The paper reports on a study aimed at developing a neural network model for detecting anomalies in the process data of the heat treatment of raw phosphate ore, which is designed to provide timely information on detected anomalies to case-based situational models. The neural network model is based on the autoencoder architecture, which is being effectively used in various applied fields to detect anomalies in data. The original feature of the proposed solution is that dual-channel processing of input process data is used. One channel uses a 1D-CNN (convolution neural network), the other one applies an LSTM recurrent network. This network architecture facilitates identification of different structures in the data and enables one to study their features more deeply when searching for anomalies. The testing of the proposed autoencoder architecture, which was performed as numerical experiments in the Matlab environment, demonstrates that, as compared to single-channel analogs, the use of dual-channel architecture provides a larger autoencoder input/output error when data with anomalies are fed to the input. This indicates the expediency of using a dual-channel version. The research results can find application in the information support for systems of identifying pre-emergency and emergency situations in the operation of processing equipment.