Method and System for Discrete Event Summarization and Prediction Using Large-Scale Language Models

Abstract

This paper relates to a method and system for predicting operational events in a manufacturing environment using a large-scale language model (LLM). [Solution] The method and system include receiving sensor data from multiple sensors at various stations that indicate machine parameters. An anomaly detection engine analyzes this data to identify anomalies. Metadata related to operational interruptions is recorded. The user enters prompts via an interface to access historical event data, anomalies, and metadata. Prompts are enhanced by historical data, output from a pattern knowledge graph, and predictions of future anomalies. The LLM processes these enhanced prompts to generate initial outputs, predict future anomalies or critical events, enhance decision-making, and reduce manual intervention.

Description

[Technical Field]

[0001] This disclosure relates to methods and systems for discrete event summarization and prediction using large-scale language models. In embodiments, these methods and systems are used for semantic pattern summarization and prediction in manufacturing environments. [Background technology]

[0002] background Recent advances in artificial intelligence have led to the development of models capable of handling complex tasks by utilizing deep learning architectures to process and generate human-like text. Despite these capabilities, these models face difficulties and can lead to inaccuracies when applied outside their direct domain. Techniques have been researched to improve performance by integrating information retrieval methods and enabling dynamic access to external databases for more accurate responses. [Overview of the project] [Problems that the invention aims to solve]

[0003] The application of these models to predict discrete events remains extremely limited, particularly in specialized fields such as manufacturing. While some modern manufacturing facilities employ sophisticated event monitoring systems to detect anomalies and issue alerts, these systems require manual assessment to determine the need for intervention, as anomalies do not necessarily lead to operational interruptions. [Brief explanation of the drawing]

[0004] [Figure 1] This figure shows a system for training a neural network according to one embodiment. [Figure 2] This figure shows an overall schematic diagram of the work line within the manufacturing equipment 200 according to one embodiment. [Figure 3A]A schematic diagram of one embodiment of a system configured to utilize a Large Language Model (LLM) to interpret the current status within a manufacturing environment and predict potential future events. [Figure 3B] A time series diagram of detected anomalies and predictions of future disruptions according to one embodiment. [Figure 4] A diagram of one embodiment of a computing platform for implementing the neural network algorithms and / or methodologies described herein. [Figure 5] A diagram of a method for predicting operational events within a manufacturing environment using an LLM according to one embodiment. **Modes for Carrying Out the Invention**

[0005] **Detailed Description** Embodiments of the present disclosure are described herein. However, it should be understood that the disclosed embodiments are merely examples and that other embodiments may take various alternative forms. The drawings are not necessarily drawn to scale, and some features may be exaggerated or reduced to show details of particular components. Accordingly, the specific structural and functional details disclosed herein should not be construed as limiting but rather as a representative basis for teaching those skilled in the art how to use the embodiments in various ways. As will be understood by those skilled in the art, the various features illustrated and described with reference to any one of the drawings can be combined with features shown in one or more other drawings to generate embodiments not explicitly illustrated or described. The combinations of features shown provide representative embodiments for typical applications. In certain applications or implementations, various combinations and modifications of features consistent with the teachings of the present disclosure may be desirable.

[0006] As used herein, "a," "an," and "the" refer to both singular and plural referents, unless the context clearly indicates otherwise. By way of example, a "processor" programmed to perform various functions refers to one processor programmed to perform any and all functions, or two or more processors collectively programmed to perform each of the various functions.

[0007] Large Language Models (LLMs) have become extremely important in dealing with complex question-and-answer tasks and skillfully selecting appropriate responses from a vast number of potential answers. LLMs such as OpenAI's GPT series are based on a deep learning architecture called the Transformer, which enables the processing and generation of text written like that of a human by learning patterns and relationships within large datasets. Despite their success, LLMs often face difficulties when operating outside their immediate domain of expertise, leading to errors or "hallucinations" that include inaccurate information generated by the model. To address these drawbacks, techniques such as Retrieval-Augmented Generation (RAG) have been studied. RAG combines the generative capabilities of LLMs with information retrieval methods, enabling the model to dynamically access an external database during the generation process to provide more accurate and contextually relevant responses.

[0008] However, the application of LLM to predict discrete events remains underdeveloped, particularly in specialized fields such as manufacturing. Manufacturing environments present significant challenges in predicting operational events due to the complexity and variability of machine parameters. Traditional systems often rely on manual monitoring and intervention, which can lead to inefficiencies and increased downtime. The need for advanced predictive analytics is emerging to improve decision-making processes and maintain operational continuity. Modern manufacturing equipment utilizes advanced event monitoring and recording systems such as Anomaly Detection Engines (ADEs). These systems capture anomalies as transient events and alert domain experts as a preemptive measure against potential critical incidents. Despite these advancements, the occurrence of anomalies does not always lead to operational interruptions, requiring manual assessment to determine the need for intervention. Furthermore, reliance on manual assessment can be slow and cumbersome, leading to delays and inaccuracies in addressing potential problems in manufacturing equipment, highlighting the limitations of current technologies in autonomously predicting and managing downtime. This gap highlights the critical need for advanced models that can autonomously describe the current status, predict operational interruptions, thereby enhancing the decision-making process and reducing reliance on manual assessment.

[0009] Accordingly, according to various embodiments herein, methods and systems for predicting operational events in a manufacturing environment using LLM are disclosed herein. By employing a zero-shot methodology, the system utilizes detailed prompts enhanced with historical event patterns and metadata. This approach enables LLM to generate accurate event summaries and predictions without specific prior training on the data. The system introduces a dual query framework, enabling domain experts to quickly check current operational status and predict future events, thereby improving operational decision-making and reducing manual intervention.

[0010] In embodiments, the method and system leverage LLM with a set of potential future events derived from historical data using frequent event pattern mining techniques. The method and system disclosed herein can enhance this dataset with additional metadata, including actual event patterns and metadata regarding event frequency and duration. This enhanced input allows the LLM to identify and select the most plausible future events based on a deep semantic understanding of the event context. The disclosed system provides a significant ability to accurately predict future operational interruptions, thereby streamlining decision-making processes within real-time manufacturing environments. This novel application of LLM in the field of discrete event prediction not only extends the utility of language models but also represents a significant advance in predictive analytics, particularly the automation of complex industrial activities.

[0011] In some embodiments, when evaluating a manufacturing process using LLM, a deep neural network (DNN) can form the foundation, enabling the model to understand and analyze complex patterns in manufacturing data. The core architecture may include transformers that excel at processing and learning from sequences, allowing the model to identify trends in time-series data, such as detecting potential disruptions in manufacturing or determining whether manufactured parts meet specifications. In some cases, a convolutional neural network (CNN) can be used to analyze visual data, such as images of parts, for quality control and extract local features for accurate analysis. In addition, a graph neural network (GNN) can be incorporated if the manufacturing process involves relationships between components or machines that can be represented as interconnected systems. This combination of neural networks enables LLM to provide accurate evaluations and insights, ensuring that the manufacturing process runs smoothly and the output meets quality standards.

[0012] Since the disclosed systems and methods rely on machine learning models such as neural networks (e.g., DNNs, GNNs), deep convolutional networks (DCNs), and CNNs, it may be useful to describe the use and training of such systems. Figure 1 shows a system 100 for training a neural network, such as a graphical neural network, which can be used with the machine learning models described herein. The neural networks illustrated and described herein are merely examples of usable machine learning networks or types of neural networks. System 100 may include an input interface for accessing training data 102 for the neural network. For example, as shown in Figure 1, the input interface can be configured by a data storage interface 104 that can access the training data 102 from data storage 106. For example, the data storage interface 104 may be a memory interface or persistent storage interface, such as a hard disk or SSD interface, or it may be a personal network interface such as a Bluetooth, Zigbee or Wi-Fi interface or an Ethernet or fiber optic interface, a local area network interface or a wide area network interface. The data storage 106 may be an internal data storage of the system 100, such as a hard drive or SSD, but it may also be an external data storage, such as a network-accessible data storage.

[0013] In some embodiments, the data storage 106 may further include a data representation 108 of an untrained version of the neural network that can be accessed from the data storage 106 by the system 100. It is understood that the training data 102 and data representation 108 of the untrained neural network may also be accessed from different data storages, for example, through different subsystems of the data storage interface 104. Each subsystem may be of the type described above for the data storage interface 104. In other embodiments, the data representation 108 of the untrained neural network may be generated internally by the system 100 based on the design parameters of the neural network and therefore may not be explicitly stored in the data storage 106. The system 100 may further include a processor subsystem 110 that can be configured to provide an iteration function in place of the layer stack of the neural network to be trained during the operation of the system 100. Here, each layer of the layer stack being replaced may have mutually shared weights and may receive as input the output of the previous layer, or, for the first layer of the layer stack, the initial activation and a portion of the input of the layer stack. The processor subsystem 110 can be further configured to iteratively train a neural network using training data 102. Here, the iterations of training by the processor subsystem 110 may include a forward propagation portion and a backpropagation portion. The processor subsystem 110 can be configured to perform the forward propagation portion by determining an equilibrium point of the iterative function at which the iterative function converges to a fixed point, where determining the equilibrium point includes finding the root solution by subtracting its input from the iterative function using a numerical root-finding algorithm, and by providing the equilibrium point instead of the output of the stack of layers in the neural network.

[0014] System 100 may further include an output interface for outputting a data representation 112 of a trained neural network. This data may also be referred to as trained model data 112. For example, as shown in Figure 1, the output interface may be configured by a data storage interface 104, which in this embodiment is an input / output ("IO") interface through which the trained model data 112 can be stored in data storage 106. For example, a data representation 108 defining an "untrained" neural network can be at least partially replaced by a data representation 112 of a trained neural network during or after training, and the parameters of the neural network, such as the neural network weights, hyperparameters and other types of parameters, can be adjusted to reflect training on the training data 102. This is also shown in Figure 1 by reference numbers 108,112, which refer to the same data records on data storage 106. In other embodiments, the data representation 112 can be stored separately from the data representation 108 defining an "untrained" neural network. In some embodiments, the output interface may be separate from the data storage interface 104, but generally, the data storage interface 104 may be of the type described above.

[0015] The structure of system 100 is an example of a system that can be used to train the neural network described herein. Additional structures for running and training machine learning models are shown in Figure 4 below.

[0016] Regarding the manufacturing process, the final product may pass through multiple workstations before the parts are completely finished or manufactured. For example, before the final product is manufactured, it may first need to be assembled with other subcomponents, painted, laser etched, strength tested, or perform other manufacturing tasks. After each station completes its task, measurements of the part can be taken to generate measurement data. This ensures that the part is fully functional, properly connected, and of sufficient size. The measurement data may include which type of station performed the measurement, what type of part was measured, and what the measurement was. The measurements may be binary values, strength values, time-series values ​​(e.g., measured response to pressure), floating-point numbers, digit sequences, integers, Booleans, statistical aggregates, etc., representing the physical state or characteristics of the part. This measurement data can be multimodal (e.g., it may include multiple types of measurements, such as those listed above as examples). This multimodal measurement data can be input into the neural network described herein. Depending on the measurements taken at the stations, the system can determine whether the part is sufficient or should be binned or discarded instead. This measurement data can also be referred to as sensor data (or a portion thereof). Sensor data can be received from one or more sensors at multiple stations in a manufacturing facility, and the sensor data indicates machine parameters associated with the machines located at those stations.

[0017] This multimodal measurement data, input into machine learning models (e.g., neural networks, GNNs, LLMs), can yield various benefits and provide a wealth of information that can assist with manufacturing lead times and logistics. For example, the output of a neural network can provide predictions about whether parts are sufficient for manufacturing or assembly into other systems, whether a station needs to be taken offline, yield time predictions, and predictions about the location where a failure may have occurred along the production line and the reason for the failure. In another example, the output of a neural network can provide predicted measurements at any station along the production line, and given this information, it is possible to eliminate dedicated stations (or procedures within those stations) for measuring the components being manufactured. This can save measurement time and costs.

[0018] Furthermore, predictive measurement of manufactured parts along the production line can reduce the costs associated with component disposal. If component measurements can be estimated within the production line (e.g., at any manufacturing station or between any manufacturing stations), this can lead to a more accurate determination of when failures or missteps will occur in manufacturing. This may mean discarding components earlier in the manufacturing process before it becomes more expensive to do so. Also, predicting component measurements before they are actually measured, depending on when the component is actually measured along the manufacturing process, allows for earlier discarding of components in the manufacturing process.

[0019] Figure 2 shows an overall schematic diagram of a work line within a manufacturing facility 200 according to one embodiment, where component 202 is manufactured at least partially. Component 202 can be any type of component undergoing a manufacturing process, and this disclosure is not limited to any particular type of component. Such components could be vehicles, parts of vehicles, gears, toothpaste containers, soap, food, etc.

[0020] Component 202 or part can move through the work line and be worked on by various machines 204. Each machine is located at a defined position (I1, I2, I3…i n The machines are located at ). Each machine also has a sensor (labeled "s") configured to sense data and generate measurement data, as described herein. Other locations (e.g., I4) can be defined along the work line, and these locations may also have their own sensors configured to generate measurement data. The sensors do not need to be equipped on a particular machine, but can generate measurement data at any defined location along the work line.

[0021] Figure 3A is a schematic diagram of one embodiment of a system 300 configured to interpret the current status and predict potential future events in the manufacturing environment using LLM. Additional details of such systems will be described further throughout, but generally, the process begins when a domain expert configures preliminary prompts using a user interface with pre-built templates or steps. The user interface can be implemented in the form of, for example, a tablet, computer, or smartphone. Simultaneously, these interfaces access historical event data from an event pattern database and other relevant sources via an application programming interface (API) incorporating comprehensive domain knowledge. After the domain expert finalizes the prompts and submits their requests, the interface forwards these prompts to the LLM. The LLM then processes the requests and generates results. Based on these initial results, the expert can refine the prompts to improve the accuracy and relevance of the LLM output. This iterative cycle of prompt creation and review continues as needed, enabling the expert to accurately assess the current state of the manufacturing process and make informed predictions about future events. In addition, the output from LLM can be integrated into other analytical tools or machine learning models for further analysis or to inspect specific stations along the production line.

[0022] The following explanation provides a more detailed description and overview of the concepts, along with the relevant terminology used in this report. Following this basic introduction, specific embodiments are explored, detailing the algorithms and methodologies used at each stage.

[0023] First, we will provide some basic definitions. The terms defined below are not intended to be limited to these definitions alone; rather, these definitions are intended to give more context to these terms.

[0024] Definition 1: Position. L={l1,l2,l3,…l n Let} be the set of positions within the manufacturing line, and each position l i It is uniquely identified by a location ID. The location ID may be, for example, spatial and temporal coordinates.

[0025] To ensure the integrity of the manufacturing process, monitoring systems such as Anomaly Detection Engines (ADEs) continuously monitor these readings from location in real time. These systems are often designed to incorporate predefined rules to detect various types of anomaly events, including outliers, abrupt changes in mean or variance, gradual mean shifts, and increases in zero values. Detected anomalies are recorded in a real-time database. Non-critical anomalies may not immediately affect manufacturing, while critical anomalies may indicate fundamental machine failures that could potentially lead to manufacturing interruptions. These interruptions can be recorded by separate systems and thus distinguished from anomalies tracked by these monitoring systems.

[0026] Definition 2: Event. Let E = A ∪ I be the set of all relevant events in the system, where A represents the set of anomalies and I represents the set of interruptions. Each event e ∈ E is characterized by a tuple (t, l, ...), where t represents the timestamp and l ∈ L represents the location of the event.

[0027] Definition 3: Anomaly. An anomaly a∈A is a special type of event characterized by a tuple (t,l,p,τ,α), where p represents the measured parameter, τ represents the detection technique, and α represents the anomaly level.

[0028] Definition 4: Interruption. An interruption i ∈ I is characterized by a tuple (t, l, d, ι, h), where d indicates the duration, ι indicates the type of interruption, and h is human-entered text describing the interruption. Data associated with an interruption can be stored as operation interruption data 416.

[0029] Recording each of these events is crucial for domain experts to manage potential disruptions. Each of these events can be detected and recorded using, for example, ADE410. However, displaying each event sequentially and alerting the user often leads to manual tracking and analysis, which can hinder ADE's goal of reducing user burden. To mitigate this, the concept of a Pattern Knowledge Graph (PKG) is introduced. By constructing event sequences from individual events within a specific period and summarizing these sequences as a graph, domain experts can easily track important events and identify correlations between anomalies and interruptions. In this disclosure, several parts of the PKG construction pipeline are utilized to construct event sequences using similar methodologies and components. Here, each sequence is defined to begin with an anomaly and further include interruptions within a specified time range and subsequent events including this interruption. This time range can be any customized or defined time range, e.g., one week, but is adjustable as needed.

[0030] Before defining event sequences, it is important to introduce the concept of event identifiers. These identifiers simplify the complex nested structures and metadata commonly found in manufacturing systems into alphanumeric codes that can be processed by pattern mining algorithms.

[0031] Definition 5: Event Identifier. An event identifier is a unique alphanumeric code used to label and distinguish different types of action events. Each identifier consists of a character prefix and a numeric suffix. The prefix indicates the type of event ("A" for abnormal, "I" for interruption), and the suffix is a unique number assigned sequentially. An abnormal event "A" can be detected via the execution of ADE410, and the data associated with an interruption can be stored as action interruption data 416.

[0032] Definition 6: Event Identifier Mapping. Let M be the mapping function from event identifiers to their semantic meanings. Each event e ∈ E is associated with an identifier id e and its semantic meaning M(id e ), which describes the specific characteristics of the event or the domain-specific interpretation. This mapping is defined as M: IDs → Descriptions, where IDs is the set of all possible event identifiers and Descriptions is the set of detailed explanations provided by domain experts.

[0033] For example, in the case of an abnormal event identifier using A1160, the mapping function M(A1160) can return the description "needle.force + deviation value", which indicates a significant deviation in the measurement of the needle's force.

[0034] Event identifiers can be used to define the form of an event sequence.

[0035] Definition 7: Positive and Negative Event Sequences. An event sequence S = <e1, e2, …, e k > is defined as above, where for each e1 ∈ E and i ∈ j, e i occurring before e j . Two categories of event sequences are introduced. A positive event sequence ends with one or more subsequent interruptions. For example, if S pos = <a1, a2, i1, i2>, where a1 and a2 are abnormal events and i1 and i2 are subsequent interruptions, then Spos This is a sequence of interruptions without any intervening anomalies.<i1,i2> Since it ends with a null, it is considered positive. In contrast, a negative event sequence does not end with a null sequence. For example, S neg =<a1,a2,a4> And if all elements are abnormal, S neg It is negative because it ends without any interruption.

[0036] The rationale for classifying event sequences based on the presence of interruptions is to identify which anomalies potentially contribute to operational failure. This distinction is crucial for informing domain experts who analyze sequences to identify critical points of failure and guide corrective actions. In addition, providing this classified data to LLMs allows them to generate more accurate summaries and predictions. Interventions signify a reset of operation in the corresponding line or location, often indicating a return to a baseline or "clean state." Thus, interruptions act as natural dividers in a continuous event stream, allowing events to be separated into coherent sequences that effectively reflect operational realities and facilitate predictive analysis.

[0037] After constructing the event sequence, we will describe the extraction of patterns. As done for constructing the PKG, the PrefixSpan algorithm can be used to mine the most common subsequence patterns from our sequence; however, alternative algorithms or libraries are equally applicable to identifying frequent event patterns. This process concludes with the formal introduction of the event patterns, as described below.

[0038] Definition 8: Positive and negative event patterns. An event pattern P is a subsequence of event sequences that meet a specific frequency criterion within a dataset, highlighting their significance or predictive power. There are two types of event patterns: positive and negative. Positive event patterns, for example,<a1,a2,i1> This is because a specific sequence of anomalies consistently precedes a series of positive events.<a1,a3> These negative event patterns characterize anomaly sequences that emerge from negative event sequences and do not result in interruptions.

[0039] The pattern mining process can be performed using the pattern knowledge graph 412, which can identify subsequences that do not end in interruptions from a positive event sequence. For example, two positive event sequences:<a1,a2,i1> and =<a1,a2,a3,i1> Therefore, miner does not end in interruption, =<a1,a2> These can be estimated as valid event patterns. While this contradicts the traditional definition of positive event patterns, recognizing these patterns can still provide valuable insights for domain experts and LLMs. It is important to provide flexibility in classifying these patterns. Users (domain experts) can determine whether to classify the pattern strictly as negative or consider it a true positive instance, depending on its importance in a particular domain or application.

[0040] In the final output, the mining algorithm generates a set of tuples, each tuple consisting of the frequency of an event pattern and the pattern itself, expressed as a list. The frequency indicates the number of sequences in which the pattern appears. These patterns are stored in a dedicated database, referred to as the pattern database, which can utilize a standard database system such as MySQL or MongoDB. For example, assuming MongoDB is used, its database configuration might include a set named "xyz_event_sequence" and another set named "xyz_event_pattern". The role of these sets is to store event sequences and patterns specific to a particular product or component referred to as "xyz" in the manufacturing plant. Within these sets, fields are maintained for event sequence and pattern data, as well as metadata such as a location ID (i.e., the location where the event sequence was observed) and timestamps marking the start and end of the event. Furthermore, it is indicated whether the event sequence and pattern are positive or negative. It is also possible to maintain two completely independent tables / sets for positive and negative patterns to improve data organization and search efficiency. These event sequences and patterns are indexed for efficient searching using one or two specific events as keys. Mechanisms for querying and searching related event patterns are detailed in the following subsection, "Searching for Related Event Patterns."

[0041] Prompts that a user can construct and submit to LLM for direct prediction or to generate code for analysis and prediction can be designed and formulated in different ways. In one embodiment, LLM first requires a comprehensive understanding of the operational context within the manufacturing plant. This includes an overview of station-specific scenarios derived from historical data collected from manufacturing equipment. Such contextual foundations are crucial to enabling accurate predictive output. For example, a prompt might read: "The following records are historically frequent patterns I have collected from several manufacturing machines that produce high-pressure injectors and high-pressure pumps."

[0042] Just as with instructing a human, it is crucial for an LLM to grasp the terminology and structural concepts associated with event identifiers and their sequences. For clarity, the user can explicitly define the format of these sequences as follows: "Each row in the data represents a pattern, where the first identifier e_current indicates the current event, and the second e_subsequent indicates a potential subsequent event. This list encapsulates past occurrences in which these identifiers were observed."

[0043] To facilitate understanding using LLM, concrete examples can be provided. For example, "The current event is identified as e_current=A1368, and the subsequent event is identified as e_subsequent=A1168. The history pattern is..."<A1234,A1364,A1368,A1163,I128> ,<A1368,A1163,A480,I51,I73,I128> This may include (e_current, e_subsequent) along with a series of other events such as >. Identifiers beginning with "A" indicate an anomaly or warning sign, while those beginning with "I" represent interruptions that should ideally be preempted.

[0044] In this structured description, e_current and e_subsequent are used to indicate specific event identifiers within the sequence, reflecting their role as either an ongoing or upcoming event, respectively. Events beginning with "A" and "I" are interpreted as anomalies and interruptions, respectively, according to Definitions 3 and 4. This format not only aids in a precise understanding of the event sequence but also in its analytical processing in predictive models.

[0045] Historical event patterns can be cataloged. This process involves compiling a detailed list of historical patterns of event sequences observed at specific locations within the manufacturing environment. For example, suppose the current event identifier is e_current=A1368, and potential future events are shown as e1=A1163, E2=I97, and e3=A1165, each followed by its respective event sequence. These sequences are in the format (e current ,e i It can be cataloged as [S1, S2, ...], where S j This represents a list of events that form a history pattern. The prompt is, for example, So far, we have the following 12 historical patterns: A1368 A1163<A1234,A1364,A1163,I128> ,<A1368,A1163,A480,I51,I73,I128> >A1368 A1165<<A1160,A1159,A481,A1368,I97> >A1368,A1165<<A1160,A1159,A1364,A1158,A1160,A1159,A1159,A1372,A2085,A481 A481,A4368,A1165,A2222M A1344,I90> >」 You can input it like this.

[0046] Here, each pattern is a subsequence of event sequences that satisfy a specific frequency criterion within the dataset, and is categorized as either a positive or negative event pattern based on its composition, as defined in Definitions 5 and 6. These patterns help the LLM recognize historical event sequences and predict potential future events based on this data. The data format can vary, depending on the complexity required, and may include simple delimited formats such as space-separated values, or more structured formats such as JSON and YAML.

[0047] A detailed description of the event types can be provided. This includes providing additional descriptions and classifications of event types so that the LLM may have an understanding or interpretation of the event sequences provided by the inventors. Specifically, additional information about the event types is provided using the event identifier mapping defined in Definition 6, as follows. Furthermore, additional explanations of the terminology used in a semantic sense are provided so that the LLM may better understand the concept or meaning of each event. Due to space limitations, only a few exemplary records are listed below: "Furthermore, the meanings of these event IDs are as follows: for example, A1160 means an instance where the parameter name is needle.force+outlier. This means monitoring the outlier of the needle's force. Some descriptions are in Turkish." A1160 Needle Force + Outliers A1159 Needle Force + Average Shift A1158 Needle. Force + Bunching A1160 Needle Force + Outliers ... I90 2500 (Presley) I97 2614(Robot Haberle me ar zas) I11 1299 (Fifostrecke Voll) I126 2950 (Kalite Problems) I51 220 (Storung Mechanik) ... In addition: Trends: Steep gradient, tendency towards acceptable range, gradual increase in NOK, ... Outliers: extreme values, drift-invariant outliers, non-Gaussian outliers, ...

[0048] Next, requests can be constructed. This component focuses on constructing queries, which can be categorized into two types: (i) status queries that prompt the LLM to summarize the event sequence and inform the expert of the current status, and (ii) predictive queries that use the current event sequence to predict future patterns. While simultaneous presentation can be implemented for these queries, presenting them separately tends to yield more reliable and complete recommendations.

[0049] A status request regarding the current event sequence requires the LLM to provide a concise summary of the specified event sequence and inform the user of the essential ongoing activity. The LLM is tasked with interpolating the descriptions associated with the events within these sequences and providing a synthetic analysis.

[0050] Definition 7: Status query. S = {s1, s2, ... s n Let} be the sequence of events, and each event s i It is associated with specific attributes and metadata. Status query function Q s Q s : A function defined as P(S) → Summary, where P(S) is the power set of S representing all possible subsequences of events in S.

[0051] Function Q sThis can be done by (1) analyzing the frequency and patterns of occurrence of each event identifier within S, (2) interpolating descriptions and synthesizing the analysis based on attributes and metadata associated with these events, and (3) generating summaries that highlight potential areas for important patterns, anomalies, and actions.

[0052] For example, sequence S =<A1160;A1159;A1364;A1158;A1160;A1159;A1372;A2085;A481;A481;A1368> Consider the following. The status query function Qs processes these events to identify and summarize dominant patterns and anomalies in the data. For example, identifiers "A1160" and "A1159" appear multiple times, indicating recurring problems related to "needle force," "outliers," and "mean shift" phenomena, respectively. The function Qs synthesizes this information and provides a summary such as: "The summary indicates a high frequency of force-related outliers and mean shift in the needle assembly process, coupled with consistent trends in press-fit window operation and compensation mechanisms, suggesting areas requiring immediate inspection and potential recalibration." Following a theoretical explanation of the status query function Qs, the following is an example of prompts used to engage LLM in summarizing event sequences from a manufacturing process. The practical example here incorporates additional constraints to limit the scope of LLM responses, addressing a common challenge where LLM might otherwise provide an overly comprehensive description not needed by the user: "We have collected a temporal event sequence <1160, A1159, A1364, A1158, A1160, A1159, A1372, A2085, A481, A481, A1268> from the manufacturing process, and we need a summary of the current situation. Based on these event sequences, please interpret them and provide your analysis. It is important to focus on synthesizing the overall semantics of these events rather than digging into event-by-event breakdowns or explaining each event individually. Your summary should be 100 words or less." This prompt is designed to elicit a concise, integrated summary that reflects a comprehensive understanding of the event sequence, while preventing the excessive fragmentation that often accompanies unrestricted LLM responses. First, the LLM provides a comprehensive overview of the situation based on the event sequence. It then informs the user of any general issues and suggests potential corrective actions.

[0053] It is possible to generate prediction requests based on the current event sequence. This request type instructs the LLM to analyze and summarize the sequence of events, enabling prediction of imminent outcomes. For example, consider a scenario where the LLM is prompted to predict the top five most likely outcomes based on a given event sequence. The user can specify the methodology to be used; for example, the user can instruct the LLM not to rely on a frequency-based counting approach that assumes the most frequent events are likely to repeat, but instead to focus on understanding the semantics and potential meaning of these sequences. This method leverages the LLM's ability to interpret complex event patterns and relationships, thus facilitating predictions beyond simple frequency analysis. Such approaches are particularly useful in situations where the accuracy of predictions depends more on semantics, including the nuances of events, than on the frequency of events. The process of leveraging semantic analysis for prediction purposes in an LLM can be formally defined as follows:

[0054] Definition 7: Predictive query. A predictive query is a function Q p :S p →P p It is formulated as, where S p P represents the space of the observed event sequence. p This represents the set of possible predictions. Query function Q p is element S p Interpreting the semantic relationships within, P p It is designed to propose predictions in this context.

[0055] Definition 8: Semantic analysis. In the context of event patterns, semantic analysis refers to the process of extracting meaningful insights from a sequence of events, going beyond statistical occurrence. This includes interpreting the significance of the event sequence and understanding the potential impact of each on future outcomes.

[0056] For example, event sequence S =<A1160,A1159,A1364,A1158,A1160,A1159,A1372,A2085,A481,A481,A1368> Let's consider the predictive query function Q. p This sequence can be analyzed to predict the most likely next event. Here, P p This represents the next possible event based on the semantic understanding of the history pattern.<A1161,A1158,A1369,A1373,A2086> This may include predictions such as those for function Q. p This approach refrains from using frequency-based approaches and instead focuses on implications and correlations revealed through semantic analysis of events in S.

[0057] Given these definitions, predictive queries can be structured to not only assume the next event but also understand potential alternative outcomes, thus providing a comprehensive foresight into possible future scenarios. The final part of the prompt, including an example query, might look like this: "Finally, the temporal event sequence from the manufacturing process<A1160,A1159,A1364,A1158,A1160,A1159,A1372,A2085,A481,A481,A1368> We have collected the following. Considering all these historical patterns, please select the top 5 most likely next events and tell us your opinion. Without using a frequency-based approach, please tell us your opinion on whether you can properly understand the meaning of these events. If none of the other events occur immediately after this sequence, what events are likely to occur next? Please tell us your opinion: It can be written as follows.

[0058] LLM can generate a list of the most likely events and identify those requiring further consideration. In the example described, five events, namely A1163, A480, A1364, I90, and A1165, are listed by LLM as the next most likely events. It is observed that the actual subsequent event is A1165, followed by interruption I90, demonstrating LLM's accurate predictive capabilities. It is also observed that the specificity of the output can vary depending on the LLM used and user-defined settings. Typically, LLM generates a list of potential subsequent events contextualized with relational insights, enabling domain experts to evaluate these predictions and make informed decisions regarding possible future events.

[0059] Another remaining challenge in this study is the effective extraction of relevant event sequences or patterns that can be used as part of an analysis prompt. To address this, we consider two different cases based on the database or setup used within the system. The first case involves the use of a vector database, where event patterns are stored in vector form, and common similarity metrics such as cosine similarity are applied to retrieve relevant event sequences. However, most vector databases or embedding techniques do not store event sequences reversibly. As a result, while partial retrieval of relevant events is implementable, reconstructing a complete event sequence from a vector embedding for prompt construction is virtually impossible. This limitation reduces the effectiveness of this approach. An alternative approach utilizes an explicit symbolic representation of event patterns stored in a conventional relational database or graph file format. This method ensures better storage and retrieval of event sequences, but still requires an efficient indexing method for collectively retrieving all relevant patterns. The operation of retrieving all relevant subsequent events from stored patterns is formalized, facilitating the analysis of nuances in event relationships based on historical data. This behavior can be customized to return only the most frequent subsequences, and depending on the analysis requirements, it can be filtered to include only positive event sequences, or both positive and negative event sequences.

[0060] Definition 9: Event pattern search. EP represents the set of all event patterns stored in the inventors' database. Given a source event e2 ∈ E, the event pattern search operation is performed for all pairs (e s ;e n Identify ) here e n This is the e in any event pattern sequence within EP. n This is the next event following. The resulting triplet is (e s ;e n;EPS) is defined as (e s ;e n This corresponds to sequences containing ). These triplets can be ordered by frequency by selecting the top k most frequent pairs and filtering them to include only positive event sequences or both positive and negative event sequences as specified.

[0061] For example, consider the following set of event patterns stored as EPs, which include the following sequences: EP={<A1234,A1364,A1368,A1163,I128> ;<A1368,A1163,A480,I51,I73,I128> ;<A479,A1567,A481,A1368,I97> ;<…,A1368,A1165,A2222,A1344,I90>} Assume A1368 is selected as the source event. Following this selection, the event pattern lookup operation identifies three subsequent events, each leading to a different sequence. The results are presented as a set of triplets showing the source event, the next event, and the corresponding sequence in which these pairs occur.

[0062] The system also introduces auxiliary behaviors designed to augment the system by facilitating the extraction of detailed metadata associated with transitions between events. This includes time intervals and their distribution within the dataset. By incorporating such metadata into the LLM prompts, the specificity of the provided event descriptions can be improved. Our observations suggest that including this detailed information improves the accuracy of the model's summarization and predictions.

[0063] Definition 10: Event sequence metadata lookup. This operation is performed on any event pair (e s ;e nThis accesses relevant metadata defined for the event and stored in the database. The retrieved metadata includes a statistical summary, which includes the minimum, maximum, and mean time intervals between cited events, along with their frequency of occurrence. This process aims to provide comprehensive temporal and quantitative insights into event pairs, thereby enhancing the analytical capabilities available to LLM. The following example illustrates the types of metadata that can be retrieved. Considering two anomalies A1368 and A1163, the operation yields a tuple as shown below. This tuple indicates that the minimum recorded time interval is 10 seconds, the maximum is 5324 seconds, the mean interval is 3015 seconds, and the events occurred twice together: 1:(A1368;A1163;10;5324;3015;2;:::).

[0064] This example highlights four specific metadata items, but additional data about these event pairs can be stored and retrieved as needed.

[0065] The disclosed system empowers domain experts by providing pre-designed templates for prompt construction. These templates ensure rapid development, maintain consistency and accuracy across queries, and adhere to the guidelines outlined in the “Prompt Structure” section. Depending on the type of query, not all parts of the prompt section may be required; for example, status queries do not always require a catalog of historical patterns.

[0066] Domain experts are presented with a chronological list of events from which they select the relevant entries to incorporate into the prompt. The event sequence specified in the query can be compiled semi-automatically by machine. To illustrate this process, let E be the set of all events in the timeline, and each event e∈E has an associated timestamp (e). current Define this as the current timestamp. δThe concept of a context window, as shown, corresponds to the duration over which related events are considered. The action of selecting an event within this context window is: W(E,t)={e∈E|t-tδ≦time(e)≦t} It can be defined as follows. Here, t δ = 7 days, and W selects events from the set E, and each event e has a timestamp in a context window ending at time t, assuming that time(e) and t are measured in days. First, a domain expert uses the W operation to specify the range of events to be analyzed, for example, to select a one-week period from the event timeline at a particular location, which is W(E,t). current This includes applying ), where abnormalities and interruptions are listed in chronological order. The user interface then transcribes the event IDs from the timeline and populates prompts accordingly. Events A1, A2, A3, A4, and A5 have occurred, and times (A3), (A4), and (A5) are W(E,t current Consider the example timeline in Figure 3B, which is within a one-week context window defined by ). In this scenario, only A3, A4, and A5 are returned by the operation, Q s Or Q p It generates A1 and A2, but they are outside the context window and cannot be selected.

[0067] The selection and popularization processes can be fully automated. For example, the interface can continuously acquire events streamed from the production line and ADE and display them on a timeline, automatically select events up to one week in advance using W, and forward prompts to the LLM to continuously update and display forecast results under the timeline. While the system automates the construction of most fields, it is important to note that domain experts should maintain the flexibility to make final adjustments before submission. Adjustments here may include authorizing the LLM to make decisions or requesting the generation of code templates to enhance analysis and forecasting. In this report, natural language text is presented as an exemplary output for readability, but in actual implementations, data formats such as JSON and YAML would be presented to support interoperability, allowing other systems or components to easily retrieve or display this data as needed.

[0068] Furthermore, all constructed prompts comply with the token limits of the LLM used, including the 4k and 16k token limits specific to the GPT-3.5 Turbo system. To navigate these limits, the system employs strategies such as prioritizing the inclusion of historical events based on their frequency, ensuring that the most relevant information is characterized within the given constraints.

[0069] While the system's primary application is within manufacturing environments utilizing temporal event data from manufacturing plants, it should be understood that the methodology and system described are highly versatile and adaptable to other domains. Potential applications include predicting future events in sports, patient monitoring, and analyzing driving logs from autonomous driving systems, all of which leverage time-series data from various sources.

[0070] Accordingly, this disclosure includes the development of an innovative method for summarizing and predicting operational events within a manufacturing environment using LLM. This method makes significant use of an enhanced dataset containing potential future events, meticulously extracted through techniques such as frequent event pattern mining from historical data. This approach enables a deeper understanding and prediction of operational scenarios, significantly improving decision-making capabilities in the manufacturing process.

[0071] Furthermore, the approach of this disclosure distinguishes itself by prioritizing semantic analysis over conventional statistical frequency methods for evaluating event sequences. This shift to deeper, meaning-based analysis facilitates more accurate predictions of future events by understanding the implications and interrelationships within event sequences, rather than simply their frequency of occurrence. This methodology not only improves the accuracy of predictions but also provides richer insights into the operational dynamics of the manufacturing system.

[0072] In addition, this disclosure systematically designs and integrates a robust framework that enables the effective acquisition, selection, and utilization of relevant event patterns. This system enhances the capabilities of the LLM by enabling the precise selection of event data based on an event timeline and its incorporation into prompts. The integration of this system supports the continuous improvement of operational intelligence in the manufacturing process and ensures that each decision is informed by the most relevant and recent data.

[0073] The neural network algorithms and / or methodologies of one or more embodiments described herein are implemented using a computing platform such as the computing platform 400 shown in Figure 4. The computing platform 400 may include memory 402, a processor 404, and non-volatile storage 406. The processor 404 may include one or more devices selected from high-performance computing (HPC) systems, including high-performance cores, microprocessors, microcontrollers, digital signal processors, microcomputers, central processing units, field-programmable gate arrays, programmable logic devices, state machines, logic circuits, analog circuits, digital circuits, or any other devices that manipulate signals (analog or digital) based on computer-executable instructions residing in memory 402. The memory 402 may include a single memory device or several memory devices, including, but not limited to, random access memory (RAM), volatile memory, non-volatile memory, static random access memory (SRAM), dynamic random access memory (DRAM), flash memory, cache memory, or any other device capable of storing information. Non-volatile storage 406 may include one or more persistent data storage devices, such as hard drives, optical drives, tape drives, non-volatile solid-state devices, cloud storage, or any other devices capable of permanently storing information.

[0074] The processor 404 can be configured to load into memory 402, reside in an embedded model 408 of non-volatile storage 406, and execute computer-executable instructions that embody one or more embedded algorithms and / or methodologies disclosed herein. The processor 404 can further load into memory 402, reside in a dynamics model 410 of non-volatile storage 406, and execute computer-executable instructions to embody the dynamics algorithms and / or methodologies described herein. The processor 404 can further load into memory 402, reside in a prediction model 412 of non-volatile storage 406, and execute computer-executable instructions that embody the prediction algorithms and / or methodologies described herein. Models 408-412 may include operating systems and applications. Models 408–412 are compilable or interpretable from computer programs written using a variety of programming languages ​​and / or technologies, including, but not limited to, Java, C, C++, C#, Objective-C, Fortran, Pascal, JavaScript, Python, Perl, and PL / SQL, either alone or in combination.

[0075] The computer-executable instructions of Models 408–412 may, when executed by the processor 404, cause the computing platform 400 to implement one or more of the neural network algorithms and / or methodologies disclosed herein. Non-volatile storage 406 may also include sensor data or measurement data 414 and interruption data 416 representing recorded metadata, for example, the frequency and duration of interruptions. Such data may also include, for example, part identification and identification of the station where the interruption occurred, the location of different components, angles between components, forces applied to components, and pressure applied to components. Measurement data 414 may include data captured or acquired from sensors located at specific stations in the manufacturing process. Measurement data 414 may also include data at specific locations (l1, l2, …l) within the manufacturing equipment. n ) can be linked to or coupled to. The sensor may be an image sensor, a laser measuring sensor, or any other type of sensor configured to generate data representing the physical quality, condition, or characteristics of the part being measured.

[0076] Computer-readable program instructions stored on a computer-readable medium can be used to instruct a computer, other types of programmable data processing devices, or other devices to function in a particular way, and as a result, the instructions stored on the computer-readable medium produce a product containing instructions that implement the functions, actions, and / or operations specified in the flowchart or diagram. In certain alternative embodiments, the functions, actions, and / or operations specified in the flowchart and diagram may be rearranged, processed sequentially, and / or simultaneously to match one or more embodiments. Furthermore, neither the flowchart nor the diagram may contain more or fewer nodes or blocks than those shown to match one or more embodiments.

[0077] As a simple example, the processor 404 can access instructions stored in memory 402 and execute one or more of LLM408, ADE410, and / or PKG412 using various data described herein, such as measurement data 414 and operation interruption data 416. Execution of these models can, for example, output predicted future potential anomalies in the manufacturing equipment.

[0078] Figure 5 shows a method for predicting operational events in a manufacturing environment using a large-scale language model (LLM) according to one embodiment. The method described herein can be executed via the computing platform 400 described above, for example, by a processor 404 that executes instructions stored in memory to perform the steps of the method.

[0079] In 502, the method first begins by receiving sensor data from a network of sensors distributed across various stations within the manufacturing facility. These sensors can be strategically placed to monitor critical mechanical parameters such as the position, angle, force, and pressure of components. For example, a sensor can measure the force applied by a robotic arm or the pressure within a hydraulic system. This data is important for maintaining operational integrity as it provides real-time insights into the performance and health of the machine. This can be stored as measurement data 414. By continuously collecting this data, the system can establish a baseline of normal operation in which anomalies can be detected.

[0080] In 504, once sensor data is collected, the Anomaly Detection Engine (ADE) 410 runs to analyze the data for any irregularities. The ADE is designed to identify various types of anomalies, including outliers, sudden changes in mean or variance, and stepped shifts. For example, if a sensor detects a sudden temperature spike that deviates from the norm, the ADE flags this as an anomaly. This may involve training a model to understand what constitutes "normal" operation, for example, sensor data within a minimum and / or threshold range established based on typical operation, and then triggering the model to flag data outside these boundaries and / or thresholds as anomalies. These anomalies are recorded as temporal events, allowing the system to alert domain experts to potential problems before they escalate into major incidents. This proactive approach helps minimize downtime and maintain smooth operation.

[0081] In 506, the method then includes recording metadata associated with the operational interruption. This metadata includes details such as the frequency and duration of the interruption, which are essential for understanding the impact of the anomaly on manufacturing. For example, if a machine frequently shuts down due to overheating, the metadata captures how often and for how long these interruptions occur. This information can be stored as operational interruption data 416. By distinguishing between anomalies and actual interruptions, the system provides a clearer picture of the operational challenges and enables more informed decision-making.

[0082] In 508, the system receives prompts entered by the user via a user interface regarding operational interruptions. The user interface provides access to sensor data, recorded anomalies, and historical event data, including metadata. This interface allows domain experts to create detailed prompts that guide the system in analyzing and predicting future events. For example, the user can enter prompts to investigate the causes of frequent machine downtime using historical data to identify patterns or trends.

[0083] In 510, the prompt is enhanced by one or more of the following: historical event data and metadata in 512, past event summaries in 514, and / or predictions about future potential anomalies. Past anomaly event summaries can be output, for example, from a pattern knowledge graph (PKG) 412. The PKG is constructed from metadata and recorded anomalies to provide past anomaly event summaries. For example, the PKG may reveal that a particular series of anomalies often precedes a machine failure. The enhanced input here enables the system to generate more accurate and contextually relevant predictions, thereby enhancing the decision-making process.

[0084] In step 518, a Large-Scale Language Model (LLM) 408 is executed based on enhanced prompts to generate initial output. The LLM processes detailed prompts, leverages its deep learning capabilities to interpret the data, and provides insights into potential future events. For example, the LLM can predict that certain anomaly patterns are likely to lead to machine failure, enabling proactive maintenance actions. This step is crucial for transforming raw data into usable intelligence.

[0085] Finally, in 520, the method includes predicting future potential anomalies or critical events in manufacturing equipment based on the execution of LLM with enhanced prompts. By leveraging the predictive capabilities of LLM, the system can predict downtime, enabling timely intervention and minimizing downtime. For example, if LLM predicts a high probability of machine failure, maintenance can be scheduled in advance to prevent manufacturing delays. This predictive analytics approach enhances the decision-making process and improves operational efficiency within the manufacturing environment.

[0086] While exemplary embodiments have been described above, these embodiments are not intended to describe all possible forms that are covered by the claims. The terms used herein are descriptive, not restrictive, and it should be understood that various modifications can be made without departing from the spirit and scope of this disclosure. As described above, features of various embodiments can be combined to form further embodiments of the invention that may not be expressly described or illustrated. Various embodiments may be described as offering advantages with respect to one or more desired characteristics, or as being preferable to other embodiments or prior art implementations, but those skilled in the art will recognize that one or more features or characteristics may be compromised to achieve desired overall system attributes that depend on a particular application and implementation. These attributes may include, but are not limited to, cost, strength, durability, life cycle cost, marketability, appearance, packaging, size, maintainability, weight, manufacturability, ease of assembly, etc. Accordingly, as long as any embodiment is described as being less desirable than other embodiments or prior art implementations with respect to one or more characteristics, these embodiments may be desirable for a particular application, and are not outside the scope of this disclosure.

Claims

1. A computer-implemented method for predicting behavioral events in a manufacturing environment using a large-scale language model (LLM), The aforementioned method, Receiving sensor data from multiple sensors at multiple stations of a manufacturing facility, wherein the sensor data indicates machine parameters associated with the machines located at the stations. The anomaly detection engine (ADE) is executed on the sensor data to monitor and record any anomalies in the machine parameters. Record metadata associated with the interruption of operation in the aforementioned manufacturing equipment, Receiving user prompt input regarding the interruption of operation via a user interface, wherein the user interface includes access to historical event data, the historical event data being associated with or including the sensor data, the recorded anomaly and the metadata, The prompt is enhanced based on historical event data and metadata associated with the interruption of the operation. Based on the enhanced prompt, the LLM is executed to generate the initial output, Based on the execution of the LLM using the enhanced prompts, predict future potential anomalies or critical events in the manufacturing equipment. Methods that include...

2. The method according to claim 1, further enhancing the prompt based on predictions of future potential anomalies without specific prior training on the historical event data.

3. The method according to claim 1, wherein the prompt is further enhanced based on output from a pattern knowledge graph (PKG) constructed from metadata associated with the operational interruption and the recorded anomaly, the output from the PKG including a summary of past interruption events.

4. The above method further, The initial output from the LLM is presented to the user via the user interface, Based on the user feedback, the prompt will be iteratively refined to improve the accuracy and relevance of the initial output, Based on the refined prompt, the LLM is executed to generate (i) an updated event summary of past events, and (ii) an updated forecast of future potential anomalies. The method according to claim 1, including the method described in claim 1.

5. The method according to claim 1, wherein the recorded metadata includes the frequency and duration of the operational interruption.

6. The method according to claim 1, wherein the ADE is configured to detect the anomaly based on the use of predetermined minimum and maximum thresholds for detecting outliers, abrupt changes in mean or variance or stepped mean shifts in the sensor data.

7. The method according to claim 1, wherein recording the metadata includes the frequency and duration of interruptions associated with a location within the manufacturing environment.

8. The above method further, (i) Provide the initial output to a verification system that verifies whether the initial output is correct, or (ii) provide the initial output to a verification system that verifies that one or more of the operation interruptions exist. The method according to claim 1, including the method described in claim 1.

9. A system that uses a Large-Scale Language Model (LLM) to predict operational events in a manufacturing environment, The aforementioned system, Processor and The memory that stores the instructions, Equipped with, When the aforementioned instruction is executed by the processor, the processor will be instructed to: A step of receiving sensor data from multiple sensors at multiple stations of a manufacturing facility, wherein the sensor data indicates machine parameters associated with a machine located at the station, The steps include: running an abnormality detection engine (ADE) on the sensor data to monitor and record abnormalities in the machine parameters; The steps include recording metadata associated with the interruption of operation in the manufacturing equipment, A step of receiving a user prompt regarding the interruption of operation via a user interface, wherein the user interface includes access to historical event data, the historical event data being associated with or including the sensor data, the recorded anomaly and the metadata, The steps include enhancing the prompt based on historical event data and metadata associated with the interruption of the operation, The steps include: executing the LLM based on the enhanced prompt to generate initial output; A step of predicting a future potential anomaly or critical event in the manufacturing equipment based on the execution of the LLM using the enhanced prompts, A system designed to perform a certain action.

10. The system according to claim 9, further enhancing the prompts based on predictions of future potential anomalies without specific prior training on the historical event data.

11. The system according to claim 9, wherein the prompting is further enhanced based on output from a pattern knowledge graph (PKG) constructed from metadata associated with the operational interruption and the recorded anomaly, the output from the PKG including a summary of past interruption events.

12. When the aforementioned instruction is executed by the processor, the processor further: The initial output from the LLM is presented to the user via the user interface, Based on the user feedback, the prompt will be iteratively refined to improve the accuracy and relevance of the initial output, Based on the refined prompt, the LLM is executed to generate (i) an updated event summary of past events, and (ii) an updated forecast of future potential anomalies. The system according to claim 9, which is for performing the following.

13. The system according to claim 9, wherein the recorded metadata includes the frequency and duration of the operational interruption.

14. The system according to claim 9, wherein the ADE is configured to detect the anomaly based on the use of predetermined minimum and maximum thresholds for detecting outliers, abrupt changes in mean or variance or stepped mean shifts in the sensor data.

15. The system according to claim 9, wherein recording the metadata includes the frequency and duration of interruptions associated with a location within the manufacturing environment.

16. When the aforementioned instruction is executed by the processor, the processor further: (i) Provide the initial output to a verification system that verifies whether the initial output is correct, or (ii) provide the initial output to a verification system that verifies that one or more of the operation interruptions exist. The system according to claim 9, which is for performing the following.

17. A non-temporary computer-readable storage medium containing instructions, When the aforementioned instruction is executed by the processor, the processor will be instructed to: Receiving sensor data from multiple sensors at multiple stations of a manufacturing facility, wherein the sensor data indicates machine parameters associated with the machines located at the stations. The anomaly detection engine (ADE) is executed on the sensor data to monitor and record any anomalies in the machine parameters. Record metadata associated with the interruption of operation in the aforementioned manufacturing equipment, Receiving user prompt input regarding the interruption of operation via a user interface, wherein the user interface includes access to historical event data, the historical event data being associated with or including the sensor data, the recorded anomaly, and the metadata, Based on the historical event data and the metadata associated with the interruption, enhance the prompt, Based on the enhanced prompt, the LLM is executed to generate the initial output, Based on the execution of the LLM using the enhanced prompts, predict future potential anomalies or critical events in the manufacturing equipment. A non-temporary, computer-readable storage medium intended for executing [a certain action].

18. The non-temporary computer-readable storage medium according to claim 17 further enhances the prompt based on predictions of future potential anomalies without specific prior training on the historical event data.

19. The non-temporary computer-readable storage medium according to claim 17 further enhances the prompt based on output from a pattern knowledge graph (PKG) constructed from the metadata associated with the operational interruption and the recorded anomaly, wherein the output from the PKG includes a summary of past interruption events.

20. When the aforementioned instruction is executed by the processor, the processor will be instructed to: The initial output from the LLM is presented to the user via the user interface, Based on the user feedback, the prompt will be iteratively refined to improve the accuracy and relevance of the initial output, Based on the refined prompt, the LLM is executed to generate (i) an updated event summary of past events, and (ii) an updated forecast of future potential anomalies. A non-temporary computer-readable storage medium according to claim 17, which is for performing the following:

Images