Driving a DSP Underground

This paper describes a general purpose digital signal processing architecture for embedded systems. The core hardware modules furnish high price/performance ratio DSP platforms on a compact footprint. Expansion and peripheral boards provide interface, communications and data acquisition subsystems, which may be attached to the DSP modules via a standardised expansion bus. Modular libraries of C and assembly code provide the final ingredient for plug-together embedded DSP solutions. This paper describes the development of several illustrative applications, including a magnetic 3-d positioning system, a centrifuge monitoring system, a 3-d optical measurement system, a tunnel profiling device and an underground vehicle for remote pipe inspection. 1. DSP-90 The DSP hardware is based on the Analog Devices ADSP-21xx family of processors. Currently ADSP-2105 and ADSP-2101 devices are supported, and development of 2181 based hardware has been initiated. The system is bus-based for expandability, and uses stack-through connectors which eliminate the need for card guides and cages in multi-board applications. Conceptually, the approach is similar to that of the PC/104 standard. This is acknowledged in the name assigned to the DSP system: DSP-90. (However the ‘90’ refers to the board dimension in millimetres rather than the number of pins in the edge connector). DSP-90 boards are designed on a circular format, but are provided with snap-off ‘ears’ to enhance the mounting options: Fig. 1(a) shows a 2105 board with the mounting lugs in place, and fig. 2(b) illustrates 2101 board with the lugs removed for mounting in a tubular instrumentation pod. Fig. 1(a) ADSP-2105 based DSP-90 Processor Board Fig. 1(b) ADSP-2101 based DSP-90 Processor Board In addition to the internal processor memory maps, external boot memory (BM), program memory (PM) and paged data memory (DM) are provided, the latter taking full advantage of paging facilities provided by Analog Devices’ programming tools. For enhanced flexibility in tailoring lower cost vs. higher performance solutions, all memory devices are socketed. Surface mount construction is used elsewhere for compactness. Programmable logic decodes the memory map and also provides the page latch and a number of control signals to the expansion bus for addressing memory mapped expansion and peripheral boards. Fig. 2 shows a DSP-90 system comprising a DSP processor board, a general purpose I/O (GPIO) board and an ethernet board. In this case, stand-offs have been fitted to facilitate mounting on a chassis plate (not shown). Fig. 2 DSP-90 System on an Ethernet Node The system illustrated in fig. 2 utilises a ribbon cable connection between the GPIO board (top) and external peripheral devices controlled via the GPIO board. This need not be the case, however. More commonly, a reversed stackthrough connector scheme provides for stacking of multiple peripheral boards on top of the GPIO board. The GPIO board furnishes two independent 16 bit FIFO input buffers, and four programmable bi-directional 8bit parallel ports. Although many of the applications that DSP-90 is targeted for involve strict real-time, the FIFO buffers provide some latitude in the allocation of processing resources between data acquisition, processing, and despatching of results (either to controlled devices or over the ethernet). FIFO devices with up to 8k words capacity may be fitted in the sockets provided on the GPIO board. The current ethernet board is designed around a highly integrated SEEQ chipset, and an NE2000 compatible board (based on a National Semiconductor chipset) is at the prototyping stage. 2. Applications A schematic illustrating a typical deployment of a stack of DSP-90 boards is given below in fig. 3. Here the application is in 2-d profiling. The instrument shown is a pipe profiler, however similar instruments are planned and under development for tunnel profiling and for navigation in autonomous guided vehicles (AGV’s). In the AGV application the instrument will be mounted to rotate about a vertical axis, and will provide accurate ranging data for objects in the vehicle environment, which in turn may be used to solve the vehicles position with respect to a known map, or to map an unknown environment. Such methods are further explored in another paper [1]. Fig. 3 Schematic of a Cross-Section Measuring Device The instrument shown in fig. 3 uses a principle of optical triangulation which determines range using the imaged position of a target spot projected by a laser. Here the DSP is used to enhance the ranging accuracy beyond the sensor pixel accuracy, by centroiding the image of the target. Sub-pixel resolutions of the order of one tenth to one twentieth of a pixel may routinely be attained. This impacts on device performance because the technique allows lower pixel-count sensors to be used, and so line data can be clocked off the charge transport faster, and hence data acquisition rates are increased proportionately. Currently a pipe profiling device with a rotational speed of 1000 rpm is under development, with accuracy in the range of tens to hundreds of micron, inside the one metre radius working envelope. A typical application of the pipe profiler is in sewer inspection. In many ways this application typifies the requirements of an embedded system: The environment is harsh; space in the equipment bay of the instrument pod is extremely limited; power requirements matter; data and control signals must be communicated to and from a remote operator; and above all, the system cost, performance gain and functionality must be clearly perceived by a very down-to-earth industry. Fig. 4(a) shows a current pipe profiling instrument. Fig. 4(b) illustrates a typical tractor vehicle used to deploy the instrument in a sewer pipe, and fig. 5 shows a number of faults which the profiler may be used to detect or quantify. Fig. 4 (a) Pipe Profiling Instrument (b) Typical Tractor Vehicle Fig. 5 (a) Hanging Debris (b) Cracked Pipes (c) Grossly Deformed Pipes (d) Blockages Assisting an engineer to decide exactly when to repair a given sewer is a task which has cost benefits if pertinent information can be provided. Traditionally CCTV surveys have been used to acquire pipe condition data. But CCTV only facilitates a subjective estimation of the degree of degradation. The pipe profiler described above allows cross-sections of sewers to be measured accurately thus enabling the engineer to make objective decisions [2]. But this is only a first step toward a new surveying philosophy. Once a DSP is embedded in the instrumentation system it is possible to process image data by implementing edge detection and Hough transforms [3]. The results may be combined with profile data to form a map of the sewer, which can be indexed with condition classifications derived automatically. Precise information on sewer faults may be archived for further analysis, whilst uninformative data may be eliminated. Fig. 6 illustrates some potential developments. Fig. 6 (a)Sobel Edge Detection (b) Reconstructed Profile Data (c) Pipe Joint Tracking Fig. 6(a) shows an edge detection scheme which could be used to enhance video data for locating missing bricks in large sewer pipes. Fig. 6(b) shows a scene reconstructed from a set of profiles obtained with the instrument of fig. 4. Fig. 6(c) shows the detection and tracking of pipe joints based on Hough transform methods. Using DSP-90 hardware and ethernet communications any and all of these techniques may be combined, with the computation distributed between the data acquisition system deployed in the sewer, the operations room in a van at the surface, and on the engineers workstation. This area of research is currently being supported by Thames Water PLC and by Morgan Collis Ltd. As indicated, there is also potential to deploy this technology in profiling larger scale civil engineering structures [4]. The tunnel profiler project applies the principles of optical triangulation to surveying of tunnels of up to 10m diameter. At these extended ranges the accuracy of measurement diminishes even when the largest linescan sensor arrays are employed in the instrument design. Again, DSP-90 based sub-pixel sampling is the enabling technology and experimental trials show that an inexpensive sensor design will be capable of sub-millimetric accuracy in this application. Sample rates of up to 20k measurements per second should be obtainable. A prototype instrument has previously been constructed to demonstrate the soundness of the method, and the current project is advancing the design of the instrument to the stage where full-scale field trials may be undertaken in order to establish a complete methodology for operating the instrument in a variety of applications. It is intended that many measuring tasks currently effected with a theodolite on a point by point basis may be partly automated. For example, the position of the profiling instrument might itself be determined with a theodolite, and then hundreds or thousands of measurements around the tunnel may be gathered by the profiler. A potential application is monitoring of deformation in tunnels being constructed using the NATM technique. The instrument may also find application in long-term monitoring of man-entry sewers and in conducting ‘Wriggle-Surveys’ of various forms of tunnel construction. 3. Further DSP 90 Applications: 3D-Net Taking the idea of distributed DSP-90 systems one step further is the concept of 3D-Net. 3D-Net was originally conceived as an extension of traditional photogrammetry, where the term ‘network’ refers to a geometrical configuration of cameras which are used to photograph an object of interest. The photographic films are analysed, knowing the camera parameters, to yield highly accurate 3-d data for the object [5, 6]. Digital photogrammetry involves the use of CCD cameras as opposed to photographic film, and the conception of 3DNet was originally as a network of camera stations where each camera contains an embedded DSP-90 system. Thus 3D-Net is a computer network as well as a photogrammetric network. Each camera/DSP-90 station transmits 2-D scene data onto the ethernet, and a central 3-D workstation processes the sets of 2-D data to yield 3-D scene data. Such data may feed into manufacturing process control systems, 3-d imaging systems, or may simply be used to automate traditional applications of photogrammetry. Local processing and intelligent target detection performed at each camera station by the DSP-90 keeps data rates well within the ethernet bandwidth. Two areas that are under current development are jigless manufacturing of advanced aerospace componentry and medical imaging. The former is backed by British Aerospace, with the intention of developing a full demonstrator system geared to jigless wing assembly. The latter is backed by the French Projet Magellan, and indirectly by EDF and the École Polytechnique. Jigless assembly involves a number of measurement cells, each of which has a known absolute position. Within each cell an independent 3D-Net segment tracks the position of componentry such as aileron hinge points. This is achieved using DSP implementations of the traditional photogrammetric methods [7]. A typical camera configuration is shown in fig. 7(a) where the compound curvature of a (marine) propeller is the subject of interest [8]. In fig. 7(b) the view of targets attached to the propeller as seen by one camera station is presented. In fig. 7(c) the rendered CAD model derived from the optical measurement process is shown. Simulation results show that DSP-90 hardware, providing embedded, distributed, numerical processing power, will enable such measurement processes to be carried out in real-time, for up to several hundred targeted locations. Fig. 7 (a) Nineteen camera stations (b) A View of Targets (c) Reconstructed CAD Model Magellan [9] involves 3-d measurement on a smaller scale. The idea is to track medical instruments in realtime during the course of an endo-nasal surgical intervention [10]. The position of the patients head is also determined in real-time. Detection currently relies on a magnetic sensor system based on the mutual coupling between sets of orthogonal loop antennae. An inverse solution of the Biot-Savart equations yields position vectors relating the co-ordinates of the sensors in 3-d space. The resulting data is used to drive a computer visualisation system which guides the surgeon. Pre-operative CT scan data is displayed as a backdrop, upon which is superimposed CAD models of the surgical instruments. Positional data ensures accurate registration between the imaged displays and the actual position of the surgical instruments. A typical display as seen by the surgeon, is presented in fig. 8. Fig. 8 Typical Magellan Display showing cuts through 3-d volume model, with superimposed instrument. The previous Magellan sensor system relied on a closely coupled mesh of transputers to solve the co-ordinates of multiple instruments plus the position of the patients head. Currently DSP-90 and 3D-Net hardware is being developed to replace the transputer system with a distributed solution. Each surgical instrument will be permanently assigned its own DSP-90 system, which in turn will be calibrated for that instrument. If the surgeon requires additional instruments during the intervention, these may simply be connected to the ethernet, as and when required. One or several Magellan workstations may also be connected to the 3D-Net to receive positional data on the net and produce the corresponding visualisation. A schematic is shown in fig. 9. The distributed DSP-90 3D-Net solution has the added advantage of inherent reliability. If one node fails, then the rest of the system should be unaffected.