||Training a model to detect patterns of interrelated events that form situations of interest can be a complex problem. This is because such situations tend to be uncommon, which means that only sparse data is avail- able. This is especially true when using input data that tends to be noisy. Our approach to this problem is a hybrid neuro-symbolic architecture based on Event Calculus that can perform Complex Event Processing (CEP), which we have called DeepProbCEP. This system allows us to use a neural network to interpret inputs while also allowing us to define rules that express the pattern of the complex event in the logic layer. This approach has multiple advantages in respect to a pure neural network approach, as (i) it needs fewer labelled data thanks to the end-to-end neural and logical learning procedure, (ii) it is robust against noise and adversarial attacks in the form of training data poisoning and (iii) it is capable of training to classify individual events even when training in an end-to-end manner. We demonstrate this by comparing the performance of DeepProbCEP and a pure neural network approach in a synthetically generated dataset based on Urban Sounds 8K.