Modern time-domain astronomy is capable of collecting a staggeringly large amount of data on millions of objects in real time. Therefore, the production of methods and systems for the automated classification of time-domain astronomical objects is of great importance. The Liverpool Telescope has a number of wide-field image gathering instruments mounted upon its structure, the Small Telescopes Installed at the Liverpool Telescope. These instruments have been in operation since March 2009 gathering data of large areas of sky around the current field of view of the main telescope generating a large dataset containing millions of light sources. The instruments are inexpensive to run as they do not require a separate telescope to operate but this style of surveying the sky introduces structured artifacts into our data due to the variable cadence at which sky fields are resampled. These artifacts can make light sources appear variable and must be addressed in any processing method.The data from large sky surveys can lead to the discovery of interesting new variable objects. Efficient software and analysis tools are required to rapidly determine which potentially variable objects are worthy of further telescope time. Machine learning offers a solution to the quick detection of variability by characterising the detected signals relative to previously seen exemplars. In this paper, we introduce a processing system designed for use with the Liverpool Telescope identifying potentially interesting objects through the application of a novel representation learning approach to data collected automatically from the wide-field instruments. Our method automatically produces a set of classification features by applying Principal Component Analysis on set of variable light curves using a piecewise polynomial fitted via a genetic algorithm applied to the epoch-folded data. The epoch-folding requires the selection of a candidate period for variable light curves identified using a genetic algorithm period estimation method specifically developed for this dataset. A Random Forest classifier is then used to classify the learned features to determine if a light curve is generated by an object of interest. This system allows for the telescope to automatically identify new targets through passive observations which do not affect day-to-day operations as the unique artifacts resulting from such a survey method are incorporated into the methods.We demonstrate the power of this feature extraction method compared to feature engineering performed by previous studies by training classification models on 859 light curves of 12 known variable star classes from our dataset. We show that our new features produce a model with a superior mean cross-validation F1 score of 0.4729 with a standard deviation of 0.0931 compared with the engineered features at 0.3902 with a standard deviation of 0.0619. We show that the features extracted from the representation learning are given relatively high importance in the final classification model. Additionally, we compare engineered features computed on the interpolated polynomial fits and show that they produce more reliable distributions than those fit to the raw light curve when the period estimation is correct.