This paper proposed neural networks to continuously provide alternative constructed signals for vehicle and wheel speed sensors utilized for the Anti-Lock Braking System (ABS), which serves as the fault tolerant control method. These alternative constructed signals are used for two purposes. The first is to generate residual signals, and the second is to be adopted instead of isolated faulty signals. The residual signal is generated by extracting the difference between the alternative constructed signals and the corresponding actual signals. These residual signals serve as an indication of fault occurrence and to express that fault severity. Whenever a fault occurrence is detected and diagnosed in one of the sensor’s signals, the faulty signal is isolated and replaced by the corresponding constructed signal to maintain the system's normal behavior under a faulty condition. The range of data covered under the proposed estimating neural networks is huge, continuous in time, and not sampled. In this work, the range of the data lies between [50 to 120 km/h] when the braking is started. That cannot be performed by any available method. These models' training process is based on the Levenberg-Marquardt (LM) algorithm, implemented and tested by MATLAB/Simulink. The results show that these models can accurately map the measured data into the desired output through the best-fit functions. The fast response of the trained models makes them suitable for real-time alternative signals for fault-tolerant purposes for speed sensors during hard or panic braking.