Temperature Compensation for the Low-Coherent Optical Settlement Sensor by Using an IGA-BP Neural Network

A low-coherent optical sensor combined with a hydrostatic leveling configuration has been developed to monitor ground settlement (GS) in engineering. However, if the monitoring sites extend in a large span, the inhomogeneous spatial temperature field can deteriorate the performance of the sensor determined by the temperature-dependent liquid density distribution in hydrostatic system. Therefore, the fluctuations of the environmental temperature cause a hydrostatic leveling error in a time-varying way. Based on the measured results, a compensation method using an improved back-propagation (BP) neural network is proposed to suppress the effects of temperature. Besides the network structure itself, a genetic algorithm incorporated with a specially designed fitness function, and Blend Crossover (BLX-α) operator is employed to optimize the weights and biases of the neural network. These treatments have improved the global searching capability, training speed, and convergence efficiency. Based on training with 9,000 samples, an improved genetic algorithm (GA) combined with a BP neural network (IGA–BP) reduces temperature-induced error by approximately 50%. The system achieves 0.3 mm accuracy across a 20 cm measurement range. It is proved by a practical testing experiment configured in a binary temperature field lasting for one month. This method could be a good reference for further practical applications.