结合XAI的AutoEncoder无监督异常侦测

近年来，随着XAI的兴起，使得AI是可解释的、可信任的、可优化的，目前将新技术XAI运用于冰机、空压机等设备的PHM的案例较少，对于实时数据监测的PHM，采用XAI进行异常解释可得到不错的效果，本文使用无监督学习中的自编码技术对时序数据进行异常侦测，采用可解释器shap进行异常原因定位

 

explainer = shap.DeepExplainer(model=model, data=torch.from_numpy(X[:200].astype(np.float32)))#得到shap基线
shap_value_single = explainer.shap_values(X=torch.from_numpy(XpredictInputData.astype(np.float32)))
shap_values_weight_mean = np.zeros(260)
name_features = anamoly_data.columns#特征名称
for j in range(len(hh)):
    input_label_index_value = hh[j]
    shap_values = np.array([], dtype=np.float64)
    weight_value = weight_values[j]
    features=[]
    values=[]
    for i in range(0, LOOKBACK_SIZE):
        tmp_shap = shap_value_single[input_label_index_value][0][i]
        shap_values = np.append(shap_values, tmp_shap)
        features.extend([s + '_tt' + str(3 * (LOOKBACK_SIZE - i)) for s in name_features])
        values.extend(anamoly_data.iloc[specific_prediction_sample_to_explain_index-(LOOKBACK_SIZE-i-1),:])
    shap_values_weight_mean += weight_value * shap_values