Supersonic inlet flow pattern identification method based on hybrid mutation non-dominated sorting genetic algorithm

Abstract

The present invention provides a supersonic inlet flow state identification method based on a hybrid variation non-dominated sorting genetic algorithm. According to the characteristics of the pressure data measured by the sensor in the supersonic inlet, the present invention performs time-mean processing on the pressure data and extracts features from both the time domain and the frequency domain. For the obtained feature set, the present invention provides a hybrid variation non-dominated sorting genetic algorithm, which optimizes the hyperparameters of a support vector machine as a classifier while performing feature selection. In order to solve the population limitation of a single-objective evolutionary algorithm used for a multi-objective evolutionary algorithm in a hybrid variation algorithm, the present invention provides a new population out-of-bounds processing method. Compared with the traditional multi-objective optimization algorithm, the single-objective optimization algorithm has advantages in both population richness and accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of supersonic inlet flow pattern recognition, and in particular to a supersonic inlet flow pattern recognition method based on a hybrid mutation non-dominated sorting genetic algorithm. Background Art

[0002] Air-breathing supersonic propulsion technology has become one of the research hotspots in the international aerospace field due to its great potential in the military and commercial fields. The air inlet is a key aerodynamic component of the air-breathing supersonic propulsion system and a core factor in the integration of the airframe or propulsion system. Its design form and working characteristics directly affect the overall performance of the propulsion system. According to whether the change of the flow pattern in the supersonic inlet affects its flow capture characteristics, the working mode can be divided into a starting state and a non-starting state. When the supersonic inlet is in the non-starting state, the total pressure recovery coefficient and the flow coefficient are significantly reduced, and violent wave motion and wall pressure oscillation are prone to occur. This not only deteriorates the thrust characteristics of the engine, but also causes structural damage to the engine due to periodic force loads and thermal loads, making the aircraft difficult to control. Therefore, it is extremely important to quickly and accurately identify the inlet flow pattern, so as to guide the inlet or engine to make corresponding adjustments to the rapid restart of the inlet.

[0003] In the identification of the non-starting state of the supersonic inlet, the current research work is relatively scattered, and most of the research directly draws on the factors that cause the non-starting state. For example, Hawkins WR et al. divided the non-starting state into two categories: too large area contraction ratio and too high back pressure, Yu DR et al. divided it into two categories: too low Mach number and too high back pressure, and Cox C et al. divided it into three categories: too high back pressure, too large contraction ratio, and too large turning angle. Specifically, Hawkins WR et al. conducted an exploratory study on the identification method of the non-starting state of the binary supersonic inlet in response to the needs of wind tunnel experiments, and proposed and practiced three non-starting state identification methods based on high-speed schlieren photography, the time-averaged static pressure relationship between the front and back of the throat, and the transient static pressure characteristics of the wall near the throat. The first method is intuitive and reliable, but requires a complex optical path system. It is only feasible in ground experiments and is obviously unrealistic in actual flight. The second method is simple to implement, but the discrimination criteria vary depending on the specific inlet design parameters and flight Mach number, making it difficult to be universal. When there is a difference between the actual parameters and the hypothetical values ​​when establishing the discrimination criteria, it is easy to make a misjudgment. The third method is to identify the inlet non-starting state from an unsteady perspective. Its main advantage is that it is based on the time domain and frequency domain characteristics of the dynamic pressure signal, and is less affected by the time-averaged static pressure prediction error. Therefore, the established identification method is expected to achieve higher accuracy and has universality. It is a new direction for the current research on non-starting state identification. Its main disadvantage is that the amount of information and calculation required is relatively large, and the difficulty of constructing the algorithm is relatively high. The feature extraction of the signal, the establishment of the discrimination criteria, the selection of the signal sampling frequency, the arrangement of the sensor, etc. all rely significantly on the understanding of the detection object. There is a large research gap in the feature extraction, feature selection, and flow pattern identification of the inlet flow pattern pressure data.

[0004] Non-dominated sorting genetic algorithms (NSGA-II) have become one of the most effective multi-objective evolutionary algorithms and have been successfully applied to various multi-objective optimization problems. Feature selection and hyperparameter optimization have always been hot topics in multi-objective optimization problems. Multi-objective optimization algorithms often obtain a set of optimal solutions when dealing with multi-objective optimization problems, which is different from the one optimal solution in single-objective optimization algorithms. According to the preference for the objective function, we can select from the optimal solution set to enrich the diversity of the solution set. The mutation operator plays a crucial and decisive role in the diversity of the solution set. The traditional real-coded NSGA-II algorithm uses polynomial mutation, and there are few studies on the use of other mutation operators. Some studies have tried to use single-objective mutation algorithms for multi-objective evolutionary processes. For example, Laumanns M et al. applied some parameter control methods of adaptive mutation intensity in single-objective optimization problems to multi-objective conditions and found that the method based on Pareto selection would lead to a low success rate and the algorithm would be difficult to converge to the Pareto set. Basseur M et al. used multiple mutation operators simultaneously to solve the ship deployment problem, and dynamically changed their ratios by comparing the statistical results of the dominance relationship between the offspring generated by each operator and the parent individuals, and selected the mutation operator with the highest ratio. Basseur M et al. improved the above by comparing the statistical results of the frontier ranking values ​​of the offspring generated by each operator and the parent individuals to dynamically change their ratios, solving the inaccurate situation when only the dominance relationship is used for statistics when the parent individuals and the offspring individuals do not dominate each other. Aggarwal V et al. designed a correlation-sensitive mutation operator using the structural characteristics of the circuit optimization problem, and demonstrated through two circuit examples that the mutation operator provides faster and better optimization performance than the traditional operator. It can be seen that researchers have focused their attention on the construction of non-dominated sets, fitness assignment, population maintenance and other issues, while there is little research on the impact of mutation operators on multi-objective optimization. Summary of the invention

[0005] Purpose of the invention:

[0006] In order to break through the limitations of the traditional supersonic inlet non-starting state recognition technology, the present invention clarifies the direction of identifying the inlet non-starting state from an unsteady perspective, takes the time domain and frequency domain characteristics of the dynamic pressure signal as the main basis, adopts a non-dominated sorting genetic algorithm with mixed mutation, and applies it to the selection of time and frequency domain features and the optimization of support vector machine hyperparameters, so as to realize the recognition of the supersonic inlet non-starting state, further enhance the real-time monitoring of the inlet flow state, and improve the engine safety performance. In addition, the present invention combines two single-objective mutation operators with traditional polynomial mutation, compares the effects of different mutation operators on multi-objective optimization, improves the classification performance of the multi-objective optimization algorithm and the richness of the optimal solution set, and is more conducive to the location selection of the sensor in the inlet.

[0007] Technical solution:

[0008] A supersonic inlet flow pattern identification method based on a hybrid mutation non-dominated sorting genetic algorithm comprises the following steps:

[0009] Step 1: Perform time-average processing on the pressure data measured by the sensor arranged in the supersonic inlet, and divide the pressure data at certain time intervals to obtain time series of fixed length. These time series correspond to the flow state in the supersonic inlet in different time periods, and then extract the flow state characteristics from each time series in both time domain and frequency domain according to the existing theoretical formula to obtain feature subsets; divide the feature set composed of feature subsets into training set, verification set and test set;

[0010] Step 2: Using the non-dominated sorting genetic algorithm NSGA-II as the benchmark algorithm, initialize the population size, number of iterations, crossover probability, mutation probability and other parameters, initialize the population and calculate the corresponding objective function value, and then perform non-dominated sorting on the initial population.

[0011] Step 3: Use the traditional tournament selection operator to select the parent population. Since the population encoding uses real number encoding, the crossover operation uses the simulated binary crossover technology SBX. Different from the traditional polynomial mutation operation, the mutation operator that combines polynomial mutation, Gaussian mutation, and Cauchy mutation is used. The three mutation operators operate in parallel to obtain the offspring population.

[0012] Step 4: Merge the obtained offspring population with the parent population and perform a non-dominated sorting operation, and then fill the population on each frontier surface according to the crowding distance until the target population size is met.

[0013] Step 5: Determine whether the number of iterations meets the requirement. If so, end the algorithm; otherwise, return to step 3.

[0014] Preferably, the flow state features extracted from the time domain in step 1 include:

[0015]

[0016]

[0017]

[0018] Where x(n) is the signal sequence when n=1,2,...,N, and N is the number of data points in a time series.

[0019] Preferably, the flow state features extracted from the frequency domain in step 1 include:

[0020]

[0021]

[0022]

[0023]

[0024] Where s(k) is the spectrum when k = 1, 2, ... K, K is the number of spectral lines, f k is the frequency value of the kth spectral line in a time series.

[0025] Preferably, in step 2, the encoding method used to initialize the population mainly includes the following contents:

[0026] The population coding of the hybrid mutation non-dominated sorting genetic algorithm mainly includes two aspects: time-frequency domain feature selection and SVM hyperparameter optimization. The hyperparameters are mainly the misclassification penalty coefficient C and the RBF kernel width γ. In order to ensure the uniformity of the coding, the real number coding method is adopted. In the feature selection part, each individual is a random number from 0 to 1. The random number greater than 0.5 represents being selected, otherwise it represents that the feature is discarded. The initialization range of the hyperparameter is [2 -10 ,2 10 ]. Taking a D-dimensional data set as an example, the i-th individual in the population is encoded as:

[0027] X i (t) = {x i,1 ,x i,2 ,...,x i,j}

[0028] x i,j ∈[0,1],j=1,2,...,D-2,i=1,2,...,N

[0029] x i,j ∈[2 -10 ,2 10],j=D-1,D,i=1,2,...,N

[0030] Where X i (t) represents the i-th individual in the t-th iteration, N represents the population size, x i,j Represents the code of the jth gene in the ith individual.

[0031] Preferably, in step 2, the objective function value is introduced as follows:

[0032] 1) Classification error.

[0033] 2) Proportion of selected features: The number of selected features is the number of feature subsets encoded as 1 in the individuals of the population.

[0034] Preferably, in step 3, the three mutation operators are defined as follows:

[0035] 1) Polynomial mutation. Polynomial mutation, as a traditional real-number-coded multi-objective evolutionary mutation operator, can be defined as: 1j (t) = x 1j (t)+Δ j ,in η u represents the distribution coefficient, u j Represents a random number between 0 and 1. 1j (t) represents the jth part of the offspring participating in the mutation operation in the tth iteration;

[0036] 2) Gaussian mutation. This mutation operator was originally used in evolutionary strategies. The one-dimensional Gaussian random number N(μ,σ) is defined as:

[0037]

[0038] Among them, μ represents the mean and σ represents the standard deviation. The form of Gaussian variation is as follows:

[0039] v' k =v k +N(0,σ)

[0040] Among them, N(0,σ) represents a one-dimensional standard distribution random number with mean 0 and standard deviation σ, v k represents the part of the pre-mutated individual, v' k represents the components in the individual after mutation, and the upper and lower limits of the components in the individual are defined as [l k ,u k ].

[0041] 3) Cauchy variation. The density function of a one-dimensional Cauchy distribution random number centered at the origin is defined as:

[0042]

[0043] Among them, σ is a measurement parameter. Compared with Gaussian distribution, Cauchy distribution is more likely to generate random numbers far away from the origin, which helps prevent the algorithm from falling into local optimality. The form of Cauchy mutation is as follows:

[0044] v' k =v k +C(0,σ)

[0045] Among them, C(0,σ) represents a Cauchy distributed random number centered at 0 and σ is the measurement parameter.

[0046] Beneficial effects:

[0047] The non-dominated sorting genetic algorithm based on hybrid mutation not only solves the problem of time-frequency domain feature screening, but also optimizes the hyperparameters of the classifier, so that the classifier based on the optimal feature set has higher classification accuracy and shorter classification time. This multi-objective optimization algorithm combines the traditional multi-objective mutation operator with the single-objective mutation operator. The three mutation operators complement each other, broaden the optimal solution set space, enrich the diversity of the population, and ensure that the algorithm achieves higher classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a supersonic inlet experimental model;

[0049] Figure 2 is the pressure sensor signal under different flow states;

[0050] Figure 3 The process of hybrid mutation non-dominated sorting genetic algorithm;

[0051] Figure 4 Design for population coding;

[0052] Figure 5 The optimal frontiers (C01-C17) of the multi-objective optimization algorithm NSGA-II using three mutation operators respectively;

[0053] Figure 6 The optimal frontiers (R01-R14) of the multi-objective optimization algorithm NSGA-II using three mutation operators respectively;

[0054] Figure 7 This is a comparison chart of the maximum test accuracy;

[0055] Figure 8 This is a comparison chart of the selected feature proportions;

[0056] Fig. 9 This is a comparison chart of the algorithm's classification time on the sensor C01-C17 test set;

[0057] Fig.10 This is a comparison chart of the classification time of the algorithm on the sensor R01-R14 test set;

[0058] Fig.11 This is the flow chart of state recognition technology. DETAILED DESCRIPTION

[0059] The experimental model adopts an external pressure supersonic inlet model such as Figure 1 As shown, the sensors are evenly distributed on the upper and lower walls of the intake duct and are numbered C01-C17 and R01-R14 respectively. The pressure signals of sensors R01 and R14 are as follows: Figure 2 As shown in Figure 2, with the development of time and throttle ratio (TR), the flow pattern in the tube undergoes four modes: steady state, small surge, mixed surge and large surge.

[0060] According to the needs and characteristics of inlet flow pattern recognition, the pressure data measured by the sensor is processed by time average, and the samples are cut at certain time intervals to obtain time series of fixed length. These time series correspond to the flow patterns in the supersonic inlet at different time periods, and then the flow pattern features are manually extracted from the time domain and frequency domain according to the formula. The obtained feature set is divided into a ratio of 6:2:2 to obtain the training set, validation set, and test set.

[0061] The hybrid mutation non-dominated sorting genetic algorithm proposed in the present invention takes the non-dominated sorting genetic algorithm NSGA-II as the benchmark algorithm and adopts mutation operators combining polynomial mutation, Gaussian mutation and Cauchy mutation. The algorithm flow is as follows: Figure 3 As shown. First, the population size is initialized to 30, the maximum number of iterations is 100, the crossover probability is 0.9, the mutation probability is 0.1, the distribution coefficient in the crossover operation is initialized to 1, and the distribution coefficient in the mutation operation is initialized to 10. Initialize the population and calculate the corresponding objective function value, perform non-dominated sorting on the population and calculate the crowding distance. The encoding method of the population is as follows Figure 4 As shown. Use the traditional tournament selection operator to select the parent population. Since the population encoding uses real number encoding, the crossover operation uses simulated binary crossover technology. Different from the traditional polynomial mutation operation, a mutation operator combining polynomial mutation, Gaussian mutation, and Cauchy mutation is used. The three mutation operators operate in parallel to obtain the offspring population. The obtained offspring population is merged with the parent population, and a non-dominated sorting operation is performed. Then, the population is filled on each frontier surface according to the crowding distance until the target population size is met. Determine whether the number of iterations is met. If it is met, the algorithm ends, otherwise it continues to iterate. Finally, the optimal individual set under the simultaneous optimization of SVM hyperparameters and feature subsets is obtained. Combining the above description of the implementation method, the state recognition technology process is obtained as follows Fig.11 shown.

[0062] This case compares the superiority of the hybrid mutation non-dominated sorting genetic algorithm and the effectiveness of the inlet duct identification technology from two aspects: the impact of different mutation operators on the optimal solution set of multi-objective optimization and the performance of the sub-classifier. In terms of the impact of different mutation operators on the optimal solution set of multi-objective optimization, the performance of the multi-objective optimization algorithm NSGA-II using polynomial mutation, Gaussian mutation, and Cauchy mutation on 31 sensor data is compared. The performance mainly includes classification time, maximum classification accuracy, average classification accuracy, and the proportion of selected features corresponding to the solution with maximum classification accuracy. The experimental results are shown in Tables 1 and 2. Figure 5 , Figure 6 As shown. On sensors C01-C17, the NSGA-II algorithm using the Cauchy mutation operator is significantly better than the algorithms using the other two mutation operators in terms of maximum classification accuracy. In terms of the proportion of selected features corresponding to the maximum accuracy solution, the polynomial mutation has a greater advantage. On sensors R01-R14, the NSGA-II algorithm using the Gaussian mutation operator achieves higher classification accuracy than the other two algorithms. The effects of the Cauchy mutation operator and the polynomial mutation operator are comparable, both in terms of maximum classification accuracy and in terms of the proportion of selected features. Figure 5 and Figure 6 In the figure, it can be observed that the distribution effect of the Pareto optimal frontier obtained by using the single-objective evolutionary mutation operator in the multi-objective evolutionary algorithm is comparable to that of the optimal frontier obtained by the multi-objective evolutionary algorithm using the traditional polynomial mutation. The optimal solution set obtained by using the Cauchy mutation has a smaller proportion of selected features in the classification error range of 0 to 0.02 than the other two mutation algorithms. The frontiers obtained by the three mutation operators all have crossover phenomena, so it is difficult to weigh the optimal method by considering the two objective function values ​​at the same time. The above proves the necessity of using a hybrid mutation algorithm.

[0063] In terms of subclassifier performance, the NSGA-II algorithm with mixed mutation, the NSGA-II algorithm with polynomial mutation (PM-NSGA-II), the NSGA-II algorithm with Gaussian mutation (GM-NSGA-II), and the NSGA-II algorithm with Cauchy mutation (CM-NSGA-II) are compared with the particle swarm algorithm. The comparison results are shown in Figure 2. Figure 7 , Figure 8 , Fig. 9 , Fig.10 As shown. It can be seen that the mixed variation NSGA-II algorithm proposed in the present invention has great advantages in terms of maximum test accuracy and selected feature proportion when performing inlet flow state recognition, while ensuring that the average classification time of the classifier based on the feature subset selected by the algorithm is small. The above proves the effectiveness of the algorithm provided by the present invention.

[0064] Table 1 Performance of NSGA-II algorithm using three mutation operators on sensors C01-C17

[0065]

[0066] Table 2 Performance of NSGA-II algorithm using three mutation operators on sensors R01-R14

[0067]

[0068]

Claims

1. A supersonic inlet flow pattern identification method based on a hybrid mutation non-dominated sorting genetic algorithm, characterized in that: The steps include: Step 1: Perform time-average processing on the pressure data measured by the sensor arranged in the supersonic inlet, and divide the pressure data into time series of fixed length. These time series correspond to the flow state in the supersonic inlet in different time periods, and then extract the flow state features from each time series in both time domain and frequency domain to obtain feature subsets; Divide the feature set consisting of feature subsets into a training set, a validation set, and a test set; Step 2: training the classifier with the training set, optimizing the classifier hyperparameters and feature subsets using a hybrid mutation non-dominated sorting genetic algorithm, and finally determining the flow state corresponding to the feature subset as the supersonic inlet flow state; The flow characteristics extracted from the time domain in step 1 include: X peak =max|x(n)|; Where x(n) is the signal sequence when n=1,2,...,N, and N is the number of data points in a time series; The flow characteristics extracted from the frequency domain in step 1 include: Where s(k) is the spectrum when k = 1, 2, ... K, K is the number of spectral lines, f k is the frequency value of the kth spectral line in a time series.

2. The method for identifying supersonic inlet flow pattern based on hybrid mutation non-dominated sorting genetic algorithm according to claim 1, characterized in that: The specific steps of step 2 include: Step 2.1: Initialize the parameters of the hybrid mutation non-dominated sorting genetic algorithm, encode the classifier hyperparameters and feature subsets, and create an initial population. The individuals in the initial population contain a feature subset and the selected classifier hyperparameters; Step 2.2: training the classifier with each feature subset in the training set, calculating the objective function value of the hybrid mutation non-dominated sorting genetic algorithm, and then performing non-dominated sorting on the initial population; Step 2.3: Select a parent population from the initial population, use three mutation operators, polynomial mutation, Gaussian mutation, and Cauchy mutation, to perform parallel operations, and then obtain a child population; Step 2.4: Merge the obtained offspring population with the parent population and perform a non-dominated sorting operation, and then fill each frontier surface with population according to the crowding distance until the target population size is met; Step 2.5: Determine whether the number of iterations meets the requirement. If so, end the algorithm; otherwise, return to step 2.

3.

3. The method for identifying supersonic inlet flow pattern based on hybrid mutation non-dominated sorting genetic algorithm according to claim 1, characterized in that: The classifier adopts SVM.

4. The method for identifying supersonic inlet flow pattern based on hybrid mutation non-dominated sorting genetic algorithm according to claim 2, characterized in that: In step 2.1, the classifier hyperparameters and feature subsets are encoded using real number encoding; each individual in the initial population is a random number between 0 and 1, and a random number greater than 0.5 indicates that it is selected, otherwise it indicates that the feature is discarded; the initialization range of the hyperparameters is [2 -10 ,2 10 ].

5. The method for identifying supersonic inlet flow pattern based on hybrid mutation non-dominated sorting genetic algorithm according to claim 4 is characterized in that: In step 2.2, the objective function value includes: 1) Classification error of the classifier; 2) Proportion of selected features: The number of selected features is the number of feature subsets coded as 1 in the individuals of the population.

6. The method for identifying supersonic inlet flow pattern based on hybrid mutation non-dominated sorting genetic algorithm according to claim 2 is characterized in that: In step 2.3, the definitions of the three mutation operators are as follows: 1) Polynomial mutation; defined as: x 1j (t) = x 1j (t)+Δ j ,in η u represents the distribution coefficient, u j Represents a random number between 0 and 1; x 1j (t) represents the jth part of the offspring participating in the mutation operation in the tth iteration; 2) Gaussian variation; the one-dimensional Gaussian random number N(μ,σ) is defined as: Among them, μ represents the mean and σ represents the standard deviation; the form of Gaussian variation is as follows: v' k =v k +N(0,σ) Among them, N(0,σ) represents a one-dimensional standard distribution random number with mean 0 and standard deviation σ, v k represents the part of the pre-mutated individual, v' k represents the components in the individual after mutation, and the upper and lower limits of the components in the individual are defined as [l k ,u k ]; 3) Cauchy variation; The density function of a one-dimensional Cauchy distribution random number with the origin as the center is defined as: Among them, σ is the measurement parameter; the form of Cauchy variation is as follows: v' k =v k +C(0,σ) Among them, C(0,σ) represents a Cauchy distributed random number centered at 0 and σ is the measurement parameter.

Images