This paper proposes a method for cable failure detection in Cable Driven Parallel Robots (CDPR) which is based on the exploitation of the load observer (LO) concept together with machine learning algorithms. By just exploiting the dynamic model of each actuator in the conditions of no load, a LO is designed for each motor to estimate the presence of a load coupled through a cable. Since the load instantaneously goes to zero for the motor with a broken cable, a simple but effective and robust signature of failure can be inferred, to provide reliable detection even in the case of various model mismatches. Additionally, the LO is not computationally demanding since just motor measurements are required, thus avoiding any direct measurement (and a dynamic model as well) on the end-effector. The detection of a failure in made through supervised classification algorithms based on artificial intelligence. The training of the machine learning algorithm is based on an “hybrid” approach: the dataset includes several failure cases which are numerically generated through a system digital twin developed through the multibody system theory, together with measurements of the real system in non-failing conditions. Different classification algorithms are considered, together with different sets of input variables to be fed to the classifier. Two numerical examples are proposed, by showing the method capability in handling both fully actuated and redundantly actuated CDPRs under cable failure.