Wideband acoustic and acceleration sensors are frequently used to monitor the condition of machines and manufacturing processes. The measurement data can be analyzed using machine learning methods to detect potential errors at an early stage. In this work a workflow is demonstrated, which allows to increase the energy efficiency of the data acquisition chain by analog signal pre-processing next to the. This is achieved by near-sensor data reduction with feature extraction in the analog domain before digitization, which allows feature sampling well below the Nyquist frequency. Especially promising and widely applicable are features from the frequency domain, which are implemented using an analog Fourier series approach. However, specifications for a circuit implementation cannot be derived directly. The impact of circuit non-idealities can typically be evaluated after classification only. Unfortunately, simulations with several gigabytes of raw input data are hardly feasible with analog circuit simulators due to long simulation times. Therefore, a workflow based on a Matlab model of the analog feature extraction was developed. This allows the extraction and evaluation of features, including non-idealities of the circuit, by statistical analysis. The highly parallelized workflow speeds up simulation time by up to four orders of magnitude. This enables the selection and parameterization of features that lead to the best classification results.