperature, vibration, and flame signals. These signals can reveal information such as pressure fluctuations, temperature distribution, vibration, and flame conditions within the combustion chamber.

[0063] Secondly, sensor placement involves rationally arranging sensors to acquire the desired signals based on the signal selection results. Sensor placement must consider performance indicators such as sensor type, sensitivity, response speed, and resistance to high temperatures and pressures, as well as the combustion chamber's structure and operating environment. Commonly used sensors include pressure sensors, temperature sensors, vibration sensors, and flame sensors. The sensor layout should enable comprehensive and accurate acquisition of signals within the combustion chamber while minimizing any impact on the combustion chamber's structure and performance.

[0064] Next, signal preprocessing involves filtering, denoising, and amplifying the collected raw signal to improve its quality and analyzability. Due to the complexity of the combustion chamber's operating environment and interference during signal transmission, raw signals often contain noise and distortion. Signal preprocessing removes these noise and distortion components, retaining useful signal components and providing an accurate signal foundation for subsequent feature extraction and fault diagnosis.

[0065] Finally, signal sampling involves discretizing the preprocessed signal to meet the requirements of digital signal processing. The sampling process requires careful consideration of the sampling frequency to ensure that the sampled signal fully retains the original signal's information. The sampling frequency should adhere to the Nyquist sampling theorem, which states that the sampling frequency should be greater than twice the highest signal frequency to avoid spectral aliasing caused by sampling.

[0066] During the fault signal acquisition process, signal transmission and storage also need to be considered. Due to the unique operating environment of aircraft engine combustion chambers, signal transmission and storage face challenges such as high temperature, high pressure, and vibration. Therefore, it is necessary to select appropriate signal transmission media and storage devices to ensure signal stability and reliability during transmission and storage.

[0067] When a system fails, the spatial distribution of its output signal energy will change accordingly compared to the normal system output, that is, the output energy contains rich fault feature information. Therefore, the present invention extracts fault feature values ​​from the distribution of signal energy in each subspace, that is, using wavelet packet decomposition, the energy of the original signal is decomposed into 2 N In the orthogonal frequency bands, the signal is analyzed in different frequency bands after multi-layer decomposition, so that the originally inconspicuous signal frequency characteristics are expressed in the form of significant energy in several subspaces with different resolutions. The main steps of the fault feature extraction method based on wavelet packets are as follows:

[0068] 1. Normalization processing of fault signals:

[0069] In order to reduce the mutual influence between the data, the fault data is first normalized. The data normalization method is:

[0070] x i (t)′=2(x i (t)-minx i (t)) / (maxx i (t)-minx i (t))-1

[0071] Where x i (t) represents the original data, x i (t)′ represents the normalized data, maxx i (t) and minx i (t) represents the maximum and minimum values ​​of all data of the same input node.

[0072] 2. Decompose the signal into N layers of wavelet packets, using X Nj Represents the Nth layer low frequency to high frequency 2 N The decomposition coefficient vector of the frequency band is obtained from the Nth layer from low frequency to high frequency 2 N The characteristic signal S of the frequency band Nj , j=1,2,…,2 N .

[0073] 3. Calculate the energy of the signal in each frequency band:

[0074]

[0075] 4. Construct the eigenvector. First, E Nj Perform normalization:

[0076]

[0077] The energy eigenvector can be determined as

[0078] The present invention uses three-layer wavelet packet decomposition to extract energy feature vectors for 120 sets of data. The energy feature vectors of any set of data under the four working conditions are shown in the following figure: Figure 2 shown.

[0079] The present invention also performs hard threshold noise reduction processing on the signal before performing wavelet transformation.

[0080] Step 2: Construct a BP neural network, use genetic algorithm to optimize the structure and parameters of the BP neural network, obtain a GA-BP neural network, and train the GA-BP neural network;

[0081] Design of BP neural network:

[0082] (1) Determination of learning samples:

[0083] The present invention carefully selects 80 groups of learning samples, of which 20 groups are selected respectively for normal system operation, bubble existence, outlet blockage, and process leakage.

[0084] It has long been theoretically proven that a network with a bias, at least one sigmoid hidden layer, and a linear output can approximate any rational function. Increasing the number of layers primarily reduces error and improves accuracy, but it also complicates the network and increases training time. Accuracy can also be improved by increasing the number of hidden layer nodes. The training effect of this is easier to observe and adjust than increasing the number of layers, so increasing the number of hidden layer nodes should be prioritized.

[0085] BP neural network has a very important theory, that is, any continuous function in a closed interval can be approximated by a BP network with a single hidden layer. Therefore, the present invention adopts a three-layer BP neural network with one hidden layer.

[0086] (2) Design of input, output, and hidden layers

[0087] The design of the input and output layers is related to the input and output variables of the actual problem, respectively. The input layer of the neural network acts as a buffer register, and its node number depends on the dimensionality of the data source. The number of nodes in the output layer is determined by the number of operating states. After feature extraction using wavelet packet transform, the data in this invention obtains signals in eight different frequency bands, so the number of nodes in the BP neural network input layer is N = 8. The four operating states of the aircraft engine were selected for signal acquisition, so the number of nodes in the neural network output layer is M = 4.

[0088] The number of hidden layer nodes is directly related to the problem requirements and the number of input / output units. Too many or too few hidden units can lead to problems such as prolonged learning time, poor fault tolerance, and non-optimal error. Therefore, the number of hidden layer nodes in a network affects network performance and, consequently, fault diagnosis effectiveness. However, determining the number of hidden layer nodes in a neural network is a very difficult problem. There is no ideal analytical formula, and it often requires the designer's experience or multiple experiments to determine the optimal number of hidden layer nodes. Alternatively, the following reference formulas can be used to select the optimal number of hidden layer nodes.

[0089] (1) Where a is a constant between [1,10];

[0090] (2)s=log2n;

[0091] (3) In the above formula, m is the number of nodes in the output layer, n is the number of nodes in the input layer, and s is the number of nodes in the hidden layer.

[0092] The present invention refers to these formulas to find the interval of the number of hidden layer nodes, and then determines the number of hidden layer nodes through the simulation results of training.

[0093] (3) Selection of activation function

[0094] Once the structure of the neural network and the training data are determined, the total error function is completely determined by the activation function. Therefore, the choice of activation function plays a very important role in the convergence of the network.

[0095] The activation function of the input layer of the present invention adopts the tansig function: The activation function of the hidden layer uses the logsig (S-type) function: The larger the parameter k, the flatter the function and the easier it is for the network to converge, but the convergence rate is slow. The smaller k, the faster the convergence rate, but it is more prone to oscillation. After repeated experiments, it was found that when k = 1, the network error is smaller and the convergence rate is faster. Therefore, the hidden layer activation function of this invention uses k = 1. The output layer activation function uses the purelin function: f(x) = x.

[0096] (4) Selection of network initial value

[0097] The initial values ​​of a neural network have a significant impact on whether learning reaches a local minimum, converges, and the duration of training. If the initial values ​​deviate significantly, causing the weighted input to fall into the saturation region of the activation function, the adjustment process will almost come to a halt. Therefore, it is generally recommended that the output value of each neuron after initial weighting be close to zero. This ensures that the weight of each neuron is adjusted at the maximum value of its sigmoid activation function.

[0098] The present invention selects a random number (-1, 1) generated by MATLAB as the initial weight and threshold of the network.

[0099] (5) Selection of learning algorithm

[0100] The present invention adopts the Levenberg-Marquardt (LM) algorithm, namely the trainlm function in MATLAB.

[0101] (6) Learning rate

[0102] The learning rate is determined by the amount of weight change generated during cyclic training. A high learning rate can lead to system instability, while a low learning rate results in longer training cycles and slower convergence, but ensures that the network's error does not exceed the bottom of the error surface and ultimately converges to the minimum error. Therefore, a low learning rate is generally preferred to ensure system stability. Generally, the learning rate range is 0.01 to 0.8.

[0103] In the present invention, the learning rate is determined by observing the results of training simulation experiments.

[0104] (7) Training stop conditions

[0105] In a neural network with many weights, too little training can affect the network's classification and generalization capabilities, while too much training can lead to poor test results.

[0106] The present invention selects the minimum error function value and the maximum number of training times as the conditions for determining the end of training and judging network convergence. The network error requirement is ε=0.0001, and the maximum number of training times is 1000.

[0107] The present invention designs a GA+BP algorithm, that is, combining a genetic algorithm with a BP neural network, and uses the genetic algorithm to simultaneously optimize the structure and parameters of the BP neural network to obtain the best network structure and the initial weights, thresholds and learning rates corresponding to the structure.

[0108] Genetic encoding and decoding:

[0109] Coding: In the GA+BP algorithm, each chromosome is decomposed into connection genes and parameter genes, and different coding methods are used to encode these two parts.

[0110] The connection gene corresponds to the number of hidden layer nodes of the neural network. It uses binary coding. The coding string consists of binary "0" and "1". Each binary number in the coding string represents a hidden layer neuron. "1" means that the neuron exists, and "0" means it does not exist. Therefore, there are as many binary numbers of 0 or 1 as there are hidden layer neurons. In other words, the binary coding length is equal to the number of hidden layer nodes.

[0111] Parameter genes use real number encoding, meaning each connection weight, threshold, and learning rate is directly represented by a real number. Since network weights and thresholds can be negative, while learning rates must be positive, parameter genes are divided into weight-threshold genes and rate genes. The encoding length of the weight-threshold gene is equal to the total number of weights and thresholds in each layer of the neural network, and the encoding length of the rate gene is one. Therefore, the encoding length of the parameter gene is the sum of the lengths of the weight-threshold and rate genes, while the total encoding length of a chromosome is the sum of the encoding lengths of the connection genes and the parameter genes.

[0112] Decoding: Assume that 1 1 0 1…1 is the code of connecting genes in a chromosome, and the length of the code is K. Then the structure of the three-layer BP neural network is as follows: Figure 3 As shown in the figure, i n is the input layer neuron, h k is the hidden layer neuron, o m is the output layer neuron, X n is the input sample, Y m is the output of the network. During decoding, the weights ω from the input layer to the hidden layer and from the hidden layer to the output layer are n,k 、ω k,m and the threshold θ of the hidden layer k As shown below.

[0113] ω n,k =[ω 1，1 ω 1，2 0ω 1，4 …ω 1，K ω 2，1 ω 2，2 0ω 2，4 …ω 2，K ...ω N，1 ω N，2 0ω N，4 …ω N，K ]

[0114] ω k,m =[ω 1，1 ω 1，2 …ω 1，M ω 2，1 ω 2，2 …ω 2，M 00…0ω4，1 ω 4，2 …ω 4，M ...ω K，1 ω K，2 …ω K,M ]

[0115] θ k =[θ1θ20θ4…θ K ]

[0116] The learning rate of the network is independent of the network structure, so it does not participate in the above decoding process, that is, no decoding is required.

[0117] The advantages of this encoding method of the GA+BP algorithm are:

[0118] (1) A BP neural network with the corresponding structure can be directly established based on the binary {0, 1} coding string connecting the genes, and the number of hidden layer nodes of the neural network is equal to the number of "1" in the coding string. In this way, the structure of the network can be clearly seen after decoding;

[0119] (2) Real-number coded parameter genes have a larger genetic search space, and the network weights, thresholds, and learning rates can be obtained without complex decoding, which improves computational efficiency.

[0120] Design of fitness function:

[0121] The fitness function of the GA+BP algorithm is based on the total error of the neural network, that is, the fitness function of each chromosome is taken as

[0122] f=1(1+E)

[0123] In the formula is the total error in the neural network, represents the ideal output, Represents the true output, and K is the number of sample sets.

[0124] Design of genetic operators:

[0125] Genetic operations include three genetic operators: selection, crossover, and mutation.

[0126] The selection operation of the GA+BP algorithm adopts a method that combines the best individual preservation and fitness ratio. For the chromosomes in the current population, the best individual preservation strategy is first adopted. The idea of ​​this method is that the individual with the highest fitness in the population does not participate in crossover and mutation operations and is directly copied to the next generation. Its advantage is that the optimal solution of a certain generation in the evolutionary process will not be destroyed by crossover and mutation operations, but it may also cause the evolution to fall into a local solution because the genetic genes of the local optimal individuals will increase rapidly. Therefore, the fitness ratio method is used for the contemporary population after the best individual preservation strategy is selected. Assuming the population size is n, the probability of individual i being selected is

[0127]

[0128] Where, f i represents the fitness function value of the i-th chromosome. In this method, the probability of each chromosome being selected is It reflects the proportion of an individual's fitness to the total fitness of all individuals.

[0129] The combination of best individual preservation and fitness ratio ensures that individuals with high fitness function values ​​have a greater chance of being selected for the next generation, while individuals with low fitness function values ​​also have a chance. This ensures the diversity of individuals in the population and keeps the fitness function values ​​of individuals in the population close to the optimal solution, preventing the algorithm from falling into a local optimum.

[0130] The GA+BP algorithm also performs crossover and mutation on the selected chromosomes. Because the linker genes and parameter genes are encoded using different methods, the crossover needs to be performed separately. The linker genes use a one-point crossover method, while the weight threshold genes and rate genes in the parameter genes use arithmetic crossover. However, the two parts crossover independently, without interfering with each other.

[0131] A single-point crossover is also called a simple crossover. A crossover point is randomly set in an individual. When the crossover is performed, the partial structures of the two individuals before or after the point are exchanged to generate two new individuals:

[0132]

[0133] Arithmetic crossover refers to the generation of a new individual by the linear combination of two individuals. Suppose an arithmetic crossover is performed between two individuals x1 and x2, then the two new individuals x1′ and x′2 after the crossover are

[0134]

[0135] Here, α is a real number between 0 and 1.

[0136] Similar to the crossover operation, the mutation operation of the GA+BP algorithm uses different mutation methods for the linker genes and parameter genes. The linker genes use a basic mutation method, while the weight threshold genes and rate genes in the parameter genes use a non-uniform mutation method. However, the two parts mutate independently without interfering with each other.

[0137] Basic variation refers to randomly selecting one or more loci from the individual coding strings in the population and performing a P-value analysis on the gene values ​​of these loci. m The mutation probability of :

[0138] Individual A1 0 1 1 011→Individual A′1 1 1 0 11

[0139] The specific operation is as follows: First, each locus of an individual is designated as a mutation point with a mutation probability; then, the gene value of each designated mutation point is inverted or replaced with another allele value, thereby generating a new generation of individuals. Uniform mutation replaces the original gene value with a uniformly distributed random number within a certain range, allowing individuals to move freely within the search space. However, this makes it inconvenient to conduct local searches in a key area. Therefore, non-uniform mutation does not replace the original gene value with a uniformly distributed random number. Instead, the original gene value is randomly perturbed and the perturbed result is used as the new gene value after mutation. Mutating each locus with the same probability is equivalent to making a slight shift in the solution space for the entire solution vector.

[0140] Selection of control parameters:

[0141] The group size can be selected between 10 and 200 according to actual conditions.

[0142] During the optimization process, the crossover probability P c Always controls the dominant crossover operator. Inappropriate crossover probability can lead to unexpected consequences. A larger crossover probability can make each generation fully crossover, but P c If P is too large, the possibility of destroying excellent individuals will increase, thus making the search random; c If it is too small, more individuals will directly enter the next generation, and the search may become stagnant. c The value range is 0.4~0.99.

[0143] Step 3: Input the energy feature vector of the extracted signal into the trained GA-BP neural network to obtain the fault type.

[0144] In actual operation, the present invention preliminarily sets the number of hidden layer nodes to 25, initializes the BP neural network, and randomly generates an initial group of 80 individuals according to the encoding method. The encoding length of each individual is 25+((8×25+25×4+25+4)+1). Among them, the first 25 are connection genes, using binary encoding, and each individual is initialized to "0" or "1"; the following ((8×25+25×4+25+4)+1) are parameter genes, using real number encoding. Among them, (8×25+25×4+25+4) are weight threshold genes, initialized to a real number (-1, 1), and the last one is a rate gene, initialized to a real number (0, 1). As we know from Section 4.3.7, the range of the learning rate is 0.01 to 0.8. Therefore, we need to define the optimized learning rate: if the learning rate is less than 0.01, it is taken as 0.01; if it is greater than 0.8, it is taken as 0.8. Therefore, the obtained learning rate is a real number (0.01, 0.8).

[0145] Based on the genetic operator design, the connected genes of each individual are subjected to one-point crossover and basic mutation. The weight threshold gene and rate gene in the parameter gene do not interfere with each other and each uses arithmetic crossover and non-uniform mutation. The fitness function is the total error of the network. The genetic operations for each chromosome based on the GA+BP algorithm are shown in Table 1.

[0146] Table 1

[0147]

[0148] We conducted simulation experiments using a GA+BP neural network on 80 selected learning samples. To maintain generality, we also conducted ten experiments using the GA+BP neural network. The results of these ten experiments, including the number of hidden layer nodes, learning rate, number of training convergence steps, runtime, and network error, are shown in Table 2.

[0149] Table 2

[0150]

[0151] In terms of workload: Using genetic algorithms to optimize the BP neural network not only obtains the optimal network structure, but also obtains the optimal initial weights, thresholds and learning rates corresponding to the structure. In this way, the structure and parameters of the network are completely obtained by computer calculation, reducing the workload of the experimenter.

[0152] In terms of training performance: As shown in Table 2, there is no training failure in the GA+BP neural network, so the fault diagnosis is more reliable.

[0153] Network Error: As can be seen from Table 2, the network error of the GA+BP neural network is almost always below 0.1. Calculation shows that the average network error of the ten diagnostic simulations is 0.0672.

[0154] In summary, the GA+BP neural network can accurately and reliably complete the fault diagnosis of aircraft engines.

[0155] Example 2:

[0156] An aircraft engine combustion chamber fault diagnosis device, comprising:

[0157] A signal acquisition unit is used to collect the fault signal of the combustion chamber of the aircraft engine, perform wavelet packet decomposition and reconstruction on the signal, and extract the energy characteristic vector of the signal;

[0158] A neural network construction unit is used to construct a BP neural network, optimize the structure and parameters of the BP neural network using a genetic algorithm to obtain a GA-BP neural network, and train the GA-BP neural network;

[0159] The signal processing unit is used to input the energy feature vector of the extracted signal into the trained GA-BP neural network to obtain the fault type.

[0160] Example 3:

[0161] An electronic device comprises a processor and a memory connected to the processor for storing instructions executable by the processor, wherein the processor is used to execute the above-mentioned method for diagnosing a combustion chamber fault of an aircraft engine.

[0162] Example 4:

[0163] A computer-readable storage medium stores a computer program, which implements the above-mentioned method for diagnosing a combustion chamber fault of an aircraft engine when executed by a processor.

[0164] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for diagnosing a combustion chamber fault of an aircraft engine, characterized in that: The steps include: Step 1: Collect the fault signal of the combustion chamber of the aircraft engine, decompose and reconstruct the signal by wavelet packet, and extract the energy feature vector of the signal; Step 2: Construct a BP neural network, use genetic algorithm to optimize the structure and parameters of the BP neural network, obtain a GA-BP neural network, and train the GA-BP neural network; Step 3: Input the energy feature vector of the extracted signal into the trained GA-BP neural network to obtain the fault type.

2. The method for diagnosing a combustion chamber fault of an aircraft engine according to claim 1, characterized in that: The aircraft engine combustion chamber fault signal in step 1 includes a pressure signal, a temperature signal, a vibration signal, and a flame signal, which are used to reflect the pressure fluctuation, temperature distribution, vibration condition, and flame state information inside the combustion chamber.

3. The method for diagnosing a combustion chamber fault of an aircraft engine according to claim 1, characterized in that: The step 1 also includes processing the collected signal, including filtering, denoising, amplifying, discretizing and performing wavelet packet denoising on the signal.

4. The method for diagnosing a combustion chamber fault of an aircraft engine according to claim 1, characterized in that: In step 1, performing wavelet packet decomposition and reconstruction on the signal, and extracting the energy feature vector of the signal includes: (1) Normalization processing of fault signals: The data normalization method is: x i (t)′=2(x i (t)-minx i (t)) / (maxx i (t)-minx i (t))-1; In the formula, x i (t) represents the original data, x i (t)′ represents the normalized data, maxx i (t) and minx i (t) represents the maximum and minimum values ​​of all data of the same input node; (2) Decompose the signal into N layers of wavelet packets, using X Nj Respectively represent the Nth layer low frequency to high frequency 2 N The decomposition coefficient vector of the frequency band is obtained from the Nth layer from low frequency to high frequency 2 N The characteristic signal S of the frequency band Nj , j=1,2,…,2 N ; (3) Calculate the energy of the signal in each frequency band: (4) Construct the feature vector: For E Nj Perform normalization: The energy eigenvector is thus determined as 5. The method for diagnosing a combustion chamber fault of an aircraft engine according to claim 1, characterized in that: In step 2, a genetic algorithm is used to optimize the structure and parameters of the BP neural network, and the GA-BP neural network is obtained, which includes: Initialize the structure, weights, thresholds and learning rate of the BP neural network; The genetic algorithm optimization capability is used to optimize the structure, weights, thresholds and learning rates of the BP neural network.

6. The method for diagnosing a combustion chamber fault of an aircraft engine according to claim 5, characterized in that: The optimization of the structure, weight, threshold and learning rate of the BP neural network by using the genetic algorithm optimization capability includes: Genetic Encoding and Decoding: Decompose each chromosome into connection genes and parameter genes; The connection gene corresponds to the number of hidden layer nodes of the neural network. The binary encoding method is adopted. The encoding string consists of binary "0" and "1". Each binary number in the encoding string represents a hidden layer neuron. "1" means that the neuron exists, and "0" means that it does not exist. The binary encoding length is equal to the number of hidden layer nodes. The parameter gene adopts real number coding. Each connection weight, threshold and learning rate are directly represented by a real number. The parameter gene is divided into two parts: weight threshold gene and rate gene. The coding length of weight threshold gene is equal to the total number of weights and thresholds in each layer of the neural network. The coding length of rate gene is one. The coding length of parameter gene is the sum of the length of weight threshold gene and rate gene. The total coding length of a chromosome is the sum of the coding length of connection gene and the coding length of parameter gene. Design fitness function: The fitness function of each chromosome is taken as f = 1 / (1+E); In the formula is the total error in the neural network, represents the ideal output, represents the true output, K is the number of sample sets; Design genetic operators, including selection, crossover, and mutation: The selection operation adopts a method combining the best individual preservation and fitness ratio; The connection gene adopts a one-point crossover method, and the weight threshold gene and rate gene in the parameter gene adopt an arithmetic crossover method. The two parts cross each other without interfering with each other. The connection gene adopts the basic mutation mode, and the weight threshold gene and rate gene in the parameter gene adopt the non-uniform mutation mode, but the two parts mutate independently without interfering with each other; Select the control parameters.

7. An aircraft engine combustion chamber fault diagnosis device, characterized in that: include: A signal acquisition unit is used to collect the fault signal of the combustion chamber of the aircraft engine, perform wavelet packet decomposition and reconstruction on the signal, and extract the energy characteristic vector of the signal; A neural network construction unit is used to construct a BP neural network, optimize the structure and parameters of the BP neural network using a genetic algorithm to obtain a GA-BP neural network, and train the GA-BP neural network; The signal processing unit is used to input the energy feature vector of the extracted signal into the trained GA-BP neural network to obtain the fault type.

8. An electronic device, comprising a processor and a memory connected to the processor for storing instructions executable by the processor, characterized in that: The processor is used to execute an aircraft engine combustion chamber fault diagnosis method as described in any one of claims 1-6 above.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for diagnosing a combustion chamber fault of an aircraft engine as described in any one of claims 1 to 6 is implemented.

Images