As part of Big Data trends, the ubiquitous use of the Internet of Things (IoT) in the industrial environment has generated a significant amount of network traffic. In this type of IoT industrial network where there is a large equipment heterogeneity, security is a fundamental issue; thus, it is very important to detect likely intrusion behaviors. Furthermore, since the proportion of labeled data records is small in the IoT environment, it is challenging to detect various attacks and intrusions accurately. This investigation builds a semisupervised ladder network model for intrusion detection in the Industrial IoT. This model considers the manifold distribution of high-dimensional data and incorporates a manifold regularization constraint in the decoder of the ladder network. Meanwhile, the feature propagation between layers is strengthened by adding more cross-layer connections in this model. On this basis, a random attention-based data fusion approach is proposed to generate global features for intrusion detection. The experiments on the CIC-IDS2018 dataset show that the proposed approach can recognize the intrusion with less false alarm rate, while model training is time efficient.