Industrial time series semantic analysis and operation and maintenance decision-making method and system based on large model

Abstract

The present invention discloses a method and system for industrial time series semantic parsing and operation and maintenance decision-making based on a large model. The method includes: collecting time series data of the normal operating status of industrial objects to construct an abnormal time series reconstruction model, reconstructing the time series data of the faulty part and screening the target variable based on the reconstruction error and the abnormal indication function; designing a time series parsing semantic library for industrial objects, and further constructing a time series semantic parsing middle platform for semantic parsing and fault knowledge recall; constructing time series semantic parsing prompt words using the time series parsing semantic library, target variables, etc. to obtain the time series feature semantic text of the target variable; recalling the corresponding fault knowledge text in the fault knowledge library through the time series semantic parsing middle platform, constructing a time series fault diagnosis prompt word input large model, obtaining the diagnosis result through multiple voting and review, and automatically generating an operation and maintenance diagnosis form. The present invention can reduce labor costs and output data-supported and interpretable diagnostic results for industrial operation and maintenance tasks.

Description

Technical Field

[0001] The present invention relates to the field of industrial intelligent operation and maintenance technology, and in particular to a large-scale model-based industrial time series semantic parsing and operation and maintenance decision-making method and system. Background Art

[0002] Industrial equipment generates massive amounts of time-series data during operation. This data contains multi-dimensional information such as equipment status, process parameters, and environmental factors. It is the foundation for critical tasks such as predictive maintenance, fault diagnosis, and energy efficiency optimization. With the widespread adoption of Industrial Internet of Things (IIoT) technologies, the volume of high-frequency, multi-source, and heterogeneous time-series data collected in real time is growing exponentially, posing significant challenges to traditional analysis methods. Existing industrial operations and maintenance methods suffer from a lack of interpretable diagnostic results, difficulty in locating the root cause of faults, and reliance on manual expert drafting of diagnostic reports, severely hindering industrial production efficiency.

[0003] With the development of artificial intelligence (AI) technology, large language models (LLMs), with their powerful natural language processing and knowledge reasoning capabilities, are increasingly being applied to industrial fault diagnosis. While LLMs demonstrate strong contextual understanding and text analysis capabilities, existing LLMs still have significant shortcomings when handling industrial fault diagnosis tasks. Industrial equipment operating data exhibits distinct time series characteristics, and equipment failures often evolve over time. Time series information, such as changes in operating parameters before a fault occurs, the time of fault occurrence, and subsequent state evolution, is crucial for fault diagnosis and intelligent operation and maintenance. While LLMs possess powerful text parsing and logical reasoning capabilities, they still face bottlenecks in effectively parsing and utilizing time series data. They struggle to capture the time series characteristics and dynamic patterns of equipment operating data, resulting in diagnostic results that are neither accurate nor timely enough to meet the demands of industrial production in complex fault diagnosis scenarios. Furthermore, industrial fault diagnosis tasks place high demands on model interpretability and reliability. Existing LLMs suffer from opaque decision-making processes, making it difficult to provide clear diagnostic evidence and explanations to industrial engineers, hindering their widespread application and widespread adoption in industrial production.

[0004] In view of the problem that large models in the industrial field have insufficient time series parsing capabilities and are difficult to adapt to fault diagnosis tasks, there is an urgent need for an industrial time series semantic parsing and operation and maintenance decision-making method and system based on large models for downstream tasks such as fault diagnosis. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for industrial time series semantic parsing and operation and maintenance decision-making based on a large model, addressing the problem that large language models in the existing technology have difficulty analyzing the time series characteristics and change patterns of time series data in industrial objects and cannot be applied on a large scale in complex industrial scenarios. The present invention mines key variable information with obvious fault characteristics in industrial objects, uses a large language model to describe the changing characteristics of time series through text, and activates the powerful reasoning ability of the large model based on text descriptions and existing fault knowledge for fault diagnosis in industrial operation and maintenance scenarios.

[0006] The object of the present invention is achieved through the following technical solutions: In a first aspect, an embodiment of the present invention provides an industrial time series semantic parsing and operation and maintenance decision-making method based on a large model, comprising the following steps:

[0007] Step 1: Build an abnormal time series reconstruction model based on the collected time series data of the industrial object under normal operation. Obtain reconstructed data based on the abnormal time series reconstruction model to determine whether the online collected time series data is abnormal. If so, skip to step 2; otherwise, end directly.

[0008] Step 2: Select time series data with anomalies to construct a test set, calculate the reconstruction error of the test sample, and screen the variables based on the reconstruction error and the anomaly indicator function to obtain the target variable;

[0009] Step 3: Obtain a time series data statistical table based on the time series data of the target variable, and use the time series parsing semantic library, industrial process description text, all measurement point information, target variable information, and the time series data statistical table to construct a time series semantic parsing prompt word. This is used as the input of the large model, and the large model outputs the time series feature semantic text of all target variables one by one;

[0010] Step 4: Build a time series semantic parsing platform. Based on the time series feature semantic text of the target variable, use the time series semantic parsing platform to recall the corresponding fault knowledge text from the industrial object fault knowledge base.

[0011] Step 5: Based on the industrial process description text, all measurement point information, recalled fault knowledge text, and the time series feature semantic text of the target variable, a temporal fault diagnosis prompt word is constructed as the input of the large model. The output results of the large model are voted and reviewed multiple times to obtain the final fault category diagnosis result and reasoning process;

[0012] Step 6: The large model integrates the fault category diagnosis results, the reasoning process, and the recalled fault knowledge text to automatically generate an operation and maintenance diagnosis sheet, which is then reviewed by experts and stored in the industrial object fault knowledge base.

[0013] Furthermore, the abnormal time series reconstruction model is constructed based on the time series data collected under normal operating conditions of the industrial objects, and reconstructed data is obtained based on the abnormal time series reconstruction model to determine whether the time series data collected online has an abnormality, specifically including:

[0014] Collect time series data of industrial objects under normal operating conditions to construct a measurement point data matrix; perform K-means clustering based on the measurement point data matrix to select n representative samples from it to construct a state matrix; calculate the weight vector between the time series data samples collected online and the state matrix, and construct an abnormal time series reconstruction model based on the weight vector and the state matrix; obtain the reconstructed data corresponding to the time series data currently collected online based on the abnormal time series reconstruction model, calculate the reconstruction error between the time series data currently collected online and the reconstructed data, and compare the reconstruction error with a preset error threshold to determine whether the time series data collected online has an abnormality.

[0015] Furthermore, the step 2 specifically includes the following sub-steps:

[0016] Step 2.1: Select time series data with anomalies to construct a test set. Use the abnormal time series reconstruction model to obtain the reconstructed data of each test sample. Calculate the corresponding reconstruction error based on each test sample and its reconstructed data. Then, divide each test sample and its corresponding reconstruction error into a baseline part and a fault part based on the fault start time and fault end time.

[0017] Step 2.2: Calculate the average reconstruction error based on the reconstruction error of the baseline part to set the abnormal threshold for each variable; calculate the absolute value of the reconstruction error of each variable at all time points in the faulty part based on the reconstruction error of the faulty part;

[0018] Step 2.3: Construct an anomaly indicator function based on the absolute value of the reconstruction error of each variable in the faulty part and its corresponding anomaly threshold, and use the anomaly indicator function to calculate the anomaly occurrence time and anomaly score of each variable;

[0019] Step 2.4: Select the top one with the largest anomaly score variables and the earliest anomaly appearance time variables to construct a set of candidate variables;

[0020] Step 2.5: Screen the candidate variables in the candidate variable set based on the variance difference between the test sample data of the baseline part and the test sample data of the fault part to obtain the target variable that can characterize the fault mode.

[0021] Furthermore, obtaining a time series data statistics table based on the time series data of the target variable specifically includes:

[0022] Based on the measured value of the target variable at the faulty part, the reconstruction value of the measured value is calculated using the abnormal time series reconstruction model, the reconstruction error is calculated based on the measured value and its reconstruction value, and the reconstruction error percentage is calculated based on the measured value and its reconstruction error;

[0023] The reconstruction error is calculated based on the measured value of the target variable in the baseline part, and the average reconstruction error and the average reconstruction error percentage of the target variable in the baseline part time window are calculated according to the reconstruction error;

[0024] For each target variable's time series data in the faulty part, the measured value, reconstruction value, reconstruction error, and reconstruction error percentage are organized into a table in the order of timestamps. The average reconstruction error and average reconstruction error percentage of the target variable in the baseline part are added to the table to finally obtain a statistical table of the target variable's time series data.

[0025] Furthermore, the time series parsing semantic library includes multiple predefined semantic units, including overall trends, inflection point characteristics, range and average characteristics, extreme value characteristics, volatility characteristics, stage characteristics, periodic or repetitive patterns, local anomalies-mutation points, local anomalies-sharp decline phenomena, initial stage characteristic summary, mid-term stage characteristic summary and late stage characteristic summary.

[0026] Furthermore, the time series semantic parsing middle platform includes a time series parsing semantic library, a time series parsing tool library and a fault knowledge recall module, wherein the time series parsing tool library includes a large model, an abnormal time series reconstruction model, principal component analysis and mean calculation; the fault knowledge recall module is used to recall the corresponding fault knowledge text from the industrial object fault knowledge base based on the existing time series feature semantic text of the target variable, wherein the industrial object fault knowledge base includes sensor measurement point data, fault diagnosis records, expert experience, fault operation and maintenance suggestions, and fault root cause analysis fault knowledge texts collected by the industrial object during its operation and maintenance process.

[0027] Furthermore, the fault knowledge recall module specifically includes:

[0028] All fault knowledge texts in the industrial object fault knowledge base and the time series feature semantic texts of the target variable are encoded into corresponding embedding vectors through the text encoder in the fault knowledge recall module;

[0029] Calculate the similarity between the embedding vector of the time series feature semantic text of each target variable and the embedding vector of all fault knowledge texts;

[0030] The fault category corresponding to the target variable is screened by judging whether the calculated similarity is greater than a preset similarity threshold, and the corresponding fault knowledge text is recalled from the industrial object fault knowledge base based on the fault category.

[0031] Furthermore, the step 6 specifically includes:

[0032] Draw a time series visualization image of the original data of the target variable containing abnormal representation and the reconstructed data of the fault part, and pass the time series visualization image, the fault category diagnosis result and reasoning process output by the large model, the recalled fault knowledge text and the time series feature semantic text of the target variable into the large model for knowledge extraction; limit the output content of the large model through prompt word instructions; use the knowledge information extracted by the large model to supplement the blank content of the operation and maintenance diagnosis form template, and realize the automatic generation of the operation and maintenance diagnosis form; and submit it to experts for review, and store the verified operation and maintenance diagnosis form in the industrial object fault knowledge base.

[0033] Furthermore, the operation and maintenance diagnosis form includes a diagnosis summary, early warning point details, measurement point trend images, possible cause analysis and operation and maintenance suggestions; wherein, the diagnosis summary includes the equipment and component path of the fault measurement point and the time of fault occurrence; the early warning point details include the source label point name, source label point description, alarm type, first occurrence time and most recent occurrence time; the measurement point trend chart includes a time series visualization image showing the original data and reconstructed data, as well as a text description and analysis chart of the time series changes.

[0034] The second aspect of an embodiment of the present invention provides an industrial time series semantic parsing and operation and maintenance decision-making system based on a large model, including one or more processors and a memory, and the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned industrial time series semantic parsing and operation and maintenance decision-making method based on a large model.

[0035] The beneficial effects of the present invention are as follows: the present invention can solve the problem that large language models cannot effectively parse and utilize time series data, so that large language models can capture the time series change laws of equipment operation data; the present invention screens time series anomaly variables by reconstructing error and anomaly indicator functions, selects target variables that can characterize abnormal patterns, and injects industrial process knowledge and time series change information into the large language model by constructing time series semantic parsing prompt words, and converts the sequence features of the time series into the form of text description; and screens possible fault types through the fault knowledge recall module, uses time series fault diagnosis prompt words to activate the reasoning ability of the large language model, and performs downstream fault diagnosis tasks; the present invention can output interpretable diagnostic results for time series data and provide support for industrial intelligent operation and maintenance; compared with the existing technology, the present invention realizes the full-link closed loop from time series data parsing to interpretable diagnostic decision-making for the first time, breaking through the semantic barriers caused by the traditional method's reliance on abstract features, reducing labor costs through automatic diagnosis order generation, and being able to output data-supported interpretable diagnostic results for industrial operation and maintenance tasks, thereby providing new ideas for the intelligent operation and maintenance of industrial equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the industrial time series semantic parsing and operation and maintenance decision-making method based on a large model of the present invention;

[0037] Figure 2 A flowchart for screening target variables capable of characterizing failure modes according to the present invention;

[0038] Figure 3 This is a flow chart of the architecture of the present invention for generating time series feature semantic text of target variables using a large model;

[0039] Figure 4 This is a flow chart of the architecture of the temporal semantic parsing platform of the present invention;

[0040] Figure 5 This is a flow chart of the timing fault diagnosis prompt word generation and large model intelligent diagnosis of the present invention;

[0041] Figure 6 A flow chart automatically generated by the operation and maintenance diagnosis sheet of the present invention;

[0042] Figure 7 A structural diagram of the industrial time series semantic parsing and operation and maintenance decision-making system based on a large model of the present invention;

[0043] Figure 8 This is an example diagram of obtaining the time series feature semantic text of the target variable in an embodiment of a three-phase flow facility in a chemical process according to the present invention;

[0044] Figure 9 This is an example diagram of the fault category diagnosis results and reasoning process obtained in an embodiment of a three-phase flow facility in a chemical process according to the present invention;

[0045] Figure 10 This is a reconstruction effect diagram of the abnormal time series reconstruction model in an embodiment of a real thermal power plant according to the present invention;

[0046] Figure 11 This is an example diagram of the fault category diagnosis results and reasoning process obtained in an actual thermal power plant embodiment of the present invention;

[0047] Figure 12 This is an example diagram of an operation and maintenance diagnosis sheet automatically generated by the present invention in an actual thermal power plant embodiment. DETAILED DESCRIPTION

[0048] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0049] The terms used in this invention are for the purpose of describing specific embodiments only and are not intended to limit the invention. The singular forms "a," "the," and "the" used in this invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0050] It should be understood that although the terms "first," "second," "third," etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are merely used to distinguish information of the same type from one another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present invention. Depending on the context, the term "if" as used herein may be interpreted as "when," "when," or "in response to determining."

[0051] The present invention will be described in detail below with reference to the accompanying drawings. Unless there is any conflict, the features of the following embodiments and implementations may be combined with each other.

[0052] See also Figure 1 The industrial time series semantic parsing and operation and maintenance decision-making method based on a large model of the present invention specifically includes the following steps:

[0053] Step 1: Build an abnormal time series reconstruction model based on the collected time series data under normal operating conditions of the industrial object. Obtain reconstructed data based on the abnormal time series reconstruction model to determine whether the online collected time series data is abnormal. If so, jump to step 2; otherwise, end directly.

[0054] It should be understood that industrial objects refer to industrial processes, including but not limited to equipment or systems, such as three-phase flow devices in chemical processes.

[0055] Furthermore, an abnormal time series reconstruction model is constructed based on the collected time series data under normal operating conditions of the industrial objects. Reconstructed data is obtained based on the abnormal time series reconstruction model to determine whether the online collected time series data has abnormalities. The specific steps include the following:

[0056] Step 1.1: Collect time series data of industrial objects under normal operating conditions to construct a measurement point data matrix.

[0057] Specifically, in this embodiment, the m measurement points of the industrial object, such as temperature, pressure, etc., can be measured by corresponding sensors or instruments. The sample collected at time i is recorded as ,in, Indicates the first measurement point data collected at time i, Indicates the second measurement point data collected at time i, Indicates the mth measurement point data collected at time i. A measurement point is a variable, and the superscript T indicates the transpose of a matrix or vector. Collect time series data under normal operating conditions to construct the measurement point data matrix of the industrial object. , where L represents the length of the time series.

[0058] Step 1.2: Perform K-means clustering based on the measurement point data matrix to select n representative samples to construct a state matrix. The representative samples are used to characterize typical temporal patterns in industrial objects.

[0059] Specifically, based on the measurement point data matrix, the K-means clustering method is used to cluster the L samples of the time series data into n clusters. The distance between each cluster center and all samples is calculated. This distance can be measured by Euclidean distance or cosine similarity. The sample closest to the cluster center is selected as the representative sample corresponding to the cluster. The state matrix is ​​constructed based on the selected n representative samples. The state matrix is ​​expressed as:

[0060]

[0061] Where n is the total number of representative samples, which also represents the sampling time; the state matrix D represents n typical time patterns under the normal operation state of the equipment.

[0062] It should be understood that K-means clustering is a commonly used clustering algorithm, the core of which is to divide data into K different clusters in an iterative manner and minimize the sum of the distances between each data point and the center of its cluster.

[0063] Step 1.3: Calculate the weight vector between the online collected time series data samples and the state matrix, and construct an abnormal time series reconstruction model based on the weight vector and the state matrix.

[0064] Specifically, for the time series data samples collected online, they are used as input samples of the abnormal time series reconstruction model. , calculate the weight vector of n time modes in the state matrix D , the input sample can be represented by a linear combination of the existing normal operating state modes in the state matrix D, where the weight vector is calculated as:

[0065]

[0066] Then the input sample is reconstructed according to the weight vector and state matrix, and the reconstructed output is , using this as the constructed abnormal timing reconstruction model, the abnormal timing reconstruction model can be expressed by the following formula:

[0067]

[0068] The time series data samples collected online are used as input samples of the abnormal time series reconstruction model , the corresponding reconstructed data can be obtained.

[0069] Step 1.4: Obtain the reconstructed data corresponding to the time series data currently collected online based on the abnormal time series reconstruction model, and calculate the reconstruction error between the time series data currently collected online and the reconstructed data. By comparing the reconstruction error with the preset error threshold, it is determined whether the time series data collected online has any abnormalities: if the reconstruction error is less than or equal to the preset error threshold, it is considered that the reconstruction effect is good, indicating that the time series data collected online is in a normal operating state; conversely, if the reconstruction error is greater than the preset error threshold, it is considered that the reconstruction effect is poor, indicating that the time series data collected online is in an abnormal operating state.

[0070] Step 2: Select time series data with anomalies to construct a test set, calculate the reconstruction error of the test sample; filter the variables based on the reconstruction error and the anomaly indicator function to obtain the target variable. Figure 2 As shown, it specifically includes the following sub-steps:

[0071] Step 2.1. Select time series data with anomalies to construct a test set. Use the abnormal time series reconstruction model to obtain the reconstructed data of each test sample. Calculate the corresponding reconstruction error based on each test sample and its reconstructed data. Then, divide each test sample and its corresponding reconstruction error into a baseline part and a fault part based on the fault start time and fault end time.

[0072] It should be noted that each test sample in the test set is divided into a baseline part and a fault part. The baseline part represents the normal operation of the industrial object in the sample data, and the fault part represents the part that may have anomalies. Therefore, based on the fault start time and fault end time, each test sample and its corresponding reconstruction error are divided into a baseline part and a fault part, which are expressed as:

[0073] ;

[0074] ;

[0075] Where, Indicates the fault start time, Indicates the fault end time. represents the test sample of the baseline part, The test sample representing the faulty part, represents the reconstruction error of the baseline part, Represents the reconstruction error of the fault part. Among them, the time series length of the baseline part is equal to the time series length of the fault part, so Start forward and select the time series length The time series data of each sampling moment is used as the baseline. In addition, the fault start time and fault end time are given by the monitoring system of the industrial object, and a certain range of deviation is allowed.

[0076] Step 2.2: Calculate the average reconstruction error based on the reconstruction error of the baseline part. The calculation formula is:

[0077]

[0078] Where b represents the average reconstruction error, ; represents the reconstruction error at time i, The absolute value of the reconstruction error of each variable at all time points in the faulty part is calculated based on the reconstruction error of the faulty part. The calculation formula is:

[0079]

[0080] Where, Represents the absolute value of the reconstruction error of variable j at all time points of the fault part, here we use Represents all time points, , ; Represents the absolute value function. Calculate the abnormal threshold of each variable j based on the average reconstruction error b ,in represents the pre-set abnormal weight coefficient, represents the average reconstruction error of variable j.

[0081] It should be noted that the fault part is a time zone divided by roughly set fault start time and fault end time, and does not mean that all points in the fault part have faults.

[0082] Step 2.3: Construct an anomaly indication function based on the absolute value of the reconstruction error of each variable in the faulty part and its corresponding anomaly threshold, and use the anomaly indication function to calculate the anomaly occurrence time and anomaly score of each variable.

[0083] Specifically, first, an abnormality indication function is constructed based on the absolute value of the reconstruction error of each variable in the fault part and its corresponding abnormal threshold. The calculation formula is:

[0084]

[0085] Where, represents the abnormal indicator function value of variable j at time t, represents the absolute value of the reconstruction error of variable j at the faulty part at time t, Represents the abnormal threshold of variable j. The abnormal indicator function can be used to determine whether the faulty data is the actual fault point. Based on the above abnormal indicator function, it can be seen that at time point t, if the absolute value of the reconstruction error of variable j in the faulty part is greater than or equal to the corresponding set abnormal threshold, the abnormal indicator function value is set to 1, indicating that a fault exists at that time point.

[0086] Secondly, if the value of the abnormal indicator function in the time window of length W starting from time point t is 1, it means that there is a fault in the time window. Based on this principle, the earliest fault occurrence time of each variable j can be obtained. The earliest fault occurrence time of variable j is the abnormal occurrence time of variable j, and its calculation formula is:

[0087]

[0088] Where, represents the abnormal occurrence time of variable j, that is, the earliest occurrence time of the fault of variable j; k represents the time point; W represents the length of the abnormal time window.

[0089] Finally, the anomaly score of variable j within the anomaly time window is calculated, and the calculation formula is:

[0090]

[0091] Where, represents the anomaly score of variable j in the abnormal time window, represents the mean function, represents the average reconstruction error of variable j, The set of absolute values ​​of the reconstruction errors of the fault points. If the absolute value of the reconstruction error of variable j at time t in the faulty part is greater than or equal to its anomaly threshold, it is included in the set and its anomaly score is calculated. If variable j does not have an actual fault point in the faulty part, its anomaly score is 0.

[0092] Step 2.4: Select the top one with the largest anomaly score variables and the earliest anomaly appearance time variables to construct a set of candidate variables.

[0093] Specifically, based on the abnormal occurrence time and abnormal score of each variable calculated in step 2.3, first select from all m variables The variables with the largest anomaly scores form a set , which is calculated as:

[0094]

[0095] In addition, you can also sort the anomaly scores of the m variables in descending or ascending order, and then select the top one with the largest anomaly score. variables constitute a set .

[0096] Then, from all m variables, select The variables with the earliest abnormal occurrence time constitute a set , and its calculation formula is:

[0097]

[0098] In addition, you can also sort the abnormal occurrence time of m variables in descending or ascending order, and then select the one with the earliest (smallest) abnormal occurrence time. variables constitute a set .

[0099] Finally, according to the set and collection Constructing a set of candidate variables , select a set of candidate variables The variables in the candidate variable set reflect the temporal characteristics of the fault to the greatest extent, which is beneficial to the downstream fault diagnosis task.

[0100] Step 2.5: Screen the candidate variables in the candidate variable set based on the variance difference between the test sample data of the baseline part and the test sample data of the fault part to obtain the target variable that can characterize the fault mode.

[0101] Specifically, the candidate variables in the candidate variable set are further screened based on the variance of the test sample data of the baseline part and the test sample data of the fault part. The specific screening method can be expressed by the following formula:

[0102]

[0103] Where, is the variable set obtained by the final screening, the set The variables in are the target variables that can characterize the failure mode. represents the variance calculation function, Represents the baseline data of variable j in the test sample data, represents the faulty data for variable j in the test sample data. If the variance of the candidate variable in the faulty data is greater than twice the variance of the candidate variable in the baseline data, the candidate variable is used as the target variable that can characterize the fault mode and is then passed to the subsequent processing flow for subsequent fault diagnosis.

[0104] Step 3: Obtain a time series data statistical table based on the time series data of the target variable, and use the time series parsing semantic library, industrial process description text, all measurement point information, target variable information and time series data statistical table to construct a time series semantic parsing prompt word, and use it as the input of the big model. Through the cross-modal reasoning ability of the big model, the time series data is described in text form as a time series feature, and the time series features and fault features of the time series data are converted into time series feature semantic text, connecting the abstract time series data and semantic understanding capabilities. Finally, the big model outputs the time series feature semantic text of all target variables one by one, such as Figure 3 shown.

[0105] It should be understood that prompt engineering is a well-known technique that aims to activate specific reasoning paths within a large model through specific text instructions. Prompt engineering is key information that allows users to convey task intent to the large model, guiding it to generate output that meets user needs. Prompt design is a prerequisite for using large models.

[0106] In this embodiment, the target variable set obtained in step 2 is The failure mode of the data in the target variable is more obvious than that of other variables, enabling fault cause analysis and reasoning based on the relevant data of the target variable's time series. Based on this, a statistical table of time series data is obtained based on the target variable's time series data. Specifically, the following steps are performed: First, based on the target variable's measured value in the faulty portion, the reconstruction value corresponding to the measured value is calculated using the abnormal time series reconstruction model. The reconstruction error is calculated based on the measured value and its corresponding reconstruction value. The reconstruction error percentage over time is calculated based on the measured value and its corresponding reconstruction error. This percentage is calculated by dividing the reconstruction error by the measured value and multiplying it by 100%. Then, based on the target variable's measured value in the baseline portion, the reconstruction value corresponding to the measured value is calculated using the abnormal time series reconstruction model. The reconstruction error is calculated based on the measured value and its corresponding reconstruction value. Finally, the average reconstruction error and average reconstruction error percentage of the target variable within the baseline portion time window are calculated. The average reconstruction error percentage is obtained by dividing the sum of the reconstruction errors of the target variable at all time points in the baseline portion by the length of the baseline portion time window. The average reconstruction error percentage is the average of the absolute values ​​of the reconstruction error percentages at all time points in the baseline portion. Secondly, the time series data of each target variable j in the fault part are sorted into a table in the order of timestamps, including the measured value, reconstruction value, reconstruction error, and reconstruction error percentage. The average reconstruction error and average reconstruction error percentage of the target variable j in the baseline part are added to the table. Finally, the statistical table of the time series data of the target variable j can be obtained. For any target variable , there is a time series data statistics table , the time series data statistics table The failure mode of the target variable j is characterized, including the timing characteristics of the failure. At the same time, the average reconstruction error and the average reconstruction error percentage of the baseline part are used to characterize the time pattern of the target variable in the normal operation state.

[0107] In this embodiment, the time series parsing semantic library includes a plurality of predefined semantic units, which can be customized to add and delete predefined semantic units in the time series parsing semantic library according to task requirements. The predefined semantic units in the time series parsing semantic library are all stored in the memory. Based on the time series characteristics of industrial objects and the needs of industrial operation and maintenance decisions, the predefined semantic units include but are not limited to the following: overall trend (increase, decrease, fluctuation); inflection point characteristics (whether it exists, whether it is concentrated); range and average characteristics (range size, average size); extreme value characteristics (number of extreme points, maximum value, minimum value); volatility characteristics (whether it is stable, the degree of fluctuation, range of range); stage characteristics (whether there is obvious segmented behavior); periodic or repetitive pattern (yes / no); local anomaly-mutation point (yes / no, more / less); local anomaly-sharp fall phenomenon (yes / no, more / less); initial stage feature summary; mid-stage feature summary; late stage feature summary. For example, Figure 4 The time series parsing semantic library shown in the figure includes 8 predefined semantic units, namely: a semantic unit containing overall trends, a semantic unit containing inflection point features, a semantic unit containing extreme value features, a semantic unit containing repeated patterns, a semantic unit containing fluctuation features, a semantic unit containing stage features, a semantic unit containing local anomalies-mutation points, and a semantic unit containing local anomalies-sharp decline phenomena.

[0108] Furthermore, the time series semantic parsing prompt words are constructed using the time series parsing semantic library, industrial process description text, all measurement point information, target variable information, and time series data statistical table. The time series semantic parsing semantic library, industrial process description text, all measurement point information, target variable information, and time series data statistical table are written into the prompt words. The final constructed time series semantic parsing prompt words are:

[0109] “Your task is to describe the deviation of the measured value of a measurement point from the ideal normal value, focusing on [SEMANTIC_FEATURES] and key anomalies;

[0110] Background information:

[0111] 1. Industrial Process: [PROCESS_INFO]

[0112] 2. All measurement points: [ALL_SENSORS]

[0113] 3. Target measurement point: [TARGET_SENSOR]

[0114] 4. Time series data statistics table (providing measured values, reconstructed values, reconstructed errors, percentage of reconstructed errors over time, and average reconstructed errors and average percentage of reconstructed errors under normal operating conditions): [TABLE]

[0115] Only focus on time periods where the reconstruction error or reconstruction error percentage significantly exceeds the corresponding error under normal operating conditions, as these periods indicate a possible fault. Time periods where the reconstruction error or reconstruction error percentage is close to normal operating conditions are not of concern. During these significant deviations, describe two things:

[0116] 1. The trend of the measured value (increasing, decreasing, stable or cyclical fluctuation)

[0117] 2. Whether the measured value is above or below the ideal normal value

[0118] If the deviation is not significant during the entire time period, it is sufficient to state that the variable has no obvious abnormality.

[0119] Provide quantitative metrics (i.e., time intervals) and qualitative insights into how measurements changed during these periods of significant deviation. Focus on observable patterns and avoid speculating and drawing conclusions about root causes. Do not use subheadings or Markdown syntax. Keep your answer concise—no more than 100 words.

[0120] In the above time series semantic parsing prompt words, [SEMANTIC_FEATURES] represents the predefined semantic unit in the time series parsing semantic library, which is used to limit the content of the time series feature semantic text output by the large model; [PROCESS_INFO] represents the descriptive text of the industrial process background and principles; [ALL_SENSORS] represents all measurement point information of the industrial object, that is, variable information, including the variable name, location, physical meaning, unit, etc.; [TARGET_SENSOR] represents the target measurement point information, that is, target variable information, including the variable name, location, physical meaning, unit, etc.; [TABLE] represents the time series data statistical table corresponding to the target variable.

[0121] Furthermore, the above-mentioned time series semantic analysis prompt words are input into the big model, and the big model will output the time series feature semantic text of all target variables one by one ,in Represents the time series feature semantic text corresponding to the target variable j, , Representing a collection The total number of target variables in the dataset. The time series semantic text of the target variables specifically includes a time series semantic text description of the target variables and possible fault analysis. The time series semantic text description is associated with predefined semantic units. For example, an example of a large model output is "Variable 1 increases steadily between time 1 and time 2, and fluctuates periodically between time 2 and time 3. Its numerical trend and abnormal fluctuations may indicate an anomaly."

[0122] Step 4: Build a time series semantic parsing middle platform. Based on the time series feature semantic text of the target variable, use the time series semantic parsing middle platform to recall the corresponding fault knowledge text from the industrial object fault knowledge base, so as to be used for building the fault diagnosis prompt words of the large model in the subsequent steps.

[0123] In this embodiment, Figure 4 As shown, the time series semantic parsing platform includes a time series parsing semantic library, a time series parsing tool library, and a fault knowledge recall module. Among them, the time series parsing semantic library includes multiple predefined semantic units, see step 3 for details, which can be added, deleted, or replaced according to the downstream tasks of the industrial object; for example, the original time series parsing semantic library includes overall trends, inflection point features, extreme value features, volatility features, and stage features. Different semantic units are selected for different attributes of downstream tasks of different industrial objects. For example, the fault of gas turbine equipment is often manifested as a high or low value, which has little correlation with features such as fluctuations and inflection points. Therefore, the volatility features and inflection point features can be deleted from the original time series parsing semantic library, and then the range and average features can be added; for example, for the monitoring of power systems, its faults are generally manifested as spikes and fluctuations. Therefore, the overall trend and extreme value features in the original time series parsing semantic library can be deleted, and then the spike features can be added. The time series parsing tool library includes a large model for parsing time series data, an abnormal time series reconstruction model, and conventional time series data analysis methods, including principal component analysis and mean calculation. The time series parsing tool library is stored in memory and can be directly called from the memory when needed. Based on the time series parsing semantic library and the time series parsing tool library, semantic text (i.e., time series feature semantic text) can be generated for time series data in industrial objects. Other tools, such as principal component analysis, can also be called based on the needs of downstream tasks of the industrial objects. For example, principal component analysis can be used to calculate the principal components of time series data and reconstruct the data to remove noise. Then, "After principal component analysis and data reconstruction..." can be added to the existing time series feature semantic text to obtain the final time series feature semantic text. The fault knowledge recall module is used to recall the corresponding fault knowledge text from the industrial object fault knowledge base based on the existing time series feature semantic text of the target variable, such as the text description of the fault, the location of the fault, etc. The fault knowledge recall module can be used to extract fault knowledge text that is strongly related to the time series feature semantic text of the target variable. The industrial object fault knowledge base includes fault knowledge text collected from multiple channels such as sensor measurement point data, fault diagnosis records, expert experience, fault operation and maintenance suggestions, and fault root cause analysis collected during the operation and maintenance process of the industrial object.

[0124] Furthermore, based on the time series feature semantic text of the target variable, the time series semantic parsing platform is used to recall the corresponding fault knowledge text from the industrial object fault knowledge base, specifically including: Figure 5As shown, firstly, all the fault knowledge texts in the industrial object fault knowledge base are recorded as ,in represents the fault knowledge text corresponding to the bth fault category in the industrial object fault knowledge base, and B represents the total number of fault categories in the industrial fault knowledge base. Then, based on the requirements of downstream tasks (such as fault diagnosis and operation and maintenance decision-making) for the industrial object, if necessary (i.e., if the given target variable's temporal feature semantic text is insufficient to infer the fault), the temporal feature semantic text of the target variable is improved using the temporal feature semantic library and temporal feature tool library of the temporal semantic parsing center. This involves invoking the corresponding tools in the temporal feature semantic library to add the target variable's temporal feature semantic text or to retrieve variable information from all measurement point variable data that the large model believes may be helpful for fault diagnosis, ultimately obtaining the improved temporal feature semantic text of the target variable. Of course, if this is not necessary, the temporal feature semantic text of the target variable does not need to be improved using the temporal feature semantic library and temporal feature tool library. Instead, the temporal feature semantic text of the target variable obtained in step 3 can be used for downstream tasks. Secondly, the fault knowledge recall module of the temporal semantic parsing center recalls the corresponding fault knowledge text from the industrial object fault knowledge base based on the existing temporal feature semantic text of the target variable.

[0125] Furthermore, the fault knowledge recall module specifically includes the following steps: First, all fault knowledge texts in the industrial object fault knowledge base and the time series feature semantic texts of the target variable are encoded into corresponding embedding vectors through the text encoder in the fault knowledge recall module. Specifically, after each fault knowledge text in the industrial object fault knowledge base is passed through the text encoder, the corresponding embedding vector can be obtained, and its calculation formula is:

[0126]

[0127] Where, Indicates fault knowledge text The corresponding embedding vector, Represents the text encoder; similarly, the time series feature semantic text of each target variable After passing through the text encoder, the corresponding embedding vector can be obtained, and its calculation formula is:

[0128]

[0129] Where, Represents the temporal feature semantic text of the target variable j Then, the similarity between the embedding vector of the time series feature semantic text of each target variable and the embedding vector of all fault knowledge texts is calculated. The calculation formula is:

[0130]

[0131] Where, Represents the similarity calculation function, Represents the modulus of the vector. Finally, the fault category corresponding to the target variable is screened by judging whether the calculated similarity is greater than the preset similarity threshold, and the corresponding fault knowledge text is recalled from the industrial object fault knowledge base based on the fault category. Specifically, if the calculated similarity between the embedding vector of the target variable and the embedding vector of a certain fault knowledge text is greater than the preset similarity threshold, it is considered that the fault category corresponding to the fault knowledge text is the possible fault category of the target variable, and the corresponding fault knowledge text is recalled from the industrial object fault knowledge base and added to the fault diagnosis prompt words of the large model in the subsequent steps, and the corresponding fault knowledge text is recorded as .

[0132] Step 5: Based on the industrial process description text, all measurement point information, recalled fault knowledge text, and the time series feature semantic text of the target variable, a time series fault diagnosis prompt word is constructed and used as the input of the large model to perform the fault diagnosis task of the industrial object. The large model outputs the fault category diagnosis result and reasoning process, and the output results of the large model are voted and reviewed multiple times to obtain the final fault category diagnosis result and reasoning process.

[0133] Specifically, if Figure 5 As shown, the time series characteristic semantic text of the target variable obtained in step 3 and the fault knowledge text recalled in step 4 are used to activate the reasoning ability of the large model. The time series characteristic semantic text of the target variable covers the trend characteristics of the time series and the fault representation, and corresponds to the fault occurrence pattern in the fault knowledge text. With its powerful context modeling and reasoning capabilities, the large model can infer the fault category diagnosis result from the time series characteristic semantic text of the target variable and trace the cause of the fault based on the time series characteristic semantic text. The constructed time series fault diagnosis prompt words are specifically:

[0134] "You are an expert in troubleshooting industrial processes. Your task is to analyze the following information to identify potential failures:

[0135] 1. Industrial Process: [PROCESS_INFO]

[0136] 2. Measurement point (list of available sensors): [ALL_SENSORS]

[0137] 3. Fault knowledge text (known faults and their characteristics): [FAULT_KNOWLEDGE]

[0138] 4. Time series observations: [TIME_DESP] (Description of the deviation between the measured and predicted values; larger absolute values ​​indicate more severe failures.)

[0139] Follow the steps:

[0140] 1. Identify critical sensors: Determine which sensors’ measurements deviate significantly from expected values.

[0141] 2. Cross-reference with fault knowledge text: Compare observed patterns with known fault characteristics to infer possible faults. Note that sometimes different faults may exhibit similar patterns on certain variables. Therefore, do not rely solely on the fault knowledge provided; you should combine the fault causes and process workflow to reasonably infer the changing trends of different variables.

[0142] 3. Determine whether the data is sufficient: If the time series observations provide sufficient information to identify the fault, output the fault number in the following format:

[0143] <reasoning> [Step-by-step analysis]< / reasoning>

[0144] <answer> number< / answer>

[0145] If the time series observations are insufficient or key variables are missing, use the get_target_table tool to query detailed data for a specific sensor. Before querying, make sure the sensor name exists in the measurement point list. Output the tool call in the following format:

[0146] <reasoning> [Step-by-step analysis]< / reasoning>

[0147] <tool> get_target_table ("[SENSOR_1]")< / tool>

[0148] Important note: Please note that sometimes different failures may show similar patterns on certain variables. Do not rely solely on the failure knowledge provided; you should combine the failure cause and process workflow to reasonably infer the changing trends of different variables.

[0149] Verify that the sensor name you are querying exists in the list of measurement points. Call get_target_table only if necessary:

[0150] 1. When time series observations are insufficient to determine the fault;

[0151] 2. When key variables used for fault diagnosis are missing from the observations. ”

[0152] In the above time series fault diagnosis prompt words, [PROCESS_INFO] represents the text description of the industrial process background and principles; [ALL_SENSORS] represents all measurement points of the industrial object, that is, variable information; [FAULT_KNOWLEDGE] represents the possible fault knowledge text recalled in step 4; and [TIME_DESP] represents the time series feature semantic text of the target variable obtained in step 3. <reasoning> [Step-by-step analysis]< / reasoning> Specify the output format of the large model, that is, <reasoning> and< / reasoning> Output the step-by-step reasoning process between two identifiers; similarly <answer> number< / answer> Indicates from <answer> and< / answer> The identifiers output the fault diagnosis classification results, i.e., the fault category diagnosis results. If the target variable does not provide enough information to support the large model to diagnose the fault type, the tool get_target_table is called. This tool will search the time series feature semantic text of all measurement point variables to find the target variable data and provide it to the large model for further analysis. <tool> get_target_table ("[SENSOR_1]")< / tool> Indicates that the data of SENSOR_1 measuring point variable is called.

[0153] Furthermore, the constructed temporal fault diagnosis prompt words are used as input to the large model, which then outputs the fault category diagnosis results and the reasoning process. The large model's output results are then voted on multiple times, i.e., repeated experiments are performed to obtain multiple fault categories output by the large model. The fault category with the most votes is selected as the final fault category diagnosis result. For example, the temporal fault diagnosis prompt words are input into the large model, the results are output, and this is repeated five times. The fault categories output in the five outputs are 4, 5, 3, 4, and 4, respectively. The number 4 with the most votes is selected as the voting result, indicating that the final fault category diagnosis result is the fourth fault category. The reasoning process corresponding to the final fault category diagnosis result is then used as the final reasoning process. Leveraging the reasoning capabilities of the large model, the temporal changes of the target variable are associated with the textual representation of fault knowledge in the industrial object fault knowledge base. The large model's reasoning process details how to infer the fault category from the semantic text of the temporal features, making the fault diagnosis results interpretable.

[0154] Step 6: The large model integrates the fault category diagnosis results and reasoning process as well as the recalled fault knowledge text and other information to automatically generate an operation and maintenance diagnosis form, which is then stored in the industrial object fault knowledge base after expert review.

[0155] Specifically, if Figure 6As shown in the figure, first, a time series visualization image of the original data of the target variable containing the abnormal representation of the fault part and the reconstructed data is drawn. Then, the time series visualization image, the fault category diagnosis results and reasoning process output by the large model obtained in step 5, the fault knowledge text recalled from the industrial object fault knowledge base obtained in step 4, and the time series feature semantic text of the target variable obtained in step 3 are input into the large model for knowledge extraction. The output content of the large model is restricted by the prompt word instruction. The prompt word input to the large model is:

[0156] "Task information: Combine the given fault diagnosis information, recall relevant knowledge from the given knowledge base, perform integrated reasoning, and output possible cause analysis, operation and maintenance suggestions, and some other information about the fault. Format requirements: The possible cause analysis should be as comprehensive as possible. The content of the provided knowledge base document has been manually analyzed and repaired, and may not be applicable to the scenario of this case. Please analyze with caution. Operation and maintenance suggestions should be as simple and straightforward as possible. The output content includes: the first occurrence time of the fault in the historical data, the possible causes of the fault, and operation and maintenance suggestions for the fault." Then, the knowledge information extracted by the large model is used to supplement the blank content of the operation and maintenance diagnosis form template to realize the automatic generation of the operation and maintenance diagnosis form. Finally, the generated operation and maintenance diagnosis form is submitted to human experts for further review, and the verified operation and maintenance diagnosis form is stored in the industrial object fault knowledge base.

[0157] Furthermore, the O&M diagnostic sheet includes five sections: a diagnostic summary, warning point details, a measurement point trend graph, a possible cause analysis, and O&M recommendations. The diagnostic summary includes the device and component path to the fault point, as well as the time of the fault. Warning point details include the source tag point name, source tag point description, alarm type, first occurrence time, and most recent occurrence time. The measurement point trend graph includes a time series visualization of the original and reconstructed data, along with a textual description and analysis of the time series changes.

[0158] See also Figure 7 An embodiment of the present invention provides an industrial time series semantic parsing and operation and maintenance decision-making system based on a large model, including one or more processors and a memory, and the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the industrial time series semantic parsing and operation and maintenance decision-making method based on a large model in the above embodiment.

[0159] The embodiment of the industrial time series semantic parsing and operation and maintenance decision system based on the large model of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, if Figure 7 As shown, this is a hardware structure diagram of any device with data processing capability where the industrial time series semantic analysis and operation and maintenance decision system based on the large model of the present invention is located. Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0160] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0161] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present invention. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0162] An embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model in the above embodiment is implemented.

[0163] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or memory. The computer-readable storage medium may also be any device with data processing capabilities, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capabilities and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0164] The method described in the present invention is applicable to industrial fields including, but not limited to, chemical engineering and energy, and can be used in any industrial production process requiring fault diagnosis and intelligent operation and maintenance. The present invention was applied to two real-world examples: fault diagnosis of a three-phase flow facility in a chemical process and operation and maintenance of equipment in a real thermal power plant, to demonstrate its effectiveness. The following examples use process descriptions, fault types, and measurement point information from real industrial processes as reference material for prompt word design. Large language models used in the present invention include, but are not limited to, GLM4, Qwen2.5, Deepseek-R1, and Deepseek-V3. It should be understood that the performance of large language models directly impacts the performance of downstream equipment operation and maintenance tasks. The best-performing large language model should be used, if the computing power of the equipment allows. Considering that large language models require training on a large corpus, much of this training data comes from online media and does not include relevant knowledge about specific industrial processes, the present invention designs prompt words for specific industrial scenarios that explain the background and principles of the industrial process, infusing the large language model with relevant knowledge from the specific field.

[0165] Example 1: Fault diagnosis of a three-phase flow facility in a chemical process

[0166] This embodiment takes a three-flow facility fault diagnosis task in a chemical process as an example to illustrate the steps of the present invention and verify its effectiveness.

[0167] The three-phase flow facility is designed to provide controlled and measurable flows of water, oil, and air to pressurized systems within chemical plants. In the supply stage, air is produced by two compressors operating in conjunction. It first flows into a buffer vessel to stabilize the pressure, then undergoes filtration and cooling. Finally, its flow rate is measured by a flow meter and controlled by a pneumatic valve before entering the mixing process. Simultaneously, water and oil are pumped from their respective storage tanks using water and oil pumps. After flow measurement and control, they also enter the mixing stage. At the mixing point, the air, water, and oil converge to form a mixture. The mixture then enters a flow circuit, where it can be transported using pipes of varying diameters and orientations depending on the requirements. The mixture then flows to the separation stage, initially reaching the top separator for initial separation of gas and liquid. The liquid portion then flows into a three-phase separator for further separation into gas, oil, and water. The gas is exhausted to the atmosphere, while the separated oil and water flow into their respective coalescers for further separation before returning to their original storage tanks, completing the entire industrial process cycle. The three-phase flow facility includes 24 fault-related measurement points, as shown in Table 1.

[0168] Table 1: Three-phase flow experimental data measurement point information

[0169]

[0170] First, data from the three-phase flow facility under various normal operating conditions is collected. Representative samples are selected and an abnormal time series reconstruction model is constructed, as described in step 1. Furthermore, the abnormal data to be analyzed is divided into a baseline portion and a fault portion. Anomaly scores are calculated, and target variables are selected as described in step 2. Key measurement point variables that can characterize the fault mode are selected from the 24 measurement point data as target variables.

[0171] Furthermore, if Figure 8 As shown, a time series semantic parsing prompt is constructed based on three-phase flow facility process information, measurement point information, and other information. This is then input into a large language model for a semantic description of the time series features. For example, the time series semantic text of the target variable output by the large model is: "Gas supply pressure (PT312, MPa): Between 102 and 200, the gas supply pressure (PT312) deviated significantly from the ideal value, showing an upward trend and periodic fluctuations, remaining consistently above the ideal value. The maximum deviation occurred at time 144, with a difference of 0.225 MPa (an error of 60.945%), indicating a possible malfunction in the gas supply system during this period."

[0172] Furthermore, the text encoder of the fault knowledge recall module encodes the semantic text of the temporal features of all target variables obtained above into an embedded vector form. The similarity is calculated with the embedded vectors of the fault knowledge text in the industrial object fault knowledge base to recall possible faults. According to the industrial object fault knowledge base, there are five fault types in three-phase flow facilities. For example, one fault knowledge text is as follows: "Fault Type 5: Surge Phenomenon in Gas-Liquid Two-Phase Flow: Surge is a transient phenomenon that can occur in a riser with multiphase flow when the velocities of gas and liquid are relatively low. Liquid tends to accumulate at the bottom of the riser, blocking the flow of gas. Due to this blockage, pressure builds at the blockage until it is sufficient to flush the liquid (and gas) out of the riser. After this surge, the remaining liquid in the riser falls back, forming a new blockage and restarting the cycle. This can cause large fluctuations in pressure and flow, which can affect or even damage the equipment."

[0173] Furthermore, if Figure 9 As shown, a fault diagnosis prompt word is constructed based on the three-phase flow facility process information, measurement point information, fault knowledge text, and the time series feature semantic text of the target variable to perform the fault diagnosis task. If the target variable information is insufficient to infer the fault, the tool call in steps 4 and 5 is performed, and the output results of the large model are voted and reviewed multiple times to obtain the final fault category diagnosis result and reasoning process. An example is as follows:

[0174] “ <reasoning>

[0175] The analysis began by identifying key sensors that were significantly deviating from their ideal values. The PT312 (supply air pressure) showed a continuous increase, while the FT305 (air flow) showed a steady decrease, both indicating a restricted air supply. This was consistent with Fault 1 (blocked air line), where manually closing the valve would reduce air flow, resulting in increased pressure (PT312↑) and decreased flow (FT305↓).

[0176] Downstream effects include increased pressure at the bottom of the riser (PT401↑) and decreased pressure at the top (PT408↓), as well as density fluctuations (FT407 and FT406). These signs indicate the presence of slugging (Fault 5). However, the slugging here is likely a secondary effect of the reduced air flow caused by Fault 1, as low air velocity promotes liquid accumulation and slugging.

[0177] If only Fault 5 were to occur, it would require a normal air supply and operating conditions resulting in low flow rates, which contradicts the observed air system anomalies. The data strongly suggests that Fault 1 is the root cause, and the slug is a secondary phenomenon rather than an independent fault. Other faults such as a water line blockage (Fault 2) or a blocked separator inlet (Fault 3) were ruled out due to sensor pattern mismatches (e.g., fluctuations in water flow on the FT104 were inconsistent with the continued manual valve closure). Fault 5 (2-inch line pressurization) could not be confirmed due to the lack of PT417 data. Therefore, the primary fault is Fault 1;

[0178] < / reasoning>

[0179] <answer> 1< / answer> ”.

[0180] The above steps enable large language models to analyze the temporal characteristics and changing patterns of time series. This approach can be applied to intelligent industrial operations and fault diagnosis tasks, replacing human experts in fault diagnosis and inferring their causes, significantly reducing labor costs. Experiments were conducted on various large language models to verify the effectiveness of this invention, and the results are shown in Table 2.

[0181] Table 2: Comparison of experimental results of different large language models

[0182]

[0183] Example 2: Equipment Operation, Maintenance, and Fault Diagnosis in a Real Thermal Power Plant

[0184] This embodiment takes a real thermal power plant equipment operation and maintenance and fault diagnosis task as an example to illustrate the steps of the present invention and verify its effectiveness.

[0185] Gas turbines are core equipment in the thermal power industry. During the gas-fired power generation phase, ambient air is compressed to form high-temperature, high-pressure gas, which mixes with natural gas and burns in the combustion chamber. The resulting expanded airflow drives the gas turbine, generating electricity. Exhaust gas from the gas turbine, exceeding 600°C, enters the waste heat boiler (HRSG), where heat exchange converts feed water into high-pressure steam. This steam drives the steam turbine, which in turn drives a second generator for secondary power generation, achieving cascaded energy utilization. Heat is transferred using water pumps. The circulating water pump maintains a low temperature in the condenser, allowing the low-pressure steam from the steam turbine to recondense into water. The condensate pump pressurizes the condensate and returns it to the HRSG. The feedwater pump continuously supplies high-pressure water to the HRSG, ensuring continuous steam generation.

[0186] The abnormal time series reconstruction model of thermal power is trained by data under normal operating conditions. The reconstruction error is obtained based on the abnormal time series reconstruction model. The target variable with obvious abnormal characteristics is selected using the reconstruction error and the abnormal indicator function. The actual reconstruction effect of the abnormal time series reconstruction model is as follows: Figure 10 As shown. For the measurement point variables of the gas turbine, the time series feature semantic text of a target variable output by the large language model is as follows: "Condensate pump sealing water pressure: Between time points 1 and 100, the sealing water pressure of condensate pump B deviates significantly from the ideal value. During this period, the measured values ​​generally show a downward trend, accompanied by periodic fluctuations. The measured values ​​are always lower than the ideal value. At time point 2, the deviation is the largest, and the difference between the measured value and the ideal value is 0.297 Pa (an error of 239.382%). This indicates that there may be a potential fault in the sealing water pressure of condensate pump B during this time period."

[0187] Furthermore, this embodiment experimentally validates eight typical faults encountered in thermal power generation: abnormal gas turbine compressor cooling air pressure, abnormal natural gas temperature in the gas turbine fuel pre-module, abnormal gas turbine combustion chamber acceleration, high condensate pump motor lower bearing temperature, abnormal condensate pump seal water pressure, abnormal waste heat boiler low-pressure side steam flow, high feedwater pump bearing (coupling end) temperature, and high circulating water pump motor stator temperature. An example of the fault information document is: "Fault Type 4: Abnormal condensate pump mechanical seal water pressure: The condensate pump is under high vacuum during standby operation, so a reliable seal is essential. The condensate pump seal water pressure continues to display low alarms, significantly lower than the estimated value."

[0188] like Figure 11 As shown in the figure, the fault knowledge text corresponding to the possible fault types is recalled from the industrial object fault knowledge base, the fault diagnosis prompt words are constructed, and the large language model is input for fault diagnosis. The fault category diagnosis results and reasoning process are shown as follows:

[0189] “ <reasoning>

[0190] 1. Key sensor identification:

[0191] - The seal water pressure of condensate pump B (30LCB12CP101.XQ01) has the largest deviation, reaching 239.382%, which is significantly lower than the ideal value;

[0192] - The thrust bearing temperature of condensate pump B (30LCB12CT109.XQ01) was higher than the ideal value, with a deviation of 4.417°C (10.574%);

[0193] - The closed cold water supply main pressure (30PGB30CP101.XQ01) is lower than the ideal value, with an error of 29.094%;

[0194] - Generator active power (30MBY10CE901.XQ01) is lower than expected, with an error of 24.409%;

[0195] 2. Compare with fault knowledge:

[0196] - Fault type 4 clearly describes "continuous low alarm of condensate sealing water pressure", which fully matches the abnormal sealing water pressure;

[0197] - The increase in thrust bearing temperature may be related to cooling or lubrication failure caused by insufficient sealing water pressure, but there is no direct corresponding item in the fault knowledge;

[0198] - Low closed cooling water pressure may affect bearing cooling, but no relevant alarm is triggered;

[0199] - The decrease in generator power may be due to abnormal sealing water pressure, which leads to reduced efficiency of the condensate pump, indirectly affecting the steam cycle;

[0200] 3. Troubleshoot other faults:

[0201] - Fault type 3 requires the bearing temperature to be "5°C above the estimate", but the observed bearing temperature is lower than expected;

[0202] - Fault type 5 requires abnormal low-pressure steam flow, but no relevant data is provided;

[0203] - Other fault types are not related to the currently observed sensor;

[0204] Conclusion: Abnormal sealing water pressure is direct evidence of fault type 4, and other observation data can be regarded as secondary related factors or indirect influences;

[0205] < / reasoning>

[0206] <answer> 5< / answer> ”.

[0207] Furthermore, the large language model is used to extract relevant knowledge from the industrial object fault knowledge base, the fault category diagnosis results and reasoning process output by the large model, and the time series feature semantic text of the target variable, and integrate information such as the fault start time, fault cause analysis, and operation and maintenance suggestions. The operation and maintenance diagnosis form is automatically generated based on the operation and maintenance diagnosis form template. The generation process and the operation and maintenance diagnosis form example are as follows: Figure 12 shown.

[0208] The fault category diagnosis results and reasoning process generated based on the large model not only illustrate the relationship between the target variable's fault characteristics and fault type, but also provide the relevant reasoning and analysis process. Furthermore, they verify that the semantic context of the time series features of the key target variables is necessary and sufficient for the diagnosis results. Overall, the method provides interpretable diagnostic results and reasoning processes, which are more interpretable than traditional fault diagnosis methods. Experiments were conducted on various large language models, and the results are shown in Table 3.

[0209] Table 3: Comparison of experimental results of different large language models

[0210]

[0211] The comparison results shown in Table 3 show that the average accuracy of the method described in this invention using four different large language models is 87.5%, 100%, 87.5%, and 100%, respectively. This indicates that in actual industrial scenarios, the method described in this invention, using different types of large language models, can describe time series data and reason about faults based on the specialized knowledge of industrial objects, revealing the correlation between time series features and faults, and demonstrating high accuracy in fault diagnosis. This demonstrates the feasibility and effectiveness of the method described in this invention.

[0212] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model, characterized by: The following steps are involved: Step 1: Build an abnormal time series reconstruction model based on the collected time series data of the industrial object under normal operation. Obtain reconstructed data based on the abnormal time series reconstruction model to determine whether the online collected time series data is abnormal. If so, skip to step 2; otherwise, end directly. Step 2: Select time series data with anomalies to construct a test set, calculate the reconstruction error of the test sample, and screen the variables based on the reconstruction error and the anomaly indicator function to obtain the target variable; Step 3: Obtain a time series data statistical table based on the time series data of the target variable, and use the time series parsing semantic library, industrial process description text, all measurement point information, target variable information, and the time series data statistical table to construct a time series semantic parsing prompt word. This is used as the input of the large model, and the large model outputs the time series feature semantic text of all target variables one by one; Step 4: Build a time series semantic parsing platform. Based on the time series feature semantic text of the target variable, use the time series semantic parsing platform to recall the corresponding fault knowledge text from the industrial object fault knowledge base. Step 5: Based on the industrial process description text, all measurement point information, recalled fault knowledge text, and the time series feature semantic text of the target variable, a temporal fault diagnosis prompt word is constructed as the input of the large model. The output results of the large model are voted and reviewed multiple times to obtain the final fault category diagnosis result and reasoning process; Step 6: The large model integrates the fault category diagnosis results, the reasoning process, and the recalled fault knowledge text to automatically generate an operation and maintenance diagnosis form, which is then reviewed by experts and stored in the industrial object fault knowledge base.

2. The large-model-based industrial time series semantic parsing and operation and maintenance decision-making method according to claim 1 is characterized in that: The abnormal time series reconstruction model is constructed based on the time series data collected under normal operating conditions of the industrial objects, and reconstructed data is obtained based on the abnormal time series reconstruction model to determine whether the time series data collected online has abnormalities, specifically including: Collect time series data of industrial objects under normal operating conditions to construct a measurement point data matrix; perform K-means clustering based on the measurement point data matrix to select n representative samples from it to construct a state matrix; calculate the weight vector between the time series data samples collected online and the state matrix, and construct an abnormal time series reconstruction model based on the weight vector and the state matrix; obtain the reconstructed data corresponding to the time series data currently collected online based on the abnormal time series reconstruction model, calculate the reconstruction error between the time series data currently collected online and the reconstructed data, and compare the reconstruction error with a preset error threshold to determine whether the time series data collected online has an abnormality.

3. The method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model according to claim 1 is characterized in that: The step 2 specifically includes the following sub-steps: Step 2.1: Select time series data with anomalies to construct a test set. Use the abnormal time series reconstruction model to obtain the reconstructed data of each test sample. Calculate the corresponding reconstruction error based on each test sample and its reconstructed data. Then, divide each test sample and its corresponding reconstruction error into a baseline part and a fault part based on the fault start time and fault end time. Step 2.2: Calculate the average reconstruction error based on the reconstruction error of the baseline part to set the abnormal threshold for each variable; Calculate the absolute value of the reconstruction error of each variable at all time points of the faulty part based on the reconstruction error of the faulty part; Step 2.3: Construct an anomaly indicator function based on the absolute value of the reconstruction error of each variable in the faulty part and its corresponding anomaly threshold, and use the anomaly indicator function to calculate the anomaly occurrence time and anomaly score of each variable; Step 2.4: Select the previous variable with the largest anomaly score and the previous variable with the earliest anomaly occurrence time to construct a candidate variable set; Step 2.5: Screen the candidate variables in the candidate variable set based on the variance difference between the test sample data of the baseline part and the test sample data of the fault part to obtain the target variable that can characterize the fault mode.

4. The method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model according to claim 1 is characterized in that: The step of obtaining a time series data statistics table based on the time series data of the target variable specifically includes: Based on the measured value of the target variable at the faulty part, the reconstruction value of the measured value is calculated using the abnormal time series reconstruction model, the reconstruction error is calculated based on the measured value and its reconstruction value, and the reconstruction error percentage is calculated based on the measured value and its reconstruction error; The reconstruction error is calculated based on the measured value of the target variable in the baseline part, and the average reconstruction error and the average reconstruction error percentage of the target variable in the baseline part time window are calculated according to the reconstruction error; For each target variable's time series data in the faulty part, the measured value, reconstruction value, reconstruction error, and reconstruction error percentage are organized into a table in the order of timestamps. The average reconstruction error and average reconstruction error percentage of the target variable in the baseline part are added to the table to finally obtain the time series data statistics table of the target variable.

5. The method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model according to claim 1 is characterized in that: The time series parsing semantic library includes multiple predefined semantic units, which include overall trends, inflection point characteristics, range and average characteristics, extreme value characteristics, volatility characteristics, stage characteristics, periodic or repetitive patterns, local anomalies-mutation points, local anomalies-sharp decline phenomena, early stage characteristic summary, mid-term stage characteristic summary and late stage characteristic summary.

6. The method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model according to claim 1 is characterized in that: The time series semantic parsing middle platform includes a time series parsing semantic library, a time series parsing tool library and a fault knowledge recall module, wherein the time series parsing tool library includes a large model, an abnormal time series reconstruction model, principal component analysis and mean calculation; the fault knowledge recall module is used to recall the corresponding fault knowledge text from the industrial object fault knowledge base based on the existing time series feature semantic text of the target variable, wherein the industrial object fault knowledge base includes the sensor measurement point data, fault diagnosis records, expert experience, fault operation and maintenance suggestions, and fault root cause analysis knowledge texts collected by the industrial object during its operation and maintenance process.

7. The method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model according to claim 6 is characterized in that: The fault knowledge recall module specifically includes: All fault knowledge texts in the industrial object fault knowledge base and the time series feature semantic texts of the target variable are encoded into corresponding embedding vectors through the text encoder in the fault knowledge recall module; Calculate the similarity between the embedding vector of the time series feature semantic text of each target variable and the embedding vector of all fault knowledge texts; The fault category corresponding to the target variable is screened by judging whether the calculated similarity is greater than a preset similarity threshold, and the corresponding fault knowledge text is recalled from the industrial object fault knowledge base based on the fault category.

8. The method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model according to claim 1 is characterized in that: The step 6 specifically includes: Draw a time series visualization image of the original data of the target variable containing abnormal representation and the reconstructed data of the fault part, and pass the time series visualization image, the fault category diagnosis result and reasoning process output by the large model, the recalled fault knowledge text and the time series feature semantic text of the target variable into the large model for knowledge extraction; limit the output content of the large model through prompt word instructions; use the knowledge information extracted by the large model to supplement the blank content of the operation and maintenance diagnosis form template, and realize the automatic generation of the operation and maintenance diagnosis form; and submit it to experts for review, and store the verified operation and maintenance diagnosis form in the industrial object fault knowledge base.

9. The method for industrial time series semantic parsing and operation and maintenance decision-making based on a large model according to claim 8 is characterized in that: The operation and maintenance diagnosis sheet includes a diagnosis summary, early warning point details, measurement point trend images, possible cause analysis and operation and maintenance suggestions; wherein, the diagnosis summary includes the equipment and component path of the fault measurement point and the time of fault occurrence; the early warning point details include the source tag point name, source tag point description, alarm type, first occurrence time and most recent occurrence time; the measurement point trend graph includes a time series visualization image showing the original data and reconstructed data, as well as a text description and analysis graph of the time series changes.

10. An industrial time series semantic parsing and operation and maintenance decision-making system based on a large model, comprising one or more processors and memories, characterized in that: The memory is coupled to the processor; wherein, the memory is used to store program data, and the processor is used to execute the program data to implement the industrial time series semantic parsing and operation and maintenance decision-making method based on a large model as described in any one of claims 1-9.

Images