In recent years, soft sensors have been widely used for real-time estimation of challenging process variables. However, their accuracy is hindered by strong nonlinearity and dynamics. To address these issues, this paper proposes KM-SAE-SEL-LSSVR, a local semisupervised soft sensor modeling method based on deep learning. This approach integrates K-means clustering using mutual information and Markov distance to localize labeled samples, reducing the impact of process nonlinearity. It also introduces a semisupervised framework with stacked autoencoders (SAEs) to fuse multirate unlabeled data effectively. Additionally, a LSSVR tailored for small sample sizes predicts target values using SAE-reconstructed samples. Model parameters are optimized using the sparrow search algorithm, and a selective integrated learning method based on statistical assumptions improves overall performance. Experimental results with real sulfur recovery unit (SRU) data and simulated penicillin fermentation process (PCA) data show KM-SAE-SEL-LSSVR outperforms traditional methods in prediction accuracy and generalization.