A non-intrusive indirect pressure estimation and early warning is proposed for high-pressure water-jet pipelines. The EfficientNet deep learning model is employed to evaluate the quality grade of steel plate surface before and after cleaning based on the images. Observable sequences are constructed by the quality grades and system-set pressure. A hidden Markov model (HMM) is then adopted to capture the latent mapping between such sequences and the pressure states of water-jet pipelines, and the pressure state sequences are decoded. Based on the characteristics of the decoded sequences, abnormal pressure can be detected and warned. Industrial experiments demonstrate that the proposed method can achieve an accuracy of 91.2% and an F1 score of 88.3% in abnormal pressure detection, with performance significantly superior to that of the threshold-based method and logistic regression. This method is of practical guiding significance for the intelligent operation and maintenance of high-pressure water-jet equipment.