Multi-feature quality prediction method for optimizing FFN based on improved Heima algorithm

Abstract

The invention discloses a multi-feature quality prediction method for optimizing an FFN based on an improved Hereum algorithm, and the method comprises the steps: obtaining technological parameter data, and constructing a sample of an HO-FFN model after feature screening; an HO-FFN model is constructed; determining a network topology structure of the FFN, and optimizing the weight and bias parameters of the FFN by using an improved He-Map algorithm; the method comprises the following steps: collecting data of a preset threshold quantity of process parameter types in a mechanical product quality processing process, performing product quality real-time prediction output by utilizing an HO-FFN model to obtain a quality prediction result, and further performing real-time adjustment on the processing process of the mechanical product. According to the method, the feed-forward neural network is optimized by improving the Heima algorithm, the process parameter prediction precision and efficiency are effectively improved, the training time is shortened, meanwhile, the local optimum problem in a traditional method is effectively avoided, and higher accuracy and stability are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of product processing quality control, and in particular to a multi-feature quality prediction method based on an improved Hippo algorithm to optimize FFN. Background Art

[0002] The quality of mechanical products is a key factor for enterprises to survive and develop in the market competition, and directly affects the reputation and economic benefits of enterprises. In order to gain a foothold in the fierce market competition, enterprises not only need to design and produce products that meet market demand, but also must ensure the stability and reliability of product quality through strict quality management. The quality of mechanical products is mainly formed in design and manufacturing, among which the manufacturing link accounts for nearly 70% of the sources of quality problems. With the diversification of customer needs and the complexity of product functions, the difficulty of controlling product quality in the manufacturing process has also increased significantly, and traditional quality management methods can no longer cope with this complex situation.

[0003] In the past, quality management mainly relied on post-inspection, that is, to determine whether the product was qualified through testing after it was completed. Although this method can detect quality problems, it often lags behind the production process and cannot effectively prevent the occurrence of problems, resulting in waste of resources and economic losses. In order to solve these problems, statistical quality control methods have begun to be applied to the manufacturing industry. SPC can predict and prevent quality problems to a certain extent by monitoring and analyzing key process parameters in the production process, but with the complexity of manufacturing processes, relying solely on SPC is no longer sufficient to cope with the current production environment.

[0004] In the actual machining process, quality is the key indicator to measure whether the product is qualified or not. In order to meet these standards, the machining process often requires frequent adjustment of process parameters and repeated measurement of the workpiece. This method leads to a lot of repetitive work and reduces production efficiency. It also causes the test results to lag behind the actual production, and the process parameters cannot be adjusted in time, increasing the risk of batch unqualified products. In addition, over-reliance on the operator's experience for adjustment is subject to greater subjectivity and uncertainty.

[0005] With the refinement of customer needs and the increase in product complexity, how to use automation to predict product quality and optimize process parameters has become an important challenge for the manufacturing industry. By establishing a mathematical model between product process parameters and quality, and combining historical data with real-time production data for prediction, the number of process adjustments can be effectively reduced, the production process can be optimized, and efficiency can be improved. Quality prediction technology can not only monitor potential quality problems in the production process in real time, but also predict the quality level under future working conditions in advance, helping companies to adjust production processes in a timely manner, avoid quality problems, improve production efficiency and reduce losses.

[0006] However, further research shows that the above technology has the following deficiencies and defects:

[0007] 1. The mechanical product processing is affected by a variety of parameter characteristics, which may include unnecessary characteristic parameters, resulting in a large error between the predicted results and the actual results;

[0008] 2. A single algorithm model has limitations and instability and cannot be applied to many other complex scenarios. Even if it is applicable, the multi-source heterogeneous data in the processing process will bring great difficulties. Summary of the invention

[0009] Technical purpose: In view of the defects in the prior art, the present invention discloses a multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN. By optimizing the feedforward neural network with the improved Hippo algorithm, the prediction accuracy and efficiency of process parameters are effectively improved, the training time is reduced, and the local optimal problem in the traditional method is effectively avoided, ensuring higher accuracy and stability.

[0010] Technical solution: In order to achieve the above technical objectives, the present invention adopts the following technical solution.

[0011] A multi-feature quality prediction method based on an improved Hippo algorithm to optimize FFN includes the following steps:

[0012] Step 1: Obtain process parameter data and construct samples of the HO-FFN model after feature screening;

[0013] Step 2: Construct the HO-FFN model; determine the network topology of FFN, and use the improved Hippo algorithm to optimize the weight and bias parameters of FFN;

[0014] Step 3: Collect data on a preset threshold number of process parameter types during the quality processing of mechanical products, use the HO-FFN model to predict and output product quality in real time, obtain quality prediction results, and then make real-time adjustments to the processing process of mechanical products.

[0015] Beneficial effects:

[0016] 1. The present invention proposes a feature screening process of SVMCV, which screens redundant process parameters by combining recursive feature elimination and cross-validation, thereby reducing the complexity of the model and improving the ability to process high-dimensional data. This method ensures that the model still has good generalization ability in complex data environments and can also adapt to different process scenarios;

[0017] 2. By improving the Hippo algorithm to optimize the feedforward neural network, this method effectively improves the accuracy and efficiency of process parameter prediction. The optimized model has the ability to converge quickly, reducing the training time, while effectively avoiding the local optimal problem in traditional methods, ensuring higher accuracy and stability;

[0018] 3. The present invention effectively improves the accuracy and efficiency of process parameter prediction results by introducing the improved Hippo algorithm to optimize the feedforward neural network and combining it with the SVMCV feature elimination. On the basis of efficient experiments, the processing parameter combination with the optimal correlation is obtained, which is convenient for subsequent quality prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention;

[0020] Figure 2 This is a process parameter feature screening flow chart in an embodiment of the present invention;

[0021] Figure 3 The figure is a flow chart of improving the Hippo algorithm to optimize FNN in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0023] Example:

[0024] As attached Figure 1 As shown, a multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN of the present invention comprises the following steps:

[0025] Step 1: Obtain process parameter data and construct samples of the HO-FFN model after feature screening;

[0026] 1. Collect multi-dimensional process parameter data as data set A: In this embodiment, process parameter data are collected from the mechanical product quality processing process through multiple sensors. The process parameter data includes data of multiple dimensions such as machine tool temperature, motor speed, cutting speed, tool wear speed, etc. to obtain data set A. In data set A, each process parameter data is evenly divided into N groups.

[0027] 2. Perform data cleaning on dataset A to obtain cleaned dataset A; use a data cleaning algorithm to remove outliers and noise from dataset A, and perform interpolation processing to fill in missing data to obtain cleaned dataset A, thereby ensuring the integrity of dataset A.

[0028] 3. Each process parameter is used as a feature, and the cleaned data set A is screened by combining support vector machine and CV cross validation to eliminate the features with low correlation with the quality of mechanical products, and obtain data set B;

[0029] Due to the high complexity and redundancy of process data in the manufacturing process of mechanical products, many weakly correlated feature dimensions will be generated, that is, process parameter types with weak correlation with the quality of mechanical products. In order to reduce the impact of redundant features on the HO-FFN model, shorten the training time, and reduce the complexity of the HO-FFN model, feature screening is performed on data set A to eliminate features with low correlation. Feature selection methods include three types: filtering, embedding, and wrapping. In this embodiment, SVMCV combined with recursive feature elimination and CV cross-validation is selected to automatically perform feature screening of process data, that is, to screen out process parameter types with strong correlation with the quality of mechanical products; SVMCV refers to the support vector machine (SVM) recursive feature elimination combined with CV cross-validation. The SVM base model calculates the degree of correlation between each feature in data set A and the quality of mechanical products, and then cross-validates the degree of correlation of feature subsets through CV to eliminate features with low correlation with the quality of mechanical products. As shown in the attached figure Figure 2 As shown, the use of SVMCV for process parameter feature screening requires the selection of a base model to obtain the importance of each feature. For high-dimensional data features, the base model in this embodiment selects the SVM basic model to perform recursive feature elimination on the cleaned data set A, and then uses CV cross-validation to evaluate the performance indicators of the current SVMCV, and eliminates features with low correlation according to the preset threshold, that is, the process parameters form a new feature subset.

[0030] Different feature subsets x are obtained in SVMCV i , i = 1, 2, 3, ..., n, n is the total number of feature subsets with the lowest correlation eliminated by SVMCV, and the above training and elimination steps are repeated; then the parameter determination based on principal component analysis is introduced: the principal component analysis method is used to reduce the dimension of the feature subsets screened out by SVMCV, and the feature set is further optimized to retain the principal component parameters that can best explain the data variance, that is, the process parameters with the highest correlation; the calculation formula of the principal component analysis method is:

[0031]

[0032] Where C i is the feature subset x iThe process parameter score, X i is the feature subset x i The principal component vector, W i is the feature subset x i Variance contribution rate of the principal component. The calculation process of the variance contribution rate of the principal component is: after the principal component analysis, its eigenvalue can be obtained, and the eigenvalue is divided by the sum of all eigenvalues ​​to obtain the variance contribution rate of the principal component. After obtaining the arrangement results of the principal component analysis, the parameter characteristics when the cumulative contribution rate of the principal component reaches 90%-95% are selected from top to bottom, and the feature subset of this parameter is used as the screened feature subset as data set B. Each feature subset contains data of only one process parameter. The present invention performs feature screening through SVMCV, and screens redundant process parameters through a combination of recursive feature elimination, cross-validation and principal component analysis, thereby reducing the model complexity of FFN and improving the ability to process high-dimensional data. This method ensures that the model still has good generalization ability in a complex data environment, and can also adapt to different process scenarios.

[0033] The base model SVM is trained. The weight vector β in the base model SVM represents the contribution of each feature to the decision boundary of the base model SVM. The larger the weight, the more important the feature is to the base model SVM. By calculating the weight change after the feature is added, the influence of the feature on the decision of the base model SVM is evaluated, and the importance of each process parameter during processing is obtained, that is, the correlation between each process parameter and the quality of the mechanical product is obtained. The calculation formula of the importance of process parameters includes:

[0034]

[0035] Among them, β i , β j are the weight coefficients of feature i and feature j in the base model SVM, w j is the importance of feature j.

[0036] Then, the performance of the base model SVM is evaluated by CV cross validation, and then the process parameters with low feature importance are eliminated to form a new feature subset, and the performance score of the base model SVM is obtained by retraining, and the above elimination and training steps are repeated until the data set is empty. Then the scores of different feature subsets are compared, and the feature subsets with the highest performance index are selected as the feature subsets with the preset threshold number. In some embodiments of the present invention, the first three groups of feature subsets are selected as the filtered feature subsets to form data set B; the preset threshold number value range here is 3 to 5. If there are too few feature types, the prediction accuracy is not high, and if there are too many feature types, the complexity of FFN is increased.

[0037] 4. Normalize data set B to obtain normalized data set B;

[0038] The data set B is dimensionally normalized to ensure that the dimensional differences between the parameter dimensions of the preset threshold number will not affect the training of the model, eliminate the dimensional differences between different process parameters, and ensure that the impact of different parameters is balanced during model training.

[0039] In data set B, the process parameter data of the same dimension are normalized, and the calculation formula for normalization is:

[0040]

[0041] Wherein, z is the normalized process parameter data, i.e., the second process parameter data, and its value range is [0, 1]; x is the original process parameter data, i.e., the first process parameter data after cleaning, x max 、x min are the maximum and minimum values ​​in the original process parameter data respectively. This process can make FFN converge faster and avoid gradient explosion.

[0042] 5. Divide the normalized data set B to obtain samples of the HO-FFN model;

[0043] The normalized data set B is divided into a training set and a validation set according to a preset ratio, such as 8:2, as samples of the HO-FFN model. In addition, the training set and the validation set are both a number of groups of sample data, each group of sample data includes a preset threshold number of process parameter type data and labels; in some embodiments of the present invention, a group of sample data includes a machine tool temperature value, a motor speed value, a cutting speed value, and a label marked as qualified;

[0044] The label is a label indicating whether the quality of the mechanical product is qualified or unqualified under the set of process parameter type data with a preset threshold number; the label can be obtained by searching historical data or by using an existing simulation model. For example, in some embodiments of the present invention, the Deform software is used to construct a mechanical processing system simulation model, and a preset threshold number of process parameter type data is input to simulate whether the quality of the mechanical product is qualified. The training set is used to train the HO-FFN model, and the validation set is used to verify the performance of the HO-FFN model. When dividing, ensure that the samples have the integrity of the time series, and ensure that the distribution of parameter characteristics is consistent with the actual situation in the production and processing of mechanical products, that is, the data of all types of process parameters are evenly distributed in the training set and the validation set.

[0045] In order to prevent the HO-FFN model from being biased towards a certain type of process state, the sample data of each type of process parameter must be balanced during the division. If there is less data on certain types of process parameters, the Generative Adversarial Network (GAN) technology is used to generate additional sample data for data supplementation.

[0046] Step 2: Construct the HO-FFN model; determine the network topology of FFN, and use the improved Hippo algorithm (HO) to optimize the weight and bias parameters of FFN.

[0047] 1. Construct the network topology of FFN; the network topology of FFN includes input layer, hidden layer and output layer, and initialize the weight and bias parameters of the network. The process parameter data is used as input to complete the initial architecture of FFN.

[0048] In the input layer, the input layer receives the process parameter data of the training set in the sample, and each process parameter data represents an input parameter feature. The input of FFN is input in chronological order to ensure that FFN can learn the trends and fluctuations in the input data.

[0049] In the hidden layer, the hidden layer adopts a multi-layer structure, each layer contains several neurons, and is designed according to the following formula:

[0050]

[0051] Among them, Nh is the total number of neurons in the hidden layer, N i is the number of neurons in the input layer, that is, the number of input process parameter types, N o is the number of neurons in the output layer;

[0052] For the problem of predicting continuous values, the output layer node is set to 1, that is, N o =1, the activation function uses ReLU (Rectified Linear Unit) to avoid the gradient vanishing problem and enhance the nonlinear expression ability of the network. The hidden layer is responsible for extracting features from the process parameter data to identify abnormal conditions in the process. The calculation formula of the activation function is as follows:

[0053] g(u) = max(0,u)

[0054] Among them, u represents the input value of the activation function;

[0055] In the output layer, the number of neurons in the output layer is 1, and the predicted value is output. The predicted value is the quality evaluation label, and the value is 1 or 0, that is, qualified or unqualified.

[0056] 2. Use the training set and test set to train and test FFN;

[0057] After the network topology of FFN is constructed, the training set and validation set are used to train and test FFN respectively until the weight and bias parameters of FFN converge to the optimal weight and bias; during the test, the accuracy and precision of the indicators are calculated, and the calculation formula is as follows:

[0058]

[0059]

[0060] Among them, TP means that the predicted value is qualified and the true value is also qualified; TN means that the predicted value is unqualified and the true value is also unqualified; FP means that the predicted value is qualified and the true value is unqualified; FN means that the predicted value is unqualified and the true value is qualified;

[0061] 3. As attached Figure 3 As shown in the figure, the improved Hippo algorithm is used to optimize the weight and bias parameters of FFN. The improved Hippo algorithm gradually optimizes the weight and bias parameters of FFN by simulating group foraging behavior. The specific steps are as follows:

[0062] 3.1. Initialize the population:

[0063] In the population initialization stage of the improved Hippo algorithm, the traditional random initialization method easily leads to uneven population coverage and local optimal problems. In this regard, the Halton sequence is introduced to improve the population initialization position formula, and the low-discrepancy sequence is used to reduce the overlap between generated points to ensure more uniform coverage.

[0064] The Halton sequence is generated by using the inverse Vandermonde transform, and its calculation formula includes:

[0065]

[0066] in, is the Halton sequence, a k It is the kth digit after the mth number in the Halton sequence is expressed in base b, where m is the index of the Halton sequence, indicating the mth point in the Halton sequence, k is also an index, indicating the current accumulated position, and b is the base, that is, the base selected by the current dimension, which is also a prime number;

[0067] A series of evenly distributed points are generated through the Halton sequence, and these points are mapped to the value range of the solution space. The mapping formula is as follows:

[0068] W pq =W min +(W max -W min )*H pq

[0069] Among them, W pq is the initial value of the qth dimension of the th individual in the population, H pq is the value generated by the Halton sequence for the ,th individual and the qth dimension, and its value range is [0, 1]; W max, W min are the minimum and maximum values ​​of the qth dimension in the solution space, respectively. A linear mapping is performed between the minimum and maximum values ​​and the [0, 1] random numbers generated by the Halton sequence.

[0070] Repeat the above process for all individuals and all dimensions to update the initial position of the entire population.

[0071] 3.2. Fitness function calculation:

[0072] The weight and bias parameters of each individual in the population are pre-trained through FFN to obtain the current parameter quality index, calculate the accuracy, and use it as the fitness value to evaluate the model performance. The accuracy is used as the fitness function, and the calculation formula is as follows:

[0073]

[0074] Among them, TP means that the predicted value is qualified and the true value is also qualified; FP means that the predicted value is qualified and the true value is unqualified;

[0075] 3.3, introducing random perturbations and Levy flight mechanism to improve foraging behavior simulation;

[0076] The foraging behavior simulation includes two updating methods, one is random individual guided updating, the other is optimal individual guided updating; the two updating methods are used alternately: random updating is to ensure the diversity of the population and the global search capability in the early stage; optimal updating is to accelerate the convergence of the algorithm in the later stage and push the solution closer to the optimal solution.

[0077] The random individual guidance update process includes:

[0078] By randomly selecting an individual in the population as a reference, other individuals are updated based on the random individual. By introducing random perturbations and Levy flight mechanisms, the Hippo algorithm improves the prey position update formula to ensure that the population is always exploring in a larger search space, and most of the step lengths are shortened, leaving only the larger jump searches at certain stages, which better avoids the population from converging to the local optimal solution too early and increases the chance of finding a better solution. The update formula is as follows:

[0079] W new =W old +ξ*(W rand -W old )+ζ*rand()+△W

[0080] b new =b old +ξ*(b rand -b old )+ζ*rand()+△b

[0081] Among them, W new With b new is the updated weight and bias of the individual, W old With b old is the current weight and bias of the individual, W rand With b rand are the weight and bias of the random individual, ξ and ζ are constants for adjusting the step size, rand() is a small random perturbation introduced to prevent the HO-FFN model from falling into the local optimum; △W is the random perturbation step size for adjusting the weight of the Levy flight mechanism, and △b is the random perturbation step size for adjusting the bias of the Levy flight mechanism;

[0082] The optimal individual guidance update process includes:

[0083] In each iteration of the hippopotamus foraging strategy, by taking the best individual as a reference point to guide the optimization of other individuals, the weights and biases will gradually move in the direction of the current optimal fitness, accelerating the convergence of the algorithm and optimizing the current solution. The update formula is as follows:

[0084] W new =W old +ξ*(W best -W old )+ζ*rand()+△W

[0085] b new =b old +ξ*(b best -b old )+ζ*rand()+△b

[0086] Among them, W best , b best is the weight and bias of the current optimal individual;

[0087] 3.4. Re-evaluate fitness:

[0088] After the weight and bias parameters are updated, the FFN is retrained and the fitness value of each individual is calculated. By comparing the changes in fitness values ​​before and after the update, it is effectively evaluated whether the updated weights and biases have improved the performance of the model.

[0089] 3.5. Termination condition judgment:

[0090] The iterative process of weight update is completed when the fitness function value tends to be stable or reaches the set maximum number of iterations T. maxAfter the end, that is, whether the fitness function value reaches the preset convergence threshold, the optimal weight and bias outputted at the end will be applied to the FFN model according to the judgment condition, otherwise it returns to the foraging behavior simulation step and continues iterative optimization. The final HO-FFN model is obtained by optimizing the FFN with the improved Hippo algorithm.

[0091] Step 3: Collect the preset threshold amount of process parameter data during the quality processing of mechanical products, use the HO-FFN model to predict the product quality in real time, obtain the quality prediction results, and then make real-time adjustments to the processing of mechanical products; when the HO-FFN model results determine that a mechanical product is qualified, it means that the mechanical product meets the design or quality standards, can be delivered normally and retain quality records; similarly, when the HO-FFN model results determine that a mechanical product is unqualified, it means that the mechanical product does not meet the quality standards, and the unqualified mechanical products need to be sorted and isolated, and analysis is used to determine whether to rework or scrap. In the actual process production process, the H0-FFN model can also be verified for effectiveness using accuracy and precision based on this prediction result and the actual quality results collected in real time on site.

[0092] The present invention proposes a method, using SVMCV feature screening model with low correlation parameter data, using SVMCV to eliminate low correlation data, and then based on principal component analysis, selecting process parameters with cumulative correlation of 90%-95%, improving the prediction performance of the prediction model and eliminating unnecessary feature data. In the Hippo algorithm, by introducing the Halton sequence to initialize the population, introducing the Levy flight mechanism and random perturbation to improve the predator position update formula, the improved Hippo algorithm can initialize the population through the Halton sequence in the initial stage of the algorithm, and can make the individuals of the initialized population more evenly distributed in the population initialization stage, which helps to increase the diversity of the population and improve the global search ability of the algorithm; introducing the Levy flight mechanism and random perturbation in the foraging stage of the algorithm can improve the optimization accuracy and efficiency of the algorithm, and avoid the situation where the algorithm falls into the local optimal solution in the late evolution stage; then the improved Hippo algorithm is applied to the FFN prediction model, and the prediction accuracy of the parameters of the FFN prediction model is improved.

[0093] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN, characterized in that: The following steps are involved: Step 1: Obtain process parameter data and construct samples of the HO-FFN model after feature screening; Step 2: Construct the HO-FFN model; determine the network topology of FFN, and use the improved Hippo algorithm to optimize the weight and bias parameters of FFN; Step 3: Collect data on a preset threshold number of process parameter types during the quality processing of mechanical products, use the HO-FFN model to predict and output product quality in real time, obtain quality prediction results, and then make real-time adjustments to the processing process of mechanical products.

2. According to claim 1, a multi-feature quality prediction method based on improved Hippo algorithm to optimize FFN is characterized in that: In step 1, process parameter data is obtained to construct a sample of the HO-FFN model, including the following steps: Collect multi-dimensional process parameter data as data set A; Perform data cleaning on data set A to obtain cleaned data set A; Each process parameter is used as a feature, and the cleaned data set A is screened by combining support vector machine and CV cross validation to eliminate the features with low correlation with mechanical product quality, and the feature subset with a preset threshold number of features with high correlation with mechanical product quality constitutes data set B; Normalize the data set B to obtain a normalized data set B; The normalized data set B is divided to obtain samples of the HO-FFN model.

3. The multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN according to claim 2, characterized in that: The cleaned data set A is subjected to feature screening by combining support vector machine with CV cross validation, including: The SVM basic model is used to perform recursive feature elimination on the cleaned data set A, and then CV cross-validation is used to evaluate the performance indicators of the current SVMCV, and features with low correlation are eliminated according to the preset threshold; repeated screening is then performed, and the feature subset with the preset threshold number of the highest performance indicators is used as the filtered feature subset through principal component analysis to form data set B.

4. According to claim 2, a multi-feature quality prediction method based on improved Hippo algorithm to optimize FFN is characterized in that: The normalized data set B is divided to obtain samples of the HO-FFN model, including: The normalized data set B is divided into a training set and a validation set according to a preset ratio as samples of the HO-FFN model. Both the training set and the validation set contain several groups of sample data. Each group of sample data includes a preset threshold number of process parameter type data and labels; the label is a label indicating whether the quality of the mechanical product is qualified or unqualified under the preset threshold number of process parameter type data.

5. The multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN according to claim 1, characterized in that: In step 2, after the network topology of FFN is constructed, the training set and test set are used to train and test FFN. During the test, the accuracy and precision of the indicators are calculated. The calculation formula is include: Among them, TP means that the predicted value is qualified and the true value is also qualified; TN means that the predicted value is unqualified and the true value is also unqualified; FP means that the predicted value is qualified and the true value is unqualified; FN means that the predicted value is unqualified and the true value is qualified.

6. The multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN according to claim 1, characterized in that: In step 2, the weight and bias parameters of FFN are optimized using the improved Hippo algorithm, including the following steps: Initialize the population and introduce the Halton sequence to improve the population initialization position; Fitness function calculation: The weight and bias parameters of each individual in the population are trained through FFN, and its fitness value is calculated, and the accuracy rate is used as the fitness value; The random perturbation and Levy flight mechanism are introduced to improve the foraging behavior simulation. The foraging behavior simulation includes random individual guidance update and optimal individual guidance update, which are performed alternately. Reassess fitness; The termination condition judgment stops.

7. The multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN according to claim 6, characterized in that: The Halton sequence is introduced to improve the population initialization position, including: The Halton sequence is generated using the inverse Vandermonde transform. The calculation formula includes: in, is the Halton sequence, a k It is the kth digit after the mth number in the Halton sequence is expressed in base b, where b is the base and is also a prime number; A series of evenly distributed points are generated through the Halton sequence, and these points are mapped to the value range of the solution space. The mapping formula includes: W pq =W min +(W max -W min )*H pq Among them, W pq is the initial value of the qth dimension of the th individual in the population, H pq is the value generated by the Halton sequence in the pth individual and the qth dimension, and its value range is between [0, 1]; W max , W min are the minimum and maximum values ​​of the qth dimension in the solution space respectively.

8. The multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN according to claim 6, characterized in that: The random perturbation and Levy flight mechanism are introduced to improve the foraging behavior simulation, which includes random individual guidance update and optimal individual guidance update. The random individual guidance update formula includes: IN new =In old +ξ*(W rand -IN old )+ζ*rand()+△W b new =b old +ξ*(b rand -b old )+ζ*rand()+△b Among them, W new With b new is the updated weight and bias of the individual, W old With b old is the current weight and bias of the individual, W rand With b rand are the weight and bias of the random individual, ξ and ζ are constants for adjusting the step size, and rand() is the introduced small random disturbance; △W is the random perturbation step size for adjusting the weight of the Levy flight mechanism, and △b is the random perturbation step size for adjusting the bias of the Levy flight mechanism; The optimal individual guidance update formula includes: IN new =In old +ξ*(W best -IN old )+ζ*rand()+△W b new =b old +ξ*(b best -b old )+ζ*rand()+△b Among them, W best 、b best are the weights and biases of the current optimal individual.

9. The multi-feature quality prediction method based on the improved Hippo algorithm to optimize FFN according to claim 6, characterized in that: The termination conditions include: The iterative process of weight update is completed when the fitness function value tends to be stable or reaches the set maximum number of iterations T. max The optimization ends when the fitness function value reaches the preset convergence threshold. The optimal weights and biases finally outputted will be applied to the FFN model according to the judgment conditions. Otherwise, the optimization returns to the foraging behavior simulation step and continues iterative optimization.

Images