This research project studies the feasibility of digital, programmable, software-based control of micro-electromechanical systems (MEMS). MEMS are extremely small physical structures capable of movement, lithographically combined with similarly-scaled electrical circuits to produce micron-scale mechanical devices. Examples of MEMS are tunneling accelerometers, gyroscopes, and micro-mirrors. Currently, analog controllers are used to control MEMS because of the very small time constants required to control devices at this scale. This project has the goal of developing a new style of embedded, software feedback controllers that communicate to sigma-delta over-sampled analog interfaces via high sample rate delta encoding. These designs are the result of research in metric-based controller decomposition that specifically seeks to achieve the potential of multi-rate, over-sampled signal processing. Direct benefits of this work are digital controllers suitable for MEMS integration that match the requirements and potential of these designs, delivering performance/power ratios that are not possible currently. Although the chosen implementation strategy is FPGA based, the decomposition strategy should lead to very small designs, competitive in power to current low-end microprocessors, with dozens to hundreds of times the performance achieved by current software based control using digital signal processing (DSP) platforms. The broader impact of this research project is the development of practical techniques for the design and modeling of over-sampled controllers for MEMS devices that will have direct impact in medical prosthetics, distributed sensing, and the vast range of applications that can benefit from battery-operated sensor array devices. The project seeks educational impact in the form of graduate curriculum development and of efforts to enable substantive undergraduate involvement in research.
This grant comprises a theoretical and practical exploration of low power signal processing implemented using a novel technique called 1-bit signal processing. It is well known that the cost of signal computation is decreasing with time, and becoming a necessary component of modern life. For example, advances in signal processing allow modern hearing aids to compensate for several types of nerve damage and for direction perception problems associated with one-sided hearing loss. These improvements come at a cost -- battery life for high-end hearing aids is measured in days, using the highest density cells available. In the processor creating the digital representation of the audio signals, it is very common to represent the data as a stream of 1-bit information which is later digitally processed into conventional binary numbers for conventional signal processing. This work explores the idea that initial bit stream can be processed directly, prior to such conversion, and that the total cost of the processing in energy can be far lower. An additional benefit is that the latency of the computation (how long to wait before it can be used) can also be reduced substantially. Unfortunately, although we can indeed do such processing easily, there are a number of difficulties that prevent wide-spread use. First, most of the intuition that designers and programmers have developed for conventional signal processing is faulty. In 1-bit, it is common for the signal level to be 100-200 times weaker than the representation noise. Most topologies for signal filtering only amplify this noise and provide no useful function. Secondly, while sigma-delta converters do indeed provide very high performance, they are also inherently unstable for un-guarded inputs. Verifying that the design is correct and is stable for all legal inputs is very difficult. On the practical side, the technology is applied to a novel MEMs based device shown in the figure. The design is a mechanical sensor with high potential sensitivity based on using quantum tunneling to position a test mass in a fixed position an measuring the current required to hold it there. Constructing a controller to perform this activity is difficult -- analog circuit controllers have inherent noise and drift issues while digital ones have high cost and complexity. This device was effectively controlled by a simple 1-bit controller in a process similar to that used by advanced MEMs manufacturers such as Analog Devices or Bosch. These techniques are rarely used in other venues due to the problems outlined above. One outcome of this work is the development of a digital circuit based verification and modeling technique that allows definitive checking of the accuracy and correctness of the design for all possible inputs, checkable in reasonable amounts of time. The second outcome is the development of efficient techniques to organize and design mixed hardware and software embedded systems such as the hearing aid mentioned above. While the first outcome enables a designer some assurance that the 1-bit design is stable, the second allows for efficient implementation of the design. While the first outcome has impact in the implementation and verification of small signal-processing systems, the second outcome has potential impact in areas as diverse as multi-media delivery (such as movies on a cell-phone) and the design cost of future generations of such devices. In the figure, the green circuit board is part of a web-appliance controller which uses the popular MATLAB simulink format for input, but creates a 1-bit controller in the FPGA (large black chip on the red board). This design was created by undergraduate EE students in my lab as a low cost teaching controller for controls classes as well as a high performance prototyping or experimental control device for research applications. The system is designed to be inexpensive (less that $400) but provide roughly 100x the performance of the CPU based controllers currently in use. When tested, these designs will be disseminated on the net at no charge.
