對於變轉速狀態下之機械設備，所產生的信號模式大部分為非穩態信號，使用頻譜分析方法（如傅立葉轉換，fast fourier transform, FFT）其信號特徵會隨分析時間長度而平均化，無法突顯重要的特徵訊號，導致在故障診斷或辨識上之困難。為了改善此缺點，本文提出時頻階次譜方法。此方法結合短時傅立葉轉換（short-time fourier transform, STFT）與轉速頻率階次方法，擷取非穩態狀態信號的階次特徵。此種信號特徵不因轉速變化而改變，可有效作為機械設備在非穩態狀態運轉下之故障辨識。此外，本文將短時傅立葉時頻階次譜結合主成份分析法進行資料量降維，提取時頻階次譜之主成份輸入倒傳遞類神經網路（back propagation neural network, BPNN），進行齒輪－轉子實驗平台於非穩態狀態運轉下之故障診斷，期望達到快速故障診斷之效果。

When the machine equipment is under the status of varying speeds, the signal patterns there are generated by the mechanical equipment are largely non-stationary. Using the order spectrum analysis method (fourier transform, fast fourier transform, FFT), the signal features are averaged in corresponding to the length of analysis time, thus making it impossible to highlight the signal characteristics and causing difficulties in identifying or diagnosing faults. For improving this weakness, in this paper, we propose an order spectrum method. This method combines the Short-Time Fourier Transform (STFT) and speed frequency ordering method to establish the non-steady state status signal characteristic. This kind of signal characteristic doesn't change because of changing the variety, and can effectively be machine equipment in the non-steady state under the status of breakdown recognition. In addition, Principal Components Analysis (PCA) is used to extract the main features of the order spectrum and reduce the volume of data. This is combined with the Back Propagation Neural Network (BPNN) to devise an artificial intelligence method for fault diagnosis in non-stationary states. Lastly, the order spectrum method is verified by using a gear-rotor test platform that proves the feasibility of the theory.