Matching Items (3)
Filtering by
- Creators: Pollat, Scott
- Creators: Cao, Yu
Description
Flow measurement has always been one of the most critical processes in many industrial and clinical applications. The dynamic behavior of flow helps to define the state of a process. An industrial example would be that in an aircraft, where the rate of airflow passing the aircraft is used to determine the speed of the plane. A clinical example would be that the flow of a patient's breath which could help determine the state of the patient's lungs. This project is focused on the flow-meter that are used for airflow measurement in human lungs. In order to do these measurements, resistive-type flow-meters are commonly used in respiratory measurement systems. This method consists of passing the respiratory flow through a fluid resistive component, while measuring the resulting pressure drop, which is linearly related to volumetric flow rate. These types of flow-meters typically have a low frequency response but are adequate for most applications, including spirometry and respiration monitoring. In the case of lung parameter estimation methods, such as the Quick Obstruction Method, it becomes important to have a higher frequency response in the flow-meter so that the high frequency components in the flow are measurable. The following three types of flow-meters were: a. Capillary type b. Screen Pneumotach type c. Square Edge orifice type To measure the frequency response, a sinusoidal flow is generated with a small speaker and passed through the flow-meter that is connected to a large, rigid container. True flow is proportional to the derivative of the pressure inside the container. True flow is then compared with the measured flow, which is proportional to the pressure drop across the flow-meter. In order to do the characterization, two LabVIEW data acquisition programs have been developed, one for transducer calibration, and another one that records flow and pressure data for frequency response testing of the flow-meter. In addition, a model that explains the behavior exhibited by the flow-meter has been proposed and simulated. This model contains a fluid resistor and inductor in series. The final step in this project was to approximate the frequency response data to the developed model expressed as a transfer function.
ContributorsHu, Jianchen (Author) / Macia, Narciso (Thesis advisor) / Pollat, Scott (Committee member) / Rogers, Bradley (Committee member) / Arizona State University (Publisher)
Created2013
Description
Speech recognition and keyword detection are becoming increasingly popular applications for mobile systems. While deep neural network (DNN) implementation of these systems have very good performance,
they have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this thesis, techniques to reduce the memory and computation cost
of keyword detection and speech recognition networks (or DNNs) are presented.
The first technique is based on representing all weights and biases by a small number of bits and mapping all nodal computations into fixed-point ones with minimal degradation in the
accuracy. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Additional reduction in memory is achieved by a technique called weight pruning,
where the weights are classified as sensitive and insensitive and the sensitive weights are represented with higher precision. A combination of these two techniques helps reduce the memory
footprint by 81 - 84% for speech recognition and keyword detection networks respectively.
Further reduction in memory size is achieved by judiciously dropping connections for large blocks of weights. The corresponding technique, termed coarse-grain sparsification, introduces
hardware-aware sparsity during DNN training, which leads to efficient weight memory compression and significant reduction in the number of computations during classification without
loss of accuracy. Keyword detection and speech recognition DNNs trained with 75% of the weights dropped and classified with 5-6 bit weight precision effectively reduced the weight memory
requirement by ~95% compared to a fully-connected network with double precision, while showing similar performance in keyword detection accuracy and word error rate.
they have large memory and compute resource requirements, making their implementation on a mobile device quite challenging. In this thesis, techniques to reduce the memory and computation cost
of keyword detection and speech recognition networks (or DNNs) are presented.
The first technique is based on representing all weights and biases by a small number of bits and mapping all nodal computations into fixed-point ones with minimal degradation in the
accuracy. Experiments conducted on the Resource Management (RM) database show that for the keyword detection neural network, representing the weights by 5 bits results in a 6 fold reduction in memory compared to a floating point implementation with very little loss in performance. Similarly, for the speech recognition neural network, representing the weights by 6 bits results in a 5 fold reduction in memory while maintaining an error rate similar to a floating point implementation. Additional reduction in memory is achieved by a technique called weight pruning,
where the weights are classified as sensitive and insensitive and the sensitive weights are represented with higher precision. A combination of these two techniques helps reduce the memory
footprint by 81 - 84% for speech recognition and keyword detection networks respectively.
Further reduction in memory size is achieved by judiciously dropping connections for large blocks of weights. The corresponding technique, termed coarse-grain sparsification, introduces
hardware-aware sparsity during DNN training, which leads to efficient weight memory compression and significant reduction in the number of computations during classification without
loss of accuracy. Keyword detection and speech recognition DNNs trained with 75% of the weights dropped and classified with 5-6 bit weight precision effectively reduced the weight memory
requirement by ~95% compared to a fully-connected network with double precision, while showing similar performance in keyword detection accuracy and word error rate.
ContributorsArunachalam, Sairam (Author) / Chakrabarti, Chaitali (Thesis advisor) / Seo, Jae-Sun (Thesis advisor) / Cao, Yu (Committee member) / Arizona State University (Publisher)
Created2016
Description
Many tasks that humans do from day to day are taken for granted in term of appreciating their true complexity. Humans are the only species on the planet that have developed such an in-depth means of auditory communication. Recreating the mechanisms in the brain that recognize speech patterns is no easy task. This paper compares and contrasts various algorithms used in modern day ASR systems, and focuses primarily on ASR systems in resource constrained environments. The Green colored blocks in Figure 1 will be focused on in greater detail throughout this paper, they are the key to building an exceptional ASR system. Deep Neural Networks (DNNs) are the clear and current leader among ASR technologies; all research in this field is currently revolving around this method. Although DNNs are very effective, many older methods of ASR are used often due to the complexities involved with DNNs; these difficulties include the large amount of hardware resources as well as development resources, such as engineers and money, required for this method.
ContributorsPetersen, Casey Alexander (Author) / Csavina, Kristine (Thesis director) / Pollat, Scott (Committee member) / Engineering Programs (Contributor) / Barrett, The Honors College (Contributor)
Created2015-12