<p>To solve the problem of ineffective extraction of advanced features that reflect the essence of customer behavior and enhance customer stickiness. A deep perception method for customer behavior in digital business halls based on multimodal data integration has been proposed. Retrieve multimodal customer information from the multimodal data platform layer of the digital business hall, including text, voice, images, and videos, as data sources. The data collection layer utilizes different collection nodes to collect this data. The data feature layer uses Deep Belief Networks (DBN) to extract features from collected multimodal data and integrate them into a multimodal data feature library to achieve feature integration of customer multimodal data. Using naive Bayes classification method, customer behavior is classified into three categories based on the characteristics of multimodal data: basic information, behavioral information, and service information. The ontology search and semantic extension technology effectively solve the semantic heterogeneity problem of information integration perception in customer basic information and service information perception, providing customers with more complete information query results. Provide customers with high-quality knowledge resources closely related to their behavior based on their level of knowledge needs. The experimental results show that the sliding window length of this method is suitable between 240 and 260 in the processing of multimodal data streams, with high data integration accuracy and strong advantages in deep perception of customer behavior information. It can significantly improve the service quality of digital business offices and enhance customer stickiness.</p>