阅读说明：本技术 空心钻磨削砂轮状态智能监测方法 (Intelligent monitoring method for grinding wheel state of hollow drill ) 是由 迟玉伦 江欢 李郝林 王赟 于 2020-06-28 设计创作，主要内容包括：本发明涉及一种空心钻磨削砂轮状态智能监测方法,首先,传感器安装及信号监测,监测磨削过程中的声发射信号、功率信号和振动信号,再进行信号特征参数提取,提取声发射信号和功率信号的时域参数以及振动信号的高频特征信息作为特征参数,然后对这些特征参数进行归一化处理并提取主要成分作为SA-SVM模型的输入样本,然后,运用模拟退火算法优化支持向量机参数的选择,并对样本进行训练学习；最后,采用SA-SVM模型智能预测,将系统分析结果与砂轮实际磨削状态做对比,判断模型的预测性能。(The invention relates to a method for intelligently monitoring the state of a grinding wheel of a hollow drill, which comprises the steps of firstly, installing a sensor and monitoring signals, monitoring an acoustic emission signal, a power signal and a vibration signal in the grinding process, then extracting characteristic parameters of the signals, extracting time domain parameters of the acoustic emission signal and the power signal and high-frequency characteristic information of the vibration signal as the characteristic parameters, then carrying out normalization processing on the characteristic parameters, extracting main components as input samples of an SA-SVM model, then optimizing the selection of parameters of a support vector machine by using a simulated annealing algorithm, and training and learning the samples; and finally, intelligently predicting by adopting an SA-SVM model, comparing the system analysis result with the actual grinding state of the grinding wheel, and judging the prediction performance of the model.)

1. The intelligent monitoring method for the grinding wheel state of the hollow drill is characterized by comprising the following steps of: firstly, mounting a sensor and monitoring a signal, monitoring an acoustic emission signal, a power signal and a vibration signal in a grinding process, extracting characteristic parameters of the signal, extracting time domain parameters of the acoustic emission signal and the power signal and high-frequency characteristic information of the vibration signal as characteristic parameters, then carrying out normalization processing on the characteristic parameters and extracting main components as input samples of an SA-SVM model, then optimizing selection of support vector machine parameters by using a simulated annealing algorithm, and training and learning the samples; and finally, intelligently predicting by adopting an SA-SVM model, comparing the system analysis result with the actual grinding state of the grinding wheel, and judging the prediction performance of the model.

2. The intelligent monitoring method for the state of the hollow drill grinding wheel according to claim 1, characterized in that: the specific method for mounting the sensor and monitoring the signal comprises the following steps: the abrasion state of the grinding wheel is monitored by adopting a power sensor, an acoustic emission sensor and a vibration sensor, the power sensor is connected in series between a spindle motor and a frequency converter, and the acoustic emission sensor is adsorbed on the left end face of a shell of the tailstock by utilizing the characteristic that the acoustic emission sensor has strong magnetism and is used for monitoring an acoustic emission signal in the grinding process; the vibration sensor adopts a three-way vibration sensor and monitors vibration signals in three directions which are vertical to each other; and signals acquired by the power sensor, the acoustic emission sensor and the vibration sensor are transmitted to data acquisition software in a computer through an amplifier and an acquisition card through signal lines for signal monitoring.

3. The intelligent monitoring method for the state of the hollow drill grinding wheel according to claim 2, characterized in that: the three-way vibration sensor has strong magnetism, and is adsorbed on the tailstock center, so that the upper surface of the three-way vibration sensor is parallel to the horizontal plane and is used for monitoring the vibration change condition of the tailstock center in the grinding process.

4. The intelligent monitoring method for the state of the hollow drill grinding wheel according to claim 1, characterized in that: the specific method for extracting the signal characteristic parameters comprises the following steps:

1) time domain feature parameters

The change of the abrasion state of the grinding wheel is reflected by calculating and analyzing the waveform, the amplitude and the time domain characteristic parameters of the power, the acoustic emission and the vibration signal;

2) wavelet packet decomposition

Wavelet packet decomposition relation:

in the formula (1), S represents an original signal, L represents a low frequency, H represents a high frequency, and the last number represents the number of layers of wavelet packet decomposition; the three-layer wavelet packet decomposition finally decomposes the original signal into 8 frequency segments, and when the number of layers of the wavelet packet decomposition is n, the decomposition can be decomposed into 2ⁿA frequency band, andthe frequency intervals of each frequency bin are identical. Assume a sampling frequency of f_sThen the frequency interval of the ith frequency band is

From the wavelet packet decomposition theory, it can be known that n layers of wavelet packet decomposition are performed on the original signal, and 2 can be obtained after decompositionⁿA sub-signal sequence of frequency bins. The wavelet packet energy value of each subsequence is

In the formula, x_iAnd (t) is a reconstruction coefficient of the ith sub-signal after wavelet packet decomposition.

2ⁿThe sub-signals can obtain energy vectors of different frequency bands with corresponding quantityNormalizing the energy values, mapping the energy data into (0,1) to obtain an energy ratio

In the formula (I), the compound is shown in the specification,

5. The intelligent monitoring method for the state of the hollow drill grinding wheel according to claim 1, characterized in that: the optimization by using the simulated annealing algorithm comprises the following specific steps:

(1) initialization: setting an initial temperature T ═ T₀Arbitrarily take an initial solution E₁Determination of Metropolis chain LengthL, which is the number of iterations per T;

(2) let T be the next value T in the cooling schedule_iRepeating steps (3) to (6) with i equal to 1,2, … and L;

(3) for the current solution E₁Random disturbance is carried out to generate a new solution E₂；

(4) Calculating delta E ═ E₂-E₁；

(5) Judging whether to accept a new solution according to a Metropolis criterion, wherein the specific rule is as follows: if Δ E<0, then the current solution E₁By new solution E₂Substituted, i.e. E₁＝E₂(ii) a If Δ E>0, then exp (- Δ E/T) is calculated_i) At the same time, randomly generating a random number rand in the (0,1) interval if exp (- Δ E/T)_i)>rand, current solution E₁Also by the new solution E₂Substitution, E₁＝E₂Otherwise, the current solution E is kept₁；

(6) If several new solutions E in Metropolis chain are in succession₂All can not replace the current solution E₁Or the set end temperature is reached, the solution E is currently solved₁The optimal solution is obtained, and the procedure is ended; otherwise, returning to the step (2) and continuing to execute step by step.

6. The intelligent monitoring method for the state of the hollow drill grinding wheel according to claim 1, characterized in that: the intelligent prediction of the SA-SVM model is combined with the establishment of an SA-SVM classification prediction model by adopting a support vector machine and a simulated annealing optimization algorithm, and the parameter optimization process of the SA-SVM model comprises the following steps: firstly, setting the upper and lower limits of initial parameters and optimizing parameters of a simulated annealing algorithm. Then, a set of solutions (C) is randomly generated₁,g₁) And forming an original SVM, training the model by using the training samples, and testing the classification prediction accuracy of the model by using the test samples. Second, in the first set of solutions (C)₁,g₁) On the basis, random disturbance is generated, parameters are updated, and the updated model is trained, classified and predicted to obtain a correct rate result. And judging whether to accept the new parameters according to the Metropolis criterion. Repeatedly executing, judging whether the parameters meet the requirements, if so, finishing the optimization, and outputting optimized parameters andclassifying results; if not, gradually reducing the annealing temperature until meeting the convergence condition to obtain the optimal parameters.

Technical Field

The invention relates to an intelligent monitoring method for the grinding wheel state of a hollow drill, in particular to an intelligent monitoring method for the grinding wheel state of the hollow drill by a simulated annealing optimization support vector machine.

Background

With the development of science and technology, higher and higher requirements are put on engineering materials, and various high-strength, high-hardness, corrosion-resistant and high-temperature-resistant engineering materials are adopted more and more, and most of the engineering materials belong to difficult-to-process materials. In order to meet the processing requirements of materials difficult to process and novel materials and ensure that the processing task is finished with high quality and high efficiency, higher requirements on the aspects of wear resistance, reliability, precision, size and the like of the cutter are provided. The hollow drill mainly comprises a cutting edge and a tool shank, and is of a hollow structure, in order to guarantee the clamping precision of the hollow drill, the cutting edge shank part needs to be ground, but the hollow drill is of a structure similar to a thin-wall part, and the grinding difficulty is greatly increased. Further, in order to improve wear resistance, heat resistance and strength of high-speed steel, various alloying elements such as W, Mo, Cr and V are added to the high-speed steel, but at the same time grindability of the high-speed steel is also lowered. Compared with a cutter for grinding common materials, the grinding tool has the advantages that in the process of grinding the high-speed steel hollow drill, the grinding force is larger, the grinding temperature is higher, and the grinding wheel is more easily abraded. The wear state of the grinding wheel directly affects the quality and efficiency of the grinding product, and the grinding wheel needs to be regularly dressed in order to ensure the quality of the processed product. The grinding wheel is frequently dressed, so that the machining efficiency is certainly influenced, the utilization rate of the grinding wheel is reduced, the machined surface roughness of a product cannot reach the standard even if the product is not dressed in time, and even chatter marks and burns occur. Therefore, how to effectively predict the abrasion state of the grinding wheel and timely finish the grinding wheel has important significance for improving the processing surface quality of the workpiece, improving the processing efficiency and optimizing the process capability.

Therefore, a grinding wheel state monitoring method based on a wavelet packet and Simulated annealing optimization support vector machine (SA-SVM) is required to solve the problems of high difficulty and high requirement in grinding high-speed steel core drill and difficulty in controlling the dressing time of the grinding wheel. The method extracts time domain characteristic parameters of acoustic emission and power signals and extracts high-frequency characteristic information of vibration signals through wavelet packet decomposition to serve as sample input of a support vector machine. Aiming at the problem that the empirical parameters of the traditional support vector machine are difficult to determine, the selection of the parameters is optimized by adopting a simulated annealing algorithm, the optimization algorithm is suitable for solving various nonlinear problems, has stronger robustness, global convergence and wide adaptability, provides an effective solution for intelligently monitoring the abrasion state of the grinding wheel during grinding of the high-speed steel hollow drill, and has important practical engineering application value.

Disclosure of Invention

The invention provides an intelligent monitoring method for the grinding wheel state of a hollow drill, which adopts the intelligent monitoring method for the grinding wheel state of the hollow drill by simulating annealing optimization support vector machine, and optimizes the core parameters of the support vector machine by using a simulated annealing method, thereby avoiding trapping in a local optimal trap and achieving global optimization.

The technical scheme of the invention is as follows: an intelligent monitoring method for the grinding wheel state of a hollow drill comprises the steps of firstly, installing a sensor and monitoring signals, monitoring acoustic emission signals, power signals and vibration signals in the grinding process, extracting characteristic parameters of the signals, extracting time domain parameters of the acoustic emission signals and the power signals and high-frequency characteristic information of the vibration signals as the characteristic parameters, carrying out normalization processing on the characteristic parameters, extracting main components as input samples of an SA-SVM model, optimizing selection of support vector machine parameters by using a simulated annealing algorithm, and training and learning the samples; and finally, intelligently predicting by adopting an SA-SVM model, comparing the system analysis result with the actual grinding state of the grinding wheel, and judging the prediction performance of the model.

Further, the specific method for sensor installation and signal monitoring comprises the following steps: the abrasion state of the grinding wheel is monitored by adopting a power sensor, an acoustic emission sensor and a vibration sensor, the power sensor is connected in series between a spindle motor and a frequency converter, and the acoustic emission sensor is adsorbed on the left end face of a shell of the tailstock by utilizing the characteristic that the acoustic emission sensor has strong magnetism and is used for monitoring an acoustic emission signal in the grinding process; the vibration sensor adopts a three-way vibration sensor and monitors vibration signals in three directions which are vertical to each other; and signals acquired by the power sensor, the acoustic emission sensor and the vibration sensor are transmitted to data acquisition software in a computer through an amplifier and an acquisition card through signal lines for signal monitoring.

Furthermore, the three-way vibration sensor has strong magnetism, and is adsorbed on the tailstock center, so that the upper surface of the three-way vibration sensor is parallel to the horizontal plane and used for monitoring the vibration change condition of the tailstock center in the grinding process.

Further, the specific method for extracting the signal characteristic parameters comprises the following steps:

1) time domain feature parameters

The change of the abrasion state of the grinding wheel is reflected by calculating and analyzing the waveform, the amplitude and the time domain characteristic parameters of the power, the acoustic emission and the vibration signal;

2) wavelet packet decomposition

Wavelet packet decomposition relation:

S＝LLL3+HLL3+LHL3+HHL3+LLH3+HLH3+LHH3+HHH3 (1)

in the formula (1), S represents an original signal, L represents a low frequency, H represents a high frequency, and the last number represents the number of layers of wavelet packet decomposition; the three-layer wavelet packet decomposition finally decomposes the original signal into 8 frequency segments, and when the number of layers of the wavelet packet decomposition is n, the decomposition can be decomposed into 2ⁿEach frequency segment, and the frequency interval of each frequency segment is identical. Assume a sampling frequency of f_sThen the frequency interval of the ith frequency band is

From the wavelet packet decomposition theory, it can be known that n layers of wavelet packet decomposition are performed on the original signal, and 2 can be obtained after decompositionⁿA sub-signal sequence of frequency bins. The wavelet packet energy value of each subsequence is

In the formula, x_iAnd (t) is a reconstruction coefficient of the ith sub-signal after wavelet packet decomposition.

2ⁿThe sub-signals can obtain energy vectors of different frequency bands with corresponding quantityNormalizing the energy values, mapping the energy data into (0,1) to obtain an energy ratio

In the formula (I), the compound is shown in the specification,is total energy and has

Further, the optimization by using the simulated annealing algorithm comprises the following specific steps:

(1) initialization: setting an initial temperature T ═ T₀Arbitrarily take an initial solution E₁Determining the Metropolis chain length L, namely the iteration number of each T;

(2) let T be the next value T in the cooling schedule_iRepeating steps (3) to (6) with i equal to 1,2, … and L;

(3) for the current solution E₁Random disturbance is carried out to generate a new solution E₂；

(4) Calculating delta E ═ E₂-E₁；

(5) Judging whether to accept a new solution according to a Metropolis criterion, wherein the specific rule is as follows: if Δ E<0, then the current solution E₁By new solution E₂Substituted, i.e. E₁＝E₂(ii) a If Δ E>0, then exp (- Δ E/T) is calculated_i) At the same time, randomly generating a random number rand in the (0,1) interval if exp (- Δ E/T)_i)>rand, current solution E₁Also by the new solution E₂Substitution，E₁＝E₂Otherwise, the current solution E is kept₁；

(6) If several new solutions E in Metropolis chain are in succession₂All can not replace the current solution E₁Or the set end temperature is reached, the solution E is currently solved₁The optimal solution is obtained, and the procedure is ended; otherwise, returning to the step (2) and continuing to execute step by step.

Further, the intelligent prediction of the SA-SVM model is combined with the establishment of an SA-SVM classification prediction model by adopting a support vector machine and a simulated annealing optimization algorithm, and the parameter optimization process of the SA-SVM model comprises the following steps: firstly, setting the upper and lower limits of initial parameters and optimizing parameters of a simulated annealing algorithm. Then, a set of solutions (C) is randomly generated₁,g₁) And forming an original SVM, training the model by using the training samples, and testing the classification prediction accuracy of the model by using the test samples. Second, in the first set of solutions (C)₁,g₁) On the basis, random disturbance is generated, parameters are updated, and the updated model is trained, classified and predicted to obtain a correct rate result. And judging whether to accept the new parameters according to the Metropolis criterion. Repeatedly executing, judging whether the parameters meet the requirements, if so, finishing the optimization, and outputting the optimized parameters and the classification result; if not, gradually reducing the annealing temperature until meeting the convergence condition to obtain the optimal parameters.

The invention has the beneficial effects that:

the invention provides an intelligent monitoring method for the grinding wheel state of a core drill based on a wavelet packet and an SA-SVM (space-support vector machine), aiming at high difficulty and high requirement in grinding high-speed steel core drill and difficulty in controlling the dressing time of a grinding wheel. The core parameters of the support vector machine are optimized by using a simulated annealing method, so that a trap trapped in local optimum can be avoided, and global optimum is achieved. Aiming at grinding high-speed steel hollow drill, the SA-SVM model can effectively judge the abrasion state of the grinding wheel, so the method has important significance for promoting the development of grinding process optimization technology and improving the machine tool machining technology level.

Drawings

FIG. 1 is a general structure and flow chart of a grinding wheel wear monitoring system;

FIG. 2 is a schematic power sensor wiring diagram;

FIG. 3 is a view of a sensor mounting location;

FIG. 4 is a wavelet packet decomposition tree;

FIG. 5 is a schematic diagram of a simulated annealing optimization algorithm;

FIG. 6 is a flow chart of SA-SVM parameter optimization.

Detailed Description

The invention is further described with reference to the following figures and examples.

An intelligent monitoring method for the grinding wheel state of a hollow drill is characterized in that an adopted grinding wheel wear state monitoring system is mainly divided into three parts: signal monitoring and characteristic parameter extraction, an SA-SVM model and a model prediction result. The method comprises the first step of monitoring an acoustic emission signal, a power signal and a vibration signal in the grinding process, extracting time domain parameters of the acoustic emission signal and the power signal and high-frequency characteristic information of the vibration signal as characteristic parameters, then carrying out normalization processing on the characteristic parameters and extracting main components as input samples of an SA-SVM model. And in the second part, the selection of parameters of the support vector machine is optimized by using a simulated annealing algorithm, and the samples are trained and learned. And the third part is to compare the system analysis result with the actual grinding state of the grinding wheel and judge the prediction performance of the model. The general structure and flow of the wheel wear monitoring system is shown in FIG. 1. And when the SA-SVM model is stably trained and learned, inputting the characteristic parameters into the SA-SVM model in real time, and judging whether the grinding wheel needs to be dressed according to the prediction result of the model. By the method, the abrasion state of the grinding wheel can be monitored in time, the grinding wheel can be dressed timely, the machining efficiency is improved, and the product quality is guaranteed.

The method comprises the following specific steps:

sensor installation and signal monitoring

The invention adopts a power sensor, an acoustic emission sensor and a vibration sensor to monitor the abrasion state of the grinding wheel. The power sensor 1 is connected in series between the spindle motor 2 and the frequency converter 3, and the wiring diagram of the power sensor is shown in fig. 2. The acoustic emission sensor 4 is adsorbed on the left end face of the tailstock shell 6 by utilizing the characteristic that the acoustic emission sensor has strong magnetism, and is used for monitoring acoustic emission signals in the grinding process. The vibration sensor 5 is a three-way vibration sensor and can monitor vibration signals in three directions perpendicular to each other. The three-way vibration sensor 5 also has strong magnetism, the vibration sensor 5 is adsorbed on the tailstock center 7, the upper surface of the vibration sensor is parallel to the horizontal plane and used for monitoring the vibration change condition of the tailstock center 7 in the grinding process, and the installation positions of the acoustic emission sensor 4 and the vibration sensor 5 are shown in fig. 3. The signals collected by the sensor are transmitted to data collection software in a computer through an amplifier and a collection card through a signal line for signal monitoring.

Second, signal characteristic parameter extraction

1) Time domain feature parameters

The change of the abrasion state of the grinding wheel can be reflected by calculating and analyzing the waveform, the amplitude and the time domain characteristic parameters of the power, the acoustic emission and the vibration signal. The common time domain characteristic parameters include mean, peak, range, standard deviation, variance, root mean square, skewness, kurtosis index, form factor, peak factor, pulse factor, margin factor, etc., and table 1 lists these characteristic parameters and their calculation formulas.

TABLE 1 time-Domain feature parameters

2) Wavelet packet decomposition

Wavelet packet decomposition is an extension of wavelet transformation, can not only further decompose the low-frequency part of a signal like wavelet transformation, but also implement re-decomposition on the high-frequency part which cannot be decomposed by wavelet transformation, and is a more precise signal time-frequency analysis method. The wavelet packet decomposition is further explained using a three-level wavelet packet decomposition tree as shown in fig. 4.

In fig. 4, S represents the original signal, L represents the low frequency, H represents the high frequency, and the last number represents the number of layers of the wavelet packet decomposition. The decomposition has the relation:

as shown in fig. 4, the three-layer wavelet packet decomposition finally decomposes the original signal into 8 frequency segments, and when the number of layers of the wavelet packet decomposition is n, the original signal can be decomposed into 2 in totalⁿEach frequency segment, and the frequency interval of each frequency segment is identical. Assume a sampling frequency of f_sThen the frequency interval of the ith frequency band is

From the wavelet packet decomposition theory, it can be known that n layers of wavelet packet decomposition are performed on the original signal, and 2 can be obtained after decompositionⁿA sub-signal sequence of frequency bins. The wavelet packet energy value of each subsequence is

In the formula, x_iAnd (t) is a reconstruction coefficient of the ith sub-signal after wavelet packet decomposition.

2ⁿThe sub-signals can obtain energy vectors of different frequency bands with corresponding quantityNormalizing the energy values, mapping the energy data into (0,1) to obtain an energy ratio

In the formula (I), the compound is shown in the specification,is total energy and has

Third, the basic principle of the simulated annealing optimization support vector machine

1) Support vector machine

The support vector machine is a machine learning method, minimizes the structure risk while minimizing the error of sample points, improves the generalization capability of the model, and has no limitation of data dimension. The method aims to search a hyperplane with maximized intervals to segment samples and convert the problem into a convex quadratic programming problem to be solved. When the classification samples are not linearly classified, the nonlinear classification is changed into the linear classification problem of a high-dimensional space through high-dimensional space transformation. Hypothesis set { (x)_i,y_i) I-1, 2, …, l is a set of training samples, where x_i(x_i∈R^d) Is the input column vector for the ith training sample,

y_i∈ R is a corresponding output sample, SVM regression maps data to a high-dimensional feature space by using a kernel function, linear regression is carried out in the high-dimensional feature space, and the learning process is converted into a convex optimization problem according to a structure risk minimization principle, namely:

where phi (x) is a non-linear mapping function, a linear insensitive loss function, ξ_i，

Is a relaxation variable; and C is a penalty factor.

Due to the computational complexity, equations are transformed into dual problems according to Lagrange's dual theory, i.e.

Solving the above problem can result in a regression equation that is ultimately expressed as:

in the formula: k (X)_i,X_j)＝Φ(X_i)Φ(X_j) Is a kernel function satisfying the Mercer condition;a_iis Lagrange multiplier in quadratic programming.

In the formula: n is a radical of_nsvThe number of the support vector machines.

Common kernel functions are linear kernel functions, polynomial kernel functions, radial basis kernel functions, and Sigmoid kernel functions. Generally, the accuracy of a training set and a test set corresponding to a linear kernel function and a Sigmoid kernel function is low, the prediction accuracy of a training set corresponding to a radial basis kernel function and a polynomial kernel function is equivalent, but the prediction accuracy of the radial basis kernel function to the test set is superior to that of the polynomial kernel function, and the method has good generalization capability, wherein the expression of the radial basis kernel function is as follows:

K(x,x_i)＝exp(-g||x-x_i||²) (8)

in the formula: g > 0.

According to the SVM algorithm principle, SVM performance is mainly influenced by a penalty factor C and a kernel function parameter g. The penalty factor C plays a role in controlling the penalty degree of the misclassified samples, and the compromise between the proportion of the misclassified samples and the algorithm complexity is realized. When the penalty factor C is smaller, the penalty of misclassification samples is reduced, and the training error is increased; when the penalty factor C is larger, the penalty of misclassification samples is increased, and the learning precision is improved. The smaller the kernel function parameter g is, the smaller the fitting error is, the longer the training time is, and the too small g can cause the overfitting of the model to reduce the generalization capability. Generally, the two parameters are selected according to experience, which is not beneficial to the support vector machine to exert the practical effect, therefore, the invention selects the simulated annealing algorithm to realize the optimization of the parameter selection.

2) Simulated annealing algorithm

The idea of Simulated Annealing (SA) is a random optimization algorithm based on a monte carlo iterative solution strategy, and when a feasible solution is iteratively updated, the algorithm receives a solution worse than the current solution with a certain probability, so that the local optimal solution may jump out to reach a global optimal solution. Taking fig. 5 as an example, assuming that the initial solution is a left blue point a, the simulated annealing algorithm will search for the local optimal solution B quickly, but will not end up after searching for the local optimal solution, and will accept the right movement with a certain probability. After several times of jumping out of the local optimal point, the global optimal point D is finally searched.

The overall steps of the simulated annealing algorithm are as follows:

(1) initialization: setting an initial temperature T ═ T₀Arbitrarily take an initial solution E₁Metropolis chain length L, i.e., the number of iterations per T, is determined.

(2) Let T be the next value T in the cooling schedule_iAnd (5) repeating the steps (3) to (6) when i is 1,2, … and L.

(3) For the current solution E₁Random disturbance is carried out to generate a new solution E₂。

(4) Calculating delta E ═ E₂-E₁。

(5) Judging whether to accept a new solution according to a Metropolis criterion, wherein the specific rule is as follows: if Δ E<0, then the current solution E₁By new solution E₂Substituted, i.e. E₁＝E₂(ii) a If Δ E>0, then exp (- Δ E/T) is calculated_i) At the same time, randomly generating a random number rand in the (0,1) interval if exp (- Δ E/T)_i)>rand, current solution E₁Also by the new solution E₂Substitution, E₁＝E₂Otherwise, the current solution E is kept₁。

(6) If several new solutions E in Metropolis chain are in succession₂All can not replace the current solution E₁Or the set end temperature is reached, the solution E is currently solved₁The optimal solution is obtained, and the program is ended. Otherwise, returning to the step (2) and continuing to execute step by step.

3) Intelligent prediction of SA-SVM model

And establishing an SA-SVM classification prediction model by combining the support vector machine and the simulated annealing optimization algorithm. The SA-SVM model parameter optimization process is shown in fig. 6.

Firstly, setting the upper and lower limits of initial parameters and optimizing parameters of a simulated annealing algorithm. Then, a set of solutions (C) is randomly generated₁,g₁) And forming an original SVM, training the model by using the training samples, and testing the classification prediction accuracy of the model by using the test samples. Second, in the first set of solutions (C)₁,g₁) On the basis, random disturbance is generated, parameters are updated, and the updated model is trained, classified and predicted to obtain a correct rate result. And judging whether to accept the new parameters according to the Metropolis criterion. Repeatedly executing, judging whether the parameters meet the requirements, if so, finishing the optimization, and outputting the optimized parameters and the classification result; if not, gradually reducing the annealing temperature until meeting the convergence condition to obtain the optimal parameters.

The punishment factor C and the kernel function parameter g in the support vector machine are optimized by using the simulated annealing optimization algorithm, so that the optimization result tends to be globally optimal, and the problem of low prediction accuracy caused by selecting the parameters by depending on experience is effectively solved.