This thesis is an exploratory work of unsupervised anomaly detection in multi-variate time-series machine data with methods originating from natural language processing. The foundation is laid by tokenizing the time-series data, i.e., converting the numeric machine data into symbolic data similar to textual data. The process of tokenization is realized by discretizing the data and assigning unique tokens to the discrete values. The symbolic sequences obtained are then inspected for anomalies with two different approaches. The first method is based on word n-gram language models. A n-gram is a sequence of words of length n. The counts of those n-grams in the data set are computed and weighed with a measure for the importance of a word to a document, the term frequency-inverse document frequency measure, to derive anomaly scores for each sequence. The second approach presented utilizes a machine learning model, more specific a masked language model with a transformer architecture at its core. Random tokens in the input sequences get masked and the transformer is trained to recreate the numeric sequences. When an input sequence that has not been used for training outputs a diverging numeric sequence, anomalies in this sequence are expected. Both anomaly detection methods were programmed and successfully applied to an unlabeled data set originating from instrumented machinery used for ground improvement of building foundations. The results indicate that both approaches are principally functional and strongly justify continued work in this area, especially on the machine learning model.