WO2024005797A1

WO2024005797A1 - A system on a chip comprising a diagnostics module

Info

Publication number: WO2024005797A1
Application number: PCT/US2022/035387
Authority: WO
Inventors: Peter Robertson
Original assignee: Siemens Industry Software Inc.
Priority date: 2022-06-28
Filing date: 2022-06-28
Publication date: 2024-01-04

Abstract

A system on a chip (SoC) (10) is provided. The SoC (10) includes an integrated circuit (12), an embedded analytics monitor (14) configured to generate analytics data by monitoring one or more interactions within the integrated circuit (12), and a diagnostics module (16) including a machine learning algorithm trained to detect an anomaly in the analytics data indicative of a fault condition of the integrated circuit (12).

Description

A SYSTEM ON A CHIP COMPRISING A DIAGNOSTICS MODULE

FIELD

[0001] The subject-matter of the present disclosure relates to a system on a chip (SoC) including a diagnostics module, a computer-implemented method of detecting a fault condition of an integrated circuit of a SoC, a computer- implemented method of training a machine learning algorithm to detect an anomaly in analytics data indicative of a fault condition of an integrated circuit of a SoC, and a computer readable medium.

BACKGROUND

[0002] A device including a SoC may be subject to anomalous behaviour either from an external source or tampering with the device itself.

[0003] Anomalous activities are inherently difficult to identify, as anomalous activities are sparse, unpredictable, and include any behaviour that is not typical of the device so are difficult to quantify. Typically, this would require data to be collected and exported off chip to be examined by an individual with significant insight into the normal working behaviour of the chip and prior knowledge that an anomalous event occurred with a rough estimate of the time at which is occurred. The collection of this data may not be possible in real world applications, as this would require knowledge of ongoing anomalous behaviour (e.g., observation of the ongoing anomalous behaviour and collection of data of the ongoing anomalous behaviour). Even if such data could be obtained, this analysis remains unable to react and prevent further unexpected behaviour on the device, as the analysis is happening off chip after the event in question.

[0004] To prevent these anomalies from causing harm to a system, the anomalies are to be identified at the time of occurrence, allowing preventative measures to be taken or real-time notifications to be made. SUMMARY

[0005] It is an aim of the present disclosure to address the issues described above and improve on the prior art.

[0006] According to an aspect of the present invention, a system on a chip (SoC) includes an integrated circuit, an embedded analytics monitor configured to generate analytics data by monitoring one or more interactions within the integrated circuit, and a diagnostics module including a machine learning algorithm trained to detect an anomaly in the analytics data indicative of a fault condition of the integrated circuit.

[0007] In an embodiment, the machine learning model may be trained to map analytics data values to a distribution.

[0008] In an embodiment, the distribution may be a Gaussian distribution.

[0009] In an embodiment, a mean of the distribution may be 0.5.

[0010] In an embodiment, the machine learning model may be configured to map a non-anomalous analytics data value to within a threshold of variation of the mean of the distribution, and may be configured to map an anomalous analytics data value to outside the threshold of variation of the mean of the distribution.

[0011] In an embodiment, the threshold may be one or more standard deviations.

[0012] In an embodiment, the machine learning model may be a critic from a generative adversarial network.

[0013] In an embodiment, the machine learning model may be an encoder from a variational autoencoder.

[0014] According to an aspect of the present invention, a computer- implemented method of detecting a fault condition of an integrated circuit of a system on a chip (SoC) includes generating analytics data by monitoring one or more interactions within the integrated circuit, and detecting, by a machine learning algorithm, an anomaly in the analytics data indicative of a fault condition of the integrated circuit.

[0015] In an embodiment, the machine learning model may be trained to map analytics data values to a distribution.

[0016] In an embodiment, the distribution may be a Gaussian distribution.

[0017] In an embodiment, a mean of the distribution may be 0.5.

[0018] In an embodiment, the computer-implemented method may further include mapping, by the machine learning model, a non-anomalous analytics data value to within a threshold of variation of the mean of the distribution, and mapping, by the machine learning model, an anomalous analytics data value to outside the threshold of variation of the mean of the distribution.

[0019] In an embodiment, the threshold may be one or more standard deviations.

[0020] In an embodiment, the machine learning model may be a critic from a generative adversarial network.

[0021] In an embodiment, the machine learning model may be an encoder from a variational autoencoder.

[0022] According to an aspect of the present invention, a computer- implemented method of training a machine learning algorithm to detect an anomaly in analytics data, generated by an embedded analytics monitor, indicative of a fault condition of an integrated circuit of a system on a chip (SoC) includes providing a plurality of analytics data values from the embedded analytics monitor of the SoC. The plurality of data values are obtained by monitoring interactions within the integrated circuit during normal, non- anomalous, operation. The computer-implemented method includes mapping the analytics data values to a distribution, determining an error between the distribution and a desired distribution, and modifying the machine learning algorithm to reduce the error.

[0023] In an embodiment, the desired distribution may be a Gaussian distribution. [0024] In an embodiment, a mean of the distribution may be 0.5.

[0025] In an embodiment, the machine learning model may include a critic of a generative adversarial network.

[0026] In an embodiment, the machine learning model may include a variational autoencoder.

[0027] According to an aspect of the present invention, a transitory, or non- transitory, computer-readable medium has instructions stored thereon that, when executed by a processor, cause the processor to perform the method as described above.

BRIEF DESCRIPTION OF DRAWINGS

[0028] The embodiments of the present inventions may be best understood with reference to the accompanying figures, in which:

[0029] Figure 1 shows a schematic block diagram of a System on a Chip, SoC, according to an embodiment of the present invention;

[0030] Figure 2 shows a schematic block diagram of a machine learning model from a diagnostics module from the SoC of Figure 1 embodied as an encoder of a variational autoencoder;

[0031] Figure 3 shows a flow chart of a computer-implemented method of training the machine learning model of Figure 2 and of Figure 4;

[0032] Figure 4 shows a schematic block diagram of a different machine learning model compared to Figure 2, embodied as a critic of a generative adversarial network; and

[0033] Figure 5 shows a flow chart of a computer-implemented method of detecting a fault condition of an integrated circuit of the SoC from Figure 1 . DESCRIPTION OF EMBODIMENTS

[0034] Any methods described herein may be computer-implemented methods. The computer-implemented methods may be provided on a transitory, or non-transitory, computer-readable medium having instructions stored thereon that, when executed by a processor, cause the processor to perform the method. The processor may be a processor of a computer that also includes storage. The non-transitory computer readable medium may be store in the storage.

[0035] With reference to Figure 1 , a SoC 10 includes an integrated circuit 12, an embedded analytics monitor 14, and a diagnostics module 16.

[0036] The SoC 10 may be any of a variety of different chips for myriad different devices. The SoC 10 may be a traffic bus, for example. A traffic bus, however, is described here for illustrative purposes only and should not be construed as the only type of SoC to which the present subject-matter is applicable.

[0037] The integrated circuit 12 may be a conventional integrated circuit where interactions between circuit components occur for the integrated circuit to perform various operations. In the case of the SoC 10 being a traffic bus, the interactions may include read and/or write interactions.

[0038] The embedded analytics monitor 14 is independent of the integrated circuit. For example, operating the embedded analytics monitor 14 does not impact on operation of the integrated circuit 12, or vice-versa.

[0039] The embedded analytics monitor 14 monitors the interactions within the integrated circuit 12. The embedded analytics monitor 14 operates in real time. The embedded analytics monitor 14 monitors the interactions of the integrated circuit to generate analytics data by sampling data values associated with the interactions over a period of time.

[0040] The diagnostics module 16 includes a machine learning algorithm. The machine learning algorithm is trained to detect an anomaly in the analytics data indicative of a fault condition of the integrated circuit 12. An anomaly in the illustrative example of a traffic bus may be detecting read data when write data is expected, or vice-versa.

[0041] With reference to Figure 2, in an embodiment, the machine learning algorithm is an encoder 30 from a variational autoencoder (VAE).

[0042] The encoder 30 is trained to map the analytics data values 23 to a distribution 32. The analytics data values 23 themselves may be viewed as a data sample from an unknown distribution to which these samples belong.

[0043] The distribution may be a Gaussian distribution. The distribution may, for example, have a mean of 0.5 (when the range of probability values are between 0 and 1 ). When the analytics data 23 is non-anomalous (e.g., when there are no errors in the analytics data due to the integrated circuit operating normally), the encoder 30 may be configured to map the analytics data to within a threshold of variation of the mean of the distribution. The threshold may be one or more standard deviations. In contrast, when the analytics data 23 is anomalous (e.g., the integrated circuit is not operating normally), the encoder 30 will map the analytics data 23 to outside the threshold variation of the mean of the distribution. In other words, the analytics data 23 will not be mapped to the Gaussian distribution.

[0044] With reference to Figure 3, the machine learning model of the embodiment of Figure 2 may be trained using a computer-implemented method such that the machine learning model detects an anomaly in analytics data, generated by an embedded analytics monitor, indicative of a fault condition of an integrated circuit of a system on a chip (SoC). The method includes providing S100 a plurality of analytics data values from the embedded analytics monitor of the SoC 10. The plurality of data values are obtained by monitoring interactions within the integrated circuit 12 during normal, non-anomalous, operation. The computer-implemented method also includes mapping S102 the analytics data values to a distribution. The computer-implemented method also includes determining S104 an error between the distribution and a desired distribution. The computer-implemented method also includes modifying S106 the machine learning algorithm to reduce the error. The desired distribution may be the Gaussian distribution, having a mean of 0.5.

[0045] By reducing the error, the non-anomalous embedded analytics data may be mapped to a Gaussian distribution for ease of comparison with anomalous data.

[0046] The decoder of the VAE is also used for training but is then discarded. Only the encoder of the VAE is to be used for inference. For example, the training method may include mapping the input analytics data to the distribution using the encoder. The method then includes decoding, using the decoder, the distribution to obtain decoded analytics data. The decoded analytics data is compared with the input analytics data, and an error is calculated. A further error is calculated between the distribution and the desired, Gaussian, distribution. Once both errors are minimised, the VAE is considered to be trained. The encoder is extracted, and the decoder is discarded.

[0047] With reference to Figure 4, in an embodiment, a machine learning algorithm in the form of a critic 40 from a generative adversarial network 42 (GAN) is provided. For the avoidance of doubt, a critic is distinguished from a discriminator in that a discriminator generates a binary output of either “real” or “fake,” whereas a critic provides a probability between 0 and 1 , for example.

[0048] The GAN 42 includes the critic 40 and a generator 44. The generator 44 receives a distribution 46 that matches the embedded analytics data 23 as an input. The generator 44 is configured to generate data based on the distribution 46. The generated data is intended to be of the same form as embedded analytics data 23 from an embedded analytics monitor. The critic 40 is configured to generate an output probability 48. When training the critic 40, the goal is such that the critic cannot determine if the embedded analytics data is distinguishable. In such cases, the critic will always output a value of 0.5. If the critic is able to determine whether the generated data is distinguishable from the embedded analytics data, the output probability will not be 0.5, or around 0.5. Instead, the output probability may tend to be 0 or 1 , for example. [0049] The generator 44 may be discarded because only the critic is required by the present embodiment. The generator 44 is required to train the critic rather than at inference time.

[0050] The critic 40, broadly speaking, may be trained according to the computer-implemented method from Figure 5.

[0051] With reference to Figure 5, the foregoing embodiments may be captured, broadly speaking, as a computer-implemented method of detecting a fault condition of an integrated circuit of a system on a chip (SoC). The method may include generating S200 analytics data by monitoring one or more interactions within the integrated circuit 12, and detecting S202, by a machine learning algorithm, an anomaly in the analytics data indicative of a fault condition of the integrated circuit 12.

[0052] Using the diagnostics module according to the foregoing embodiments provides that anomalies may be detected on the SoC itself rather than having to take the embedded analytics data offline. This provides that the anomalies may be flagged and potentially resolved in real-time. Such teal-time alerting/flagging and potential resolution is particularly important for systems such as autonomous vehicles.

[0053] Attention is directed to all papers and documents that are filed concurrently with or previous to this specification in connection with this application and that are open to public inspection with this specification, and the contents of all such papers and documents are incorporated herein by reference.

[0054] All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and/or all of the steps of any method or process so disclosed, may be combined in any combination, except combinations where at least some of such features and/or steps are mutually exclusive.

[0055] Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent, or similar purpose, unless expressly stated otherwise. Thus, unless expressly stated otherwise, each feature disclosed is one example only of a generic series of equivalent or similar features.

[0056] The invention is not restricted to the details of the foregoing embodiment(s). The invention extends to any novel one, or any novel combination, of the features disclosed in this specification (including any accompanying claims, abstract and drawings), or to any novel one, or any novel combination, of the steps of any method or process so disclosed.

Claims

1 . A system on a chip (SoC) (10) comprising: an integrated circuit (12); an embedded analytics monitor (14) configured to generate analytics data by monitoring one or more interactions within the integrated circuit (12); and a diagnostics module (16) comprising a machine learning algorithm trained to detect an anomaly in the analytics data indicative of a fault condition of the integrated circuit (12).

2. The SoC of Claim 1 , wherein the machine learning model is trained to map analytics data values to a distribution.

3. The SoC of Claim 2, wherein the distribution is a Gaussian distribution.

4. The SoC of Claim 2 or Claim 3, wherein a mean of the distribution is 0.5.

5. The SoC of any of Claims 2 to 4, wherein the machine learning model is configured to map a non-anomalous analytics data value to within a threshold of variation of the mean of the distribution, and is configured to map an anomalous analytics data value to outside the threshold of variation of the mean of the distribution.

6. The SoC of Claim 5, wherein the threshold is one or more standard deviations.

7. The SoC of any of Claims 2 to 6, wherein the machine learning model is a critic from a generative adversarial network.

8. The SoC of any of Claims 2 to 6, wherein the machine learning model is an encoder from a variational autoencoder.

9. A computer-implemented method of detecting a fault condition of an integrated circuit of a system on a chip (SoC), the computer-implemented method comprising: generating (S100) analytics data by monitoring one or more interactions within the integrated circuit (12); and detecting (S102), by a machine learning algorithm, an anomaly in the analytics data indicative of a fault condition of the integrated circuit (12).

10. A computer-implemented method of training a machine learning algorithm to detect an anomaly in analytics data, generated by an embedded analytics monitor, indicative of a fault condition of an integrated circuit of a system on a chip (SoC) (10), the computer-implemented method comprising: providing a plurality of analytics data values from the embedded analytics monitor of the SoC (10), the plurality of data values obtained by monitoring interactions within the integrated circuit (12) during normal, non-anomalous, operation; mapping the analytics data values to a distribution; determining an error between the distribution and a desired distribution; and modifying the machine learning algorithm to reduce the error.

11. The computer-implemented method of Claim 10, wherein the desired distribution is a Gaussian distribution.

12. The computer-implemented method of Claim 10 or Claim 11 , wherein a mean of the distribution is 0.5.

13. The computer-implemented method of any of Claims 10 to 12, wherein the machine learning algorithm comprises a critic of a generative adversarial network.

14. The computer-implemented method of any of Claims 10 to 12, wherein the machine learning algorithm comprises a variational autoencoder.

15. A transitory, or non-transitory, computer-readable medium having instructions stored thereon that when executed by a processor, cause the processor to perform the method of any of Claims 9 to 14.