| With the rapid development of urban construction and the rise of urban high-rise buildings,the demand for elevators is becoming increasingly dependent.However,at present,passengers are not aware of safe elevator rides,especially for the phenomenon of users pushing electric vehicles to ride elevators,which is not uncommon in people’s lives.However,people do ignore the many hazards of the electric car entering the car.For example,it may crash the elevator car door when entering the car,and the uneven force of the elevator motor may cause various safety problems of the elevator.At present,the monitoring of electric vehicles entering the car is mainly focused on the use of smart cameras to perform image recognition on them to determine whether there is an electric vehicle.Judging by only a single camera,the comprehensive accuracy rate and false alarm rate of the elevator car in a closed and complex environment are not satisfactory.In view of the above problems,the main work of this thesis is as follows:(1)Combining the actual application scenarios of this paper,a model combining multiple neural networks and improved Dempster-Shafer Evidence Theory(D-S evidence theory)is proposed.According to the proposed model,the appropriate sensor is selected,and the data set is prepared in combination with the actual complex situation of the elevator.(2)Aiming at the current method of judging that the electric car enters the car only by image recognition,this paper uses VGG19 neural networks,Back Propagation Neural Network(BPNN)and improved Res Net101 neural networks to identify and classify the collision sound of the electric car,the three-axis gravity acceleration and the image data.(3)Aiming at the shortcomings in the traditional D-S evidence theory,this paper proposes an evidence synthesis method based on cosine similarity and Euclidean distance,and builds an improved algorithm model.In order to verify the effectiveness of the method proposed in this paper,this paper carried out different comparative experiments based on three neural networks.In the pattern recognition of the collision sound of electric vehicles,the logarithmic mel spectrogram contains more characteristic information than the wavelet packet decomposition,and the highest accuracy rate of 66.8% is obtained in the VGG19;in the three-axis gravitational acceleration pattern recognition Among them,the comprehensive training of BP by using the three-axis combined acceleration peak and the three-axis gravitational acceleration change amplitude is the best,with an accuracy rate of 71.3%;in the image pattern recognition,the SE-Res Net101 network has an accuracy rate compared with Res Net101 and Alex Net.As high as 81.6%.The improved evidence theory algorithm verifies the feasibility of the method through case analysis and comparison with previous classics.And using the data set of this article for statistical analysis,the accuracy rate is increased by 14.7% compared with the traditional D-S evidence theory algorithm,which verifies the practicability of the method.It provides an important theoretical basis for the future smart elevator Internet of Things. |
