System and method for data autonomy through inference path in task orchestration platform

Abstract

The present disclosure relates to a sharp end mission orchestration platform, aiming to address the shortcomings in mission data management and processing, ensuring the autonomy and integrity of data. The platform involves a new application of a multi-layered reasoning path protocol for autonomous arbitration of artificial intelligence models to manage and process mission data. This advanced approach of deploying a multi-layered reasoning path protocol facilitates extremely fast data self-organization, enhances data independence in the mission platform, maintains high data integrity, and accelerates the extraction of actionable information from various data types obtained from mission orchestration.

Description

Technical Field

[0001] This disclosure relates to a leading task orchestration platform designed to address shortcomings in task data management and processing, ensuring data autonomy and integrity. The task orchestration platform involves a novel application of a multi-layered inference path protocol for autonomous arbitration of artificial intelligence (AI) models to manage and process task data. This advanced approach to deploying the multi-layered inference path protocol facilitates extremely rapid data self-organization, enhances data independence within the task platform, maintains high data integrity, and accelerates the extraction of actionable information from the diverse data types acquired in task orchestration. Background Technology

[0002] In the realm of task orchestration platforms, particularly those using modern databases such as, but not limited to, Hadoop, MongoDB, ElasticSearch, or Cassandra, numerous operational challenges exist, primarily due to the management of large volumes of heterogeneous data types, such as video, images, text, audio, and other non-standardized data types. The labor-intensive nature of running and maintaining these systems is significant, as it involves comprehensive management of diverse data storage and complex data retrieval operations. Furthermore, systems can bottleneck during periods of high data inflow, as rapid data surges can saturate system capacity. Tools capable of handling various data formats—structured, unstructured, semi-structured, and benchmark data—are typically used to detect patterns that represent anomalies and / or events. The challenges associated with managing such extensive datasets stem from the sheer volume, diversity, and speed at which the data must be processed.

[0003] For example, once data patterns and / or insights are successfully identified in a task orchestration platform, the platform will then generate corresponding notifications for operators. However, the volume of notifications often exceeds what operators can manage effectively and efficiently. The processes involved in analyzing these notifications typically require detailed examination and may involve the analysis of subject-specific information and any specific instructions used to process the notification content. Furthermore, operators often spend a significant amount of time and effort to fully understand the nature and impact of detected events, which often leads to severe fatigue and a high attire rate.

[0004] In particular, these tasks place extremely high demands on the cognitive abilities of operators. The main challenge stems from the sheer volume and rapid growth of this data, which existing task orchestration platforms often struggle to efficiently handle. This necessitates not only scalable storage solutions but also robust mechanisms to ensure data integrity and redundancy, significantly increasing infrastructure cost and complexity. From an analytical perspective, extracting actionable insights from the data triggering notifications requires advanced analytics tools and substantial computing resources. When system entropy reaches a critical level, most current processes tend to slow down, become unavailable, or even crash abruptly.

[0005] To address the challenges associated with managing large amounts and diverse types of data in task orchestration platforms, several approaches have been proposed by those skilled in the art.

[0006] Of the approaches proposed to date, the application of AI models is a particularly effective alternative to the manual analysis and triage of notification content by routine human operations in task orchestration platforms. While the basic concepts of AI—including the steps of feeding computers training data to autonomously develop predictive models—are easily understood, their practical implementation faces significant challenges. For example, multiple versions and / or types of AI models are ultimately designed for similar tasks, and each model tends to use different technologies, programming languages, and libraries. These differences pose obstacles to ensuring compatibility between these AI models. Furthermore, the performance of these models can vary significantly due to the types of input data they process.

[0007] Existing solutions for task orchestration platforms that enhance various data streams typically involve the direct application of artificial intelligence models, with minimal preprocessing to adapt the data to a model-compatible format. These models, tasked with inference based on received data, often involve complex algorithms and are resource-intensive. Therefore, industry experts have been seeking more adaptive and autonomous learning systems that can efficiently perform both atomic and macroscopic correlations of data. These autonomous systems tend to offer more effective solutions to the complex challenges prevalent in modern data management and analytics operations. Summary of the Invention

[0008] This disclosure aims to transform the technology used for processing large volumes of data in task orchestration platforms by adopting a novel inference path protocol to autonomously manage the analysis of these data sets. The primary objective is to reduce human involvement and the need for continuous human oversight and constant monitoring by empowering machines to effectively examine and process data patterns and insights generated from data analytics.

[0009] The technology presented in this paper automates the complex processes of data association, analysis, and triage, which significantly improves operational efficiency and accuracy, reduces human error, and decreases reliance on large-scale human supervision.

[0010] This technology introduces a novel approach for integrating multiple artificial intelligence (MLAI) models into a unified framework within a task orchestration platform. This modular approach not only facilitates seamless processing and autonomous decision-making across diverse data types but also ensures dynamic adaptation to structured and data environments. Furthermore, each module can autonomously adjust its operating parameters to optimize its processing methods for the specific types of data it encounters, whether the data is a real-time video stream or batch-processed text data.

[0011] At the heart of the system architecture lies the implementation of sophisticated artificial intelligence models designed to replace conventional manual monitoring methods. These models continuously receive diverse data inputs, enabling them to iteratively learn and enhance their predictive capabilities. To maintain a high level of accuracy and performance, the platform incorporates mechanisms to ensure compatibility and interoperability between different models that utilize various technologies, programming languages, and libraries. This interoperability is crucial for maintaining the system's robustness and flexibility, allowing it to adapt to the rapid evolution of data processing technologies.

[0012] According to one aspect, this disclosure describes a computational module for autonomously verifying and categorizing data in a database. The disclosed computational module includes a processing unit and a non-transitory medium readable by the processing unit. The medium stores instructions that, when executed by the processing unit, cause the processing unit to extract data from the database, process the data using a feature extraction model to identify multiple features in the data, and assign a ranking to each identified feature according to a preset ranking criterion. For each feature, sorted in descending order based on its assigned ranking, the processing unit uses a first classification model selector to select an inference path protocol from a set of inference path protocols, wherein the first classification model selector is trained based on a set of rules related to the characteristics and data type of the feature, and classifies the feature using the selected inference path protocol. Each inference path protocol includes a hierarchical, multi-level path structure, including a basic node for receiving input, and multiple classification models set as route nodes at each level of the path structure, such that each route node processes the input data and branches into further paths until a terminal node is set with a multi-class classification model that produces the final classification output. Then, the processing unit aggregates the final classification output generated by the inference path protocol, selects a trained artificial intelligence (AI) model from a collection of AI models stored in the database using a second classification model selector, based on the aggregated final classification output, and performs an inference task using the selected AI model, based on the aggregated final classification output. The processing unit then performs operations on the data based on the results of the inference task.

[0013] According to another aspect, this disclosure describes a method for autonomously verifying and categorizing data in a database using a computational module. The method includes the following steps: extracting data from the database, identifying features in the data using a feature extraction model, and ranking each identified feature. In this method, for each feature in the data, and sorted in descending order based on the corresponding ranking for each feature, a first classification model selector selects an inference path protocol from a set of inference path protocols, and classifies the feature using the selected inference path protocol. The first classification model selector is trained based on a set of rules, and each inference path protocol includes multiple classification models set in a multi-level path structure. The input of each inference path protocol is provided to a basic node of the corresponding path structure, and the final classification output is generated by a multi-class classification model provided at the terminal node of the corresponding path structure, where each layer of the path structure has route nodes including classification models. Then, the method uses a second classification model selector to select a trained AI model from a set of trained AI models based on a combination of the final classification outputs generated by all inference path protocols, performs an inference task based on the combination of classifications, and then performs operations on the data based on the result of the inference task. Attached Figure Description

[0014] Different embodiments of this disclosure will be described below with reference to the following figures:

[0015] Figure 1 A block diagram of a system for autonomous verification and data triage in a system database, according to an embodiment of the present disclosure, is shown.

[0016] Figure 2 A block diagram is shown of a computing module for autonomously verifying and processing data groups in a database, according to an embodiment of the present disclosure;

[0017] Figure 3 A block diagram illustrating a processing system for performing embodiments of the present disclosure is shown; and

[0018] Figure 4 A flowchart illustrating a process for autonomously verifying and processing data groups in a system according to an embodiment of this disclosure is provided. Detailed Implementation

[0019] The following detailed description will be made with reference to the accompanying drawings, which illustrate details and embodiments of the present disclosure for illustrative purposes. Features described in the context of one embodiment may be adapted accordingly to the same or similar features in other embodiments, even if not explicitly described in those other embodiments. Additions and / or combinations and / or substitutions to features described in the context of one embodiment may be adapted accordingly to the same or similar features in other embodiments.

[0020] In the context of different embodiments, the articles “a,” “an,” and “the” used with respect to features or elements include references to one or more features or elements.

[0021] In the context of different embodiments, the term “about” or “approximately” applied to numerical values ​​covers exact values ​​and reasonable differences as commonly understood in the relevant art, such as within 10% of a specified value.

[0022] The term “and / or” as used herein includes any and all combinations of one or more of the listed related items.

[0023] As used herein, “including” means, but is not limited to, anything that follows the word “including”. Therefore, the use of the term “including” indicates that the listed elements are required or mandatory, but other elements are optional and may or may not be present.

[0024] As used herein, “composed of” includes and is limited to anything contained within the phrase “composed of”. Therefore, the use of the phrase “composed of” indicates that the listed elements are required or mandatory, and that no other elements exist.

[0025] As used in this article, “data” refers to various data formats, including but not limited to structured data, unstructured data, semi-structured data, and benchmark data.

[0026] As used herein, when certain components, modules, etc. are described as being “set up” to perform certain functions, it should be understood that such setting can be implemented, for example, by electronic circuits or hardware designed to perform such functions, and this can be done by software programming or by controlling the function of electronic circuits in a particular way to perform the function or a combination of functions.

[0027] Furthermore, those skilled in the art will recognize that certain functional units in this description are labeled as modules throughout the specification. They will also recognize that modules can be implemented as circuits, logic chips, or any type of discrete component. Additionally, they will recognize that modules can be implemented as software, which can then be executed by different processor architectures. In embodiments of this disclosure, modules may also include computer instructions or executable code that can instruct a computer processor to perform a series of events based on received instructions. The choice of module implementation is a design choice made by those skilled in the art and does not in any way limit the scope of the claimed subject matter.

[0028] Furthermore, those skilled in the art will recognize that the detailed workings, internal structures, techniques, and training datasets of the artificial intelligence models, neural networks, and anomaly assertion models mentioned in this disclosure have not been disclosed or shown in detail, as such information is known to those skilled in the art and is therefore omitted for the sake of brevity.

[0029] Artificial intelligence (AI) models, such as but not limited to convolutional neural networks (CNNs), deep neural networks (DNNs), recurrent neural networks (RNNs), or other types of AI models, are developed through a structured training process. This process involves feeding training data from the content to be learned to a learning algorithm—the mechanism by which the model learns. The training data typically contains correct responses, often referred to as "ground facts," which are called the target or target attributes. During training, the learning algorithm feeds this data to the AI ​​model, which then produces corresponding outputs. The algorithm then evaluates the performance of the AI ​​model by calculating the loss or error relative to the correct answer and subsequently adjusts the model's parameters to minimize that loss or error. Essentially, the learning algorithm identifies patterns in the training data that associate input data attributes with the target, resulting in a trained AI model that encompasses these identified patterns.

[0030] In task orchestration platforms that manipulate and process large datasets, critical operations such as rapid response, data analysis, and monitoring heavily rely on the effective management of diverse and massive amounts of data from various sources. These sources tend to generate large volumes of data, which may include video, audio files, temperature measurements, pressure measurements, olfactory measurements, images, text data, and non-text data, where these different types of data files typically include multiple or various formats. In embodiments of this disclosure, video data can be obtained using image acquisition devices, such as cameras configured to acquire and record real-time video feeds; audio data can be obtained using audio acquisition devices, such as microphones and audio sensors configured to acquire real-time audio feeds, audio recordings, or ambient sound data; temperature and pressure data can be obtained using measurement devices, such as temperature or pressure sensors configured to provide real-time temperature or pressure readings; and text data can be obtained from various sources, such as reports, event logs, text messages, and emails. Those skilled in the art will recognize that various other methods can be used to acquire and record other types of data.

[0031] Specifically, video streams provide visual perspectives of various areas, while audio feeds offer auditory cues and context. Simultaneously, temperature and pressure sensors transmit critical environmental data. Furthermore, text data sources provide insights and reports related to specific matters. The robust setup of the task orchestration platform enables seamless integration and analysis of all these data streams, providing a comprehensive picture of the ecosystem to ensure that no potential anomalies or irregularities are overlooked.

[0032] To handle large amounts of information from diverse sources, the data collected from these sources is aggregated into multiple data groups. Each group stores the data as a single entity in the database system, thus accommodating the storage of large amounts of data elements, such as multimedia files, without requiring the data to conform to the strict types required by regular database fields.

[0033] In embodiments of this disclosure, data sets are temporarily stored in a cache, such as a memory buffer, instead of storing the data sets directly in the database system. The data cache serves several purposes: it improves the speed of data retrieval and processing, allowing for faster operational responses; it reduces the load on the storage system; and it prevents redundant processing of the same data multiple times.

[0034] Figure 1 A block diagram of a system for autonomous verification and data triage according to an embodiment of the present disclosure is shown. System 100 includes a computing module 102 configured to process data 101 before providing the processing results to a data management system 114. In embodiments of the present disclosure, computing module 102 includes a database 103, a feature extraction module 104, a ranking module 106, a classification model selector module 108, and databases 110 and 112.

[0035] In embodiments of this disclosure, database 103 may include a cache configured to store data group 101. The cache may include high-speed memory components, which are typically smaller and faster than the main memory of computing module 102, and operate at processing speeds closer to those of the processor in computing module 102, thereby reducing latency and improving the overall performance of the computing module. Therefore, the cache may include static random-access memory (SRAM) or dynamic random-access memory (DRAM), depending on the cost and performance requirements of the application.

[0036] In embodiments of this disclosure, the feature extraction module 104 is configured to select a feature extraction model from a set of feature extraction models stored in a database within module 104. Each of these models is configured to extract features from different types of media, such as, but not limited to, images, videos, audio clips, and text logs. This set may include AI models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs) and their variants, and natural language processing (NLP) models. The convolutional neural networks can be used to identify objects, patterns, and activities in image and video data. The recurrent neural networks and their variants are, for example, long short-term memory (LSTM) networks used to analyze audio clips for speech recognition or anomaly detection. The NLP models are used to process and extract meaningful information from text logs, reports, and emails. Detailed descriptions of the operation of these feature extraction AI models are omitted for brevity, as they are well-known in the art.

[0037] Based on the format and / or type of the monitoring data to be processed by the calculation module, the feature extraction module 104 will autonomously select an appropriate feature extraction model to be applied to the data. This selection can be based on a preset, rule-based method, which may involve preset criteria such as data type (i.e., data characteristics), source, or specific analysis requirements. For the sake of brevity, implementation details of such rule-based methods are omitted, as they are known to those skilled in the art.

[0038] In a further improvement to the above embodiments, when the data block retrieved from database 103 includes image files, feature extraction module 104 can retrieve a convolutional neural network (CNN) type feature extraction model from its database, since CNNs are typically used to automatically detect important features without any human supervision. In a typical CNN, the input image passes through a series of convolutional layers that apply different filters (kernels) to detect features such as edges, colors, textures, and other spatial levels. These features are then pooled (typically by max pooling) to reduce dimensionality, and the image is classified or encoded into a feature vector through multiple fully connected layers. This vector captures the basic features of the image and can be used for further processing or analysis. In embodiments of this disclosure, a You-Only-Look-Once (YOLO) object detection model, which constructs object detection as a single regression problem, can also be used, or a Single Shot Multi-box Detector (SSD) model can also be used to detect and extract features from the image.

[0039] In a further embodiment, when the input data includes video files, the feature extraction module 104 can retrieve a feature extraction model from a database within module 104. This feature extraction model includes a CNN architecture (as described above), which has been further extended to encompass 3D convolutional neural networks, CNNs adapted for spatial feature extraction, recurrent neural networks (RNNs), or long short-term memory networks (LSTMs). 3D CNNs are typically used in situations requiring simultaneous processing of multiple frames to capture both spatial and temporal features, thereby identifying patterns related to both time and space. Alternatively, a CNN can process a single frame to extract spatial features, and then utilize an RNN or LSTM to analyze the sequence of feature vectors generated by the CNN, thereby capturing dynamics and movement related to time.

[0040] In another embodiment of this disclosure, when the input data includes an audio file, the feature extraction module 104 can retrieve a feature extraction model from a database based on Mel-frequency cepstral coefficients (MFCCs) or a CNN-based spectrogram. MFCC is a feature extraction model that effectively represents the short-term power spectrum of sound based on the linear cosine transform of the logarithmic power spectrum at a nonlinear Mel-frequency scale. MFCC is widely used in speech and audio processing because it more closely simulates the response of the human auditory system than the linearly spaced frequency bands used in ordinary Fourier transforms. Alternatively, the audio file can be converted into a spectrogram, and features can be extracted from the spectrogram using a feature extraction model including the CNN described above, similar to the processing of image data files. This extracts robust features of both acquired spectral and temporal characteristics.

[0041] In embodiments of this disclosure, ranking module 106 may include a ranking model configured to rank items, features, or entities based on their relevance or importance relative to a particular query or task. A basic ranking model includes a scoring function configured to assign a numerical score to each feature based on its importance or relevance, and this function may include, but is not limited to, a linear combination of weighted features or a deep neural network trained to process input features in a non-linear manner. In general, ranking models can be categorized into three types based on how they assess the relevance between items: pointwise, pairwise, and listwise. Pointwise models treat the ranking problem as a regression or classification task, where they predict relevance scores or independently classify the relevance of individual items. For example, RankNet's pairwise models focus on correctly ranking pairs of items by learning from comparisons; they predict that one item in a pair is more relevant than the other and are commonly used in search engine results ranking settings. Listwise models consider the entire list of items and aim to optimize the overall order of all items; examples include ListNet and LambdaRank, which are particularly effective in scenarios where the overall order of items affects performance, such as in search engine ranking or recommendation lists. Those skilled in the art will recognize that the choice of ranking model depends on the characteristics to be ranked.

[0042] In embodiments of this disclosure, the classification model selector module 108 may include various selector models and / or benchmarking models, which facilitate the management and optimization of the selection and deployment of AI models within one model group or inference path protocols within another model group. Each selector model may be configured to select one or more models from the model group to process specific input data using different strategies. This selection may be random, ordered (e.g., round-robin), rule-based, or based on more complex rules, such as frequency-based partitioning, attribute-based partitioning, or multi-armed bandit strategies for A / B testing. Furthermore, the selector module may be configured to dynamically improve its selection rules based on prior performance data. This dynamic selection is typically guided by specific business objectives or performance metrics, such as feedback or accuracy derived from operators, which serves as a scoring metric for evaluating the effectiveness of the deployed models. For brevity, detailed operations of the classification model selector are omitted, as these are known to those skilled in the art.

[0043] In a further embodiment of this disclosure, the classification model selector module 108 may include an evaluator model. The evaluator model can be configured to evaluate different models offline before the selector module selects a model. The evaluator model scores the outputs of the selected AI models on a test dataset using accuracy metrics such as confusion matrices and the area under the receiver operation characteristic (ROC) curve, thereby achieving offline evaluation of different models. This evaluation helps compare different models or different versions of the same model to determine the best-performing model or version under given conditions. The evaluation process may also include visualization tools provided by the artificial intelligence platform to assist in decision-making. The results generated by the evaluator module can then be used to influence the selection process of the selector model.

[0044] In embodiments of this disclosure, database 110 may include multiple inference path protocols. At this stage, it is noteworthy that in many applications, such as object recognition and classification use cases, it may be necessary to classify input into a large number of categories, potentially exceeding tens of thousands. Training a single classification model to handle such a large array of categories presents significant challenges, in addition to the often difficult convergence, excessively long training time, and high demands on computational resources and memory. Furthermore, even if training is technically feasible, the practicality of using a single model for such a wide range of classification or inference tasks is generally questionable due to these limitations.

[0045] Therefore, to address these challenges, in embodiments of this disclosure, an inference path protocol or model can be used instead of conventional large and complex AI models. The inference path protocol employs multiple classification models organized into several layers within an inference path structure, where input is provided to basic nodes of the path structure, and where a final output is generated after the input has selectively traversed a series of branch paths. Each node within the branches of this multi-layered path structure includes a distinct classification model (i.e., a binary classification model), which can be of any type, ranging from linear classifiers such as logistic regression or naive Bayes to more complex models such as support vector machines (SVMs), decision trees, nearest neighbor classifiers, neural networks, or large language models. These models can be independently selected, trained, and updated according to the needs of each node on the branch paths. Typically, each node, except for the terminal node (i.e. the final node at the bottom of the inference path structure), operates as a binary classification model, which divides the input into two categories or groups, ideally containing an equal number of classes, but not necessarily an equal number of classes.

[0046] The inference path structure allows setting the path depth (i.e., the number of path branches) and the number of classes processed by each route node, typically with constraints to make training manageable and efficient (e.g., fewer than 50 layers, where each route node processes fewer than 20 classes per group). Currently, there are numerous classification models, each trained to classify specific types of objects and / or features. Without instantiation of the inference path protocol described above, these multiple models cannot be applied efficiently. The inference path framework facilitates the appropriate selection of a suitable classification model and its application to the thoroughly investigated objects. Therefore, the framework enables parallel processing across the entire path, starting with the identification of the highest-ranking object.

[0047] The introduction of this inference path protocol not only simplifies the classification process but also optimizes computational resources by enabling simultaneous processing of multiple paths. By systematically prioritizing object identification based on object ranking, the efficiency of the classification system is significantly improved, highlighting the importance of this model in object classification methods.

[0048] In other words, this structure ensures that the models at the route nodes are relatively easy to train, and that the overall system does not have excessive depth, thus allowing for a fast and predictable classification process. This multi-level path approach not only enhances the manageability of training with a large number of categories, but also improves the speed and accuracy of classification tasks, making it well-suited for applications that require distinguishing many categories, such as systems designed to understand and respond to the intentions of different users based on their input messages.

[0049] In most implementations of inference path protocols, the path structure for each protocol is established before the selection and training of the classification model at each node. This process typically begins by classifying all available classes into two roughly equal groups. These groups are then assigned to two child nodes at subsequent layers of the path model. This splitting process is repeated at each layer, continuously dividing multiple classes into nearly equal subsets until the number of classes at each node falls below a predetermined threshold.

[0050] This method effectively creates a near-binary pathway structure. At each non-terminal node in this pathway structure, a binary classification model is selected and trained. These binary classifiers have the advantage of being faster to train and using less memory compared to models handling larger sets of classes. This simplified approach facilitates a fast and efficient training process, which is crucial for managing the extensive and complex datasets typically encountered in such systems.

[0051] At the end points of the path structure—the terminal nodes—a multi-class classification model is implemented. Using a multi-class model at these points is strategically significant because it limits the overall complexity or depth of the path structure, ensuring that the classification process remains efficient even as it reaches the finer granularity required to reach these terminal nodes. This setup allows the system to handle matrices of various classes while maintaining manageable computational demands. Furthermore, the model at each node of the tree can be trained in parallel or sequentially.

[0052] When querying the inference path structure, the input to the structure is provided to a base node. Starting from this base node, the input is processed by the corresponding route nodes (i.e., the classification model), and the input is guided through layers based on the logic set at each node until the relevant terminal node or the required depth is reached. This method allows for efficient data filtering and extraction because each node serves as a decision point that progressively refines the search or decision-making process, directing the query path towards the most relevant subset of data.

[0053] The ability to compartmentalize and organize data in a multi-level manner significantly improves the effectiveness of query inference path structures, which essentially reduces the search space and improves query response time. This multi-level feature allows for logical, progressive refinement during the query process, ensuring that each step is focused on and based on criteria accumulated from previous nodes, thereby optimizing the query process for both accuracy and efficiency.

[0054] In embodiments of this disclosure, database 112 may include AI models trained to perform higher-level classification based on multiple inputs. Specifically, these trained AI models are configured to determine events or anomalies in multiple feature representations provided to the models, such as unauthorized access, intrusion, or any other anomaly. These trained AI models may include, but are not limited to, Support Vector Machines (SVMs), CNNs, Random Forests, Neural Networks, Anomaly Detection Models, etc. In the context of identifying anomalous events, SVMs can classify complex features extracted from images into categories such as "normal" and "abnormal" by finding a hyperplane that best separates different categories with maximum margins, while CNNs can be trained on labeled data of both normal events and compromising incidents to directly learn discriminative features. Random Forest models are particularly advantageous in situations where reliable decision-making is required under changing and potentially noisy conditions. Ensemble methods help address overfitting and provide a practical way to assess the importance of different features for classification, while neural network models can learn from large amounts of labeled data (images labeled as normal or representing events) to discern subtle patterns that can represent anomalies. Finally, when large amounts of images cannot be labeled, anomaly detection models can be used. These models learn the meaning of “normal” combinations of features and then label deviations from that norm as potential events. Anomaly detection models can use autoencoders, which reconstruct normal events and highlight anomalies by detecting reconstruction errors; this type of autoencoder is a neural network.

[0055] refer to Figure 2 During operation, when continuous input of data 101 is provided to the computing module 102, the data 101 is converted into data groups, such as groups 101a to 101d, and then groups 101a to 101d are stored in the database 103 as data groups 203a to 203n. In embodiments of this disclosure, each data group may include data with the same format and / or the same type, or may be grouped according to data markers used to indicate the start point of the sequence of received data. The feature extraction module 104 may then extract the data groups in an ordered manner or based on a set of rules. The feature extraction module 104 then autonomously selects an appropriate feature selection model to be applied to each extracted data group based on the format and / or type of the data.

[0056] Once module 104 has identified features of a set of data (e.g., data 203a), these identified features are ranked by ranking module 106 based on preset ranking criteria. Module 106 is configured to classify items, features, or entities based on their relevance or importance to a specific query or task. For example, if module 106 is ranking objects / features generated from image files, it can rank features based on their size or position in the image; if module 106 is ranking features derived from video files, it can rank features based on their frequency of occurrence across several video frames, in addition to their size or position; and if module 106 is ranking features derived from audio files, it can rank features based on their decibel level or frequency of occurrence within a timeframe. Those skilled in the art will recognize that these are merely examples, and module 106 can utilize other parameters that can be determined by those skilled in the art to rank features of a data set.

[0057] In embodiments of this disclosure, the ranking module may include a convolutional neural network (CNN) because such neural networks are highly effective for tasks involving image ranking due to their inherent ability to autonomously extract and learn features directly from image data. These networks consist of convolutional layers that apply filters to the input, which extract fundamental features such as edges, textures, and more complex patterns at deeper layers. This capability makes CNNs particularly well-suited for analyzing visual content in images, which is crucial for ranking images based on various criteria such as aesthetic quality, size, relevance, or specific content. The training process involves using a large dataset of pre-ranked or rated images, where the network learns to associate specific visual features with their corresponding rankings through a series of training and validation steps to effectively tune the network's parameters. Once training is complete, the CNN can then be deployed to dynamically perform the image ranking task.

[0058] Then, the calculation module 102 will classify each feature identified by the module 104 based on the ranking assigned to each of these features. In embodiments of this disclosure, features with more important rankings (i.e., features with higher ranking scores) will be classified first, and more computational resources may be allocated to those higher-ranking features as needed. Those skilled in the art will recognize that the order can be switched or changed, and such changes in processing order are still covered in this disclosure.

[0059] In the process of classifying each feature, the computation module 102 can utilize the classification model selector module 108 to select an inference path protocol from the database 110. This inference path protocol will be used to classify the feature selected from multiple identified features. For example, if five different unique features need to be classified, the selector module 108 can select up to five different types of inference path protocols from the database 110, where each model is specifically designed to classify one of these unique features.

[0060] After classifying all identified features of the data set using their respective inference path protocols, all these features are then provided to the classification model selector module 108. The classification model selector module 108 continues to run to select from the database 112 the most suitable AI model for performing the inference task based on the features contained in the data set. The computation module 102 then performs operations based on the results generated by the selected, trained AI model and subsequently provides the processed data set to the data management system 114. This process is repeated for the next set of data (i.e., on data 203b) until all data sets have been processed as described above.

[0061] It should be noted that this approach is beneficial when a dataset contains multiple different types of objects; dedicated sub-models can be deployed to leverage data autonomy, independently focusing on specific types of features within the dataset. For example, if the dataset represents image files, one type of inference path protocol can be used to identify vehicles, while another type is better suited for detecting people. These models operate autonomously, applying their specialized knowledge only to relevant parts of the images, and then the results are aggregated at a higher level using more sophisticated AI models contained in database 112 to produce a cohesive output.

[0062] In summary, this method allows for autonomy in inference paths because individual parts of a set of data can be processed separately to identify unique attributes, which are then integrated into higher layers of the model to form a unified understanding of the data set. This method is particularly beneficial for datasets containing dense information, where different objects can overlap or interact in ways imperceptible to the human eye.

[0063] In another embodiment of this disclosure, where the data includes an image file, each classification object derived from the image file may also be tagged with the location of the classification object within the image file before the step of providing the identified features, which have been classified by their respective inference path protocols, to the classification model selector module 108.

[0064] Once this is complete, a graph of these categorized objects can be constructed, with each categorized object corresponding to a node in the graph. Each of these nodes can encompass different attributes of the object, including the object's dimensions, coordinates, and any other features extracted for the reasoning path used to categorize the object. For example, if an image file contains five unique objects categorized in the previous step, each categorized object would then be represented as a node in the graph.

[0065] In the diagram, the edges between nodes are established based on the spatial relationships between each categorized object. For example, if two objects are close enough to each other, or their bounding boxes overlap, a connection or edge is formed between them to represent their spatial relationship. These edges encode information about the proximity and relative position of objects within the image. For instance, an edge can connect two objects in the image that are adjacent to each other or close in proximity.

[0066] A graph neural network (GNN) architecture can be obtained from database 112 and then used to process the graphs of detected objects. The GNN takes node features (object attributes) and edge features (spatial relationships) as input, propagating information between nodes and edges to capture spatial dependencies. During inference, the trained GNN model predicts spatial relationships between multiple objects based on the features of multiple objects and contextual information encoded in the graph structure. For example, a message-passing GNN architecture can be configured to iteratively update node representations based on information received from neighboring nodes, allowing it to infer complex spatial relationships between multiple objects. The GNN used in this embodiment can be trained based on previously generated annotated data, where spatial relationships between different types of objects are already defined. For brevity, the detailed working process of the GNN is omitted, as the structure of standard GNNs is well known to those skilled in the art.

[0067] Then, the previously identified features, classified through their respective inference path protocols, and the predicted spatial relationships between the classified objects can be provided to the classification model selector module 108. The classification model selector module 108 then continues to run to select from the database 112 an AI model capable of performing or carrying out an operable inference task based on the provided input. The computation module 102 then performs operations based on the results generated by the selected, trained AI model.

[0068] The combination of the final classification output obtained by the inference path protocol and the spatial relationships between the classified objects represents a significant advancement in artificial intelligence methods. The inference path protocol excels at breaking down complex classification challenges into more manageable parts, enabling accurate identification of objects of different categories. By further combining these detailed classification outputs with spatial relationship data, the resulting AI model not only gains the ability to identify objects but also to understand their interactions or relative positions within a given space. This two-layer approach significantly improves the accuracy and contextual depth of the analysis. By understanding the context of how multiple objects are spatially interconnected, the model can make informed predictions and decisions that take into account the dynamic environment of multiple objects.

[0069] Furthermore, the ability to recognize complex patterns arising from both object classification and spatial relationships enables machine models to perform more sophisticated analyses, as pattern recognition can be extended beyond simple object recognition to include object grouping, temporal movement patterns, and spatial distribution analysis. Moreover, the incorporation of spatial relationships helps reduce the inherent biases of models trained primarily on appearance features. These biases can lead to misclassification or incorrect predictions when models rely solely on visual identifiers without considering the contextual meaning of objects within their environment. By integrating spatial context, the model's reliance on potentially biased visual features is reduced, improving the model's generalizability and accuracy.

[0070] In another embodiment of this disclosure, prior to providing the identified features, already classified through their respective inference path protocols, to the classification model selector module 108, the computation module 102 may perform chaining across multiple paths using different inference path protocols to identify any paths from the multiple inference path protocols that could lead to the same final classification output for different features. Chaining across multiple paths essentially involves identifying paths in different inference path protocols that result in the same final classification outcome. Once convergent paths of different inference path protocols have been identified, the computation module 102 aggregates information from each route of each different inference path protocol to identify features or attributes that influence the classification decisions at nodes along each path. The computation module 102 then analyzes the consistency of classification decisions across these multiple paths to generate a confidence value in the final classification result. If the aggregation of information across different paths shows significant consistency in the reasoning or feature importance used in different AI models at each route node, this means that the classification is likely performed correctly and the result is reliable. The features that have been classified through their respective inference path protocols can then be provided to the classification model selector module 108 to perform operations as described above.

[0071] Conversely, if multiple routes exhibit significant differences in their logical analysis or feature importance, it may indicate potential weaknesses or ambiguities in the classification criteria used in the respective paths. The previously obtained classification results for associated features can then be discarded, and the computation module 102 can trigger the classification model selector module 108 to select other inference path protocols from the database 110 to perform another round of classification on the affected features. This feedback loop, triggered by the chained operation across multiple paths with different inference path protocols, ensures that new changes or complexities in the data can be handled more accurately and efficiently. Features already classified by their respective inference path protocols can then be provided to the classification model selector module 108 for operation, as described above.

[0072] According to embodiments of this disclosure, Figure 3 A block diagram illustrating components of processing system 300 is shown. These components may be provided in computing module 102 or any module provided within computing module 102 to perform any of the functions described above. Those skilled in the art will recognize that the exact settings of each processing system provided within these modules may differ, and the exact settings of processing system 300 may be varied. Figure 3 The settings shown are provided as examples only.

[0073] In embodiments of the invention, the processing system 300 may include a controller 301 and a user interface 302. The user interface 302 is configured to enable human interaction between the user and the computing modules as needed, and for this purpose, the user interface 302 includes user input instructions to provide each of these modules with the input / output components required for updates. Those skilled in the art will recognize that the components of the user interface 302 may vary from embodiment to embodiment, but generally include one or more of a display 340, a keyboard 335, and optical devices 336.

[0074] The controller 301 communicates with the user interface 302 via a bus 315. The controller 301 includes a memory 320, a processor 305 mounted on a circuit board, an operating system 306, an input / output (I / O) interface 330 for communicating with the user interface 302, and a communication interface. The processor 305 processes instructions and data for executing the methods of this embodiment. In this embodiment, the communication interface is in the form of a network interface card (NIC) 350. For example, the NIC 350 can be used to send data from these modules to other processing devices via wired or wireless networks, or to receive data via wired or wireless networks. Wireless networks that the NIC 350 can use include, but are not limited to, Wi-Fi, Bluetooth, Near Field Communication (NFC), cellular networks, satellite networks, telecommunications networks, and Wide Area Networks (WANs).

[0075] Memory 320 and operating system 306 communicate with CPU 305 via bus 310. The memory components include volatile and non-volatile memory, and multiple memories within each type, including Random Access Memory (RAM) 323, Read Only Memory (ROM) 325, and mass storage device 345, which includes one or more solid-state drives (SSDs). Those skilled in the art will recognize that the aforementioned memory components include non-transitory computer-readable media and should be considered to include all computer-readable media except for transiently propagated signals. Typically, instructions are stored in the memory components as program code, but they can also be hardwired. Memory 320 may include a kernel and / or programming modules, such as software applications that may be stored in volatile or non-volatile memory.

[0076] The term "processor" as used herein is generally used to refer to any device or component capable of processing the instructions, and may include: microprocessors, microcontrollers, programmable logic devices, or other computing devices. In other words, processor 305 may be provided by any suitable logic circuitry for receiving input, processing the input according to instructions stored in memory, and generating output (e.g., output to a memory component or output to a display 340). In this embodiment, processor 305 may be a single-core or multi-core processor with memory addressable space. In one example, processor 305 may be multi-core, including, for example, an 8-core CPU. In another example, processor 305 may be a cluster of multiple CPU cores operating in parallel to accelerate computation.

[0077] Figure 4 The process of identifying abnormal structures in real time based on monitoring data is illustrated, wherein, according to embodiments of the present disclosure, method 400 can be executed by a computing module in conjunction with modules contained within the computing module.

[0078] Method 400 begins at step 402, where it extracts data from a database. Then, method 400 proceeds to identify features in the data using a feature extraction model in step 402. Then, at step 404, method 400 uses the feature extraction model to identify features in the data. Then, method 400 generates a ranking for each identified feature, and this is performed in step 406. Then, method 400 selects the feature with the highest ranking in step 408. Then, method 400 uses a first classification model selector to select an inference path protocol based on the features selected from the set of inference path protocols, and proceeds to classify the selected features in step 410.

[0079] If method 400 determines in step 412 that there is another unclassified feature related to the data, then method 400 proceeds to step 414 to select the next feature with the next highest ranking. Then, method 400 proceeds to step 410, where method 400 uses a first classification model selector to select an inference path protocol from the set of inference path protocols based on the selected feature, and continues to classify the selected feature in step 410.

[0080] In embodiments of this disclosure, the first classification model selector can be trained based on a set of rules, and each inference path protocol can include multiple classification models set in a multi-layered path structure. The input to each inference path protocol is provided to a basic node of the corresponding path structure, and the final classification output is generated by a multi-class classification model located at a terminal node of the corresponding path structure. Furthermore, each layer of the path structure has route nodes that can include binary classification models.

[0081] For example, when a feature to be classified is provided to a base node of the inference path protocol, the base node, which includes a binary classification model, then evaluates whether the feature should align with a first set of classes or a second set of classes. The result of this evaluation can then be represented by two probabilities—the probabilities representing the likelihood that the feature belongs to either set of classes. The method then moves to another route node associated with a class having a higher probability value (e.g., the first set of classes), located at another layer in the inference path structure. At this other route node, another binary classification model is used to evaluate whether the input feature should align with a first or second subset of classes in the first set of classes. The method then continues through the layers of the inference path structure until it reaches a terminal node, where it proceeds to generate the final classification output.

[0082] If method 400 determines in step 412 that there is another data-related feature that has not yet been classified, then method 400 continues to step 414. Steps 414 to 412 are repeated until all features identified in step 404 have been classified.

[0083] Then, method 400 proceeds to step 416, where method 400 uses a second classification model selector to select a trained AI model from the set of trained AI models based on the combination of the final classification outputs generated in step 410. Then, method 400 performs an inference task based on the combination of classification outputs. In step 418, method 400 performs operations on the data based on the results of step 416 and appends the results to the data for further processing.

[0084] In embodiments of this disclosure, each classification model for route nodes in the set of each inference path protocol is trained to classify the input into two groups of categories, and a multi-class classification model is trained to classify the input into a single category.

[0085] In embodiments of this disclosure, the extracted data may include image files, and method 400 may proceed to determine the spatial relationships between all final classification outputs generated by the inference path protocol prior to the step of selecting a trained AI model from the set of trained AI models.

[0086] In a further embodiment, method 400 may use the spatial relationships between all final classification outputs and the final classification outputs generated by all inference path protocols to select the trained AI model from a set of trained AI models.

[0087] In another embodiment, when method 400 classifies features using a selected inference path protocol, method 400 may perform a chaining operation on multiple paths of inference path protocols that generate similar final classification outputs. Then, when a difference is determined from the chaining operation results in the linked paths, method 400 may proceed to discard the generated, similar final classification outputs. Method 400 then proceeds to trigger a first classification model selector to select another inference path protocol from the set of inference path protocols to classify the features associated with the discarded final classification output.

[0088] In another embodiment of this disclosure, the classification model at the route nodes may include a support vector machine classifier, a neural network classifier, a nearest neighbor classifier, a decision tree classifier, a logistic regression classifier, or a Naive Bayes classifier, and the multi-class classification model may include a neural network classifier, a support vector machine classifier, a K-nearest neighbor classifier, or a large language model. Method 400 may also be configured to perform the classification of each feature of the data in parallel.

[0089] Those skilled in the art can identify many other changes, substitutions, alterations, and modifications, and this application is intended to cover all such changes, substitutions, alterations, and modifications that fall within the scope of the appended claims.

Claims

1. A calculation module for autonomously verifying and triaging data in a database, the calculation module comprising: Processing unit; as well as The processing unit can read a non-transitory medium, the medium storing instructions, which, when executed by the processing unit, cause the processing unit to: Extract data from the database; The data is processed using a feature extraction model to identify multiple features in the data; Based on a preset ranking standard, each of the identified features is ranked. For each feature, based on its labeled rank, a first classification model selector is used to select an inference path protocol from a set of inference path protocols, wherein the first classification model selector is trained based on a set of rules related to the characteristics and data type of the feature. The features are classified using the selected inference path protocol, wherein each inference path protocol includes: A hierarchical, multi-level path structure, comprising a basic node for receiving input, and... Multiple classification models are set as route nodes at each layer of the path structure, such that each route node processes the input data and branches into further paths until a terminal node is set with a multi-class classification model that produces the final classification output is reached. Aggregate the final classification output generated by the inference path protocol; Using a second classification model selector, a trained artificial intelligence model is selected from a collection of artificial intelligence models stored in a database, the selection being based on the final classification output of the aggregation; The selected artificial intelligence model is used to perform an inference task based on the final classification output of the aggregation; and Based on the results of the inference task, operations are performed on the data.

2. The computing module according to claim 1, characterized in that, Each of the plurality of classification models, which are set as route nodes in the set of the inference path protocol, is trained to classify the input into two groups of categories, and is characterized in that the multi-class classification model is trained to classify the input into a single category.

3. The computing module according to claim 1, characterized in that, The retrieved data includes image files, and is characterized in that, prior to the instruction to select a trained artificial intelligence model from a set of trained artificial intelligence models, the computing module further includes instructions for instructing the processing unit to perform the following operations: Determine the spatial relationships among all the final classification outputs generated by the inference path protocol, and provide the spatial relationships to the second classification model selector.

4. The computing module according to claim 3, characterized in that, The instructions for selecting the trained artificial intelligence model from the set of trained artificial intelligence models also include instructions for instructing the processing unit to perform the following operations: Using the spatial relationships of the determinations among all the final classification outputs and the final classification outputs generated by all the inference path protocols, the trained artificial intelligence model is selected from the set of trained artificial intelligence models.

5. The computing module according to claim 1, characterized in that, The instructions for classifying the features using the selected inference path protocol also include instructions for instructing the processing unit to perform the following operations: Chaining operations are performed on multiple inference path protocols that generate similar final classification outputs.

6. The calculation module according to claim 5, characterized in that, The instructions for performing chained operations along multiple paths also include instructions for instructing the processing unit to perform the following operations: When the chained operation results indicate that there are differences in the linked paths, the generated, similar final classification outputs are discarded; and The first classification model selector is triggered to select another inference path protocol from the set of inference path protocols to classify the features associated with the discarded final classification output.

7. The computing module according to claim 1, characterized in that, The multiple classification models set as route nodes can each include a support vector machine classifier, a neural network classifier, a nearest neighbor classifier, a decision tree classifier, a logistic regression classifier, a Naive Bayes classifier, or a large language model.

8. The computing module according to claim 1, characterized in that, The multi-class classification model may include a neural network classifier, a support vector machine classifier, a K-nearest neighbor classifier, or a large language model.

9. The computing module according to claim 1, characterized in that, The processing unit performs the classification of each feature of the data in parallel.

10. A method for performing self-checking and triage of data in a database using a computing module, the method comprising: Extract data from the database; The data is processed using a feature extraction model to identify multiple features in the data; Based on a preset ranking standard, each of the identified features is ranked. For each feature, an inference path protocol is selected from a set of inference path protocols using a first classification model selector, wherein the first classification model selector is trained based on a set of rules related to the characteristics and data type of the feature. The features are classified using an implementation-selected inference path protocol, wherein each inference path protocol includes: A hierarchical, multi-level path structure, comprising a basic node for receiving input, and... Multiple classification models are set as route nodes at each layer of the path structure, such that each route node processes the input data and branches into further paths until a terminal node is set with a multi-class classification model that produces the final classification output is reached. Aggregate the final classification output generated by the inference path protocol; Using a second classification model selector, a trained artificial intelligence model is selected from a collection of artificial intelligence models stored in a database, the selection being based on the final classification output of the aggregation; The selected artificial intelligence model is used to perform an inference task based on the final classification output of the aggregation; and Based on the results of the inference task, operations are performed on the data.

11. The method according to claim 10, characterized in that, Each of the multiple classification models, which are set as route nodes in the set of inference path protocols, is trained to classify the input into two categories, and is characterized in that the multi-category classification model is trained to classify the input into a single category.

12. The method according to claim 10, characterized in that, The retrieved data includes image files, and the method is characterized by further comprising the following steps prior to the step of selecting a trained artificial intelligence model from a set of trained artificial intelligence models: Determine the spatial relationships among all the final classification outputs generated by the inference path protocol, and provide the spatial relationships to the second classification model selector.

13. The method according to claim 12, characterized in that, The step of selecting the trained artificial intelligence model from the set of trained artificial intelligence models further includes the following steps: Using the spatial relationships between all the final classification outputs, the determinations, and the final classification outputs generated by all the inference path protocols, the trained artificial intelligence model is selected from the set of trained artificial intelligence models.

14. The method according to claim 10, characterized in that, The step of classifying the features using the selected inference path protocol further includes the following steps: On a protocol that generates multiple inference paths with similar final classification outputs, chain operations are performed on multiple paths.

15. The method according to claim 14, characterized in that, The steps for performing chained operations on multiple paths also include the following: When a difference is determined in the path of the chained operation based on the results, the generated, similar final classification output is discarded; and The first classification model selector is triggered to select another inference path protocol from the set of inference path protocols to classify the features associated with the discarded final classification output.

16. The method according to claim 10, characterized in that, The multiple classification models set as route nodes can each include a support vector machine classifier, a neural network classifier, a nearest neighbor classifier, a decision tree classifier, a logistic regression classifier, a Naive Bayes classifier, or a large language model.

17. The method according to claim 10, characterized in that, The multi-class classification model may include a neural network classifier, a support vector machine classifier, a K-nearest neighbor classifier, or a large language model.

18. The method according to claim 10, characterized in that, The classification of each feature in the data occurs in parallel.

Images