Robust artificial intelligence inference in edge computing devices

Abstract

A system for supporting artificial intelligence inference in an edge computing device associated with a physical process or plant, the system comprising a neural network training module, a neural network testing module, and a digital twin of the physical process or plant. The neural network training module is configured for training a neural network model for deployment to the edge computing device based on data received from the edge computing device including baseline training data and live data. The neural network testing module is configured for validating the trained neural network model prior to deployment to the edge computing device by utilizing the digital twin of the physical process or plant.

Description

Technical Field

[0001] This disclosure generally relates to techniques for supporting artificial intelligence (AI) inference in edge computing devices. Embodiments of this disclosure provide a system and method for performing robust AI inference in edge computing devices associated with a process or plant by leveraging a digital twin of that process or plant. Background Technology

[0002] Recent advances in processor technology and AI have enabled real-time operating systems running on edge computing devices to efficiently perform predictions (also known as "inference") from neural network models residing within the AI ​​core. These models are trained on sample data from a research population. Given this, ensuring the accuracy and performance of the neural network model after deployment is challenging. For example, in dynamically changing production environments, the neural network model may face inputs that have not been extensively trained and may produce inaccurate predictions. Summary of the Invention

[0003] Various aspects of this disclosure relate to a technique for robust artificial intelligence reasoning in edge computing devices associated with a process or plant by utilizing a digital twin of the process or plant.

[0004] According to a first aspect of this disclosure, a system is provided for supporting artificial intelligence inference in an edge computing device associated with a physical process or plant. The system includes a neural network training module configured to train at least one neural network model for deployment to the edge computing device based on data received from the edge computing device, including baseline training data and field data. The system also includes a neural network testing module configured to evaluate the readiness of the trained neural network model prior to deployment to the edge computing device. The system further includes a digital twin of the physical process or plant, comprising a simulation platform configured to perform simulations of the physical process or plant. The neural network testing module is configured to: provide simulated input to the digital twin, the simulated input including one or more test scenarios involving the trained neural network model, the test scenarios being generated using the field data, and validate the trained neural network model based on simulated output obtained from the digital twin.

[0005] According to a second aspect of this disclosure, a computer-implemented method is provided for supporting artificial intelligence inference in an edge computing device associated with a physical process or plant. The method includes training at least one neural network model for deployment to the edge computing device based on data received from the edge computing device, including baseline training data and field data. The method also includes evaluating the readiness of the trained neural network model by employing a digital twin of the physical process or plant prior to deployment to the edge computing device. The digital twin includes a simulation platform configured to perform simulations of the physical process or plant. Evaluating the readiness of the trained neural network model includes providing simulated inputs to the digital twin, the simulated inputs including one or more test scenarios involving the trained neural network model, the test scenarios being generated using the field data, and validating the trained neural network model based on simulated outputs obtained from the digital twin.

[0006] Other aspects of this disclosure implement the features of the above-described methods in computing systems and computer program products.

[0007] Additional technical features and benefits can be achieved through the technology disclosed herein. Embodiments and aspects of this disclosure are described in detail herein and are considered part of the claimed subject matter. For a better understanding, refer to the detailed description and accompanying drawings. Attached Figure Description

[0008] The foregoing and other aspects of this disclosure are best understood from the following detailed description when read in conjunction with the accompanying drawings. For ease of identification of any element or action discussed, the most significant one or more digits in the reference numerals refer to the drawing number in which the element or action is first introduced.

[0009] Figure 1 A schematic block diagram of a system for remotely supporting AI inference in an edge computing device, according to an embodiment of the present disclosure, is shown.

[0010] Figure 2 An example of a routine for validating a computer implementation of a trained neural network model for deployment to an edge computing device is shown.

[0011] Figure 3 An exemplary routine implemented by a smart edge computing device is shown.

[0012] Figure 4 Exemplary computing environments in which embodiments of the present disclosure can be implemented are shown. Detailed Implementation

[0013] Cloud computing services can be used to train neural network (NN) models deployed to edge computing devices. Cloud resources also enable the storage of large amounts of data or the execution of computationally intensive software. In industry, digital twins provide digital images or replicas of physical processes or factories as a means of optimizing their performance. Such digital images can also reside in the cloud. In other areas, digital twins can leverage powerful simulation software to validate and optimize production characteristics.

[0014] To enable neural network models deployed on edge computing devices to deliver high-accuracy results, it is desirable to periodically retrain the models on newly collected data. This is particularly relevant in industrial settings where high performance accuracy is required in most use cases. For example, in industrial robotics applications, a movement in the path of a robotic arm that collides with critical equipment could have catastrophic consequences. Embodiments of this disclosure, as illustrated herein, provide a technique for leveraging cloud resources such as neural network training capabilities, digital twins, and simulation software to robustly handle inference in edge computing devices.

[0015] Now for reference Figure 1 A system 102 is illustrated for supporting AI inference in an edge computing device 120 associated with a physical process or plant 122. In the illustrated embodiment, system 102 is implemented in a cloud computing environment to take advantage of considerable computing power, large storage capacity, and the ability to run high-fidelity simulation software. In alternative embodiments, system 102 can be implemented on a dedicated server without residing in the cloud. System 102 broadly includes the following functional modules: a neural network training module 104 (hereinafter referred to as training module 104) configured to train one or more neural network models to be deployed to edge computing device 120; a neural network testing module 108 (hereinafter referred to as testing module 108) configured to evaluate the readiness of the trained neural network models obtained from training module 104 before deployment to edge computing device 120; and a digital twin 110 of the physical process or plant 122, which can be used by testing module 108 to validate the trained neural network models before deployment.

[0016] Digital twin 110 can include a high-fidelity model of process or plant 122, referred to herein as process / plant model 112. Process / plant model 112 can utilize, for example, CAD models of physical devices (referred to as field devices) representing process or plant 122. Process / plant model 112 can also include up-to-date sensor data associated with the individual physical devices. For example, a CAD drawing package can be used to create a digital model of process or plant 122, which the process control system can then use to create process / plant model 112. This software can provide a link between digital twin sensors and real-world controls. In other embodiments, instead of using a CAD model, sensor data such as imaging data can be used to generate digital representations of the field devices. Digital twin 110 can also include a simulation platform 114 configured to perform simulations of the physical process or plant 122 using process / plant model 112. Simulation platform 114 can include appropriate physics libraries for solving physical and / or process simulations of the physical process or plant 122 based on defined constraints.

[0017] Edge computing device 120 can include any device with computing capabilities deployed close to a field device in a physical process or plant 122. For example, edge computing device 120 can include a programmable logic controller (PLC) or a computing module coupled to a PLC. Edge computing device 120 can be configured to receive sensor input 142 from sensors associated with the field device, perform AI inference or prediction based on the sensor input 142 using one or more neural network models, and generate an output task 146 that can include commands for actuators associated with the field device. In one embodiment, edge computing device 120 can include a hardware accelerator (AI accelerator) specifically designed to efficiently perform prediction or inference from a deep neural network. The computing power of these AI accelerators allows them to run inference (forward pass) on a trained neural network model in an embedded device. A non-limiting example of an AI accelerator suitable for this embodiment is the SIMATIC manufactured by Siemens AG. TM TM NPU (Neural Processing Unit).

[0018] Training module 104 is capable of generating one or more trained neural network models 126 using one or more neural network model skeletons 124 (i.e., neural network models before training) and data 150 from data repository 106. Training data 150 can include baseline training data used in the initial training of the neural network models, as well as field data 130 received from edge computing device 120. Field data 130 can be derived from sensor data 142 associated with field devices. Field data 130 relayed by edge computing device 120 can include at least field data identified by edge computing device 120 as “fault data.” “Fault data” can include, for example, input data from sensors that results in inaccurate or low-confidence outputs (typically predictions or inferences) from the neural network model running on edge computing device 120. Inaccurate inference will cause one or more field devices to fail to perform the task generated based on that inference in a satisfactory manner. In some embodiments, if communication bandwidth allows, field data 130 relayed by edge computing device 120 can also include “success data” in addition to “fault data.” Such "success data" can include, for example, input data from sensors, which results in high-precision or high-confidence outputs from a neural network model running on the edge computing device 120. The edge computing device 120 can be configured to label relayed field data 130 as "failure data" or "success data".

[0019] The basic training of training module 104 can include, for example, data preparation (e.g., including normalization, filtering, and feature generation, among other steps) and neural network weight computation via backpropagation. Training module 104 can additionally employ advanced or post-training techniques, including, for example, data augmentation for inaccurately predicted data samples (as part of the model retraining specification, as described below), model pruning, regularization, etc. The trained neural network model 126 is then passed to testing module 108.

[0020] Test module 108 may include, for example, a high-fidelity sandbox capable of performing safety tests on a newly trained neural network model 126 of the digital twin 110 of the physical process or plant 122. Test module 108 may generate one or more test scenarios involving the trained neural network model 126. Test scenarios may be generated using field data 130. In one embodiment, “failure data” (i.e., input data that leads to previously inaccurate reasoning) may be specifically emphasized in generating test scenarios. Once the test scenarios are prepared, they are provided as analog inputs 132 to the digital twin 110.

[0021] Based on simulation input 132, digital twin 110 can perform simulations of physical processes or factories 122 using the latest process / factory model 112 by simulation platform 114. Specifically, simulations can be performed based on constraints, such as tasks / actions derived from predictions or inferences generated by trained neural network models 126 in conjunction with a test scenario. Digital twin 110 returns simulation output 134 to test module 108. Simulation output 134 can include, for example, performance metrics associated with each trained neural network model 126. Performance metrics can quantify the accuracy or confidence level of predictions or inferences generated by the trained network models 126, for example, based on the simulated response to the predictions or inferences.

[0022] Based on the simulated output 134 from the digital twin 110, the testing module 108 can determine whether the test is satisfactory. For example, the testing module 108 can determine whether the performance metrics in the simulated output 134 are acceptable based on defined thresholds (e.g., quantified confidence levels or accuracy). If the performance metrics are deemed unacceptable, the testing module 108 can request the training module 104 to retrain the trained neural network model based on a retraining specification 136. The retraining specification 136 can represent the methods that the training module 104 will consider in the next training iteration. An exemplary item in the methods could be a request to perform data augmentation on some underperforming data in the simulated input 132 to make the neural network model robust relative to such data. Underperforming data could include, for example, input data that causes faults in the output, as determined by simulation. In such cases, the retraining specification 136 can request data augmentation on the faulty input data to more closely fit the faulty conditions, thereby allowing the neural network model to produce correct predictions or inferences for these conditions.

[0023] The retraining of the neural network model by the training module 104 and the validation of the retrained neural network model by the testing module 108 using the digital twin 110 can be performed iteratively. The testing module 108 can evaluate the acceptability of the retrained neural network model, for example, based on the performance metrics described above, and terminate the retraining process after a specified number of iterations or after achieving acceptable performance metrics. Once validated, the neural network model is considered deployment-ready. The deployment-ready neural network model 138 can be pushed to the model memory 118. The model memory 118 can contain a set of neural network models that have been validated and are ready for deployment to the edge computing device 120. A set of neural network models 140 can be deployed from the model memory 118 to the edge computing device 120. Depending on the capacity of the edge computing device 120, the set of deployed neural network models 140 can be a subset or the entire set of validated neural network models in the model memory 118. The model memory 118 can be periodically updated, for example, by adding newly validated neural network models and / or discarding outdated or unused models.

[0024] As an additional feature supporting AI inference in edge computing device 120, system 102 can be configured to receive high-precision output requests 146 from edge computing device 120 under certain circumstances, such as bandwidth within time tolerance or low-confidence outputs from deployed neural network models. In such cases, system 102 can use digital twin 110 to return high-precision inference 148 to edge computing device 120. For this purpose, digital twin 110 can include an undeployed neural network model 116, which may have greater weights (i.e., computational intensity) than the neural network model deployed to edge computing device 120. The heavy-duty neural network model 116 can be used to generate high-precision inference, which can be tested using the simulation capabilities of digital twin 110 before being returned to edge computing device 120.

[0025] Figure 2 A flowchart is shown for a method 200 for validating a trained neural network model for deployment to an edge computing device associated with a physical process or plant. Method 200 can include computer-implemented routines. Method 200 can be implemented as part of a maintenance service for an edge computing device associated with a physical process or plant, for example, in a cloud computing environment. As part of the maintenance service, method 200 can be executed periodically at a specified frequency, or it can be executed in an event-based manner, such as being triggered by a plant operator. Figure 2 The flowchart is not intended to indicate that the boxes of method 200 will be executed in any particular order, or that all boxes of method 200 will be included in every case. Furthermore, method 200 is capable of including any suitable number of additional operations.

[0026] Box 202 of method 200 relates to training a neural network model. Training can be implemented using data including baseline training data and field data received from the edge computing device. The field data used in training can include at least the field data identified as “fault data” by the edge computing device. Method 200 then uses a digital twin of the physical process or plant to evaluate the readiness of the trained neural network model to be deployed to the edge computing device. The digital twin can include a simulation platform that performs simulations of the physical process or plant.

[0027] Box 204 of method 200 relates to generating one or more test scenarios involving a trained neural network model. Test scenarios can be generated using field data, and in some embodiments, "fault data" is specifically emphasized, i.e., input data that led to previously inaccurate inference.

[0028] In box 206, the test scenario is sent as a simulated input to the digital twin.

[0029] Box 208 of method 200 relates to performing simulations of physical processes or plants via digital twins. This can include performing simulations of physical processes or plants based on inference generated by a trained neural network model in conjunction with one or more test scenarios.

[0030] Box 210 of method 200 relates to obtaining a simulated output from a digital twin. The simulated output 134 can include, for example, performance metrics associated with a trained neural network model.

[0031] Box 212 of method 200 relates to a decision regarding whether a performance metric is acceptable. This decision can be made, for example, based on a defined threshold.

[0032] If the performance metrics are deemed unacceptable at box 212, control proceeds to box 214, which involves generating a retraining specification. Control then returns to box 202, which includes retraining the neural network model based on the retraining specification. The retraining process is performed in an iterative loop until it is determined at box 212 that the retrained neural network model produces acceptable performance metrics. Alternatively, the retraining loop can be terminated after a specified number of iterations. The neural network model is then validated.

[0033] Box 216 of method 200 relates to deploying a validated neural network model to an edge computing device.

[0034] Figure 3 Examples of routines or methods 300 that can be implemented by intelligent edge computing devices associated with physical processes or factories are shown. Figure 3The flowchart is not intended to indicate that the boxes of method 300 will be executed in any particular order, or that all boxes of method 300 will be included in every case. Furthermore, method 300 is capable of including any suitable number of additional operations.

[0035] Box 302 of method 300 relates to receiving input data from field devices in a process or plant. The input data can be collected from sensors connected to the field devices and prepared in an edge computing device, for example, using vision and / or signal processing techniques.

[0036] Box 304 of method 300 involves generating predictions or inferences from input data by running one or more neural network models deployed to an edge computing device.

[0037] Box 306 of method 300 relates to evaluating the quality of inference or prediction provided by a locally deployed neural network model before providing a prediction or inference to the output interface of an edge computing device. Specifically, box 306 relates to a decision on whether to indicate high-precision inference. High-precision inference can be indicated in one or more of the following non-limiting exemplary scenarios: (1) the output of the next action can tolerate the increased latency of cloud requests, (2) the output of the neural network model of the edge computing device has a confidence level below a specified threshold, and (3) the edge computing device has successfully attempted multiple times on the same input data and requires a high-quality solution to exit the loop.

[0038] If it is determined at box 306 that no high-precision inference is indicated, control proceeds to box 312. If it is determined at box 306 that high-precision inference is indicated, control proceeds to box 308, which involves sending a high-precision output request to the digital twin of the physical process or plant. The digital twin can reside in a cloud computing environment. When a high-precision output request is received in the cloud, the digital twin can use its current model of the process or plant, high-precision simulation tools, and potential heavy-duty neural network models to compute and test inference. Box 310 of method 300 involves receiving high-precision inference from the digital twin.

[0039] Box 312 of method 300 relates to converting prediction or inference into an output task, which can include commands for actuators associated with field devices.

[0040] Figure 4 An exemplary computing environment 400 in which various embodiments of the present disclosure can be implemented is shown. Computers and computing environments, such as computer system 402 and computing environment 400, are known to those skilled in the art and are therefore briefly described herein. It should be noted that some items included in the exemplary computing environment 400 (e.g., keyboard 436) may not be applicable when implementing embodiments of the present disclosure in a cloud computing environment.

[0041] like Figure 4 As shown, computer system 402 may include a communication mechanism such as system bus 404 or other communication mechanisms for transmitting information within computer system 402. Computer system 402 also includes one or more processors 406 coupled to system bus 404 for processing information. Processor 406 may include one or more central processing units (CPU), graphics processing units (GPUs), or any other processor known in the art.

[0042] Computer system 402 also includes system memory 408 coupled to system bus 404 for storing information and instructions to be executed by processor 406. System memory 408 may include computer-readable storage media in the form of volatile and / or non-volatile memory, such as read-only memory (ROM) 910 and / or random access memory (RAM) 912. System memory RAM 412 may include other dynamic storage devices (e.g., dynamic RAM, static RAM, and synchronous DRAM). System memory ROM 410 may include other static storage devices (e.g., programmable ROM, erasable PROM, and electrically erasable PROM). Furthermore, system memory 408 may be used to store temporary variables or other intermediate information during processor 406 instruction execution. A basic input / output system 414 (BIOS) containing basic routines that facilitate the transfer of information between elements within computer system 402, such as during startup, may be stored in system memory ROM 410. System memory RAM 412 may contain data and / or program modules that are immediately accessible by processor 406 and / or currently being operated by processor 406. System memory 408 may additionally include, for example, an operating system 416, application programs 418, other program modules 420, and program data 422.

[0043] Computer system 402 also includes a disk controller 424 coupled to system bus 404 to control one or more storage devices, such as magnetic hard disks 426 and removable media drives 428 (e.g., floppy disk drives, optical disk drives, magnetic tape drives, and / or solid-state drives). Storage devices can be added to computer system 402 using appropriate device interfaces, such as Small Computer System Interface (SCSI), Integrated Device Electronics (IDE), Universal Serial Bus (USB), or FireWire.

[0044] Computer system 402 may also include a display controller 430 coupled to system bus 404 to control display 432 (such as a cathode ray tube (CRT) or liquid crystal display (LCD) for displaying information to a computer user. Computer system 402 includes a user input interface 434 and one or more input devices (such as a keyboard 436 and a clicking device 438) for interacting with the computer user and providing information to one or more processors 406. Clicking device 438 may be, for example, a mouse, light pen, trackball, or pointer, for transmitting directional information and command selections to one or more processors 406 and for controlling cursor movement on display 432. Display 432 may provide a touchscreen interface that allows input to supplement or replace the communication of directional information and command selections by clicking device 438.

[0045] Computer system 402 is capable of performing some or all of the processing steps of embodiments of the present disclosure in response to one or more processors 406 executing one or more sequences of one or more instructions contained in memory such as system memory 408. Such instructions can be read into system memory 408 from another computer-readable medium, such as magnetic hard disk 426 or removable media drive 428. Hard disk 426 can contain one or more data repositories and data files used by embodiments of the present disclosure. Data storage contents and data files can be encrypted to enhance security. Processor 406 can also be used in a multiprocessor configuration to execute one or more sequences of instructions contained in system memory 408. In alternative embodiments, hard-wired circuitry can be used in place of or in combination with software instructions. Therefore, embodiments are not limited to any particular combination of hardware circuitry and software.

[0046] Computer system 402 may include at least one computer-readable medium or memory for storing instructions programmed according to embodiments of the present disclosure and for containing data structures, tables, records, or other data described herein. As used herein, the term "computer-readable medium" refers to any medium that participates in providing instructions to one or more processors 406 for execution. Computer-readable media can take many forms, including but not limited to non-transitory, non-volatile, volatile, and transmission media. Non-limiting examples of non-volatile media include optical discs, solid-state drives, magnetic disks, and magneto-optical discs, such as magnetic hard disk 426 or removable media drive 428. Non-limiting examples of volatile media include dynamic memory, such as system memory 408. Non-limiting examples of transmission media include coaxial cables, copper wires, and optical fibers, including conductors constituting system bus 404. Transmission media may also take the form of acoustic or optical waves, such as acoustic or optical waves generated during radio wave and infrared data communication.

[0047] The computing environment 400 may also include a computer system 402 operating in a networked environment using a logical connection to one or more remote computers, such as a remote computing device 444. The remote computing device 444 can be a personal computer (laptop or desktop), mobile device, server, router, network PC, peer-to-peer device, or other common network node, and typically includes many or all of the elements described above with respect to the computer system 402. When used in a networked environment, the computer system 402 may include a modem 442 for establishing communication over a network 440 such as the Internet. The modem 442 may be connected to the system bus 404 via a network interface 446 or via other suitable mechanisms.

[0048] Network 440 can be any network or system known in the art, including the Internet, intranet, local area network (LAN), wide area network (WAN), metropolitan area network (MAN), direct connection or a series of connections, cellular telephone network, or any other network or medium that facilitates communication between computer system 402 and other computers (e.g., remote computing device 444). Network 440 can be wired, wireless, or a combination thereof. Wired connections can be implemented using Ethernet, Universal Serial Bus (USB), RJ-6, or any other wired connection known in the art. Wireless connections can be implemented using Wi-Fi, WiMAX and Bluetooth, infrared, cellular networks, satellite, or any other wireless connection method known in the art. Furthermore, several networks can operate independently or communicate with each other to facilitate communication within network 440.

[0049] The embodiments of this disclosure can be implemented using any combination of hardware and software. Furthermore, the embodiments of this disclosure can be included in an article of manufacture having, for example, a computer-readable non-transitory medium (e.g., one or more computer program products). This medium contains, for example, computer-readable program code for providing and facilitating the embodiments of this disclosure. This article of manufacture may be included as part of a computer system or sold separately.

[0050] While various aspects and embodiments have been disclosed herein, other aspects and embodiments will be apparent to those skilled in the art. The various aspects and embodiments disclosed herein are for illustrative purposes and are not intended to be limiting.

[0051] The functions and processing steps described herein can be executed automatically, wholly or partially, in response to user commands. Automatically executed activities (including steps) are performed in response to one or more executable instructions or device operations without the user directly initiating the activity.

[0052] The systems and processes illustrated in the accompanying drawings are not exclusive. Other systems, processes, and menus can be derived from the principles of this disclosure to achieve the same purpose. Although this disclosure has been described with reference to specific embodiments, it should be understood that the embodiments and variations shown and described herein are for illustrative purposes only. Modifications to the present design can be made by those skilled in the art without departing from the scope of this disclosure.

Claims

1. A system for supporting artificial intelligence inference in an edge computing device associated with a physical process or factory, the system comprising: A neural network training module, configured to train at least one neural network model for deployment on the edge computing device based on data received from the edge computing device, including baseline training data and field data; A neural network testing module configured to evaluate the readiness of the trained neural network model before deployment to the edge computing device; and The digital twin of the physical process or plant, the digital twin including a simulation platform configured to perform simulations of the physical process or plant; The neural network testing module is configured to: The digital twin is provided with simulated inputs, the simulated inputs including one or more test scenarios involving the trained neural network model, the test scenarios being generated using the field data, and The trained neural network model is validated based on the simulated output obtained from the digital twin. The digital twin is configured to generate the simulation output by performing a simulation of the physical process or factory based on inference generated by combining the trained neural network model with one or more of the test scenarios.

2. The system according to claim 1, wherein, The field data received from the edge computing device includes at least the field data identified as "fault data" by the edge computing device.

3. The system according to claim 2, wherein, The test scenario was generated using the "fault data".

4. The system according to claim 1, in, The simulated output from the digital twin includes performance metrics of the trained neural network model, and The neural network testing module is configured to verify the trained neural network model by determining whether the performance metric is acceptable based on a defined threshold.

5. The system according to claim 4, wherein, If the performance metrics are unacceptable, the neural network testing module is configured to request the neural network training module to retrain the neural network model based on the retraining specification.

6. The system according to claim 5, wherein, The retraining specification includes a request to perform data augmentation on underperforming data in the simulated input.

7. The system according to claim 5, wherein, The neural network testing module is configured to iteratively request retraining of the neural network model and verify the retrained neural network model before deployment to the edge computing device, until: Perform a specified number of iterations, or The performance metrics have been achieved to an acceptable level.

8. The system according to claim 1, wherein, The digital twin is configured to provide high-precision inference directly to the edge computing device in response to a request from the edge computing device. The high-precision inference is generated by combining one or more undeployed neural network models with the digital twin's simulation of the physical process or plant.

9. The system according to claim 1, wherein, The system is implemented in a cloud computing environment.

10. A computer-implemented method for supporting artificial intelligence inference in an edge computing device associated with a physical process or factory, the method comprising: At least one neural network model for deployment on the edge computing device is trained based on data received from the edge computing device, including baseline training data and field data. as well as Prior to deployment to the edge computing device, the readiness of the trained neural network model is evaluated using a digital twin of the physical process or factory, the digital twin including a simulation platform configured to perform simulations of the physical process or factory, wherein evaluating the readiness of the trained neural network model includes: The digital twin is provided with simulated inputs, which include one or more test scenarios involving the trained neural network model, the test scenarios being generated using the field data. The digital twin performs a simulation of the physical process or factory by inference generated based on the trained neural network model combined with one or more test scenarios, thereby generating a simulation output. The trained neural network model is validated based on the simulated output obtained from the digital twin.

11. The method according to claim 10, wherein, The field data received from the edge computing device includes at least the field data identified as "fault data" by the edge computing device.

12. The method of claim 11, wherein the method includes generating the test scenario using the "fault data".

13. The method according to claim 10, wherein, The simulated output from the digital twin includes performance metrics of the trained neural network model, and The method includes validating the trained neural network model by determining whether the performance metric is acceptable based on a defined threshold.

14. The method of claim 13, comprising: If the performance metrics are unacceptable, the neural network model is retrained based on the retraining specification.

15. The method according to claim 14, wherein, The retraining specification includes a request to perform data augmentation on underperforming data in the simulated input.

16. The method of claim 14, wherein the method comprises iteratively retraining the neural network model and validating the retrained neural network model prior to deployment to the edge computing device, until: Perform a specified number of iterations, or The performance metrics have been achieved to an acceptable level.

17. The method of claim 10, comprising: In response to a request from the edge computing device, the digital twin provides high-precision inference to the edge computing device. The high-precision inference is generated by combining one or more undeployed neural network models with the digital twin's simulation of the physical process or plant.

18. A non-transitory computer-readable storage medium, the computer-readable storage medium comprising instructions that, when executed by a computer, cause the computer to: Based on data received from the edge computing device, including baseline training data and field data, at least one neural network model is trained for deployment on the edge computing device associated with a physical process or factory; and Before deployment to the edge computing device, the readiness of the trained neural network model is evaluated using a digital twin of the physical process or factory, the digital twin including a simulation platform configured to perform simulations of the physical process or factory, wherein... The assessment of the readiness of the trained neural network model includes: The digital twin is provided with simulated inputs, which include one or more test scenarios involving the trained neural network model, the test scenarios being generated using the field data. The digital twin performs a simulation of the physical process or factory by inference generated based on the trained neural network model combined with one or more test scenarios, thereby generating a simulation output. The trained neural network model is validated based on the simulated output obtained from the digital twin.

Images