Meas. Sci. Technol.11 (2000) 157–166. Printed in the UK PII: S0957-0233(00)08404-6

A new apparatus for non-destructiveevaluation of green-state powdermetal compacts using theelectrical-resistivity method

Gene Bogdanov, Reinhold Ludwig and William R Michalson

Department of Electrical and Computer Engineering, Worcester Polytechnic Institute,Worcester, MA 01609, USA

Received 5 October 1999, in final form and accepted for publication 16 November 1999

Abstract. This paper presents a new apparatus developed for non-destructive evaluation(NDE) of green-state powder metal compacts. A green-state compact is an intermediate stepin the powder metallurgy (PM) manufacturing process, which is produced when a metalpowder–lubricant mixture is compacted in a press. This compact is subsequently sintered in afurnace to produce the finished product. Non-destructive material testing is most costeffective in the green state because early flaw detection permits early intervention in themanufacturing cycle and thus avoids scrapping large numbers of parts. Unfortunately,traditional NDE methods have largely been unsuccessful when applied to green-state PMcompacts. A new instrumentation approach has been developed, whereby direct currents areinjected into the green-state compact and an array of spring-loaded needle contacts recordsthe voltage distributions on the surface. The voltage distribution is processed to identifypotentially dangerous surface and sub-surface flaws. This paper presents thecustom-designed hardware and software developed for current injection, voltage acquisition,pre-amplification and flaw detection. In addition, the testing algorithm and measurementresults are discussed. The success of flaw detection using the apparatus is established byusing controlled samples, which are PM compacts with dielectric inclusions inserted.

Keywords: nondestructive evaluation, green-state powder-metallurgy compacts,instrumentation, array sensor, resistivity testing

(Some figures in this article appear in black and white in the printed version.)

1. Introduction

1.1. Powder metallurgy

In a powder-metallurgy (PM) manufacturing process, themetal parts are formed by compressing metal powder athigh pressure. The resulting ‘green-state’ compacts are thensintered in a furnace to produce the final products [1, 2].This manufacturing process is fully automated, very fastand efficient. However, the PM manufacturing process isin need of quality assurance, because the occurrence offlaws in the compacts can significantly reduce the outputefficiency, adversely affecting cost. The main quality hazardin PM compacts is cracking. Cracks occur mainly duringcompression and ejection of the green-state compact [3].Unfortunately, quality assurance in powder metallurgy hasbeen successfully applied only to the finished state [4]. Thedelay from compaction to quality-assurance inspection canrange between hours and days. Therefore, a large numberof flawed parts produced during that period may have to bescrapped before the process is corrected. In order to improve

this situation, it is desirable that flaws in PM compacts bedetected early in the process, preferably in the green state.

Despite considerable efforts, traditional non-destructiveevaluation (NDE) methods have been largely unsuccessfulat detecting flaws in green-state PM parts [5–8]. Ultrasonictesting does not render repeatable results because the green-state PM materials strongly attenuate the elastic waves.Additionally, the individual powder particles randomlyscatter the sound waves, further reducing accuracy. Eddy-current testing encounters limited field–medium interactionbecause the PM material has a very low conductivitycompared with that of metals. The random particledistribution complicates the induced eddy current patternsand also degrades the reproducibility of the measurements.X-ray imaging, although it is usable for flaw detection inPM parts, cannot easily detect small near-surface and cornercracks. However, these locations are the preferred siteswhere in practice most flaws occur [3]. Thermal imagingis hindered by the relatively low thermal conductivity of thegreen-state compact. Because of these difficulties, classical

0957-0233/00/020157+10$30.00 © 2000 IOP Publishing Ltd 157

G Bogdanovet al

Figure 1. The measurement apparatus designed for non-destructive testing of green-state PM compacts using the electrical-impedancemethod.

NDE methods are difficult, if not impossible, to use as aninspection technique for green-state PM parts.

1.2. Electrical-resistivity testing

The apparatus developed uses an alternative method based onelectrical-resistivity testing to detect flaws in green-state PMcompacts. This method was first proposed to a PM industryconsortium in 1996 [4]. The apparatus, shown in figure 1,uses a sensor consisting of many spring-loaded probes tocontact the sample under test. Direct current is injectedthrough two of the probes and voltages are measured usingthe remaining probes deployed over the surface. A flawin the material produces a characteristic voltage responsethat can be detected. This method allows fast inspectionbecause the sensor has to be positioned only once on thesurface of typically sized parts. Although the method is mostsensitive to surface-breaking flaws, subsurface flaws can alsobe detected, down to a certain depth. The apparatus presentedis relatively inexpensive and perhaps the only drawback isthe need to build different sensor configurations for parts ofdifferent geometries.

The NDE method implemented with the new apparatusis based on the four-wire resistivity-measurement method[9]. This method is illustrated in figure 2. Four probescontact the surface of a material of unknown, but assumedconstant, conductivity. Current is injected though the twoouter probes and the voltage is measured through the innerprobes. If the sample is sufficiently large compared with thesensor, boundary conditions can be neglected and the solid isapproximately modelled as a half-space. For this simplifiedsituation, the relationship between the conductivityσ and thevoltage and current measurements is

σ = I

2πV

[(1

r1− 1

r2

)−(

1

r3− 1

r4

)](1)

whereV is the measured voltage,I is the injected currentandr1 to r4 are distances indicated in figure 2.

Figure 2. The four-probe resistivity-measurement arrangement.

The NDE testing apparatus depicted in figure 1 uses asimilar type of measurement, but the sensor contains manymore probes, thus covering the entire area of the sample.The three main components of the measurement apparatusare the press, the sensor and the part holder. The press,consisting of a platform moved by a stepper motor, lowersthe sensor onto the surface of the part while the part rests inan appropriately shaped holder at the base of the fixture. Thesensor, shown in figure 3(a), is a planar array of spring-loadedpins that establish direct electrical contact. The probes on theperiphery of the sensor array are used for current injection.The remaining probes, including the current-injection probesthat do not carry current, are utilized for voltage recording. Toconduct measurements, a constant direct current is injectedthrough a pair of outer probes in one of four directions,illustrated in figure 3(b), and voltages between adjacentvoltage probes in the direction of current flow are measured.

To demonstrate the flaw-detection method, a simplifiedten-probe sensor configuration is shown in figure 4(a).

158

Evaluation of powder metal compacts

(a) (b)

Figure 3. The layout of the planar multi-probe sensor utilized as part of the electrical-impedance NDE apparatus: (a) voltage and currentcontact points and (b) the four current-flow directions.

(a) (b)

Figure 4. The simplified ten-probe impedance measurement set-up and the resulting current-flow pattern for (a) an unflawed part and(b) a part with a surface-breaking flaw.

Figure 5. Surface voltage distributions measured on an unflawedand a flawed part using the ten-probe array depicted in figure 4.

Figure 4(b) depicts the effect of a surface-breaking flaw on theinternal current-flow pattern. The voltage distribution on thesurface of the part, sampled at the voltage-probe locations,is shown in figure 5. The presence of a flaw perturbs thecurrent flow and produces a higher than normal voltage dropin its vicinity [10]. The perturbation due to the flaw is moreevident in the differential voltage distribution seen in figure 6.The differential voltages between adjacent pins are measured.The magnitude of the differential voltage peak near the flawlocation depends on the size and, if the crack is a subsurfaceone, its depth. The peak can be detected using a statisticalcomparison with unflawed parts. As a first step in thetesting process, differential voltage distributions from a set

Figure 6. Differential voltage distributions measured on anunflawed and a flawed part using the ten-probe array shown infigure 4. Differential voltages between adjacent voltage probes aremeasured.

of unflawed parts are collected. These voltages are combinedinto one voltage distribution, called a baseline, with standard-deviation information for each discrete voltage sample.Voltage distributions taken from parts of unknown qualitycan then be compared against the established baseline. If atleast one voltage recording deviates by more than a predefinednumber of standard deviations from the baseline, the part isconsidered flawed. This so-called statistical-distance methodis used with small modifications as the standard method forflaw detection in PM compacts [10–12].

159

G Bogdanovet al

Figure 7. The block diagram of the customized hardware developed for the instrument for non-destructive testing of PM compacts.

Figure 8. The assembly of the NDE apparatus for the inspection of green-state PM compacts.

2. The design approach

The system consists of three basic components: a sensormounted on a mechanical fixture, the front-end hardwareand the control and flaw-detection software. The sensorcaptures the measurement information from the PM compactbeing tested. The front-end electronic hardware processesthe measurement and delivers it to the host PC for storageand additional signal analysis. The software residing inthe PC governs the data-acquisition process, performs theactual flaw detection and controls communication with theuser. Figure 7 illustrates the structure of the data-acquisition-hardware system and figure 8 shows the system in its fully

assembled state with its key modules. The operation of theindividual modules is discussed in the following sections.

2.1. Mechanical instruments

The mechanical press is responsible for holding the PM partand bringing the sensor into physical contact with it. Thefixture shown in figure 1 is custom tailored to accommodatea particular geometrical size of the PM compacts being tested.To carry out the NDE inspection, the PM part is placed in thepart holder located at the bottom of the fixture depicted infigure 1. A mould made of insulating material is used to holdthe part in place. Two calibrated dials allow adjustment of

160

Evaluation of powder metal compacts

the horizontal position of the part relative to the sensor. Thesensor itself is mounted on a platform that translates by meansof a vertical, threaded shaft powered by a stepper motor. Thestepper motor allows accurate positioning of the sensor toensure reproducibility of measurements. Both the sensor andthe part holder can be reconfigured to accommodate differentPM compact geometries.

The press is designed for a pin-pointing accuracy of0.025 mm and a part-positioning accuracy of 0.025 mm.Because the pins tend to slightly translate in their socketsand make contact in a slightly different place every timethe sensor descends onto the part, the 0.025 mm pin-placement inaccuracy translates into approximately 1%average measurement error for a pin spacing of 2.54 mm.That this mechanical inaccuracy is the largest source ofmeasurement error in the system will become evident later.

2.2. The front-end electronics

The front-end electronics performs the tasks of injectingcurrent into the PM sample, collecting the voltage data anddelivering it to the PC. The block diagram in figure 7 showsthe major hardware modules (i.e. the current source, currentmultiplexer, voltage multiplexer and amplifier) and theirinterconnection. The current source generates a constantdirect current that is injected into the PM sample, initiating avoltage distribution across the surface of the green-state PMmaterial. Multiple current-injection directions are achievedthrough current multiplexing using a multiplexer–routercombination. The voltages recorded at the sensor probesare multiplexed (by the voltage multiplexer), amplified (bya switched-gain amplifier), sampled (by the data-acquisitionboard) and further processed by the software. Additionally,the front-end hardware includes a stepper-motor controllerand driver that allow computerized control of the press.

The current necessary to create voltage drops on thesurface of the PM compact is supplied by the current-source module. It is a voltage-controlled current sourcethat is specified to deliver direct currents from−1 to +1 Ato a 5� resistive load with an accuracy of 0.1%. Thevoltage-control signal for the current source is supplied by adigital-to-analogue converter (DAC) output located on thedata-acquisition board, which is under the control of thesoftware. The current magnitude of 1 A was chosen asthe most suitable for measurements on PM compacts forthe following reasons. The current has to be sufficientlylarge to create measurable voltage drops on the surface ofthe highly conductive PM samples. It was established thatthe conductivity of an average steel PM sample is of the orderof 15 000 S m−1, resulting in voltage drops of approximately0.1–1 mV. Smaller voltages become increasingly difficult tomeasure accurately because of an increase in noise. Thecurrent also has to stay within reasonable limits so as notto damage the probes due to heating at the current-injectionpoints and not to present problems to the multiplexing androuting circuits.

Instead of having four separate current sources to providecurrent flow in four different injection directions, a singlecurrent source is multiplexed between the different current-injection probes. This arrangement saves space, cost and

calibration efforts. The current multiplexer circuit consistsof analogue switches that can independently connect theoutput of the current source and ground to eight output lines(four current-source and four ground lines). The multiplexeroutputs also pass through an 8 to 100 current router onthe way to the sensor probes. The current router is a unitthat directly connects the current multiplexer outputs to anyof the 100 sensor probes. This current multiplexer–routerconfiguration was selected as an alternative to having a verylarge multiplexer. The router placed inside the sensor headprovides the ability to use any of the 100 probes for currentinjection while reducing the number of interconnection wires.The current source and multiplexer cannot be placed insidethe sensor head due to their high power dissipation ofapproximately 10 W. Four current-injection directions werefound to be sufficient to conduct the non-destructive testing[11]. A fifth current-source output of the multiplexer (notshown in figure 7) connects to a resistor that is used to measurethe current strength. However, it primarily serves as a loadfor the current source when the sensor is not in contact witha PM sample.

The voltage multiplexer is similar in function to thecurrent multiplexer. It selects any two sensor probes andoutputs the voltages as a differential signal. The voltagemultiplexer consists of two identical 104 to 1 analoguemultiplexers that share their inputs but are controlled byseparate digital signals. In addition to sensor signals, thevoltage multiplexers admit four additional signals from avoltage reference. The reference signals are used to calibratethe gain of the amplification stages that follow the voltagemultiplexer.

The signal-conditioning module amplifies and filters thevoltage signal at the output of the voltage multiplexer. Theideal configuration would be to sample this voltage responseand convert it into digital form before transmitting it over aninterface cable to the PC. For the prototype design, however,the analogue-to-digital conversion is left to a commercialdata-acquisition card installed inside the PC. The amplifierallows switching its gain to obtain the maximum allowablesignal magnitude over the cable in order to reduce the effectof noise generated by the cable itself. Combined with aswitched-gain amplifier in the data-acquisition card, thisconfiguration permits one to obtain the maximum possibleaccuracy of the sampled voltage. The gains of the amplifierare calibrated in software using the voltage accessed throughthe voltage multiplexer. The signal-conditioning module alsocontains a first-order passive filter intended to remove high-frequency noise from the voltage signal.

The PC interface consists of the digital interface moduleand the data-acquisition card. The digital interface moduleconnects to the PC parallel port and provides the controlsignals for all the analogue switches present in the custom-built front end. The data-acquisition card provides up toeight differential analogue inputs and two outputs. One of theanalogue outputs controls the current source, while one of thedifferential analogue inputs receives the voltage signal fromthe signal-conditioning module. Another analogue inputis used to measure the voltage on the current-source loadresistor. A third analogue input is employed to monitor thevoltage across the reference resistor, ensuring that the currentsource does not saturate.

161

G Bogdanovet al

2.3. Software development

The software residing in the PC is responsible for performingdata acquisition, data processing and communication withthe user. The design of the software is modular, allowinga change of the underlying hardware and the user interfacewithout interference with the operation of the main portionof the software responsible for the testing of PM parts.

The software environment was designed with futurechanges and expansion in mind. Figure 9 shows thehierarchical block diagram of the software and hardwaresystems. The environment is subdivided into layers, eachresponsible for a particular, independent task. The hardwareaccess layer communicates with the custom hardware andhides the details of the hardware implementation from thelayer above it. If the underlying hardware is modified, thehardware access layer is adjusted accordingly while the restof the software remains unchanged. The measurement layerautomatically makes accurate voltage measurements andhides the details of amplifier gain selection and measurementaveraging from the layer above. The data-storage and data-acquisition layer performs the measurement sweep and storesthe voltage data on disk in a consistent manner. The testingalgorithm layer performs flaw detection using the acquireddata. Testing algorithm modules may be added and modifiedin the future without disturbing the rest of the software.The user-interface layer communicates with the user and,if the need should arise, can be replaced by a different userinterface.

Figure 10 shows a snapshot of the non-destructive testinggraphical user interface (GUI) with an example of a baselineand a voltage data document open. Toolbar buttons orfoot pedals (shown in figure 8) are used to perform thebasic operations of non-destructively testing the PM parts(lowering the sensor, acquiring voltage data, testing thesample and raising the sensor). To perform the NDE task,the software uses the statistical-distance algorithm mentionedpreviously. The test result is displayed as an intensityplot, showing the location and magnitude of the flaw. Formanufacturing-floor convenience, the test results may alsobe observed on indicator lights: green for unflawed andred for flawed. In addition, the software also displaysthe voltage data that can be used primarily for researchpurposes.

The statistical-distance testing algorithm, brieflydescribed in the introduction, first establishes a baseline.Approximately 20 known unflawed samples are measuredand the data are stored on disk. Next, the baseline calculationroutine (invoked through the menu) averages the voltage datafrom these samples and calculates standard deviations foreach differential voltage, thus creating a baseline. Eachbaseline voltage is calculated as an average over the set ofunflawed samples according to

V̄i = 1

N

N∑j=1

V measi (j) (2)

whereV̄i is theith baseline voltage,V measi (j) represents theith voltage measurement from thej th sample andN is thenumber of samples included in the baseline. The indexi

Figure 9. The block diagram of the data-acquisition and testingsoftware system. All major hardware components are shown aspart of a hierarchical layer organization.

ranges from 1 to 342(=90× 2 + 81× 2), the total numberof differential voltage measurements when four current-flowdirections are utilized.

Each baseline voltage is assigned a standard deviation,calculated as follows:

σVi =(

1

N − 1

N∑j=1

[V measi (j)− V̄i ]2

)1/2

(3)

whereσVi is the standard deviation assigned to theith baselinevoltage. Using the mean and standard-deviation data, we canstatistically determine how far the voltages deviate from theirexpected values. For each voltage measurementi conductedon a PM sample of unknown quality, astatistical distanceSDi is calculated:

SDi = |Vmeasi − V̄i |σVi

. (4)

The larger the statistical distance the farther the voltagedeviates from the established standard and the less likely is theevent that the variation is due to measurement noise or smallconductivity imperfections in the sample. If the statisticaldistance exceeds a user-selected threshold, the sample isidentified as flawed. The locations where the statistical

162

Evaluation of powder metal compacts

Figure 10. The user interface of the data-acquisition and testing software. Two documents are shown (baseline and voltage data). Thebaseline is a template calculated using unflawed samples against which all unknown samples can be compared. The voltage-data documentdisplays the result of a comparison with the baseline for a horizontally oriented flaw.

distance exceeded the threshold are shown on an intensityplot signifying the surface of the sample under the multi-probe sensor. As can be seen in figure 10, this intensity plotclearly identifies the location, orientation and even relativesize of the flaw.

An important property of the statistical-distance methodis the dependence on the standard deviation of the baselinevoltages. The greater the variation inherent in the baselinethe larger the standard deviation, making the algorithmless sensitive to the presence of flaws. Baseline standarddeviations depend on material composition and measurementerror. Materials have a large influence on baseline standarddeviations. Among several powders studied [12], pureiron exhibits the lowest standard deviations of 5.5% of themeasured voltages, steel 10–15% and brass 15%. Differentstainless steel powders give results ranging from 7% to20%. Measurement error with the current instrument maycontribute as much as 1–1.5% to the baseline standarddeviations.

In addition to the measurement and testing software,more advanced characterization algorithms have beendeveloped as stand-alone packages. These include estimationof the average conductivity of a PM sample, improvedflaw-detection methods, an experimental dipole model forflaw characterization and an experimental two-dimensionalconductivity-mapping algorithm that allows reconstructionof the conductivity distribution inside the sample from thesurface voltage data [12].

3. Instrumentation results

3.1. Measurement accuracy

The performance of the data-acquisition hardware ischaracterized by measurement accuracy. The higher theaccuracy the better the system will be able to distinguishvoltage responses from small and deep subsurface flawsfrom background noise. The statistical-distance algorithmimplemented in software provides the means to calculate thestandard deviation from a set of voltage measurements. Thisis done through a baseline calculation. The baseline standarddeviation can then be used as an indication of the voltage-measurement accuracy.

Three distinct error sources can be examined using thistechnique: electrical noise, errors due to pin placement anderrors due to the alignment of the sensor with respect to thepart. The electrical noise is the noise generated by the data-acquisition circuits combined with external noise picked upby the wiring. The error associated with pin placement isdue to pin misalignment each time the sensor is lowered ontoa PM sample. Part-placement inaccuracy causes changes inthe surface texture under the sensor, thereby increasing thepin-placement errors.

The first test of accuracy involves establishing a baselinefrom multiple measurements made on the same compactwithout moving the sensor. The sensor was positioned on thesurface of the part and ten consecutive measurements wereacquired and included in a baseline calculation. The fact thatthe sensor remained stationary ensured that no mechanicalerrors were factored into the measurements. The resulting

163

G Bogdanovet al

Table 1. Measurement-related error sources of the instrument.

MeasurementType of measurement Noise source standard deviation (%)

Sensor unmoved Electrical 0.076Lifting the sensor Above plus pin placement 0.63Removing the part Above plus part placement 1.56

Figure 11. Process steps in the manufacturing of controlled samples for various flaw configurations.

average baseline standard deviation is listed in table 1 under‘sensor unmoved’. The accuracy of the electrical system inthis case is better than 0.1%, making it the smallest source oferror in the system.

A second test is conducted by measuring the samesample ten times, lifting and lowering the sensor prior to eachmeasurement, but not moving the part. This test captures theerrors due to pin placement but not sensor-to-part alignment.Ten measurements are acquired and a baseline is calculated.The average baseline standard deviation is listed in table 1under ‘lifting the sensor’.

The third test is conducted by measuring the samesteel PM sample ten times, removing the sample from themeasurement apparatus and replacing for each measurement.This test combines all the errors due to the measurementsystem. The average baseline standard deviation for this testis listed in table 1 under ‘removing the part’. The standarddeviation is now greater than 1%, but is still significantlyless than the standard deviation of a baseline that includesmeasurements of PM samples manufactured during differentdays of the working week. The baseline standard deviationsfor controlled PM compacts range from 5.5% to 20%. A1.56% error due to the testing equipment is acceptable underthese testing conditions.

3.2. Flaw-detection performance

The feasibility of the electrical-impedance method of flawdetection has already been established [10, 13]. The currenttask is to determine the flaw detection accuracy of the newapparatus. Both production and controlled samples areavailable to assess the detection performance. Althoughproduction samples present the most realistic test for the

apparatus, there is no direct means of determining whether asample is indeed flawed or unflawed.

Controlled samples of simplified shape permit thecreation of flaws through use of dielectric inserts [12]. Themanufacturing process for such a controlled sample is shownin figure 11. During the first step, a portion of the powderallocated for the sample is measured and set aside. Theamount of powder set aside is calculated such that it producesa layer equal in thickness to the flaw-depth specification. Therest of the powder is poured into the die and the plastic flawis inserted vertically into the powder. In the second step, theremaining amount of powder is added and then compacted.Part 3 of figure 11 shows the resulting controlled sample witha flaw located near the top surface. The advantage of sucha procedure is that the location and size of a flaw can becontrolled with a good degree of accuracy, allowing precisemeasurement of the instrument’s sensitivity.

The controlled samples were manufactured usingatomized stainless steel powder with an additive of 0.75%lithium lubricant. The samples were compacted to a nominaldensity of 6.45 g cm−3. The overall sample size of67.31 mm× 67.31 mm× 19.05 mm is selected to eliminatecorner effects. The simulated flaw, cut from plastic shimstock, is 0.1 mm× 10.16 mm× 2.54 mm in size. Flawswere placed at the surface and at four subsurface locationsat depths of 1.3, 2.5, 5.1 and 7.6 mm respectively. Threesamples for each flaw depth were fabricated. In addition tothe flawed samples, 20 unflawed samples were manufacturedto provide an accurate baseline for the testing algorithm.

The controlled samples were analysed using boththe standard statistical-distance method and an improvedstatistical-distance method that takes into account theaverage sample conductivity according to the four-probe test

164

Evaluation of powder metal compacts

1 100

20

40

60

80

100

standard methodimproved method

Threshold

% f

als

e id

en

tific

atio

n r

ate

Figure 12. The theoretical false-identification rate versus thethreshold setting in standard deviations for unflawed parts.

(discussed later). A quantity useful for evaluation of flaw-detection performance is the so-called false-identificationrate. It is the percentage of parts identified incorrectlyby the algorithm, for example unflawed parts identified asflawed. The false-identification rate can be plotted against thestatistical-distance threshold to observe the effect of choosingdifferent threshold levels. Separate false-identification-ratecurves are plotted for flawed and unflawed samples. Asthe threshold increases, more unflawed parts are identifiedcorrectly, but at the same time more flawed parts are identifiedas unflawed. This is a property of the statistical-distancealgorithm. The choice of the optimal threshold is normallymade experimentally by plotting false-identification rates forparts of known quality. The false-identification-rate curve forunflawed samples can also be modelled theoretically [12],yielding the result shown in figure 12. For the standardmethod a conservative threshold is found to be approximately4, and for the improved method, approximately 5 standarddeviations. These conservative threshold values are biasedtowards the correct identification of most unflawed samplesat the expense of admitting some marginally flawed samplesas unflawed.

Figure 13 shows the false-identification rates for thecontrolled samples obtained using the standard statistical-distance method. 20 unflawed parts are used to calculate abaseline, giving an average baseline standard deviation of11.5%. The legend in figure 13 identifies the flaw depthfor each flawed sample. Although there is some variation,the rule that the deeper the flaw the smaller its voltageresponse is generally followed. From the false-identification-rate diagram, we can observe that all flaws down to a depthof 2.5 mm can be detected. However, at the depth of 5.1 mm,only one part out of three is identified correctly. Most ofthe samples with flaws 5.1 mm deep and deeper cannot bedistinguished from the unflawed samples.

Certain improvements to the original statistical-distancemethod can be designed to enhance the algorithm’s sensitivityto subsurface flaws. Specifically the determination of thesample’s conductivity and its incorporation into the baseline

Figure 13. False-identification rates obtained using the standardstatistical-distance algorithm for a set of controlled sample data.The left-hand curve shows the baseline established from 20unflawed parts. If the threshold level is set to correctly identify allunflawed parts (3.8), certain subsurface flaws can no longer bedetected.

normalization improves the depth sensitivity. Becauseconductivity varies with density, different PM compacts mayexhibit different average conductivities because the amountof powder used to make them fluctuates. This conductivityvariation does not significantly affect the quality of the part,but it interferes with the detection method. Estimatingthe conductivity of the sample and then normalizing allmeasurements with respect to it removes the dependenceon conductivity, thus improving the sensitivity to smallflaws. The conductivity estimation is accomplished usingan advanced version of the method based on equation(1). All measured voltages are then normalized throughmultiplication by the conductivity of the sample. Voltagesare normalized prior to calculating the baseline and beforecalculating the statistical distances.

Another improvement to the standard testing method isthe use of differential voltages spanning longer distances inthe baseline. Subsurface flaws produce voltage responseswith peaks separated by a large distance, far beyond 2.54 mm.Using these longer distance measurements makes smallvoltage responses of the subsurface flaws more visible to thealgorithm.

Figure 14 shows the false-identification rates obtainedfor controlled samples using the standard and the improvedalgorithm. The improved algorithm implements all theimprovements described above. By examining the curvesobtained by using the improved algorithm, we can find athreshold around 3.8 that allows us to identify all partscorrectly. A conservative threshold around 5, accordingto figure 12, still permits detection of all flaws down to adepth of 5.1 mm and one out of three 7.6 mm deep flaws.This is a significant improvement over the original method.More advanced detection and characterization algorithms arecurrently under development [10]. These algorithms usetomographic reconstruction and neural-network approachesthat are likely to significantly improve the detection accuracyin the future.

165

G Bogdanovet al

Figure 14. A direct comparison between the standard andimproved statistical distance algorithms based on controlledsample data.

4. Conclusions

A new non-destructive testing system based on directcontact resistivity measurements using direct current hasbeen developed. The electrical-resistivity method is so far theonly non-destructive evaluation method that has been appliedsuccessfully to green-state PM compacts. A prototypeinstrument was designed with the ultimate goal of industrialimplementation. The hardware and software developed canbe used as a prototype for an industrial system.

The instrument demonstrated good voltage recording ac-curacy for iron- and steel-based compacts with conductivitiesof the order of 15 000 S m−1 (measured voltages are 0.1–1 mV). A measurement error of approximately 1% is suffi-cient for flaw detection using the current statistical-distancealgorithm. If future detection algorithms require higher ac-curacy, the voltage-measurement hardware can be improved,but most importantly the mechanical accuracy of the sen-sor pins and the press must be improved because the largestsource of error is the mechanical system.

Flaw-detection performance of the apparatus isestablished by using controlled samples, which are PMcompacts with controlled dielectric inclusions simulating

flaws. The tests show that the apparatus reliably detects flawsthat break the surface and subsurface flaws down to a depth of2.5 mm. Using an improved detection method, it is possibleto detect flaws at depths of 5.1 mm, and often as deep as7.6 mm. The detectability of the flaw at this depth largelydepends on the defect geometry and the sample configuration.

References

[1] German R M 1994Powder Metallurgy Science(Princeton,NJ: Metal Powder Industries Federation)

[2] Klar E (ed) 1983Powder Metallurgy: Applications,Advantages and Limitations(Metals Park, OH: AmericanSociety for Metals)

[3] Zenger D and Cai H 1996The Common Cracks in Green P/MCompacts Handbook(Worcester, MA: Powder MetallurgyResearch Center, Worcester Polytechnic Institute)

[4] Ludwig R, Zenger D and Zhang R 1995 Review of variousnon-destructive testing methods for detecting cracks ingreen P/M componentsReport(Worcester, MA: CarlGunnard Johnson Powder Metallurgy Research Center,Worcester Polytechnic Institute)

[5] Lewis A and Bush D 1991 Resistivity measurement forevaluation of coating thicknessMater. Evaluation491231–9

[6] Frise P R and Bell R 1992 Improved probe array for ACPDcrack measurementsBr. J. Non-dest. Testing3415–19

[7] Berry M D, Brisker H C and Cuneo A A 1989USA Patent4888546

[8] Witt G C 1992USA Patent5136252[9] Runyan W R 1975Semiconductor Measurements and

Instrumentation(New York: McGraw-Hill)[10] Ludwig R, Makarov S and Apelian D 1998 Theoretical and

practical investigations of electric resistivity testing ofgreen-state P/M compactsJ. Nondest. Evaluation17153–66

[11] Stander J G 1996 A non-destructive evaluation method for‘green’ powder metal parts using electrical impedanceM.S. Thesis.(Worcester, MA: Department of Electricaland Computer Engineering, Worcester PolytechnicInstitute)

[12] Bogdanov G 1999 Theoretical basis and practicalimplementation of electrical impedance materialinspection of powder metallurgy compactsM.S. Thesis(Worcester, MA: Department of Electrical and ComputerEngineering, Worcester Polytechnic Institute)

[13] Ikeda K, Yoshimi M and Miki C 1991 Electric potential dropmethod for evaluating crack depthInt. J. Fracture4725–38

166