Xuelang Industrial Data Challenge
Using big data and artificial intelligence technology to automate motor abnormal sound detection, solve the pain points that artificial detection cannot accurately and reliably identify abnormal sounds, and help upgrade lean manufacturing and intelligent manufacturing
Single sample 4 columns of audio with a length of 79999 and a sampling rate of 51200.The duration is 1.5625s.
After visualization, it can be seen from the figure that the number of glitches in the positive sample is larger, and the glitches are larger.In the time-frequency plot of the data of the positive and negative samples, the dominant frequencies are quite different.
In the original data (530 groups), there was a serious imbalance between positive and negative samples. To solve this problem, we enhanced the positive samples. We use Gaussian white noise for enhancement, on the one hand, it reduces the impact on the original spectrum and at the same time, it reduces the impact on statistical characteristics. Then we used the lightgbm model and the stft-cnn model to analyze the statistical and spectral characteristics, respectively.
π_ππ’π (π₯)=πβmaxβ‘( π(π₯))βππππππ()
where:
πProportion for data enhancement
π (π₯) is the feature column that needs to be enhanced, such as ai1, ai2
maxβ‘ (π (π₯)) The maximum value of the feature column
ππππππ () is a random number generation function, a Gaussian distribution with a random number of 0-1
Selection of sliding window size: The sound itself is not a stable random sequence. Set the length of the window to extract long and short audio information.
Tree model construction features: Use the values of the leaf nodes output by the tree model as features
Feature-rich, including time domain and frequency domain features The model runs fast, and 5 minutes from the feature extraction to the model run is enough The results are stable, there is not much difference between the preliminary and live competition scores, the model has strong generalization performance and strong practicability The model is simple and does not need to adjust parameters