答辩博士：SUNUSI IDRIS IDRIS
论文题目：Research on Intelligent Terrain Identification for Traction Control in Distributed Drive Off-Road Electric Vehicle
The need to reduce climate-related emissions is one of the driving force for the electrification of automobile systems. In addition to reducing emissions, electrification of vehicles can also improve traction efficiency and vehicle stability. Unlike conventional vehicles, which use an internal combustion engine (ICE) to drive one or two axles; electric vehicles (EV) can have independent motors to drive each wheels, this makes it possible to use electric motors as the traction control actuators. Because the speed and torque characteristics of electric motors in electric vehicles are good indicators of the terrain type being traverse by the vehicle, these give rise to the development of simple method of online terrains identification. The need for terrain identification is more apparent in autonomous vehicles that work in off-roadenvironments because of the absence of dedicated operator to observe the terrains. A literature review shows that a wide range of control methods are used for traction control in EVs. However, the major challenge remained the online identification of terrain types and conditions. There are different methods applied in the identification, depending on the terrain types; off-road or on-road. For off-road situations, terrain trafficability parameters estimation or classification of the terrain are the most use methods. In on-road situations, slip-based surface identification is the most common. In either case, developing methods that relies on inexpensive sensors that are already available in the vehicles is an important research area. This work is centred on terrain identification for traction control in electric vehicles. It is sectioned into terrain classification, terrain parameter estimation, and optimum friction coefficient estimation. The concise summary on each section is as follows:
The performance of vehicle systems, such as traction control and path planning, are strongly affected by the terrain being traversed by the vehicle. Various means of terrain identification were introduced to assist such systems in working efficiently. Although some methods have comparatively high accuracy, expensive sensors are involved. In line with objective of exploring the means of terrain identification that does not involve the use of expensive sensors, experiments were conducted to develop intelligent online terrain classification algorithms for distributed drive electric tractors used in agricultural environments. An instrumented robotic tractor experimental platform collected torque and slip ratio data from six different terrains; concrete, dense grass, sparse grass, firm soil, soft soil, and tarred road. Because of the presence of localized variation in the field, such as potholes and rock pebbles, which cause outliers in the raw datasets, a MATLAB function for removing and smoothening data was applied to the data. The rolling resistance values were calculated using the vertical load on the wheels, the wheel radius, and the measured torque values. Then data were hand-labelled and programmatically segmented into the six corresponding terrains based on the terrains where it was collected. For each terrain, 235 data windows were selected for uniformity purposes because the lowest dataset has 235 windows. The data was sampled at 20 Hz in the first experiment, while in the second, third, fourth, and fifth experiment, the data were sampled at 30, 40, 50, and 60 Hz, respectively. The 20 Hz preprocessed data were then transformed from a time domain to a frequency domain using fast furrier transform (FFT) and power spectral density (PSD). Two experimental analyses were conducted on the transformed feature vectors of rolling resistance, slip ratio, and the vector formed by the rolling resistance and slip ratio concatenation. In the first experiment, the transformed data were directly used to train classifiers, while in the second experiment, the feature dimension of the transformed data was reduced by principal component analysis (PCA) before the training operations. Four classifiers, SVM, KNN, NB, and ANN, were trained using the feature vectors, and the performance of the classifiers were finally evaluated. The same procedure was repeated for 30, 40, 50, and 60 Hz, respectively; in each case, the total accuracy was recorded. Because In all the cases, the total classification accuracy increase from 20-50Hz data and started to decrease after 50Hz data, the rest of the analysis in the research was based on 50Hz frequency. The algorithms could classify terrains with accuracy as high as 99%. The work's nobility lies in the algorithm's ability to not only classify terrains with high accuracy, but to also consider and compare classification accuracy based on different terrains conditions. The terrains condition considered were firmness (firm or soft soil) and grass density (sparse or dense grass cover); this is because the terrain conditions greatly influence trafficability and vehicle mobility.
In the secondsection, a method for online estimation of terrain trafficability parameters for the agricultural tractor using feedforward neural network was investigated. Data that can be measured without the need of expensive sensors were used for training the algorithm: they are driving torque, vertical wheel load, and wheel slip ratio. Traction force and wheel sinkage measurements were only used for validation purposes, they are not required in the development of the estimation algorithm. The input data for the neural network were slip ratio, wheel traction force, and the wheel vertical force, while the output were cohesion, internal friction angle, shear deformation modulus, lumped pressure sinkage. The training data for the neural network were generated by simplying Bekker classical terramechanics models and the use of soil trafficability parameters available in the literature. In the first stage of generating the training data, the wheel traction and vertical forces were calculated from the simplified form of the terramechanic models, and the used of values soil trafficability parameters from literatures, a total of 2596 sets of training input and output data was prepared. The values of the slip ratio was fixed in the range of 0.1 to 0.8 with step length 0.1. To evaluate and validate the performance of the algorithms, traction, vertical force, and slip ratio data were collected from the field using the agricultural tractor experimental platform. Using the 220 collected data as inputs to the trained feedforward neural network model, 220 values of cohesion, internal friction angle, shear deformation modulus, and lumped pressure sinkage were estimated. Also, soil shear test, and insitu pressure sinkage equipment were used to measure the soil parameters of the experimental field. The soil parameters values estimated by the neural network were compared with the values measured from field by the soil shear test and insitu pressure sinkage equipments. The field estimated values of cohesion, internal friction angle, lumped pressure sinkage, and shear deformation modulus were in the range of 5.702- 6.829, 26.003 - 29.633, 898.055 - 1016.430, and 0.025 - 0.028 respectively, while the measured parameters were in the range of 5.680-7.102, 27.288-30.023, 987.271-1040.959. Soil shear modulus was not measured because of the lack of dedicated equipment, however, it estimated values fall within the range of values for clay loamy soil, which is the soil of the experimental field. Considering the closeness between the measured and the estimated trafficability parameters, it can be confirm the trained neural network model is accurate and can be employ in estimating soil trafficability parameters.
In the thirdsection, considering that estimation and control of wheel slip is a critical consideration in preventing loss of traction, minimizing power consumptions, and reducing soil disturbance in vehicles. An approach to wheel slip estimation and control, which is robust to sensor noises and modeling imperfection has been investigated in this study. The proposed method uses a simplified form of wheels longitudinal dynamic and the measurement of wheel and vehicle speeds to estimate and control the optimum slip. The longitudinal wheel forces were estimated using a robust sliding mode observer. A straightforward and simple interpolation method, which involves the use of Burckhardt tire model, instantaneous values of wheel slip, and the estimate of longitudinal force, was used to determine the optimum slip ratio that guarantees maximum friction coefficient between the wheel and the road surface. An integral sliding mode control strategy was also developed to force the wheel slip to track the desired optimum value. The algorithm was tested in MATLAB/SIMULINK environment and later implemented on an autonomous electric vehicle test platform. Results from simulation and field tests on surfaces with different friction coefficients have proved that the algorithm can detect an abrupt change in terrain friction coefficient; it can also estimate and track the optimum slip. More so, the result has shown that the algorithm is robust to bounded variations on the weight on the wheels and rolling resistance. During simulation and field test, the system reduced the slip from non-optimal values of about 0.8 to optimal values of less than 0.2. The algorithm achieved a reduction in slip ratio by reducing the torque delivery to the wheel, which invariably leads to a reduction in wheel velocity.
In general, the algorithms, FFT, PSD, and PCA combination were able to classify terrainsfor off-road vehicleswith accuracy as high as 99%. The trafficability parameters can be estimate accurately without the knowledge of the wheel sinkage. And soil friction coefficient was estimated simply and accurately for traction control using integral sliding mode strategy.