Windsor, Berkshire-based Domain Dynamics Ltd claims that it has successfully combined its signal-processing technology with NeuralDesk, the neural network package developed by Southampton, Hampshire-based Neural Computer Sciences Ltd, thereby providing the basis for intelligent, fully-automatic signal recognition systems. Applications for the new technology include security devices using voice recognition; sophisticated intruder detection systems that can differentiate between different types of intruders man or animal, and locate them with far fewer sensors than conventional systems; and monitors for fixed wing and helicoptor aircraft engines.
By Lynn Stratton
Domain, which supplies signal processor addons for Sun Microsystems Inc workstations, has for the past three years been working on its Time-Encoded Signal Processing and Recognition system, TESPAR at its Swindon research base. TESPAR is a processor-based circuit, developed on a Sun workstation, which can be built into a single custom chip, application specific integrated circuit, or hybrid device. It works by capturing and converting sound waves into codes that it then outputs as matrices. From these, it generates an archetype matrix to which new data can be referred in order to recognise patterns. Hitherto, this process of comparison has involved the use of statistical correlation techniques, but neural networks, which can be trained to respond to particular types of data, in particular ways, provide a means of automatically comparing new input to the reference matrix and determining that which matches – with a remarkable 100% success rate according to Domain’s trials. Domain used Neural Computer’s NeuralDesk package to design its network for signal recognition. NeuralDesk, released in October 1991, is a development package for users of Microsoft Windows 3.X. It enables the construction of network models via a graphical user interface. Developers enter data on spreadsheet tables, and then select the appropriate algorithms and parameters for their application in order to create their network. This done, they can direct and interrogate their model on screen – the system’s engine, Neurun, providing a window graph to plot training progress. Neural-net calculations are performed in time with the application software providing the data via a data direct entry facility. The recommended configuration for running NeuralDesk is a 4Mb 80386 system (although it will run on an 80286) with 80387 coprocessor, 40Mb disk, VGA monitor and mouse. The system supports a maximum of 65,536 neurons and 65,536 synapses, 200 layers and a 16Mb training set, and it is claimed to perform at a rate of 4,500 connections per second; the system costs UKP1,000.
Domain exploits technology developed at Bath
Domain Dynamics Ltd is a UK company started up in 1990 to market signal processing technology developed at the University of Bath and at RMCS Cranfield at Shrivenham. This technology – Time Encoded Signal Processing and Recognition or TESPAR – simplifies the data acquisition and analysis phases of signal processing and recognition. The company supplies a range of hardware and software for applications developers and researchers using signal capture, analysis and processing tools in their work, and also functions as an applications developer and consultant in this field. It is currently developing signal processing techniques for condition monitoring and diagnosis; ‘smart’ acoustic and vibration sensors; self-learning pattern recognition systems; speech recognition; signal analysis; sonar and seismic signature identification; and industrial process control. These are being used for applications as diverse as the monitoring of mining progress to food crispness; the monitoring of industrial machinery, for mechanic imbalances, worn gears, component failure and cracks; and for aircraft health and usage monitoring systems, HUMS, which analyse the vibrations of hel-icopter and plane engines. A variety of security applications have also been developed, from intruder detection to speaker verificati
on and identification systems for controlling access to premises and facilities, and for financial transactions.
Neural Sciences explores neural networks
Neural Computer Sciences Ltd was set up in 1989 by electronics engineer Andy Edmonds, who was inspired by the potential he saw for neural network technology. The first product developed by the company ran under MS-DOS and served as a low cost entry point for users to evaluate and experiment with neural networks. It also inspired the development of the NeuralDesk package which integrated neural networks with the Windows 3 operating environment. Since its release in September 1991, this product has attracted over 5,000 enquiries and around 120 copies have been sold to companies in the UK, France, Germany, Italy, Switzerland, Eire, Australia and the US. The company is currently supporting design-ins for a range of applications and has introduced an accelerator board, NeuSprint as an enhancement to the package. The board increases processing speed 100-fold using signal processing technology to provide a fast 32-bit processor with 640Kb of SRAM on a standard board, and has an optional C compiler. It costs UKP1,500. Another recent introduction is NeuDesk, also for Windows 3, and targeted at the first time user wanting to experiment with neural networks on a personal computer. The package provides all the core neural network development and training capabilities of NeuralDesk and includes several technical enhancements and interface improvements. It costs UKP385 and is upwards-compatible to NeuralDesk. Currently under development, and scheduled for release in November, is NeuForecast, for users wanting to predict the behaviour of continuously varying data. The main uses envisaged for the product are the prediction of financial markets, economic trends and the processing of scientific data. Again the package is for use with Windows 3 and is supplied with a signal processing accelerator board. Neural Computer Sciences is now seeking partners with the appropriate financial and technical expertise to help develop the product.
