AUTOMATIC VALIDATION OF SENSOR TRACES USING MACHINE LEARNING

Description

CROSS REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 62 / 982,568, entitled "AUTOMATIC SENSOR TRACE VALIDATION USING MACHINE LEARNING" filed February 27, 2020.TECHNOLOGY FIELD

[0002] The present disclosure relates to automatic validation of a sensor trace signal using machine learning techniques. More particularly, the present disclosure relates to a method and a system for detecting a failure of a device through validation of a sensor trace signal.BACKGROUND

[0003] An immunoassay and clinical chemistry analyzer is an integrated machine, which relies on the reliable performance of each individual component. For instance, an aspirate probe is used to aspirate a certain desired volume of a liquid and dispense it. One or more sensors are often provided to validate whether an expected amount of volume is aspirated. If the validation fails on a regular basis, it may suggest that the aspirate probe malfunctions and requires maintenance.

[0004] A set of rules specifying safety margins are typically provided to automate the validation process. When a trace signal is beyond the safety margins, the task (e.g., aspiration) may fail. However, these sensors can also potentially collect ambient noises, and thus the algorithmic analysis must be invariant to such noises. To deal with variations caused by the noises, rule engines often become increasingly complex, and gradually become difficult to maintain.

[0005] Document US2014 / 276566 A1 discloses a computer-implemented method, wherein a pump is connected to a sensor. A digital vibration signal representative of the sensed vibrations is generated, and a trace status of a first tubing between the vibration generator and the vibration sensor is determined based on the digital vibration signal.

[0006] Document US2018 / 214634 A1 discloses a computer-implemented method for detecting a failure of a medical pump device, wherein the pump is connected to a piezoresistive pressure sensor.SUMMARY

[0007] Embodiments of the present invention address and overcome one or more of the above shortcomings and drawbacks, by providing a method and system of automatically detecting a device failure using sensor trace data.

[0008] The disclosure provides a computer-implemented method according to the wording of claim 1.

[0009] The disclosure further provides a system according to the wording of claim 8.

[0010] The dependent claims define further advantageous embodiments.

[0011] Additional features and advantages of the invention will be made apparent from the following detailed description of illustrative embodiments that proceeds with reference to the accompanying drawings.BRIEF DESCRIPTION OF THE DRAWING FIGURES

[0012] The foregoing and other aspects of the present invention are best understood from the following detailed description when read in connection with the accompanying drawings. For the purpose of illustrating the invention, there is shown in the drawings embodiments that are presently preferred. Included in the drawings are the following Figures: FIG. 1 illustrates a block diagram of the failure automatic detection system, in accordance with an embodiment described herein; FIG. 2 illustrates a block diagram of the failure automatic detection system, in accordance with another embodiment described herein; FIG. 3 illustrates a valid sample trace signal from a sensor connected to an aspirator probe, in accordance with an embodiment described herein; FIG. 4 illustrates a block diagram of a machine learning model, in accordance with an embodiment described herein; FIG. 5A illustrates an original trace signal indicative of a failure, in accordance with an embodiment described herein; FIG. 5B illustrates a reconstructed trace signal, in accordance with an embodiment described herein; FIG. 5C illustrates an original trace signal highlighting a discrepancy between the original trace signal and the reconstructed trace signal, in accordance with an embodiment described herein; FIG. 6 illustrates a flow chart of a method for training the machine learning model, in accordance with an embodiment described herein; FIG. 7 illustrates a flow chart of a method for detecting a failure of a device, in accordance with an embodiment described herein; and FIG. 8 illustrates an example of a parallel processing memory architecture, in which aspects of the illustrative embodiments are implemented. DETAILED DESCRIPTION

[0013] The following disclosure describes the present invention according to several embodiments directed to a method and system of automatically detecting a device failure using sensor trace data. A machine learning model is used to analyze trace signals from one or more sensors connected to a device (e.g., aspirate probe) and determine whether a task (e.g., aspiration) associated with the device is completed as desired. In an embodiment, if the task fails, then the machine learning model can further indicate a cause of the failure. In an embodiment, a deep neural network is used in the machine learning model to reconstruct trace signals, while simultaneously detecting task failures. The deep neural network can highlight one or more portions of the original trace signal which may be indicative of the failure.

[0014] FIG. 1 illustrates a block diagram of a failure automatic detection system 100, in accordance with an embodiment described herein. The failure automatic detection system 100 includes a device 102 (e.g., aspirate probe) for performing a task (e.g., aspiration); a sensor 104 connected to the device 102; and a machine learning model 106 (e.g., deep neural network) trained to identify a failure of the task, and a cause of the failure. The sensor 104 monitors the task performed by the device 102 and generates a trace signal 108, which is input into the machine learning model 106. The machine learning model 106 has been trained by a large number of trace signals including valid trace signals indicative of successful tasks and trace signals indicative of failed tasks. Thus, the trained machine learning model 106 can identify a trace signal 108 indicative of a failed task. In some embodiments, information is provided with the failed trace signals 108 indicating the cause of failure. This information is used to train the machine learning model 106 to further provide a category of the failure (i.e., cause of the failure) when a trace signal 108 is classified as corresponding to a failed task. In an embodiment, the trained machine learning model 106 can identify whether the trace signal 108 is valid or invalid (i.e., failure) 114. In another embodiment, the trained machine learning model 106 can further output a reconstructed trace signal 110. If the reconstructed trace signal 110 is different from the trace signal 108, the trained machine learning model 106 can further create a display of the two trace signals that highlight regions where one or more discrepancies 112 reside. In another embodiment, the trained machine learning model 106 can further output an estimated parameter 116 associated with the device 102, to further support the valid or invalid identification 114. The parameter 116 can be any parameter associated with the function of the device 102. For example, if the device is an aspirate probe, the parameter 116 can be a volume aspirated by the aspirate probe, or the number of aspiration operations, etc.

[0015] In general, the sensor 104 can be any sensor which can monitor a task performed by the device 102, e.g., a pressure sensor, a temperature sensor, or a voltage sensor, etc. A trace signal 108 is output from the sensor 104 and input into the machine learning model 106.

[0016] FIG. 2 illustrates a block diagram of the failure automatic detection system 200, in accordance with another embodiment described herein. The failure automatic detection system 200 includes an aspirate probe 202 for performing an aspiration task; a pressure sensor 204 connected to the aspirate probe 202; and a machine learning model 106 (e.g., deep neural network) trained to identify a failure of the aspiration, and a cause of the failure. The pressure sensor 204 monitors aspiration operations by the aspirate probe 202 and generates a trace signal 108, which is input into the trained machine learning model 106. The trained machine learning model 106 can identify a trace signal 108 indicative of a failed aspiration and category of the failure (e.g., no aspiration occurs; no dispensation occurs; no trace signals observed, etc.). In an embodiment, the trained machine learning model 106 can identify whether the trace signal 108 is valid or invalid (i.e., indicative of failure) 114. In another embodiment, the trained machine learning model 106 can further output a reconstructed trace signal 110, together with one or more discrepancies 112 in case of a failure. In another embodiment, the trained machine learning model 106 can further output an estimated volume 206 aspirated by the aspirate probe 202. FIG. 3 illustrates a valid sample trace signal from a sensor connected to the aspirate probe 202, in accordance with an embodiment described herein. The sample trace signal can include time-series measured data collected from any sensor (e.g., the pressure sensor 204).

[0017] FIG. 4 illustrates a block diagram of the machine learning model 106, in accordance with an embodiment described herein. In this embodiment, the machine learning model 106 is implemented as a deep neural network. As is generally understood in the art, the term "deep neural network" refers to an artificial neural network that uses multiple layers to progressively extract higher-level features from input data. The machine learning model 106 shown in FIG. 4 includes an encoder network 302, a decoder network 304, a classification network 306, and a parameter estimation network 308. In this example, the encoder network 302 encodes the trace signal 108 into one or more compact representations 310. In one embodiment, the compact representations 310 comprise a 64-dimensional manifold. The representations 310 are used in other networks 304, 306, and 308 for trace signal reconstruction, failure identification, and parameter estimation, respectively. In an embodiment, the representations 310 can be input to the classification network 306 to classify the trace signal 108 as a valid or invalid trace signal. In a further embodiment, the representations 310 can be input to the decoder network 304 to reconstruct a trace signal 110. If the trace signal 108 is identified as an invalid signal, then the reconstructed trace signal 110 is different from the original trace signal 108. One or more discrepancies 112 between the reconstructed trace signal 110 and the original trace signal 108 are highlighted in a display. For example, in one embodiment, the reconstructed trace signal 110 and the original trace signal 108 are displayed in a graphical user interface (GUI) and the discrepancies 112 are shown in red color, so that the user can be notified of the failure and quickly locate the possible failure cause. The discrepancies 112 can be identified and highlighted based on the reconstructed trace signal 110 from the decoder network 304 and gradients 312 from the classification network 306. Conventional backpropagation techniques may be used to compute gradients 312 in a classification network 306 more efficiently. FIG. 5A illustrates an original trace signal 502 indicative of a failure, in accordance with an embodiment described herein. FIG. 5B illustrates a reconstructed trace signal 504, in accordance with an embodiment described herein. FIG. 5C illustrates the original trace signal 502 highlighting a discrepancy 112 between the original trace signal and the reconstructed trace signal, in accordance with an embodiment described herein.

[0018] Continuing with the example machine learning model 106 of FIG. 4, the representations 310 are input to the parameter estimation network 308 for parameter estimation. For example, where a machine learning model 106 is included in the failure automatic detection system 200 with the aspirate probe 202, the parameter estimation network 308 can regress the representations 310 to obtain an estimated parameter 116 (e.g., estimated volume 206 aspirated by the aspirate probe 202). The estimated volume 206 is compared with an expected volume. If the difference between the estimated volume and the expected volume is over a predefined threshold, then the aspiration is considered a failure. If the difference between the estimated volume and the expected volume is within the predefined threshold, then it can be concluded that a sufficient volume has been aspirated, irrespective of ambient noises that may be observed in the original trace signal 108.

[0019] If a trace signal indicates a failure, then the classification network 306 can further categorize the failure into a failure category depending on a failure cause. For example, if a failure associated with the aspirate probe 202 is detected, the failure can result from a different factor. For instance, the aspirate probe 202 may fail to aspirate any liquid; the aspirate probe 202 may fail to dispense any liquid after aspiration; the pressure sensor 204 may malfunction because no trace signals can be observed. The classification network 306 can categorize the failure into a failure category, in addition to valid or invalid identification 114. In an embodiment, a failure notification will be sent to a system (e.g., an immunoassay and clinical chemistry analyzer) controlling the aspirate probe 202. This failure notification may then be presented on a display for review by an operator or other use. Alternatively, the system may automatically take one or more measures to remediate the failure (e.g., re-aspirate the liquid).

[0020] In an embodiment, a plurality of sensors can be provided and connected to the device 102. For example, a pressure sensor, a temperature sensor, and a current sensor are connected to the device 102. Each sensor generates a different trace signal 108, and all the trace signals are input to the machine learning model 106 for analysis.

[0021] Although the machine learning model 106 discussed above with reference to FIG. 4 is implemented as a deep neural network, it should be understood that other machine learning models may be utilized in other embodiments of the present invention. Various machine learning models generally known in the art may be utilized including decision tree, gradient-boosted tree, logistic regression, and random forest models.

[0022] The machine learning model 106 is trained with a large number of trace signals, with a good sampling of both valid or invalid (i.e., indicative of failure) signals. For example, in one embodiment, the machine learning model 106 is trained with about 80,000 trace signals, including nearly 10,000 trace signals indicative of failures. According to proof of concept testing performed with this training set, the deep neural network can achieve a classification accuracy of 0.9942, which indicates that the failure automatic detection system 100 can effectively detect a failure of a task using sensor trace data.

[0023] FIG. 6 illustrates a flow chart of a method 600 for training the machine learning model 106, in accordance with an embodiment described herein. At step 602, a large number of trace signals 108 are provided to train the machine learning model 106. Ideally, the trace signals 108 include a large number of trace signals, with a sampling of valid and invalid trace signals representative of various scenarios under which a corresponding device operates. At step 604, each trace signal 108 is labeled by humans (e.g., machine learning experts with respect to the device 102 and the sensor 104 that generates the trace signal 108). The labeling tags can indicate whether this trace signal 108 is valid or invalid. The labeling tags can also provide other information, for example, the volume aspirated by the aspirate probe 202, the number of aspiration operations, etc. At step 606, the labeled trace signals are input into the machine learning model 106 for training. With supervised learning, the trained machine learning model 106 can identify whether a new trace signal is valid or invalid. The trained machine learning model 106 can further estimate one or more parameters from the new trace signal, such as an estimated volume aspirated by the aspirate probe 202, the estimated number of aspiration operations, etc. The trained machine learning model 106 can further reconstruct the new trace signal, which is compared with the original new trace signal. One or more discrepancies 112 between the reconstructed trace signal and the original new trace signal can be highlighted.

[0024] FIG. 7 illustrates a flow chart of a method for detecting a failure of a device 102, in accordance with an embodiment described herein. At step 702, a new trace signal is inputted into the trained machine learning model 106. At step 704, the encoder network 302 in the trained machine learning model 106 encodes the new trace signal into vector representations 310, which form a 64-dimensional manifold in this example. At step 706, the classification network 306 identifies whether the new trace signal is valid or invalid. At step 708, the decoder network 304 reconstructs the new trace signal based on the vector representations 310. At step 710, one or more discrepancies 112 between the reconstructed trace signal and the original new trace signal are highlighted based on the gradients 312 provided by the classification network 306. At step 712, the parameter estimation network 308 estimates one or more parameters (e.g., an estimated volume aspirated by the aspirate probe 202, the estimated number of aspiration operations, etc.) associated with the device 102. The order of steps 706-712 can be randomly adjusted in some embodiments; for example, step 712 can be performed prior to step 706.

[0025] FIG. 8 provides an example of a parallel processing memory architecture 800 that may be utilized to implement the machine learning model 106 and other aspects of the various workflows discussed herein. This architecture 800 may be used in embodiments of the present invention where NVIDIA CUDA ™< (or a similar parallel computing platform) is used. The architecture includes a host computing unit ("host") 805 and a graphics processing unit (GPU) device ("device") 810 connected via a bus 815 (e.g., a PCIe bus). The host 805 includes a processor or a central processing unit (referred to as "CPU") (not shown in FIG. 8), and host memory 825 accessible to the CPU. The device 810 includes the graphics processing unit (GPU) and its associated memory 820, referred to herein as device memory. The device memory 820 may include various types of memory, each optimized for different memory usages. For example, in some embodiments, the device memory includes global memory, constant memory, and texture memory.

[0026] Parallel portions of a big data platform and / or big simulation platform (see FIG. 8) may be executed on the architecture 800 as "device kernels" or simply "kernels." A kernel comprises a parameterized code configured to perform a particular function. The parallel computing platform is configured to execute these kernels in an optimal manner across the architecture 800 based on parameters, settings, and other selections provided by the user. Additionally, in some embodiments, the parallel computing platform may include additional functionality to allow for automatic processing of kernels in an optimal manner with minimal input provided by the user.

[0027] The processing required for each kernel is performed by a grid of thread blocks (described in greater detail below). Using concurrent kernel execution, streams, and synchronization with lightweight events, the architecture 800 of FIG. 8 (or similar architectures) may be used to parallelize portions of the model-based operations performed in training or utilizing the machine learning model discussed herein.

[0028] The device 810 includes one or more thread blocks 830 which represent the computation unit of the device 810. The term thread block refers to a group of threads that can cooperate via shared memory and synchronize their execution to coordinate memory accesses. For example, in FIG. 8, threads 840, 845 and 850 operate in thread block 830 and access shared memory 835. Depending on the parallel computing platform used, thread blocks may be organized in a grid structure. A computation or series of computations may then be mapped onto this grid. For example, in embodiments utilizing CUDA, computations may be mapped on one-, two-, or three-dimensional grids. Each grid contains multiple thread blocks, and each thread block contains multiple threads. For example, in FIG. 8, the thread blocks 830 are organized in a two-dimensional grid structure with m+1 rows and n+1 columns. Generally, threads in different thread blocks of the same grid cannot communicate or synchronize with each other. However, thread blocks in the same grid can run on the same multiprocessor within the GPU at the same time. The number of threads in each thread block may be limited by hardware or software constraints.

[0029] Continuing with reference to FIG. 8, registers 855, 860, and 865 represent the fast memory available to thread block 830. Each register is only accessible by a single thread. Thus, for example, register 855 may only be accessed by thread 840. Conversely, shared memory is allocated per thread block, so all threads in the block have access to the same shared memory. Thus, shared memory 835 is designed to be accessed, in parallel, by each thread 840, 845, and 850 in thread block 830. Threads can access data in shared memory 835 loaded from device memory 820 by other threads within the same thread block (e.g., thread block 830). The device memory 820 is accessed by all blocks of the grid and may be implemented using, for example, Dynamic Random-Access Memory (DRAM).

[0030] Each thread can have one or more levels of memory access. For example, in the architecture 800 of FIG. 8, each thread may have three levels of memory access. First, each thread 840, 845, 850, can read and write to its corresponding registers 855, 860, and 865. Registers provide the fastest memory access to threads because there are no synchronization issues and the register is generally located close to a multiprocessor executing the thread. Second, each thread 840, 845, 850 in thread block 830, may read and write data to the shared memory 835 corresponding to that block 830. Generally, the time required for a thread to access shared memory exceeds that of register access due to the need to synchronize access among all the threads in the thread block. However, like the registers in the thread block, the shared memory is typically located close to the multiprocessor executing the threads. The third level of memory access allows all threads on the device 810 to read and / or write to the device memory. Device memory requires the longest time to access because access must be synchronized across the thread blocks operating on the device.

[0031] The embodiments of the present disclosure may be implemented with any combination of hardware and software. For example, aside from parallel processing architecture presented in FIG. 8, standard computing platforms (e.g., servers, desktop computers, etc.) may be specially configured to perform the techniques discussed herein. In addition, the embodiments of the present disclosure may be included in an article of manufacture (e.g., one or more computer program products) having, for example, computer-readable, non-transitory media. The media may have embodied therein computer-readable program code for providing and facilitating the mechanisms of the embodiments of the present disclosure. The article of manufacture can be included as part of a computer system or sold separately.

[0032] The scope of the present invention is indicated by the following appended claims.

[0033] An executable application, as used herein, comprises code or machine-readable instructions for conditioning the processor to implement predetermined functions, such as those of an operating system, a context data acquisition system or other information processing system, for example, in response to user command or input. An executable procedure is a segment of code or machine-readable instruction, sub-routine, or other distinct section of code or portion of an executable application for performing one or more particular processes. These processes may include receiving input data and / or parameters, performing operations on received input data and / or performing functions in response to received input parameters, and providing resulting output data and / or parameters.

[0034] A graphical user interface (GUI), as used herein, comprises one or more display images, generated by a display processor and enabling user interaction with a processor or other device and associated data acquisition and processing functions. The GUI also includes an executable procedure or executable application. The executable procedure or executable application conditions the display processor to generate signals representing the GUI display images. These signals are supplied to a display device that displays the image for viewing by the user. The processor, under control of an executable procedure or executable application, manipulates the GUI display images in response to signals received from the input devices. In this way, the user may interact with the display image using the input devices, enabling user interaction with the processor or other devices.

[0035] The functions and process steps herein may be performed automatically or wholly or partially in response to a user command. An activity (including a step) performed automatically is performed in response to one or more executable instructions or device operation without user direct initiation of the activity.

[0036] The present description and claims may make use of the terms "a," "at least one of," and "one or more of," with regard to particular features and elements of the illustrative embodiments. It should be appreciated that these terms and phrases are intended to state that there is at least one of the particular features or elements present in the particular illustrative embodiment, but that more than one may also be present. That is, these terms / phrases are not intended to limit the description or claims to a single feature / element being present or require that a plurality of such features / elements be present. To the contrary, these terms / phrases only require at least a single feature / element with the possibility of a plurality of such features / elements being within the scope of the description and claims.

[0037] In addition, it should be appreciated that the following description uses a plurality of various examples for various elements of the illustrative embodiments to further illustrate example implementations of the illustrative embodiments and to aid in the understanding of the mechanisms of the illustrative embodiments. These examples are intended to be nonlimiting and are not exhaustive of the various possibilities for implementing the mechanisms of the illustrative embodiments.

[0038] Although the invention has been described with reference to exemplary embodiments. The scope of the invention is defined by the following claims.

Claims

1. A computer-implemented method for detecting a failure of a device (102), wherein the device (102) is connected to a sensor, characterized in that, the method comprising: receiving, by a machine learning model (106), a trace signal (108) from the sensor (104) indicating a status of the device (102), wherein the trace signal (108) is input into the machine learning model (106); encoding, by the machine learning model (106) using an encoder network (302), the trace signal (108) into a plurality of vector representations (310); determining, by the machine learning model (106) using a classification network (306), whether the trace signal (108) is valid or invalid based on the plurality of vector representations (310), wherein the representations (310) are input to the classification network (306) to classify the trace signal (108) as a valid or invalid trace signal; reconstructing, by the machine learning model (106) using a decoder network (304), the trace signal (108), highlighting one or more discrepancies (112) between a reconstructed trace signal (110) and the trace signal (108) in a graphical user interface (GUI), wherein the representations (310) are input to the decoder network (304) to reconstruct the trace signal (110); wherein the trained machine learning model (106) outputs the reconstructed trace signal (110); and estimating one or more parameters (116) associated with the device (102) based on the plurality of vector representations (310), wherein the representations (310) are input to a parameter estimation network (308) for parameter estimation, wherein the trained machine learning model (106) outputs one or more estimated parameters (116) associated with the device (102), wherein the machine learning model (106) is a deep neural network.

2. The method of claim 1, further comprising: regressing the plurality of vector representations (310) to estimate the one or more parameters (116).

3. The method of claim 1, further comprising: if the trace signal (108) is invalid, identifying, by the machine learning model (106) using the classification network (306), a category of the failure.

4. The method of claim 1, further comprising: receiving, by the machine learning model (106), a plurality of training trace signals (108) from the sensor (104); and receiving, by the machine learning model (106), one or more labelling tags from a human for each training trace signal (108), wherein the one or more labelling tags indicate whether each training trace signal (108) is valid or invalid.

5. A computer-implemented method, as claimed in claim 1, for detecting a failure of an aspirate probe (202), wherein the aspirate probe (202) is connected to a sensor (104), characterized in that, the method comprising: receiving, by a machine learning model (106), a trace signal (108) from the sensor (104) indicating a status of the aspirate probe (202); encoding, by the machine learning model (106) using an encoder network (302), the trace signal (108) into a plurality of vector representations (310); and determining, by the machine learning model (106), whether the trace signal (108) is valid or invalid based on the plurality of vector representations (310); if the trace signal (108) is invalid, identifying, by the machine learning model (106) using a classification network (306), a category of the failure associated with the aspirate probe (202).

6. The method of claim 5, further comprising: regressing, by the machine learning model (106), the plurality of vector representations (310) to estimate one or more parameters (116) associated with the aspirate probe (202), and wherein preferably the one or more parameters (116) include a volume (206) aspirated by the aspirate probe (202), or the number of aspiration operations.

7. The method of claim 5, wherein the category of the failure is one of no aspiration, no dispensation, and no trace signals.

8. A system for detecting a failure of a device (102), wherein the device (102) is connected to a sensor (104), characterized in that, the system comprising: an encoder network (302) for encoding a first trace signal (108) from the sensor into a plurality of vector representations (310), wherein the trace signal (108) is input into a machine learning model (106); a decoder network (304) for reconstructing the plurality of vector representations (310) into a second trace signal (110), wherein the decoder network (304) is configured for identifying one or more discrepancies (112) between the first trace signal (108) and the second trace signal (110), wherein the representations (310) is input to the decoder network (304) to reconstruct the trace signal (110), and wherein the trained machine learning model (106) outputs the reconstructed trace signal (110); a classification network (306) for identifying whether the first trace signal (108) is valid or invalid, wherein the representations (310) are input to the classification network (306) to classify the trace signal (108) as a valid or invalid trace signal; a parameter estimation network (308) for estimating one or more parameters (116) associated with the device (102) based on the plurality of vector representations (310), wherein the representations (310) are input to the parameter estimation network (308) for parameter estimation, wherein the trained machine learning model (106) outputs one or more estimated parameters (116) associated with the device (102); wherein the machine learning model (106) is a deep neural network.

9. The system of claim 8, wherein the system is adapted to estimate the one or more parameters (116) by regressing the plurality of vector representations (310).

10. The system of claim 8, wherein the plurality of vector representations (310) is included in a 64-dimensional manifold, wherein the encoder network (302) in the trained machine learning model (106) is adapted to encode the new trace signal into vector representations (310), which form the 64-dimensional manifold.

11. The system of claim 8, wherein preferably the one or more discrepancies (112) are identified based on the second trace signal (110) and a plurality of gradients (312) output from the classification network (306).