At this stage,coal mining enterprises are mainly judged by monitoring whether there are abnormal changes in CO concentration to prevent mine fires.Due to the existence of many changeable interference factors in the mine,the CO sensor generates abnormal data,and the abnormal data interferes with the operation of the alarm system,which further affects the control and decision-making of the entire coal mine safety monitoring system.Therefore,the abnormality in the CO sensor data is accurately diagnosed.Data has important practical significance for the safe production of coal mining enterprises.Therefore,this paper studies the abnormal diagnosis method for mine CO sensor.Given the difficulty in collecting abnormal samples of CO sensors and the variety of changes,a CO sensor abnormality diagnosis model is designed in this paper,which consists of two sub-models,an abnormality detection model based on VAE-GRU and an abnormality classification model based on GRU-BP.The anomaly detection model based on VAE-GRU combines the VAE model and the GRU model and adopts the idea of probability reconstruction.The encoder projects the data into the latent space in the form of a probability distribution and samples the probability distribution to obtain a latent vector.GRU receives the hidden vector.vector and predict the hidden vector of the next time window,the decoder restores the hidden vector to data and calculates the Euclidean distance with the actual data of the next time window to judge whether there is an abnormality;and then combines GRU and BP algorithm to design abnormal data Classification model,GRU extracts the dependencies between data,and BP algorithm completes abnormal data classification.The CO anomaly diagnosis model developed in this paper is 93.9%,91.7%,and 92.8%in the accuracy rate,recall rate,and F1 value of data identification respectively;The recognition rates of,interference data and calibration data can reach 95%,93.1%,75.4%,and 99%respectively,the recognition rate of fault data and interference data is the first among similar comparison methods,and the average recognition accuracy reaches 94.8%.In order to solve the problem of long inference time of the abnormal diagnosis model of CO sensor,the parameters of the model are quantified with low precision to im prove the inference speed of the model.The method adopts a channel-by-channel sharing strategy in the selection of the shared quantization parameter range,and allocates the quantization parameters according to the number of channels;in the selection of the dynamic range of the model parameters,the entropy strategy is used to calculate the KL dispersion of the fixed-point and floating-point numbers before and after the quantization model parameters.In the process of model training,all parameters in the model are quantized in real time,and the quantization loss is calculated and added to the calculation of the model loss to reduce the impact on the network.impact on accuracy.Compared with the original model,the F1 value decreased by 0.011,the model size was reduced by 2.454 times,and the inference acceleration ratio reached 213.9%. |