An industrial time series data quality evaluation method, system, device and storage medium based on large model adaptive zooming

Abstract

The application discloses an industrial time series data quality evaluation method, system, device and storage medium based on a large model adaptive zooming, comprising obtaining measurement point description information of a target measurement point and original time series data of the target measurement point in a first time range, and generating first time series analysis data based on the original time series data; inputting the measurement point description information and the first time series analysis data into a large model for data quality reasoning, so that the large model outputs a final evaluation result or a zooming request containing a target time interval. The method can quickly filter out abnormal data on one hand, and automatically match the appropriate resampling granularity according to the data characteristics of each measurement point on the other hand, effectively solving the adaptability problem of the fixed fragmentation number strategy in different scenes such as high-frequency vibration and smooth trend.

Description

Technical Field

[0001] This invention relates to the field of industrial data quality evaluation technology, specifically to a method, system, device, and storage medium for evaluating the quality of industrial time-series data based on large-model adaptive zoom. Background Technology

[0002] Existing data quality assessment schemes based on large models all require data resampling before inputting into the large model. Resampling strategies typically employ a fixed number of slices to uniformly divide short-period, densely sampled data. However, the characteristics of data from different industrial measurement points vary significantly. High-frequency fluctuating data requires more slices to retain detailed features of anomalous changes, while data with stable changes can accurately represent trends with fewer slices. When faced with diverse measurement point types, the fixed-slice-number strategy either results in overly coarse slices that cause key anomalous features to be smoothed out during resampling, or overly dense slices that fill the LLM input window with redundant information, impacting analysis efficiency and accuracy.

[0003] Furthermore, existing solutions indiscriminately call upon large models for deep analysis of all measurement points when conducting data quality assessments. However, simple anomalies such as dead points, incorrect data types, and missing descriptions can be quickly identified using preset rules, without consuming the inference resources of large models. This indiscriminate approach leads to a waste of computational resources and fails to meet the efficiency requirements of industrial scenarios.

[0004] Meanwhile, the output of large models in existing solutions is usually unstructured natural language text, making it difficult for downstream systems to automatically parse and integrate the judgment results, which limits the application of evaluation results in industrial data quality management systems. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is that existing fixed resampling methods suffer from the smooth loss of abnormal features and the redundancy of input information. Furthermore, existing solutions call large models for analysis without distinguishing all measurement points, resulting in high computational resource overhead.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a method for evaluating the quality of industrial time-series data based on large-model adaptive zoom, comprising: acquiring measurement point description information of a target measurement point and the original time-series data of the target measurement point within a first time range, and generating first time-series analysis data based on the original time-series data; inputting the measurement point description information and the first time-series analysis data into a large model for data quality inference, so that the large model outputs a final evaluation result or a zoom request containing a target time interval; responding to the zoom request, generating second time-series analysis data within the target time interval based on the original time-series data, wherein the target time interval is smaller than the first time range, and the data processing granularity of the second time-series analysis data is finer than that of the first time-series analysis data; inputting the second time-series analysis data into the large model for further data quality inference, and when the large model outputs a zoom request again, generating time-series analysis data within the corresponding target time interval based on the newly output zoom request and continuing data quality inference, until the large model outputs the final evaluation result.

[0008] As a preferred embodiment of the industrial time-series data quality evaluation method based on large-model adaptive zoom described in this invention, the first time-series analysis data and the second time-series analysis data both include multiple data units arranged in chronological order and time information corresponding to each data unit; wherein each data unit is generated based on the original sampled data within the corresponding time slice, and the time information includes the time point or time interval of the corresponding data unit; the data processing granularity of the second time-series analysis data is finer than that of the first time-series analysis data, including at least one of the following: within the same time range, the second time-series analysis data includes more data units than the first time-series analysis data; the time span corresponding to each data unit in the second time-series analysis data is smaller than the time span corresponding to each data unit in the first time-series analysis data; the amount of original sampled data retained by the second time-series analysis data per unit time is greater than the amount of original sampled data retained by the first time-series analysis data per unit time.

[0009] As a preferred embodiment of the industrial time-series data quality evaluation method based on large model adaptive zoom described in this invention, the method includes: in response to the zoom request, generating second time-series analysis data within the target time interval based on the original time-series data, comprising: re-querying the original sampling data within the target time interval from the time-series database storing the original time-series data according to the measurement point identifier of the target measurement point, the start time and end time of the target time interval; and generating the second time-series analysis data based on the re-queried original sampling data.

[0010] As a preferred embodiment of the industrial time-series data quality evaluation method based on large-model adaptive zoom described in this invention, the following steps are taken: when generating the time-series analysis data for each round, data statistical characteristics are calculated based on the original sampled data within the corresponding time range, and the resampling parameters for the current round are determined according to the data statistical characteristics. The time-series analysis data for the current round is then generated according to the resampling parameters. The data statistical characteristics include at least one of standard deviation, mean absolute difference, first-order autocorrelation coefficient, and extreme value density. The resampling parameters include at least the number of slices. Determining the resampling parameters for the current round based on the data statistical characteristics includes: the number of slices taking values ​​between a preset lower limit and a preset upper limit. The number of slices increases with the increase of the mean absolute difference, decreases with the increase of the first-order autocorrelation coefficient, and increases with the increase of the extreme value density.

[0011] As a preferred embodiment of the industrial time-series data quality evaluation method based on large-model adaptive zoom described in this invention, the following steps are included: inputting the second time-series analysis data into the large model for further data quality inference, comprising: retaining the inference context corresponding to the previous round of data quality inference; inputting the inference context and the second time-series analysis data into the large model together; wherein the inference context includes at least one of the following: measurement point description information input in the previous round, time-series analysis data input in the previous round, output content of the large model in the previous round, and analysis time interval corresponding to the previous round.

[0012] As a preferred embodiment of the industrial time-series data quality evaluation method based on large model adaptive zoom described in this invention, the method further includes, before generating the first time-series analysis data based on the original time-series data: acquiring sparse sampling data of the target measurement point within a second time range; wherein the second time range is larger than the first time range, and the sampling interval of the sparse sampling data is larger than the sampling interval of the original time-series data; performing rule detection on the target measurement point based on the sparse sampling data according to preset rules, wherein the preset rules include at least a dead point detection rule for detecting data that remains unchanged for a long time; in response to the rule detection detecting a data quality anomaly of a preset type, directly outputting the evaluation result corresponding to the data quality anomaly of the preset type, and terminating subsequent data quality inference; in response to the rule detection not detecting the data quality anomaly of the preset type, executing the step of generating the first time-series analysis data based on the original time-series data.

[0013] As a preferred embodiment of the industrial time-series data quality evaluation method based on large-model adaptive zoom described in this invention, the step of inputting the measurement point description information and the first time-series analysis data into the large model for data quality inference includes: calculating the semantic embedding vector of the measurement point description information; based on the semantic embedding vector, retrieving a preset number of historical samples with the highest semantic similarity to the target measurement point from the historical sample library, wherein the historical samples include the measurement point description information of the historical measurement point, historical time-series data fragments, and corresponding historical evaluation results; and inputting the retrieved historical samples, the measurement point description information, and the first time-series analysis data into the large model for data quality inference.

[0014] Another objective of this invention is to provide an industrial time-series data quality assessment system based on large-model adaptive zoom, which can execute the aforementioned industrial time-series data quality assessment method based on large-model adaptive zoom, and solve the problems of smooth loss of abnormal features and redundant input information caused by fixed resampling strategies in existing industrial time-series data quality assessment.

[0015] As a preferred embodiment of the industrial time-series data quality evaluation system based on large-model adaptive zoom according to the present invention, it includes: a data processing module, a large-model inference module, and a zoom control module; the data processing module is used to acquire the measurement point description information of the target measurement point and the original time-series data of the target measurement point within a first time range, and generate first time-series analysis data based on the original time-series data; the large-model inference module is used to input the measurement point description information and the first time-series analysis data into the large model for data quality inference, so that the large model outputs a final evaluation result or a zoom request containing the target time interval; the zoom control module is used to respond to the zoom request and control the data processing module. The module generates second time-series analysis data within the target time interval based on the original time-series data, and controls the large model inference module to input the second time-series analysis data into the large model for further data quality inference. The target time interval is shorter than the first time range, and the data processing granularity of the second time-series analysis data is finer than that of the first time-series analysis data. The zoom control module is also used to control the data processing module to generate corresponding time-series analysis data within the target time interval based on the newly output zoom request when the large model outputs a zoom request again, and to control the large model inference module to continue data quality inference until the large model outputs the final evaluation result.

[0016] Another object of the present invention is to provide an industrial time-series data quality assessment device based on large model adaptive zoom, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the industrial time-series data quality assessment method based on large model adaptive zoom.

[0017] Another object of the present invention is to provide a storage medium for industrial time series data quality assessment based on large model adaptive zoom, wherein a computer program is stored thereon, and when the computer program is executed by a processor, the steps of the industrial time series data quality assessment method based on large model adaptive zoom are implemented.

[0018] The beneficial effects of this invention are as follows: This invention employs a dynamic resampling parameter adaptive strategy based on data statistical characteristics, enabling the system to automatically match appropriate resampling granularity according to the data characteristics of each measurement point. This effectively solves the adaptability problem of fixed-slice strategies in different scenarios such as high-frequency vibration and stable trends. While ensuring the complete preservation of abnormal features, it maximizes the utilization efficiency of the LLM input window. Furthermore, through a multi-scale data sampling strategy and an intelligent agent iterative scaling feedback mechanism, it significantly improves the evaluation accuracy of complex industrial time-series data quality problems. It can effectively identify anomalies such as bad pixels and conversion errors that are difficult to handle by traditional methods, balancing evaluation efficiency and analysis depth. The rule engine completes the filtering of simple anomalies in milliseconds, and the LLM is only called when necessary, greatly reducing the consumption of computing resources. This allows the system to be implemented in real-time industrial scenarios. The use of XML format to strictly constrain the output results of large models ensures the structured and parsable nature of the judgment results, facilitating downstream system integration and automated processing. Moreover, through a mechanism of manual confirmation of closed loops and continuous accumulation of historical sample databases, the system's evaluation capability can be continuously improved with use. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 The present invention provides an overall flowchart of an industrial time-series data quality evaluation method based on large-model adaptive zoom.

[0021] Figure 2 The flowchart of dynamic resampling and zoom iteration is provided for an industrial time series data quality evaluation method based on large model adaptive zoom, which is provided by the present invention.

[0022] Figure 3The flowchart of the rule-based pre-filtering, historical sample retrieval enhancement, and manual confirmation closed-loop process of an industrial time-series data quality evaluation method based on large model adaptive zoom provided by the present invention.

[0023] Figure 4 This is a trend chart of raw data from normal measurement points, provided by an embodiment of an industrial time-series data quality evaluation method based on large-model adaptive zoom, according to the present invention.

[0024] Figure 5 This invention provides an example of an industrial time-series data quality evaluation method based on large model adaptive zoom, showing the original data trend of dead point measurement points. (a) is a 24-hour short-cycle data chart of dead point measurement points, and (b) is a 45-day long-cycle data chart of dead point measurement points.

[0025] Figure 6 This is a trend chart of raw data of bad point measurement points, provided by an embodiment of an industrial time-series data quality evaluation method based on large model adaptive zoom.

[0026] Figure 7 The original data trend chart of the type error measurement point is provided in the embodiment of the industrial time series data quality evaluation method based on large model adaptive zoom provided by the present invention, wherein (a) is the 24-hour short-cycle data chart of the type error measurement point, and (b) is the 45-day long-cycle data chart of the type error measurement point.

[0027] Figure 8 The image shows the results of a measurement point quality evaluation system, which is an embodiment of an industrial time-series data quality evaluation method based on large-model adaptive zoom provided by the present invention.

[0028] Figure 9 This invention provides a system flowchart for evaluating the quality of industrial time-series data based on large-model adaptive zoom. Detailed Implementation

[0029] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0030] Example 1, referring to Figure 1This invention provides an embodiment of an industrial time-series data quality evaluation method based on large-model adaptive zoom, comprising: acquiring measurement point description information of a target measurement point and original time-series data of the target measurement point within a first time range, and generating first time-series analysis data 100 based on the original time-series data; inputting the measurement point description information and the first time-series analysis data 100 into a large-model 200 for data quality inference, so that the large-model 200 outputs a final evaluation result or a zoom request containing a target time interval; responding to the zoom request, generating second time-series analysis data 300 within the target time interval based on the original time-series data, wherein the target time interval is smaller than the first time range, and the data processing granularity of the second time-series analysis data 300 is finer than the data processing granularity of the first time-series analysis data 100; inputting the second time-series analysis data 300 into the large-model 200 for further data quality inference, and when the large-model 200 outputs a zoom request again, generating time-series analysis data within the corresponding target time interval based on the re-output zoom request and continuing data quality inference until the large-model 200 outputs a final evaluation result.

[0031] In this embodiment, the large model 200 does not perform a one-time quality judgment on the input data. Instead, when the current data processing granularity is insufficient to form a final evaluation result, the large model 200 actively determines the target time interval that needs further analysis and outputs a zoom request. The system generates time series analysis data with finer data processing granularity for the target time interval and inputs it into the large model 200 again, thereby forming a zoom evaluation process that progresses step by step from coarse-grained global analysis to fine-grained local analysis.

[0032] When comparing the granularity of data processing for different time-series analyses, the number of data fragments formed within a unit time range, or the average time span covered by a single data fragment, is used as the granularity measurement indicator. The more data fragments within a unit time range, or the shorter the average time span covered by a single data fragment, the finer the data processing granularity.

[0033] The adaptive zoom referred to in this embodiment means that the large model 200 autonomously determines whether it is necessary to increase the data processing granularity within a local time range based on the current time series analysis data, and outputs a zoom request to determine the local analysis range when it is necessary to increase the data processing granularity. The system automatically generates time series analysis data within the corresponding target time interval based on the zoom request, without the need for manual specification of the time interval to be focused in advance. In addition, it can also carry the desired analysis granularity prompt to instruct the system to generate more fine-grained time series analysis data to continue inference.

[0034] Therefore, this implementation does not merely utilize the large model 200 to classify anomalies in the pre-generated data, but rather further uses the inference output of the large model 200 as control information for subsequent time-series data processing. The target time interval output by the large model 200 is used to control the system to extract data from the original time-series data and regenerate finer-grained time-series analysis data. The regenerated time-series analysis data then serves as the input for the next round of inference by the large model 200, thus forming a closed loop of mutual feedback between the large model inference process and the time-series data processing process.

[0035] In this embodiment, the large model 200 is a pre-trained large language model with natural language understanding and structured content generation capabilities. It can perform data quality inference based on the test point description information, time series analysis data, and anomaly rule definitions in the prompt words, and generate the final evaluation result or zoom request according to a preset output format. This invention is not limited by the specific network structure, parameter scale, or training method of the large model 200, and can be implemented using a locally deployed or remotely invoked large language model.

[0036] Specifically, this implementation method includes the following steps.

[0037] S101. Obtain the measurement point description information and the original time series data within the first time range.

[0038] The target measurement points for data quality evaluation are identified, and their description information is retrieved from the measurement point information database. This description information characterizes the equipment object, the measured physical quantity, and the data semantics corresponding to the target measurement point. For example, it may include one or more of the following: measurement point name, equipment name, measurement point type, physical meaning, engineering unit, measurement range, and the location of the associated equipment.

[0039] Obtain the raw time series data of the target measurement point within the first time range from the time series database. The first time range is the nearest h. st The original time series data for each hour is short-period densely sampled data within that time range, denoted as: , in, x represents the set of short-period densely sampled data of the target measurement point within the first time range. i This represents the measurement value of the i-th original sampled data arranged in chronological order of sampling time, and n represents the n-th original sampled data.

[0040] If n st This represents the number of raw sampled data obtained within the first time frame, expressed in h. st This indicates the duration of the first time range. In one example, h st Take 24 hours, n stTake 5000. In this embodiment, the original sampling data are also associated with corresponding sampling times.

[0041] S102. Generate the first time series analysis data 100 based on the original time series data.

[0042] P st Resampling is performed to adapt to the input length and analysis granularity of the large model 200, generating the first time series analysis data 100.

[0043] Specifically, P st Divide into n evenly in chronological order rs Each data partition is denoted as: , Where, n rs Let be the number of data slices. For the i-th data slice, take the sample value with the largest deviation from the slice mean as the representative value of that slice. And record the time point t corresponding to the representative value. rs i To obtain the representative value sequence and time series Simultaneously, the standard deviation of the sampled data within each data segment is calculated to obtain the standard deviation sequence. When multiple sampled values ​​have the same maximum absolute deviation, the earliest or latest sampled value can be selected as the representative value.

[0044] Time series Representative value sequence and standard deviation series The data is arranged in chronological order and formatted as a CSV table, serving as the first time series analysis data 100. The data unit of the first time series analysis data 100 is a data row consisting of a time point, a representative value, and a standard deviation. Each data unit is generated based on the original sampled data within the corresponding data slice, i.e., the time slice.

[0045] The data processing granularity referred to in this embodiment refers to the degree to which time-series analysis data represents the details of the original sampled data within a unit time range. In the case of n slices... rs At a given time, the time span h covered by the fragmentation st / n rs The smaller the value, the finer the data processing granularity; when the time range is fixed, the number of partitions n rs The larger the value, the finer the granularity of data processing.

[0046] In this step, the number of fragments n rs It can be a preset fixed value; in a preferred embodiment, n rs Based on P stThe statistical characteristics of the data are dynamically determined.

[0047] S103. Input the measurement point description information and the first time series analysis data 100 into the large model 200 to perform data quality inference.

[0048] The measurement point description information, the preset anomaly rule definition, the first time series analysis data 100CSV data table, and the output format constraints are assembled into prompt words, which are then input into the large language model to evaluate and infer the data quality of the target measurement point.

[0049] In one implementation, the prompt includes a system prompt and a user prompt. The system prompt includes at least the data quality assessment task, anomaly rule definition, field definition of the final evaluation result, field definition of the zoom request, and output format constraints; the user prompt includes at least the measurement point description information of the target measurement point and the first time-series analysis data 100 arranged in chronological order.

[0050] The anomaly rule definition is used to explain to the large model 200 the types of data quality anomalies to be identified and their meanings, such as bad point anomalies, conversion anomalies, and other anomalies that do not conform to the physical meaning of the measurement points and the data change patterns; the output format constraint is used to limit the output of the large model 200 to adopt a parsable structured format. In this embodiment, the XML format is used, and the output of the large model 200 is limited to one of the following two types:

[0051] The first is the final evaluation result, which uses the result label. It includes the is_good field to indicate whether the quality of the measurement point is good, the reason field to determine the reason, the fault_type field to indicate the type of abnormality, with values ​​of bad point, conversion abnormality or NULL, and the fault_range field to represent the abnormal range, including the start and end time of the abnormal range.

[0052] The second is the zoom request, which uses the zoom tag and includes the start and end fields, representing the start and end times of the target time interval for the requested focus analysis, respectively.

[0053] The large model 200 combines the semantic representation of the measurement points by the measurement point description information and the data change characteristics represented by the first time series analysis data 100 to perform data quality inference. When the large model 200 can determine the data quality of the target measurement point based on the current input, it outputs the final evaluation result and the process ends. When the large model 200 finds suspicious change characteristics in the first time series analysis data 100, but the current data processing granularity is insufficient to determine whether the suspicious change characteristics constitute a data quality anomaly, it may include at least one of the following situations: there are suspicious data units in the current time series analysis data that differ from adjacent data units by more than a preset degree; the representative values ​​or standard deviations of multiple consecutive data units show an abnormal change trend; the current time series analysis data can indicate the existence of anomaly candidate intervals, but is insufficient to determine the specific start and end times of the anomaly; or the measurement point description information cannot rule out the possibility that the suspicious change belongs to normal equipment operating conditions. Under the constraint of the prompt words, the large model 200 determines the target time interval based on the time point corresponding to the suspicious data unit and outputs a zoom request containing the target time interval, proceeding to step S104.

[0054] The target time interval can be determined by the time range corresponding to one or more consecutive suspicious data units, and a preset time can be extended on both sides of the start and end times of the suspicious data units to retain context data before and after the suspicious change occurs. The start and end fields in the zoom request should be within the time range corresponding to the current round of input time-series analysis data, with start preceding end.

[0055] For example, when the measurement point description information indicates that the target measurement point is a bearing temperature measurement point, the large model 200 can judge the zero value and sudden drop value in the data by combining the engineering unit and physical meaning of temperature; when the measurement point description information indicates that the target measurement point is a device speed measurement point, the large model 200 can judge the situation where the speed drops to zero by combining the device start-up and shutdown status, thereby avoiding isolated judgment based solely on the numerical value.

[0056] S104. In response to the zoom request, generate second time-series analysis data 300.

[0057] Parse the zoom request, read the `start` and `end` fields from the `zoom` tag to obtain the target time interval. The target time interval is within the first time range, and its length is less than the length of the first time range, i.e., `h`. zoom <h st Each zoom cycle extracts sampled data within the target time interval from the original time-series data, rather than resampling from the previous resampled data, in order to avoid the accumulation and loss of local abnormal features caused by multiple resampling cycles.

[0058] From the original time series data P stThe sampled data located within the target time interval is extracted, and the extracted data is resampled according to the resampling method in step S102 to generate second time series analysis data 300: the sampled data within the target time interval is evenly divided into new n rs Each data segment is divided into 100 data segments, and the representative value, representative time point, and standard deviation of each data segment are determined. The data segments are combined into a CSV format data table. Each data row of the first time series analysis data 100 can include three fields in sequence: "representative time point, representative value, and segment standard deviation". The large model 200 identifies the data change process of the target measurement point within the first time range according to the time order of each data row.

[0059] Because the length h of the target time interval zoom The length h is less than the first time range st Even if the number of fragments n of the second time series analysis data 300 rs 'Number of fragments n compared to the first time series analysis data 100 rs Similarly, the time span h covered by each data fragment in the second time series analysis data 300 is... zoom / n rs It also needs to meet the requirement that the time span h covered by each data fragment in the first time series analysis data 100 is less than the time span h of each data fragment. st / n rs This allows for a more detailed characterization of local jumps, extreme values, or short-term anomalies within the target time interval; that is, the data processing granularity of the second time-series analysis data 300 is finer than that of the first time-series analysis data 100. In other embodiments, the data processing granularity can be further refined by increasing the number of slices or increasing the retention ratio of the original sampled data.

[0060] S105. Input the second time series analysis data 300 into the large model 200 to continue reasoning and iterating until the final evaluation result is output.

[0061] The measurement point description information, anomaly rule definition, second time series analysis data 300, and output format constraints are assembled into prompt words and input into the large model 200, enabling the large model 200 to continue data quality inference for the target time interval. The output of the large model 200 in this round is still one of the final evaluation result or zoom request: when the second time series analysis data 300 can characterize the local details of suspicious change features, the large model 200 outputs the final evaluation result, and the zoom iteration ends; when the second time series analysis data 300 is still insufficient to support the final judgment, the large model 200 outputs a zoom request again containing a new target time interval, and the new target time interval is located within the previous target time interval and its length is further shortened; in response to the zoom request output again, the system generates time series analysis data with finer data processing granularity within the corresponding target time interval based on the original time series data, and inputs it into the large model 200 again for data quality inference.

[0062] In this embodiment, the large model 200 is not only the main body generating data quality evaluation results, but also the decision-making body for the scope of subsequent time series data processing. The large model 200 outputs a target time interval based on the time series analysis data at the current granularity. The system converts the target time interval into a truncation and fine-grained representation operation of the original time series data, thereby generating the time series analysis data required for the next round of inference. Therefore, an executable feedback relationship is formed between the inference output of the large model 200 and the time series data processing operation. This allows the system to focus on suspicious time intervals round by round without pre-exhaustively listing all local time intervals and without always using the highest processing granularity for all original data. This improves the ability to identify short-term anomalies and abnormal intervals while controlling the amount of input data to the large model 200, and reduces the probability of the large model 200 exhibiting illusions.

[0063] Through the above process, in each round of data quality inference, the large model 200 determines whether it needs to acquire finer-grained local data based on the current granularity of time-series analysis data. It then progressively narrows the analysis time range and increases the data representation granularity through zoom requests until the final evaluation result is output. In a preferred embodiment, a maximum number of iterations L can be set. When the number of zoom iterations reaches L, the large model 200 is forced to output the final evaluation result based on the acquired time-series analysis data, preventing the iteration process from failing to converge.

[0064] Example 2, refer to Figure 2 This embodiment of the invention, based on Embodiment 1, further explains the parsing of zoom requests and the re-querying of original sampled data, the dynamic determination of resampling parameters based on data statistical characteristics, the preservation of inference context, and the termination control of zoom iteration. Specifically, this embodiment includes the following steps.

[0065] S201. Obtain the measurement point description information and the original sampling data within the first time range.

[0066] Identify the target measurement points for data quality evaluation and obtain their identifiers and descriptions. The identifier is used to uniquely index the sampled data of the target measurement point in the time-series database. The description includes one or more of the following: equipment object, physical meaning, engineering unit, range, and measurement point type. Measurement point types include analog measurement points, switch measurement points, status measurement points, or packaged points.

[0067] Following step S101 in Example 1, the target measurement point is obtained in the nearest h... st Short-period dense sampling data P within an hour st ={x1,x2,…,x n}, and process P according to step S102. stResampling is performed to generate the first time series analysis data 100, which is then input into the large model 200 for the first round of data quality inference.

[0068] S202, parse the zoom request and re-query the original sampling data.

[0069] When the inference output of the large model 200 is a zoom request, the XML structure of the zoom request is parsed, and the start time indicated by the start field and the end time indicated by the end field within the zoom tag are read to obtain the target time interval.

[0070] Based on the target measurement point's identification, start time, and end time, the original sampling data within the target time interval is retrieved again from the time series database storing the original time series data, denoted as P. zoom ={x1',x2',…,x m '}, where m is the number of original sampled data within the target time interval.

[0071] In this embodiment, the original sampling data is re-queried for zoom requests, rather than the first time-series analysis data 100 is processed again. This ensures that the data details on which the subsequent fine-grained analysis is based are not affected by the smoothing of the previous resampling process, and the position and amplitude information of abnormal jumps are completely preserved.

[0072] S203. Dynamically determine resampling parameters based on data statistical characteristics.

[0073] Before each resampling cycle, including the initial resampling and resampling after each zoom request response, statistical indicators reflecting the data characteristics are first calculated for the sampled data within the current time range. Let P... zoom For example, during the first round of resampling, P st Performing the same calculations, the statistical indicators include: firstly, the overall degree of data fluctuation, expressed as standard deviation σ. and mean absolute difference Δ Joint measurement. The formula for calculating the mean absolute difference is: , The mean absolute difference (MAD) characterizes the average variation between adjacent sampling points and is more sensitive to local abrupt changes than the standard deviation.

[0074] Secondly, the temporal correlation of the data is assessed using the first-order autocorrelation coefficient. The measurement and calculation formula is as follows: , in, This represents the mean of the current sampled data. A lower autocorrelation coefficient indicates stronger independence between adjacent data points and more frequent fluctuations, requiring finer granular data segmentation.

[0075] Third, the extreme value density γ of the data, that is, the number of samples that exceed the mean. The proportion of data points within ±c times the standard deviation is calculated using the following formula: , Where |{}| represents the number of sampled values ​​in the set that meet the conditions, and c is the preset extreme value judgment coefficient. The higher the extreme value density, the more candidate points of anomalies there are, requiring more refined sharding to retain the location information of these potential anomalies. Therefore, the shard size can be adjusted by adjusting c as needed.

[0076] Based on the above statistical indicators, the number of shards n in the current round is determined according to the following qualitative rules. rs Number of fragments n rs At the preset lower limit value n min With the preset upper limit value n max Values ​​between; n rs It increases with the increase of the mean absolute difference Δ, and with the increase of the first-order autocorrelation coefficient. The value decreases as the extreme value density γ increases, and increases as the extreme value density γ increases. That is: the more drastic the data fluctuation, the weaker the temporal correlation, and the more outlier candidate points, the larger the number of slices should be to retain more detailed features; the more stable the data, the smaller the number of slices should be to simplify the input of the large model 200.

[0077] In a preferred embodiment, the number of slices n rs Calculated using the following function: , Where, n min and n max These are the lower and upper limits for the number of fragments, respectively, where α is the normalized scaling factor, and Δ is the upper limit. max As a normalized baseline value, This indicates rounding down. When the calculated result is less than n... min When n min , higher than n max When n max Normalized baseline value Δ max The settings can be configured based on the range of the target measurement point or the fluctuation statistics of historical data. In one example, n min Take 20, n max Take 500, α is 1, Δ max Take 10.

[0078] According to the determined number of fragments n rs , for P zoom Execute the resampling process in step S102: P zoom Divide evenly into n rsThe data is divided into several data shards. The sampled value with the largest deviation from the shard mean within each shard is taken as the representative value. When multiple sampled values ​​have the same maximum absolute deviation, the earliest or latest sampled value can be selected as the representative value. The representative time points are recorded, and the shard standard deviation is calculated. These data are combined into a CSV format data table as the time series analysis data for the current round. The shard standard deviation is the standard deviation of each sampled value within the corresponding data shard relative to the shard mean.

[0079] In the zoom cycle, the final number of segments used should also meet the data processing granularity conditions specified in step S104, so that the time span covered by each segment in the current cycle is less than that in the previous cycle; if the number calculated by the formula does not meet this condition, the number of segments should be adjusted within the preset range.

[0080] It should be noted that after each zoom and focus, the data segments within the new target time interval have different statistical distributions. The system recalculates the aforementioned statistical indicators and redetermines the number of segments, thus adopting the resampling granularity that best matches the current data characteristics in each iteration. For example, when the zoomed data segment set contains anomalous jumps, its mean absolute difference and extreme value density increase significantly, and the number of segments automatically increases accordingly, preserving the complete local details of the anomalous features. When there is no original sampled data within the target time interval, the system prompts the large model 200 to adjust the target time interval or outputs the final evaluation result based on existing time series analysis data. When the number of original sampled data, m, is less than 2, the mean absolute difference and first-order autocorrelation coefficient are no longer calculated, and the original sampled data is directly retained as the time series analysis data for the current round. When the denominator of the first-order autocorrelation coefficient calculation formula is 0, the current sampled data is determined to be fluctuation-free data, and the first-order autocorrelation coefficient is set to 1.

[0081] S204. Preserve the reasoning context and input the large model 200 to continue reasoning.

[0082] When inputting the current round of time series analysis data into the large model 200, the inference context corresponding to the data quality inference of the previous round is retained, and the inference context and the current round of time series analysis data are combined to form a prompt word for input into the large model 200. The inference context includes at least one of the following: the measurement point description information input in the previous round, the time series analysis data input in the previous round, the output content of the large model 200 in the previous round, and the analysis time interval corresponding to the previous round.

[0083] By preserving the reasoning context, the large model 200 can understand the global data features already acquired in the previous round and the reasoning basis for generating zoom requests when performing fine-grained analysis. This avoids repeated semantic understanding of the same measurement point and enables multiple rounds of reasoning to form a coherent analysis chain, improving the accuracy of anomaly interval location.

[0084] In this embodiment, the resampling parameters are not kept fixed during each zoom cycle, but are recalculated based on the statistical characteristics of the original sampled data obtained by re-querying within the target time interval determined by the large model 200. Therefore, the selection of the target time interval by the large model 200 not only changes the scope of the next round of data analysis, but also further affects the number of slices and the granularity of data processing in the next round by changing the data set participating in the statistical calculation, so that the interval decision of the large model and the adjustment of the resampling parameters are linked.

[0085] S205, Termination control of zoom iteration.

[0086] The system counts the number of zoom requests output by the large model 200. When the number reaches a preset threshold L, it stops responding to new zoom requests and instead forces the large model 200 to output a final evaluation result based on the time-series analysis data obtained in each round, using a prompt word. Therefore, even if the large model 200 remains uncertain about suspicious features, the system can converge its output within a limited computational overhead, avoiding excessive inference resources consumed during iteration. In one example, L is set to 3 to 5 times.

[0087] S206. Analysis and output of the final evaluation results.

[0088] When the large model 200 outputs the final evaluation results, its XML structure is parsed, and the quality status of the `is_good` field, the reason for the judgment in the `reason` field, the anomaly type in the `fault_type` field, and the anomaly range in the `fault_range` field are extracted. These are then output and saved as the data quality evaluation results for the target measurement points. When `fault_type` is NULL, `fault_range` is empty, indicating that the data quality of the target measurement points is good. When `fault_type` is a bad point or a conversion anomaly, the start and end times in `fault_range` define the time interval corresponding to the anomaly.

[0089] Example 3, Reference Figure 3 This embodiment of the invention, based on Embodiment 1, further describes the rule-based pre-filtering process and case retrieval enhancement process before data quality inference in the large model 200. In this embodiment, data quality evaluation adopts a hierarchical processing architecture of "rules + cases + large model 200": rule detection quickly filters simple anomalies, case retrieval provides reference context for filtered measurement points, and the large model 200 only performs zoom evaluation on measurement points that require in-depth analysis, thereby reducing the overall computational overhead while ensuring evaluation accuracy. Specifically, this embodiment includes the following steps.

[0090] S301. Acquire long-period sparse sampling data and perform rule detection.

[0091] Before executing step S102 to generate the first time series analysis data 100, long-period sparse sampling data of the target measurement point within the second time range are obtained from the time series database, denoted as: , in, This represents long-period sparse sampling data; This represents the measurement value of the i-th long-cycle sampled data arranged in chronological order of sampling time; i represents the sampling data sequence number.

[0092] If we take Indicates the number of sampled data within the second time range, in Indicates the duration of the second time range; Indicates the duration of the first time range; This represents short-period, densely sampled data within the first time frame. In one example, Take 1080 hours, which is 45 days. Take 5000.

[0093] Based on P lt Based on the measurement point description information, the target measurement points are sequentially subjected to rule detection according to preset rules. The preset rules include at least one of the following: First, the dead-point detection rule: when the proportion of a continuous time interval in which the sampled data remains unchanged exceeds a preset proportion threshold, a dead-point anomaly is determined to exist. In one example, the preset proportion threshold is set to 60%.

[0094] Secondly, the data type detection rule: When the target measurement point is labeled as an analog measurement point, but the actual collected data exhibits discrete characteristics of switch quantities, status quantities, or packaged points, a data type error is determined to exist. In one example, the discrete characteristics are determined by the following combined conditions: the proportion of unique values ​​is less than a preset proportion threshold, all sampled values ​​are integers, and the absolute value of the maximum value among the sampled values ​​is less than a preset numerical threshold; the preset proportion threshold is, for example, 10%, and the preset numerical threshold is, for example, 64. The proportion of unique values ​​refers to the ratio of the number of deduplicated sampled values ​​to the total number of sampled data.

[0095] Third, the description missing detection rule: when the description field of the measurement point description information is empty or lacks valid content, it is determined that there is a description missing anomaly.

[0096] If any preset rule is hit during rule detection, i.e., a preset type of data quality anomaly is detected, the evaluation result corresponding to the preset type of data quality anomaly is directly output, including the anomaly type, the cause of the anomaly, and the time interval of the anomaly, and subsequent processing is terminated, and the large model 200 is no longer called for data quality inference; if no preset rule is hit during rule detection, the subsequent steps are executed.

[0097] Because the above rules are based on P lt It determines the long-term trend, with simple and clear logic and minimal computational overhead. It can quickly filter simple anomalies with low computational overhead, allowing a large number of measurement points that do not require in-depth analysis to exit the processing flow in advance. The large model 200 only calls the measurement points that cannot be determined by rule detection.

[0098] S302, Enhanced Case Search.

[0099] For target measurement points that pass the rule detection, before inputting the measurement point description information and the first time series analysis data 100 into the large model 200 in step S103, historical samples with semantic similarity to the target measurement point are retrieved from the historical sample library to provide a reference context for data quality inference of the large model 200. Specifically, this includes: inputting the description text corresponding to the measurement point description information of the target measurement point into the semantic embedding model to obtain the target embedding vector q; inputting the measurement point description text of the j-th historical sample in the historical sample library into the same semantic embedding model to obtain the historical sample embedding vector sj, and finally obtaining the complete corresponding embedding vectors q and sj. Here, the historical samples include the measurement point description information of the historical measurement points, historical time series data fragments, and corresponding historical evaluation results; the semantic similarity between the target measurement point and each historical sample is measured by cosine similarity, and the calculation formula is: , Where q represents the embedding vector corresponding to the measurement point description text of the target measurement point, and s j Let represent the embedding vector corresponding to the measurement point description text of the j-th historical sample, where j represents the historical sample number. This represents the dot product of two vectors. and These represent the magnitudes of the two vectors, respectively. The greater the cosine similarity, the closer the target measurement point is to the corresponding historical sample in terms of equipment object, measured physical quantity, or data semantics.

[0100] The top k most similar historical samples are selected as reference cases. The historical time series data fragments of the reference cases and the corresponding historical evaluation results are combined with the measurement point description information, the definition of anomaly rules, the first time series analysis data 100 and the output format constraints to form prompt words, which are then input into the large model 200 for data quality inference.

[0101] Enhanced by case retrieval, the large model 200 can refer to historical judgments of similar equipment and physical quantities to make analogical reasoning on suspicious features of the current measurement point, thereby improving the accuracy of identifying complex anomalies such as bad points and conversion anomalies, while reducing misjudgments caused by semantic understanding bias of the measurement point.

[0102] S303, Manual verification and historical sample database update.

[0103] After outputting the final evaluation results of the target measurement points, the final evaluation results and the corresponding original time series data change trends are provided to the human for confirmation. When the final evaluation results are confirmed to be correct, the measurement point description information, abnormal data fragments, abnormality types, abnormality causes, and human confirmation results of the target measurement points are saved to the historical sample database as historical samples that can be retrieved in subsequent measurement point data quality evaluations.

[0104] Through manual verification of the closed loop and the continuous accumulation of the historical sample database, the coverage of the historical sample database expands with the use of the system, and the reference quality of the enhanced case retrieval is improved accordingly, enabling the system's evaluation capabilities to have a mechanism for continuous optimization.

[0105] Example 4, Reference Figure 9 As an embodiment of the present invention, an industrial time-series data quality evaluation system based on large model adaptive zoom is provided, including a data processing module 400, a large model inference module 500, and a zoom control module 600.

[0106] During system operation, the data processing module 400 first retrieves the corresponding measurement point description information from the measurement point information database based on the measurement point identifier of the target measurement point to be evaluated, and then retrieves the original time series data of the target measurement point within the first time range from the time series database. The measurement point description information may include at least one of the following: equipment object, physical meaning, engineering unit, range, and measurement point type.

[0107] The data processing module 400 segments, resamples, or extracts features from the original time-series data according to a preset data processing method, generating first time-series analysis data 100. The first time-series analysis data 100 includes multiple data units arranged in chronological order and the time information corresponding to each data unit, so as to characterize the overall change trend of the time-series data within the first time range while controlling the amount of data input.

[0108] Subsequently, the data processing module 400 sends the measurement point description information and the first time series analysis data 100 to the large model inference module 500. The large model inference module 500 inputs the above information into the large model 200, enabling the large model 200 to perform data quality inference by combining the semantic information of the target measurement point and the change characteristics of the time series data.

[0109] When the large model 200 is able to complete the data quality judgment based on the current input information, the large model inference module 500 outputs the final evaluation result. The final evaluation result may include the data quality status of the target measurement point, the anomaly type, the evaluation reason, and the anomaly time interval.

[0110] When the large model 200 detects suspicious changes in the first time-series analysis data 100, but the current data processing granularity is insufficient to form a final judgment, the large model inference module 500 outputs a zoom request containing the target time interval. The target time interval is located within the first time range but is smaller than the first time range.

[0111] The zoom control module 600 receives and parses the zoom request, obtains the start and end times of the target time interval, and controls the data processing module 400 to obtain the original sampling data located within the target time interval from the original time series data or the time series database.

[0112] The data processing module 400 processes the raw sampled data within the target time interval with a finer granularity than when generating the first time series analysis data 100, generating the second time series analysis data 300. For example, the data processing module 400 can reduce the time span corresponding to each data slice, increase the number of data units per unit time, or increase the retention ratio of the raw sampled data within the target time interval, thereby allowing the second time series analysis data 300 to retain more local data details.

[0113] The zoom control module 600 controls the large model inference module 500 to input the second time-series analysis data 300 into the large model 200, enabling the large model 200 to continue data quality inference for the target time interval. In one embodiment, the input content, output content, or analysis time interval of the previous round of data quality inference can also be used as the inference context and input into the large model 200 together with the second time-series analysis data 300.

[0114] If the large model 200 can determine the data quality of the target measurement point based on the second time series analysis data 300, it will output the final evaluation result; if the large model 200 still believes that the current data is insufficient to complete the judgment, it will output a zoom request containing a smaller target time interval again.

[0115] Based on the zoom request output again, the zoom control module 600 repeatedly controls the data processing module 400 to acquire the raw sampling data within the corresponding target time interval and generate time series analysis data with finer data processing granularity. At the same time, it controls the large model inference module 500 to continue to perform data quality inference until the large model 200 outputs the final evaluation result.

[0116] For example, for bearing temperature measurement points, the first time-series analysis data 100 covers data from the most recent 24 hours. The large model 200 detects abnormally low values ​​in certain time periods, but the current data granularity is insufficient to determine the duration and change process of these abnormally low values. Therefore, it outputs a zoom request for the suspected time period. Based on this, the zoom control module 600 controls the data processing module 400 to reacquire densely sampled data within that time period and generate fine-grained second time-series analysis data 300. The large model 200 further determines whether the abnormally low values ​​conform to the physical change law of bearing temperature based on the second time-series analysis data 300 and outputs the corresponding data quality evaluation results.

[0117] Therefore, this embodiment achieves adaptive zoom of time-series data driven by the inference results of the large model 200 through the coordinated operation of the data processing module 400, the large model inference module 500 and the zoom control module 600, enabling the system to gradually focus from coarse-grained analysis over a large time range to fine-grained analysis over suspicious time intervals.

[0118] Example 5, refer to Figures 4 to 7 As an embodiment of the present invention, four industrial measurement points are used as experimental objects to verify the aforementioned industrial time series data quality evaluation method based on large model 200 adaptive zoom.

[0119] The experiment selected four measurement points, corresponding to four data quality types: normal, dead, bad, and type error. For each measurement point, short-period dense sampling data and long-period sparse sampling data were acquired.

[0120] Short-period dense sampling data P st Take the nearest h at the corresponding measuring point st =n within 24 hours st =5000 sampling points are used to characterize local changes and anomaly details in the measurement point data; long-period sparse sampling data P lt Take the nearest h at the corresponding measuring point lt =n within 1080 hours, or 45 days lt =5000 sampling points were used to characterize the long-term changing trend of the measurement data. The data collection results of the four measurement points are as follows: Figures 4 to 7 As shown.

[0121] Following the method of step S301 in Example 3, based on P lt Rule detection is performed on the four measuring points along with the measuring point description information. (Reference) Figures 4 to 8Point 1 failed to meet the dead point detection, data type detection, and missing description detection rules, and proceeded to the subsequent large model 200 data quality inference. To illustrate the execution process of the dead point detection rules, point 2 was further illustrated. 5000 sampling time points were evenly distributed over a 45-day analysis period, with the first 3519 sampled values ​​remaining continuously at 0, while the remaining sampled values ​​exhibited non-constant variations. The duration was determined by the difference between the first and last sampling times of the constant value interval. This interval accounted for approximately 70.37% of the analysis time range, exceeding the preset threshold of 60%. Therefore, point 2 was determined to have a dead point anomaly. Point 3 did not match any rules and proceeded to the subsequent large model 200 data quality inference. Point 4 was configured as an analog quantity, but the actual sampled data was always integer, with a unique value ratio of less than 10% and the absolute value of the maximum value being less than 64. This matched the data type detection rule and was determined to be a data type error anomaly. The evaluation reason was "the measurement point may be a switch quantity or a packaged point, but the actual configuration is an analog quantity; there are 7 unique values, namely 2.0, 5.0, 4.0, 3.0, 8.0, 7.0, and 6.0".

[0122] Therefore, points 2 and 4 directly output evaluation results after rule detection, without calling the large model 200; points 1 and 3 enter the large model 200 analysis stage.

[0123] Following step S302 in Example 3, based on the measurement point names and descriptions of points 1 and 3, semantic embedding vectors are calculated and cosine similarity is calculated with samples in the historical sample library to retrieve the historical sample with the highest similarity. Specifically, for point 1 (HLDK_AI000002), the retrieval of the generator speed at Haladaokou A01 yields the case of normal generator speed (HLDK_AI000025) and the case of abnormal generator speed (HLDK_AI000324) at Haladaokou A15; for point 3 (WRZS-FJ-1@0025), the retrieval of the temperature of bearing 1 of motor 1 (for fan 1#) yields the case of normal temperature (WRZS-FJ-2@0089) of bearing 1 of motor 2 (for fan 2#) and the case of abnormal temperature (WRZS-FJ-1@0035) of bearing 1 of motor 2 (for fan 1#).

[0124] Following the method of step S203 in Example 2, n is set min =20, n max =500, α=1, Δ max =10, calculate P for points 1 and 3 respectively. st Statistical indicators and determine the number of dynamic fragments n rs The calculation result for point 1 is σ. =14.38, α is 1. =9.74, =0.4838, γ=0.0514, finally determine n rs =273; the calculation result for point 3 is σ =3.2, α is 1. =0.05, =0.9468, γ=0.0054, the calculated number of slices is lower than the lower limit n. min Therefore, we take n. rs =20.

[0125] As can be seen, the generator speed data at point 1 fluctuates greatly and has weak time correlation, so the number of dynamic segments is increased accordingly to retain local fluctuations and equipment start-up and shutdown changes in the speed data; the bearing temperature data at point 3 is generally stable and has a high autocorrelation coefficient, so the number of segments approaches the lower limit, and at the same time, the lower limit constraint ensures that the time series analysis data after resampling contains sufficient data units.

[0126] According to the determined n rs P for points 1 and 3 respectively st Perform resampling: P st Divide evenly into n rs The data is divided into several parts, and the sampled value with the largest deviation from the mean of each part is taken as the representative value. Record the representative time point t rs i And calculate the standard deviation σ of each piece rs i Sometimes, for ease of calculation and presentation, variance can be used instead. When needed, the square root of the variance of each piecewise segment is taken to obtain the corresponding standard deviation σ of that segment. rs i The aforementioned data were combined into a first time-series analysis dataset 100 in CSV format. Part of the first time-series analysis dataset 100 for point 1 is shown in Table 1: Table 1: First Time Series Analysis Data for Point 1 (Table 100)

[0127] The first time series analysis data 100 for point 3 is shown in Table 2: Table 2: First Time Series Analysis Data for Point 3 (Table 100)

[0128] The piecewise variances listed in Tables 1 and 2 are intermediate statistical results; when constructing the CSV data for the input large model, the square root of the piecewise variance is used to obtain the corresponding piecewise standard deviation σ. rs i .

[0129] The results above show that the amount of data input to the large model 200 after resampling is significantly reduced compared to the original 5000 sampling points. At the same time, the suspicious abnormal data with a value of 0.0 in point 3 and its corresponding time information are still retained.

[0130] The measurement point description information of point 1, the definition of anomaly rules, the first time series analysis data 100, the retrieved historical samples, and the output format constraints are all input into the large model 200. The large model 200 analyzes the physical meaning of point 1 as a generator speed measurement point and outputs the final evaluation result: is_good is true, fault_type is NULL, and the reasons for the judgment include: "The historical data fluctuation range is within a reasonable range, with no obvious sudden changes or outliers; the overall data shows a downward trend, consistent with equipment shutdown or low load conditions; no conversion errors or sensor anomalies were found." Point 1 is determined to be a normal measurement point, and zoom analysis is not required.

[0131] The measurement point description information of point 3, the anomaly rule definition, the first time series analysis data 100, the retrieved historical samples, and the output format constraints are all input into the large model 200. The large model 200 finds that the representative value of some data units is 0.0, which is significantly different from the data approximately 35℃ to 37℃ before and after. However, the first time series analysis data 100 covers 24 hours, and the current data processing granularity is insufficient to determine whether the 0.0 value belongs to a continuous bad point and its specific anomaly range. Therefore, a zoom request is output: <zoom> <start> 2026-03-30 07:12:05< / start> <end> 2026-03-30 10:48:07< / end> < / zoom> After parsing the zoom request, based on the measurement point identifier of point 3 and the start and end times in the zoom request, the original sampling data within the range of 07:12:05 to 10:48:07 on March 30, 2026, is re-queried from the time series database. Following the method of step S203 in Example 2, the statistical indicators of the re-queried original sampling data are recalculated and the number of slices in the current round is determined. Then, resampling processing is performed to generate second time series analysis data 300 with a data processing granularity finer than the first time series analysis data 100.

[0132] Subsequently, the measurement point description information from the first round of data quality inference, the first time-series analysis data 100, the zoom request output by the large model 200, and the corresponding analysis time interval are retained as the inference context. The second time-series analysis data 300 and the inference context are then input into the large model 200 to continue inference. The evaluation results output in this round of inference are: is_good is false, fault_type is bad, and fault_range is from 2026-03-30 09:17:23 to 2026-03-30 09:46:29.

[0133] In the above process, the target time interval is not predetermined by a preset sliding window or manual operation, but is actively output by the large model 200 based on the suspicious low-value features in the first time series analysis data 100. The system converts the time interval output by the large model 200 into the original sampled data for re-query and fine-grained time series analysis data generation operation, and feeds the generation result back to the large model 200 for further inference, thereby realizing an adaptive zoom closed loop driven by the inference result of the large model to focus the data.

[0134] Experimental results of this implementation show that rule-based detection can directly identify dead-point anomalies and data type error anomalies, enabling simple anomalies to be quickly filtered with low computational overhead. For dead-point anomalies that are difficult to identify by rule-based detection, the large model 200 can combine measurement point description information and historical samples to find suspicious time intervals, and drive the system to re-query the original sampled data by outputting zoom requests. After improving the granularity of data processing, dead-point identification and anomaly interval location are completed. The adaptive adjustment of the number of dynamic slices according to data characteristics, with 273 for point 1 and 20 for point 3, verifies the matching relationship between resampling parameters and data fluctuation, time correlation, and extreme value density. This verifies the feasibility of the industrial time-series data quality evaluation method based on large model 200 adaptive zoom.

[0135] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the industrial time-series data quality evaluation method based on large model adaptive zoom proposed in the above embodiment.

[0136] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the industrial time-series data quality evaluation method based on large model adaptive zoom proposed in the above embodiment.

[0137] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0139] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0140] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for evaluating the quality of industrial time-series data based on large-model adaptive zoom, characterized in that, include: Obtain the target measurement point description information and the original time series data of the target measurement point within a first time range, and generate the first time series analysis data based on the original time series data (100). The measurement point description information and the first time series analysis data (100) are input into the large model (200) for data quality inference, so that the large model (200) outputs the final evaluation result or a zoom request containing the target time interval; In response to the zoom request, second time series analysis data (300) within the target time interval is generated based on the original time series data, wherein the target time interval is smaller than the first time range, and the data processing granularity of the second time series analysis data (300) is finer than the data processing granularity of the first time series analysis data (100). The second time series analysis data (300) is input into the large model (200) to continue data quality inference. When the large model (200) outputs a zoom request again, the time series analysis data within the corresponding target time interval is generated based on the zoom request and the data quality inference continues until the large model (200) outputs the final evaluation result.

2. The industrial time-series data quality assessment method based on large-model adaptive zoom as described in claim 1, characterized in that: Both the first time series analysis data (100) and the second time series analysis data (300) include multiple data units arranged in chronological order and time information corresponding to each data unit; Each of the data units is generated based on the original sampled data within the corresponding time slice, and the time information includes the time point or time interval of the corresponding data unit. The data processing granularity of the second time-series analysis data (300) is finer than that of the first time-series analysis data (100), including at least one of the following: Within the same time frame, the second time series analysis data (300) includes more data units than the first time series analysis data (100). The time span corresponding to each data unit in the second time series analysis data (300) is smaller than the time span corresponding to each data unit in the first time series analysis data (100); The amount of original sampled data retained by the second time series analysis data (300) per unit time is greater than the amount of original sampled data retained by the first time series analysis data (100) per unit time.

3. The industrial time-series data quality assessment method based on large-model adaptive zoom as described in claim 1 or 2, characterized in that: In response to the zoom request, second time-series analysis data (300) within the target time interval is generated based on the original time-series data, including: Based on the measurement point identifier of the target measurement point, the start time and end time of the target time interval, the original sampling data within the target time interval is re-queried from the time series database storing the original time series data; The second time-series analysis data (300) is generated based on the re-queried original sampled data.

4. The industrial time-series data quality evaluation method based on large-model adaptive zoom as described in claim 3, characterized in that: When generating the time series analysis data for each round, the data statistical characteristics are calculated based on the original sampled data within the corresponding time range, and the resampling parameters for the current round are determined according to the data statistical characteristics. The time series analysis data for the current round is then generated according to the resampling parameters. The statistical characteristics of the data include at least one of standard deviation, mean absolute difference, first-order autocorrelation coefficient and extreme value density, and the resampling parameters include at least the number of slices. The resampling parameters for the current round are determined based on the statistical characteristics of the data, including: The number of fragments takes a value between a preset lower limit and a preset upper limit; The number of segments increases with the increase of the mean absolute difference, decreases with the increase of the first-order autocorrelation coefficient, and increases with the increase of the extreme value density.

5. The industrial time-series data quality assessment method based on large-model adaptive zoom as described in claim 1, characterized in that: The second time-series analysis data (300) is input into the large model (200) to continue data quality inference, including: Retain the reasoning context corresponding to the previous round of data quality reasoning; The inference context and the second time series analysis data (300) are input into the large model (200); The reasoning context includes at least one of the following: the measurement point description information input in the previous round, the time series analysis data input in the previous round, the output content of the large model (200) in the previous round, and the analysis time interval corresponding to the previous round.

6. The method for evaluating the quality of industrial time-series data based on large-model adaptive zoom as described in any one of claims 1, 2, 4, and 5, characterized in that: Before generating the first time series analysis data (100) based on the original time series data, the method further includes: Acquire sparse sampling data of the target measurement point within a second time range; Wherein, the second time range is greater than the first time range, and the sampling interval of the sparse sampled data is greater than the sampling interval of the original time series data; Based on the sparse sampling data, the target measurement point is subjected to rule detection according to preset rules. The preset rules include at least dead point detection rules for detecting data that remains unchanged for a long time. In response to the rule detection detecting a preset type of data quality anomaly, the evaluation result corresponding to the preset type of data quality anomaly is directly output, and subsequent data quality inference is terminated. In response to the rule detection not detecting the preset type of data quality anomaly, the step of generating first time series analysis data (100) based on the original time series data is executed.

7. The industrial time-series data quality assessment method based on large-model adaptive zoom as described in claim 6, characterized in that: The step of inputting the measurement point description information and the first time series analysis data (100) into the large model (200) for data quality inference includes: Calculate the semantic embedding vector of the measurement point description information; Based on the semantic embedding vector, a preset number of historical samples with the highest semantic similarity to the target measurement point are retrieved from the historical sample library. The historical samples include measurement point description information of the historical measurement point, historical time series data fragments, and corresponding historical evaluation results. The retrieved historical samples, the measurement point description information, and the first time series analysis data (100) are input into the large model (200) for data quality inference.

8. An industrial time-series data quality assessment system based on large-model adaptive zoom, employing the industrial time-series data quality assessment method based on large-model adaptive zoom as described in any one of claims 1 to 7, characterized in that: include: Data processing module (400), large model inference module (500) and zoom control module (600); The data processing module (400) is used to acquire the measurement point description information of the target measurement point and the original time series data of the target measurement point within a first time range, and generate the first time series analysis data (100) based on the original time series data. The large model inference module (500) is used to input the measurement point description information and the first time series analysis data (100) into the large model (200) to perform data quality inference, so that the large model (200) outputs the final evaluation result or a zoom request containing the target time interval; The zoom control module (600) is used to respond to the zoom request by controlling the data processing module (400) to generate second time series analysis data (300) within the target time interval based on the original time series data, and controlling the large model inference module (500) to input the second time series analysis data (300) into the large model (200) to continue data quality inference, wherein the target time interval is smaller than the first time range, and the data processing granularity of the second time series analysis data (300) is finer than the data processing granularity of the first time series analysis data (100); The zoom control module (600) is also used to control the data processing module (400) to generate time series analysis data within the corresponding target time interval based on the zoom request output again when the large model (200) outputs a zoom request again, and to control the large model inference module (500) to continue to perform data quality inference until the large model (200) outputs the final evaluation result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the industrial time-series data quality evaluation method based on large model adaptive zoom as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the industrial time-series data quality evaluation method based on large model adaptive zoom as described in any one of claims 1 to 7.

Images