Determining Anomalies in Radar Data for Seedbed Tine Harrow Operation

Abstract: The agricultural industry is constantly evolving with automation as one of the current main focuses. This thesis involves the automation of a seedbed tine harrow, specifically the control of the tillage depth. The tillage depth is instrumental to farming as it determines the quality of the tilth, how well clods are broken up, and how well the soil aggregates are sorted. Poor control of the tillage depth could result in a bad harvest for the farmer. To control the tillage depth, several pulse radar sensors are installed on the harrow. The sensors measure the distance from the tines of the harrow to the ground. This distance is used in a control-loop that controls the hydraulic actuators that lifts and pushes down the frame of the harrow. Because of the rough working conditions of the tine harrow, the pulse radar sensors are in danger of being damaged or disturbed. A sensor not working as intended will lead to poor control of the tillage depth or even an unstable control system. The purpose of this thesis is to develop diagnosis systems to detect and generate an alarm if the output of a sensor is faulty. Four different systems are developed, three machine learning approaches and one model based approach. To be able to test and train models without having to go out on a field with a real harrow, a test rig is available. The test rig emulates a harrow driving on a field and the tests are designed to imitate plausible sensor errors. The models trained on and tuned to the test rig data are validated with data gathered from a real tine harrow.  The validation data from the harrow reveal that the main difference between the field data and test rig data are the vibrations and the sensor heights. The test rig produces negligible amounts of vibrations whereas the vibrations on a real harrow are immense. These differences affect the performances of the models and some tuning have to be done to the models to accommodate for the vibrations. The performance of the model based approach is good and no larger adjustments have to be made to it. The machine learning models created from the test rig data do not work in the field and new models are trained using field data. The new models are accurate and show great potential; albeit, it would be necessary to collect a lot more data for further training. Specifically, training the machine learning models on varying heights. In conclusion, the test rig data is similar to the field data but the vibrations in the system is missing and the heights differ. The missing vibrations results in that the models do not work as intended on field data. The conventional diagnostics approach works, but the generated alarms are binary meaning that the alarm only reveal if the signal is good or bad and does not provide any nuance. The machine learning models does provide nuance, meaning that the model can detect errors, what is causing the error, and warn if an error is about to occur. However, the machine learning models need a lot of data to train on to make this happen.