A component abnormality warning and diagnosis method

Abstract

The present invention discloses a component abnormality warning and diagnosis method, which includes the following steps: setting component abnormality warning indicators for a pre-trained warning model based on machine component position warning indicator information; obtaining data information of each component of each machine; inputting the data information of each component into the pre-trained warning model, and outputting the probability distribution of abnormal states of easily-worn components; and quickly matching abnormal causes and treatment measures in a knowledge graph based on the probability distribution of abnormal states of easily-worn components.

Description

Technical Field

[0001] The present invention relates to a component abnormality early warning and diagnosis method, belonging to the technical field of component management in the tobacco industry. Background Art

[0002] As tobacco manufacturers' production equipment continues to grow in complexity and intelligence, their production capacity is also increasing. Equipment maintenance and fault warnings are crucial for ensuring the proper operation of production equipment and reducing operating costs. The complex and tightly coupled electromechanical structure of production equipment leads to chain reactions and mutual influences between its components, increasing the difficulty of detecting abnormal faults. However, current fault warning methods for production equipment components are inaccurate, often preventing maintenance personnel from developing appropriate equipment maintenance plans.

[0003] During the production process, wear and tear of machine components can lead to increased product quality defects, which are then reflected in corresponding process indicators. By developing an early warning model for key component failures based on the real-time data collected by these production devices and improving the sensitivity of early warnings for wear-prone components, we can, to a certain extent, address the issue of delayed component failure warnings. Summary of the Invention

[0004] The purpose of the present invention is to provide a component abnormality warning and diagnosis method to solve the defect of the prior art that component failure warning is not timely.

[0005] A component abnormality warning and diagnosis method, the method comprising the steps of:

[0006] According to the machine component position warning indicator information, the pre-trained warning model is used to set the component abnormality warning indicator selection;

[0007] Obtain data information of each component of each machine;

[0008] Input the data information of each component into the pre-trained early warning model and output the probability distribution of abnormal status of the wear-prone parts;

[0009] Quickly match abnormal causes and treatment measures in the knowledge graph based on the probability distribution of abnormal states of wearable parts.

[0010] Furthermore, the machine component position warning indicator information includes:

[0011] Set the location information of each component of each machine;

[0012] Set the early warning indicator information for each component position of each machine.

[0013] Furthermore, the selection of component abnormality warning indicators includes:

[0014] By analyzing the historical real-time data of the equipment production process, we use the indicator feature extraction to find the component abnormality warning indicators that affect the product quality status, analyze the status changes of each indicator data, and verify the status changes of the indicators in real-time production;

[0015] The correlation weight of each indicator is calculated based on the entropy method, and the weight of each indicator is sorted. Then, those indicators with very low weights are excluded. Then, the weight of each indicator is recalculated for the remaining indicators based on historical data to ensure the effectiveness of each indicator in warning of component abnormalities and confirm the correlation between product quality status and the characteristics of each indicator.

[0016] Furthermore, the component abnormality warning includes:

[0017] Monitor the index values ​​of the fault-causing components of the pilot machine models in real time, and regularly predict the abnormal status distribution of components of each machine based on the component early warning model. If an abnormal state is found, record the index data of the component abnormality and immediately push the one or more components that caused the component abnormality. Match the abnormal causes and treatment measures in the knowledge graph through the component abnormality early warning.

[0018] Furthermore, the step of obtaining data information of each component of each machine includes:

[0019] A data acquisition rule for each component position of each machine is set, and the data information of each component of each machine is acquired to meet the acquisition rule.

[0020] Furthermore, the training method of the early warning model includes:

[0021] Obtain the valid historical data of each component position of each machine and assemble it into a directly usable DataFrame;

[0022] Use the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data;

[0023] Assign each piece of data independently to a data cluster;

[0024] Calculate the distance between two data in two-dimensional space;

[0025] Sort by the distance between the data to get a list of relationships from near to far;

[0026] Circulate the relationship list and merge the two closest data clusters into one data cluster;

[0027] Repeat the merging steps until the data clusters in the list reach a discrete state;

[0028] Print discrete clusters of data on a two-dimensional chart and output the data to a file;

[0029] Display two-dimensional data to assist manual decision making and determine the accuracy of data clusters;

[0030] Read the clustered data and label it based on manual and machine-assisted decisions;

[0031] Use the decision tree algorithm to randomly divide the data into two parts: training data and verification data;

[0032] The SVC algorithm is used to load the training data and train the early warning model.

[0033] Furthermore, the method also includes optimizing the early warning model, specifically by:

[0034] Convert data into training data according to data preprocessing rules;

[0035] Divide the newly generated reliable data into two categories: training data and verification data according to the proportion;

[0036] Add the newly generated training data to the original training set and retrain the prediction model;

[0037] Use the newly generated verification data to verify the accuracy of the early warning model, and adjust the scale of training data and data preprocessing rules according to the accuracy;

[0038] Adjust the training data in the training set to adapt to the new data preprocessing rules;

[0039] Retrain the pretrained model again.

[0040] Furthermore, the SVC algorithm is a support vector machine algorithm, which includes:

[0041] Establish an optimal decision hyperplane so that the distance between the two types of samples closest to the plane on both sides is maximized. For a multidimensional sample set, the system randomly generates a hyperplane and moves it continuously to classify the samples until the sample points belonging to different categories in the training samples are exactly located on both sides of the hyperplane. Find a hyperplane that meets the conditions and maximizes the blank area on both sides of the hyperplane, thereby achieving optimal classification of linearly separable samples.

[0042] Compared with the existing technology, the present invention has the following advantages: when a component is in an abnormal state, it uses data analysis algorithms to infer the probability of the abnormal state, thereby quickly matching the abnormal cause and treatment measures in the knowledge graph and pushing them to business personnel. This achieves the purpose of rapid warning and diagnosis of component abnormalities, rapid search for spare parts, and rapid repair.

[0043] The invention is applicable to all grassroots production departments such as cigarette rolling and packaging workshops and silk-making workshops of all cigarette industrial enterprises, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the steps of the method of the present invention;

[0045] Figure 2 This is an example of the t-SNE algorithm of the present invention;

[0046] Figure 3 It is an example of the SVM algorithm of the present invention;

[0047] Figure 4 It is a component abnormality warning diagram of the present invention;

[0048] Figure 5 This is a diagram of an abnormal warning framework for easily worn parts based on a model algorithm of the present invention;

[0049] Figure 6 It is a correlation analysis diagram of various process parameters of abnormal components of the present invention;

[0050] Figure 7 This is a fluctuation trend analysis diagram of key process parameters of the present invention;

[0051] Figure 8 This is a flowchart of the component abnormality warning and diagnosis management business of the present invention;

[0052] Figure 9 The present invention is based on component evaluation of multi-level warning. DETAILED DESCRIPTION

[0053] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0054] like Figure 1 The present invention discloses a component abnormality warning and diagnosis method, the method comprising the steps of:

[0055] Step 1: According to the machine component position warning indicator information, set the component abnormality warning indicator selection for the pre-trained warning model;

[0056] Step 2: Obtain data information of each component of each machine;

[0057] Step 3: Input the data information of each component into the pre-trained early warning model and output the probability distribution of abnormal status of the wear-prone components;

[0058] Step 4: Quickly match the abnormal causes and treatment measures in the knowledge graph based on the probability distribution of abnormal states of wearable parts.

[0059] In this embodiment, step 1 is further explained. First, according to the machine component position warning indicator information, the following steps are performed: setting the warning indicator information of each component position of each machine. The use case function description is shown in the following table: Setting the warning indicator information of the equipment component:

[0060] name Set early warning indicator information for equipment components Function Set early warning indicator information for the component location of each machine Executor Management personnel Prerequisites Prepare early warning indicator information for component locations that need to be set Postcondition Selection of component abnormality warning indicators enter Warning indicator information of the set component location Process Description Set early warning indicator information for each component position of each machine Output Warning indicator information of the set component location Business Rules none

[0061] Secondly, set the component abnormality warning indicators for the warning model, and the specific settings are as follows:

[0062] By analyzing the historical real-time data of the equipment production process, the indicator feature extraction is used to find out the component abnormality warning indicators that affect the product quality status, analyze the status changes of each indicator data, and verify the status changes of the indicators in real-time production. At the same time, the correlation weight of each indicator is calculated based on the entropy method, and the weight of each indicator is sorted. Then, those indicators with very low weights are excluded, and then the weight of each indicator is recalculated again based on the historical data for the remaining indicators to ensure the effectiveness of each indicator for component abnormality warning, so as to confirm the correlation between the product quality status and each indicator feature. The use case function description is as follows: Selection of equipment component abnormality warning indicators:

[0063]

[0064]

[0065] In this embodiment, step 2 is specifically described as follows:

[0066] Before acquiring data, we must first set the data acquisition rules for each component position of each machine, such as the frequency of data acquisition, the steady-state vehicle speed and the duration of data acquisition, etc. The data cleaning rule settings are shown in the following table:

[0067] name Data cleaning rule settings Function Set data acquisition rules for each component position on each machine Executor Management personnel Prerequisites Prepare data cleaning rules Postcondition Data cleaning management enter Data cleaning rules content Process Description Set data acquisition rules for each component position on each machine Output Data cleaning rules content Business Rules none

[0068] Set the location information of each component of each machine, as shown in the following table:

[0069] name Set device component information Function Set the component location information for each machine Executor Management personnel Prerequisites Prepare the component location information that needs to be set Postcondition Selection of component abnormality warning indicators enter Set component position information Process Description Set the position information of each component on each machine Output Set component position information Business Rules none

[0070] Equipment component indicator settings:

[0071] Set the early warning indicator information for each component position of each machine. The use case function description is shown in the following table to set the early warning indicator information of equipment components:

[0072] name Set early warning indicator information for equipment components Function Set early warning indicator information for the component location of each machine Executor Management personnel Prerequisites Prepare early warning indicator information for component locations that need to be set Postcondition Selection of component abnormality warning indicators enter Warning indicator information of the set component location Process Description Set early warning indicator information for each component position of each machine Output Warning indicator information of the set component location Business Rules none

[0073] In this embodiment, the training method of the early warning model includes:

[0074] Obtain the valid historical data of each component position of each machine and assemble it into a directly usable DataFrame;

[0075] Use the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data;

[0076] Assign each piece of data independently to a data cluster;

[0077] Calculate the distance between two data in two-dimensional space;

[0078] Sort by the distance between the data to get a list of relationships from near to far;

[0079] Circulate the relationship list and merge the two closest data clusters into one data cluster;

[0080] Repeat the merging steps until the data clusters in the list reach a discrete state;

[0081] Print discrete clusters of data on a two-dimensional chart and output the data to a file;

[0082] Display two-dimensional data to assist manual decision making and determine the accuracy of data clusters;

[0083] Read the clustered data and label it based on manual and machine-assisted decision-making;

[0084] Use the decision tree algorithm to randomly divide the data into two parts: training data and verification data;

[0085] Use the SVC algorithm to load the training data to train and obtain the early warning model;

[0086] After obtaining the early warning model, in order to further improve the accuracy of data warning, it is necessary to further optimize and adjust the early warning model and improve the breadth and depth of training data. The system can optimize the early warning model by following the steps below:

[0087] S121: Develop a daily data collection plan and add valid data to the dataset of known abnormal data and known normal data based on the workshop operation status;

[0088] S122: Converting the data into training data according to data preprocessing rules;

[0089] S123: Divide the newly generated credible data into two categories: training data and verification data according to the proportion;

[0090] S124: Add the newly generated training data to the original training set and retrain the prediction model;

[0091] S125: Use the newly generated verification data to verify the accuracy of the early warning model, and adjust the scale of the training data and data preprocessing rules according to the accuracy;

[0092] S126: Adjust the training data in the training set to adapt to the new data preprocessing rules;

[0093] S127: Retrain the preprocessing model again.

[0094] like Figure 2 As shown in the figure, a brief description of some artificial intelligence algorithms used in building the model is as follows:

[0095] B1: The t-SNE algorithm is the t-distributed stochastic neighbor embedding (t-SNE) clustering algorithm

[0096] t-SNE (tSNE) converts the similarity of data points into probabilities. Similarity in the original space is represented as a Gaussian joint probability (converted to a probability based on the similarity between the data points), while similarity in the embedded space is represented as a Student's t distribution. This makes t-SNE particularly sensitive to local structure and offers several other advantages over existing techniques.

[0097] t-SNE is currently the most effective data dimensionality reduction and visualization method, but it has drawbacks, such as high memory usage and long runtime. However, when we want to classify high-dimensional data and are unsure whether the dataset is well separable (i.e., small gaps between similar classes and large gaps between different classes), we can use t-SNE to project the data into a 2D or 3D space.

[0098] The quality of the visualization is evaluated by the Kullback-Leibler (KL) divergence of the joint probability between the original space and the embedded space. This means that a function related to the KL divergence is used as the loss function, which is then minimized through gradient descent until convergence is achieved. In practice, the t-SNE algorithm can be optimized by adjusting multiple parameters to achieve optimal visualization quality.

[0099] like Figure 3 As shown in B2: the SVC algorithm is a support vector machine (Support Vector Machine, SVM) algorithm.

[0100] The main idea of ​​the Support Vector Machine (SVM) is to establish an optimal decision hyperplane that maximizes the distance between the two closest samples on either side of the plane, thereby providing good generalization capabilities for classification problems. For a multidimensional sample set, the system randomly generates a hyperplane and continuously moves it, classifying the samples until the sample points belonging to different categories in the training sample are exactly on both sides of the hyperplane. There may be many hyperplanes that meet this condition. The SVM formally finds a hyperplane that maximizes the blank area on both sides of the hyperplane while ensuring classification accuracy, thereby achieving optimal classification of linearly separable samples.

[0101] The advantages of SVM are as follows:

[0102] 1. They don't require many samples. This doesn't mean the absolute number of training samples is small. Rather, it means that compared to other classification algorithms, SVMs require fewer samples for the same problem complexity. Furthermore, because SVMs incorporate kernel functions, they can easily handle high-dimensional samples.

[0103] 2: Minimize structural risk. This risk refers to the cumulative error between the classifier’s approximation of the true model of the problem and the true solution to the problem.

[0104] Nonlinearity means that SVM is good at dealing with situations where sample data is linearly inseparable. This is mainly achieved through slack variables (also called penalty variables) and kernel function technology. This part is also the essence of SVM.

[0105] B3:LightGBM (Light Gradient Boosting Machine) algorithm

[0106] The LightGBM is a new member of the boosting ensemble model. Like XGBoost, it is an efficient implementation of GBDT. In principle, it is similar to GBDT and XGBoost. Both use the negative gradient of the loss function as the residual approximation of the current decision tree to fit a new decision tree.

[0107] The Gradient Boosting Decision Tree (GBDT) is a long-standing model in machine learning. Its key concept is to iteratively train weak classifiers (decision trees) to obtain an optimal model. This model offers advantages such as good training performance and low overfitting resistance. GBDT is widely used in industry, often for tasks such as click-through rate prediction and search ranking.

[0108] The LightGBM algorithm described above performs better than XGBoost in many aspects and can speed up the training of the GBDT model without compromising accuracy. LightGBM optimizes the traditional GBDT algorithm as follows:

[0109] B4: Decision tree algorithm based on Histogram.

[0110] Gradient-based One-Side Sampling (GOSS): Using GOSS can reduce the number of data instances with only small gradients, so that only the remaining data with high gradients can be used when calculating information gain. Compared with XGBoost traversing all eigenvalues, it saves a lot of time and space overhead.

[0111] Exclusive Feature Bundling (EFB): Using EFB, many mutually exclusive features can be bundled into one feature, thus achieving the purpose of dimensionality reduction.

[0112] Leaf-wise growth strategy with depth restriction: Most GBDT tools use an inefficient level-wise decision tree growth strategy because it treats leaves at the same level indiscriminately, incurring unnecessary overhead. In fact, many leaves have low splitting gains, making searching and splitting unnecessary. LightGBM uses a leaf-wise growth algorithm with depth restriction.

[0113] According to the data cleaning rules, valid data of the specified machine component position and whether there are abnormal results are obtained within a period of time. The use case function description is shown in the following table Data Cleaning Management:

[0114]

[0115]

[0116] Based on the valid historical data of each component location on each machine, combined with machine learning algorithms, the component abnormality warning model is trained and the accuracy of the model is tested. The use case function description is shown in the component abnormality warning model management table below:

[0117]

[0118] like Figure 4As shown, the indicator values ​​of the fault-causing components of the pilot machine models are monitored in real time. The abnormal status distribution of components of each machine is regularly predicted based on the component warning model. If an abnormal state is found, the indicator data of the component abnormality is recorded and the one or more components causing the component abnormality are immediately pushed. The abnormal cause and treatment measures in the knowledge graph are matched through the component abnormality warning. The use case function description is shown in the component abnormality warning table below:

[0119]

[0120] like Figure 5 As shown, the system uses data modeling to analyze part status. It analyzes data label points and component correlations by analyzing equipment module maintenance and replacement records, historical data in real-time libraries, and expert consultations. It establishes relationships between 55 types of easily worn parts and the data points with high correlations with those parts within the knowledge graph, forming feature vectors for these parts. Using the real-time data of these feature vectors as input, it creates an abnormality prediction model for easily worn parts and outputs a probability distribution for abnormal states of easily worn parts. Training and adjusting the abnormality prediction model for easily worn parts improves model prediction accuracy. As a result, the system uses data analysis algorithms to infer the probability of an abnormality when a easily worn part exhibits an abnormal state, and quickly matches the abnormality cause and treatment measures within the knowledge graph to business personnel.

[0121] Although the wear and abnormal conditions of parts and components during the equipment production process can be reflected in product quality defects and the corresponding process indicator data can be found. However, it is still not easy to identify specific easily-worn parts and corresponding process parameter indicators. This requires a large amount of data analysis of process parameter feature associations and parameter trend changes and empirical evidence to determine their correlation. In addition, to determine whether the corresponding process parameter indicators of parts and components are "normal" and "abnormal", it is necessary to cluster the process parameter indicator data. According to the clustering results, a class is defined for each cluster, and then a component abnormality classification model is trained based on these classes to determine the abnormal status of equipment parts and components in daily production operations, and to promptly push abnormal warning information of easily-worn parts to business personnel. Based on the equipment failure knowledge graph, it recommends handling measures for component abnormalities to help business personnel quickly deal with component abnormalities, reduce product quality defects, and improve product quality. To this end, the project members carried out the following work:

[0122] like Figure 6 As shown, S910: clarify the correlation between easy-to-wear parts and process parameter indicators

[0123] By analyzing historical data from the production process and using feature extraction to identify the main process parameters that affect product defects (empty cigarettes), the trends of each process parameter are identified. These trends are verified in real-time production, and the process parameters are sorted according to their correlation weights to confirm the correlation between components and the characteristics of each process parameter.

[0124] In the process of functional design and development, we took the ZJ116 winder component (such as the front-end leveling plate device) as an example to illustrate how to determine the correlation between this easily worn component and process parameter indicators through big data analysis. The steps are as follows:

[0125] S911: Obtain all process parameters of the current machine, historical real-time data in various time periods before and after the front-end flattening plate components are replaced and before and after the front-end flattening plate components are in maintenance status, perform data cleaning according to data cleaning rules, and obtain steady-state valid data for related analysis such as parameter characterization status.

[0126] S912: Compare the preliminary process parameters affected by the front-end leveling disc components with the process parameters after big data analysis, optimize and adjust them according to the different weights of each process parameter, and understand the correlation between each process parameter.

[0127] S913: When performing data analysis on process parameters of abnormal front-end flattening disk components, it is also necessary to perform fluctuation trend analysis and parameter characterization status analysis on historical real-time process parameters. If the parameter trend is abnormal, an abnormal prompt will be triggered to confirm whether the status of each process parameter is truly the process parameter affected by the abnormal front-end flattening disk components.

[0128] like Figure 7 As shown in S914: After the above steps, we obtained the first 30 preliminary indicators that have a significant impact on the front-end flattening plate components. The process parameters that cause abnormalities in the front-end flattening plate components were preliminarily confirmed, and then reviewed and confirmed by manual experience with equipment experts. In actual situations, continuous verification and optimization were carried out to ensure the effectiveness of the parameters. The process parameter indicators related to the front-end flattening plate components were obtained as shown in the following table:

[0129]

[0130] S915: Using the same method, we selected 40 important components and their corresponding process parameters of the No. 5 winder of the ZJ116 model, including the front-end flattening plate device, the rear-end flattening plate device, the front-end suction belt, the rear-end suction belt, and the VE negative pressure, as well as 15 important components and their corresponding process parameters of the No. 5 packaging machine of the ZB48 model. The relevant sample table is shown in the following table:

[0131]

[0132] S920: Obtain real-time valid data according to data cleaning rules

[0133] Based on the selected process parameter indicators related to each component and in accordance with the confirmed data cleaning rules (for example: output greater than 20,000 pieces, vehicle speed greater than a steady-state speed of 2,000 pieces / minute, continuous operation time greater than 3 minutes, and the unit time frequency of data acquisition, such as 10 minutes), the data of the process parameter indicators are cleaned and preprocessed, invalid data and abnormal data are filtered or marked, valid data for a time period (real-time value or incremental value) is obtained, and the data is saved in a time series database.

[0134] The effective real-time data of the process parameter indicators of the components obtained are shown in the following table:

[0135]

[0136] S930: Construction of component abnormality warning model

[0137] We can use the confirmed process parameter indicators for a particular wear-prone component (pre-process leveling plate) to obtain valid data from a real-time database over a long period of time (including data before and after repair or replacement). Based on the current valid data, we use the entropy method to calculate the different weights of all process parameter indicators corresponding to the component. The current valid data also serves as the training data sample (including the validation data set) for the valid data of the warning leveling plate component. For this valid data, we use the t-SNE clustering algorithm to perform multi-dimensional dimensionality reduction clustering. Clustering can be used for dimensionality reduction and vector quantization, compressing high-order features into two columns. The data is then discretized, grouped, and labeled (divided into two categories: normal and abnormal) using the SVM classification algorithm. The component anomaly prediction model is trained using the Light GBM decision tree algorithm, and the model parameters are continuously tuned in actual testing to obtain the optimal component anomaly prediction model. In the same way as described above, component anomaly warning models can be established for all wear-prone components.

[0138] S940: Predict the probability of abnormalities of wearable parts affecting short positions based on real-time data

[0139] Based on the real-time data of various process parameter indicators of the machine's easily worn parts, the model is used to warn the accuracy of normal and abnormal conditions of each easily worn part, and then the different weights of multiple indicators of each part are calculated to obtain the overall abnormality probability of each part; the different abnormality probability information of multiple parts is pushed according to the abnormality push rules.

[0140] S950: Improve the indicators of easily worn parts

[0141] For the prediction results of easily worn parts, if there is a deviation from the actual situation of the easily worn parts, the range of the process parameter indicator TAG points of the selected parts can be expanded or adjusted according to the actual situation; historical valid data with abnormalities (such as the valid data of a shift before the part is replaced) can also be selected and added to the model training data sample; according to the entropy method, multiple process parameter indicators that have an important impact on the parts are screened to form the training data sample to train the new model again. The third step of building an abnormal warning model for easily worn parts can be repeated, and through testing and verification, a more complete abnormal warning model for easily worn parts can be obtained.

[0142] S960: Verification results of component abnormality warning model

[0143] After we have established a component abnormality warning model, we will obtain the machine process indicator parameter data for a period of time when a certain component was replaced in the past from the real-time database to test the component abnormality warning model.

[0144] S1010: Automatic repair reporting and maintenance information push function based on shutdown code

[0145] During the production process, equipment downtime can provide a direct indicator of component reliability. Downtime due to specific reasons is likely a reflection of decreased component reliability. A downtime code is a message containing the equipment fault number and specific details of the fault, reported when a fault occurs during operation. Key information about downtime codes includes expert evaluation of ZJ116 and ZB48 models to determine whether the downtime code triggers a repair notification and the option to set a repair notification.

[0146] Multi-level warning is a quality monitoring application developed by the cigarette factory based on CPS. The technical department sets limits for key indicators. When these limits are exceeded, multi-level alarms are triggered. Based on this, the knowledge graph system links alarm information with equipment components. When a multi-level warning message appears during production, the machine's warning event information is promptly issued. If the current warning event requires repair, a repair call request is automatically sent to the equipment management system. Based on the equipment knowledge graph library and the equipment fault diagnosis network model, the backend service pushes the most likely fault phenomenon, cause, and corrective measures when a multi-level warning event requires repair.

[0147] S1020: Device abnormality diagnosis based on event monitoring

[0148] like Figure 9As shown, based on the equipment maintenance records, the downtime event information within 3 days of the equipment maintenance occurs is obtained, and the corresponding relationship between the equipment maintenance part BOM and the corresponding components, quality defects, fault phenomena, fault causes, treatment measures, various downtime codes, multi-level warning codes and occurrence times is established. Finally, the relationship between equipment components and downtime codes is formed, and then the red and yellow judgment standards for maintenance or replacement of equipment components, downtime codes and downtime times are established. Then, based on this component maintenance judgment standard, the current maintenance status of the equipment component is obtained. The process of determining component maintenance by downtime code is as follows:

[0149] S1021: Based on the equipment maintenance record, obtain downtime event information within 3 days when the equipment maintenance occurs;

[0150] S1022: Find the corresponding component based on the equipment maintenance part BOM, establish the corresponding relationship between the component and the quality defect, fault phenomenon, fault cause, treatment measure, each stop code and the number of stoppages, and finally establish the relationship between the equipment component and the stop code;

[0151] S1023: Based on the above relationship data, the red and yellow judgment criteria for equipment component maintenance or replacement can be automatically generated (the conditions are the relationship factors of the shutdown code and the number of shutdowns under the current component);

[0152] S1024: Then, according to the component maintenance judgment standard, the maintenance status of each component of each device under the specified conditions is obtained (status: no maintenance, maintenance, replacement).

[0153] Through the above steps, when a component experiences an abnormal state, the data analysis algorithm is used to infer the probability of the abnormal state. This allows for rapid matching of the abnormality cause and treatment measures in the knowledge graph and push notifications to business personnel. This enables timely warning and diagnosis of component abnormalities, rapid search for spare parts, and rapid repairs.

[0154] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A component abnormality warning and diagnosis method, characterized in that: The method comprises the steps of: According to the machine component position warning indicator information, the pre-trained warning model is used to set the component abnormality warning indicator selection; Obtain data information of each component of each machine; Input the data information of each component into the pre-trained early warning model and output the probability distribution of abnormal status of the wear-prone parts; Quickly match abnormal causes and treatment measures in the knowledge graph based on the probability distribution of abnormal states of wearable parts; The selection of component abnormality warning indicators includes: By analyzing historical real-time data from the equipment production process, we use indicator feature extraction to identify component abnormality warning indicators that affect product quality status, analyze the status changes of each indicator data, and verify the status changes of indicators in real-time production; The correlation weight of each indicator is calculated based on the entropy method, and the weight of each indicator is sorted. Then, those indicators with very low weights are excluded. Then, the weight of each indicator is recalculated for the remaining indicators based on historical data to ensure the effectiveness of each indicator in warning of component abnormalities and confirm the correlation between product quality status and the characteristics of each indicator.

2. The component abnormality warning and diagnosis method according to claim 1, characterized in that: The machine component position warning indicator information includes: Set the location information of each component of each machine; Set the early warning indicator information for each component position of each machine.

3. The component abnormality warning and diagnosis method according to claim 1, characterized in that: The component abnormality warning includes: Monitor the index values ​​of the fault-causing components of the pilot machine models in real time, and regularly predict the abnormal status distribution of components of each machine based on the component early warning model. If an abnormal state is found, record the index data of the component abnormality and immediately push the one or more components that caused the component abnormality. Match the abnormal causes and treatment measures in the knowledge graph through the component abnormality early warning.

4. The component abnormality warning and diagnosis method according to claim 1, characterized in that: The acquisition of data information of each component of each machine includes: A data acquisition rule for each component position of each machine is set, and the data information of each component of each machine is acquired to meet the acquisition rule.

5. The component abnormality warning and diagnosis method according to claim 1, characterized in that: The training method of the early warning model includes: Obtain the valid historical data of each component position of each machine and assemble it into a directly usable Data Frame; Use the t-SNE algorithm to reduce the dimension and calculate the two-dimensional coordinates of the data; Assign each piece of data independently to a data cluster; Calculate the distance between two data in two-dimensional space; Sort by the distance between the data to get a list of relationships from near to far; Circulate the relationship list and merge the two closest data clusters into one data cluster; Repeat the merging steps until the data clusters in the list reach a discrete state; Print discrete clusters of data on a two-dimensional chart and output the data to a file; Display two-dimensional data to assist manual decision making and determine the accuracy of data clusters; Read the clustered data and label it based on manual and machine-assisted decisions; Use the decision tree algorithm to randomly divide the data into two parts: training data and verification data; The SVC algorithm is used to load the training data and train the early warning model.

6. The component abnormality warning and diagnosis method according to claim 5, characterized in that: The method further includes optimizing the early warning model, specifically by: Convert data into training data according to data preprocessing rules; Divide the newly generated reliable data into two categories: training data and verification data according to the proportion; Add the newly generated training data to the original training set and retrain the prediction model; Use the newly generated verification data to verify the accuracy of the early warning model, and adjust the scale of training data and data preprocessing rules according to the accuracy; Adjust the training data in the training set to adapt to the new data preprocessing rules; Retrain the pretrained model again.

7. The component abnormality warning and diagnosis method according to claim 5, characterized in that: The SVC algorithm is a support vector machine algorithm, which includes: Establish an optimal decision hyperplane so that the distance between the two types of samples closest to the plane on both sides is maximized. For a multidimensional sample set, the system randomly generates a hyperplane and moves it continuously to classify the samples until the sample points belonging to different categories in the training samples are exactly located on both sides of the hyperplane. Find a hyperplane that meets the conditions and maximizes the blank area on both sides of the hyperplane, thereby achieving optimal classification of linearly separable samples.

Images