A Roller Fault Early Warning Method Based on T-test Data Analysis

By constructing high-dimensional feature vectors and using T-test analysis, combined with the Mset algorithm, the problem of high false alarm rate in continuous casting machine roller fault monitoring was solved, achieving more accurate fault warning and stable equipment operation.

CN117591892BActive Publication Date: 2026-11-17SHANGHAI BAOSTEEL IND TECHNOLOGICAL SERVICE
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311457465.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-03
Publication Date
2026-11-17
Estimated Expiration
2043-11-03

AI Technical Summary

Technical Problem

In continuous casting machine roller fault monitoring, due to insufficient fault cases and low data dimensionality, existing technologies are unable to accurately identify the roller status, resulting in a high false alarm rate and insufficient robustness for single-parameter threshold alarms.

Method used

By collecting roller motor current data, a high-dimensional feature vector is constructed. The effectiveness of the feature vector is verified using the T-test method. An alarm model is established by combining the Mset algorithm, calculating the similarity between the observed feature vector and the estimated feature vector, and setting a similarity threshold to output an alarm.

Benefits of technology

It improves the accuracy of roller fault early warning, reduces the false alarm rate, enhances the robustness of the alarm model, reduces equipment maintenance workload, and ensures normal equipment operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117591892B_ABST
    Figure CN117591892B_ABST
Patent Text Reader

Abstract

This invention discloses a roller fault early warning method based on T-test data analysis. The method collects current data of the corresponding drive motor of the roller, removes null and zero data, and retains non-zero operating state data; it divides fault data, normal data, baseline data, and comparison data according to maintenance records; it constructs observation feature vectors through observation samples, uses baseline data as samples to construct baseline vectors, and builds baseline vectors and corresponding observation feature vectors for different cases; it uses a dual-population T-test to verify the corresponding baseline vectors and observation feature vectors respectively; it constructs a memory matrix and uses the Mset algorithm to establish the model; and it calculates the observation feature vector F. obs and estimate the eigenvector F est similarity value S obs The alarm output is obtained by comparing the data with a similarity threshold. This method analyzes the current signal data of the roller motor over a longer period, enabling the identification of the current data distribution of the roller motor under different states and improving the robustness of the alarm model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of equipment testing technology, and in particular to a method for early warning of roller faults based on T-test data analysis. Background Technology

[0002] If the population sample follows a normal distribution, and the population standard deviation is unknown and the sample size is less than 30, the t-distribution of the deviation statistic between the sample mean and the population mean will be the result of a t-distribution. T-test data analysis uses the t-distribution theory to infer the probability of differences occurring, and then compares whether the differences between the means of two sets of data are significant. The T-test is based on the assumption that the distributions of the two samples follow a normal distribution, and its null hypothesis is that the means of the two population samples are the same.

[0003] This method is applied to the health status monitoring of the sector section rollers in continuous casting machines. Since the probability of roller failure is relatively low, the number of failure cases that can be collected is relatively insufficient, making it difficult to use failure sample data for modeling. At the same time, since the only available effective data is the current data of the roller motor, the data dimension is low, and the rationality of using single-parameter thresholds for status judgment is not good. The accuracy and robustness of the judgment are easily affected by interference, such as impact data, which can easily cause unnecessary false alarms.

[0004] Taking the arc-shaped section roller of the continuous casting machine as an example, as shown in Table 1, as of July 2022, a total of two recorded fault cases were obtained. The fault on June 12th was determined to be an accurate alarm, while the fault on June 28th was determined to be a false alarm. Through data analysis, the following results were obtained: Figure 1a and Figure 1b The data distribution shown illustrates this: Case 1 exhibits a clear overall upward trend, while Case 2 shows a less pronounced upward trend with some abrupt jumps in data. This demonstrates that setting thresholds for single parameters can easily lead to false alarms.

[0005] Table 1. Statistics of cases involving the arc section of continuous casting machines.

[0006] Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a roller fault early warning method based on T-test data analysis. This method analyzes the current signal data of the roller motor over a long period, constructs a high-dimensional feature vector, and uses the T-test method to verify the effectiveness of the constructed feature vector. This enables the identification of the current data distribution of the roller motor under different states, solves the problem of single signal and excessive false alarms caused by setting threshold alarms for single signals, and improves the robustness of the alarm model.

[0008] To address the aforementioned technical problems, the roller fault early warning method based on T-test data analysis of this invention includes the following steps:

[0009] Step 1: Collect the current data of the corresponding drive motor of the roller. The data interval is 10 seconds. Rearrange the data in time order to ensure the validity of the data. Remove null and zero data and keep the non-zero working status data.

[0010] Step 2: Mark the data according to the fault time point marked in the maintenance record. Mark the data from t days before the fault time to the fault time as fault data, and mark the data before t days before the fault time as normal data. Select a segment of data under normal operating conditions from the normal data as the baseline data, and randomly select a segment of data as the comparison data.

[0011] Step 3: Construct the observation feature vector. Select data from t′ days prior to a certain moment as the observation sample. The data length is l, and the data is divided into d windows of equal size using a sliding window method. The windows are denoted as W1, W2, ..., Wd. d With a step size of p, calculate the standard deviations σ1, σ2, ..., σ of the data within the window. d Construct the feature vector F = [σ1, σ2, ..., σ] d ];

[0012] Step 4: Construct a baseline vector, using baseline data as samples, with a data length of l. b And construct the reference vector F using the sliding window method in step three. b =[σ b1 , σ b2 ,......,σ bd ];

[0013] Step 5: Prepare case studies and data, and construct the baseline vector F for different case studies. b1 F b2 ,F b3 F bn And the corresponding observed feature vectors F1, F2, F3, ..., F n ;

[0014] Step 6: Use the dual-population T-test function to verify [Fx1, F1] and [Fx2] respectively. b2 [F2], ..., [F bn F nTo determine the homogeneity of variance of an array, first perform a homogeneity of variance test on the sample. Set a homogeneity of variance parameter for the T-test. If the result of the homogeneity of variance test is greater than the set parameter, the result is True; otherwise, the result is False. Input the result of the homogeneity of variance test into the T-test function, which outputs a p-value. If the p-value is greater than 0.05, the null hypothesis is accepted, and F = 0. b If the sample means and distributions are the same as those of F, then the null hypothesis is rejected if the p-value is less than 0.05. b The mean and distribution of the samples are different from those of F.

[0015] Step 7: Construct the memory matrix. Using the Mset algorithm, establish an alarm model based on the drive motor using normal data. Select baseline data by randomly selecting m data samples of a certain length, each with a length of l′, where l′ < l. Construct the feature vectors of the samples using a sliding window method, denoted as F. s1 F s2 F sm Construct the memory matrix D = [F s1 F s2 F sm ];

[0016] Step 8: Alarm value calculation. Using data from time t′ prior to a given time point as the observed data, calculate the feature vector F of the observed data using a sliding window method. obs The feature vector is used as input to the alarm model. Based on the Mset algorithm and combined with the memory matrix, the estimated feature vector F is calculated. est The observed feature vector F is calculated by combining Euclidean distance and cosine distance. obs and estimate the eigenvector F est similarity value S obs By comparing with a preset similarity threshold S t In comparison, if S obs <S t An alarm will be output in real time.

[0017] Because the roller fault early warning method based on T-test data analysis of this invention adopts the above-mentioned technical solution, namely, this method collects the current data of the corresponding drive motor of the roller, removes null and zero data, and retains non-zero working state data; divides fault data, normal data, baseline data, and comparison data according to inspection and maintenance records; constructs observation feature vectors through observation samples, uses baseline data as samples to construct baseline vectors, and constructs baseline vectors and corresponding observation feature vectors for different cases; uses a dual-population T-test to verify the corresponding baseline vectors and observation feature vectors respectively; constructs a memory matrix, uses the Mset algorithm to establish a model; and calculates the observation feature vector F. obsand estimate the eigenvector F est similarity value S obs The alarm output is obtained by comparing the signal with a similarity threshold. This method analyzes the long-cycle current signal data of the roller motor, constructs a high-dimensional feature vector, and uses the T-test method to verify the effectiveness of the constructed feature vector. This enables the identification of the current data distribution of the roller motor under different states, solves the problem of too many false alarms caused by setting threshold alarms for single signals, and improves the robustness of the alarm model. Attached Figure Description

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments:

[0019] Figure 1a This is a schematic diagram showing the current data of the lower roller reducer in the arc section of a continuous casting machine.

[0020] Figure 1b A schematic diagram showing the current data of the upper roller reducer in the arc section of a continuous casting machine;

[0021] Figure 2 This is a flowchart of the roller fault early warning method based on T-test data analysis according to the present invention. Detailed Implementation

[0022] Implementation, for example Figure 2 As shown, the roller fault early warning method based on T-test data analysis of the present invention includes the following steps:

[0023] Step 1: Collect the current data of the corresponding drive motor of the roller. The data interval is 10 seconds. Rearrange the data in time order to ensure the validity of the data. Remove null and zero data and keep the non-zero working status data.

[0024] Step 2: Mark the data according to the fault time point marked in the maintenance record. Mark the data from t days before the fault time to the fault time as fault data, and mark the data before t days before the fault time as normal data. Select a segment of data under normal operating conditions from the normal data as the baseline data, and randomly select a segment of data as the comparison data.

[0025] Step 3: Construct the observation feature vector. Select data from t′ days prior to a certain moment as the observation sample. The data length is l, and the data is divided into d windows of equal size using a sliding window method. The windows are denoted as W1, W2, ..., Wd. d With a step size of p, calculate the standard deviations σ1, σ2, ..., σ of the data within the window. d Construct the feature vector F = [σ1, σ2, ..., σ] d ];

[0026] Step 4: Construct a baseline vector, using baseline data as samples, with a data length of l. b And construct the reference vector F using the sliding window method in step three. b =[σ b1 , σ b2 ,......,σ bd ];

[0027] Step 5: Prepare case studies and data, and construct the baseline vector F for different case studies. b1 F b2 ,F b3 F bn And the corresponding observed feature vectors F1, F2, F3, ..., F n ;

[0028] Step 6: Use the dual-population T-test function to verify [F] separately. b1 [F1], [F] b2 [F2], ..., [F bn F n To determine the homogeneity of variance of an array, first perform a homogeneity of variance test on the sample. Set a homogeneity of variance parameter for the T-test. If the result of the homogeneity of variance test is greater than the set parameter, the result is True; otherwise, the result is False. Input the result of the homogeneity of variance test into the T-test function, which outputs a p-value. If the p-value is greater than 0.05, the null hypothesis is accepted, and F = 0. b If the sample means and distributions are the same as those of F, then the null hypothesis is rejected if the p-value is less than 0.05. b The mean and distribution of the samples are different from those of F.

[0029] Step 7: Construct the memory matrix. Using the Mset algorithm, establish an alarm model based on the drive motor using normal data. Select baseline data by randomly selecting m data samples of a certain length, each with a length of l′, where l′ < l. Construct the feature vectors of the samples using a sliding window method, denoted as F. s1 F s2 F sm Construct the memory matrix D = [F s1 F s2 F sm ];

[0030] Step 8: Alarm value calculation. Using data from time t′ prior to a given time point as the observed data, calculate the feature vector F of the observed data using a sliding window method. obs The feature vector is used as input to the alarm model. Based on the Mset algorithm and combined with the memory matrix, the estimated feature vector F is calculated. estThe observed feature vector F is calculated by combining Euclidean distance and cosine distance. obs and estimate the eigenvector F est similarity value S obs By comparing with a preset similarity threshold S t In comparison, if S obs <S t An alarm will be output in real time.

[0031] Specific case analysis: Select the cases shown in Table 2, where Case 1 and Case 3 are fault cases, and Case 2 and Case 4 are false alarm cases.

[0032] Table 2. Statistics on Continuous Casting Machine Cases

[0033]

[0034] Select the parameters shown in Table 3 and calculate the p-value for each case in Table 2.

[0035] Table 3 Statistical Analysis of Continuous Casting Machine Failure Cases and Results

[0036]

[0037]

[0038] As can be seen from the p-values ​​in Table 3, Case 1 and Case 3 accepted the null hypothesis, indicating that the abnormal data and the baseline data did not follow the same distribution; Case 2 and Case 4 rejected the null hypothesis, indicating that the abnormal data and the baseline data followed the same distribution. Therefore, this method can accurately identify anomalies based on known fault samples, avoiding false alarms caused by single-threshold alarms.

[0039] This method uses T-test data analysis to construct feature vectors in a reasonable manner, enabling accurate assessment of the roller status even under objective conditions where the original data has low dimensionality. Compared with single-parameter alarms, it reduces the false alarm rate, decreases the workload of equipment maintenance personnel, ensures the normal operation of equipment, and improves the economic benefits of enterprises.

Claims

1. A roller fault early warning method based on T-test data analysis, characterized in that... This method includes the following steps: Step 1: Collect the current data of the corresponding drive motor of the roller. The data interval is 10 seconds. Rearrange the data in time order to ensure the validity of the data. Remove null and zero data and keep the non-zero working status data. Step 2: Mark the data according to the fault time point marked in the maintenance record. Mark the data from t days before the fault time to the fault time as fault data, and mark the data before t days before the fault time as normal data. Select a segment of data under normal operating conditions from the normal data as the baseline data, and randomly select a segment of data as the comparison data. Step 3: Construct the observation feature vector. Select data from t′ days prior to a certain moment as the observation sample. The data length is l, and the data is divided into d windows of equal size using a sliding window method. The windows are denoted as W1, W2, ..., Wd. d With a step size of p, calculate the standard deviations σ1, σ2, ..., σ of the data within the window. d Construct the feature vector F = [σ1, σ2, ..., σ] d ]; Step 4: Construct a baseline vector, using baseline data as samples, with a data length of l. b And construct the reference vector F using the sliding window method in step three. b =[σ b1 , σ b2 ,......,σ bd ]; Step 5: Prepare case studies and data, and construct the baseline vector F for different case studies. b1 F b2 F b3 , ..., F bn And the corresponding observed feature vectors F1, F2, F3, ..., F n ; Step 6: Use the dual-population T-test function to verify [F] separately. b1 [F1], [F] b2 [F2], ..., [F bn F n To check the homogeneity of variance of the array, first perform a homogeneity of variance test on the sample, set the homogeneity of variance parameter of the T-test, and if the result of the homogeneity of variance test is greater than the set homogeneity of variance parameter, it is True; otherwise, the result of the homogeneity of variance test is False. Input the homogeneity of variance test results into the T-test function, which outputs the p-value. If the p-value is greater than 0.05, the null hypothesis is accepted. b If the sample means and distributions are the same as those of F, then the null hypothesis is rejected if the p-value is less than 0.

05. b The mean and distribution of the samples are different from those of F. Step 7: Construct a memory matrix, adopt the Mset algorithm, use normal data to establish an alarm model with the drive motor as the basic object, select reference data, use a random point scattering method to select m data samples of a certain length, each sample has a length of l′ where l′<l, and construct the feature vectors of the samples by a sliding window method, which are respectively denoted as F s1 , F s2 , ..., F sm ; construct the memory matrix D=[F s1 , F s2 , ..., F sm ; Step 8: Alarm value calculation. Using data from time t′ prior to a given time point as the observed data, calculate the feature vector F of the observed data using a sliding window method. obs The feature vector is used as input to the alarm model. Based on the Mset algorithm and combined with the memory matrix, the estimated feature vector F is calculated. est The observed feature vector F is calculated by combining Euclidean distance and cosine distance. obs and estimate the eigenvector F est similarity value S obs By comparing with a preset similarity threshold S t In comparison, if S obs t An alarm will be output in real time.​

Citation Information

Patent Citations

  • Method for hierarchically diagnosing fault of rod-comprising oil pumping system based on indicator diagram

    CN102011576A

  • Intelligent early warning and online diagnosis method and system for equipment fault of air and smoke system

    CN113390641A