رگرسیون لجستیک و طبقه بندی شبکه عصبی سوابق لرزه ای

The identification of seismic records in seismically active mines is examined by considering logistic regression and neural network classification techniques. An efficient methodology is presented for applying these approaches to the classification of seismic records. The proposed procedure is applied to mining seismicity from two mines in Ontario, Canada, and compared based on an analysis of the receiver operating characteristic curve as well as a number of performance metrics related to the contingency matrix. The logistic and neural network models presented excellent performance for identifying blasts, seismic events and reported events in the training and testing datasets for both mining seismicity catalogues. Operated under their respective optimal decision threshold values, the logistic and neural network models, accuracy was higher than 95% for classification of seismic records. In general, the logistic regression and neural network methods had close overall classification accuracies. The ability of the models to reproduce the frequency-magnitude distribution of the testing dataset was used as a signature of classification quality. The logistic and neural network models reproduced the reference distribution in a satisfactory manner. The advantages and limitations pertaining to the two classifiers are discussed.

Microseismic monitoring is employed in many mining operations in Canada as a tool to identify potential ground control safety hazards to workers. The full-waveform seismic systems employed at these mines provide real-time seismic parameters of seismic records. From these seismic parameters, it is possible to characterize the rockmass response around mining excavations, particularly to the blasting cycle which triggers most of the seismicity as aftershocks. Occasionally large magnitude events are triggered, caused by the interaction of mining and geological structures at depth. Following large seismic events or blasts there is a short-term increase in levels of seismicity that over time decays to background levels. One of the applications of the microseismic data is to enhance workplace safety by restricting access to the affected zones of the mine for sufficient time to allow this decay of aftershock events. This is the re-entry protocol [1] and [2]. A key aspect of re-entry policies is the triggering of re-entry incidents, i.e., when should a re-entry protocol be invoked? Based on a survey on current re-entry practices at 18 seismically active mines, it was established that 90% of re-entry incidents are triggered by blasting [3]. Therefore, it is necessary to accurately identify the origin time and location of blasts. The microseismic technologist at the mine has an idea when blasts are scheduled, but exact times are not recorded in blast notices or daily blast logs. It is up to the technologist to manually match blasts to the recorded seismicity; therefore, automating this procedure is an invaluable labour saving device [4]. Some of the guidelines used at the surveyed mines for invoking a re-entry protocol after large magnitude events, measured in the Nuttli magnitude scale (Mn), are: 1. Any seismic event with a Mn≥3.0, regardless of location and whether or not there was damage to mine excavations. 2. Any seismic event with a Mn≥1.5 and affecting the main accesses (e.g. ramp, footwall drifts), which could require workers to be confined to underground refuge stations and/or could require the evacuation of workers. 3. Any seismic event with a 3.0>Mn≥1.5, located within 30 m from mine excavations and/or main infrastructure (e.g. cross-cuts, ramp, refuge station, electrical sub-station, garage, crusher station). Guidelines such as these, which are based on a correlation between event magnitude and damage, require a history of seismicity and careful calibration. However, large magnitude events are not that frequent in all mining operations in Ontario, and mines that are just starting to experience seismicity and rockbursting are faced with the difficulty of having to develop their own guidelines for invoking a re-entry protocol without the benefit of significant local experience. What is needed is an approach that enables the type of seismic record (blast, microseismic event, trigger of a re-entry protocol) to be classified based on the real-time information provided by the microseismic monitoring system. This paper examines the applicability of logistic regression and neural network-based classifiers for the identification of blasts, microseismic events and events that may trigger a re-entry protocol by using multiple seismic parameters. This linkage implicitly assumes that there is a direct correlation between the seismic parameters of an individual event and the consequences as observed underground.

In this paper, approaches to develop classifiers for seismicity are evaluated and discussed. The method involves the use of logistic regression and neural network techniques, making it possible to incorporate a third class (reported incidents) into the analysis for re-entry protocol development. Logistic regression and artificial neural networks has proved to be an efficient and reliable tool for the classification of microseismic records. Operated under their respective optimal decision threshold values, the logistic and neural network models, accuracies were higher than 95%. The predictive ability of the artificial neural network was found to be comparable to that of the logistic regression for both seismic datasets. However, when a third class was included into the analysis the logistic model generalized better than the neural network model. Additional studies using other mining seismicity catalogues in different mining environments may further clarify the differences between logistic and neural network models for classifying seismic records. The fact that logistic regression can be developed quickly without overfitting the data makes it an efficient classifier that can be easily retrained as additional data becomes available. The successful predictions of the logistic and neural network models invite further refinement of these models in the identification of blasts and reported incidents. Other parameters can be introduced in the algorithm, such as: distance from mining excavations and to the microseismic array, time relative to the last blast, and triggering of the strong ground motion system. Other types of neural network architectures and classification algorithms may be tested as well.