Source-linked AI summary
FPGA Based Design And Implementation of Neural Network Compensators for Nanocrystalline MEMS Pressure Sensor
Somsubhra Talapatra, Sudip Ghosh, C Roy Chaudhuri, Biplab K Sikdar, Hafizur Rahaman
TL;DR
Nanocrystalline MEMS pressure sensors have nonlinearity and temperature effects that require compensation. The paper trains a feedforward neural network and implements it on an FPGA using a pipelined architecture; simulations report 98% error reduction, with a 13.477 ns worst-case delay.
Problem
Nanocrystalline silicon MEMS pressure sensors exhibit nonlinearity and temperature effects that need compensation, despite their use in low-pressure biomedical applications.
Source: §1 · p. 1, §2 · pp. 1–2Method
The paper trains an MLP feedforward network using back propagation and implements the trained network on an FPGA with a pipelined architecture.
Source: §1 · p. 1, §3 · p. 2, §4.2 · p. 4Results
98% error reduction was observed in MATLAB simulation, and the FPGA implementation had a worst-case delay of 13.477 ns.
Source: §7 · p. 6Takeaways & Limitations
The reported compensator operates over the stated 0–1 bar pressure and 25–80°C temperature ranges, with the pipelined FPGA design intended for parallel sensing operation.
Source: §1 · p. 1, §7 · p. 6The design assumes that the interval between consecutive interrupts exceeds 42 pipeline cycles.
Source: §4.2 · p. 5
Abstract
from arXiv · showhide
In this paper the effects of nonlinearity and temperature of nanocrystalline MEMS pressure sensor have been compensated using neural network algorithm which has been implemented on a FPGA chip. The neural network employs back propagation algorithm to train the multilayer perceptron feed forward network that causes an error reduction due to temperature upto the level of 98%. It is realized on FPGA with pipelined architecture for parallel and fast sensing operation. The pressure sensor along with the FPGA based compensation results in an intelligent sensor system capable of high speed and accurate pressure measurement which has been tested for a typical biomedical application.
1. Introduction
Nanocrystalline silicon MEMS pressure sensors offer improved sensitivity but exhibit nonlinearity and temperature effects that require compensation. The paper uses a back-propagation-trained feedforward neural network and an FPGA pipelined implementation to address these effects.
- Nanocrystalline silicon pressure sensors have three times the sensitivity of conventional silicon sensors for a particular nanowire dimension.Their increased piezoresistive coefficient is cited as a reason for interest.Source: §1 · p. 1
- Nonlinearity and considerable temperature effects in these sensors need compensation for efficient utilization.Source: §1 · p. 1
- A back-propagation-trained multilayer perceptron feedforward network is implemented on a pipelined FPGA architecture for parallel operation and improved speed.The network is trained offline using MATLAB ANN toolbox; its datapath is optimized for area.Source: §1 · p. 1
2. Sensor Design, Fabrication and Results
The section describes a nanocrystalline MEMS pressure sensor for low-pressure biomedical applications and identifies temperature dependence as a compensation concern. It gives typical respirator and blood-pressure sensor specifications.
- The sensor exhibits significant temperature dependence, which the paper identifies as needing compensation.Source: §2 · p. 2, Figure 2 · p. 2
- Typical respirator or blood-pressure sensor specifications include a 0–1 bar pressure range, 30 mV V–1 bar–1 desired sensitivity, and nonlinearity below 0.5% of full-scale output.The listed resolution is 1 mbar and the operative temperature range is 25–80°C.Source: §2 · pp. 1–2
- Figure 1 presents a cross-sectional view of the sensor.Source: Figure 1 · p. 2
- Figure 2 presents output voltage variation with temperature for different pressure values.Source: Figure 2 · p. 2
3. Compensation using ANN
The compensation scheme uses an inverse neural-network model trained on experimental sensor input-output data. Simulation studies normalize the sensor variables and use seven pressure-specific data sets for training and testing.
3. Compensation using ANN
- The inverse ANN compensation model takes sensor output and ambient temperature as inputs and uses an MLP feedforward network with one output.The hidden layer uses a tansigmoidal activation function and the output layer is linear; hidden-neuron count is selected heuristically to minimize error.Source: §3 · p. 2
- Back propagation trains the MLP, with hidden-layer neuron count and weights optimized for minimum error.Training starts with the minimum number of hidden neurons for a fixed epoch; back propagation is described as a generalization of LMS.Source: §3 · p. 2
- Figure 3 depicts the proposed compensation model, and Figure 4 compares pressure variation with temperature for compensated and uncompensated sensors.Source: Figure 3 · p. 2, Figure 4 · p. 2
3.1 Simulation Studies
- Sensor pressure, ambient temperature, and detection-circuit voltage are normalized to the range {-1, 1}, and seven pressure-specific data sets are generated for training and testing.The data sets are derived from the graph in Figure 2; simulation uses the MATLAB ANN toolbox environment.Source: §3.1 · p. 2
4. FPGA Based Implementation
The ANN’s two-layer organization is decomposed into independent computations to motivate a pipelined FPGA implementation. The design reports agreement with MATLAB simulation and improved CLB utilization compared with simple behavioral modeling.
4.1 Motivation for Pipelined Architecture Design for the above ANN
- The first layer has 12 tansigmoidal neurons receiving weighted pressure, temperature, and bias inputs; a second, linear neuron combines their outputs.A dedicated bias input also contributes to the second-layer sum.Source: §4.1 · p. 3
- Twelve independent computational segments can be calculated in parallel, motivating a pipeline intended to run faster than sequential computation.The timing diagram for the proposed pipeline is identified as Figure 6.Source: §4.1 · p. 4
4.2 Design and implementation of the pipeline architecture
- The pipeline datapath uses three multiplications and three additions and is area-optimized for implementation where silicon area matters.Source: §4.2 · p. 4
- One 16-bit signed multiplier and one 16-bit adder serve successive pipeline stages, which are issued three cycles apart to avoid resource conflicts.The datapath also uses eleven 16-bit registers; weights and biases are stored in four ROMs, and the tansigmoid function in a LUT ROM.Source: §4.2 · p. 4
- The controller processes 12 consecutive segments and completes an operation in 42 pipeline cycles, assuming successive interrupts are more than 42 cycles apart.Source: §4.2 · pp. 4–5
- The FPGA-implemented trained network matches the MATLAB simulation, and the pipelined design enables optimum Spartan 3 CLB utilization unavailable with simple behavioral modeling.The synthesis summary and timing results are identified in Tables 2 and 3, respectively.Source: §4.2 · p. 5, Table 2 · p. 5, Table 3 · p. 6
7. Conclusion
The compensator targets a nanocrystalline porous-silicon MEMS pressure sensor over 0–1 bar and 25–80°C. The authors report 98% error reduction in MATLAB studies and an area-optimized pipelined FPGA implementation with a 13.477 ns worst-case delay.
- 98% error reduction is reported for the ANN compensator in MATLAB simulation studies.Source: §7 · p. 6
- The 12-hidden-neuron network was trained offline and implemented on FPGA with an area-optimized pipeline and 13.477 ns worst-case delay.Source: §7 · p. 6
- The reported compensation covers a 0–1 bar pressure range and a 25–80°C temperature range for a nanocrystalline porous-silicon MEMS pressure sensor.The pipelined design is also reported to enable optimum Spartan 3 CLB utilization compared with simple behavioral modeling.Source: §7 · p. 6