The operation conditions of multi-product pipeline changes frequently and it is difficult to judge the operation state accurately. Therefore, the recognition and monitoring by on-site personnel is easy to cause misjudgment. In order to realize the accurate recognition of pipeline operation conditions, considering the physical spatial characteristics of the pipeline, the operation parameters (pressure, flow rate and density) of each station are sorted out. Considering the time series characteristics of pipeline operation, operating data matrix is formed to overcome the transient disturbance at a single moment based on the SCADA data. Aiming at the high-dimensional and non-linear characteristics of pipeline operating data, the powerful feature compression and reconstruction capabilities of the convolutional autoencoder (CAE) are used to reduce the noise of pipeline data. T-distributed stochastic neighbor embedding algorithm (T-SNE) is used to perform dimensionality reduction and clustering processing on pipeline data, and finally the model based on CAE-TSNE for pipeline operation condition recognition is established. Taking two real multi-product pipeline as example, the mainstream machine learning nonlinear classification models (ANN, DT and RF) were compared with the proposed method. The results show that the operating condition identification model based on CAETSNE has the highest accuracy, and the recognition rate of clustering identification of operating data after noise reduction can reach 99%, which can guide the operation and management of on-site pipelines.