Method and device for assisting a technician in data evaluation

Abstract

Method (10) for assisting a technician in evaluating available data, characterized by the following features: - the technician sends a query (11) related to the data to a language model, - the language model converts the query (11) into a program that performs the evaluation (12), and - the language model provides the technician with a representation (14) of the data.

Description

The present invention relates to a method for assisting a technician in data evaluation. The present invention furthermore relates to a corresponding device, a corresponding computer program and a corresponding storage medium.The so-called data science (data science) is a field of sciences that is gaining in importance in many industries, including the automotive industry. In particular, in the detection, management and analysis of large amounts of data generated by vehicles and their various systems, this disclin plays an increasing role. It uses various programming languages and tools to access data, analyze it, and obtain insight therefrom.One of the most common tools for data analysis is the Python programming language with its abundant library of modules and packages for data analysis and processing. Python scripts allow databases to be accessed, processed and analyzed, statistical models to be created, and complex visualizations to be generated that allow engineers and other technicians to identify and interpret patterns and trends in the data.In addition to Python, data scientist uses, within the scope of what is known as business intelligence (BI), auxiliary means such as Tableau and Microsoft PowerBI. These tools allow them to create complex and interactive dashboards that can use participants to look into the data and make informed decisions.However, data analysis is a complex process that requires skills in various fields, including statistics, data processing, and visualization.Large language models (LLM) are a concept of artificial intelligence and, in particular, machine learning. They are based on technologies such as the transformer model and are trained with large amounts of text data to generate human-like texts. LLMs such as GPT-4 from OpenAI are capable of understanding complex semantic and syntactic structures and generating relevant and coherent texts based on these structures.US2023032612A1 presents a multi-modal assistant that uses large language models to support mechanics and technicians in military stores. It combines conversational AI with image segmentation to retrieve and process information from various teachers. The system integrates two separate models - text recognition and recognition - supplemented by custom code and user feedback. Pre- and post-processing are used to convert the model outputs to structured formats such as step-by-step instructions. In addition, the system supplements the Rich Media advice from its databases for comprehensive, user-friendly guidance.US20230350929A1 describes a method and system for generating responses by virtual agents using an LLM that analyzes a knowledge base and intent behind a user request. It involves extraction of topics and topics from historical conversations, which may include interactions between human agents and users to understand the intentions of the users and to give personalized responses. In addition, a computer programmable product is described that includes a non-transitory computer readable storage medium having program code instructions for carrying out this method. The knowledge base is modeled in a graph-based structure that visually represents the relationships between different concepts to enable more efficient information navigation and retrieval. This graph-based structure is intended to support the retrieval and identification of information relevant to the extracted topics.EP4220434A1 discloses systems and methods for enhancing a large language model to generate suggestions for automatically completing queries over an interactive user interface. This involves continuous training of the language model for software-related queries using masking techniques. In addition, the system dynamically updates the model based on user interactions with the automatic completion suggestions. A unique function is a range of results that asynchronously indicates the effects of selected auto-completion suggestions on the underlying code. Although the presented example relates to SQL, the system may be adapted to various computer languages.US11769017B1 discloses a system that employs an LLM to generate natural language abstracts in response to queries, wherein the LLM processes both the content of the query and additional related content. A context engine determines the query context based on factors such as previous queries, user profile data, and device location. The system refines the LLM inputs from user interactions and retrieved data to adapt the summary to its individual preferences without additional user effort. It takes into account various types of queries, including related, recently occurring, and implied queries derived from the context or profile data. Finally, a rendering engine on the client device presents these sums acoustically or visually via speakers, displays, or projectors.US20230244678A1 discloses a system and method for improving prediction accuracy of a machine learning-based virtual call agent. In this case, an authenticated context exus is calculated from the embeddings of user requests, said context exus containing a categoric and a subcategoric parameter. From these parameters search logic parameters are created to perform a contextual search in a structured corpus of data. A response to the user's request is then generated by extracting a candidate from the structured data. In addition, the system may output the values of the context nexus parameters as metadata that may be used in subsequent searches within the same query session.US20230350936A1 discloses a query processing system that receives a query comprising a string of input tokens and a data element of a different modality to generate an output token string. It allows for a series of queries where subsequent inputs may include the output strings from previous interactions, facilitating coherent dialog over a particular data item. In addition, the system may output executable code in response to a query input. The kernel of the system consists of token processing layers that operate as a speech model trained on a comprehensive database of natural language data. This model is designed to provide relevant responses to the input token strings.WO2020160264A1 discloses a system for organization and access of data by a so-called feature graph. Nodes in this graph represent measurable entities such as objects or features, while edges identify the statistical associations between those nodes based on data retrieved from different sources. The graph links to data sets that support the statistical associations or measure the variables it represents. Mathematical, linguistic, and visual techniques are employed to convey the integrity and reliability of the information underlying the statistical associations. In addition, the system employs techniques such as optical character recognition, image processing, and natural language processing to enhance the processing of the collected data.One problem is that given the complexity and skill required for data analysis, the cost of setting and training data sciences is considerable. In addition, the BI tools that use data scientist are often complicated to manage and adapt, especially when considering changes in the data structure. The data generated by vehicles and their systems may be very different and may change over time. Therefore, the tools used to analyze these data must be flexible enough to adapt to these changes. However, the complexity of managing and adapting these tools may result in data analysis becoming inefficient and time consuming.Another problem is that the existing data analysis pipelines are often not flexible enough to accommodate additional analysis requests. Data scientist therefore typically must go through a series of pre-fabricated analysis processes that are not always able to meet specific or unexpected analysis requests. This can result in having to spend valuable time and resources to adapt the pipelines or re-create them to meet the specific requirements.Finally, it may also be a challenge to present the results of the data analysis in a manner that is readily understood and interpreted by technicians and other participants. The creation of dashboards and visualizations often requires additional time and skill, and even then, it may be difficult to display the results in a manner that is comprehensible and meaningful to non-experts. This can impair the efficiency and value of the data analysis and lead to the overlooking or misunderstanding of important findings.The described problem is solved by a method for supporting a technician during data evaluation, a corresponding device, a corresponding computer program and a corresponding storage medium according to the independent claims.This approach has the advantage that it opens up new possibilities for the evaluation of data in a multiplicity of scenarios. The LLM-based method allows technicians to make the data analysis efficient and effective by making the natural language queries that understand and act accordingly on the model.By converting the request into a programmatic code that performs the requested analysis, the LLM can establish a direct and seamless connection between the technician's request and the generated data. This can make the process both time-saving and resource-saving, since no additional steps or tools are required to interpret the request or write the code.In addition, the LLM may provide a representation of the data that is matched to the technician's request. This could be in the form of a table, diagram, or other visual representation that allows the technician to understand the data directly and intuitively. This function can eliminate the need to use separate tools for data visualization or rely on the capabilities of a data scientist to interpret and present the data.In cases where the technician does not accept the presentation of the data, the LLM may also be able to further specify and adjust the request to meet the technician's specific requirements or preferences.The integration of LLMs into the data analysis may also improve the flexibility and adaptability of the analysis pipelines. By having the LLM understand and implement complex requests, the system can simplify the process of adapting and managing the pipelines and increase the efficiency of data analysis.Finally, the use of LLMs in data analysis could reduce the need for data science managers and minimize the cost and challenges associated with their recruitment, training, and management.Further advantageous embodiments of the invention are specified in the dependent patent claims.Brief Description of the DrawingsFIG. 1 shows a conventional data science flow. FIG. 2 shows a data scientific procedure based on LLMs. FIG. 3 shows an illustrative example of the graphical output of a computer program according to the invention.FIG. 1 illustrates a conventional data science flow ( 10). A technician makes a request (11), for example for the normal charging duration of a specific vehicle model. In the analysis phase (12), the data scientist must decide which features are needed for the examination, which statistical methods should be applied, how the results could be visualized and which data source is to be used. The implementation (13) then requires writing python code, adding features and examining the data. Finally, the data scientist provides a representation (14) of the data in the form of refined visualizations or a dashboard. The technician checks this representation and either accepts it or rejects it (15). In the event of rejection, the data scientist must adapt and repeat the process.Figure 2, on the other hand, shows a data scientific flow based on LLMs as proposed in the present invention. Here too, the technician makes a request (11). However, the assistant analyst, based on the LLM, performs the tasks of feature selection, method decision, execution, and visualization. The LLM generates the code, performs the analysis, and creates a visual representation of the data. If the representation is rejected by the technician, the LLM interacts with the technician to make a more specific request and adjust the process accordingly.Consider the following example of tabular data relating to charging cycles for a fleet of electric vehicles:12022Cayenne50752510022023Taycancan25352615032022Taycancan1580351542022Cayenne2580152552022Taycan3311001850.....................10000002023Taycancan15301545Here, the columns of the header designate the following features: 1. "ChargingCycleld"-the unique identification number of each load cycle. 2. "Year" - the year in which the charge cycle took place. 3. "CarType"- the type of vehicle that has been loaded. 4. "SoC_Start(pc)"- the starting state of the battery charge in percent. 5. "SoC_End (pc)"- the final state of battery charge in percent. 6."Temp (C)" - The temperature during the charge cycle in degrees Celsius. 7. "Current (A)"- The current used during the charging cycle in amperes.A user could request, such as to determine the distribution of temperature for model "Taycan" vehicles during AC charging at the home charging port (at a charging current below 50 A) when more than half of the battery capacity is charged ("SoC_End-SoC_Start>50"). This is just one example of the type of requests that the LLM can handle and it shows how it is able to put the request in context and provide appropriate analysis and representation of the data.After issuing such a request, the LLM performs a series of operations. It would first filter the data to select only the charging cycles that meet the criteria indicated: the vehicle is of the "Taycan" type, the charging is at home with alternating current at a current of less than 50 A, and the state of charge at the end of the cycle is more than 50 percentage points higher than at the beginning.To do this, it selects the relevant features from the data, calculates the desired statistics, and generates the following code that implements the request: > import pandas as pd > df = pd.read_csv('gpt_sample.csv') > df = df [(df.CarType = = = 'Taycan') & (df.SoC_End - df.SoC_Start > 50) & (df. Current<50)]>ax= df. Temperature.plot.hist(bins=30 title='Distribution of temperature of Taycan AC home charging')> ax.set_xlabel("temperature", fontsize=12)> ax.set_ylabel('# of charging cycles', fontsize=12)This program draws the histogram shown in Figure 3 which shows the frequency of the various temperatures measured. If the user wishes to change the representation of the results, for example by changing the title, labels, colors or graph type, the LLM can understand and implement these requirements. In addition, the LLM may handle other requests, such as more accurately reviewing the data (e.g., using only data from 2022) or merging the data with other data sources.Moreover, the LLM may understand the context of the data, even though the table is in a different form or different column designations, physical units, or variable types are used. For example, the LLM concludes from the context that the model "Taycan33" falls below the generic term "Taycan".A user's inquiry about the maximum difference in the state of charge for the model results in the following process chain: > Ent new AgentEx chain,... That: To find the maximum difference in SoC (State of Charge) for Taycan, we need to filter the data frame for rows when CarType is 'Taycan', Then calculate the difference action input: df[df['CarType']== 'Taycan']['SoC_End'].subtract(df[df['CarType']= 'Taycan']['SoC_Start']).max () Observation: 69 Throught:The maximum difference in SoC for Taycan is 69 Final Answer: 69> Finish chain. {'input': 'whatman is the maximum difference in soc for taycan?. Please use action: python_repl_ast.''output': '69'}As representation (14), the LLM in this case provides the number 69.Further exemplary requests are as follows: request=""Bar plot me a graph of currents that are bigger thann 280"""'request="""Select the top 5 max 'Temperature (C)' rows. Bar plot their current (A). Use blue, red and grey colors.'''' request = ''''''''''''''''''''''''''''''''List of reference characters10 Method 11 Request 12 Evaluation 13 Implementation 14 Representation 15 DecreaseReferences included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Patent Literature citedUS 2023032612A1

[0007] US 20230350929A1

[0008] EP 4220434A1

[0009] US 11769017B1

[0010] US 20230244678A1

[0011] US 20230350936A1

[0012] WO 2020160264A1

[0013]

Claims

Method (10) for supporting a technician during an evaluation of present data, characterized bythe following features: - the technician directs a request (11) relating to the data to a speech model, - the speech model converts the request (11) into a program which carries out the evaluation (12), and - the speech model provides the technician with a representation (14) of the data.Method according to Claim 1, characterized bythe following feature: - if the technician refuses to decrease (15) the representation (14), the request (11) is made concrete to the speech model.Method according to claim 1 or 2, characterized byat least one of the following features: - the speech model is a generative transformer and - the speech model is pre-trained with reinforcement learning by human feedback.Method according to one of Claims 1 to 3, characterized bythe following features: - the data have headers, - the language model uses the headers in order to place the request (11) in a context, and - the language model gives the technician instructions relating to the context.Method according to one of Claims 1 to 4, characterized bythe following features: - the context is managed in a volatile memory of the data model, and - the request (11) is stored in the memory for a later reference.Method according to one of Claims 1 to 5, characterized bythe following feature: - the request (11), evaluation (12) and display (14) of the data takes place in an automated process chain.Method according to claim 6, characterised byat least one of the following features: - the process chain comprises a software agent for implementing the request (11), and - the process chain comprises an interpreter for executing the program.Device, characterized bythe following features: - the device is set up to carry out a method according to one of Claims 1 to 7.Computer program which is configured to carry out all the steps of a method according to one of Claims 1 to 7.A machine readable storage medium having stored thereon a computer program according to claim 9.

Images