This thesis investigates the application of machine learning techniques to predict rotor temperature in induction motors, a crucial parameter for enhancing the safety, efficiency, and longevity of electric machines. It focuses on the development, training, and evaluation of three neural network models: Multilayer Perceptron (MLP), Thermal Neural Network (TNN), and 1-D Convolutional Neural Network (1-D CNN). These models are assessed for their ability to predict rotor temperature using measurable inputs such as rotor speed, slip speed, current, voltage, and cooling conditions, both with and without stator temperature measurements.

Experimental results show that the TNN model, which incorporates heat-transfer principles, outperforms other models in terms of accuracy and reliability. The analysis is based on a comprehensive dataset comprising 2,876,126 data points collected over approximately 82 hours, covering various operational scenarios represented by 87 profile IDs. The models were trained on rotor temperatures ranging from 21.6°C to 259.6°C, with 22°C representing idle room temperature and 250°C the rotor's critical threshold.

The findings suggest that machine learning models offer a more adaptable and precise approach to temperature estimation in electric motors, with significant potential to improve operational efficiency and safety protocols in the automotive and industrial sectors. The TNN model was particularly effective, achieving mean squared errors of 4.2 and 23.3 in two different normal driving scenarios. Additionally, including motor housing temperature data, both inside and outside, significantly enhanced rotor temperature prediction accuracy, potentially reducing costs by eliminating the need for stator temperature sensors.

Data collection and preprocessing were done in collaboration with Volvo Cars. Including motor housing temperature significantly improved rotor temperature estimation accuracy, suggesting cost savings by eliminating the need for stator temperature sensors. This research highlights the superiority of ML models over deterministic models for predicting EM temperature, enhancing safety, efficiency, and reliability. The TNN model’s versatility in various driving conditions sets the groundwork for future ML integration into real-time Motor Control Systems.

This research demonstrates the superiority of machine learning models over deterministic models for induction motor temperature prediction, improving safety, efficiency, and reliability. The TNN model's ability to accurately predict rotor temperature under various conditions highlights the potential for integrating machine learning into real-time motor control systems, setting the stage for future advancements in this field.

Today a lot of companies that produces processing machines is competing to make their product more efficient than their competitors in order to improve sales. This paper is focused on processing machines that measure weight and uses mechanical force to rotate goods to increase surface contact with a processing agent. This study will investigate combinations of algorithms that is most suitable for extracting a gradient from a noisy load signal containing a periodic disturbance. A better extraction of the gradient, which is the flowrate of the processing agent, would allow the processing machine to make better predictions for its process control. 

This study applies quantitative research methods through simulations and one experiment in order to get an understanding of each algorithm’s behavior. In each test they are tested according to certain criteria, these are phase-lag, similarity, accuracy and precision of both load and gradient. 

The results of this study suggest three different algorithms that are likely to perform well in the mentioned criteria. Kalman CV is the best choice for gradient extraction on one second intervals when tracking a periodic signal is not necessary. Kalman CV-Kalman PLI is the best choice for both load and gradient extraction when tracking a periodic signal is needed for process control. It is also the best choice for extracting a gradient on sample-to-sample intervals overall. Finally, Chebyshev is the best choice for load extraction if tracking a periodic signal is not desired. 

Since this paper could only determine which algorithm is more likely to perform better, it can only partially answer its research questions. Despite only a partial answer, this paper concludes that its findings could still be used by companies to find suitable algorithms for their processing machines or be used by those new to the field to get an introduction to the presented algorithms.

Abstract

This report is about the creation of a test system for development of a control system for a off the shelf RC-helicopter, which can be bought for about two thousand SEK. The test system consists of both electronics for controlling the helicopter and mechanical constrainer. The constrainer is used for mounting the helicopter so that it won´t crash.A circuit board including processor and sensors was manufactured. This was used for doing test of the different control algorithms. In this way were we able to analyze the quality of the sensors chosen, and what other sensors that could contribute to the control of the system.Basic helicopter control have been studied and presented in the report. Also different kinds of regulation algorithm were presented. These have also been implemented and tested in the test system.Experiments were made to identify how signals in the radio controller were mixed and translated before they were sent to the receiver, to be used to control the helicopter. This could later be used to translate the control signal from the regulators.Demonstrations, with the help of the test system, revealed the pros and cons of the different sensors and other hardware. Different regulation techniques were also tested and analyzed.

In this contribution we present a new computational method for coupled bulk-surface problems on time-dependent domains. The method is based on a space–time formulation using discontinuous piecewise linear elements in time and continuous piecewise linear elements in space on a fixed background mesh. The domain is represented using a piecewise linear level set function on the background mesh and a cut finite element method is used to discretize the bulk and surface problems. In the cut finite element method the bilinear forms associated with the weak formulation of the problem are directly evaluated on the bulk domain and the surface defined by the level set, essentially using the restrictions of the piecewise linear functions to the computational domain. In addition a stabilization term is added to stabilize convection as well as the resulting algebraic system that is solved in each time step. We show in numerical examples that the resulting method is accurate and stable and results in well conditioned algebraic systems independent of the position of the interface relative to the background mesh.

The majority of all power cuts that affects individuals and communities are caused by earth faults on the power transmission lines at 10kV and 20kV level [1]. If these power cuts could be eliminated, less disturbances and interrupts would lead to large amounts of money savings.

Swedish Neutral has developed such a protection system. When an earth fault occurs power is injected into the neutral point of the transformer. The RCC (Residual Current Compensation) protection calculates a compensation current exactly 180 degrees out of phase to the fault current. Doing this, the voltage at the fault location becomes very close to zero, without affecting the power transmission.

The protection system can only compensate automatically for the fundamental frequency (50Hz), and manually for the 3^rd, 5^th, 7^th and 9^th harmonics. In most cases, when the harmonics are very small it is not necessary to compensate for them. There are though cases when compensation for the harmonics is necessary.

This thesis focuses on finding the best method to extract the content of a sampled signal regarding both simplicity and speed. Both amplitude and phase of each harmonic must be calculated. Is the proposed method suitable for the current computer system and how can it easily be implemented.

Because the fundamental frequency is known and the harmonics are all multiples of the fundamental frequency it makes the task less complex. It is not necessary to use the FFT algorithm. The DFT can be calculated using correlation. Both phase and amplitude can be calculated very precisely with few samples and not so many computer operations.

This study aims at investigating a central mounted Indoor Positioning System based onBluetooth RSSI values and Kalman filtering where the Bluetooth module is mounted on arotating arm around an axle. The purpose is to investigate how the length of the arm and therotational speed influence on the possible positioning accuracy of the artefact. As requested bythe company Combitech AB, the artefact will also be evaluated whether the design is areasonable alternative to de most common Bluetooth based Indoor Positioning Systems. Thecase study is performed using design science as an artefact is developed and evaluated. Theartefact is tested in a predetermined test environment, in this case an office room. In the testenvironment, measurements will be realized on the different iterations of the artefact. Theresult of the study is the achieved positioning accuracy for the artefact in two and threedimensions. Linear equations describing the relationship between the length of the arm andthe achieved accuracy of the artefact will also be proposed. The best accuracy achieved in twodimensions was 510 centimeters. The best achieved accuracy in three dimensions was 730 centimeters. The observed relationship between the length of the arm and the uncertainty ofthe trilaterated position is that when the length of the arm increases the uncertainty decreasesin both two and three dimensions.

In this report, some 40 papers related to backlash control have been studied, and their results have been arranged into a systematic structure, where the main categories are linear, passive and active nonlinear controllers. The emphasis of the study is on control strategies suitable for automotive powertrain applications, but the reported results are not limited to this field. The reported results are based on partly different system structures, depending on how the backlash is incorporated in the system. Different backlash models are also used. Therefore, the report contains a section that describes these differences as well.

Backlash is the source of an uncomfortable acceleration phenomenon, which in the automotive industry is called “shunt and shuffle”. Until recently, only simple controllers have been used to reduce the discomfort, and the tuning of these is made quite subjectively. A number of approaches to control of systems with backlash are reported in the literature, but very few are directed towards control of automotive powertrains. This paper evaluates four different controllers for powertrains with backlash, by means of simulation. The results show that active nonlinear controllers have a potential for improved backlash control. However, the robustness of these controllers needs further investigation.