Generalized and alignment model for repair recommendation

Abstract

To provide generalized and alignment models for repair recommendation.SOLUTION: Systems and methods described herein can involve: training a first generative artificial intelligence (AI) model for a general domain, the first generative AI model being trained using standard information components of the general domain; training a second AI model for a specific domain from the first generative AI model, the training of the second AI model being based on the use of the standard information components, non-standard information components of the specific domain and available label data of the specific domain; and fine-tuning the second AI model to align with preferences of the specific domain to maximize reward and minimize error.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates generally to repair systems, and more particularly to generalized and consistent models for repair recommendations. [Background technology]

[0002] Maintenance is a process whose objective is to keep equipment in an operational, efficient, and cost-effective condition. The maintenance process is accomplished by performing actions on the equipment necessary to accomplish one or more of these objectives. These actions include, but are not limited to, inspection, adjustment, repair, and overhaul of the equipment or its components.

[0003] When complex equipment breaks down, it can sometimes be difficult and time-consuming to diagnose the equipment, recognize the root cause of the problem, and determine what repair actions need to be taken to fix the equipment. This problem is solved using a data-driven system for predictive maintenance (also known as a repair recommendation system) that guides maintenance personnel on how to repair the equipment when it is brought in for repair. Summary of the Invention [Problem to be solved by the invention]

[0004] Data-driven repair recommendations using machine learning models have traditionally relied on the availability of large amounts of data for every instance on which they are created and developed. Such a requirement makes the system, as well as the machine learning models trained using the data, non-transferable. Related art implementations lack the ability to learn from different customers / instances in a similar field, which makes such implementations unusable for customers / instances with less data. Furthermore, related art repair recommendation models do not have the ability to adapt to customer preferences (i.e., no room for personalization or tuning to customer preferences).

[0005] Figure 1 shows an example of a related art data-driven repair recommendation system. Data-driven repair recommendation (i.e., recommending a course of repair actions based on multi-modal input) follows the flow shown in Figure 1. In this related art flow, a model is trained using historical repair data. [Means for solving the problem]

[0006] The implementation solves the problems of the related art by developing generative models in standard information components that are transferable for different customers / instances of the system and developing a data-driven model update mechanism based on engineer feedback to personalize / tune the model to customer preferences. The proposed algorithmic system accommodates multiple modalities where the data includes high dimensional sparse and dense data components. The algorithm includes a training mechanism for consistent structured output prediction. The algorithm leverages the standardized information components to train the generative model. The algorithm includes a mechanism for instantiating the generative model and training the fine-tuned model using available data for the customer / instance. The algorithm also includes a mechanism for personalizing the model by using customer feedback to update the model parameters to match customer preferences.

[0007] Data-driven repair recommendation systems have the following problems:

[0008] Lack of knowledge reuse: In repair recommendation systems, there is an information component for an asset class that is standardized and used to train the model. For example, fault codes generated in the automotive field following Society of Automotive Engineers (SAE) standards such as the J1939 protocol can be used. In related art repair recommendation systems, the model is trained using such standard information as well as other information available for the instance / customer. The standard information has a wealth of knowledge associated with it and is not limited to a particular instance or customer. Such knowledge should be reused and transferred to different instances or customers. Currently, no special modeling techniques are used to handle standard information and develop a generalized AI on knowledge with the intention of reuse and transfer to different instances / customers.

[0009] Model inapplicability to small data: Existing repair recommendation systems assume that there is a large amount of data available for training the model per instance or customer. However, this assumption is often incorrect and many customers / use cases cannot utilize the power of the repair recommendation system. Current systems do not apply unique techniques to solve this problem. Knowledge reuse as described above and the methods developed in the implementations described herein mitigate this situation.

[0010] Lack of personalization and tuning: Repair recommendation models trained using a lot of historical data are never perfect and can make mistakes. Instances / use cases / customers can have preferences regarding model behavior. These preferences map to personalization. Current repair recommendation systems do not provide any algorithmic room to tune model parameters for instance / use case / customer specific personalization. The problem of personalization is further heightened when less historical data is available to train the model.

[0011] The implementation described herein solves all the above problems by developing generalized models for standard information components in the form of generative models and making such models transferable from one instance / customer to another, thereby enabling knowledge reuse. The proposed implementation also mitigates the data scarcity problem by using the generalized models and other data available from the customers to develop custom model instances for each instance / customer. Finally, the system allows updating of model parameters leading to personalization and tuning to customer / instance / use case preferences.

[0012] Aspects of the disclosure may include a method, the method may include training a first generative artificial intelligence (AI) model for a general domain, where the first generative AI model is trained using standard information components for the general domain; training a second AI model for a specific domain from the first generative AI model, where training the second AI model is based on use of the standard information components, non-standard information components for the specific domain, and available label data for the specific domain; and fine-tuning the second AI model to match preferences for the specific domain to maximize reward and minimize error.

[0013] Aspects of the disclosure may include a computer program, the computer program may include instructions involving training a first generative artificial intelligence (AI) model for a general domain, the first generative AI model being trained using standard information components for the general domain, training a second AI model for a specific domain from the first generative AI model, the training of the second AI model being based on the use of standard information components, non-standard information components for the specific domain, and available label data for the specific domain, and fine-tuning the second AI model to match preferences for the specific domain to maximize rewards and minimize errors. The computer program and instructions may be stored in a non-transitory computer-readable medium and executed by one or more processors.

[0014] Aspects of the disclosure may include a system, the system including means for training a first generative artificial intelligence (AI) model for a general domain, where the first generative AI model is trained using standard information components for the general domain; means for training a second AI model for a specific domain from the first generative AI model, where training the second AI model is based on use of the standard information components, non-standard information components for the specific domain, and available label data for the specific domain; and means for fine-tuning the second AI model to match preferences for the specific domain to maximize reward and minimize error.

[0015] Aspects of the disclosure may include an apparatus, the apparatus including a processor configured to: train a first generative artificial intelligence (AI) model for a general domain, where the first generative AI model is trained using standard information components for the general domain; train a second AI model for a specific domain from the first generative AI model, where training of the second AI model is based on use of the standard information components, non-standard information components for the specific domain, and available label data for the specific domain; and fine-tune the second AI model to match preferences for the specific domain to maximize reward and minimize error. [Brief description of the drawings]

[0016] [Figure 1] FIG. 1 shows an example of a related art data-driven repair recommendation system.

[0017] [Diagram 2] FIG. 2 illustrates an example of components and flow of an example implementation described herein.

[0018] [Diagram 3] FIG. 3 illustrates the system flow according to one implementation.

[0019] [Figure 4] FIG. 4 illustrates multiple physical systems networked to a management device according to one implementation.

[0020] [Diagram 5] FIG. 5 illustrates an example computing environment with an example computing device suitable for use with some implementations. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] The following detailed description provides details of the drawings and implementation examples of the present application. Redundant element reference numbers and descriptions between the drawings are omitted for clarity. The terms used throughout the description are provided as examples and are not intended to be limiting. For example, the use of the term "automatic" may include a fully automatic implementation or a semi-automatic implementation including user or administrator control over certain aspects of the implementation, depending on the desired implementation of the person skilled in the art practicing the implementation of the present application. The selection may be made by a user through a user interface or other input means, or may be implemented by a desired algorithm. The implementations described herein may be utilized alone or in combination, and the functionality of the implementations may be implemented by any means depending on the desired implementation.

[0022] FIG. 2 illustrates an example of components and flow of an example implementation described herein.

[0023] Generalized Models for Standard Information Sets: The first component (Component 1) develops generalized models for standardized information sets. An example of a standardized information set is fault codes in the automotive field that follow SAE standards (such as the J1939 protocol). Although each fault code is standard, a set of fault codes can lead to different faults and associated repairs. Every customer may have the same set of fault codes, but different faults with different sequences of fault codes and therefore associated repairs. A generative model trained on such a standard information set allows for transfer of the model to different customers / instances, knowledge reuse across instances / customers, and combination with custom information available for fine-tuning and instantiation.

[0024] Scaling Instances / Customers: In the second component (Component 2), to instantiate a repair recommendation model, the generalized model is combined with information available on the use case / customer / instance to form a fine-tuned custom repair recommendation model, thereby achieving knowledge reuse of standardized information, reduced dependency on availability of large amounts of historical data per instance / customer, and scaling of repair recommendation model instances.

[0025] Personalization / Preference Alignment: The third component (Component 3) involves personalizing each instance of the repair recommendation model by aligning it with customer preferences. An example of a preference is that if the repair recommended by the model is wrong, a penalty of $100 will occur for customer 1, while a penalty of $10 will occur for customer 2. Preferences can have various attributes such as cost of wrong repair, safety aspect of wrong repair, etc., according to the desired implementation.

[0026] Figure 3 shows the system flow according to one implementation example. The following are the detailed steps, as shown in Figure 3. Steps S301 and S304 relate to component 1, steps S302 and S303 relate to component 2, and steps S305 to S309 relate to component 3.

[0027] Step S301: In step S301, this implementation trains a first generative artificial intelligence (AI) model for a general domain, where the first generative AI model is trained using standard information components for the general domain. Using a standard information repository, this step trains a generalized model for each piece of information in the set. The generalized model is a generative model and is trained in an unsupervised / self-supervised manner.

[0028] In one implementation, training can be done by inputting partial information (e.g., partial sequences) from known information of standard information components of a general domain, using a first generative AI model to output a prediction of the remaining information of the known information from the partial information, and utilizing unsupervised learning to reduce the error between the predicted remaining information and the known information. Specifically, a method for training a generative model can be one in which a part of the sequence is masked to be partial information, and the unmasked sequence is used to predict the masked part of the sequence. This error is used to train a generalized model. There may be other methods for training such generalized models, and their learned feature space is rich, given their generative capabilities. The standardized information set is continuously evolving, and information from different customers / instances is added to the set. Given that this is industry standard information related to an asset class (e.g., automobiles), the data does not violate any privacy or ethical and financial guidelines. The generalized model, being trained in an unsupervised / self-supervised manner, is transferable to different instances, thus enabling knowledge reuse and transfer. This first generative AI model is configured to output a first repair recommendation for the general domain. Such repair recommendations may include a series of repair actions, where each repair action in the series includes a location for repair and a repair action.

[0029] The generalized model is trained continuously based on the schedule and the availability of new information from new instances / customers.

[0030] Step S302 and Step S303: In steps S302 and S303, this implementation trains a second AI model for the specific domain from the first generative AI model, where the training of the second AI model is based on using standard information components, non-standard information components for the specific domain, and available label data for the specific domain. Such available label data can include a repair code indicative of the system to be repaired, and a repair action associated with the system to be repaired.

[0031] These steps together include creating a repair recommendation model instance. The repair recommendation model instance depends on the availability of data about the use case / customer. The available data can be of various modalities in addition to standard information. Furthermore, the available data does not need to be large-scale, and even a small amount of data can be applied to create the repair recommendation model. The available standard information undergoes the same data processing steps as creating the generalized model, while other data may have their own data processing steps. Thus, to train a second AI model for a specific domain from the first generative AI model, feature engineering can be performed on the non-standard information components of the specific domain, the features of the non-standard information components can be encoded, and the available label data can be used to combine the first generative AI model with the encoded non-standard information components to generate the second AI model.

[0032] By combining the first generative AI model with the encoded non-standard information component using the available label data, the repair recommendation model is composed of an instance of the generalized model and another model for handling other additional data modalities. The output of this model is a set of repair actions. During training, this combined model (i.e., the repair recommendation model) uses learning, such as supervised learning, to reduce the error of the second AI model from the available label information. Such an implementation allows for optimizing model parameters using the error between the prediction and the ground truth. This second AI model can be configured to output a second repair recommendation for the specific area. Such repair recommendation can include a set of repair actions, each repair action in the set of repair actions including a location for repair and a repair action.

[0033] Using composite models allows for knowledge reuse of standardized information. It should be noted that the collective knowledge is used for the generalized model and the specific repair recommendation model instance, as standard information is collected from all customers / use cases / instances and continues to grow (S304).

[0034] Step S304: This step involves archiving and adding new standards information to the existing standards repository set. This new data is used to retrain the generalization model (S301).

[0035] Step S305: Deploy the repair recommendation model instance to the field. By doing this, the model is used to generate repair recommendations that can be used by technicians. In addition, the actual repairs made by the technicians are also recorded in case the recommendations are incorrect. To fine-tune the second AI model, step S305 deploys the second AI model for a period of time to collect model predictions of the second AI model and actual repairs made during the period of time.

[0036] Step S306: This step includes collecting adjustment feedback between the predictions and the actual repairs made in the field. This step includes collecting the model predictions of the second AI model and the actual repairs made during a period of time. The adjustment feedback can be simply true or false, or a composite such as the distance between them in the coding space.

[0037] Step S307: This step determines domain-specific preference attributes from actual repairs. Customers have preferences regarding repair recommendations. For example, for customer 1, the cost of repeated repairs in case of inaccurate predictions may be large, and therefore customer 1 is sensitive to that. For customer 2, it may be a safety issue. Preferences can have multiple attributes, which, when combined, may result in an understanding of whether a particular prediction creates good or bad value for the customer. Step S307 also includes training a reward model using the model predictions, actual repairs made, and preference attributes. Values ​​are encoded by the reward model in the form of a reward score. The reward model may be a deterministic model or a machine learning model. The reward model considers recommendations, actual repairs, and attributes and generates a scalar score.

[0038] Step S308: This step includes generating a reward score using a reward model. Depending on the desired implementation, the error of the reward model may be determined from the difference between the model prediction and the selected actual repair. In such a case, the reward model may be configured to generate a reward if the model prediction is the same as the selected actual repair.

[0039] Step S309: The reward score is used to update the repair recommendation model parameters. The second AI model is then fine-tuned from the rewards generated from the reward model. Iteratively performing steps S305 to S309 results in a model that is personalized and tailored to the preferences of the customer / instance / use case. Thus, this implementation can fine-tune the second AI model to match the preferences of the specific domain to maximize rewards and minimize error. The steps described in S301 to S309 allow the implementation to transfer knowledge from a general domain to a specific domain, allow training of a model even when limited data is available from the specific domain, and allow personalization of the model to the specific domain with reduced error compared to training from scratch, thereby providing a technical improvement over implementations of the related art.

[0040] The current state of the art of data-driven repair recommendation methods relies on the availability of large amounts of data per instantiation, leaving no room for knowledge reuse and no room for personalization. The implementations described herein reduce the dependency on the availability of large amounts of data, thus opening up deployment opportunities for customers who do not have large amounts of data.

[0041] Furthermore, given the modular nature of modeling, the implementation example emphasizes knowledge reuse of standard information across different instantiations / customers. Knowledge reuse can benefit smaller customers by using the knowledge and experience gained by established customers without violating any contractual obligations. On the other hand, knowledge reuse also assists model service providers by scaling their deployments seamlessly across many customers.

[0042] Data-driven repair recommendation models may make mistakes. Existing related art repair recommendation model algorithms lack the ability to adjust for personalization. The implementations described herein introduce a personalization component that is customizable and adapts the model to customer needs and preferences, creating operational value.

[0043] 4 illustrates multiple physical systems networked to a management device according to one implementation. One or more physical systems 421 that manage a customer database with repair code information along with its specific field are communicatively connected to a network 420 (e.g., a local area network (LAN), a wide area network (WAN)) via corresponding network interfaces of the physical systems 421 that are connected to the management device 422. The one or more physical systems 421 may or may not be associated with sensors to identify repair codes for a given task depending on the desired embodiment. The management device 422 manages a database 423 that includes repair codes for general fields. Models may be trained or deployed on the management device 422 on behalf of the physical systems 421 or may be provided to the physical systems 421 for execution depending on the desired implementation.

[0044] Figure 5 illustrates an example computing environment having an example computing device suitable for use in some implementations, such as management device 422 illustrated in Figure 4. The computing device 505 in the computing environment 500 can include one or more processing units, cores, or processors 510, memory 515 (e.g., RAM, ROM, etc.), internal storage 520 (e.g., magnetic, optical, solid state storage, and / or organic), and / or I / O interface 525, any of which may be connected to a communication mechanism or bus 530 to communicate information or may be embedded in the computing device 505. The I / O interface 525 may also be configured to receive images from a camera or provide images to a projector or display, depending on the desired implementation.

[0045] The computing device 505 may be communicatively connected to an input / user interface 535 and an output device / interface 540. Either or both of the input / user interface 535 and the output device / interface 540 may be wired or wireless interfaces and may be removable. The input / user interface 535 may include any physical or virtual device, component, sensor, or interface that can be used to provide input (e.g., buttons, touch screen interface, keyboard, pointing / cursor control, microphone, camera, Braille, motion sensor, optical reader, etc.). The output device / interface 540 may include a display, television, monitor, printer, speaker, Braille, etc. In some implementations, the input / user interface 535 and the output device / interface 540 may be embedded in the computing device 505 or may be physically connected to the computing device 505. In other implementations, other computing devices may function as or provide the functionality of the input / user interface 535 and the output device / interface 540 for the computing device 505.

[0046] Examples of computing devices 505 may include, but are not limited to, highly mobile devices (e.g., smart phones, devices in vehicles and other machines, devices carried by humans and animals, etc.), mobile devices (e.g., tablets, notebooks, laptops, personal computers, portable televisions, radios, etc.), and devices not designed for mobility (e.g., desktop computers, other computers, information kiosks, televisions, radios with one or more embedded and / or connected processors, etc.).

[0047] Computing device 505 may be communicatively connected (e.g., via I / O interface 525) to external storage 545 and a network 550 for communicating with any number of networked components, devices, and systems, including one or more computing devices of the same or different configurations. Computing device 505, or any connected computing device, may function, provide services, or be referred to as a server, client, thin server, general purpose machine, special purpose machine, or another label.

[0048] I / O interface 525 may include, but is not limited to, wired and / or wireless interfaces using any communication or I / O protocol or standard (e.g., Ethernet, 802.11x, Universal System Bus, WiMax, modem, cellular network protocols, etc.) to exchange information with at least all connected components, devices, and networks in computing environment 500. Network 550 may be any network or combination of networks (e.g., the Internet, a local area network, a wide area network, a telephone network, a cellular network, a satellite network, etc.).

[0049] The computing device 505 may use and / or communicate using computer usable or computer readable media, including transitory and non-transitory media. Transitory media include transmission media (e.g., metallic cables, optical fibers), signals, carrier waves, etc. Non-transitory media include magnetic media (e.g., disks and tapes), optical media (e.g., CD ROM, digital video disks, Blu-ray disks), solid-state media (e.g., RAM, ROM, flash memory, solid-state storage), and other non-volatile storage or memory.

[0050] In some examples of computing environments, the computing device 505 may be used to implement techniques, methods, applications, processes, or computer-executable instructions. The computer-executable instructions may be obtained from a transitory medium or stored on and obtained from a non-transitory medium. The executable instructions may be from one or more of any programming language, scripting language, and machine language (e.g., C, C++, C#, Java, Visual Basic, Python, Perl, JavaScript, etc.).

[0051] The processor 510 can run under any operating system (OS) (not shown) in a native or virtual environment. One or more applications can be implemented, including a logic unit 560, an application programming interface (API) unit 565, an input unit 570, an output unit 575, and an inter-unit communication mechanism 595 for the various units to communicate with each other, with the OS, and with other applications (not shown). The described units and elements can be modified in design, function, configuration, or implementation and are not limited to the description provided. The processor 510 can be in the form of a hardware processor, such as a central processing unit (CPU), or in the form of a combination of hardware and software units.

[0052] In some implementations, when information or instructions to execute are received by API unit 565, it may be communicated to one or more other units (e.g., logic unit 560, input unit 570, output unit 575). In some implementations described above, logic unit 560 may be configured to control information flow between units and direct services provided by API unit 565, input unit 570, output unit 575. For example, the flow of one or more processes or implementations may be controlled by logic unit 560 alone or in combination with API unit 565. Input unit 570 may be configured to obtain inputs for calculations described in the implementations, and output unit 575 may be configured to provide outputs based on the calculations described in the implementations.

[0053] The processor 510 may be configured to execute a method or instructions, which may include training a first generative artificial intelligence (AI) model for a general domain, where the first generative AI model is trained using standard information components for the general domain, training a second AI model for a specific domain from the first generative AI model, where training the second AI model is based on use of standard information components, non-standard information components for the specific domain, and available label data for the specific domain, as described with respect to S301 to S305 of FIG. 3, and fine-tuning the second AI model to match preferences for the specific domain to maximize reward and minimize error.

[0054] Depending on a desired implementation, the first generative AI model is configured to output a first repair recommendation for a general domain and the second AI model is configured to output a second repair recommendation for a specific domain. Such first and second repair recommendations may include a series of repair activities, where each repair activity in the series may include a location (e.g., on the object to be repaired) for the repair and a repair action.

[0055] As described with respect to S301 to S303 of FIG. 3, the processor 510 may be configured to execute the above method or instructions, wherein training the first generative AI model for the general domain includes inputting partial information (e.g., a sequence) from known information of a standard information component of the general domain, outputting a prediction of the remaining information of the known information from the partial information using the first generative AI model, and utilizing unsupervised learning to reduce an error between the predicted remaining information and the known information.

[0056] The processor 510 may be configured to execute the above-described methods or instructions, and as described with respect to S301 to S304 of FIG. 3, training a second AI model for a particular domain from the first generative AI model may include performing feature engineering on non-standard information components of the particular domain (e.g., by any feature extraction process), encoding features of the non-standard information components (e.g., from features generated from the feature engineering), combining the first generative AI model and the encoded non-standard information components using available label data to generate a second AI model, and using supervised learning to reduce an error of the second AI model from the available label information.

[0057] The processor 510 may be configured to execute the above-described methods or instructions, and fine-tuning the second AI model may include deploying the second AI model for a period of time, collecting model predictions of the second AI model and actual repairs performed during the period of time, determining preference attributes for a particular area from the actual repairs, training a reward model using the model predictions, the actual repairs performed, and the preference attributes, and fine-tuning the second AI model from rewards generated from the reward model, depending on the desired implementation, an error of the reward model is determined from the difference between the model predictions and the actual repairs, and the reward model is configured to generate rewards for cases where the model predictions are the same as the actual repairs to ensure that the reward model provides rewards in response to "correct" predictions.

[0058] As described herein, the available label data may include a repair code indicative of the system being repaired and the repair action associated with the system being repaired.

[0059] Some portions of the detailed description are presented in terms of algorithms and symbolic representations of operations within a computer. These algorithmic descriptions and symbolic representations are the means used by those skilled in the data processing arts to convey the substance of their innovations to others skilled in the art. An algorithm is a sequence of defined steps leading to a desired end state or result. In one implementation, the steps performed require physical manipulations of tangible quantities to achieve a tangible result.

[0060] Unless otherwise specified, as will be apparent from the description, the description utilizing terms such as "processing," "calculating," "operating," "determining," "displaying," and the like throughout the description will be understood to include the actions and processes of a computer system or other information processing device that manipulates and converts data represented as physical (electronic) quantities in the registers and memory of the computer system into other data similarly represented as physical quantities in the memory or registers of the computer system, or other information storage, transmission, or display device.

[0061] The implementations may also relate to an apparatus for performing the operations of the present specification. The apparatus may be specially constructed for the required purposes, or may include one or more general-purpose computers selectively activated or reconfigured by one or more computer programs. Such computer programs may be stored in a computer-readable medium, such as a computer-readable storage medium or a computer-readable signal medium. The computer-readable storage medium may include tangible media, such as, but not limited to, optical disks, magnetic disks, read-only memories, random access memories, solid-state devices and drives, or any other type of tangible or non-transitory medium suitable for storing electronic information. The computer-readable signal medium may include media such as carrier waves. The algorithms and displays presented herein are not inherently related to any particular computer or other apparatus. The computer programs may include pure software implementations that include instructions to perform the operations of the desired implementation.

[0062] Various general-purpose systems may be used with the programs and modules according to the examples herein, or it may prove convenient to construct more specialized apparatus to perform the desired method steps. Additionally, the implementations are not described with reference to any particular programming language. It will be appreciated that a wide variety of programming languages ​​may be used to implement the techniques of the implementations described herein. Instructions in a programming language may be executed by one or more processing devices, such as a central processing unit (CPU), processor, or controller.

[0063] As is known in the art, the operations described above may be performed by hardware, software, or some combination of software and hardware. Various aspects of the implementations may be implemented using circuits and logic devices (hardware), while other aspects may be implemented using instructions stored on a machine-readable medium (software) that, when executed by a processor, cause the processor to perform a method for performing the implementation of the present application. Furthermore, some implementations of the present application may be performed exclusively in hardware, while other implementations may be performed exclusively in software. Furthermore, the various functions described may be performed in a single unit or may be distributed among several components in any number of ways. When performed by software, the method may be executed by a processor, such as a general purpose computer, based on instructions stored on a computer-readable medium. If desired, the instructions may be stored on the medium in compressed and / or encrypted format.

[0064] Moreover, other implementations of the present application will be apparent to those skilled in the art from a consideration of the specification and practice of the techniques of the present application. Various aspects and / or components of the described implementations may be used alone or in any combination. It is intended that the specification and implementations be considered as examples only, with a true scope and spirit of the present application being indicated by the following claims. [Explanation of symbols]

[0065] 420 Networks, 421 Physical systems, 422 Management equipment, 423 Databases, 500 Computing environments, 505 Computer devices, 510 Processors, 515 Memory

Claims

1. training a first generation artificial intelligence (AI) model for a general domain, the first generation AI model being trained using standard information components for the general domain; training a second AI model for a particular domain from the first generative AI model, where training the second AI model is based on the use of the standard information component, the non-standard information component for the particular domain, and available label data for the particular domain; Fine-tuning the second AI model to match the preferences of the particular domain to maximize rewards and minimize errors. A method comprising:

2. 2. The method of claim 1, wherein the first generative AI model is configured to output first repair recommendations for the general domain and the second AI model is configured to output second repair recommendations for the specific domain.

3. The method of claim 2 , wherein the first repair recommendation and the second repair recommendation comprise a series of repair activities, each repair activity in the series of repair activities comprising a location for repair and a repair action.

4. Training the first generative AI model for the general domain includes: inputting partial information from known information of said standard information component of said general field; outputting a prediction of the remaining information of the known information from the partial information using the first generative AI model; Using unsupervised learning to reduce the error between the predicted residual information and the known information. The method of claim 1 , comprising:

5. Training the second AI model for the particular domain from the first generative AI model includes: performing feature engineering on the non-standard information components of the particular domain; encoding a characteristic of said non-standard information component; combining the first generated AI model with the encoded non-standard information component using the available label data to generate the second AI model; Using supervised learning to reduce an error of the second AI model from the available label information. The method of claim 1 , comprising:

6. Fine-tuning the second AI model includes: deploying the second AI model for a period of time; collecting model predictions of said second AI model and actual repairs performed during said period of time; determining preference attributes for said particular domain from said actual repairs; training a reward model using the model predictions, the actual repairs made, and the preference attributes; and fine-tuning the second AI model from the rewards generated from the reward model. The method of claim 1 , comprising:

7. 7. The method of claim 6, wherein an error of the reward model is determined from a difference between the model prediction and the actual repair, and the reward model is configured to generate a reward for cases where the model prediction is the same as the actual repair.

8. The method of claim 1 , wherein the available label data includes a repair code indicative of a system being repaired and a repair action associated with the system being repaired.

9. A non-transitory computer-readable medium storing instructions for carrying out a process, comprising: The instruction: training a first generation artificial intelligence (AI) model for a general domain, the first generation AI model being trained using standard information components for the general domain; training a second AI model for a particular domain from the first generative AI model, where training the second AI model is based on the use of the standard information component, the non-standard information component for the particular domain, and available label data for the particular domain; and Fine-tuning the second AI model to match the preferences of the particular domain to maximize rewards and minimize errors. A non-transitory computer readable medium comprising:

10. 10. The non-transitory computer-readable medium of claim 9, wherein the first generative AI model is configured to output first repair recommendations for the general domain and the second AI model is configured to output second repair recommendations for the specific domain.

11. 11. The non-transitory computer-readable medium of claim 10, wherein the first repair recommendation and the second repair recommendation comprise a series of repair actions, each repair action in the series of repair actions comprising a location for repair and a repair action.

12. Training the first generative AI model for the general domain includes: inputting partial information from known information of said standard information component of said general field; outputting a prediction of the remaining information of the known information from the partial information using the first generative AI model; Using unsupervised learning to reduce the error between the predicted residual information and the known information.

10. The non-transitory computer readable medium of claim 9, comprising:

13. Training the second AI model for the particular domain from the first generative AI model includes: performing feature engineering on the non-standard information components of the particular domain; encoding a characteristic of said non-standard information component; combining the first generated AI model with the encoded non-standard information component using the available label data to generate the second AI model; Using supervised learning to reduce an error of the second AI model from the available label information.

10. The non-transitory computer readable medium of claim 9, comprising:

14. Fine-tuning the second AI model includes: deploying the second AI model for a period of time; collecting model predictions of said second AI model and actual repairs performed during said period of time; determining preference attributes for said particular domain from said actual repairs; training a reward model using the model predictions, the actual repairs made, and the preference attributes; and fine-tuning the second AI model from the rewards generated from the reward model.

10. The non-transitory computer readable medium of claim 9, comprising:

15. 15. The non-transitory computer-readable medium of claim 14, wherein an error of the reward model is determined from a difference between the model prediction and the actual repair, and the reward model is configured to generate a reward for cases where the model prediction is the same as the actual repair.

16. 10. The non-transitory computer readable medium of claim 9, wherein the available label data includes a repair code indicative of a system to be repaired and a repair action associated with the system to be repaired.

17. training a first generation artificial intelligence (AI) model for a general domain, the first generation AI model being trained using standard information components for the general domain; training a second AI model for a particular domain from the first generative AI model, where training the second AI model is based on the use of the standard information component, the non-standard information component for the particular domain, and available label data for the particular domain; and Fine-tuning the second AI model to match the preferences of the particular domain to maximize rewards and minimize errors. a processor configured to:

13. An apparatus comprising:

Images