Research On Fault Diagnosis For Reciprocating Compressors Based On Optimized Neural Networks

Reciprocating compressors are widely used in the industrial production,and the production efficiency is directly influenced by their operating conditions.Therefore,fault diagnosis for the reciprocating compressor has been paid attention all the time.Analysis for vibration signal is the most common and effective method in the area of mechanical diagnosis.However,because compressor's vibration signal is nonlinear,non-stationary and impacting-response,a few of useful features have always been drowned by noise.That make it difficult to analyze the vibration signal of reciprocating compressor.Time-frequency analysis,on behalf of short time Fourier transform(STFT),Wigner-Ville distribution(WVD)and wavelet transform(WT),has positioning function in time and frequency field.It is a useful processing method of non-stationary vibration signal whose frequency changes over time.STFT adds time windows to the vibration signal when it is carried on Fourier transform.The wider time window,the higher the frequency resolution but the lower the time resolution,and vice versa.Scilicet,STFT cannot obtain perfect resolution at both time and frequency.Besides,adding windows can also cause Picket Fence Effect.Wigner-Ville distribution has good performance in processing chirp signal because of its invariance when time or frequency is shifted and its similarity when time or frequency is extended as well as compressed.But Wigner-Ville distribution doesn't satisfy the Superposition Principle.There will be a cross-term when calculates signal's Wigner-Ville distribution which obtained by adding two signals.That will interfere analysis.Although adding time windows can restrain cross-terms in certain degree,some excellent properties such as margin features will lose.In contrast,wavelet transform has better adaptive and shifting windows,as well as "zoom" in time and frequency field.Therefore,it is one of the most widely used time-frequency method.The Morlet wavelet is one of the most commonly used wavelet basis.However,it's difficult to gain the best waveform parameter and scale parameter,and conventional methods ether are calculation-complicated or lack of physical meaning.This thesis proposes two respective methods about that.Conventional BP neural networks' learning methods converge slow,and they are easy to return local optimal solution.As a comparison,Particle Swarm Optimization(PSO)is one of the most used-commonly evolutionary algorithm for its high-precision,fast-convergence and easy-to-implement.But if all particles use same inertia weights when update speed vectors and position vectors,some particles will move towards the direction which is not conducive to their own optimization.So this thesis proposes an adaptive inertia weights algorithm to improve PSO method.If a particle's current fitness is better than the last iteration,then the inertia weight is assigned value normally;if not,then the inertia weight is assigned zero so that the particle's direction is determined only by its direction.In the end,this thesis applies the improved PSO algorithm to BP neural networks' training.Then use the BP neural networks to the fault diagnosis for reciprocating compressor.The tests show that the algorithm proposed by this thesis not only can speed up the learning process,but also gain a better diagnosis result than conventional PSO.