Prediction of brake pad wear and remaining useful life considering varying vehicle mass and an experimental holistic approach

Abstract

Estimation of passenger car brake pad wear is considered, including real-time mass estimation to account for variations in vehicle mass. A longitudinal dynamics model is used to estimate mass, the application of the brakes, and a longitudinal braking force. Wear is modelled from brake work based on the product of the braking force and distance travelled during braking. Data is provided by standard vehicle on-board diagnostics data and a sensor fusion based on an inertial measurement unit and satellite position data, sampled from a single car with 10,200 km of on-road driving, including 31 brake pad wear measurement samples, and driving with up to 23% increased vehicle mass. The mass is accurately estimated within ±5% of the actual. Braking states determined from comparisons of measured and modelled deceleration provide accuracies of ≈91%, precisions of ≈96%, and recall rates of ≈86%, with inclusion of varying mass only resulting in minor differences. Brake work is estimated with (a) constant mass, (b) estimated mass, (c) actual mass, and (d) actual mass and braking states. Compared to (a), (b) and (c) provide 9% and 11% increased work, while (d) provides an increase of about 5% compared to (c). Wear coefficients are fitted on the initial half of the wear data considering (a), (b), and (d). Here (a) and (b) provide similar wear rates of k≈−1.22 mm/GJ, while (d) is about 5% lower. Validating on the second half of the data, (b) and (d) provide very similar wear estimates, with overall lower errors than (a). At the final sample, however, (a) provides a slightly lower absolute error, overestimating the wear by about 2%, while (b) and (d) underestimate it by about 3%. RUL, in terms of distance, is estimated across about 9,500 km of data by estimating a running ratio between cumulated work and distance. Here only (a) and (b) are considered, with (a) overestimating the RUL by 600 km at the actual wear limit, while (b) underestimates it by 190 km. A distance-based model is also considered, resulting in a constant RUL error of about 2,000 km.