Production anomaly prediction method and system based on large model and MES system

Abstract

The invention relates to the field of data processing, in particular to a production anomaly prediction method and system based on a large model and an MES (Manufacturing Execution System), and the method comprises the steps: training a prediction model based on operation data sets of each batch of products produced by production equipment in history, each operation data set comprising an operation data sequence of each dimension of the production equipment; and obtaining a current input window of the operation data of each dimension when the production equipment produces a current batch of products, inputting the current input window of the operation data of each dimension based on the trained prediction model, outputting a prediction quality deviation index corresponding to the current input window, and when the prediction quality deviation index is greater than a preset threshold value, the MES system gives an abnormal alarm. According to the method, the optimal input window is determined through data driving, it is ensured that the data of the input model has the most information content in the time sequence length, and an optimal data basis is provided for building a high-performance prediction model.

Description

Technical Field

[0001] This invention relates to the field of data processing. In particular, it relates to a method and system for predicting production anomalies based on a large model and a MES system. Background Technology

[0002] In the precision manufacturing processes of automotive parts, the health of production equipment directly determines the stability of product quality. To achieve timely early warning of production anomalies, predictive models based on equipment operating data (such as torque, vibration, and temperature) are typically used to predict the risk of product quality deviations. The Manufacturing Execution System (MES), as the core of production management, is responsible for collecting this real-time data and issuing early warnings.

[0003] Currently, when training predictive models using equipment runtime sequence data, a fixed-length sliding window is commonly used to capture the input data. However, the operating status of production equipment is dynamic, and the effective data length corresponding to its abnormal development process (such as progressive tool wear) is not constant. A fixed window cannot adaptively capture abnormal features at different stages and with varying severity. An excessively long window introduces a large amount of normal data unrelated to the current anomaly, diluting key anomaly signals; a window that is too short may fail to cover a complete anomaly development cycle, preventing the model from learning effective anomaly patterns. Therefore, the predictive models used in existing production anomaly prediction methods have low accuracy and weak generalization ability. Summary of the Invention

[0004] To address the technical problems of low accuracy and weak generalization ability of the prediction models used in the existing production anomaly prediction methods, this invention provides solutions in the following aspects.

[0005] In the first aspect, a production anomaly prediction method based on a large model and MES system includes: The prediction model is trained based on the historical production batches of products from the production equipment. Each production dataset includes a sequence of production data for each dimension of the production equipment. The system obtains the current input window of the operating data of each dimension when the production equipment is producing the current batch of products. Based on the trained prediction model, it inputs the current input window of the operating data of each dimension and outputs the prediction quality offset index corresponding to the current input window. When the prediction quality offset index is greater than the preset threshold, the MES system issues an abnormal alarm. The training of the prediction model based on the historical production data of each batch of products from the production equipment includes: for a single production dataset, calculating the anomaly value of each production data in a single-dimensional production data sequence; determining the starting point of the candidate window based on the anomaly value and using the end of the production data sequence as the ending point to form several candidate windows; calculating the volatility index of each candidate window based on the anomaly value of the production data, and selecting the candidate window with the largest volatility index as the optimal candidate window for that dimension; selecting the maximum window length among the optimal candidate windows of all dimensions as the input length; and using the input length and the end of the production data sequence for each dimension as the end of the input window to extract the input windows for each dimension to train the prediction model, and outputting the prediction quality offset index corresponding to the production dataset.

[0006] Preferably, determining the starting point of the candidate window based on the anomaly degree value includes: calculating the average anomaly degree value of all running data in the running data sequence; and taking the running data in the running data sequence whose anomaly degree value is greater than the average value as the starting point of the candidate window.

[0007] Preferably, the method for calculating the anomaly value of each running data in the running data sequence includes: calculating the mean and standard deviation of all running data in the running data sequence; for a single running data, calculating the absolute value of the difference between the running data and the mean, and using the ratio of the absolute value to the standard deviation as a first ratio; calculating the average of the first derivatives of all running data within a neighborhood of a specified radius centered on the running data, and using the ratio of the first derivative of the running data to the average of the aforementioned first derivatives as a second ratio; and calculating the product of the absolute values ​​of the first ratio and the second ratio to obtain the anomaly value of the running data.

[0008] Preferably, the calculation of the volatility index of each candidate window based on the anomaly value of the running data includes: for a single candidate window, calculating the ratio of the maximum anomaly value of the running data within the candidate window to the maximum anomaly value of the running data in the running data sequence corresponding to the candidate window, as a third ratio; calculating the ratio of the standard deviation of the anomaly value of the running data within the candidate window to the standard deviation of the anomaly value of the running data in the running data sequence corresponding to the candidate window, as a fourth ratio; and multiplying the third ratio and the fourth ratio as the volatility index of the candidate window.

[0009] Preferably, after training the prediction model, the method further includes: calculating the influence weight of each dimension based on the fluctuation index of the optimal candidate window in the same dimension of all running datasets and the error between the predicted quality offset index and the true quality offset index of all running datasets.

[0010] Preferably, the method for calculating the influence weight of each dimension includes: for a single dimension, obtaining the volatility index of the optimal candidate window for that dimension in all running datasets, and arranging them in the order of the collection time of the running datasets to generate a volatility index sequence; obtaining the error between the predicted quality offset index and the true quality offset index of all running datasets, and arranging them in the order of the collection time of the running datasets to generate an error sequence; calculating the standard deviation of the volatility index sequence and the covariance between the volatility index sequence and the error sequence, calculating the ratio of the aforementioned covariance to the standard deviation, and performing negative correlation normalization on the ratio to obtain the influence weight of that dimension.

[0011] Preferably, the method for obtaining the current input window includes: setting a basic window length; obtaining the previous input window and the error between the corresponding predicted quality offset index and the actual quality offset index of the operating data sequence of each dimension when the production equipment produces the current batch of products; calculating the fluctuation index of the previous input window of each dimension; and, based on the influence weight of each dimension, weighted summing of the fluctuation indices of the previous input windows of all dimensions to obtain a sum value; determining a reference error based on the error sequence; calculating the ratio of the reference error to the error corresponding to the previous input window; calculating the product of the basic window length, the aforementioned sum value and ratio to obtain the length of the current input window; and extracting the current input window based on the length of the current input window.

[0012] Preferably, the method for setting the basic window length includes: performing negative correlation processing on the error between the predicted quality offset index and the true quality offset index of each running dataset to obtain the weight coefficient of each running dataset; and performing a weighted average of the input window lengths of all running datasets based on each weight coefficient to obtain the basic window length.

[0013] Secondly, a production anomaly prediction system based on a large model and a MES system includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned production anomaly prediction method based on a large model and a MES system is implemented.

[0014] The present invention has the following effects: 1. This invention calculates the degree of anomaly of data points and determines the optimal window accordingly. It can automatically and accurately extract the data segments most relevant to product quality deviations from the long original equipment operation data sequence as training input. It effectively eliminates the interference of a large amount of stable and normal redundant data on the model, realizes the purification of training data, and enables the model to focus on learning the truly key abnormal feature patterns.

[0015] 2. Since the prediction model is trained based on the high-value input window that best represents the abnormal state, the features it learns have a stronger and more direct correlation with the product quality deviation, laying a foundation for high accuracy from the source of model training. At the same time, the method adaptively determines the length of the input window applicable to different historical segments, enabling the model to learn the abnormal patterns that appear at different time scales, thereby significantly improving its generalization ability for unknown data.

[0016] 3. This invention fundamentally solves the problem of difficulty in selecting a fixed window length, avoiding feature dilution (introducing too much noise) caused by an excessively long window, and feature incompleteness (inability to capture complete abnormal cycles) caused by an excessively short window. This data-driven optimal input window determination mechanism ensures that the data input to the model is the most informative in terms of time series length, providing the optimal data foundation for building high-performance prediction models. Attached Figure Description

[0017] Figure 1 This is a flowchart of steps S1-S3 in a production anomaly prediction method based on a large model and MES system according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of steps S10-S12 in a production anomaly prediction method based on a large model and MES system according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] Reference Figure 1 A production anomaly prediction method based on a large model and MES system includes steps S1-S3, as follows: S1: Train the prediction model based on the historical production batches of products from the production equipment. Each production dataset includes a sequence of production data for each dimension of the production equipment.

[0022] Before training the predictive model, training data needs to be acquired. The lifecycle of the production equipment is considered as the period from the replacement of the cutting head to the equipment's shutdown for maintenance. The training dataset consists of the operational dataset for each batch of products produced within this lifecycle. Each batch of products' operational dataset includes a sequence of operational data across various dimensions of the production equipment.

[0023] In a scenario using automotive parts manufacturing equipment as an example, the quality deviation of the produced parts is often related to multi-dimensional operating data of the equipment (such as torque, vibration, temperature, etc.), and the degree of influence of each dimension on the quality deviation varies. Therefore, it is necessary to perform collaborative analysis of these multi-dimensional data to determine the optimal input window for prediction.

[0024] Reference Figure 2 The training of the prediction model based on the historical production data of each batch of products from the production equipment includes steps S10-S12, as follows: S10: For a single running dataset, calculate the anomaly degree value of each running data in the single-dimensional running data sequence, determine the starting point of the candidate window based on the anomaly degree value, and use the end of the running data sequence as the ending point to form several candidate windows.

[0025] The method for calculating the anomaly score of each data point in the operational data sequence includes: calculating the mean and standard deviation of all operational data points in the sequence; for a single data point, calculating the absolute value of the difference between the data point and the mean, and using the ratio of this absolute value to the standard deviation as the first ratio; calculating the average of the first derivatives of all operational data points within a specified radius centered on the data point, and using the ratio of the first derivative of the data point to the average of the aforementioned first derivatives as the second ratio; and calculating the product of the absolute values ​​of the first and second ratios to obtain the anomaly score of the operational data point. This is specifically expressed by the following formula: In the formula, Indicates the first The first running dataset The first dimension of the running data sequence The degree of anomaly in the running data; Indicates the first The first running dataset The first dimension of the running data sequence One set of running data; Indicates the first The first running dataset The average value of all running data in a multi-dimensional running data sequence; Indicates the first The first running dataset The standard deviation of all running data in a multi-dimensional running data sequence; Indicates the first The first running dataset The first dimension of the running data sequence The first derivative of each running data; Indicates the first Centered on operational data, and its neighborhood The average of the first derivatives of all running data is used. The neighborhood radius is set to 2 units. The neighborhood radius is a hyperparameter and can be adjusted based on empirical values.

[0026] Reflects the first The degree of outlier of a running data point relative to the overall data indicates that the running data deviates further from the normal range. However, it may also be noise, and needs to be evaluated in conjunction with the continuity of outliers. Reflects the first The outlier continuity of the data from the first running data point measures the... The consistency between the trend of the data and its surrounding local trends; if this value is close to 1, it indicates that the trend of the data is consistent with the trend of the surrounding local trends. If the changes in the data show a certain continuity within the neighborhood, it is more likely to be a genuine anomaly; if the value deviates significantly from 1, it indicates that the... The change in the operational data is inconsistent with the surrounding trend, which is an isolated point of mutation and is more likely to be noise.

[0027] Abnormality value The larger the value, the higher the value. The more a piece of operational data deviates significantly from the normal range and the more persistent its abnormal changes are, the greater the likelihood that it is a real anomaly and has a potential impact on product quality.

[0028] Determining the starting point of the candidate window based on the anomaly severity value includes: calculating the average anomaly severity value of all running data in the running data sequence, and taking the running data in the running data sequence whose anomaly severity value is greater than the average value as the starting point of the candidate window.

[0029] S11: Calculate the volatility index of each candidate window based on the abnormality value of the running data, and take the candidate window with the largest volatility index as the optimal candidate window for that dimension.

[0030] The volatility index for each candidate window is calculated based on the anomaly severity values ​​of the running data. This involves: for a single candidate window, calculating the ratio of the maximum anomaly severity value of the running data within that window to the maximum anomaly severity value of the running data in the corresponding running data sequence, as the third ratio; calculating the ratio of the standard deviation of the anomaly severity values ​​within that window to the standard deviation of the anomaly severity values ​​of the running data in the corresponding running data sequence, as the fourth ratio; and multiplying the third and fourth ratios to obtain the volatility index for that candidate window. The specific formula is as follows: In the formula, Indicates the first The first running dataset The window length for each dimension is The volatility index of the candidate window; Indicates the first The first running dataset The window length for each dimension is The maximum anomaly value of all running data in the candidate window; Indicates the first The first running dataset The maximum anomaly value of all running data in the running data sequence of each dimension; Indicates the first The first running dataset The window length for each dimension is The standard deviation of the outlier values ​​of all running data in the candidate window; Indicates the first The first running dataset The standard deviation of the anomaly values ​​of all running data in the running data sequence of each dimension.

[0031] The window length was measured as The ratio of the maximum anomaly value of all running data in the candidate window to the maximum anomaly value of all running data in the entire running data sequence is used to determine the quality of the product. The closer this ratio is to 1, the more significant the anomaly points in the entire running data sequence are contained in the candidate window, and the greater the potential impact on product quality.

[0032] The window length was measured as The ratio of the dispersion of the degree of anomaly of all running data in the candidate window to the dispersion of the degree of anomaly of all running data in the entire running data sequence. A higher ratio indicates that the running data in the candidate window is unstable, fluctuates wildly, and has a complex state. This high degree of disorder is usually a manifestation of unstable equipment operation and also increases the risk of deviation in product quality.

[0033] Selecting the candidate window with the largest volatility index as the optimal candidate window for this dimension means that the selected window simultaneously meets the two key conditions of containing significant anomalies and being in a chaotic state, thus most effectively reflecting the potential impact of the operating data of this dimension on product quality deviation.

[0034] S12: Select the maximum window length of the optimal candidate window in all dimensions as the input length. Based on the input length, use the end of the running data sequence in each dimension as the end of the input window, extract the input window of each dimension to train the prediction model, and output the prediction quality offset index corresponding to the running dataset.

[0035] After obtaining the optimal candidate window for each dimension, the maximum window length among all optimal candidate windows is selected as the input window length. Based on the input window length, and using the end of the running data sequence for each dimension as the end of the input window, the input window for each dimension is extracted, thus obtaining the input dataset for a single running dataset.

[0036] The Quality Deviation Index (QDII) is a comprehensive metric used to quantitatively assess the degree to which the final output of a production batch deviates from its preset quality standards. It obtains the true QDII for each batch of products produced within a single lifecycle of the production equipment.

[0037] The input dataset of a single batch of products produced by the production equipment is used as the input to the prediction model, and the true quality deviation index of the batch of products is used as the supervision label of the prediction model. By minimizing the mean square error between the predicted quality deviation index output by the prediction model and the true quality deviation index of the batch of products, the parameters of the prediction model are optimized, thereby training the prediction model to obtain the ability to accurately predict the product quality deviation index.

[0038] By independently calculating the optimal candidate window for each dimension, it is ensured that the most representative data fragments of its own abnormal state are extracted from each dimension. Subsequently, the largest window among the optimal candidate windows of all dimensions is selected as the uniform input length. This satisfies the requirement of the prediction model for uniform input data format while ensuring that key abnormal signals from any dimension are not lost due to window truncation, thus providing the prediction model with training samples that cover complete abnormal information from all dimensions.

[0039] Moreover, the predictive model can learn the collaborative changes and concurrent patterns among anomalous features of different dimensions, rather than viewing changes in a single dimension in isolation. Based on this, the predictive model trained on this basis makes judgments based on a multi-dimensional and comprehensive chain of anomalous evidence, thereby significantly improving the accuracy and robustness of the predictive model when facing complex industrial scenarios.

[0040] S2: Calculate the influence weight of each dimension based on the volatility index of the optimal candidate window in the same dimension of all running datasets and the error between the predicted quality offset index and the true quality offset index of all running datasets.

[0041] Based on the data obtained in step S1, for a single dimension, the volatility index of the optimal candidate window for that dimension is obtained from all running datasets, and a volatility index sequence is generated by arranging the running datasets in the order of their collection time. The error between the predicted quality offset index and the true quality offset index of all running datasets is obtained, and an error sequence is generated by arranging the running datasets in the order of their collection time. The standard deviation of the volatility index sequence and the covariance between the volatility index sequence and the error sequence are calculated. The ratio of the covariance to the standard deviation is calculated, and this ratio is negatively normalized to obtain the influence weight of that dimension. The specific formula is as follows: In the formula, Indicates the first The influence weights of each dimension; This indicates the first of all running datasets. The volatility index sequence is formed by the volatility index of the optimal candidate window in each dimension. This represents the error sequence formed by the error between the predicted quality offset index and the true quality offset index of all running datasets; It represents the covariance between the volatility index series and the error series, and is used to measure the consistency of their changing trends; The standard deviation represents the volatility index series and is used to quantify the dispersion of the volatility index in this dimension.

[0042] Characterized the first The change in prediction error caused by a unit change in the volatility index of each dimension, i.e., the sensitivity of the data volatility in that dimension to the model's prediction error. The larger this ratio, the more significant the impact of the volatility in that dimension on the prediction accuracy, and therefore it should be given a higher impact weight.

[0043] The normalization function normalizes the ratios calculated for each dimension, ensuring that the sum of the influence weights of all dimensions is 1, thus forming a standardized weight allocation.

[0044] After obtaining the influence weights of all dimensions, they are applied to the subsequent real-time prediction process to guide the model to dynamically and selectively process input data of different dimensions.

[0045] S3: Obtain the current input window of the data sequence of each dimension when the production equipment is producing the current batch of products. Based on the trained prediction model, input the current input window of each dimension and output the prediction quality offset index corresponding to the current input window. When the prediction quality offset index is greater than the preset threshold, the MES system issues an abnormal alarm.

[0046] The method for obtaining the current input window includes: setting a basic window length; obtaining the previous input window and the error between the corresponding predicted quality offset index and the actual quality offset index of the data sequence of each dimension when the production equipment produces the current batch of products; calculating the fluctuation index of the previous input window for each dimension; and, based on the influence weight of each dimension, weighted summing of the fluctuation indices of the previous input windows for all dimensions to obtain a sum; determining a reference error based on the error sequence; calculating the ratio of the reference error to the error corresponding to the previous input window; calculating the product of the basic window length, the aforementioned sum, and the ratio to obtain the length of the current input window; and extracting the current input window based on its length. In this embodiment, the reference error is the average of all errors in the error sequence. It is understood that in other embodiments, the reference error may also be the minimum error in the error sequence.

[0047] The method for setting the base window length includes: negatively correlating the errors between the predicted quality offset index and the true quality offset index of each running dataset to obtain the weight coefficients for each running dataset; and then weighting the input window lengths of all running datasets based on these weight coefficients to obtain the base window length. The specific formula is as follows: In the formula, Indicates the base window length; Indicates the first Weight coefficients for each running dataset; Indicates the first The length of the input window for each running dataset; This represents the total number of datasets run; where, , Indicates the first The error between the predicted quality offset index and the true quality offset index for each running dataset. It is a very small positive number, used to prevent the denominator from being zero.

[0048] The weighting coefficients of the running dataset are negatively correlated with the prediction error. The smaller the prediction error, the more effectively the corresponding input window length can extract equipment operation features strongly correlated with quality offset, and therefore it is given a higher weight, contributing more to the formation of the base window length. Conversely, the larger the prediction error, the lower the reference value of its corresponding input window length, and the lower its weight. Through this selective weighted averaging, the resulting base prediction window length incorporates high-value experience from historical data, providing a more representative and robust initial benchmark and laying a reliable foundation for subsequent dynamic adjustments to real-time predictions.

[0049] The formula for calculating the length of the current input window is as follows: In the formula, Indicates the length of the current input window; Indicates the base window length; This represents the average value of all errors in the error sequence. This represents the error between the predicted quality offset index and the actual quality offset index of the previous input window. Indicates the first The influence weights of each dimension; Indicates the first The fluctuation index of the previous input window for each dimension is calculated in the same way as the fluctuation index of the candidate window in step S1. When calculating the abnormality value of each running data in the running data sequence of each dimension when the production equipment is producing the current batch of products, the end of each running data sequence is the same as the end of the previous input window. This indicates the total number of dimensions.

[0050] This ratio reflects the deviation of the prediction accuracy of the previous input window from the historical average. When the ratio is greater than 1, it indicates that the prediction error of the previous input window is less than the average historical error, and the prediction accuracy is high. In this case, the system tends to increase the length of the current input window to utilize more extensive data to consolidate and optimize the prediction results and improve its robustness. When the ratio is less than 1, it indicates that the prediction error of the previous input window is greater than the average historical error, and the prediction accuracy is not ideal. In this case, the system tends to shorten the length of the current input window to focus more on recent equipment dynamic changes, quickly adapt to new operating states, and thus improve the sensitivity and accuracy of the current prediction.

[0051] This sum reflects the degree of abnormality in the data from the previous input window. A higher sum indicates that the equipment is currently in a significantly unstable or abnormal state. Given that production anomalies are usually persistent, expanding the input window at this time helps to smooth out any possible transient disturbances (noise), thereby more accurately capturing and confirming the true trend of persistent anomalies. A lower sum indicates that the equipment has been operating smoothly recently. In this case, shortening the input window allows the system to maintain a higher response speed to any new changes in the equipment status and avoid interference from outdated data.

[0052] By combining prediction error feedback with real-time production equipment status assessment, a closed-loop adaptive adjustment mechanism is constructed. This mechanism can intelligently determine the optimal data horizon for the next prediction based on the specific performance of each prediction and the latest status of the equipment. This method effectively overcomes the inherent limitations of fixed window length in prediction, dynamically balancing the robustness and sensitivity of the model, thereby significantly improving the accuracy and precision of each prediction and the applicability of the entire prediction model in real production environments.

[0053] After obtaining the length of the current input window, the current input window is extracted from the operational data sequence of each dimension during the production of the current batch of products. Based on the trained prediction model, the current input window for each dimension is input, and the predicted quality offset index corresponding to the current input window is output. When the predicted quality offset index is greater than a preset threshold, it indicates that there is an anomaly in the production equipment in the current prediction result, causing a large deviation in product quality. The MES system needs to issue an anomaly alarm to promptly investigate and troubleshoot the abnormal production equipment failure. After obtaining the true quality offset index of the current input window, the length of the next input window is adaptively adjusted based on the above method.

[0054] This application also discloses a production anomaly prediction system based on a large model and a MES system. The system includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the production anomaly prediction method based on a large model and a MES system according to the above embodiments of the present invention is implemented.

[0055] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0056] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A production anomaly prediction method based on a large model and a MES system, characterized in that, include: The prediction model is trained based on the historical production batches of products from the production equipment. Each production dataset includes a sequence of production data for each dimension of the production equipment. The system obtains the current input window of the operating data of each dimension when the production equipment is producing the current batch of products. Based on the trained prediction model, it inputs the current input window of the operating data of each dimension and outputs the prediction quality offset index corresponding to the current input window. When the prediction quality offset index is greater than the preset threshold, the MES system issues an abnormal alarm. The training of the prediction model based on the historical production data of each batch of products from the production equipment includes: for a single production dataset, calculating the anomaly value of each production data in a single-dimensional production data sequence; determining the starting point of the candidate window based on the anomaly value and using the end of the production data sequence as the ending point to form several candidate windows; calculating the volatility index of each candidate window based on the anomaly value of the production data, and selecting the candidate window with the largest volatility index as the optimal candidate window for that dimension; selecting the maximum window length among the optimal candidate windows of all dimensions as the input length; and using the input length and the end of the production data sequence for each dimension as the end of the input window to extract the input windows for each dimension to train the prediction model, and outputting the prediction quality offset index corresponding to the production dataset.

2. The production anomaly prediction method based on a large model and MES system according to claim 1, characterized in that, The step of determining the starting point of the candidate window based on the anomaly degree value includes: calculating the average anomaly degree value of all running data in the running data sequence; and taking the running data in the running data sequence whose anomaly degree value is greater than the average value as the starting point of the candidate window.

3. The production anomaly prediction method based on a large model and MES system according to claim 1, characterized in that, The method for calculating the anomaly value of each running data in the running data sequence includes: calculating the mean and standard deviation of all running data in the running data sequence; for a single running data, calculating the absolute value of the difference between the running data and the mean, and using the ratio of the absolute value to the standard deviation as the first ratio; calculating the average of the first derivatives of all running data within a neighborhood of the running data with a specified radius as the center, and using the ratio of the first derivative of the running data to the average of the aforementioned first derivatives as the second ratio; and calculating the product of the absolute values ​​of the first ratio and the second ratio to obtain the anomaly value of the running data.

4. The production anomaly prediction method based on a large model and MES system according to claim 1, characterized in that, The calculation of the volatility index for each candidate window based on the anomaly value of the running data includes: for a single candidate window, calculating the ratio of the maximum anomaly value of the running data within the candidate window to the maximum anomaly value of the running data in the running data sequence corresponding to the candidate window, as the third ratio; calculating the ratio of the standard deviation of the anomaly value of the running data within the candidate window to the standard deviation of the anomaly value of the running data in the running data sequence corresponding to the candidate window, as the fourth ratio; and multiplying the third ratio and the fourth ratio as the volatility index of the candidate window.

5. The production anomaly prediction method based on a large model and MES system according to claim 1, characterized in that, After training the prediction model, the process also includes: calculating the influence weight of each dimension based on the fluctuation index of the optimal candidate window in the same dimension of all running datasets and the error between the predicted quality offset index and the true quality offset index of all running datasets.

6. The production anomaly prediction method based on a large model and MES system according to claim 5, characterized in that, The calculation method for the influence weight of each dimension includes: for a single dimension, obtaining the volatility index of the optimal candidate window for that dimension in all running datasets, and arranging them in the order of the collection time of the running datasets to generate a volatility index sequence; obtaining the error between the predicted quality offset index and the true quality offset index of all running datasets, and arranging them in the order of the collection time of the running datasets to generate an error sequence; calculating the standard deviation of the volatility index sequence and the covariance between the volatility index sequence and the error sequence, calculating the ratio of the aforementioned covariance to the standard deviation, and performing negative correlation normalization on the ratio to obtain the influence weight of that dimension.

7. The production anomaly prediction method based on a large model and MES system according to claim 6, characterized in that, The method for obtaining the current input window includes: setting a basic window length; obtaining the previous input window and the error between the corresponding predicted quality offset index and the actual quality offset index of the operating data sequence of each dimension when the production equipment produces the current batch of products; calculating the fluctuation index of the previous input window of each dimension; and, based on the influence weight of each dimension, weighted summing of the fluctuation indices of the previous input windows of all dimensions to obtain a sum value; determining the reference error based on the error sequence; calculating the ratio of the reference error to the error corresponding to the previous input window; calculating the product of the basic window length, the aforementioned sum value and ratio to obtain the length of the current input window; and extracting the current input window based on the length of the current input window.

8. The production anomaly prediction method based on a large model and MES system according to claim 7, characterized in that, The method for setting the basic window length includes: performing negative correlation processing on the error between the predicted quality offset index and the true quality offset index of each running dataset to obtain the weight coefficient of each running dataset; and performing a weighted average of the input window lengths of all running datasets based on each weight coefficient to obtain the basic window length.

9. A production anomaly prediction system based on a large model and a MES system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the production anomaly prediction method based on a large model and MES system according to any one of claims 1-8.

Images