Main ventilator state parameter early warning method, system, medium and equipment

Abstract

The invention relates to a main ventilator state parameter early warning method, system, medium and equipment, relates to the technical field of safety engineering, analyzes the basic principle of an MSET algorithm, obtains the key of the algorithm, and proposes two dynamic memory matrix methods based on K nearest neighbor and improved K nearest neighbor. Temperature data of a motor stator with an abnormal state in a main ventilator is selected for an experiment, and a result shows that the root mean square error based on the improved K-nearest neighbor dynamic MSET algorithm is increased from 0.1944 of a normal stage to 1.0563 of an abnormal stage, which shows that the algorithm has higher recognition capability on abnormal points. And after the residual error is smoothed by introducing an exponential weighted moving average value, early warning is given out 112 minutes earlier than that of a K-nearest neighbor dynamic MSET algorithm based on the improved K-nearest neighbor dynamic MSET algorithm.

Description

Technical Field

[0001] The present invention relates to the field of safety engineering technology, and in particular to a main fan state parameter early warning method, system, medium and equipment. Background Art

[0002] In the health assessment system of the main ventilation fan in the mine, high-precision state prediction and timely fault warning are the key to ensuring safe production in coal mines. Multivariate state estimation technology (MSET) has become an important technical means in this field due to its excellent nonlinear prediction ability and high generalization performance. Multivariate state estimation technology (MSET) is an advanced nonlinear multivariate prediction and diagnosis technology developed by Argonne National Laboratory. This technology aims to explore the intrinsic connection between the health data of the equipment under normal operating conditions and the actual operation monitoring data, and evaluate the actual operating status of the equipment according to the degree of correlation.

[0003] The MSET algorithm uses data from normal operating conditions to construct a memory matrix. If the equipment is operating normally, the predicted value output by the model will be close to the measured value, and the residual between the two will be small. On the contrary, if the equipment is operating abnormally, the predicted value output by the model will be significantly different from the measured value, that is, the residual is large. By comparing the residual with the given threshold, the operating status of the equipment can be accurately evaluated. In practice, the MSET algorithm has been successfully applied to equipment status detection in many fields, including nuclear power plants, wind turbines, aerospace, etc., due to its wide applicability and low computational complexity.

[0004] The memory matrix is ​​the key to the MSET algorithm. The traditional fixed memory matrix cannot effectively utilize historical data. Current research on the construction of the memory matrix is ​​usually carried out in the offline stage. The above two methods are used to extract a part of the data from the training sample to form a fixed memory matrix. Although this method is simple, the utilization rate of data is relatively low. The fixed memory matrix composed of data extracted from a large amount of normal historical data is limited by the computational complexity of the model. The size of the memory matrix will be significantly reduced compared to the original historical data, which means that a large amount of data will be permanently discarded. If the construction method fails to extract key information and incorporate it into the memory matrix, it will have a significant impact on the prediction accuracy of the MSET algorithm. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art that data cannot be processed effectively and comprehensively and prediction accuracy cannot be improved, the main purpose of the present invention is to provide a main fan status parameter early warning method, system, medium and equipment.

[0006] To achieve the above object, the present invention adopts the following technical solution, a main fan state parameter early warning method, comprising:

[0007] The temperature parameters, electrical parameters, vibration parameters and fan performance parameters of the main ventilation fan equipment in the mine under normal operation are obtained, and the feature data after dimension reduction are obtained by processing, and divided into training set and validation set;

[0008] Collect the temperature parameters, electrical parameters, vibration parameters and fan performance parameters of the real-time status of the mine main ventilation fan, and obtain the observation vector through multiple identical feature extraction;

[0009] Introducing interval sampling to improve the dynamic MSET algorithm of K nearest neighbors, obtaining an improved dynamic MSET algorithm of K nearest neighbors, and using a training set and a validation set to train and validate the improved dynamic MSET algorithm of K nearest neighbors;

[0010] By improving the dynamic MSET algorithm of K nearest neighbors, the state data that meets the set conditions is obtained from the real-time state data, and a dynamic memory matrix is ​​constructed based on the state data that meets the set conditions;

[0011] Input the observation vector and the dynamic memory matrix into the dynamic MSET algorithm to obtain the prediction vector;

[0012] The residual is obtained by comparing the predicted vector with the observed vector, and the residual is smoothed using the exponentially weighted moving average to obtain the average residual value. If the average residual value is greater than the parameter warning threshold, a warning signal is issued.

[0013] The method of obtaining detection data of the equipment under normal operation state through the monitoring and data acquisition control system of the mine main ventilation fan, obtaining the feature data after dimension reduction and processing it includes the following steps:

[0014] When the equipment is in normal operation, the real-time detection data of the equipment operation status is collected through the monitoring and data acquisition control system of the mine main ventilation fan;

[0015] Real-time detections are normalized and feature dimension reduction is performed using principal component analysis.

[0016] The temperature parameters include motor stator temperature and bearing temperature;

[0017] The electrical parameters include three-phase voltage, two-phase current and active power of the motor;

[0018] The vibration parameters include horizontal vibration and vertical vibration of the motor;

[0019] The fan performance parameters include at least wind pressure, flow rate and static pressure efficiency.

[0020] The method of introducing interval sampling to improve the dynamic MSET algorithm of K nearest neighbors to obtain an improved dynamic MSET algorithm of K nearest neighbors includes the following steps:

[0021] After preprocessing the data, the distances between all data points are calculated, and the dynamic MSET algorithm is constructed using the minimum spanning ellipse tree algorithm, where each data point is connected to the K nearest neighbors to form a dynamically updated ellipse tree structure;

[0022] Confirm the interval sampling type, perform interval sampling on the data, obtain the sampling sub-set, apply the dynamic MSET algorithm to the sampling sub-set, and build a dynamic MSET model for interval sampling;

[0023] K nearest neighbor search is performed in the sampling subset and the dynamic MSET model of interval sampling is used to obtain the K nearest neighbor distance, forming a dynamically updated elliptical tree structure of the sampling subset, and then obtaining the improved K nearest neighbor dynamic MSET algorithm.

[0024] The method of obtaining state data meeting set conditions from real-time state data by improving the dynamic MSET algorithm of K nearest neighbors and constructing a dynamic memory matrix based on the state data meeting the set conditions includes the following steps:

[0025] Get the Euclidean distance between the observation vector and each data point in the training set;

[0026] Sort the training set in ascending order according to the Euclidean distance;

[0027] Select the first K training samples after the arrangement as the memory matrix D1, and extract M training samples according to the sampling interval from the remaining ascending training samples after the selection as the memory matrix D2;

[0028] The memory matrix D1 is combined with the memory matrix D2 to obtain a dynamic memory matrix.

[0029] A state parameter early warning system based on MSET, comprising:

[0030] The data acquisition module is used to obtain the detection data of the equipment under normal operation through the monitoring and data acquisition control system of the mine main ventilation fan, obtain the feature data after dimension reduction, process it, and divide it into a training set and a verification set;

[0031] The data processing module is used to collect the real-time state detection data of the main ventilation fan for mining, and obtain the observation vector by extracting multiple identical features; introduce interval sampling to improve the dynamic MSET algorithm of K nearest neighbors, and obtain the improved dynamic MSET algorithm of K nearest neighbors; obtain the state data that meets the set conditions from the real-time state data by improving the dynamic MSET algorithm of K nearest neighbors, and construct a dynamic memory matrix based on the state data that meets the set conditions; input the observation vector and the dynamic memory matrix into the dynamic MSET algorithm to obtain the prediction vector;

[0032] The early warning processing module is used to compare the predicted vector with the observed vector to obtain the residual, and use the exponentially weighted moving average to smooth the residual to obtain the average residual value. If the average residual value is greater than the parameter early warning threshold, an early warning signal is issued.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: through real-time monitoring and data analysis, it is possible to promptly detect signs of abnormal status of the main fan, thus avoiding equipment shutdown and production interruption due to unexpected failures. Make full use of the large amount of historical data archived by the SCADA system of the main fan in the mine to improve the utilization rate of the archived data; use the improved K nearest neighbor dynamic multivariate state estimation technology to achieve timely and accurate early warning of the abnormal status of the main fan, and the early warning capability is significantly improved; use the exponentially weighted moving average processing method to smooth the residual, improve the model's response speed to the abnormality of the main fan, and be able to detect the abnormality of the main fan more quickly. By combining real-time data acquisition, efficient data processing and predictive analysis technology, the ability to monitor and warn the operating status of the main fan is effectively improved, which plays a key role in the safety and stable operation of mining equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0035] Figure 1 It is a schematic diagram of the framework structure of the present invention;

[0036] Figure 2 It is a schematic diagram of the time series of the abnormal point of the constant temperature of the first-level motor in the present invention;

[0037] Figure 3 It is a schematic diagram of the effect of the memory matrix size on RMSE and operation time in the present invention;

[0038] FIG4 is a diagram showing the prediction results of the conventional MSET algorithm and the dynamic MSET algorithm on the validation set and the test set in the present invention; FIG4(a) is a diagram showing the prediction residual of the 1-level motor constant temperature 3, and FIG4(b) is a diagram showing the prediction residual of the 1-level motor constant temperature 3;

[0039] Figure 5 It is a schematic diagram of the construction process of the present invention based on the K nearest neighbor dynamic memory matrix;

[0040] FIG6 is the prediction results of different dynamic MSET algorithms in the present invention on the validation set and the test set, wherein FIG6(a) is the measured value and the predicted value of the fixed temperature 3 of the first-level motor, and FIG6(b) is the prediction residual of the fixed temperature 3 of the first-level motor;

[0041] Figure 7 It is a warning diagram of different MSET algorithm residuals in the present invention;

[0042] Figure 8 It is a schematic diagram of the prediction process structure of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0044] In the health assessment system of the main ventilation fan in the mine, high-precision state prediction and timely fault warning are the key to ensuring safe production in coal mines. Multivariate state estimation technology (MSET) has become an important technical means in this field due to its excellent nonlinear prediction ability and high generalization performance. Multivariate state estimation technology (MSET) is an advanced nonlinear multivariate prediction and diagnosis technology developed by Argonne National Laboratory. This technology aims to explore the intrinsic connection between the health data of the equipment under normal operating conditions and the actual operation monitoring data, and evaluate the actual operating status of the equipment according to the degree of correlation.

[0045] The MSET algorithm uses data from normal operating conditions to construct a memory matrix. If the equipment is operating normally, the predicted value output by the model will be close to the measured value, and the residual between the two will be small. On the contrary, if the equipment is operating abnormally, the predicted value output by the model will be significantly different from the measured value, that is, the residual is large. By comparing the residual with the given threshold, the operating status of the equipment can be accurately evaluated. In practice, the MSET algorithm has been successfully applied to equipment status detection in many fields, including nuclear power plants, wind turbines, aerospace, etc., due to its wide applicability and low computational complexity.

[0046] The memory matrix is ​​the key to the MSET algorithm. The traditional fixed memory matrix cannot effectively utilize historical data. Current research on the construction of the memory matrix is ​​usually carried out in the offline stage. The above two methods are used to extract a part of the data from the training sample to form a fixed memory matrix. Although this method is simple, the utilization rate of data is relatively low. The fixed memory matrix composed of data extracted from a large amount of normal historical data is limited by the computational complexity of the model. The size of the memory matrix will be significantly reduced compared to the original historical data, which means that a large amount of data will be permanently discarded. If the construction method fails to extract key information and incorporate it into the memory matrix, it will have a significant impact on the prediction accuracy of the MSET algorithm.

[0047] In order to overcome the above problems in the prior art, a main fan status parameter early warning method is specifically as follows:

[0048] The SCADA monitoring and data acquisition control system data set for the main ventilation fan in the mine and the basic specification parameters of the corresponding main ventilation fan point out that the SACDA data set only contains the monitoring data of the running fan, and the data set needs to be divided into working conditions.

[0049] Analyze the causes of abnormal data, use corresponding processing methods for different data types, use the Laida criterion to process data that conforms to the normal distribution, and use quartile processing for data that does not conform to the normal distribution. Correct the temperature data for ambient temperature, and standardize the processed data to eliminate the impact of dimension.

[0050] Interval sampling is a classic statistical sampling method that selects sample elements from the target population at intervals. The number of sample elements can be flexibly adjusted according to the interval. Interval sampling can maintain the original distribution of the target population while significantly reducing the number of samples. However, for an unordered target population, samples obtained using interval sampling may be biased. By introducing interval sampling to obtain data containing normal state distribution information, the data distribution of the memory matrix is ​​more comprehensive. The construction process is as follows: Figure 3 -6.

[0051] In the training sample X according to the observation vector X obs The Euclidean distance XT②X obs After ascending order, extract the first K training samples of the arranged training samples as memory matrix D1. Then extract M training samples from the remaining ascending training samples according to the sampling interval as memory matrix D2, merge memory matrix D1 and D2, and finally obtain a dynamic memory matrix D consisting of K+M training samples. Substitute the constructed dynamic memory matrix D into the MSET model algorithm to obtain the predicted value X est .

[0052] The dynamic memory matrix based on improved K nearest neighbor not only contains a large amount of valid data close to the current observed data, but also contains some distributed data that can reflect normal operating conditions. This enables the dynamic MSET algorithm based on improved K nearest neighbor to accurately predict the actual value of the parameters when the mine main ventilation fan is operating normally, and the state parameter prediction has extremely high accuracy; and when the state parameters of the mine main ventilation fan are abnormal, the MSET algorithm can promptly identify the difference between them and normal operating conditions, and reflect the degree of parameter abnormality through the size of the residual value.

[0053] Example:

[0054] (1) Dataset selection

[0055] This example uses the coal mine 1 data set for analysis. The data set composed of the detection data is processed and analyzed by the mine main fan SCADA monitoring and data acquisition control system to determine the characteristic parameters. By observing the characteristic parameters selected in the data set, it is found that the stator temperature of the 1st level motor is higher than that at other times around 15:45 on October 23, 2023 and 15:00 on October 26, 2023. Figure 2The fixed temperature of the first-level motor in the figure has been corrected. Although the temperature at these two time points has not reached the alarm upper limit of the motor stator temperature, it is about 4.8℃ higher than the maximum value of the motor fixed temperature at other times, which has produced a significant difference and can be regarded as an abnormal state.

[0056] The amount of data in the coal mine 1 data set after preprocessing is 274,078, which is too large. Considering that the acquisition cycle of the SCADA monitoring and data acquisition control system is 1 minute, and the values ​​of the parameters between the connected data do not change much, the acquisition cycle can be appropriately increased to reduce the amount of data. Here, the sampling cycle of the data set is increased to 4 minutes, and 68,520 data are finally obtained. The training set selects normal operation data from June 25 to October 1, 2023, totaling 34,890 data, which are used to build the MSET state monitoring model for the main ventilation fan of the mine. The test set and validation set select a total of 12,000 data from October 2 to November 2, 2023. The validation set selects the first 7,500 normal operation data of the main ventilation fan to evaluate the state prediction accuracy of the model. The test set selects 7,501 to 12,000 data containing abnormal motor constant temperature to test the model's ability to identify and warn of abnormal states.

[0057] (2) Impact of memory matrix size on the MSET algorithm

[0058] The size of the memory matrix has a significant impact on the accuracy and computational complexity of the MSET algorithm. If the memory matrix is ​​too small, although the operation time is fast, some key information will be lost when extracting data from the training set, resulting in low prediction accuracy; if the memory matrix is ​​too large, although it can provide more information and help improve the prediction accuracy of the algorithm, more data points need to be processed when making state predictions, the operation time will be greatly increased, and the system resource usage will also be too large. In practice, considering the system hardware configuration, it is necessary to balance the prediction accuracy and operation time. On the premise of meeting the prediction accuracy, the size of the memory matrix should be reduced to reduce the operation time and reduce the system load. The platform configuration parameters used in the experiment of this application are shown in Table 3-1. The model code is written in Python, version 3.9.7, and the code runs in multiple processes to improve computing efficiency.

[0059] Table 2-1 Experimental environment

[0060] Configuration Name Model version operating system Windows 11 CPU 12thGenIntel(R)Core(TM)i7-12700H GPU NVIDIA GeForceRTX3070TiLaptop8GB Running Memory 16 GB Number of cores 14 Python 3.9.7

[0061] The size of the dynamic memory matrix is ​​set to increase from 500 to 2500, with a step size of 500. Some data are selected from the normal operation data for repeated experiments. The prediction accuracy of the 1-level motor constant temperature output by the MSET algorithm model composed of memory matrices of different sizes is tested, and the operation time of the MSET algorithm model is recorded. The root mean square error RMSE is used as an indicator to measure the prediction accuracy of the model, and the change trend of RMSE and operation time with the change of memory matrix size is plotted, as shown in Figure 2. Figure 3 shown.

[0062] from Figure 3 It can be seen that as the size of the memory matrix increases, the RMSE of the MSET algorithm gradually decreases, while the operation time gradually increases. When the memory matrix size increases from 500 to 1000, the RMSE value decreases significantly, from 1.317 to 0.599, indicating that the memory matrix at this time contains too little effective information, and the increase in the size of the memory matrix can significantly improve the prediction accuracy. However, when the size of the memory matrix is ​​further increased,

[0063] The decreasing trend of RMSE becomes gentle, indicating that the memory matrix already contains enough effective information. Excessive increase in the size of the memory matrix has little effect on improving the prediction accuracy, but the required computing time is doubled. After the memory matrix size changes from 2000 to 2500, the RMSE decreases by 0.069, but the required computing time increases from 1599s to 2840s, and the computational complexity of the model increases significantly. Considering the hardware configuration of the experiment, it is decided to set the size of the memory matrix to 2000, which will not cause too much pressure on computing resources while the prediction accuracy meets the experimental requirements.

[0064] (3) Comparative experiment between conventional memory matrix and K-nearest neighbor dynamic memory matrix

[0065] The conventional memory matrix is ​​constructed using the interval sampling method proposed in the literature, and the dynamic memory matrix uses K nearest neighbor. The MSET state monitoring model of the main ventilation fan for mines is constructed respectively, and the size of the memory matrix is ​​2000. 12000 data of the validation set and the test set are input into the MSET model, and the predicted value and residual of the fixed temperature 3 of the first-level motor are obtained as shown in Figure 4.

[0066] As shown in Figure 4(a), among the 7500 data in the validation set, the measured values ​​of the stator temperature are mostly between 51 and 58°C, and the temperature of a few points reaches 59°C. Overall, it is still at normal working temperature. The output of the MSET model using the K nearest neighbor to construct the dynamic memory matrix and the output of the MSET model using interval sampling to construct the fixed memory matrix are very close to the measured values, especially the output of the dynamic MSET basically coincides with the measured values. As can be seen from Figure 4(b), the residual value of the dynamic MSET model constructed using the K nearest neighbor is significantly smaller than the residual value of the conventional MSET model constructed with the fixed memory matrix. The residual value of the dynamic MSET model is around 0.17 and has a small fluctuation, while the residual value of the conventional MSET model fluctuates between 0 and 0.8, with many glitches, and is not as good as the dynamic MSET model algorithm in terms of prediction accuracy and stability.

[0067] It can be clearly seen in the test set that the stator temperatures at two locations are higher than those at other time periods. By enlarging the comparison chart of these two locations, it can be found that the measured values ​​and the predicted values ​​have deviated and no longer have a trend of overlap. This indicates that the state parameters at this time have deviated from the state parameter values ​​under normal operation, causing the output values ​​of the MSET model to deviate from the measured values. Among them, the output value of the dynamic MSET model constructed using K nearest neighbors is closer to the measured value than the output value of the conventional MSET model. It can also be seen from the residual graph that the temperature residual value at the corresponding position of the temperature anomaly point has increased significantly. The residual change trend of the dynamic MSET model output is basically the same as that of the conventional MSET model, but the residual value of the dynamic MSET algorithm is smaller than that of the conventional MSET algorithm, indicating that the dynamic MSET algorithm is not sensitive enough to the abnormal state and may underestimate the degree of abnormality of the main fan state parameters.

[0068] Example 2

[0069] As for the construction of the dynamic memory matrix, the dynamic MSET model based on K nearest neighbors has small residuals and high prediction accuracy when the main fan is operating normally. However, it can also be seen from Figure 4 that when the main fan enters an abnormal state, the residual value output by the dynamic MSET model based on K nearest neighbors is significantly smaller than that of the conventional MSET model. Although it can be distinguished from the residual value during normal operation, its change is not significant enough, which may cause the evaluation result to weaken the abnormality of the main fan, which will have a negative impact on the accuracy of the main fan state monitoring. The reason is that the dynamic memory matrix constructed by the K nearest neighbor algorithm only contains normal state data close to the input data, resulting in a large difference between the data distribution of the memory matrix and the normal training samples. From the perspective of the information contained in the data, the training data closer to the observed data contains more effective information and can improve the prediction accuracy of the algorithm, while the training data farther away from the observed data can reflect the data distribution of the normal operating state, which helps the algorithm respond to and accurately evaluate the abnormal state data. Therefore, when constructing the dynamic memory matrix, the K nearest neighbor algorithm needs to be improved so that it can not only obtain training data that is close to the input data, but also include some training data that is farther away. The final dynamic memory matrix can reflect the data distribution in a normal state to a certain extent.

[0070] Interval sampling is a classic statistical sampling method, which selects sample elements from the target population at intervals, and the number of sample elements can be flexibly adjusted according to the intervals. Interval sampling can maintain the original distribution of the target population while greatly reducing the number of samples, but for an unordered target population, the samples sampled using interval sampling may be biased. The data of the SCADA monitoring and data acquisition control system for the main ventilation fan in the mine is time series data, but the construction of the dynamic memory matrix only extracts key data from the historical database based on the current real-time data, and does not pay attention to the time characteristics of the data. Therefore, the SCADA monitoring and data acquisition control system data can be regarded as unordered data. Therefore, interval sampling cannot be directly applied to the construction of the dynamic memory matrix, and the training set in the SCADA monitoring and data acquisition control system data must be ordered first. In the previous section, the data in the training set is arranged according to the Euclidean distance from the real-time data, which is an ordered processing of the training set. At the same time, this processing process is also part of the K nearest neighbor algorithm, which avoids repeated processing of training data and improves the efficiency of online construction of the memory matrix.

[0071] In view of the difference between the data distribution of the dynamic memory matrix constructed based on K nearest neighbors and the normal state, a dynamic memory matrix construction method based on improved K nearest neighbors is proposed. This method is an improvement on the K nearest neighbor algorithm. By introducing interval sampling to obtain data containing normal state distribution information, the data distribution of the memory matrix is ​​more comprehensive. The construction process is as follows: Figure 5 shown.

[0072] After the training samples X are arranged in ascending order according to the Euclidean distance XT②Xobs with the observation vector Xobs, the first K training samples after arrangement are extracted as the memory matrix D1. Then, M training samples are extracted from the remaining ascending training samples according to the sampling interval as the memory matrix D2, and the memory matrices D1 and D2 are merged to finally obtain a dynamic memory matrix D consisting of K+M training samples. The constructed dynamic memory matrix D is substituted into the MSET model algorithm to obtain the predicted value Xest.

[0073] The dynamic memory matrix based on the improved K nearest neighbor not only contains a large amount of valid data close to the current observed data, but also contains some distributed data that can reflect normal operating conditions. This enables the dynamic MSET algorithm based on the improved K nearest neighbor to accurately predict the actual value of the parameters when the mine main ventilation fan is operating normally, and the state parameter prediction has extremely high accuracy; when the state parameters of the mine main ventilation fan are abnormal, the MSET algorithm can promptly identify the difference between them and the normal operating conditions, and reflect the degree of parameter abnormality through the size of the residual value. The response to abnormal conditions is not sensitive enough, which may underestimate the degree of abnormality of the state parameters of the main ventilation fan.

[0074] Example 3

[0075] This example will conduct a comparative experiment between the dynamic MSET algorithm based on improved K nearest neighbors and the dynamic MSET algorithm based on K nearest neighbors. The data is still the coal mine 1 data set, and the preprocessing stage is the same as the previous section. In the memory matrix construction stage, the dynamic memory matrix based on improved K nearest neighbors proposed in this section is used to construct

[0076] The dynamic memory matrix construction method based on K nearest neighbor proposed in Section 3.2 is used to build two different dynamic MSET prediction models. The memory matrix size is 2000, where K=1000 and M=1000 in the improved K nearest neighbor construction process. A total of 12,000 running data in the validation set and the test set are input into the two MSET models respectively.

[0077] The comparison between the measured value and the predicted value of the constant temperature 3 of the first-level motor and the prediction residuals of the two MSET models are shown in Figure 6.

[0078] As shown in Figure 6, after removing the abnormal areas of the two 1-level motors with constant temperature 3, the predicted values ​​of the dynamic MSET algorithm based on the improved K nearest neighbor are basically coincident with the measured values, and the residual value obtained is very small compared with the K nearest neighbor dynamic MSET algorithm, which is due to the fact that the constructed memory matrix contains many training data with close distance and high similarity to the moment. In the temperature abnormality area, the dynamic MSET algorithm based on the improved K nearest neighbor can well distinguish the data at that moment from the normal state data. In the local magnified image, it can be found that the predicted value and the measured value have obvious differences, indicating that the training set under normal operation does not contain data with high similarity to the moment. Although the constructed dynamic memory matrix selects some data with relatively close distance from the training set, the distance is significantly increased compared with the normal operation state. At the same time, the training data obtained by equal-interval sampling contains some data with long distance and low similarity, which will further amplify the difference between the moment and the normal state, so that the final residual will not be insignificant due to the small distance between the dynamic memory matrix and the abnormal data.

[0079] The evaluation indicators are root mean square error RMSE, mean absolute percentage error MAPE and mean absolute error MAE, and the prediction accuracy of the conventional MSET algorithm, K nearest neighbor MSET algorithm and improved K nearest neighbor MSET algorithm under normal and abnormal conditions is quantitatively evaluated. All sample data in the validation set are selected for normal conditions, and 100 sample data before and after the first temperature abnormal point are selected for abnormal conditions. The evaluation results are shown in Table 2.

[0080] Table 2 Performance of different MSET models under different working conditions

[0081]

[0082] As shown in Table 2, under normal working conditions, the RMSE, MAPE, and MAE of the K-nearest neighbor MSET algorithm are the smallest, indicating that its prediction accuracy is the highest. The RMSE of the improved K-nearest neighbor MSET algorithm is 0.1944, which is only 0.024 larger than the K-nearest neighbor, and the difference is very small. The prediction accuracy is still at a high level. The RMSE, MAPE, and MAE of the conventional MSET algorithm are still small in value, but they are significantly larger than the other two MSET algorithms, and its prediction accuracy is far inferior to that of the K-nearest neighbor and improved K-nearest neighbor algorithms.

[0083] Since the memory matrix only contains normal working condition data, the prediction accuracy of the MSET algorithm is high under normal working conditions, but the prediction accuracy will drop significantly under abnormal working conditions. The decline in prediction accuracy is due to the lack of training data similar to the abnormal state data in the memory matrix. The difference of the MSET algorithm is reflected in the construction of the memory matrix. Under the same scale, the better the memory matrix is ​​constructed, the better the prediction accuracy of normal working conditions. At the same time, the difference in prediction accuracy between normal and abnormal working conditions is greater. As can be seen from Table 2, under abnormal working conditions, the prediction accuracy of the conventional MSET algorithm is the lowest, but due to its low prediction accuracy under normal working conditions, the RMSE difference is 0.7319, and the accuracy difference is not as good as the improved K nearest neighbor MSET algorithm. The RMSE of the K nearest neighbor MSET algorithm is 0.6671 in abnormal conditions, and the prediction accuracy is still maintained at a high level, with a RMSE difference of only 0.4967. Part of the reason is that the abnormal state is not obvious enough, but the main reason is that the memory matrix is ​​composed entirely of training data closest to the current state data. The data selected under abnormal conditions are all training data close to the edge of the normal operating range, resulting in a subtle decrease in model accuracy and blurring the difference between normal and abnormal conditions. The improved K-nearest neighbor MSET algorithm showed a significant decline in prediction accuracy under abnormal conditions, with RMSE increasing to 1.0563. The difference in prediction accuracy between normal and abnormal conditions was the largest, with an RMSE difference of 0.8619. This shows that the improved K-nearest neighbor MSET algorithm is sensitive to abnormal data while maintaining high prediction accuracy, can identify abnormal equipment conditions in a timely manner, and is suitable for abnormal state parameter warning.

[0084] Example 4

[0085] This embodiment is a type of warning example analysis. The first 500 data when the stator temperature parameter reaches the maximum value for the second abnormality are selected for residual analysis. The corresponding time is from 5:29 on October 21, 2023 to 14:45 on October 22. The MSET algorithm uses conventional MSET, K nearest neighbor MSET and improved K nearest neighbor MSET. There are differences in the prediction accuracy of different algorithms, and the residual warning threshold needs to be divided separately. After calculation, the conventional MSET threshold Td = 1.00066°C, the K nearest neighbor MSET threshold Td = 0.41909°C, and the improved K nearest neighbor threshold Td = 0.53026°C. The exponentially weighted moving average algorithm with a smoothing coefficient α = 0.1 is used to smooth the stator temperature residuals. The residual trend changes as shown in the figure. Figure 7 shown.

[0086] from Figure 7It can be seen that the residual fluctuation after exponential weighted moving average processing is slightly flat, and it is maintained in a lower value area during the normal operation stage. When the motor constant temperature is abnormal, the residual begins to increase. Among them, the improved K nearest neighbor dynamic MSET model first reaches the warning threshold and exceeds the threshold at 10:21 on the 22nd. The conventional MSET model exceeds the warning threshold at 11:13, but its prediction accuracy is not high under normal operation. Although the K nearest neighbor dynamic MSET model has the highest prediction accuracy under normal operation, it does not respond to abnormal conditions in time and exceeds the warning threshold at 12:13, which is 112 minutes later than the improved K nearest neighbor dynamic MSET model. The motor constant temperature reaches the maximum value at 14:45, and the improved K nearest neighbor dynamic MSET model issues a warning for the abnormal constant temperature 264 minutes in advance. In summary, the improved K nearest neighbor dynamic MSET model has the fastest warning response to abnormal conditions while maintaining high prediction accuracy.

[0087] The basic working principle of a main fan status parameter early warning method is:

[0088] The abnormal warning process of the state parameters of the main ventilation fan in the mine based on the improved K-nearest neighbor dynamic MSET algorithm is obtained by using multivariate state estimation technology, principal component analysis, improved K-nearest neighbor dynamic memory matrix construction and exponentially weighted moving average processing method. First, the historical data collected from the SCADA monitoring and data acquisition control system of the main ventilation fan in the mine is preprocessed to filter out the historical health data. Then, the principal component analysis is used to reduce the dimension of the characteristic parameters, and the key parameters are extracted to reduce the complexity of the model. The parameters after feature dimension reduction are standardized to eliminate the influence of order of magnitude and dimension differences, and the standardized data can constitute the initial data set. The real-time data is converted into the observation vector of the model input through the same feature extraction. The improved K-nearest neighbor algorithm is used to select qualified data from the initial data set to form a dynamic memory matrix. The dynamic MSET algorithm calculates the prediction vector based on the observation vector and the memory matrix, and compares it with the observation vector to obtain the residual. The residual is smoothed using an exponentially weighted moving average to see if the residual value exceeds the parameter warning threshold. If not, it is considered that the mine main ventilation fan is operating normally, and the real-time data is updated to start a new round of warning process. If the residual value exceeds the parameter warning threshold, an alarm is issued. The specific process is as follows: Figure 8 shown.

[0089] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0090] The above embodiments are merely examples of the present invention and do not limit the protection scope of the present invention. All designs that are the same or similar to the present invention fall within the protection scope of the present invention.

Claims

1. A main fan status parameter early warning method, characterized in that: The following steps are involved: The temperature parameters, electrical parameters, vibration parameters and fan performance parameters of the main ventilation fan equipment in the mine under normal operation are obtained, and the feature data after dimension reduction are obtained by processing, and divided into training set and validation set; Acquire the temperature parameters, electrical parameters, vibration parameters and fan performance parameters of the real-time status of the mine main ventilation fan, and obtain the observation vector through multiple identical feature extraction; Introducing interval sampling to improve the dynamic MSET algorithm of K nearest neighbors, obtaining an improved dynamic MSET algorithm of K nearest neighbors, and using a training set and a validation set to train and validate the improved dynamic MSET algorithm of K nearest neighbors; By improving the dynamic MSET algorithm of K nearest neighbors, the state data that meets the set conditions is obtained from the real-time state data, and a dynamic memory matrix is ​​constructed based on the state data that meets the set conditions; Input the observation vector and the dynamic memory matrix into the dynamic MSET algorithm to obtain the prediction vector; The residual is obtained by comparing the predicted vector with the observed vector, and the residual is smoothed using the exponentially weighted moving average to obtain the average residual value. If the average residual value is greater than the parameter warning threshold, a warning signal is issued.

2. The main fan status parameter early warning method according to claim 1, characterized in that: The method of obtaining detection data of the equipment under normal operation state through the monitoring and data acquisition control system of the mine main ventilation fan, obtaining the feature data after dimension reduction and processing it includes the following steps: When the equipment is in normal operation, the real-time detection data of the equipment operation status is collected through the monitoring and data acquisition control system of the mine main ventilation fan; Real-time detections are normalized and feature dimension reduction is performed using principal component analysis.

3. The main fan status parameter early warning method according to claim 1, characterized in that: The temperature parameters include motor stator temperature and bearing temperature; The electrical parameters include three-phase voltage, two-phase current and active power of the motor; The vibration parameters include horizontal vibration and vertical vibration of the motor; The fan performance parameters include at least wind pressure, flow rate and static pressure efficiency.

4. The main fan status parameter early warning method according to claim 1, characterized in that: The method of introducing interval sampling to improve the dynamic MSET algorithm of K nearest neighbors to obtain an improved dynamic MSET algorithm of K nearest neighbors includes the following steps: After preprocessing the data, the distances between all data points are calculated, and the dynamic MSET algorithm is constructed using the minimum spanning ellipse tree algorithm, where each data point is connected to the K nearest neighbors to form a dynamically updated ellipse tree structure; Confirm the interval sampling type, perform interval sampling on the data, obtain the sampling sub-set, apply the dynamic MSET algorithm to the sampling sub-set, and build a dynamic MSET model for interval sampling; K nearest neighbor search is performed in the sampling subset and the dynamic MSET model of interval sampling is used to obtain the K nearest neighbor distance, forming a dynamically updated elliptical tree structure of the sampling subset, and then obtaining the improved K nearest neighbor dynamic MSET algorithm.

5. The main fan status parameter early warning method according to claim 1, characterized in that: The method of obtaining state data meeting set conditions from real-time state data by improving the dynamic MSET algorithm of K nearest neighbors and constructing a dynamic memory matrix based on the state data meeting the set conditions includes the following steps: Get the Euclidean distance between the observation vector and each data point in the training set; Sort the training set in ascending order according to the Euclidean distance; Select the first K training samples after the arrangement as the memory matrix D1, and extract M training samples according to the sampling interval from the remaining ascending training samples after the selection as the memory matrix D2; The memory matrix D1 is combined with the memory matrix D2 to obtain a dynamic memory matrix.

6. A state parameter early warning system based on MSET, characterized in that: include: The data acquisition module is used to obtain the detection data of the equipment under normal operation through the monitoring and data acquisition control system of the mine main ventilation fan, obtain the feature data after dimension reduction, process it, and divide it into a training set and a verification set; The data processing module is used to collect the real-time state detection data of the main ventilation fan for mining, and obtain the observation vector by extracting multiple identical features; introduce interval sampling to improve the dynamic MSET algorithm of K nearest neighbors, and obtain the improved dynamic MSET algorithm of K nearest neighbors; obtain the state data that meets the set conditions from the real-time state data by improving the dynamic MSET algorithm of K nearest neighbors, and construct a dynamic memory matrix based on the state data that meets the set conditions; input the observation vector and the dynamic memory matrix into the dynamic MSET algorithm to obtain the prediction vector; The early warning processing module is used to compare the predicted vector with the observed vector to obtain the residual, and use the exponentially weighted moving average to smooth the residual to obtain the average residual value. If the average residual value is greater than the parameter early warning threshold, an early warning signal is issued.

7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the method according to any one of claims 1 to 5 is implemented.

Images