ABSTRACT In the process of modern industrial production, soft sensing technology is often used to predict the target variables that are difficult to be directly measured by hard instruments. However, the input variables used for prediction are not all closely related to the output. Feature selection (FS) aims to select the features highly related to the target variables and discard the redundant features. How to select the optimal feature subset and reduce the operation cost while ensuring the prediction accuracy becomes a key problem. A multi‐strategy fusion binary Zebra optimization algorithm (MFBZOA) was proposed to select the optimal feature subset in a prediction model. Firstly, the Cauchy inverse cumulative distribution function is used to mutate individual positions in the defense stage of the zebra optimization algorithm (ZOA), and then reproductive behavior is introduced into the algorithm to increase the diversity of the solution set and improve the overall quality of the solution. Finally, the Cauchy mutation strategy is introduced to disturb the worst individuals in the zebra population and increase the probability of jumping out of the local optimum. Firstly, the proposed improved ZOA is combined with the Zebra optimization algorithm, Golden sine algorithm, Whale optimization algorithm, Frilled lizard optimization, Human evolution optimization algorithm, Coatis optimization algorithm, and Goose optimization algorithm to perform CEC2022 function optimization simulation experiments to verify its effectiveness. Then, MFBZOA and the above comparison algorithms are used as search strategies respectively, combined with the wrapper FS method driven by a multi‐layer perceptron to solve the FS problem of four industrial process data and build the corresponding prediction model. Then, the optimal feature subset selected by each algorithm is used in the prediction experiment. The simulation results show that MFBZOA can effectively select the optimal feature subset, improve the global search ability and local search ability, and maintain good prediction accuracy and generalization performance.

ABSTRACT Cloud computing system task scheduling optimization has garnered substantial attention because it directly influences resource utilization, service quality, and system energy consumption. To address the demands of cloud computing environments, an enhanced seahorse optimizer using alpha balance factor and t‐distribution mutation is proposed. Firstly, the alpha balance adaptive mechanism was proposed. The original alpha balance factor exhibits a relatively smooth downward trend. To overcome this limitation, the standard normal distribution random numbers are considered to add, which introduces greater volatility and enhances the algorithm's ability to jump out of the local optima. Secondly, a multi‐dimensional t‐distribution mutation operator was designed, taking into account both exploration and exploitation. This promotes the dynamic exploration and exploitation of seahorse individuals during the movement behavior stage, enhances population diversity, reduces the possibility of aggregation at local points, and accelerates the convergence speed of the algorithm. The optimal variant HTSHOt2 was selected by validation on the CEC‐2022. Finally, HTSHOt2 was applied to small‐scale and large‐scale cloud computing task scheduling scenarios for the first time, achieving coordinated optimization of task completion time, resource utilization rate, and energy consumption. The results show that the total cost of HTSHOt2 in small‐scale scenarios ranges from 2.28E‐01 to 2.62E‐01, reducing by 12.37% to 24.25%. The total cost range in large‐scale scenarios is between 2.75E‐01 and 2.87E‐01, reducing by 4.33% to 7.72%. In addition, HTSHOt2 spends very little on price cost, load cost, and time cost, all of which are far superior to other comparison algorithms. This indicates that HTSHOt2 can effectively solve the problem of task scheduling optimization in cloud computing systems and has powerful performance.

Long and short-term memory (LSTM) has been used in soft-sensor modeling of industrial processes in recent years. However, LSTM still has many defects for soft-sensor. This article proposes a variational autoencoder bidirectional LSTM soft-sensor modeling method based on batch training (Bt-VAEBiLSTM). First, the training samples are divided into multiple batches according to the time series, in order to reduce the influence of abnormal points and noise, the variational autoencoder is then used to reconstruct the training samples in each batch in order to solve the problem of the global LSTM model discarding critical data information during training; this article proposes a batch training method that is to say the reconstructed samples are trained in batches according to the time series. After the training of a batch samples is completed, the structural parameters of the previous local bidirectional LSTM (BiLSTM) model are shared with the next local BiLSTM model as the initial parameters to retain important state information. At the same time, in order to prevent the Bt-VAEBiLSTM model from overfitting, the L2 regularization term is introduced in the loss function. The effectiveness of the proposed method is verified by simulation experiments on the grinding and classifying process.