Devices and methods for the detection of counterfeit and defective electronic components using machine learning

Abstract

Counterfeit microelectronics have been estimated to cause more than $250 billion dollars; of damage1 annually to commercial and government entities, including threatening national security. Traditional methods of detection are not sufficient as they take too long, are expensive, are destructive to the tested devices or are not flexible enough for many device types. This technology addresses and detects recycled and remarked counterfeits without the need for destructive tests. Optical analysis when coupled with other sensor data cart reveal anomalies between an authentic and a counterfeit device in a quick handheld test without destroying the device under test (DUT). This data can be analyzed by several machine learning algorithms to indicate whether a component is considered counterfeit.

Description

[0001] nil E

[0002] Devices and Methods for the Detection of Counterfeit and Defective Electronic Components Using Machine Learning

[0003] FIELD OF THE IN VEN' T ION

[0004] The invention pertains to electronic components and methods for identification of counterfeit electronic components and for confirmation that electronic components are authentic. More specifically, the invention pertains to authentication of electronic components.

[0005] BACKGROUND

[0006] There is a need for electronics manufacturers and electronic product manufacturers to have methods and devices capable of detecting whether an electronic component is authentic, counterfeit. and / or defective.

[0007] Counterfeit electronics are a large and growing problem in global supply chains for electronics manufacturers. A rnetastudy found that consumer and industrial businesses are losing approximately $250 billion each year due to the purchase of counterfeit electronic components including computer chips. After manufacturing and incorporation into downstream products, counterfeit or defective chips can cause the entire product to fail, resulting in safety issues and / or loss of value. In addition, experts reported that -45% of the spare and replacement parts the Pentagon bought in the market were counterfeit, a major threat to national security. However, existing anti-counterfeiting methods are unpractical. Current techniques are based on either slow electrical testing, subject mater expert (SME) analysis or expensive large equipment (e.g., X-Rays).

[0008] Current industry methods for determining whether art electronic component is authentic involve a costly, labor intensive procedure that has to be performed by highly trained counterfeit detection technicians on expensive equipment, often destroying the device under test (DUT) in the process.

[0009] These conventional methods have several issues. Firstly, as there are differences in how each technician, interprets the test results and there is discretion in accepting or rejecting a chip or electronic component, the process is subject io human error. Secondly, these methods take significant time to ship the components to specialized centers. Counterfeit detection labs test components on many pieces of equipment valued in the tens of thousands of dollars. These conventional methods cost up to $5,000 per tested batch because of the time spent by trained lab workers and the use of expensive equipment Finally, only a small number ofchips can be tested and manv of those that are tested are destroyed in the testing process. This means that if a given set of components tad only 10% counterfeits mixed, it is unlikely they could be identified.

[0010] SUMMARY

[0011] This invention rises a visual and mass-based inspection and an electronic assessment to make a digital fingerprint with hundreds of individual data points of the characteristics of an authentic electronic component or chip. As used herein, drip and electronic component are used interchangeably.

[0012] Embodiments of the device and method may then automatically identify if a component is in a tolerance range of the digital fingerprint and rejects or accepts the component. The cost of our solution will be in the hundreds of dollars range as compared to price tags of greater than $25K for some counterfeit detection machines.

[0013] in an embodiment of our machine learning solution, the method comprises moving images through a segmentation model as an initial processing step to determine more about components before putting them through our primary classifier. Second, embodiments of the method may comprise using additional sensor inputs like spectral analysis and an ambient light sensor to find tricky features such as engraving depth and exact color. And third, embodiments of the method may comprise using a combination of a supervised binary classifier and an unsupervised anomaly detection model so the device and method can make classifications of “Authentic” on electronic components without having trained the data on anything other than software -generated counterfeits.

[0014] DESCRIPTION OF THE FIGURES

[0015] Figure 1 shows a possible flow of machine learning models of the device. (A) is the physical device and measurement step. (B) and (C) am the image normalization and image segmentation steps respectively. These will process the images to generate features. The output of this step is a normalized image with the pins, labels and body of a component segmented. Feature generation will be processing of the raw data into information for the models (D) and (E) is a supervised model that will determine if rite chip is Authentic or Not Authentic. Next, the device and method may send the normalized and segmented images into the unsupervised anomaly detection model (F). (G) is the final step, in which we will assess the probabilities of both models and apply a classification threshold to create a class label of Authentic or Not Authentic; Figure 2 depicts a prototype of the device. There is a camera mounted on the top of the enclosure. There is lighting throughout the enclosure inside. There is a control board to the rigid of the enclosure that handles interaction with the user. There is also a scale at the bottom of the enclosure that has a load cell exposed to the device; and

[0016] Figure 3 shows the electrical measurement interface. There is a lever arm that presses the chip down into a matrix of connecting pads. These sit on a padded platform. Each pad is connected with a wire to the control board.

[0017] DESCRIPTION

[0018] Embodiments of the devices and methods for the detection of counterfeit and defective electronic components using machine learning may comprise some or all of the following steps in any order:

[0019] 1. Component enters or is placed in the machine.

[0020] 2. Operator enters any metadata or scans a code to enter the metadata into the device memory or cloud database about the test run including at least one of a lot number, lot size, invoice number, date, condition, operator name or other items, for example.

[0021] 3. Operator initiates the test.

[0022] 4. The device may coordinate lighting, engage a digital imaging device to produce a raw image file, and a scale to measure the weight of the component The machine may also be configured to measure proximity (via an IR proximity sensor) and spectrum analysis for color (via an ambient light and spectral analysis chip).

[0023] 5. The device may send data from the tested component to a processing unit in the cloud, other networked computer, or use an internal processing unit to compare the measured data to that of a known good component or compare the data in its memory,

[0024] a. The software may perform several intermediary functions, if desired, including but not limited to:

[0025] i. Raw image processing - e.g., crop, rotate, as well as other processes, if applicable;

[0026] ii. Segmentation of the chip image into elements including, but not limited to, pins, background, package, and text via a machine learning model, for example; and / or

[0027] iii. Optical character recognition (OCR) on the label of the chip. As described above, the processing unit is in communication with a database of chip, component or board infonnation. The chip, component, or board information includes information that defines the authenticity and functionality. This information may include, but is not limited to, at least one of a chip identification code of the authentic electronic component, a weight range of the authentic electronic component identified by the chip identification code, a color range of the authentic electronic component identified by the chip identification code, label information of the authentic electronic component

[0028] identified by file chip identification code, label placement coordinates of the label on the authentic

[0029] electronic component identified by the chip identification code, a size of the authentic electronic

[0030] component identified by the chip identification code, a size of various elements of the authentic electronic component identified by the chip identification code, for example.

[0031] The chip, component, or board information may further include, but are not limited to, various elements that define the chip, component, or board a length, width, and height of a body, the label,

[0032] characters of the label, a depth of the characters of the label, pins, solder residue, or other features of the authentic electronic component.

[0033] 6. Embodiments of the device and method may use the following validator information and methods in this order to identify if the chip is genuine:

[0034]

[0035]

[0036] 2. If enough validators fail (are not within defined specifications or tolerances for the authentic component) or specific validators fail, then the chip or other component will be rejected, and an indicator will alert an operator or take some other action to indicate that the chip or other electronic component is defective or counterfeit, if all validators are considered to be within tolerances or otherwise successful, then the chip wifi be accepted, and an approval of She chip or other electronic component will be indicated. 8. The metadata and / or measured information of the chip may be posted to the cloud or a database on the device or other remote location and used in a variety of ways including, but not limited to:

[0037] 1. Creating a “fingerprint” including specifications and tolerances of the data of a known “good” or authentic chip;

[0038] 2. The approved chips may be automatically posted to a marketplace for resale;

[0039] 3. Recording serial numbers of the failed and authenticated chips or components in a database;

[0040] 4. Creating, maintaining and revising metrics on inbound quality of chip supply for each component or ehip;

[0041] 5. Adding chip data to a database which may be used to improve validation procedures; 6. Developing; maintaining; and improving a “trust score” for the supplier of the chips; and 7. Determining the value and condition of the chips based upon prior sales, current market; and / or condition of' he chip or component.

[0042] Machine Learning Steps

[0043] The device and method may include macltine learning based upon the information gathered during testing. For example, for each difference (measured value compared to the specifications and tolerances), the method may compare the value universally (e.g., across many chips) or on a SKU-basis (meaning that comparing with only the same part number by the same manufacturer), or both. For example, bent and corroded pins may be determined based upon and identifiable by a single- machine learning algorithm across multiple SKUs having similar pins designs.

[0044] Label locations, on the other hand, may be located at various locations on different chips based upon chip or components size and manufacturer. Labels may be in the same location only for the same SKU or across a few SKUs. The device and method may use the universal method to determine Authentic or Not Authentic score for common features or differences, this score will be more accurate because the device and method may use a much larger data set than just the values available for a single SKU. For SKU-based differences, the device and method may generate a comparison model for each SKU (a unique manufacturer, part continuation).

[0045] In order to capture the differences above, embodiments of the method may comprise (one or more of the following steps as depicted in Figure I) obtaining an image, processing the image by measuring 1 portions of the image such as the chip to be assessed, proximity (via an IR. proximity sensor), and spectrum analysis for color (via an ambient light and spectral analysis chip, for example). Embodiments 3 of the device may comprise a camera, camera mounting system, and camera control module configured to 4 take basic stationary images at the right resolution and right lighting source for single chips (A), The 5 device and method may do the image normalization (B) and image segmentation model (C) that uses a U- 6 Net convolutional neural network, for example. The output of this step is a normalized image with the 'I pins, labels and body of an elec-tonic component segmented.

[0046] 8 Next, foe method may comprise generating features using combinations of the sensor inputs and 9 feeding specifications of the features into the supervised binary classifier (D and E). Next, the method 0 may comprise sending the normalized and segmented images into the unsupervised anomaly detection model (F). G may be the final step, in which the method comprises assessing the probabilities of both 2 models and applying a classification threshold to create a class label of Authentic or Not Authentic. 3

[0047] 4 Model E:

[0048] 5 - Type: Supervised binary classifier model where the target class is “Authentic” Returns: Likelihood that 6 an unknown sample is included in foe “Authentic” class.

[0049] 7 Algorithms that may be used, in this step, for example:

[0050] 8® A Logistic Regression model with LI regularization

[0051] 9® Support Vector Machine

[0052] 0® XGBocst

[0053] 1 ® Random Forest (RndF)

[0054] 2® An anomaly detection algorithm as described by Model F

[0055] 3 Model F:

[0056] 4 - Type: Anomaly detection model with input of the segmented and raw images- Returns: Anomaly 5 detection likelihood

[0057] 6 Algorithms that may be used in this step, for example:

[0058] 7® OneClassSVM

[0059] 8® Robust Random Cut Forest

[0060] 9® isolation forest The overall verification model and method will return an indication of False if' either the binary classifier or the anomaly detection describes an anomalous condition. The best combination of models may be assessed by the one with the maximum area under the precision-recall curve (PR AUC), for example.

[0061] All methods and devices may also be configured and used to assess a PCBA or printed circuit board assembly to determine if it is acceptable. For example, the device and method may assess whether either 1) the overall structure of the board is correct or incorrect (e.g., placement of components, overall design, soldering, etc,) or 2) a specific chip on the board is counterfeit bat the overall structure of the board is intact. Tims, the method may perform all checks above at the PCB level. In another embodiment, the method comprises subdividing the board or an image tile of the board into chips and putting each ehip image through the chip authenticity algorithm as described herein to determine its authenticity.

[0062] Relationship Between Components

[0063] Additional embodiments of the method and the device may be used to determine whether a component is authentic and / or functional. Alternatively, the method and device may be used for a calibration ran which will capture data about a chip that is known to be valid, to this case, steps 7.1 to 7.13 will be the same, hut then the machine will only indicate that the process is complete and proceed to process its data instead of doing any validation steps (8.1 only), since the chip, component, or board is known to be authentic and fimctionak

[0064] The steps may be ordered to have the fastest and most definitive tests first, for example. This will improve the overall speed as the device and the method may make a determination that the chip, component, or board is most likely to be counterfeit or defective at any step after foe chip or electronic component. For example, the weight test, for some components, may be toe most effective; heat damage, plastic type discrepancies and most functional chip differences will alter the weight of the component and cause failure of this test concl sively.

[0065] Therefore, in some embodiments, a failure of the weight comparison test (step 7.1) is foe simplest method of ascertaining whether a chip is not genuine or not functional. Weighing the component and comparing the weight to the weight range of an authentic functioning component is the least processing-mtensive, fastest, and also quite effective. If that method fails to determine that the chip or component is out of specification, the visual tests steps 7.2-7.5 may be performed as the next steps in the method. These image processing steps may include a machine learning algorithm to run to segment the 2 image file to make the image processing steps much more effective. These image processing steps are 3 effective in finding counterfeits that have been created via blacktcppiug or printing of poor quality. 4 However, visual methods cannot identity all defective authentic chips.

[0066] 5 Therefore, the method or machine may include checks for pin damage or solder residue and, in 6 some embodiments, the device or method may include a test to compare the text on the label to the 7 expected text stored in the database.

[0067] 8 Finally, an electrical test: may be performed, At any point if a step fails, then die chip will he 9 rejected. The electrical step (7.10) may he the most complicated step. The orientation of the chip may be 10 visually mapped by using either the position of the positioning dot on the package or the lettering on the device. The algorithm identifies which of the pins of the component are attributed and / or electronically 12 connected to each of the pins of the test rig bed. If any of the pins are misaligned such that the test may 13 not be completed, the operator is alerted to rotate the chip to orient it correctly or the device may 14 automatically align the chip with a matrix of pin connectors. The device may then apply a voltage to one 15 or more pins of the chip according to a series of voltages determined earlier in the calibration of the good 16 chip. For example, the good chip might have received the following where I indicates high voltage and 0 indicates low voltage.

[0068] 18 ® Series 1 input, Pin I [1, 0, 0, 1 ], Pin 2 [1, 1, 0, 01

[0069] 19 * Series 2 input, Pin 1 [1, I, I, 1], Pin 2 [I, 1, I, 0]

[0070] 20 It then may have exhibited the following: expected output:

[0071] 21 • Series I output, Pin 3 [1. 1, 0, 1 ], Pin 4 [0, 0, 0, 01

[0072] 22 ® Series 2 output. Pin 3 [1, 0, 0, 1], Pin 4 [0, 0, 0, 0]

[0073] 23 In this case, when testing the questionable chip, the machine will replicate the Scries 1 and 2 24 inputs and listen tor the Series 1 and 2, expected output. If there are delays in timing, variations in 25 response or unexpected additional responses, the validation will fail. If ail the output is as expected, the 26 chip will pass. 1 here may be many more input voltages included in each series and many more series of '>7 inputs included than the example above, The expected output may he on a chip by chip basis or on 28 similarly related chips.

[0074] 29 The calibration step for determining whether or not a component is authentic and functional 30 involves Identifying the correct set of input and output voltage series to use in validation. When a chip is 31 entered for calibration, the machine may test multiple input and output pairings or series to determine authenticity and functionality. For example, some pins are power or output pins and, therefore, will not respond to input voltages. This means that feeding an input voltage to these pins will not generate meaningful output voltages. In the calibration step, the machine will pass in many more series of inputs than are needed in the final validation in order to find sequences that characterize the function of' the chip and are non-trivial. It will store the final sequences in a database.

[0075] The inputfoutput pairs to use on a component may be a secret key used for authentication. For example, the manufacturer may provide an '"‘electronic hologram” of sorts to aid the validation of the chip. The masjufacturer may program in a specific series of input voltages that trigger an unexpected output voltage series that may not be related to the function of the chip but a specific output may be used to validate it.

[0076] Embodiments of the method and device are meant to test electronic components in environments such as integrated circuits, passive components, and other embodiments of the device may be configured and used for any electronic component or chip.

[0077] Embodiments inchide a device or a machine that executes the above methods and steps automatically, a piece of software that runs the tests as above and an overall methodology for automatically checking authenticity and function of electronic components.

[0078] The machine is composed of several items or some of the following components: a camera or other digital imaging device, such as a digital camera with a macro lens, an enclosure that standardizes the lighting of the chip or electronic component, a series of individually addressable LEDs or sets of lights that can be turned on and oil its sequence to generate ditlsrent lighting conditions, a scale that is sufficiently accurate to distinguish authentic, functioning components from defective or counterfeit. components, such as a scale accurate to 0.0001 grams, for example, an electrical contact matrix or set of contact pins, a rig arm, and a control board that relays information to its storage and / or processing service.

[0079] The software is composed of three intermediary functions and 13 validator functions as above. Other validator tests can be added, such as fraud checks on the supplier of the components, data on its provenance, additional electrical function tests or validation of the serial number with external parties.

[0080] The test steps can be performed and will work in any order as long as the intermediary steps are completed. The tests can be partially completed as well.

[0081] %

Claims

CLAIMS1. A device for authenticating an electronic component, comprising:at least one digital imaging device, wherein the digital imaging device outputs digital image data; at least one light for illuminating a staging area within the device, wherein the staging area is within a field of view of at least one of the at least one digital imaging device;a scale with a digital output; anda central processing unit in electrical communication with the at least one digital imaging device and the scale.

2. The device of claim 1, wherein the at least one light is a plurality of lights.

3. The device of claim 2, wherein the plurality of lights comprises at least two sets of individually controlled lights.

4. The device of claim 3, wherein the set of individually controlled lights are controlled by the central processing unit.

5. The device of claim 4, wherein the individually controlled lights are controlled during an imaging process of the digital imaging device.

6. Hie device of claim I, wherein the scale is capable of measuring the weight reads the weight of an object placed.

7. The device of claim 6, wherein the scale is accurate to 0.0001 grams.

8. The device of claim 1, wherein the central processing unit is in communication with a database of chip information,9. The device of claim 8, wherein the chip information comprises at least one of a chip identification code of the authentic electronic component, a weight range of the authentic electronic component identified by the chip identification code, a color range of the authentic electronic component identified by the chip identification code, label information of foe authentic electronic component identified by the chip identification code, label placement coordinates of the label on the authentic electronic component identified by the chip identification code, a size of the authentic electronic component identified by the chip identification code, a size of various elements of the authentic electronic component identified by the chip identification code.

10. The device of claim 1, wherein the various elements may include a length, width, and height of a body, the label, characters of the label, a depth of the characters of the label, pins, solder residue, or other features of the authentic electronic component11. The device of claim 1, comprising a matrix of pin connections.

12. The device of claim 11, wherein each of the pin connections is individually connected to the central processing unit.

13. The device of claim 12, wherein at least a portion of the pin connectors receive an input signal from the central processing unit or a signal generator.

14. The device of claim 13, wherein at least a portion of the pin connectors output a signal to the central processing unit.

15. The device of claim 11, comprising a platform or other member to press an electronic component to the matrix to provide electrical communication between at least one pin of the electronic component to a pin connection.

16. The device of claim 15, wherein the platform or other member is pivoted by a lever arm.

17. A method of determining the authenticity of an electronic component, comprising: weighing the electronic components to determine a weight of the electronic component; and comparing the weight of the electronic component to determine whether the weight is within an expected weight range of an authentic electronic component.

18. The method of claim 17, comprising determining an expected weight range of the authentic electronic component by referencing an electronic component identification code from a database.

19. The method of claim 17, comprising imagining the electronic component with a digital imaging device to prepare a raw image file.

20. The method of claim 19, comprising coordinating a series of lights illuminating the electronic component as the digital Imaging device images the electronic component to produce at least one raw image file.

21. The method of claim 17, comprising processing the raw image file to determine segments or elements of the electronic component.

22. The method of claim 17, comprising reimaging the electronic device to prepare a second raw imaging file.

23. The method of claim 19, wherein the elements of the image file may comprise an image of pins, background, package of the electronic component, and label text on the electronic component.

24. The method of claim 19, wherein processing the raw image file comprises a machine learning model.

25. The method of claim 23, wherein the label text on the electronic component is determined by optical character recognition.

26. The method of claim 19, comprising comparing a color of the electronic component determined from the raw image file to a color range of the authentic electronic component referenced with an electronic component identification code from a database.

27. The method of claim 19, comprising comparing a label location of the electronic component determined from the raw image file to a label location of the authentic electronic component referenced with an electronic component identification code from a database.

28. The method of claim 19, comprising comparing a size of the electronic component determined from the raw image file to a size of the authentic electronic component referenced with an electronic component identification code from a database.

29. The method of claim 19, comprising comparing a size of elements of the electronic component determined from the raw image file to a size of the elements of the authentic electronic component referenced with an electronic component identification code from a database.

30. The method of claim 19, comprising comparing a label font of the electronic component determined from the raw image file to a label font of the authentic electronic component referenced with an electronic component identification code from a database.4 31. The method of claim 19, comprising comparing a depth of a label characters of a label of 5 the electronic component determined from the raw image file to a depth of a label characters of a label of 6 the authentic electronic component referenced with an electronic component Identification code from a database.8 32. The method of claim 19, comprising comparing a pin of the electronic component 9 determined from the raw image file to a pin of the authentic electronic component referenced with an 0 electronic component identification code from a databas33. The method of claim 19, determining whether solder is visible on the electronic component fam the raw image file to the authentic electronic.

34. The method of claim 19, comprising comparing label characters of the electronic component determined from the raw image file to label characters of the authentic electronic component referenced with an electronic component identification code from a database.

35. The method of claim 19, comprising creating a database of metadata of an authentic-electronic component from the weight and information from the raw image file.

36. The method of claim 35, comprising determining an authenticity trust score based upon the comparing of the properties and characteristics of the electronic component determined fam the raw image file and die properties and characteristics from the database.

37. The method of claim 17, comprising providing an input into one or more input pins of the electronic component and measuring air output of one or more output pins of the electronic component.

38. The method of claim 37, comprising signaling an operator if the input cannot be provided to the one or more input pins or if the output of the one or more output pins cannot be measured.

39. The method of claim 37, comprising determining if the input cannot be provided due to misalignment of the electronic component by processing the raw image file.

40. The method of claim 38, comprising comparing the output of the output pins with an expected output based upon the input into the input pins.

41. The method of claim 1, comprising comparing a depth of a label characters of a label of the electronic component determined from raw image files with varying shadow depths from angled light to a depth of a label characters of a label of the authentic electronic component referenced with an electronic component identification code from a database.

42. The method of claim 1, comprising comparing a depth of an electronic component determined from raw image files with varying shadow depths from angled light to a depth of an authentic electronic component referenced with an electronic component identification code from a database.

43. The method of claim 19, wherein processing the raw image file comprises a machine learning model that segments the image into components representing pins, package, engraving and other components.

44. The method of claim 17, comprising mapping machine pins to the pins in an image and / or to the pins in the calibration run.

45. The method of claim 17, comprising identifying a subset of input and output voltages that are relevant to the verification of tire components.

46. The method of claim 45, comprising of using a machine learning model to identify input and output relevant pairs.

46. The method of claim 17, comprising of identifying an undisclosed input and output voltage pair established during the manufacture of the chip.

47. The method of claim 17, comprising of identifying a chip without external knowledge of pin mapping, function, input and output voltages or other information, only the physical chip itself.

48. The method of claim 38, comprising creating series of input voltages to test either by randomizing voltage series or using a machine learning model to improve the creation of these voltage series.

47. A device for authenticating an electronic component, comprising:a central processing unit in electrical communication with the at least one digital imaging device and a scale; anda plurality of pin connections.

48. The device of claim 47, wherein each of the pin connections is individually connected to the central processing unit.

49. The device of claim 48, wherein at least a portion of the pin connectors receive an input signal from the central processing unit50. The device of claim 49, wherein at least a portion of the pin connectors output a signal to the central processing unit.

51. The device of claim 47, comprising a platform or other member to press an electronic component to the matrix to provide electrical communication between at least one pin of the electronic component to a pin connection.

52. The device of claim 51, wherein the platform or other member is pivoted by a lever arm.

Images