There can be multiple approaches or trial and error methods to get some solution but it is always better to go step by step while solving such problems. Probably, before going to model training, the data must be explored and checked whether the data is clean enough for analysis.
After cleaning the data, sometimes exploratory data analysis helps in deciding the useful features. Depending on the dimensions, one can go ahead and plot some charts or perform statistical tests to detect anomalies.
In your case, the exploratory data analysis might give some indications of which brands are more prone to errors (just an example). This analysis will be completely dependent on the nature of the data.
After understanding the data, one can start with basic models like decision trees or if one wants to start with LSTMs then, a single layer of LSTM can be used and model architecture can be improved gradually depending on results. If there is a similar problem that was solved using any pre-defined architecture then one can proceed with it.
There is no fixed answer to this question because there the nature of data is not known, but for any time series or sequence-related problems like text or speech, LSTMs, BiLSTMs or GRUs will perform well.