| In mechanical equipment,bearings play the role of supporting the main shaft and transmitting torque.Due to the high frequency of use,making the bearings become one of the most easily damaged parts.Once the failure occurs,it may lead to more serious consequences,light equipment due to the failure of the chain reaction leading to equipment damage,or may cause personal safety issues.Therefore,real-time,accurate and rapid fault diagnosis of bearings can effectively prevent accidents.In this paper,based on the characteristics of rolling bearings presented under amplitude frequency sampling when faults occur,a combination of deep learning and ZYNQ deployment technology is investigated to achieve bearing fault diagnosis.The main research work is as follows:Firstly,a bearing fault diagnosis model based on particle swarm optimisation fused with convolutional neural network was designed to address the problems of low quality of feature extraction in the traditional convolutional neural network for bearing fault diagnosis as well as the degradation of recognition accuracy under different working conditions.The model firstly adapts the multiple hyperparameters of the model through particle swarm optimization,then introduces residual connections to prevent the gradient from disappearing,and at the same time adopts global average pooling to replace the fully connected layer part of the common model,which effectively reduces the training parameters of the model and strengthens the generality of the network,and finally adds a Dropout layer to prevent the network from overfitting.The experimental results show that the model under particle swarm optimisation hyperparameters can achieve 100%recognition rate in three of the four different working conditions,and 99.5% accuracy in the other one.Meanwhile,compared with the traditional diagnostic method,the optimised model has higher accuracy under 6 variable load conditions between the 4 conditions,and the average accuracy under 6 variable load experiments reaches more than 95%.Secondly,a bearing fault diagnosis model based on particle swarm optimisation fused with multi-scale convolutional neural network is proposed on the basis of this model while retaining the above optimisation strategy,the model adopts an improved multi-scale convolutional structure to replace the single convolutional kernel deepening the structure of the network in the original model.Experiments show that the model outperforms the pre-optimisation model in small-sample training experiments,noise experiments,and variable load condition experiments.In the small-sample training experiment,the average recognition accuracy of the model is close to 95%,in the noise experiment,the model can maintain more than 95% recognition accuracy when the signal-to-noise ratio(SIGNAL NOISE RATIO,SNR)is greater than or equal to-4d B,and the average accuracy of the model reaches more than 97% in the six load-change experiments.Finally,to address the problems of cumbersome deep learning-based bearing faults in industrial practical deployment,platform limitations,and slow operation speed on platforms that are relatively easy to deploy,this paper designs a general-purpose convolutional neural network model based on ZYNQ.The model can be deployed on the ZYNQ platform by using a high-level synthesis tool to design IP cores for the convolutional and pooling layers,and deploying the trained convolutional neural network to speed up the operation of the model.The design adopts a general-purpose design to facilitate the deployment of networks of different sizes.When the network to be deployed is changed,the deployment of the new network can be achieved by only slightly adjusting the network structure when performing the SDK development.Finally,by validating the2 diagnostic models proposed in this paper,the experimental results obtained show that the model improves the speed of convolution and pooling operations by 3.65 times and2.31 times,respectively,compared to the ARM platform with only a minimal loss of accuracy. |
