A pattern recognition feature selection optimization method based on simulated annealing algorithm
Through the simulation annealing algorithm, the feature selection is optimized, combined with Fisher score mapping and adaptive threshold factor, the problem of large amount of feature selection and low accuracy in phase-sensitive photoreflectometers is solved, and more efficient feature screening and pattern recognition accuracy is achieved.
Patent Information
- Application Number
- CN202310902340.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-21
- Publication Date
- 2025-08-29
- Estimated Expiration
- 2043-07-21
AI Technical Summary
The existing feature selection methods have problems such as huge calculation amount, large result deviation or slow speed in phase-sensitive photoreflectors, especially in high-dimensional intrusion signal feature vectors, which leads to low pattern recognition accuracy.
The feature selection optimization method based on the simulated annealing algorithm is adopted, and the Fisher score mapping and adaptive threshold factor are combined with the LIBSVM classification accuracy as the appropriate function. Metropolis is used to accept the new solution criteria and the temperature attenuation mechanism to achieve global convergence and rapid convergence, and select the optimal intrusion signal feature combination.
The pattern recognition classification accuracy of phase-sensitive photo-time domain reflectometer is improved, the probability of missing selection of feature combinations is reduced, and the local optimal solution is jumped out of the process, achieving more efficient feature selection.
Smart Images
Figure CN116933867B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of pattern recognition of a phase-sensitive optical time domain reflectometer, and particularly relates to an optimization method for pattern recognition feature selection based on an analog annealing algorithm. Background Art
[0002] The distributed optical fiber sensing system based on a phase-sensitive optical time domain reflectometer has significant advantages such as high sensitivity, strong reliability, and low cost, and has been widely applied in the fields of industrial production process detection, oil and gas pipeline maintenance, vibration quantitative measurement, building structure health monitoring, perimeter security, etc. The phase-sensitive optical time domain reflectometer can determine the position of an intrusion signal by analyzing the backscattered Rayleigh optical signal, and then accurately judge and identify different intrusion signals through pattern recognition. Among them, in the research on pattern recognition of intrusion signals collected by the phase-sensitive optical time domain reflectometer, a series of data to be processed are often described by the feature vectors of high-dimensional intrusion signals. Each component of these high-dimensional intrusion signal feature vectors represents a certain feature of the data. In fact, among the features of numerous intrusion signals, some features are irrelevant to solving the problem, while some features are redundant. In order to improve the efficiency of subsequent calculations (such as classification, prediction, etc.) and save the storage space of data, it is necessary to screen out these irrelevant and redundant features from the feature vectors of high-dimensional intrusion signals and select the features that are truly important for solving the problem.
[0003] Feature selection requires selecting d (d < M) features that are most conducive to subsequent operations from M original features (that is, the dimension of the original data is M). For the exhaustive search method [Hyafil L, et al. Information Processing Letters, 5(1): 15 - 172, 1976], the optimal feature combination can be selected. Since each set of original features has (2 M - 1) (excluding the empty set) feature combinations, the results of subsequent operations on these (2 M - 1) feature combinations can be run out and compared to select the feature combination that can obtain the best result. This is practical for feature selection problems with a relatively small M value. However, as M increases, at least (2 M - 1) calculations need to be performed and the calculation results need to be sorted, which makes the computational complexity of feature selection extremely large.
[0004] In order to solve this problem, researchers have tried to find feasible feature selection methods. According to different feature evaluation criteria and evaluation methods, many feature selection methods can be roughly divided into two categories: filtering and encapsulation. Filtering-based feature selection methods such as mutual information [D. Koller, et al. In ICML, 284-292, 1996] and Fisher Score [RODuda, et al. Wiley-Interscience Publication, 2001] mainly use certain criteria (evaluation functions) to give each feature a "score", and then set a threshold θ to select features with scores higher than the threshold. However, such methods usually select redundant features and easily filter out some feature combinations that should be retained. Encapsulation-based feature selection methods such as hill climbing and forward selection [Battiti, et al. IEEE Trans, 5(4): 537–550, 1994] directly use the performance of subsequent operations (for example, the accuracy of subsequent classification) as the evaluation criteria of the feature combination to encapsulate the feature selection in the search process. The accuracy of encapsulated feature selection is higher than that of filtered feature selection, but since a subsequent operation must be run for each feature combination, the operation speed of this type of method is much slower than that of the filtering method.
[0005] The above-mentioned filtering and encapsulation feature selection methods are mainly characterized by the following characteristics: the filtering feature selection method runs fast, but the results of subsequent operations have large deviations; while the encapsulation feature selection method makes the results of subsequent operations more accurate but runs slowly. Therefore, researchers naturally think of combining the two types of methods for feature selection and propose hybrid feature selection methods, such as the Relief-Wrapper method and the Relief-GA-Wrapper method. The above-mentioned methods remove irrelevant features by selecting features with high relief scores and filtering out features with the lowest relief scores. However, sometimes different features have low relief scores, but their combination performs well in subsequent operations, making it impossible to see the overall effect of the feature combination. In addition, in the subsequent encapsulation method, in order to reduce the amount of computation and complexity, the solution will be expanded from a local perspective, resulting in the result being at the local extreme value rather than the global minimum, thus falling into a local optimum. Summary of the Invention
[0006] To address the above shortcomings, the present invention proposes a pattern recognition feature selection optimization method based on a simulated annealing algorithm to ensure the global convergence of the algorithm, accelerate the convergence speed, reduce the probability of missing intrusion signal feature combinations, escape from local optimality, and improve the pattern recognition classification accuracy of phase-sensitive optical time-domain reflectometry.
[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a pattern recognition feature selection optimization method based on a simulated annealing algorithm, comprising the following steps:
[0008] S1. Encode the intrusion signal characteristic parameters and then initialize the parameters;
[0009] S2. Calculate the Fisher score of each intrusion signal feature value;
[0010] S3, performing linear transformation on the calculated Fisher score and comparing it with the threshold factor to obtain the initial intrusion signal feature combination;
[0011] S4. Determine the fitness function;
[0012] S5. Calculating an initial fitness function value based on the initial intrusion signal feature combination;
[0013] S6. Generate a new intrusion signal feature combination based on the inversion of the G feature vector codes in the previous intrusion signal feature combination, and calculate the fitness function value of the new intrusion signal feature combination;
[0014] S7. If the fitness function value of the new intrusion signal feature combination is greater than the fitness function value solution of the current intrusion signal feature combination, then the new solution is directly accepted; otherwise, the Metropolis new solution acceptance criterion is used to determine whether to accept the new solution;
[0015] S8. Repeat steps S6 to S7 at temperature T until the number of iterations is reached.
[0016] S9, update the temperature according to the temperature attenuation coefficient; repeat steps S6-S8 until the temperature is less than or equal to the end temperature T min When , the optimal intrusion signal feature combination is determined according to the optimal fitness function value.
[0017] In step S1, the intrusion signal characteristic parameters are encoded using 0 / 1, where 1 represents selected and 0 represents unselected.
[0018] In step S2, the calculation formula of the Fisher score of the intrusion signal characteristic value is:
[0019]
[0020] Among them, n k represents the number of samples in the kth category, f j,i represents the value of the jth sample in the i-th feature, μ fi represents the i-th feature f i The average value of all samples in , μkfi represents the i-th feature f iThe average value of all samples belonging to the kth category, c represents the number of categories of samples, F(f i ) represents the i-th feature f i Fisher score.
[0021] The specific method of step S3 is:
[0022] S301. Map the Fisher score of each intrusion signal characteristic value to the closed interval [ε, 1-ε], where ε represents a positive real number parameter and 0<ε<0.1. The calculation formula is:
[0023]
[0024] Among them, L(F(f i )) represents the i-th feature f i The mapping value of Fisher score; F max 、F min Respectively represent the maximum and minimum values of the Fisher score of each intrusion signal feature value;
[0025] S302. Randomly generate M real-number threshold factors ω i ,ω i ∈U(0,1); M represents the number of feature parameters to be selected;
[0026] S303: Determine the mapping value of the Fisher score of each feature and the corresponding threshold factor ω i If L(F(f i ))>ω i , then select the intrusion signal feature; otherwise, L(F(f i ))≤ω i , then the intrusion signal feature is not selected, i = 1, 2, ..., M, and the selected intrusion signal feature is used as the initial intrusion signal feature combination.
[0027] In step S4, the LIBSVM classification accuracy of the high-dimensional intrusion signal feature vector is used as the fitness function.
[0028] In step S6, the new intrusion signal feature combination is generated by:
[0029] S601. Randomly generate M real numbers ρ i ,ρ i ∈U(0,1);
[0030] S602: Determine the mapping value of the Fisher score of each feature and the real number ω i If L(F(f i ))>ρ i, then select the intrusion signal feature; otherwise, L(F(f i )) ≤ ρ i , then do not select the intrusion signal feature, i = 1, 2, …, M; L(F(f i )) represents the mapped value of the Fisher score of the i-th feature f i ;
[0031] S603. Randomly select G feature vectors from the selected intrusion signal features, and perform an inversion operation on the encoding of these G feature vectors in the previous intrusion signal feature combination to obtain a new intrusion signal feature combination.
[0032] The judgment method in step S7 is specifically as follows:
[0033] (1) If the fitness function value of the new intrusion signal feature combination is greater than the fitness function value solution of the current intrusion signal feature combination, then update the intrusion signal feature combination;
[0034] (2) If the fitness function value of the new intrusion signal feature combination is less than or equal to the fitness function value solution of the current intrusion signal feature combination, then calculate the probability. The calculation formula is:
[0035]
[0036] T represents the current temperature, h(S′) represents the fitness function value of the new feature combination, and h(S) represents the fitness function value of the current feature combination; then take a random number 0 < r < 0.1. If r < P, then update the feature combination, otherwise do not update.
[0037] In step S9, the method for updating the temperature is:
[0038] T′ = KT;
[0039] Among them, T represents the temperature before update, T′ represents the temperature after update, and K represents the temperature decay coefficient.
[0040] In step S1, the initialization parameters include the initial temperature and the number of iterations at each temperature value.
[0041] The present invention has the following beneficial effects compared with the prior art:
[0042] 1. The present invention provides an optimization method for pattern recognition feature selection based on the simulated annealing algorithm. By setting an adaptive threshold factor, it solves the limitation problem of selecting intrusion signal features through a threshold in the existing Fisher Score method, enabling all intrusion signal features to have the opportunity to be selected, so that the overall effect of the intrusion signal feature combination can be seen, and reducing the probability of missing selection of the intrusion signal feature combination;
[0043] 2. The method of selecting the initial intrusion signal feature combination in the present invention accelerates the convergence speed of the simulated annealing algorithm. The random factors introduced in the search process accept poor solutions with a certain probability to jump out of the local optimal solution, reach the global optimal solution to complete the selection of the intrusion signal feature value, and improve the pattern recognition and classification accuracy of the phase-sensitive optical time-domain reflectometer. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 A schematic flow chart of a method for optimizing feature selection for pattern recognition based on a simulated annealing algorithm provided in an embodiment of the present invention.
[0045] Figure 2 A schematic diagram of a specific flow chart of a pattern recognition feature selection optimization method based on a simulated annealing algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0046] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0047] In this embodiment, Figure 1-2 As shown, the embodiment of the present invention discloses a pattern recognition feature selection optimization method based on a simulated annealing algorithm, comprising the following steps:
[0048] S1. Encode the intrusion signal characteristic parameters and then initialize the parameters.
[0049] In this embodiment, the initialization parameters include the initial temperature and the number of iterations L at each temperature value. Initialization temperature T = T max In addition, let the initial intrusion signal feature vector be X={x1,x2,…,x M}, M is the number of feature parameters to be selected. They are encoded using 0 / 1, with 1 representing selected and 0 representing unselected. For example, S = {1,1,1,0,0,0,0,…,0}, indicating that the three feature parameters x1, x2, and x3 are selected.
[0050] S2. Calculate the Fisher score of each intrusion signal feature value.
[0051] In step S2, the calculation formula of the Fisher score of the intrusion signal characteristic value is:
[0052]
[0053] Among them, n k represents the number of samples in the kth category, f j,i represents the value of the jth sample in the i-th feature, μ fi represents the i-th feature f i The average value of all samples in , μkfi represents the i-th feature f i The average value of all samples belonging to the kth category, c represents the number of categories of samples, f j,i ∈k represents the sample belonging to the kth category in the i-th feature. F(f i ) represents the Fisher score of the i-th feature.
[0054] If a feature is discriminative, then the variance between the feature and samples of the same category should be as small as possible, while the variance between the feature and samples of different categories should be as large as possible, so that it is conducive to subsequent operations such as classification and prediction. As can be seen from formula (1), F(f i ) value is larger, it can explain the feature f to a certain extent. i The more relevant it is to the problem being solved. The main idea is that the features with strong discrimination performance are characterized by the smallest possible intra-class distance and the largest possible inter-class distance.
[0055] S3. Perform linear transformation on the calculated Fisher score and compare it with the threshold factor to obtain the initial intrusion signal feature combination.
[0056] The specific method of step S3 is:
[0057] S301. Map the Fisher score of each intrusion signal characteristic value to the closed interval [ε, 1-ε], where ε represents a positive real number parameter and 0<ε<0.1. The calculation formula is:
[0058]
[0059] Among them, L(F(f i )) represents the i-th feature f i The mapping value of Fisher score; F max 、F min Respectively represent the maximum and minimum values of the Fisher score of each intrusion signal feature value. max =max{F(f1),F(f2),…,F(f M )},F min =min{F(f1),F(f2),…,F(f M )}.
[0060] In this embodiment, the Fisher score of each intrusion signal feature [F(f1), F(f2), ..., F(f M)], through a linear function L(F(f i )), F(f i )(i=1,2,…,M) are all mapped to the closed interval [ε,1-ε]
[0061] S302. Randomly generate M real-number threshold factors ω i ,ω i ∈U(0,1); M represents the number of feature parameters to be selected;
[0062] S303: Determine the mapping value of the Fisher score of each feature and the corresponding threshold factor ω i If L(F(f i ))>ω i , then select the intrusion signal feature, that is, x i =1; if L(F(f i ))≤ω i , then the intrusion signal feature is not selected, that is, x i =0; i=1,2,…,M, the selected intrusion signal features are used as the initial intrusion signal feature combination.
[0063] In this embodiment, unlike the simulated annealing algorithm which generates the initial intrusion signal feature combination in a random manner, the present invention uses the mapped Fisher score, i.e., L(F(f i )) is used to generate the initial intrusion signal feature combination of the simulated annealing algorithm. In addition, in the process of generating the initial intrusion signal feature combination, F(f i ) are all mapped to the closed interval [ε, 1-ε] (0<ε<0.1), rather than the interval [0,1]. This ensures that all intrusion signal features, regardless of their Fisher score, have a chance of being selected into the initial intrusion signal feature combination used to generate the simulated annealing algorithm. Furthermore, intrusion signal features with high Fisher scores have a higher probability of being selected into the initial intrusion signal feature combination than intrusion signal features with low Fisher scores. The closer ε approaches 0, the more likely intrusion signal features with high Fisher scores are to be selected into the initial intrusion signal feature combination, while the more likely intrusion signal features with low Fisher scores are not to be selected. Conversely, the closer ε approaches 0.1, the less likely intrusion signal features with high Fisher scores are to be selected into the initial intrusion signal feature combination, while intrusion signal features with low Fisher scores have a relatively higher chance of being selected into the initial intrusion signal feature combination.
[0064] S4. Determine the fitness function.
[0065] The choice of fitness function directly affects the convergence speed of the simulated annealing algorithm and its ability to find an optimal solution. Generally speaking, its design adheres to the principles of single-valued, continuous, maximal, low computational complexity, and high versatility. The goal of feature selection is to select the features most relevant to the problem being solved, thereby improving the accuracy of feature classification.
[0066] Specifically, in this embodiment, the LIBSVM classification accuracy of the high-dimensional intrusion signal feature vector is used as a fitness function to evaluate the quality of the selected intrusion signal feature combination. Its expression is:
[0067] h(x g )=accuracy(x g ); (3)
[0068] Among them, x g is the g-th generation selected intrusion signal characteristic value. The larger the fitness function value is, the better the selected intrusion signal characteristic value combination is.
[0069] S5. Calculate the initial fitness function value h(S) based on the initial intrusion signal feature combination S.
[0070] S6. Generate a new intrusion signal feature combination S′ (new solution) based on the inversion of the G feature vector codes in the previous intrusion signal feature combination, and calculate the fitness function value h(S′) of the new intrusion signal feature combination S′.
[0071] In step S6, the new intrusion signal feature combination is generated by:
[0072] S601. Randomly generate M real numbers ρ i ,ρ i ∈U(0,1);
[0073] S602: Determine the mapping value of the Fisher score of each feature and the real number ω i If L(F(f i ))>ρ i , then select the intrusion signal feature; otherwise, L(F(f i ))≤ρ i , then the intrusion signal feature is not selected, i=1,2,…,M;L(F(f i )) represents the i-th feature f i The mapping value of Fisher score;
[0074] S603: Randomly select G feature vectors from the selected intrusion signal features, and perform an inversion operation on the G feature vector codes in the previous intrusion signal feature combination to obtain a new intrusion signal feature combination.
[0075] In this embodiment, G = [M / 10]*2 + 1.
[0076] S7. Determine whether to accept the update of the feature combination based on the comparison result of h(S) and h(S′) and in combination with the Metropolis acceptance criterion for new solutions.
[0077] The specific judgment method in step S7 is as follows:
[0078] (1) If the fitness function value of the new intrusion signal feature combination is greater than the fitness function value of the current intrusion signal feature combination, that is: h(S′) > h(S), then change the current solution to the new solution with a probability of 1, that is, update the intrusion signal feature combination;
[0079] (2) If the fitness function value of the new intrusion signal feature combination is less than or equal to the fitness function value of the current intrusion signal feature combination, then calculate the probability. The calculation formula is:
[0080]
[0081] T represents the current temperature, h(S′) represents the fitness function value of the new feature combination, and h(S) represents the fitness function value of the current feature combination. Then take a random number 0 < r < 0.1. If r < P, then update the feature combination, otherwise do not update and still retain the original feature combination S.
[0082] In this embodiment, the greater the current temperature T, the greater the value in the exponential function parentheses, and the greater the probability P of accepting this new value. As the number of iterations increases, it will become smaller and smaller, and the probability P of accepting the new value will also become smaller and smaller, which is equivalent to gradually cooling down and tending to a stable state.
[0083] S8. At temperature T, repeat steps S6 - S7 until the number of iterations is reached.
[0084] S9. Update the temperature according to the temperature decay coefficient; repeat steps S6 - S8 until the temperature is less than or equal to the end temperature T min At this time, determine the optimal intrusion signal feature combination S″ according to the optimal fitness function value h(S″).
[0085] In step S9, the method for updating the temperature is:
[0086] T′ = KT; (6)
[0087] Among them, T represents the temperature before update, T′ represents the temperature after update, and K represents the temperature decay coefficient, which can control the cooling rate of the solid annealing process, that is, the decay rate of temperature T.
[0088] Through the above steps, the selection of an intrusion signal feature combination with a better overall effect can be achieved. Specifically, the present invention uses a pattern recognition feature selection optimization method based on a simulated annealing algorithm, compared with other intelligent algorithms: on the one hand, with the introduction of a threshold factor, intrusion signal features with high or low Fisher scores have the opportunity to be selected, and the overall effect of an intrusion signal feature combination can be seen. On the other hand, the use of the Fisher Score method to select the initial intrusion signal feature combination speeds up the convergence speed of the simulated annealing algorithm, and the random factors introduced in the search process accept the difference solution with a certain probability to jump out of the local optimal solution, reach the global optimal solution to complete the selection of the intrusion signal feature value, and can eliminate redundant intrusion signal features. The selected intrusion signal features can improve the pattern recognition classification accuracy of the phase-sensitive optical time domain reflectometer.
[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A pattern recognition feature selection optimization method based on simulated annealing algorithm, characterized in that: The following steps are involved: S1. Encode the characteristic parameters of the intrusion signal collected by the phase-sensitive optical time domain reflectometer and then initialize the parameters; S2. Calculate the Fisher score of each intrusion signal feature value; S3, performing linear transformation on the calculated Fisher score and comparing it with the threshold factor to obtain the initial intrusion signal feature combination; S4. Determine the fitness function; S5. Calculating an initial fitness function value based on the initial intrusion signal feature combination; S6, according to the previous intrusion signal feature combination G The feature vector codes are inverted to generate a new intrusion signal feature combination, and the fitness function value of the new intrusion signal feature combination is calculated; S7. If the fitness function value of the new intrusion signal feature combination is greater than the fitness function value solution of the current intrusion signal feature combination, then the new solution is directly accepted; otherwise, the Metropolis new solution acceptance criterion is used to determine whether to accept the new solution; S8. Repeat steps S6 to S7 at temperature T until the number of iterations is reached. S9, update the temperature according to the temperature attenuation coefficient; repeat steps S6-S8 until the temperature is less than or equal to the end temperature T min When , the optimal intrusion signal feature combination is determined according to the optimal fitness function value; The specific method of step S3 is: S301, map the Fisher scores of each intrusion signal characteristic value to a closed interval [ ε , 1- ε ] above; among them, ε represents a positive real number parameter, and 0 < ε < 0.1; the calculation formula is: ； in, Indicates the i Features f i The mapping value of Fisher score; Respectively represent the maximum and minimum values of the Fisher score of each intrusion signal feature value; S302, random generation M A real-number threshold factor ω i , ω i ∈U(0,1); M represents the number of feature parameters to be selected; S303: Determine the mapping value of the Fisher score of each feature and the corresponding threshold factor ω i The relationship between L ( F ( f i )) > ω i , then select the intrusion signal feature; otherwise, L ( F ( f i )) ≤ ω i , then the intrusion signal feature is not selected. i= 1,2,…,M, the selected intrusion signal features are used as the initial intrusion signal feature combination.
2. The method for pattern recognition feature selection optimization based on simulated annealing algorithm according to claim 1, characterized in that: In step S1, the intrusion signal characteristic parameters are encoded using 0 / 1, where 1 represents selected and 0 represents unselected.
3. The method for pattern recognition feature selection optimization based on simulated annealing algorithm according to claim 1, characterized in that: In step S2, the calculation formula of the Fisher score of the intrusion signal characteristic value is: ； in, n k Indicates the k The number of class samples, f j,i Indicates the i Among the features j The value of the sample, μ f i Indicates the i Features f i The average value of all samples in , μk fi Indicates the i Features f i Belong to the k The average value of all samples in the categories, c represents the number of categories of samples, Represents the i-th feature f i Fisher score.
4. The method for pattern recognition feature selection optimization based on simulated annealing algorithm according to claim 1, characterized in that: In step S4, the LIBSVM classification accuracy of the high-dimensional intrusion signal feature vector is used as the fitness function.
5. The method for pattern recognition feature selection optimization based on simulated annealing algorithm according to claim 1, characterized in that: In step S6, the new intrusion signal feature combination is generated by: S601, random generation M real numbers ρ i , ρ i ∈U(0,1); S602: Determine the mapping value of the Fisher score of each feature and the real number ω i relationship, if L ( F ( f i )) > ρ i , then select the intrusion signal feature; otherwise, L ( F ( f i )) ≤ ρ i , then the intrusion signal feature is not selected. i= 1,2,…, M ; Indicates the i Features f i The mapping value of Fisher score; S603, randomly select from the selected intrusion signal features G feature vectors, for this intrusion signal feature combination G The new intrusion signal feature combination is obtained by performing the inversion operation on the feature vector encoding.
6. The method for pattern recognition feature selection optimization based on simulated annealing algorithm according to claim 1, characterized in that: The determination method in step S7 is specifically as follows: (1) If the fitness function value of the new intrusion signal feature combination is greater than the fitness function value solution of the current intrusion signal feature combination, the intrusion signal feature combination is updated; (2) If the fitness function value of the new intrusion signal feature combination is less than or equal to the fitness function value solution of the current intrusion signal feature combination, the probability is calculated using the following formula: ； T Indicates the current temperature. h (S′) represents the fitness function value of the new feature combination, h ( S ) represents the fitness function value of the current feature combination; then take a random number 0 < r < 0.1, if r < P , then the feature combination is updated, otherwise it is not updated.
7. The method for pattern recognition feature selection optimization based on simulated annealing algorithm according to claim 1, characterized in that: In step S9, the method for updating the temperature is: T ′ = KT ； in, T Indicates the temperature before updating, T′ Indicates the updated temperature, K Represents the temperature attenuation coefficient.
8. The method for pattern recognition feature selection optimization based on simulated annealing algorithm according to claim 1, characterized in that: In step S1, the initialization parameters include the initial temperature and the number of iterations at each temperature value.