Computer-implemented method, and system

Abstract

A computer-implemented method in which a digital identity (10), in particular that of an organization, is created and / or improved, in particular using a machine learning model (40; 42) and / or for use with a machine learning model (40; 42).

Description

[0001] NN 17624 WO

[0002] December 23, 2025

[0003] Computer-implemented method, system

[0004] State of the art

[0005] The invention relates to a computer-implemented method according to claim 1 and a system according to claim 20.

[0006] Computer-implemented methods for data processing with patent specifications are already known from the prior art.

[0007] The object of the invention is, in particular, to provide a generic method with advantageous properties with regard to computational efficiency and / or output data quality. This object is achieved according to the invention by the features of claims 1 and 20, while advantageous embodiments and further developments of the invention can be found in the dependent claims.

[0008] Advantages of the invention

[0009] According to the invention, a computer-implemented method is proposed in which a digital identity, in particular of a company, is created and / or improved, especially using artificial intelligence and / or for use with artificial intelligence.

[0010] The computer-implemented method according to the invention advantageously minimizes computational efficiency, since the digital identity enables efficient further data processing. Furthermore, a digital identity, preferably technical, can advantageously represent a company's interest in a, preferably technical, context.

[0011] Autonomous, further processing can be reflected. Furthermore, a digital identity can advantageously capture a company's collective experience, particularly across individuals, departments, and / or projects, and establish best practices across all areas. Additionally, creating a digital identity can advantageously provide company-specific, context-related caching, especially for transformer-based artificial intelligence systems. This leads to significantly reduced computing time and / or optimized system scalability, at least compared to systems without caching. In particular, using artificial intelligence to create and / or improve the digital identity allows for the provision of data of the highest possible quality.

[0012] Preferably, the digital identity describes at least one attribute of an entity, most preferably at least one company, particularly a manufacturing and / or development company. However, individual persons, particularly manufacturing and / or development companies, or groups of persons are also conceivable as entities. Preferably, the digital identity comprises a structured collection of gathered data, in particular a cluster dataset, which preferably contains data of different types. Most preferably, a technical digital identity is created and / or improved, where the attribute describes a technical behavior of the entity, particularly the company. In principle, other attributes are also conceivable, such as social behavior, strategic behavior, and / or economic behavior, as well as a combination of different attributes."Behavior" should be understood to mean at least a temporal, relevance-dynamic, categorical and / or spatial development of the property. NN 17624 WO.

[0013] Particularly preferred is the provision of selective data processing via the digital identity. Preferably, the digital identity provides selective processing, at least on the basis of an evaluation, preferably temporal, relevance-dynamic, categorical and / or spatial, preferably by a user and / or a machine learning model, and is then made available, in particular, to a user interface and / or for use with the machine learning model.

[0014] Preferably, the digital identity comprises at least one database, in particular a relational database, graph-based and / or vector-based database, wherein technical, economic and / or social data are preferably stored in the database. In an alternative embodiment, the digital identity comprises a data mask, in particular factors which, when applied to any database of the company, evaluate the data temporally, dynamically in terms of relevance, categorically and / or geographically.

[0015] Preferably, at least the input data, preferably company-related and particularly preferably describing the technical behavior, is entered via the system's user interface. Furthermore, inputs, in particular questions, ratings, and / or statements, preferably relating to the entity, especially the company, projects, and / or the digital identity, are preferably entered by the user via the user interface. Preferably, the digital identity is created and / or enhanced using the input data. Preferably, output data is generated using the digital identity, which is then output, in particular, via the user interface.

[0016] Preferably, when creating the digital identity, all data constituting the digital identity are analyzed, processed, and / or transformed into a structured form, particularly at least partially using machine learning models and / or manually. This preferably involves data from various sources, such as company documents, projects, and experiences.

[0017] and processes, extracted and semantically linked and / or structured. The machine learning model is preferably used to capture relationships between the data, eliminate redundant information, and generate a coherent representation of the digital identity.

[0018] Preferably, when improving the digital identity using a machine learning model, the quality and / or currency of the data constituting the digital identity is monitored and processed, preferably continuously. In doing so, patterns in the existing data are preferably recognized, new data describing the digital identity is added, and / or outdated and / or inconsistent data is removed, particularly using the machine learning model. Furthermore, recommendations for the structure and / or processing of data can preferably be generated using the machine learning model, preferably via the user interface, in order to adapt the digital identity specifically to new requirements or developments.

[0019] Preferably, when using the digital identity with the machine learning model, access to relevant information is dynamically optimized and adapted to specific questions or contexts. In particular, the digital identity is provided to the machine learning model, at least partially, as context, especially to answer a user request via the user interface.

[0020] The digital identity is particularly preferred, preferably after its creation, to be used throughout with the machine learning model and improved with the results of the machine learning model.

[0021] A "machine learning model" is understood to be, at least in essence, a sequence of computational operations on the computing device that preferably cause the computer to manipulate data and / or control system components. The machine learning model can be based on various well-known principles; for example, the machine learning model can be a CNN (Convoluted Neural Network) algorithm.

[0022] The machine learning model may be an LSTM (Long Short-Term Memory Network) algorithm, an autoencoder algorithm, an SVM (Support Vector Machine) algorithm, a random forest algorithm, or another known machine learning algorithm. A particularly preferred machine learning model is at least a transformer-based neural network. A particularly preferred machine learning model, especially a transformer-based neural network, is a text processing model, particularly a text generation model, preferably designed for natural language processing (NLP) and / or other sequences. For each of the above examples, there are open-source libraries and frameworks that enable a person skilled in the art to implement the machine learning model.

[0023] In particular, the digital identity is created, improved, and / or used with a system comprising at least one computing device. A "computing device" is understood to be, in particular, a device with information input, information processing, and information output. Advantageously, the computing device comprises at least one processor, a memory unit, a user interface comprising input and output means, further electrical components, an operating program, control routines, and / or calculation routines. The memory unit of the computing device preferably includes a computer program with program code for executing the computer-implemented method. The computer-implemented method, in particular the computer program with program code, is most preferably executed by the processor. Preferably, the computing device is at least part of a computer.Preferably, the components of the computing device are arranged on a common circuit board and / or advantageously in a common housing. Alternatively, the computing device can also be designed as a distributed, in particular virtual, computing device, such as a cloud. NN 17624 WO.

[0024] The term "intended" should be understood to mean specifically programmed, designed, and / or equipped. The fact that an object is intended for a specific function should be understood to mean, in particular, that the object fulfills and / or executes this specific function in at least one application and / or operating state.

[0025] Furthermore, it is proposed that a dedicated, preferably technical, document sequence and / or image sequence be used to create and / or improve the digital identity. This allows the digital identity to be more precisely tailored to the user's, and especially the company's, technical expertise or innovation, resulting in a more robust representation. Preferably, the document sequence and / or image sequence describes a technical product, system, and / or process.

[0026] Preferably, the document sequence comprises a logically and / or functionally structured sequence of data from at least one document that describes, in particular, technical characteristics of a product, system and / or process, such as, but not limited to, intellectual property documents, in particular patent specifications, utility model specifications, published designs and / or design specifications, technical verification documents, in particular test reports, protocols, in particular experimental procedures, measurement data and / or software documentation, and / or technological descriptions, in particular specification documents, circuit diagrams, system architectures, process descriptions and / or scientific reports.Preferably, the image sequence comprises a temporally, spatially and / or thematically ordered sequence of at least one visual data set, in particular consisting of static and / or dynamic image data that describe technical properties of a product, system and / or process, such as, but not limited to, technical visualizations, in particular CAD drawings, exploded views, drawings of intellectual property documents, FEM simulations and / or structural models, NN 17624 WO.

[0027] Photographic and graphic representations, in particular images of prototypes and technical developments, test facilities, microscopic views and / or production steps, dynamic image data sets, in particular filmed and / or animated process sequences, video sequences of products and / or experiments and / or time-resolved simulations.

[0028] Furthermore, it is proposed that the document sequence and / or image sequence constitute at least one proprietary intellectual property document. This allows the digital identity to be tailored as individually as possible to the user, particularly the company, which is advantageous because the company's proprietary intellectual property documents represent a significant portion of its relevant developments over time. Furthermore, the structured language and structure of the proprietary intellectual property documents can be efficiently processed by the machine learning model and / or prepared for the machine learning model. Preferably, the proprietary intellectual property document is a utility model specification, a published patent application, a design specification, and / or, most preferably, a patent specification, in particular a patent application and / or a patent.The term "own intellectual property document" means that the intellectual property document is filed, registered, and / or owned by the user, in particular the company, itself. Preferably, the digital identity comprises at least a substantial portion, and preferably the entirety, of the user's, in particular the company's, intellectual property portfolio, preferably including applications, grants, expired, and / or abandoned intellectual property documents. The term "substantial portion" means, in particular, at least 50%, preferably at least 65%, advantageously at least 75%, most preferably at least 85%, and most advantageously at least 95% of a total value. Intellectual property documents, preferably the user's own intellectual property document, and most preferably the user's own intellectual property portfolio are particularly preferred, especially in the creation and / or [NN 17624 WO].

[0029] Improvement is used as a basic building block, in particular so that at least in essence the further data are directly and / or indirectly linked to the own intellectual property document, especially to the intellectual property portfolio.

[0030] Furthermore, it is proposed that intellectual property documents researched in relation to the company's own intellectual property document be used to create and / or improve the digital identity. This allows the digital identity to be technically enriched by incorporating researched intellectual property documents, thereby creating a comprehensive picture of the relevant technical contexts and the innovation environment. Preferably, the intellectual property documents researched in relation to the company's own intellectual property document are identified based on the content, claims, and / or classifications of the company's own intellectual property documents.Preferably, the intellectual property documents searched are all intellectual property documents that have been searched for relating to the applicant's own intellectual property document and / or to the product, system, and / or process contained in the intellectual property document. This includes, in particular, searches conducted by at least one examining division in examination proceedings concerning the applicant's own intellectual property document and / or by a competitor, for example, in opposition proceedings, invalidity proceedings, infringement proceedings, or similar proceedings, and / or searches conducted on the product, system, and / or process prior to filing the intellectual property application. It is particularly preferred that at least some of the intellectual property documents, especially those relating to competitors and / or specific technology areas, are integrated into the digital identity and stored cyclically, for example, monthly or quarterly, preferably automatically, such as via an API interface with a patent office.

[0031] Furthermore, it is proposed that for the creation and / or improvement of the digital identity, intellectual property documents and at least one external intellectual property classification, in particular an international NN 17624 WO, should be used.

[0032] Intellectual property classification can be used. This can advantageously achieve stronger networking and technological classification of the user's innovation, particularly that of the company. Furthermore, the structured storage advantageously enables context-specific information such as claims, abstracts, or definitions to be selectively retrieved and integrated into dynamic contexts for the machine learning model. Preferably, the intellectual property classification is a standardized and / or systematic scheme intended, in particular, for the content-related classification of the product and / or process, whereby the intellectual property documents are assigned to thematic categories, preferably based on technical features, fields of application, and / or technologies. An International Patent Classification (IPC) and / or a Common Patent Classification (CPC) is particularly preferred.

[0033] Furthermore, it is proposed that an evaluation of features from a feature breakdown of patent claims in an intellectual property document be used to create and / or improve the digital identity. This advantageously enables the linking of several different intellectual property documents of the digital identity via relevantly evaluated features, tailored to the individual needs of the user. Preferably, the feature breakdown is a systematic division, in particular an ontology, of a patent claim into individual technical features, especially individual technical terms, which preferably each define specific technical functions or properties of the claimed subject matter.

[0034] Preferably, the evaluation of the features is an analysis and / or weighting of the features of the feature outline, such as with regard to relevance, novelty, inventive step and / or technical and / or legal significance for the user, wherein the evaluation is in particular automated, preferably with the machine learning model to improve the digital identity, and / or preferably manually, preferably via the NN 17624 WO

[0035] User interface. In particular, the evaluation of the characteristics of the intellectual property documents, which are preferably added to the digital identity during the creation and / or improvement of the digital identity, is particularly preferred to determine a probability of relevance and, in particular, a weighting of a further intellectual property document, which is preferably added to the digital identity, preferably by means of a relevance comparison.

[0036] Furthermore, it is proposed that a task definition and / or a description of benefits from a proprietary intellectual property document be used to create and / or improve the digital identity. This allows the digital identity to be enriched in a technically sound manner by linking intellectual property documents across tasks and / or benefits. Preferably, the task definition of the proprietary intellectual property document includes the technical problem that the product and / or process claimed in the document is intended to solve. Preferably, the description of benefits in the proprietary intellectual property document includes the technical advantages that the claimed product and / or process is intended to achieve.Preferably, the task definition and / or benefit description is used to determine the probability of relevance and, in particular, the weighting of a further intellectual property document, which is preferably added to the digital identity, preferably with relevance matching. In this process, relevance and, in particular, a weighting are determined based on the similarity of the task definition and / or benefit description.

[0037] Preferably, a relevance comparison determines the relevance and / or weighting of a relevant text element, preferably a feature, from the task description and / or the benefit description, of the further intellectual property document that supplements the digital identity, particularly based on the intellectual property documents of the digital identity. Preferably, in one process step, the text elements relevant for the relevance comparison are extracted from the further intellectual property document, preferably using natural language processing. Preferably, in a further process step, the relevant text elements of the further intellectual property document are numbered, preferably vectorized, for example, using TF-IDF (Term Frequency-Inverse Document Frequency), word embeddings, and / or other methods known to a person skilled in the art.Preferably, in a further process step, the similarity, and in particular the relevance, of the relevant text elements of the further intellectual property document is determined using a comparison algorithm, for example by means of cosine similarity and / or preferably with a machine learning model, in particular with a support vector machine (SVM), gradient boosting model, for example XGBoost, and / or neural networks. Preferably, based on the similarity, and in particular the relevance, the relevant text elements of the further intellectual property document are weighted and added to the digital identity.In an alternative design, the relevant text modules of the intellectual property documents of the digital identity are semantically compared with the further intellectual property document, and in particular the relevance of the further intellectual property document is determined based on a matching number of words with the relevant text modules of the digital identity, whereby a weighting is carried out in particular.

[0038] Furthermore, it is proposed that word matching between several technical document sequences and / or image sequences be used when creating and / or improving the digital identity. This allows the digital identity to advantageously capture the relevance of specific key technical technologies or innovation areas from other intellectual property documents added during the creation and / or improvement of the digital identity. Furthermore, it is advantageously possible to [NN 17624 WO]

[0039] In particular, compared to similarity calculations using a machine learning model, computational efficiency is increased. Specifically, word matching involves comparing text-based content between a proprietary, preferably technical, document sequence and another, preferably technical, document sequence, which is preferably input into the digital identity. Similarities, differences, and / or correlations are identified at the level of individual terms, word groups, or phrases. Particularly preferably, the word matching process includes a statistical quantitative evaluation of word frequencies in the document sequences being compared. This evaluation preferably determines the relevance, and in particular the weighting, of the other document sequence in the creation and / or improvement of the digital identity, depending on the degree of similarity.

[0040] Furthermore, it is proposed that company-specific keywords be used to create and / or improve the digital identity. This can advantageously enable a targeted and authentic presentation of the company's technical capabilities and orientation, allowing the digital identity to be individually tailored to the company. Preferably, company-specific keywords include terms, phrases, and / or technical terminology that are particularly characteristic of a company's innovation and technology fields, such as industries, product classes, or project classes. Ideally, the keywords should also reflect the company's strategic orientation, such as corporate goals, new technology platforms, future technology directions, economic orientations, and / or social orientations.In particular, company-specific keywords are defined manually via the user interface. Additionally or alternatively, company-specific keywords are determined automatically, taking from the company's own document sequences, especially intellectual property documents, documentation, research reports, and technical documents (NN 17624 WO).

[0041] Specifications, product documentation, and / or development protocols are derived, with frequently used terms being captured as keywords. Preferably, the company-specific keywords are assigned to the document sequences and / or image sequences, preferably technical ones, and one or more of the keywords are added to each additional document sequence and / or image sequence, preferably technical ones.

[0042] In particular, historical development is recorded for a category of keywords, such as industries, product classes, project classes, company goals, new technology platforms, future technology orientations, economic orientations, and / or social orientations. The historical development of at least one category of keywords is especially preferred as an input parameter for the machine learning model in the creation, use, and / or utilization of the digital identity.

[0043] Furthermore, it is proposed that technical sensor data be used, in particular for the creation and / or improvement of the digital identity and / or during its use. This can advantageously increase the authenticity and meaningfulness of the digital identity and dynamically adapt it to the entity's actual technical capabilities and states. Specifically, technical sensor data from the entity's processes, machines, and / or devices will be recorded and collected, describing, in particular, an operating mode and / or structure, preferably technical characteristics.Preferably, the technical sensor data includes measured values ​​acquired by sensors to monitor physical, chemical and / or technical parameters, for example, physical data such as temperature, pressure, acceleration or vibration, chemical data such as pH values, concentrations or gas compositions, and technical condition data such as position, rotational speed, current and / or wear parameters and / or particularly preferably NN 17624 WO.

[0044] Recorded image and / or video data. Preferably, the technical sensor data is evaluated using data processing methods such as data analysis, pattern recognition, and / or machine learning. In particular, the technical sensor data is supplemented with the digital identity data and, especially, ontologically assigned. It is particularly preferred that, when using the digital identity to process input data, the technical sensor data is linked with the data from document sequences and / or image sequences. It is particularly preferred that, when creating and / or improving the digital identity, the sensor data, especially as unstructured data, is manually and / or automatically, especially by a machine learning model, preferably ontologically structured and preferably assigned at least to company-specific keywords.

[0045] In a further implementation, sensor data from individuals within the entity, particularly the company, is collected, specifically audio and video recordings and / or user behavior analyses, such as keystrokes, user movements, visited websites and / or screen recordings. Video conferences and / or customer conversations are particularly favored for recording and / or transcribing as sensor data.

[0046] Furthermore, it is proposed that the sensor data be recorded using a mobile device, in particular a smartphone. This advantageously provides a cost-effective, compact, and easy-to-use solution for acquiring technical sensor data, especially without the need for specialized measuring instruments. Alternatively, the mobile device can be a tablet, camera, and / or similar device. Preferably, the sensor data is processed and visualized directly on the mobile device and / or transmitted to the digital identity via an interface. Preferably, the sensors integrated into the mobile device, for example, accelerometers and gyroscopes for measuring movements, are used.

[0047] Positions, cameras for optically capturing technical parameters such as color measurements or visual analyses, and microphones for recording acoustic signals, for example for vibration or noise measurements, capture sensor data. In particular, it is conceivable that external sensors could be connected to the mobile device via interfaces such as Bluetooth or USB to capture further technical sensor data.

[0048] Furthermore, it is proposed that the mobile device be used for digital identity-specific analysis. This advantageously enables fast and privacy-friendly analysis, as processing takes place directly at the point of data collection. The specific analysis, particularly the processing, analysis, and / or interpretation of the collected sensor data in connection with the digital identity, is preferred. Specifically, the analysis is performed using hardware and / or software functions of the mobile device, for example, through data preprocessing such as raw data filtering, noise reduction, or measurement normalization. It is particularly preferred that the sensor data be segmented and / or classified using the mobile device, especially based on company-internal keywords.

[0049] Furthermore, it is proposed that the digital identity be used to assign company-specific keywords to document and / or image sequences. This would allow for a more precise organization and categorization of document sequences, enabling a faster assessment of their technical relevance. The assignment of these keywords to the document and / or image sequences is preferably carried out using automated analysis with text analysis and / or natural language processing (NLP) algorithms, which recognize relevant keywords within the document and / or image sequences. Additionally, synonyms and / or related terms are preferably identified based on company-specific word lists and / or ontologies.

[0050] In particular, information stored in the digital identity, NN 17624 WO

[0051] Such as preferably technological expertise, patent specifications and / or research areas, are used to assess the relevance of the keywords in the document. Furthermore, the identified keywords are preferably linked automatically and / or manually to the corresponding sections and / or categories within the document sequence and / or image sequence.

[0052] Furthermore, it is proposed that the digital identity be used for technical development support, onboarding, and / or monitoring optimization. This will accelerate development processes, streamline onboarding processes, and allow for more precise adjustments to technical monitoring. Technical development support is particularly enhanced by targeted access to relevant technical data. Ideally, the digital identity is used as a context within a machine learning model to generate output that supports development, onboarding, and / or monitoring. Preferably, automated recommendations regarding technical resources, technologies, and / or solutions are provided based on the data stored in the digital identity.Preferably, the digital identity is integrated into development platforms to identify technical skills for team building and / or solving specific problems. In particular, the digital identity is used to provide new employees and / or partners with access to relevant intellectual property documents, technical specifications, and / or training materials tailored to their stored digital skills. Personalized onboarding to technologies, processes, and / or systems is especially preferred, based on the specific skills and experience stored in the digital identity. Monitoring optimization is preferably enabled through precise adaptation and monitoring of processes based on the specific technical requirements and parameters of the digital identity. For example, NN 17624 WO.

[0053] Relevant sensor and / or process data are analyzed, and optimization potential is identified. Real-time adjustments and / or warnings based on stored technical parameters and expertise are given particular priority.

[0054] Furthermore, it is proposed that a task definition and / or a benefit description be used in development support. This can advantageously enable a structured and efficient approach to development projects. Development resources are advantageously prioritized and deployed strategically based on the relevance of the task definition and / or the expected benefits. In particular, starting from a task definition and / or a benefit description, which is entered primarily via the user interface, solutions for development are comprehensively identified across intellectual property documents stored in the digital identity, preferably the company's own intellectual property documents, which exhibit at least essentially the same task definition and / or benefit description.

[0055] Furthermore, it is proposed that a comparison of a user's own task definition and / or benefit description with a digital task definition and / or benefit description of the digital identity be used, particularly for technical development support. This can advantageously enable a precise alignment of development processes with documented competencies, technologies, and solutions stored in the digital identity. Preferably, the user's own task definition and / or benefit description is entered into the user interface.

[0056] Preferably, the comparison involves comparing the user's own task definition and / or benefit description, preferably of document sequences and / or image sequences, particularly preferably of intellectual property documents, with the tasks definition and / or benefit descriptions of the digital identity, whereby solutions from the document sequences and / or image sequences, preferably NN 17624 WO

[0057] Intellectual property documents will be issued, depending on the degree of conformity.

[0058] The matching process is preferably carried out using automated analysis and comparison, in which algorithms analyze the terms, objectives, and technical parameters of both the company's own and the digital tasks, as well as descriptions of benefits. This process preferably identifies similarities, complementary elements, and / or potential conflicts. In particular, the matching is used to identify synergies, for example, when the digital benefit description offers additional approaches and / or solutions for the company's own task. The matching process particularly contributes to the optimization of development goals by integrating relevant information from the digital identity.

[0059] Development goals are specified and / or adapted to ensure targeted and innovative support for technical development processes. Preferably, the comparison of the client's own task definition and / or benefit description is carried out according to the relevance assessment.

[0060] Furthermore, it is proposed that results from development work, onboarding processes, and / or monitoring optimizations be used to create and / or improve the digital identity. This allows the digital identity to be enriched with current data obtained from these processes, thereby increasing its relevance and timeliness. Additionally, linking the digital identity to real-world results, particularly in conjunction with intellectual property documents, can create an authentic and verifiable basis for demonstrating technical expertise and progress. Specifically, results from development work include documented findings such as technological advancements, new process parameters, developed prototypes, and / or optimized product solutions. Innovations and / or technical improvements relevant to intellectual property rights are particularly preferred.

[0061] These results are achieved particularly within the context of development work and incorporate measurement data and / or simulation results. Preferably, results from onboarding processes include key competencies identified, training outcomes, and / or learned technologies, which are documented, especially during onboarding. Preferably, the onboarding processes utilize process optimizations and / or adjustments initiated by the integration of new employees and / or partners.

[0062] Results from monitoring optimizations include, in particular, monitoring parameters and / or algorithms developed within the framework of optimizing monitoring systems, especially for monitoring intellectual property documents. Preferably, the results from development work, onboarding processes, and / or monitoring optimizations are analyzed, structured, and integrated into the digital identity.

[0063] Furthermore, it is proposed that a vector representation of the digital identity be created and / or modified during its creation and / or improvement. This advantageously enables a precise and structured representation of the digital identity that can be easily analyzed, visualized, and compared. Moreover, the mathematical representation of the vector representation can advantageously facilitate efficient computations and comparisons with input vectors, particularly by a user, for example, through vector space analyses or similarity searches. Additionally, a pre-encoding method can reduce the computation time for transformer models, as the data does not need to be re-encoded with each query. Preferably, the vector representation of the digital identity is processed with the transformer as a machine learning model.In particular, the digital identity, at least partially, preferably the document sequences and / or image sequences of the digital identity, especially in raw text form and / or in latent representations, is stored in the vector image.

[0064] Latent representations with an encoder are particularly preferred, NN 17624 WO

[0065] Preferably, transformer encoders, for example using PatentBERT and / or SciBERT, are generated. Preferably, particularly for the creation and / or processing of the vector image, the document sequences and / or image sequences of the digital identity are decomposed into their structural components, such as title, abstract, background of the invention, problem statement, description of advantages, detailed description, technical descriptions and / or definitions of products and / or processes and / or claims. Preferably, the structural components are decomposed into subword units, for example using byte-pair encoding, whereby intellectual property document-specific and / or technology-specific terms are treated as independent tokens.Preferably, the structural components are transformed into latent representations by the transformer, wherein the latent representations contain semantic and / or structural information in the document sequence and / or image sequence, and particularly preferably positional embeddings. Preferably, the latent representations, forming the vector image, are stored in a vector memory such as FAISS and / or Milvus. Preferably, the vector image, and in particular the latent representations, is categorized. Technical and origin-based metadata are particularly preferred for the categorization, taking into account categories such as technologies, technical components, origin, benefit, task, and / or document types.In particular, a multi-level categorization is conceivable, whereby technical definitions and / or descriptions of products and / or processes form a category and are further subdivided into different technical categories. Preferably, the metadata is integrated directly into the vector storage and enables dynamic filtering and / or navigation within the stored latent representations. Preferably, the categorization of the vector image, especially the latent representations, is performed manually by the user and / or automatically, particularly rule-based and / or mathematically, for example, using Hierarchical Navigable Small World and / or clustering methods such as k-means, HDBSCAN, and / or Deep Clustering.Particularly preferred are the document sequences and / or image sequences of the digital identity stored as vector images and additionally in a metadata database, wherein preferably the machine learning model accesses the vector storage and / or the metadata database to answer the request, particularly depending on the user's request via the user interface.

[0066] Furthermore, it is proposed that at least two machine learning models be used at different levels. This advantageously enables dynamic adaptation to specific requirements, whether through locally specialized data processing or global knowledge integration. Furthermore, with two machine learning models, one model can advantageously process the data efficiently and in compliance with data protection regulations by reducing it to a precise and relevant dataset, while the other machine learning model, as a global model, further processes the provided data. Preferably, a first machine learning model, in particular a local and / or smaller model, is directly linked to the digital identity and is especially preferably used for data preprocessing and / or data selection.

[0067] Preferably, the first machine learning model is used for data preprocessing and / or data selection, which includes, in particular, the input, conversion, and / or organization of relevant digital identity data, such as document sequences, image sequences, sensor data, and / or company-specific keywords, as well as conversion into latent representations, data categorization and selection, pattern recognition, and / or reduction of the data set to relevant content. Preferably, a second machine learning model, in particular a global and / or large-scale model, most preferably a large-scale language model, is used. Most preferably, the second machine learning model generates outputs for the user, in particular analyses and / or the generation of results. Particularly preferred are NN 17624 WO

[0068] The second machine learning model uses the data provided by the first machine learning model. Preferably, the first machine learning model processes the input data, particularly data provided by a user, and the second machine learning model generates the output data.

[0069] Furthermore, it is proposed that at least one machine learning model be used for data backup. This allows for the best possible protection of sensitive data, particularly within a company. Additionally, the efficiency of the entire system can be advantageously increased, as the amount of data is reduced by preprocessing one of the machine learning models, thus accelerating the analysis of the second model. The use of a machine learning model for data backup preferably means that this model exhibits at least a higher level of security compared to the second machine learning model. This includes, in particular, security protocols, preferably a company-specific encryption technique, access control (especially via third parties), logging, data masking, and / or a physical arrangement.In particular, the physical arrangement is achieved, for example, by an at least partially, preferably substantially, local arrangement of the computing device for processing the digital identity data using the machine learning model. Furthermore, it is conceivable that several machine learning models are used for data backup, wherein the machine learning models preferably have different security requirements and / or the machine learning models access at least substantially different datasets of the digital identity.

[0070] Furthermore, it is proposed that at least a first machine learning model provides data of the digital identity to a second machine learning model for processing, whereby the second machine learning model only uses at least a part of the data of the digital identity to identify the characteristics identified by NN 17624 WO.

[0071] The first machine learning model has access to the data provided. This allows for advantageous optimization of security and data protection requirements, as the second model has, at least substantially, no direct access to the digital identity data, thus protecting sensitive information, particularly that of the company. A portion of the digital identity data should preferably be understood to include particularly sensitive data of the entity, especially the company, which should specifically not be provided to the second machine learning model in its unprocessed form.

[0072] In particular, this portion of the data comprises at least 10%, preferably at least 25%, advantageously at least 50%, particularly preferably at least 75%, and particularly advantageously at least 90% of the data. Furthermore, it is conceivable that the second machine learning model accesses data from the digital identity only via the first machine learning model. Preferably, the first machine learning model is designed to anonymize, mask, and / or encrypt at least the portion of the digital identity data provided to the second machine learning model.

[0073] Furthermore, it is proposed that the digital identity data be assigned a data security category, at least in part, and that the subsequent machine learning model accesses the digital identity data depending on this data security category. Combining data security categories with controlled access advantageously ensures flexible and secure data processing. Preferably, the data security category, particularly during categorization, is determined at least based on the sensitivity, confidentiality, and / or origin of the digital identity data. For example, public data, such as published intellectual property documents, are assigned a low data security category, while personal data and / or trade secrets are assigned a high data security category. Preferably, the second machine learning model accesses the data depending on the data security category.

[0074] The second machine learning model accesses the digital identity data directly at a low data security category, while at a high data security category, the second machine learning model only accesses the data via the first machine learning model, and not directly.

[0075] Furthermore, it is proposed that a Retrieval Augmented Generation model be used to create, improve and / or use the digital identity, employing at least two machine learning models.

[0076] This allows for the advantageous generation of context from the digital identity with increased content relevance and accuracy, and reduced hallucination and content misinterpretation. Preferably, the digital identity, particularly the document sequences and / or image sequences, is transformed into the vector image as latent representations, especially using the first machine learning model as an embedding model. Specifically, the data derived from the digital identity, preferably from document sequences, image sequences, intellectual property documents, sensor data, and / or company-specific keywords, is first transformed into a text-based raw dataset. Preferably, the digital identity data, especially the text-based raw dataset, is then segmented into chunks, particularly preferably with contextual data.Preferably, the chunks are converted into latent vector representations using the first machine learning model, in particular a local embedding model. Specifically, all vector representations created with the embedding model form the vector image. In particular, the vector representations are stored in the vector memory. Preferably, a query is performed on user input via the user interface, in which the user input is first converted into a user vector representation using the first machine learning model, in particular the local embedding model. The generated user vector representation is combined with the vector representations of NN 17624 WO.

[0077] Vector images are compared, capturing semantically relevant chunks from the vector storage. Furthermore, it is conceivable that, in determining the relevant chunks, external data sources, in particular intellectual property databases, official registers, sensor data, or up-to-date development data, are also considered alongside the digital identity data.

[0078] The captured chunks are preferably subjected to anonymization and / or masking before being passed on to the second machine learning model, in particular a generative language model.

[0079] Furthermore, it is conceivable that the chunks provided to the second machine learning model are compressed before being made available to the second machine learning model, preferably by removing redundant information, creating summary representations, and / or replacing synonyms by normalizing them to taxonomic terms. Based on the context data thus provided, the second machine learning model generates a user-specific output, in particular a linguistic response. Specifically, it is conceivable that the chunks captured from the vector memory are weighted before being made available to the second machine learning model using a context evaluation procedure, preferably taking into account positional information within the document structure, a weighting of the data source, ontology mappings, and / or semantic relevance assessments.Furthermore, it is proposed that a prompt template be selected depending on the type of user input, preferably differentiating between different types of requests, in particular freedom-to-operate analyses, novelty checks, development assistance, onboarding processes, research or monitoring tasks.

[0080] Furthermore, it is proposed that the retrieval augmented generation model utilize both the vector image and categorizations and / or ratings from a vector storage and / or metadata database. This advantageously allows scaling to large datasets with at least 17624 WO in the NN.

[0081] This enables a substantially constant response latency. A hybrid retrieval method is particularly preferred, whereby a subset of the vector representations is preferably formed by filtering based on categories and / or ratings, in particular intellectual property classifications, tasks, benefit descriptions, feature ratings, company-specific keywords, technical parameters, and / or time periods. In particular, a vector-based similarity comparison is performed within this subset using the embedding representations, so that the most semantically relevant chunks are identified from the filtered dataset. The filtering is particularly preferably rule-based and / or performed with the first machine learning model, taking into account technical ontologies and taxonomic structures from the metadata.Preferably, the filtering is derived from user input, with the user defining at least some of the filters and / or the filters derived from the user input being captured by the first machine learning model. In particular, it is conceivable that at least some of the categories or ratings, such as an intellectual property rights rating, are considered in the hybrid retrieval process for at least a significant proportion of queries, depending on a classification by the first machine learning model, and especially for all queries. Alternatively or additionally, the filtering is performed in such a way that a context-relevant surplus of chunks is generated when comparing the user vector representation with the vector representations of the digital identity, preferably reducing the number of chunks available to the second machine learning model through filtering.Alternatively or additionally, sub-question queries are used in semantic response generation. In this process, a user input is broken down into several sub-questions using the first machine learning model. Each of these sub-questions is then separately converted into a user vector representation and, in particular, compared with the vector representations of the vector image to capture the chunks. In a further embodiment, preferably using the NN 17624 WO.

[0082] User input is used, employing a text-to-SQL method that generates queries against at least one relational metadata database and / or graph-based metadata database. Furthermore, particularly with the first machine learning model, agent-based controls are preferably used, implementing adaptive query logic within a retrieval-augmented-generation pipeline. The agent-based control preferably manages the switching between filters, vector retrieval, sub-question queries, and / or the text-to-SQL method, especially depending on the type of user input and the complexity of the requested information.

[0083] It is particularly preferred that user input, ratings, and / or corrections be used to adjust relevance scores, fine-tune the embedding model, optimize the filter logic, and / or the agent-based control, preferably with a continuous learning process to improve the system. Furthermore, a system for carrying out the computer-implemented procedure is proposed. This system can advantageously increase computational efficiency, as the digital identity enables efficient further data processing. Moreover, a digital identity, preferably of a technical nature, can reflect a company's interest, preferably a technical one, in further processing, preferably autonomously.Furthermore, a digital identity can advantageously capture a company's collective experience, particularly across individuals, departments, and / or projects, and establish best practices across the organization. Additionally, creating a digital identity can provide company-specific, context-aware caching, especially for (preferably transformer-based) artificial intelligence systems. This leads to significantly reduced computing time and / or optimized system scalability, at least compared to systems without caching. (See also: NN 17624 WO.)

[0084] Artificial intelligence can be used to create and / or improve digital identities, providing data of the highest possible quality.

[0085] In particular, the system includes the computing equipment, preferably communication interfaces and preferably the sensors.

[0086] Furthermore, it is proposed that distributed computing capacities be provided. This allows for the advantageous flexible adaptation of computing resources to changing requirements, such as processing large volumes of data or performing complex analyses. Distributing tasks between local and / or external units also ensures a high level of data security, as sensitive data can be processed locally while less critical tasks are outsourced to external resources. Computing capacities are preferably distributed across multiple physical and / or virtual units, such as local servers, cloud environments, or hybrid structures. Local units are particularly preferred for the preprocessing, storage, and management of sensitive data, such as intellectual property documents or sensor data, especially to meet data protection requirements.Preferably, external units, such as cloud resources, are used for complex analyses, such as pattern recognition, machine learning, and / or AI-supported result generation. In particular, at least the first machine learning model and / or the second machine learning model is executed on the external units. Furthermore, the system includes a control unit for distributing tasks among the distributed computing resources, especially local and external ones. In an alternative or additional embodiment, the control unit is designed for dynamic load balancing. In particular, user input data and / or digital identity data, especially when processing large datasets such as sensor data, are processed in parallel across multiple units, preferably using dynamic load balancing.

[0087] Furthermore, it is proposed that the system include a vector memory designed to store the digital identity, at least partially, in a vector-based manner. This would advantageously enable the rapid processing of large volumes of complex data. The vector memory is preferably a configured memory area that accommodates and manages multidimensional vectors. In particular, each vector represents, as a latent representation, specific features or properties of the digital identity in a mathematical space, such as technical competencies, intellectual property information, or sensor data. Preferably, the vector memory is designed for metric spatial operations, such as similarity comparisons, clustering, or dimensional transformations.

[0088] Furthermore, it is proposed that the system include a metadata database designed to store at least part of the digital identity, particularly relationships between data. This would advantageously enable a structured and traceable organization of the data and its connections. Moreover, the metadata database would allow for the direct analysis of data relationships, thereby supporting optimization of the digital identity. The metadata database preferably stores structured information about the digital identity data, specifically its properties, origin, context, and / or links, such as relationships, hierarchies, and / or dependencies.

[0089] A relational and / or graph-based metadata database is particularly preferred. Examples of stored links include linking intellectual property documentation with specific keywords, classifications, and benefit descriptions, or connecting sensor data with the context of its acquisition, such as timestamps or device type, as well as the associated analysis results.

[0090] The computer-implemented method and system according to the invention are not intended to be limited to the application and embodiment described above. In particular, the NN 17624 WO

[0091] The computer-implemented methods and the system according to the invention, for fulfilling a mode of operation described herein, may have a different number of individual elements, components and units than the number mentioned herein.

[0092] Drawings

[0093] Further advantages will become apparent from the following description of the drawings. The drawings illustrate an embodiment of the invention. The drawings, the description, and the claims contain numerous features in combination. A person skilled in the art will expediently consider the features individually and combine them into meaningful further combinations.

[0094] They show:

[0095] Fig. 1 a system for carrying out a computer-implemented procedure for development work, onboarding processes and / or monitoring optimizations with a digital identity, Fig. 2 the system for creating, improving and / or using the digital identity,

[0096] Fig. 3 shows the computer-implemented method for creating the digital identity,

[0097] Fig. 4 shows the computer-implemented method for improving digital identity,

[0098] Fig. 5 shows the computer-implemented method for using the digital identity,

[0099] Fig. 6 shows the creation and use of a vector representation of the digital identity.

[0100] Fig. 7 shows a word comparison of another technical document sequence with the digital identity,

[0101] Fig. 8 shows a relevance comparison between two technical

[0102] Document sequences, Fig. 9 the computer-implemented procedure for recording an injury, and

[0103] Fig. 10 shows the computer-implemented method using a retrieval-augmented-generation model for creating, improving and / or using the digital identity.

[0104] Description of the exemplary implementations

[0105] Figure 1 schematically depicts a system 12 for carrying out a computer-implemented procedure. The system 12 comprises a digital identity 10, which preferably describes at least one property of an entity. This entity can most preferably be a manufacturing and / or developing company; however, individual persons, particularly those involved in manufacturing and / or development, or groups of persons are also conceivable. The digital identity 10 preferably comprises a structured collection of gathered data, in particular a cluster dataset containing data of different types. A technical digital identity is most preferably created and / or improved, wherein the property describes a technical behavior of the entity. In principle, other properties, such as social, strategic, and / or economic behavior, as well as combinations thereof, are also conceivable."Behavior" is understood to mean at least a temporal, relevance-dynamic, categorical and / or spatial development of the respective property.

[0106] In the illustrated embodiment, sensor data 34 are recorded using a mobile device 26, preferably a smartphone 28 (process step 100). The sensors 30 integrated into the mobile device 26 can record various technical parameters as needed, such as accelerations, positions, color measurements, visual or acoustic signals, and the like. Alternatively or additionally, it is conceivable that external sensors can be connected to the smartphone 28 via interfaces such as Bluetooth or USB to record further technical sensor data 34. PreferablyNN 17624 WO

[0107] The sensor data 34 thus acquired are processed directly on the mobile device 26, visualized and / or transmitted via an interface to the digital identity 10 (process step 102). In this context, an object 44 can be, for example, a tire 46 whose condition, properties or operating parameters are measured.

[0108] The digital identity 10 preferably provides selective data processing. This preferably occurs based on an evaluation – for example, temporal, relevance-dynamic, categorical, and / or spatial – which is carried out either by a user and / or by a machine learning model. Subsequently, the data filtered or analyzed in this way are made available to a user interface 20 and / or the machine learning model for further processing (process step 103).

[0109] Furthermore, distributed computing resources are provided to efficiently handle large volumes of data. These computing resources can be distributed across multiple physical and / or virtual units, such as local servers, cloud environments, or hybrid structures. Local units are particularly preferred for the preprocessing, storage, and management of sensitive data, such as intellectual property documents or sensor data, to meet data protection requirements. For complex analyses, such as pattern recognition, machine learning, and / or AI-supported result generation, external units (e.g., cloud servers) can be used.

[0110] Cloud resources) are provided. In particular, at least one machine learning model is executed on the external units. Furthermore, the system 12 has a control unit responsible for distributing tasks among the distributed (especially local and external) computing capacities. In an alternative or additional configuration, this control unit is provided for dynamic load balancing. In this way, for example, the user's input data 32 and / or the digital identity data 10, especially with large data volumes (such as sensor data 34), NN 17624 WO

[0111] processed in parallel in multiple units, preferably using dynamic load distribution.

[0112] In this way, the system 12 shown in Figure 1 enables the acquisition of technical sensor data 34 using a mobile device 26, the selective processing of the data via the digital identity 10, and its distribution to local and external computing units according to requirements. In particular, temporal, relevance-dynamic, categorical, and / or spatial evaluations are incorporated, thus ensuring a technically sound, efficient, and data protection-compliant execution of the computer-implemented procedure.

[0113] Figure 2 shows a schematic representation of a system 12 for creating, improving, and / or using a digital identity 10. The system 12 comprises at least one computing device 14, which includes a processor 18, a storage unit 16, and a user interface 20 with input and output means. The computing device 14 can be a standalone unit, distributed within a local IT infrastructure, or configured as a virtual cloud infrastructure. A computer-implemented process is executed within this computing device 14, which is controlled in particular by a computer program with program code stored in the storage unit 16.

[0114] Digital identity 10 preferably includes all data that describe at least one characteristic of an entity, such as a manufacturing and / or development company. This data can be of a technical, economic, and / or social nature. Alternatively, instead of or in addition to a dedicated database, a data interface is provided that offers factors for evaluating any company data in terms of time, relevance, categories, and / or location.

[0115] Input data, in particular input data 32, which preferably describes the technical behavior of the entity, is entered via the user interface 20. Furthermore, users can use this NN 17624 WO

[0116] The interface allows further information to be entered, for example, user input 36 in the form of questions, ratings, and / or statements about specific projects or document sequences. The digital identity 10 is created and / or improved based on this input data 32 and user input 36.

[0117] Subsequently, 10 output data sets can be generated using the digital identity and output via the user interface.

[0118] System 12 also includes a vector memory 22, which is designed to represent and manage the digital identity 10, at least partially, in the form of vectors. Each vector represents a latent representation of specific features or properties of the digital identity 10 in a multidimensional space, such as technical skills, intellectual property information, or sensor data 34. Ideally, the vector memory 22 supports metric spatial operations such as similarity comparisons, clustering, or dimensional transformations.

[0119] Furthermore, the system 12 comprises a metadata database 24, in which structured information about the data of the digital identity 10 is preferably stored. This particularly concerns its properties, origin, context, and / or links, such as relationships, hierarchies, and dependencies between individual data records. The metadata database 24 is particularly preferably implemented in a relational and / or graph-based manner. Examples of such links include the assignment of intellectual property documents to specific keywords 62, classifications 58, and benefit descriptions, or the storage of sensor data 34 with associated timestamps, device types, and analysis results.

[0120] In this way, the system 12 shown in Figure 2 enables structured and efficient management of the digital identity 10: It processes input and sensor data 32, 34, organizes them using the metadata database 24, and stores them partly vectorized in the vector memory 22. At the same time, users can input information and make queries via the user interface 20.

[0121] and retrieve the results generated by the processor 18 or other computing units - such as new, improved or updated data sets of the digital identity 10 - as output data 38.

[0122] Figure 3 shows a multi-step process for creating the digital identity 10. In a first step 100, all data constituting the digital identity 10 are analyzed, processed, and transformed into a structured form, preferably at least partially using a machine learning model 40 and / or manually. In a second step 102, data from various sources, such as company documents, projects, experiences, and processes, are extracted and subsequently semantically linked and / or structured. In a third step 104, the machine learning model 40 is used to identify relationships between the data, eliminate redundant information, and generate a coherent representation of the digital identity 10.

[0123] In a subsequent process step 106, at least one separate, preferably technical, document sequence 50 and / or image sequence 52 is considered for the creation and / or improvement of the digital identity 10. The document sequence 50 can, for example, include intellectual property documents (patents, utility models, etc.), technical verification documents (test reports, measurement data, etc.), and / or technological descriptions (specification documents, circuit diagrams, process descriptions, scientific reports). The image sequence 52, in turn, can contain technical visualizations (e.g., CAD drawings), photographic representations (prototypes, test setups), or dynamic image datasets (video sequences of products and / or experiments).

[0124] Furthermore, in procedural step 108, it is proposed to use 10 intellectual property documents for the creation and / or improvement of the digital identity, which were researched into a separate intellectual property document.

[0125] Preferably, these researched intellectual property documents are identified using NN 17624 WO.

[0126] their contents, claims and / or intellectual property classifications 58, for example an international intellectual property classification 60. These documents are integrated and stored either fully or partially on a cyclical basis (e.g. monthly, quarterly) in the digital identity 10, in particular also with the help of automated searches (e.g. via an API interface with a patent office).

[0127] In a further process step 110, an evaluation of features from a feature breakdown of the patent claims within the intellectual property documents is carried out. This can be done either automatically (e.g., using a machine learning model 40) or manually via a user interface 20. This allows, in particular with the help of a relevance comparison, the determination of the probability of relevance and weighting of further intellectual property documents that are added to the digital identity 10. In this step, tasks and / or benefit descriptions from the applicant's own intellectual property documents are also included in order to determine their relevance for further documents and, if applicable, to integrate them into the digital identity 10.

[0128] Also in process step 110, a word comparison can be performed between different technical document sequences 50 and / or image sequences 52 to identify similarities, differences and / or correlations of individual terms, word groups or phrases. This supports the automatic or manual weighting of new documents and their assignment to the digital identity 10.

[0129] Furthermore, in process step 112, company-specific keywords 62 are used to create and / or improve the digital identity 10. These keywords are characteristic of a company's innovation and technology fields (e.g., industries, product classes, project classes) and can be defined manually or automatically extracted from existing document sequences 50. Preferably, the keywords 62 are assigned to the technical document sequences 50 and / or image sequences 52; beiNN 17624 WO

[0130] When adding further sequences, the appropriate keywords can be added directly. In process step 114, these keywords are also tracked historically and used as input parameters for the machine learning model 40.

[0131] In a subsequent process step 116, technical sensor data 34 are used to create, improve, and / or utilize the digital identity 10. This data originates from processes, machines, and / or devices (including data from people within the entity, such as audio, video, and behavioral recordings). Preferably, the sensor data 34, particularly if unstructured, are automatically or manually ontologically structured and assigned to the company-specific keywords 62. In process step 118, the sensor data 34 are optionally linked to existing document sequences 50 and / or image sequences 52, particularly for contextualization when using the digital identity 10.

[0132] In process step 120, the digital identity 10 is further used for development support, onboarding new employees, and / or optimizing monitoring. This can involve providing automated recommendations on technical resources, technologies, and solutions, identifying team-building competencies, or making targeted information accessible to relevant individuals. Finally, in process step 122, results from development work, onboarding processes, and / or monitoring optimizations are evaluated and integrated into the digital identity 10. This includes, among other things, documented findings, measurement data, simulation results, and newly identified key competencies.

[0133] In this way, Figure 3 illustrates the procedure for creating the digital identity 10, which in particular involves the recording and structural merging of technical, economic and / or social data from NN 17624 WO

[0134] various sources, including intellectual property documents, sensor data 34 and company-specific keywords 62.

[0135] Figure 4 shows a method for improving the digital identity 10, in which, in a first process step 200, the quality and currency of the data constituting the digital identity 10 are monitored and evaluated, preferably continuously and in particular using a machine learning model 40. In process step 202, patterns are recognized, new data describing the digital identity 10 are added, and / or outdated or inconsistent data are removed. In process step 204, recommendations can be generated via a user interface 20 to adapt the digital identity 10 to new requirements or developments.

[0136] For example, the structuring of the document sequences 50 and / or image sequences 52 can be refined in process step 206, the weighting of certain features adjusted, or new keywords 62 added. In process step 208, the sensor data 34 and its assignment are updated, if necessary, especially if new measurement or user inputs 36 are available. Finally, the changes made are transferred to the digital identity 10 in process step 210, so that a continuously optimized data set is available. Further steps such as 212, 214, 216, 218, 220, and 222 can be used, according to the creation of the digital identity, to map additional subprocesses or alternative paths of improvement (e.g., extended data cleansing, additional evaluations, or revalidations).

[0137] Figure 5 illustrates a procedure for using the digital identity 10, in which, in a process step 300, the machine learning model 40 is used to dynamically optimize access to relevant information and to answer context-specific questions. In process step 302, the digital identity 10 is provided with at least partial context specifying which data and content are available for a particular query NN 17624 WO

[0138] are relevant. For example, in process step 304, an evaluation of document sequences 50, image sequences 52 or sensor data 34 can be carried out, possibly taking into account company-specific keywords 62, in order to provide the user with a focused output 38.

[0139] In a process step 306, the digital identity 10 is preferably continuously improved based on the results of the machine learning model 40. For this purpose, analysis results can be reintegrated into the digital identity 10, thus further refining its structure, relevance weights, and categorizations. In this way, a self-improving system is available that allows consistent, context-based, and secure data processing, for example, by segmenting or classifying sensor data 34, applying multiple machine learning models (local and global), and operating with different security levels depending on the data security category 64.

[0140] Thus, Figures 3, 4 and 5 each illustrate the phases of creating, improving and using the digital identity 10, ensuring a comprehensive life cycle for handling and evaluating business-relevant, especially preferably technical, data.

[0141] Figure 6 shows a method for creating a vector image 80 of the digital identity 10, in which, in a first step, the digital data—preferably document sequences 50 and / or image sequences 52—are stored at least partially in raw text form and / or as latent representations in a vector image 80. A transformer, for example a transformer encoder (e.g., PatentBERT or SciBERT), is particularly preferably used as a machine learning model 40 for generating and / or processing this vector data.

[0142] Preferably, in particular for the creation and / or processing of the vector image 80, the document sequences 50 and / or image sequences 52 inNN 17624 WO are used.

[0143] Its structural components, such as title, abstract, background, task statement, benefits description, detailed description, technical descriptions, definitions, or claims, are broken down. These components are advantageously structured into subword units using byte-pair encoding, with patent-document-specific or technology-specific terms treated as independent tokens. The transformer converts these into latent representations, which, in addition to semantic and / or structural information, can contain, in particular, positional embeddings.

[0144] The latent representations are stored in a vector storage system 22 (e.g., FAISS or Milvus) and form the vector image 80 of the digital identity 10. The vector image 80 is then categorized, for example, according to technical and / or origin-based metadata such as technologies, components, origin, benefits, tasks, and / or document types. This categorization can be single- or multi-stage and can be performed manually or automatically (e.g., rule-based, mathematical, using clustering methods such as k-means or HDBSCAN). Furthermore, the metadata can be directly integrated into the vector storage system 22, enabling dynamic filtering and / or navigation within the stored latent representations.The procedure shown in Figure 6 thus offers a scalable and structured storage and processing of the digital identity 10; the machine learning model 40 can access the vector image 80 and / or a metadata database 24 depending on the user request and processing requirements.

[0145] Figure 7 illustrates an example of a word matching 70 between at least two technical document sequences that are added to or compared with the digital identity 10. In a first step, this can (e.g.,

[0146] In process step 100 (or a corresponding process step), text-based content is captured and analyzed for word distributions 72, 74, and 76. The dashed lines 72, 74, and 76 can represent different frequency distributions or weightings, which are determined through statistical comparison or NLP (Natural Language Processing) analysis. A high NN 17624 WO

[0147] Overlap of certain keywords or phrases suggests increased relevance. In particular, this determines the extent to which the submitted document or image sequence has similarities or differences to data records already present in the digital identity. This word matching preferably results in recommendations for classifying, determining relevance, or updating the digital identity.

[0148] Figure 8 schematically illustrates a relevance comparison between at least two technical document sequences to determine the relevance or weighting of relevant text elements when creating or improving the digital identity 10. In a first process step 200, the relevant text elements—for example, features, task descriptions, or benefit descriptions—are extracted from another intellectual property document. This is followed in process step 202 by the numbering of these text elements, e.g., using TF-IDF, word embeddings, or other methods. In the next process step 204, the similarity or relevance of the relevant text elements is determined using a comparison algorithm, such as cosine similarity or a machine learning model like SVM, gradient boosting (e.g., XGBoost), or neural networks.Depending on this calculated similarity, the relevant text modules are weighted in process step 206; if they are sufficiently similar, a high relevance for the digital identity 10 is assumed. Finally, in process step 208, these evaluated text modules are integrated into the digital identity 10. In a possible alternative design, the text modules of the existing intellectual property documents in the digital identity 10 are semantically compared with the new document, so that its relevance can be determined, for example, from the number of matching words.

[0149] This relevance check, shown in Figure 8, ensures that new documents, sequences, or attributes are precisely integrated into the digital identity 10 and correctly weighted within the system. (ThisNN 17624 WO)

[0150] facilitates precise searching, categorization and ultimately the use of the digital identity 10 in further processes, such as intellectual property strategies, development support, onboarding or monitoring.

[0151] Figure 9 shows a method for detecting damage to an object 44, which here is exemplified as a tire 46 with a tire tread 48. The system 12 serves for data acquisition and analysis in order to identify possible damage or deviations on the tire 46.

[0152] In a first step, the tire profile 48 is examined via the system 12, preferably by means of at least one sensor 30 (not shown) or other detection device on the object 44. Technical sensor data 34, such as measurements of tread depth, structure, cracking or wear, can be recorded.

[0153] Subsequently, in a further step, an evaluation of these sensor data 34 is carried out, for example by comparison with reference values ​​or using a machine learning model 40, in order to identify irregularities in the tire tread 48. In this way, damage (e.g. cracks, cuts or deviations from the target tread depth) can be detected automatically.

[0154] A categorization of the detected damage according to its severity and type is particularly preferred. Depending on the result, the system 12 can issue a warning message and forward the relevant information – such as timestamp, location, or extent of the damage – directly to the digital identity 10 or a connected database. In this way, the condition of the tire 46 can be continuously monitored and documented.

[0155] The procedure shown in Figure 9 is particularly suitable for continuous condition monitoring of vehicle tires in industrial processes or fleet management systems. It can also be applied to other objects with a similar structural design to enable early detection of wear.

[0156] To detect injuries, wear and tear or abrasion and to initiate appropriate measures (e.g. maintenance, replacement, repair).

[0157] Figure 10 shows the computer-implemented procedure using a Retrieval Augmented Generation Model. The Retrieval Augmented Generation Model is used to create the digital identity. The Retrieval Augmented Generation Model is used to improve the digital identity. The Retrieval Augmented Generation Model is used to utilize the digital identity. The Retrieval Augmented Generation Model is used to create, improve, and / or utilize the digital identity, employing at least two machine learning models.40, 42

[0158] During extraction step 400, a text-based raw data set is obtained from input data 32 for the digital identity 10. Extraction step 400 involves obtaining a text-based raw data set from document sequences 50, image sequences 52, intellectual property documents 54, 56, sensor data 34, and keywords 62. The raw data set is divided into chunks 402. Metadata such as source, timestamp, semantic context, and associated categories 66 are assigned to each chunk 402. A first machine learning model 40, which acts as an embedding model, converts each of these chunks 402 into a latent vector representation in the embedding step 404. The vector representations together form the vector image 80 and are stored in vector memory 22.

[0159] The retrieval augmented generation model uses both the vector image 80 and categorizations and / or ratings 66, 68 from the vector memory 22. The retrieval augmented generation model also uses both the vector image 80 and categorizations and / or ratings 66, 68 from the metadata database 24.

[0160] An agent-based controller 406 processes user input 36 entered via the user interface 20. The controller 406 analyzes the input and decides which downstream procedures in a retrieval augmented neural network 17624 WO should be used.

[0161] The generation pipeline of the Retrieval Augmented Generation Model is applied. A Sub-Question Query Module 408 breaks down potentially complex user input into several sub-questions that can be processed separately. A Text-to-SQL Module 410 generates database-compatible queries to the metadata database 24 and forwards the resulting structured information directly to a compression module 424, so that it can be supplied as supplementary context to the second machine learning model 42.

[0162] A hybrid retrieval method 412 is used to generate a pre-selected subset of vector representations from the vector memory 22. This hybrid retrieval method 412 performs filtering based on categories 66, ratings 68, intellectual property classifications 58, task descriptions, benefit descriptions, company-specific keywords 62, technical parameters, time periods, and / or considering positional information within the document structure, ontology mappings, an evaluation of the data source, and semantic relevance assessments 68. The filtering can be rule-based and / or performed by the first machine learning model 40, taking into account, in particular, ontology mappings and taxonomic structures from the metadata database 24.

[0163] In an embedding step 414, the user input 36 is converted into a user vector representation. In an embedding step 414, the first machine learning model 40 is used to convert the user input 36 into a user vector representation. The user vector representation is then compared in a chunk retrieval step 416 with all vector representations of the vector image 80. As a result, semantically relevant chunks 420 are identified from the vector memory 22. External data sources such as intellectual property databases, official registers, current sensor data, or development data can also be included.

[0164] A hybrid retrieval method 418, applied after chunk retrieval 416, evaluates the captured chunks 420 with regard to their contextual relevance. EsNN 17624 WO

[0165] A further reranking of the chunks takes place. In the hybrid retrieval procedure 418, a reranking is carried out taking into account categories 66, ratings 68, intellectual property classifications 58, tasks, benefit descriptions, company-specific keywords 62, technical parameters, time periods and / or taking into account position information within the document structure, ontology assignments, an evaluation of the data source and semantic relevance evaluations 68.

[0166] The extracted chunks 420 are transformed into text chunks 422 using the first machine learning model 40. These text chunks 422 are then converted into a compression 424. This compression 424 removes redundant information, generates summary representations, and replaces synonyms with standardized terms based on taxonomic structures. Additionally, the SQL data extracted in step 410 is integrated into this compression 424 to generate a complete response context.

[0167] The data generated during compression (424) undergoes a data backup (426) in which the information is anonymized, masked, and / or versioned. Compliance with data security category 64 is ensured.

[0168] The second machine learning model 42 receives the context data generated by compression 424 and data backup 426 and generates a user-specific output from it. The generated output is provided as input data 38. Reference sign

[0169] 10 Digital Identity

[0170] 12 System

[0171] 14 Computer equipment

[0172] 16 storage units

[0173] 18 processor

[0174] 20 User interface

[0175] 22 Vector memory

[0176] 24 Metadata database

[0177] 26 Mobile device

[0178] 28 Smartphones

[0179] 30 Sensor

[0180] 32 Input data

[0181] 34 sensor data

[0182] 36 User input

[0183] 38 Initial data

[0184] 40 machine learning models

[0185] 42 Model of machine learning

[0186] 44 objects

[0187] 46 tires

[0188] 48 Tire profile

[0189] 50 document sequence

[0190] 52-image sequence

[0191] 54 Intellectual Property Document

[0192] 56 Intellectual Property Document

[0193] 58 Classification of intellectual property rights

[0194] 60 International Classification of Intellectual Property Rights 62 Keywords

[0195] 64 Data security category

[0196] 66 Category Relevance Word Matching Word Distribution Word Distribution Word Distribution Vector Image External Unit Processor Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step Process Step NN 17624 WO

[0197] Procedure step

[0198] Procedure step

[0199] Procedure step

[0200] Procedure step

[0201] Procedure step

[0202] Procedure step

[0203] Procedure step

[0204] Procedure step

[0205] Procedure step

[0206] Procedure step

[0207] Procedure step

[0208] Procedure step

[0209] Extract

[0210] Chunks

[0211] Embedding

[0212] Agent-based control

[0213] Sub-Question Queries

[0214] Text-to-SQL

[0215] Hybrid retrieval procedure

[0216] Embedding

[0217] Chunk Retrieval

[0218] Hybrid retrieval procedure

[0219] Chunks

[0220] Text chunks

[0221] Compression

[0222] Data backup

Claims

December 23, 2025 Claims 1. Computer-implemented process in which a digital identity (10), in particular of an enterprise, is created and / or improved, in particular using a machine learning model (40; 42) and / or for use with a machine learning model (40; 42).

2. Computer-implemented method according to claim 1, characterized in that a proprietary, preferably technical, document sequence (50) and / or image sequence (52) is used to create and / or improve the digital identity (10).

3. Computer-implemented method according to claim 2, characterized in that the document sequence (50) and / or image sequence (52) is at least one separate intellectual property document (54).

4. Computer-implemented method according to claim 3, characterized in that intellectual property documents (56) are used to create and / or improve the digital identity, which were researched for the user's own intellectual property document (54). - 50 - NN 17624 WO 5. Computer-implemented method according to one of claims 3 or 4, characterized in that at least one external intellectual property classification (58), in particular an international intellectual property classification (60), is used to create and / or improve the digital identity intellectual property documents (54; 56).

6. Computer-implemented method according to one of the preceding claims, characterized in that an evaluation of features of a feature breakdown of patent claims of an intellectual property document (54; 56) is used to create and / or improve the digital identity (10).

7. Computer-implemented method according to one of claims 3 to 5, characterized in that a task definition and / or a description of advantages of a proprietary intellectual property document (54) is used to create and / or improve the digital identity (10).

8. Computer-implemented method according to claim 2, characterized in that a word matching (70) between several technical document sequences (50) and / or image sequences (52) is used in the creation and / or improvement of the digital identity (10).

9. Computer-implemented method according to one of the preceding claims, characterized in that company-specific keywords (62) are used to create and / or improve the digital identity (10).- 51 NN 17624 WO 10. Computer-implemented method according to one of the preceding claims, characterized in that technical sensor data (34) are used, in particular for the creation and / or improvement of the digital identity (10) and / or in the use of the digital identity (10).

11. Computer-implemented method according to claim 10, characterized in that the sensor data (34) are recorded with a mobile device (26), in particular with a smartphone (28).

12. Computer-implemented method according to one of claims 10 or 11, characterized in that the mobile device (26) is used for evaluation specific to a digital identity (10).

13. Computer-implemented method according to one of the preceding claims, characterized in that the digital identity (10) is used to assign company-specific keywords (62) to the document sequences (50) and / or image sequences (52).

14. Computer-implemented method according to one of the preceding claims, characterized in that the digital identity (10) is used for technical development support and / or onboarding and / or monitoring optimization.

15. Computer-implemented method according to claim 14, characterized in that a problem statement and / or a description of advantages is used in the development support. - 52 - NN 17624 WO 16. Computer-implemented method according to one of the preceding claims, characterized in that a comparison of an own task definition and / or a own advantage description and a digital task definition and / or advantage description of the digital identity (10) is used, in particular for technical development support.

17. Computer-implemented method according to one of the preceding claims, characterized in that results from development work, onboarding processes and / or monitoring optimizations are used to create and / or improve the digital identity (10).

18. Computer-implemented method according to one of the preceding claims, characterized in that a vector image (80) of the digital identity (10) is created and / or edited during the creation and / or improvement of the digital identity.

19. Computer-implemented method according to one of the preceding claims, characterized in that at least two machine learning models (40; 42) are used at different stages.

20. Computer-implemented method according to one of the preceding claims, characterized in that at least one machine learning model (40) is used for data backup. NN 17624 WO 21. Computer-implemented method according to one of the preceding claims, characterized in that at least a first machine learning model (40) provides data of the digital identity (10) to a second machine learning model (42) for processing, wherein the second machine learning model (42) has access to only the data provided by the first machine learning model (40) from at least a part of the data of the digital identity (10).

22. Computer-implemented method according to one of the preceding claims, characterized in that the data of the digital identity (10) is at least partially assigned a data security category (64) and the further machine learning model (42) accesses the data of the digital identity (10) depending on the data security category (64).

23. Computer-implemented method according to claim 19, characterized in that a retrieval augmented generation model is used to create, improve and / or use the digital identity (10), wherein the at least two machine learning models (40; 42) are used.

24. Computer-implemented method according to claim 23, characterized in that the retrieval augmented generation model uses both the vector image (80) and categorizations and / or ratings (66; 68) from a vector memory (22) and / or a metadata database (24).

25. System (12) for carrying out a computer-implemented method according to one of the preceding claims.

26. System (12) according to claim 20, characterized in that distributed computing capacities are provided.

27. System (12) according to one of claims 21 or 22, characterized by a vector storage device (22) which is designed to store the digital identity (10) at least partially in a vector-based manner.

28. System (12) according to one of claims 21 to 23, characterized by a metadata database (24) which is intended to store the digital identity (10) at least partially, in particular relationships between data.

Images