TABLE OF CONTENTS

• Introduction
• The timing requirements in automated NDE
• The SENDAS concept
  • Task organization
  • Fast supporting hardware
  • Distributed processing
• Experiments in laboratory
• Considerations about execution speed
• Conclusions

Achieving high productivity is of main concern to the cost of the NDE process, especially in the aircraft manufacturing industry applications where large parts have to be inspected. To that purpose, systems with multiple channel configurations, driven by high speed scanning robots are frequently used (automated NDE). Usually, C-scan images are formed from the unprocessed signals, which allows a preliminary flaw assessment. This way it has been feasible to reach high inspection rates (several square meters per hour), with a good spatial resolution.

On the other hand, NDE is a field that can take advantage from the numerous Digital Signal Processing (DSP) algorithms developed over the years [1]. However, in the context of automated NDE, speed constraints have prevented the widespread use of these techniques. Furthermore, the huge data volume produced by the inspection process itself, makes the recording and post-processing of signals an impractical issue. The only valid alternative to this situation is to carry out DSP functions on the fly, simultaneously with the scanning process.

This paper addresses the subject of DSP in the hard real-time context of automated NDE by means of a hardware-based approach: the SENDAS concept. SENDAS (which stands for the Spanish "Sistema para Evaluación No Destructiva de Arquitectura Segmentada"), is a pipeline of hardware processors or Processing Modules (PM), specialized in functions relevant to the NDE applications. All the PMs configured in the pipeline perform concurrently at a fixed speed of 10 MS/s, which allow for real-time processing in most applications. SENDAS performance is evaluated by means of a set of experiments over a thick CFRP composite test piece, involving high dynamic range DAC amplification, digitalization, non linear spatial filtering for EMI noise cancellation, deconvolution, digital envelope extraction and multigated - multipeak detection.

The geometry of the tested parts, acoustic physics and, occasionally other parameters, determine the maximum pulse repetition frequency (prf) useable in a given application. This must account for the following:

• Delay time, (T_d), which is the time interval from transducer firing to the beginning of data acquisition. It is the water path transit time in pulse-echo applications, being the time required by the ultrasonic pulse to reach the receiver end in the through-transmission method.
• Acquisition time (T_q), which determines the depth of the inspected zone. Signals reaching the receiver are conditioned and digitized during this interval. The acquisition time window is usually widened more than strictly needed to accommodate for variations in the water path, sound velocity and material inhomogeneities.
• Supplementary time (T_s), which must be included because of physical (unattenuated sound), mechanical or safety constraints.

 In any application, these time intervals are non overlapping, so that their sum:

T_a = T_d + T_q + T_s (1)

define the application time period T_a. This yields the maximum physically attainable scanning speed, v_p, as the ratio of the required spatial resolution in the scanning direction, d, and T_a for a single channel:

v_p = d/T_a (2)

No matter the fast the underlying equipment is, a given application cannot perform at a higher rate that indicated by Eq. (2). Rather, the operation of the equipment might slow down the application if it were not properly designed.

 

The operation of a high performance UT equipment requires cycling through several steps:

• Channel selection and parameter programming (different channels may require different DAC curves, gain factors, time gates, etc.).
• Signal acquisition during a predetermined time window, maybe synchronized to some event (interface echo, completion of the delay time interval, etc.).
• Digital signal processing, which may involve several algorithms.
• Strong data reduction, which must provide all the relevant information required by the image formation and the flaw assessment processes.
• Transfer of results to an evaluation computer.

Conventionally, these tasks are carried out sequentially, making it very difficult to achieve high speed. Usually, only the signal acquisition and, perhaps, the data reduction (peak detection) processes are carried out in real-time. This gives no allowance for DSP functions unless some well-known concepts are built into the UT system architecture:

Task organization At the higher level, tasks that can be run concurrently with the time constraints of the application may be identified and distributed accordingly. In SENDAS, tasks are organized as follows:

Fig 1: Working principle of the decoupled programming in SENDAS.

• Channel selection and parameter programming, which is done using most of the delay time interval, when no relevant information is present at the receiver. Only those operations that do not fit inside this time interval precede the trigger pulse. This process (decoupled programming), is carried out by a local processor which takes and programs the required parameters from a local memory at high speed (Fig. 1). Besides, giving usefulness to this time interval, decoupled programming releases the host computer from these low-level tasks, making allowance for higher level functions.

• Parallel signal acquisition and data processing. In SENDAS, digital signal processing starts as soon as the validating event is found (interface echo, completion of the delay time, etc.). A programmable number of samples preceding this event can be acquired and processed (this allows, for example, synchronization with the bottom echo). Digitized data are fed into the pipeline of DSP processors usually at a fixed rate of 10 MS/s (which will only be lowered if the sampling rate is below this figure). On the other hand, all PMs accept incoming data and most of them provide results at this rate. Thus, a sustained acquisition and processing rate of several thousands of A-scan traces per second can be achieved, which allow to reach the maximum physical scanning speed in many applications (Fig. 2).

Fig. 2.- Parallel operation of SENDAS.

• Delayed result transfer, which is carried out concurrently with the acquisition and processing of the next A-scan. This provides the whole application time interval for data transfer, which constitutes a bottleneck with most conventional approaches. Furthermore, SENDAS may be programmed in automatic mode, providing results from all the configured channels at once. Packing the information that is sent to the host computer, avoids some software management overheads.

Fast supporting hardware. At the middle and lower level, the key to success in high speed UT is a fast computing support. This is achieved in SENDAS at two hierarchical levels, the architectural framework and the processing hardware:

• SENDAS architecture: DSP functions are arranged in a pipeline of dedicated processors, the Processing Modules (Fig.3).

Fig. 3.- SENDAS architecture.

 Every PM has access to two buses: the control bus, which is conventional, driven asynchronously and which is mainly used for programming purposes, and the signal bus. This bus is segmented, providing data from a PM to the succeeding one in the chain. To this purpose, a time slot is opened every 100 ns. If a PM has a result to transfer to the following one, it is sent; if the result is not yet available, it waits to the next time slot. This simple mechanism guarantees an orderly transfer of data that is handled by three control signals (Fig. 4):

INI: INItial data sample.

/DAV: Data AVailable.

FIN: FINish of a processed A-scan trace.

 

PMs follow a mechanical and electrical standard. They are 10x5 cm. size, having connectors to a motherboard that, besides providing mechanical support, also plays an important role in the architecture.

Fig. 4.- Timing diagram of SENDAS data transfers between PMs.

• Motherboard: Contains logic to control all the PMs configured in a given application. In particular, it has a local processor which handles the decoupled programming and the result transfer tasks. Also, it has provisions to capture all the data traffic between two processing modules. This has proved to be invaluable for debugging purposes of the PMs; but, moreover, it allows getting data partially processed in real applications. It is used, for example, in calibrating the optimum parameters of the application, by providing the original A-scan or the result of any configured DSP function. Two versions of the motherboard have been developed. One is ISA bus compatible, for local applications, while the other has facilities for remote operation through a specially designed full-duplex serial data link (5 Mbit/s). Up to six PMs fit in a single motherboard.
• Processing Modules: Each one is specialized in a given function (analog or digital) oriented towards NDE applications. Currently they include: pulsers of different nature, linear and log amplification with signal conditioning, multi-DAC curve generators, digitalization, FIR digital filtering, spatial trace compounding, general calculator, data reduction, split-spectrum processing and beam formers with 5 ns resolution. Combined operation of the analog and digital modules allows for dynamic ranges in excess of 90 dB, in spite of the 8 bit wide signal bus. Some modules have internal pipelines to reach the nominal processing rate (10 MS/s; only the SSP module performs at a lower speed). Furthermore, most PMs can be reconfigured on-line for a different function. This is possible because the extensive use of in-system programmable logic, mainly RAM-based FPGAs, which allows system reconfiguration on the fly. For example (currently operative functions):
• The same trace-compounding module can be used for spatial averaging, non-linear filtering for EMI noise cancellation or intelligent A-scan reduction.
• The FIR filter (up to 64 coefficients) can be reconfigured for real-time deconvolution, band-pass filtering or other purposes.
• The general calculator may be configured to compute the envelope of a RF signal, its instantaneous power or its derivative.
• The data reduction module implements a multiple gate (up to 4) monopeak detector or a double gate multipeak detector (up to 16 peaks/gate).

Distributed processing. Obviously, not only the UT system affects the performance of a given application. It is customary to combine its operation with the robot-axis control, image formation and evaluation tasks. These tasks are synergically combined in the SARA [2] application set, directed towards the inspection of large aircraft manufactured parts, whose main components are shown in Figure 5.

Fig. 5.- The main components in a SARA system for automatic inspection of large parts.

Following this scheme, the control computer is responsible of trajectory planning and robot positioning, reduced data acquisition, parameter programming and image formation. This is sent to the evaluation computer, where a full set of software tools help the operator in the flaw assessment process. In the whole context, SENDAS cooperates by means of the decoupled programming, digital signal processing and data reduction functions, freeing the control computer from these tasks. Critical to the overall operation is to keep synchronization between the spatial position of the scanning robot and the data provided by the UT system. It is worth to note that, although the result transfer is delayed from the SENDAS system, this is done by a fixed amount that is taken into account by the control computer. Communication between control and evaluation computer is carried out by means of a standard Ethernet link. This architectural approach provides a quite efficient system at a very reasonable cost, mainly determined by the mechanics of the robot arrangement. These subjects are addressed with more detail in [6].

 

A set of experiments has been arranged to show the performance of some real-time DSP functions in NDE applications using the SENDAS approach. Besides conventional functions, like variable gain amplification (DAC) and digitalization, we have included the following ones:

• Cancellation of EMI noise spikes, through a non-linear filtering process involving order statistics [3]. It uses the trace compounding PM.
• Deconvolution applied to increase the axial resolution of the A-scan traces, as described in [4]. It uses the FIR filter PM.
• Envelope extraction of the processed RF signals in real time [5]. It uses the CALculator PM.
• Multi-peak, multi-gate data reduction for the formation of B-scan images. It uses the PIK PM.

To this purpose, a test piece of laminated CFRP composite, 30 mm. thick, is inspected in pulse-echo in an immersion tank, using a 5 MHz transducer. Several flat bottom holes (FBH) of 4 mm. diameter have been drilled to different depths in the CFRP piece. The hole nearest to the surface is at 0.5 mm. depth. SENDAS is configured with the following Processing Modules (Fig. 6):

• Wideband pulser.
• Linear amplifier (A).
• Distance-Amplitude Compensation (DAC).
• Analog to digital converter (A/D), sampling at 33 MHz.
• A Trace compounding module configured for EMI noise cancellation (EMI).
• A digital FIR filter module, configured for deconvolution (DCV).
• A calculator module, configured to extract the envelope of the signal (ENV).
• A multigated-multipeak detector module (DPK).

Fig. 6.- Experimental arrangement.

The SPES software support, developed for our experimental work, is used thoroughly and shown in the subsequent pictures, which have been directly captured from the computer screen (Figure 7 shows the first interface echo under SPES). Signal acquisition and processing are triggered by the completion of the delay time interval (about 25 m s), which also starts the DAC curve generator. The curve is first edited using FBH at different depths to get similar peak indications in the A-scan. Note that, due to the material characteristics, gain must be increased non-linearly from 12 to 65 dB, which, in this case, is done in 0.3 dB steps and with a temporal resolution of 400 ns. White noise effects are reduced at the amplifier stage by means of analog filters, wide enough (Q=1.4) to do not distort the signals.

Fig. 7.- Interface echo shown by SPES.

Furthermore, to simulate real conditions frequently found in industrial environments, we located a poorly filtered DC-controlled motor near the experimental arrangement. This motor produces EMI noise spikes, indistinguishable from a real echo signal, as shown in Figure 8, where the top and bottom interface echoes are also shown (after DAC amplification and using the VIDeo mode). Obviously, these spikes are more harmful at depths that require higher amplification factors (they are amplified about 60 dB at that depth). Figure 9 shows the output of the trace compounding module, which has removed the EMI noise.

Fig. 8.- Effect of electromagnetic interference (EMI) caused by a motor. Note the noise spikes near the bottom echo.
Fig. 9.- Output of the EMI filtering module, where the noise spikes have been removed.

As it is well known, deconvolution increases axial resolution.This function must be applied to the RF A-scan, because the envelope extraction is not a linear function (although this fact is not fully recognized by some authors). This gives applicability to the digital envelope detector, located after the deconvolver module (Fig. 6). Figure 10 shows the input to the deconvolver when the transducer is located directly over the FBH nearest to the interface; as it can be seen, it is quite difficult to separate both echoes. This signal is deconvolved and the resulting envelope is extracted. Fig. 11 shows the result of this process, which has allowed to separate clearly the two echoes.

Fig. 10.- RF input to the deconvolver module, when the transducer is located directly over the nearest FBH.
Fig. 11.- Output of the digital envelope detector module after deconvolution. The two reflecting surfaces are visible.

 

Finally, just a word about execution speed of this application. As seen in the pictures above, a total of 940 samples taken at 33 MHz compose a A-scan (about 30 ms, the round-trip time across the test piece being around 20 ms). Thus, the minimum delay time must be 15 ms to avoid the second interface echo entering the acquisition window (we used 25 ms). Also, this inspection requires a supplementary time interval of, at least, 5 times the delay time to allow attenuation of the subsequent interface echoes. The resulting application time interval is (Eq. 1):

T_a = T_d + T_q + T_s = 15 + 30 + 5·15 = 120 m s (3)

Thus, maximum physically attainable prf is 8.3 KHz for this application. At the nominal processing rate (10 MS/s), SENDAS takes about 94 m s to acquire and process a single A-scan trace, including all the configured processing functions. This is below the 30+5·15=105 ms time allowed for these tasks by the application (see Fig. 2). Parameter programming can be easily fitted into the delay interval of 15 ms (in this specific case no parameters have to be modified between pulses; in a multichannel application, only memory bank switching would normally be required to change DAC curves). Finally, results transfer (3 bytes position-amplitude pairs for every detected peak), is easily achieved in the 120 ms application time interval.

The conclusion is clear. SENDAS performs transparently to this application, reaching the maximum physically attainable scanning speed of more than 8000 traces/s. By the way, in our lab we have no means of scanning that piece at several m/s, but the results show it would be possible in a real, perhaps multichannel, application....

It is worth to compare these results with software based approaches. To this purpose, we have built a library of functions in MATLAB which emulate the SENDAS functions, using the most effective algorithms. Using a Pentium PC at 200 MHz, we registered the following execution times:

Trace compounding for EMI noise cancellation	2.4 ms
Deconvolution by inverse filtering using FFT	8 ms
Envelope extraction through the Hilbert transform	20.3 ms
Monopeak detection on a single time window	1,5 ms

Since execution of these algorithms by software is a sequential process, we add the times together to obtain 32.2 ms of total execution time. This would prevent to achieve this application at more than about 30 traces/s, that is, 250 times below the SENDAS performance.

 

The SENDAS concept has been presented and its performance in digital signal processing of ultrasonic signals for NDT has been evaluated. It has been shown that the proposed approach allows reaching the maximum physically attainable scanning speed in a given application. In this sense, SENDAS is transparent to the application, while providing, at the same time, an improvement in the information content of the signals through DSP functions executed in real time. Modular design and the incorporation of advanced architectural concepts allow a great flexibility of application.

SENDAS is currently commercialized by a Spanish industry (TECAL, S.A.), being used by Construcciones Aeronáuticas, S.A. (C.A.S.A.), in their manufacturing facilities. It has been certified by Mc Donell Douglas and Boeing for the inspection of different aircraft structural parts, including honeycomb cores and laminated CFRP.

It must be highlighted that a team of researchers and engineers, whose work has concentrated on the implementation of every SENDAS function, has developed all the processing modules. The defined SENDAS standard has provided the glue to join all these efforts into an operational high performance system.

Grants CYCIT TAP-94-0812 and C.A.M I+D 52/94 have supported this work, jointly with TECAL, S.A. Also, some results from PACE project PC144, supported by the European Community and the Spanish Ministry of Industry and Energy, have been incorporated to the SENDAS concept.

The authors want also to thank the stimulating disposition of colleagues at C.A.S.A., whose positive criticism and cooperating behavior helped (and is helping) to improve many of the SENDAS performance.

Part of the work above has been extracted from our paper: "SENDAS: An Approach to Modular Digital Signal Processing for Automated NDE", published on Proc. 2^nd. Conf. on NDE Applied to Process Control of Composite Fabrication, Oct. 1-2, St. Louis, Missouri, pp.163-181. We want to thank Prof. George A. Matzkanin for its cooperative attitude to this subject.

[1] G.P. Singh, J.L. Schmalzel, S.S. Udpa: The application of digital signal processing and pattern recognition to ultrasonic and electromagnetic nondestructive testing and evaluation. NTIAC-90-2, Southwest Research Institute, San Antonio, Texas, Feb. 1991.

[2] C. Valdecantos: SARA 10: A system for high speed automatic NDE of large composite parts. Proc. of the 2^nd. Conf. on NDE applied to Process Control of Composite Fabrication. October, 1-2, 1996. St. Louis, Missouri, pp. 195-210. Abstract

[3] I. Pitas, A.N. Venetsanopoulos: Nonlinear Digital Filters, Kluwer Academic Pub., Norwell, Mass. 02061, USA, 1995.

[4] J.J. Anaya, L.G. Ullate, C. Fritsch: A Method for Real-Time Deconvolution, IEEE Trans. on Instrum. and Meas., vol. IM-41-3, pp. 413-419, June 1992.

[5] C. Fritsch, A. Ibáñez, M. Parrilla: An Envelope Detection Filter (in revision, IEEE Trans. on Instrumentation and Measurement).

[6] A. Ibáñez, et al.: A Distributed Architecture for Automatic NDI of Aircraft parts. This UTonline Journal.

AUTHOR

C. Fritsch, L.G.Ullate, T.Sánchez, A.Ibáñez, M. Parrilla.
R.Mateos, M.A.G.Izquierdo, O.M.Graullera, J.Giménez, E.Villanueva
Instituto de Automática Industrial (CSIC)
La Poveda (Arganda) 28500 Madrid (Spain)
Email: carlos@iai.csic.es
http://www.iai.csic.es/users/end/

For more information see: NDT in Aerospace - UTonline 11/97