Xuelang_AI—Diagnosis-of-Motor

Xuelang Industrial Data Challenge

Question background

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

Data exploration

Training set:

There were 30 positive samples and 500 negative samples. The samples were severely uneven.

Test set:

Sample 5738 sets

Data length:

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.

𝑓_𝑎𝑢𝑔 (𝑥)=𝜆∗max⁡( 𝑓(𝑥))∗𝑟𝑎𝑛𝑑𝑜𝑚()
where:
 
  𝜆Proportion for data enhancement
  𝑓 (𝑥) is the feature column that needs to be enhanced, such as ai1, ai2
  max⁡ (𝑓 (𝑥)) The maximum value of the feature column
  𝑟𝑎𝑛𝑑𝑜𝑚 () is a random number generation function, a Gaussian distribution with a random number of 0-1

Feature extraction

Sliding window construction features:

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.

Sliding window 1: Window 2000, steps 1000

Sliding window 2: Window 3000, steps 1000

Sliding window 3: Window 10240, steps 10240

Sliding window 4: Window 12800, steps 12800

Statistical characteristics: Maximum, minimum, range, mean, variance, skewness, kurtosis

Glitch features: ai1 and ai2 glitch number, ratio, and period

Frequency domain characteristics: STFT transform, using CNN to extract spectrogram features

Timing characteristics: autocorrelation, binned entropy

Tree model construction features: Use the values of the leaf nodes output by the tree model as features

Model

ResNet50

Idea: structure data volume overfitting

Do short-time Fourier transform to extract stft features

Splicing feature channels and inputting the resnet network structure model

lightgbm model

Idea: Further tuning

No adjustment of parameters

5-fold cross-validation output (fast and stable)

Model advantages

Model advantages and program potential:

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

Additional promising

Try more different data enhancement methods to increase the amount of data

To further improve the CNN model, in addition to using the cosine annealing method to train the model, you can try to use Zhejiang University's AdaBound optimization method in 2019

Further try to use the most recent octave convolution to speed up the calculation

Tuning the tree model for better results

try more methods

Environment

Operating environment: 8g memory, windows, python3.6

Dependent packages: lightgbm-2.2.3, xgboost-0.82, tsfresh-0.11.2, sklearn, multiprocessing, tqdm

Generate results process:

1. The training data is stored in data / Motor_tain and the test data is stored in data / Motor_testP.

2. Run run.py in the order of steps 1, 2, 3, 4, 5;

3. The final result is generated in the data root directory, named sub_final.csv

Directory architecture:

root: [xuelang]

+-data

| +-Motor_tain

| | +-Negative

| | +-Positive

| +-Motor_tain_tutu

| | +-Positive

| +-Motor_testP

| +-stack

| +-test

| +-train

+-method.py

+-read_me.txt

+-run.py

Final result

Recall rate 100%

Accuracy 89.11%