The objective of this project is to design and develop a cost-effective commercial fire detector system that monitors multiple characteristics of fires/flames in residential environments for a zero error early fire detection and identification system. The detector integrates existing current state of the art smoke detectors with an intelligent wireless solar-blind dual-band photodetector system for advanced early fire detection/recognition, controlled by FPGA type portable circuitry, with neural network-based identification capability.

The three main characteristics of a fire are an increase in temperature, generation of smoke particles and optical emission. Fires initiated by exposure of combustible material to temperatures around or slightly above the flashpoint of the material, typically tend to generate smoke before they ignite into flame. However, in some situations where the ignition temperature is greater than the flashpoint of the object, complete combustion occurs with very little smoke. Other situations, such for example one where a draft exists prior to the fire, may also hinder the capability of the smoke alarm system, , removing most if not all smoke particles, causing tragic delay in fire alarm. In such cases early fire detection is practically not possible with current state of the art ionization smoke alarms or optical smoke alarms. In such scenarios integration of existing smoke detectors with optical emission sensing and identification technology is a paramount for a false free and nuisance-alarm free fire detector.

Fires produce emissions ranging from ultraviolet to IR. Such emissions can only be detected over the wide-range of ambient light background, by fast multi-range optical detectors allowing time- and spectrally-resolved measurements in particular optical regions. As a result, not only the spectral range, but also the detector speed, spatial resolution and alignment become critical for fast fire detection as well as for avoiding costly false alarms. Currently used photo-multiplier tubes (PMTs) have high sensitivity. However, they are bulky, require high voltage operation, have low mechanical and temperature strength, and cannot be easily integrated into current fire detectors. The recently developed dual-band detectors that are composed of discrete UV and IR solid-state components are bulky, not capable of detecting the multi-band optical signal with high spatial resolution, and are not suitable for networking.

Employment of a miniature, chip-based dual-color high-temperature visible- or even solar-blind optical sensor system would allow for fast and false alarm-free fire detection and recognition, thus providing a fast and reliable response in separated UV and IR bands with high spatial resolution, and “smart”, artificial neural networks (ANN)-based signal analysis. Moreover, development of such sensors promotes fabrication of multi-pixel dual-band UV/IR focal plane arrays with a visible- or solar-blind imaging capability.

There are two primary approaches of integrating the optical sensor system with existing state of the art smoke detectors. The first method is to remotely locate high sensitivity dual-band UV/IR focal plane arrays and smoke detectors in areas that are prone to possible fires, such as kitchens and bedrooms. These devices will then communicate with one central control system that analyses the nature and type of flame and sound an alarm accordingly. The second method is to integrate the smoke and the high sensitivity dual-band UV/IR focal plane array detector into a unit controlled by one system, and then place them in a close proximity of possible fire sources.

Group III-nitride materials are superior for advanced UV detector fabrication, due to their wide direct band gap along with high thermal, chemical, mechanical, and radiation tolerance. Research and development performed by several groups, indicate that effective optical emission and detection can be achieved in a wide spectral band ranging from 200 to 1770nm, which includes also the near IR range.  The Radio Frequency Molecular Beam Epitaxy (RF MBE) method used in our laboratory for nitride material growth allows fabrication of multilayer structures that incorporate binary, ternary, or even quaternary nitride compounds, with a precise control over the layer, thickness, chemical composition, crystalline quality, and doping during a single-process growth on commercial  sapphire or silicon substrates. Our preliminary data from GaN, AlGaN, and InGaN based photodiode structures grown on Si and sapphire, indicate that sensitivity in both the UV and IR ranges can be achieved from a single structure. Measurements performed in our laboratory on GaN/InGaN-based heterostructure chips, show that they can be operated at temperatures over 300°C without internal or external cooling. The challenge is to take advantage of all the advanced nitride material growth and processing capabilities, as well as the unique optical, chemical, and thermal properties of the nitrides, in order to develop wireless, miniature, inexpensive, and reliable integrated multi-band solar-blind fire detectors.  We own two US patents  (US 7,381,966 and US 7,566,875) on this technology.

Other preliminary results are from the development of chip integrated optoelectronic multi-band chemical sensors. In this project, an integrated device structure based on wavelength-selective LED and photodetector chips (Figure 4) is controlled by a FDMA-based circuitry with an Artificial Neural Network (ANN)-based signal acquisition and analysis. Variable signal patterns are generated by combined effects of fluorescence, absorption, and scattering resulting from interaction of the multi-wavelength optical emission with the analytes.  ANN was employed for the categorization of different analytes of various concentrations using Stuttgart Neural Network Simulator (SNNS) tool (Figure 4a). For 8 different analytes at 4 or 5 different concentrations, totaling 35 different samples, after 2000 cycles of training the network the results were:  96% accuracy for the testing set and 100% accuracy for the training set (Figure 4b). The current efforts on this project are directed towards the development of an intelligent portable multifunctional bio-chemical sensor system with time-resolved capability in a ps time resolution range.

The schematic of a single-pixel dual-band UV/IR photodetector is shown in Figure 5. Double side-polished n-type <111> Si wafers are used as substrates. AlN and GaN buffer layers are grown by RF MBE on Si to compensate for the lattice mismatch and reduce the effect of the substrate material on the active AlGaN film grown on the top of the structure. The content of Al in this film determines the UV cut-off wavelengths of the device and can be varied between 30 and 60%. Spectroscopic ellipsometry is used to determine the band gap of the AlGaN layer. Reflection High Energy Electron Diffraction (RHEED) is used to monitor the crystalline quality of the layers during growth. Post-growth characterization includes photoluminescence (PL), optical transmission, spectroscopic ellipsometry, and Hall effect measurements. In order to form the IR sensitive part of the photodetector, a Pt /Au layer is deposited by e-beam evaporation on the backside of Si and patterned to form Pt Schottky barrier contacts to n-type Si. Then Ti / Au dots are deposited through a stencil mask in between the Pt/Au dots in order to form the ohmic contacts to n-type Si. The UV sensitive part of the photodetector is processed as follows. A silicon dioxide (SiO2) layer is first deposited on top of the AlGaN layer by a PECVD method in order to provide insulation for the Schottky barrier contacts. Rows of round openings aligned with the metal dots on the bottom of the Si wafer are formed then in the SiO2 layer using photolithography. Conductive transparent tin oxide (SnO2) layer is deposited then by spray pyrolysis on top of the patterned SiO2 layer. This layer is also patterned by photolithography in order to produce round contacts aligned with the Pt/Au dots on the bottom of the wafer, and with the corresponding windows in the SiO2 layer in every second row. Finally Ti /Au contacts are deposited by e-beam evaporation on the top of the wafer. Figure 6a shows a diced 4 pixel array before mounting on the AlN chip (view insert in Fig. 2), indicating the UV pixels, ohmic contacts and bulk contacts.  Figure 6b shows the array of UV diodes and ohmic contacts fabricated on a silicon wafer.

A standard TO-8 housing with a 5 mm opening in the cap is used for packaging. The Pt contacts on the backside of the silicon chip are bonded using thermo-compressive bonding to the Au pads deposited and patterned on top of a thermally-conductive electrically insulating AlN ceramic carrier plate. The Au pads on the ceramic plate are then micro-bonded with a 30 µ thick Au wire with two of the TO-8 housing legs, while the Ti/Au contacts on top of the chip are micro-bonded with other two legs of the housing.

We should optimize Si substrate properties, III nitride growth parameters, and device processing in order to achieve higher efficiency of the dual-band UV/IR photodetectors. The optimization of the substrate parameters is done first by using theoretical simulations that will be directed towards reduction of the leakage between the UV and the IR structures currently resulting in some background sensitivity to the visible (or solar) light. For this purpose Si substrates with guarding p-n-junctions can be applied. The boron implantation parameters are modeled using SILVACOTM device simulation software. The thickness, doping type and concentration of the Si substrate and the nitride layers will be optimized.  The Poisson and Schrödinger equations representing the device structures should be simultaneously solved to determine the electric field distribution and calculate the responsivity of each diode.

The nitride growth parameters are experimentally optimized mainly for two purposes: a) to accomplish more effective and reproducible doping of the nitride layers, b) reduce the defect density that greatly affects the device efficiency. The device processing optimization  focuses on providing higher mechanical and thermal stability as well as more efficient utilization of the advanced electrical, optical, and semiconductor properties of all materials used in the diode fabrication process. In particular, loss resulting from the reflection of the light passing multiple interfaces can be reduced by employment of anti-reflection coatings. The metal combinations used for contact fabrication should be selected to sustain elevated temperatures and harsh environments.



We also consider various approaches for fabrication of micro-miniature surface-mount dual band UV/IR photodetectors that can be simply integrated into large networks and mounted into hard-to-reach areas. In addition, we investigate the possibility for integration of such detectors with micro-miniature Si-based wireless transmitter chips.


Two main issues targeted in this project are early fire detection and elimination of nuisance alarms. The earliest alarm time reported for ionization type smoke detectors is 37s, which in turn affects the egress time. However, with early fire detection technology using photodetectors as discussed in this proposal, alarm time can be as low as 1s. This allows for possible extinction of fire before tragic loss of lives or property damage can be done. Secondly by monitoring UV and IR emissions from the source of alarm nuisance alarms can be eliminated. The only major drawback of a photodetector based alarm system is that the source of alarm has to be in the line of sight of the detector. Thus, emphasizing the need for an integrated photodetection and smoke detection system.

For effective fire detection the signals generated by the photodetector in response to the flame, need to be captured continuously or with a very high sampling rate. Our approach is based on two different measurements: steady state (SS) and time-resolved (TR) measurements. The system’s output (fire-alarm) is based on data from both measurements that can be classified by ANN in real-time.

In steady-state photocurrent measurements, the common method of capturing and digitizing a DC signal is to use a switched integrator in combination with an analog to digital converter (ADC). The principle is based on collecting the signal to an integration capacitance for an integration period selected by the user (Figure 7) followed by digitizing. Current ADCs are very accurate in digitizing low-level currents. In our photo-detection set up, we envision currents in the range from picoamperes for very low-emission levels to a few hundred nanoamperes for photodetectors placed in close proximity to fires and bright sparks. To measure such a wide dynamic range of current-inputs, an ADC with a high dynamic range will be necessary. In this part, we will design, fabricate and test a suitable ADC, which can perform the necessary capturing of low intensity level flames and sparks. This can be performed separately from the TR measurement, which will be combined with the SS measurement at a later stage. The measurements also require necessary ADC control signals from a portable setup. There are many different ways of controlling signals and acquiring data by using a portable design based on employment of microcontrollers, FPGAs, PLDs and other programmable devices. Utilization of time-resolved measurements for flame identification requires employment of FPGA based control capable of providing fast response-time and stable operation. A suitable FPGA should be selected depending on the requirements for time resolution in the TR measurements, and to have a sufficient capacity for  performing ANN and other related tasks.

Based on our preliminary flame dynamics studies, a TR system with a resolution in the microsecond range is required for recognizing flame dynamics patterns used for early fire detection. Employment of fast Schottky barrier structures provide an additional benefit for the TR measurements. Amplification of the photodetector signals (in pA or nA) plays an important role in these measurements.

We have already demonstrated a method of amplifying such signals using a bootstrapped-cascoded technique[1]. This technique was successfully employed for amplification of current pulses with 5-10ns fall/rise time. Reduction of noise background is also a key criterion, which is investigated in detail during this task. Optimization of photodiode amplifier circuit parameters should be performed for the amplification of low-level current from low level fires with the elimination of background noise at sampling high-speeds. Another key is to combine the TR and SS measurement controls in a single FPGA chip, overcoming noise from different sources.  A sample set-up proposed for combining the steady-state and time-resolved measurements is shown in Figure 8[2]. Here an EEPROM is required to start the system in a portable mode. The SS and TR measurements are made separately and stored in FPGA, which performs the ANN algorithm for fire detection.

[1] C. Joseph, M. Boukadoum, J. Charlson, D. Starikov, and A. Bensaoula, “High-speed front end for LED-Photodiode based fluorescence lifetime measurement system”, Proc. IEEE International Symposium on Circuits and Systems, May 2007 (ISCAS 2007), pp. 3578-358. http://ieeexplore.ieee.org/document/4253454/

[2] Data sheet for DDC101 from Texas Instruments. http://www.ti.com/lit/ds/sbas029/sbas029.pdf

We use several approaches for early error-free fire detection and identification. One of the approaches is employment of solar/visible blind photodetectors insensitive to common background light sources. The second approach is miniaturization of these photodetectors in order to integrate them into large networks parts of which will be placed in close proximity from potential fire sources and at places that cannot be monitored by conventional fire protection equipment The third approach is based on independent simultaneous detection of the optical flame emission in two separate UV and IR spectral bands. This would remove most false alarms related to optical emissions from lightning, welding, and electrical arcs, as well as from various heat sources and lighting. Finally these devices will be integrated into existing commercial smoke detectors to ensure an early false free residential alarm system. Figure 9 shows an example a design approach would that can be used for such integration.

We strongly believe that in addition to the above approaches, the dual band photodetector capability combined with intelligent analysis can be used: a) to distinguish a flame from other light/heat sources; b) to identify the fire source. Such features can be enabled after investigation of various typical optical emission scenarios taking place during an interval that includes time preceding the ignition and the time right after the ignition. For example, if an electrical spark is the cause of the ignition, a short, high-intensity, broad-band light pulse followed by the flame emission light will be detected by the sensor. If the fire is caused by excessive heat, a strong, slow, monotonically increasing IR signal followed by a mixed UV/IR signal- after flame ignition, will be detected. In addition, flames from sources of different chemical composition will have distinct optical signatures (time and energy). Depending on which gas radicals are excited in the burning process, the intensities of the UV and IR peaks will change during the initial time period- after ignition, before normal flickering takes effect.


The proposed time resolution acquisition system should easily discriminate between signal patterns, characteristic to various potential fire sources, such as fuels, plastics, wood, electrical isolation, etc Understanding and classifying these signal patterns is an important goal for this project. Flame flickering at a specific frequency (typically from few to hundreds of Hz) will be used, in addition to the initial signal pattern recognition, to confirm presence of a flame. To the best of our knowledge, none or very little research has been performed on optical flame dynamics from various ignition sources. All existing relevant data attributed to combustion processes, that are quite different from flames, can be used only for identification of various gas radicals that are excited during combustion (Table 1[1],[2]).

In our preliminary studies, a repeatable delay between the IR and the UV signal components from a matchstick flame was measured first with a high-speed memory oscilloscope (Figure 10). In order to further investigate different flame dynamics we used an analog to digital data acquisition card (DAC USB 1208LS). The instrument was interfaced with LABVIEW™ and data was first collected and stored on the card FIFO buffer before being read. This allowed for the card to rapidly monitor the flame using commercial UV and IR photodiodes with peak wavelengths of 375, and 850 nm,

respectively. In order to detect very low light intensities we designed and implemented a photodiode amplifier circuit using low noise high gain OPA177GP Op Amps. Table 2 indicates that delays between the IR and UV signal components measured with the setup from flames of aviation JP-8 fuel and gasoline are positive and negative, respectively, and also have differing average values.

[1] Measurement and analysis of OH emission spectra following laser-induced optical breakdown in air Parigger, Christian G. (Center for Laser Applications, University of Tennessee, Space Institute); Guan, Guoming; Hornkohl, James O. Source: Applied Optics, v 42, n 30, Oct 20, 2003, p 5986-5991.

[2] Photographic Observation and Emission Spectral Analysis of HCCI Combustion, Combustion Science and Technology. v 177, pg1699-1723, 2005

Further work should be undertaken to characterize flames from other materials relevant to residential environments e. g. ignition of fuel, oil, plastic, etc. For this purpose a higher speed data acquisition card with a resolution in a nanosecond range, currently used in our time-resolved fluorescence sensors, will be employed. Once we gain sufficient knowledge on flame dynamics, we will, in the Phase II project, design and fabricate a low cost fire flame simulator based on electronically controlled mixing of light from different LEDs with emission bands ranging from UV to IR. Such simulator will be used for certification and testing of various photodetectors for fire/flame detection without the hassle and safety concerns related to testing of real flames.

For early-fire detection, a very important task is to integrate the data from both SS and TR measurements and classify the flame using the trained ANN, in real-time. The SS data contains the flame signals from UV and IR photodetectors. This alone is partially implemented in some current systems, however they are still subjected to false alarms due to effects from various ambient heat and light sources. ANN is very effective in recognizing patterns which are reproducible. As described above, the repeatable features of flames during their initial formation and development can be used for distinguishing flames from other light (heat) sources. Other features that depend on the chemical composition of the burning substance can be used for source identification.

Data from different possible burning sources will be collected and used for training the ANN. This data will be employed for error-free identification of the flame source. The data contains information collected during the period from just before the ignition to complete development of the flame. The flow-chart in Figure 11 shows the different sources of data, with a view on how the problem could be approached.

We will be also investigating different ways of implementing the fire-detection and fire-prevention methods. The ANN based decision would be evaluated as a stand-alone method or as a second-layer of fire-detection in addition to SS-method and frequency counting. The architecture required to integrate the data from different sources (SS, TR and Frequency counter) will be designed and developed, for an effective method of fire detection and prevention. In addition to architecture design for integration of data and analysis, the selection and optimization of ANN algorithms and architecture will be performed. The final algorithm with all the coefficients derived from the trained ANN will be implemented in an FPGA to allow for a complete portable device.


Radical OH H2O CO, CO2 CH
Wavelength (nm) 306-322 809.7 350-450 431.5



JP-8 0.23 0.32 0.12 0.33 0.34 0.27 0.31 0.33 0.2 0.22
Gasoline -0.35 -0.21 -0.37 -0.35 -0.28 0.36 -0.32 -0.28 -0.26 -0.3