Tension roller set failure prediction method based on anomaly detection and classification algorithm

Abstract

This invention discloses a fault prediction method for tension roller groups based on anomaly detection and classification algorithms. The method involves installing vibration sensors on both sides of the motor and rollers of the tension roller group to collect vibration data at different measuring points and frequencies within the shaft system. Time-frequency domain data, frequency band proportions, and trend characteristic data are calculated. Through clustering, anomaly, and classification algorithms, combined with fault inspection and maintenance record data, an optimal fault identification model is formed. The latest real-time vibration data of the tension roller group is obtained, and the optimal fault identification model is used to predict the real-time vibration data, deriving the fault risk probability value. This method employs dimensionality reduction and anomaly detection algorithms to extract features, ensuring a strong correlation between features and faults. By pre-constructing a fault identification model, the accuracy of predicting potential faults and the efficiency of supporting maintenance decisions are improved.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis and predictive maintenance technology, and in particular to a method for predicting the faults of tension roller groups based on anomaly detection and classification algorithms. Background Technology

[0002] Tension roll assembly equipment occupies a very important position in metallurgical cold rolling lines. The normal operation of tension rolls ensures the orderly progress of production and the quality of cold-rolled sheets. However, if tension rolls malfunction and are not repaired in time, it will cause unplanned shutdowns of the production line and the quality of the produced products will not meet the standards, resulting in huge economic losses.

[0003] Traditional tension roll inspection and maintenance largely relies on on-site technicians observing the rolls' operation in real time. They determine whether the roll set is malfunctioning by checking the surface quality of the cold-rolled sheet and observing any unusual noises from the motor. This method heavily depends on the subjective judgment of on-site technicians, and it fails to detect subtle changes during roll operation. Therefore, implementing intelligent equipment monitoring, using artificial intelligence to assess the tension roll set's operating status and provide real-time feedback, and proactively predicting malfunctions, significantly improves the efficiency of determining roll malfunctions and enhances the ability to detect potential faults. This is an essential stage in the development of intelligent operation and maintenance for manufacturing enterprises.

[0004] However, it is important to clarify that even in the era of artificial intelligence, the ability to establish accurate fault identification models and predict faults remains the most pressing consideration. Tension roller faults are of many types, such as tension roller coupling failures, imbalances, rolling bearing wear, and poor lubrication. Furthermore, process and operating condition data are often difficult to obtain. Existing fault identification models are insufficient in terms of accuracy and speed, especially when operating condition and process data are lacking. They cannot promptly and accurately identify faults in tension roller groups, nor do they possess the ability to detect faults in advance.

[0005] Taking the tension roll group of the 2RAP cold rolling line at Taiyuan Iron & Steel Group's base as an example:

[0006] For this type of tension roller assembly equipment, although vibration and temperature data have been measured, vibration signal data can fluctuate due to factors such as start-up, steel grade switching, specification switching, linear speed changes, and shutdown during operation. The vibration signal data is unstable, irregular, and difficult to distinguish between processes, failing to accurately reflect the state of the steel plate during transport. Currently, due to the lack of synchronized process and operating condition data, only fault alarms for the tension roller assembly are implemented. There is no adaptive and accurate method for determining fault types, let alone an accurate analytical method for predicting potential faults. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a fault prediction method for tension roller groups based on anomaly detection and classification algorithms. This method uses dimensionality reduction and anomaly detection algorithms to extract features, ensuring a strong correlation between features and faults, and pre-constructs a fault identification model, thereby improving the accuracy of predicting potential faults and the efficiency of supporting maintenance decisions.

[0008] To address the aforementioned technical problems, the tension roller fault prediction method based on anomaly detection and classification algorithms of this invention includes the following steps:

[0009] Step 1: Install vibration sensors on the bearing seats of the motor free side, load side, and roller working side and operating side of the tension roller assembly to collect horizontal, vertical and axial acceleration vibration signals of the motor free side, load side and roller working side and operating side;

[0010] Step 2: Obtain raw vibration data for 3 months prior to the fault occurrence, from the fault occurrence to shutdown, and 1 month after fault handling. Calculate time-domain characteristics for vibration data of different directions and signal types at each measuring point to generate the first batch of characteristic data. Convert the data into frequency-domain signals using Fourier transform, calculate the frequency conversion and harmonic components, as well as the peak value of the low-frequency spectrum. Use continuous wavelets to calculate the frequency band energy ratio to generate the second batch of characteristic data. Merge the first batch of characteristic data to generate the third batch of characteristic data.

[0011] Step 3: Based on the fault diagnosis and maintenance records, label the obtained fault causes and the third batch of feature data extracted in Step 2. The data from the time the fault started to the time the fault was started to be marked as 1, the data after the fault was handled as -1, and the data before the fault occurred as floating-point values ​​between 0.99 and 0.01.

[0012] Step 4: Perform cluster analysis on the first batch of feature data before the fault occurred, select the optimal number of categories based on the data characteristics, classify the third batch of feature data, and generate the fourth batch of feature data.

[0013] Step 5: Classify the fourth batch of feature data segmented in Step 4, perform feature data transformation, and logarithmic processing on the feature data; select a time window to generate the fifth batch of feature data, which includes moving averages, mean, and standard deviation calculated by sliding calculation, and calculate the slope and intercept using a linear fitting method.

[0014] Step 6: Use principal component analysis to reduce the dimensionality of the fifth batch of feature data. The data after dimensionality reduction is the sixth batch of feature data.

[0015] Step 7: Using an anomaly detection algorithm, calculate the threshold of the sixth batch of feature data at the moment the fault begins to occur, and label the sixth batch of feature data according to the threshold to generate the seventh batch of feature data.

[0016] Step 8: Divide the seventh batch of feature data into training set, validation set and test set according to each category. Use LightGBM classification algorithm to establish a fault early warning identification model. The output of the fault early warning identification model is to predict the test set to obtain the category and probability, which are used to measure whether there are potential faults in the tension roller group and the severity of the fault risk.

[0017] Step 9: Use a Bayesian network to optimize the hyperparameters of the LightGBM classification algorithm, including: learning rate, number of leaf nodes, node depth, feature selection ratio, and sample selection ratio, to train the optimal fault warning model.

[0018] Step 10: Obtain the latest real-time vibration signal data of the tension roller group, call the optimal fault early warning model, predict the real-time vibration data, and obtain the fault risk probability value.

[0019] Furthermore, the time-domain features in step two include the margin index, kurtosis index, and quantile.

[0020] Furthermore, in step ten, the risk probability value is set to three levels: A, B, and C. The minimum risk probability for level A is 55%, for level B it is 75%, and for level C it is 95%.

[0021] Because the fault prediction method for tension roller groups based on anomaly detection and classification algorithms in this invention adopts the above-mentioned technical solution, namely, vibration sensors are set on both sides of the motor and rollers of the tension roller group to collect vibration data at different measuring points and frequencies of the shaft system, calculate time-frequency domain data, frequency band proportions, and trend feature data, and model the fault identification model by combining clustering, anomaly, and classification algorithms with fault inspection and maintenance record data, an optimal fault identification model is formed through optimization, the latest real-time vibration data of the tension roller group is obtained, the optimal fault identification model is called, and the real-time vibration data is predicted to obtain the fault risk probability value. This method uses dimensionality reduction and anomaly detection algorithms to extract features, ensuring a strong correlation between features and faults, and pre-constructs a fault identification model, thereby improving the accuracy of predicting potential faults and the efficiency of supporting inspection and maintenance decisions. Attached Figure Description

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments:

[0023] Figure 1 This is a flowchart of the tension roller fault prediction method based on anomaly detection and classification algorithms of the present invention. Detailed Implementation

[0024] Implementation, for example Figure 1 As shown, the tension roller fault prediction method based on anomaly detection and classification algorithms of the present invention includes the following steps:

[0025] Step 1: Install vibration sensors on the bearing seats of the motor free side, load side, and roller working side and operating side of the tension roller assembly to collect horizontal, vertical and axial acceleration vibration signals of the motor free side, load side and roller working side and operating side;

[0026] Step 2: Obtain raw vibration data for 3 months prior to the fault occurrence, from the fault occurrence to shutdown, and 1 month after fault handling. Calculate time-domain characteristics for vibration data of different directions and signal types at each measuring point to generate the first batch of characteristic data. Convert the data into frequency-domain signals using Fourier transform, calculate the frequency conversion and harmonic components, as well as the peak value of the low-frequency spectrum. Use continuous wavelets to calculate the frequency band energy ratio to generate the second batch of characteristic data. Merge the first batch of characteristic data to generate the third batch of characteristic data.

[0027] Step 3: Based on the fault diagnosis and maintenance records, label the obtained fault causes and the third batch of feature data extracted in Step 2. The data from the time the fault started to the time the fault was started to be marked as 1, the data after the fault was handled as -1, and the data before the fault occurred as floating-point values ​​between 0.99 and 0.01.

[0028] Step 4: Perform cluster analysis on the first batch of feature data before the fault occurred, select the optimal number of categories based on the data characteristics, classify the third batch of feature data, and generate the fourth batch of feature data.

[0029] Step 5: Classify the fourth batch of feature data segmented in Step 4, perform feature data transformation, and logarithmic processing on the feature data; select a time window to generate the fifth batch of feature data, which includes moving averages, mean, and standard deviation calculated by sliding calculation, and calculate the slope and intercept using a linear fitting method.

[0030] Step 6: Use principal component analysis to reduce the dimensionality of the fifth batch of feature data. The data after dimensionality reduction is the sixth batch of feature data.

[0031] Step 7: Using an anomaly detection algorithm, calculate the threshold of the sixth batch of feature data at the moment the fault begins to occur, and label the sixth batch of feature data according to the threshold to generate the seventh batch of feature data.

[0032] Step 8: Divide the seventh batch of feature data into training set, validation set and test set according to each category. Use LightGBM classification algorithm to establish a fault early warning identification model. The output of the fault early warning identification model is to predict the test set to obtain the category and probability, which are used to measure whether there are potential faults in the tension roller group and the severity of the fault risk.

[0033] Step 9: Use a Bayesian network to optimize the hyperparameters of the LightGBM classification algorithm, including: learning rate, number of leaf nodes, node depth, feature selection ratio, and sample selection ratio, to train the optimal fault warning model.

[0034] Step 10: Obtain the latest real-time vibration signal data of the tension roller group, call the optimal fault early warning model, predict the real-time vibration data, and obtain the fault risk probability value.

[0035] Preferably, the time-domain features in step two include margin indices, kurtosis indices, and quantiles.

[0036] Preferably, in step ten, the risk probability value is set to three levels: A, B, and C. The minimum risk probability for level A is 55%, for level B it is 75%, and for level C it is 95%.

[0037] Typically, predicting potential faults in tension rollers requires close integration of operating conditions and process data. However, the differences in data characteristics under different operating conditions and processes, along with the differences between potential faults and normal data, can make it impossible to judge and predict. Since there is no operating condition and process data available, this method uses the start time of the fault occurrence stage as the end point and provides a fault prediction method for tension roller groups based on anomaly detection and classification algorithms.

[0038] This method involves installing vibration sensors on the free and load sides of the tension roller motor, and on the working and operating sides of the roller, respectively. The sensors output raw vibration waveforms, which are then modeled and stored as fault identification model files. After acquiring 30 days of monitoring data following fault handling, the fault identification model is used to output fault probability values. If the fault warning threshold is reached, a corresponding message is sent to the intelligent operation and maintenance platform to remind staff to schedule equipment inspection and maintenance.

[0039] This method collects vibration signal data from different measuring points and frequencies in the shaft system, calculates time-frequency domain data, frequency band proportions, and trend characteristic data, avoiding diagnostic errors caused by single data features. Through clustering, anomaly, and classification algorithms, it models and outputs the predicted types of potential faults and the probability of unplanned downtime caused by the faults during operation phases where there are no manual inspection results, no abnormalities are found by traditional methods, or potential faults exist. This improves the accuracy of predicting potential faults and the efficiency of maintenance decision support. If the output shows a high probability of risk, staff can choose to carry out maintenance work during the most recent routine downtime maintenance period to avoid unplanned downtime, thereby improving production line stability and increasing production capacity.

[0040] This method uses data prior to the occurrence of a failure as model training data. Theoretically, it can provide early warning of unplanned downtime of tension roller sets more than 7 days in advance, allowing production line maintenance to be scheduled ahead of time. This enables the replacement of rollers that are about to fail during the most recent routine maintenance, reducing unplanned downtime. This is of great significance for enhancing production line stability and increasing production capacity.

Claims

1. A method for predicting tension roller group faults based on anomaly detection and classification algorithms, characterized in that... This method includes the following steps: Step 1: Install vibration sensors on the bearing seats of the motor free side, load side, and roller working side and operating side of the tension roller assembly to collect horizontal, vertical and axial acceleration vibration signals of the motor free side, load side and roller working side and operating side; Step 2: Obtain raw vibration data for 3 months before the fault occurred, from the fault occurrence to the shutdown, and 1 month after the fault was resolved. Calculate time-domain characteristic quantities for vibration data of different measurement points and different signal types to generate the first batch of characteristic data. The signal is converted into a frequency domain signal by Fourier transform, and the frequency conversion and harmonic components and the peak value of the low frequency band spectrum are calculated. The frequency band energy ratio is calculated by continuous wavelet, and the second batch of feature data is generated. The first batch of feature data is merged to generate the third batch of feature data. Step 3: Based on the fault diagnosis and maintenance records, label the obtained fault causes and the third batch of feature data extracted in Step 2. The data from the time the fault started to the time the fault was started to be marked as 1, the data after the fault was handled as -1, and the data before the fault occurred as floating-point values ​​between 0.99 and 0.

01. Step 4: Perform cluster analysis on the first batch of feature data before the fault occurred, select the optimal number of categories based on the data characteristics, classify the third batch of feature data, and generate the fourth batch of feature data. Step 5: Classify the fourth batch of feature data segmented in Step 4, perform feature data transformation, and logarithmic processing on the feature data; select a time window to generate the fifth batch of feature data, which includes moving averages, mean, and standard deviation calculated by sliding calculation, and calculate the slope and intercept using a linear fitting method. Step 6: Use principal component analysis to reduce the dimensionality of the fifth batch of feature data. The data after dimensionality reduction is the sixth batch of feature data. Step 7: Using an anomaly detection algorithm, calculate the threshold of the sixth batch of feature data at the moment the fault begins to occur, and label the sixth batch of feature data according to the threshold to generate the seventh batch of feature data. Step 8: Divide the seventh batch of feature data into training set, validation set and test set according to each category. Use LightGBM classification algorithm to establish a fault early warning identification model. The output of the fault early warning identification model is to predict the test set to obtain the category and probability, which are used to measure whether there are potential faults in the tension roller group and the severity of the fault risk. Step 9: Use a Bayesian network to optimize the hyperparameters of the LightGBM classification algorithm, including: learning rate, number of leaf nodes, node depth, feature selection ratio, and sample selection ratio, to train the optimal fault warning model. Step 10: Obtain the latest real-time vibration signal data of the tension roller group, call the optimal fault early warning model, predict the real-time vibration data, and obtain the fault risk probability value.

2. The method for predicting tension roller group faults based on anomaly detection and classification algorithms according to claim 1, characterized in that: The time-domain features in step two include the margin index, kurtosis index, and quantiles.

3. The method for predicting tension roller group faults based on anomaly detection and classification algorithms according to claim 1, characterized in that: In step ten, the risk probability value is set to three levels: A, B, and C. The minimum risk probability for level A is 55%, for level B it is 75%, and for level C it is 95%.

Images