MACHINE FIELD DATA ANALYSIS AND PREDICTION FOR
SEMI-AUTOMATION OF CATERPILLAR'S EXCAVATOR MACHINE
This project was done by me at advanced research section in Caterpillar India, Engineering Design Centre - 1. This was a month-long short project. The primary objective of the project was to improve the performance of Caterpillar's excavator machine. The performance of existing excavator machines of caterpillar depends heavily upon the manual joystick (gear-stick like Instrument used to control the excavator) control by the machine operator. The excavator operator has control on 4 types of motion using his joystick control - Swing, Bucket, Stick and Boom.
My primary research topic was to improve the performance of the machine and enable the machine to assist the operator by predicting the next loading and unloading cycle as the machine moves from point A to point B. The results of my research would be used by the Caterpillar team to prepare the machine’s hydraulic system for the next cycle. This foresighted preparation would increase the efficiency of the machine by avoiding any sudden jerks and vibrations caused in the operation due to control by the machine operator, lead to a smooth movement of the machine. This also proves to be a great advantage for the machine’s mechanical system and the operator. Since the machine can predict the next cycle, even a rough movement by the operator is sufficient to move the machine to precise locations. Hence it also prevents work fatigue of the operator.
Subsequently, a closed form relationship for Force and Velocity was derived using data of other available parameters. The test machine had Force and Velocity sensor built in it. Since these sensors are very expensive and tough to instrument, I was looking at the possibility of eliminating the need for these sensors.
Currently, all these movements are physically implemented using a centralized hydraulically powered system. Bucket, Stick and Boom motion involves the activation of the respective hydraulic piston. Swing motion is the rotation of upper structure with respect to the under-carriage. The machine operator controls the joysticks that deliver hydraulic signal pressures to the main hydraulic valve system to provide actuation of each of the implement and swing systems (i.e., they were hydraulic over hydraulic systems). The boom, stick and bucket systems used two-way hydraulic cylinders for actuation while the upper structure system used a hydraulic motor, or swing motor, to provide left and right rotation relative to the undercarriage. The hydraulic system included a nitrogen-filled gas type accumulator to provide oil to the circuit as makeup oil during combined operations and for lowering implements immediately after the engine has been stopped.
Data obtained from 90 degree Truck Loading operation of 608 Machine was used. The raw data of the lever signals were very noisy. So the data was initially filtered using a Butterworth low pass filter as an effort to remove the high-frequency noises present in the data. By visual inspection, it was found that the cycle time is around the order of around 20s (0.05 Hz). To find a rough time period of each cycle, Discrete Fast Fourier Transform was applied to the data. Thus the data was converted from time domain to that of a frequency domain. This gave a rough idea of time period of the cycle.
Based on the rough cycle time, we now split the cycles using some rules which are specific to the operation. Discrete Fast Fourier Transform of the individual cycles was carried out with the hope that these value might show a trend. Based on this trend, next set of FFT values could be computed and the next cycle could be predicted through Inverse Discrete Fourier Transform. However, since the cycles were varying in an inconsistent manner it there was no specific trend which was observable in the FFT values of the individual cycles. Given the varying nature of cycles, a Dirichlet wavelet transform was seen to be a good solution. Since we are able to split the cycles of the data, a wavelet transform was applied to each of the cycles and some significant features were extracted. A regression algorithm (based on Levenburg-Marquardt gradient method ) was used to predict the next set of features for the wavelet and the inverse transform was performed on the wavelet to yield the next cycle. This method proved to be very efficient and was able to predict the next cycle to above 90% accuracy.