| The detection of dynamic liquid level depth in oil wells has always been a hot issue in oil extraction.The dynamic liquid level depth directly reflects the liquid supply capacity of underground oil layers and the relationship between underground supply and discharge,which is of great significance for predicting oil well production,understanding the actual situation of underground operations,and ensuring safe and efficient production.At present,the acoustic detection method is commonly used to measure the depth of oil well dynamic liquid level.However,due to the complex underground conditions,the collected sound reflection echoes are mixed with environmental noise and mechanical noise generated by the pump rod,resulting in significant calculation errors.Therefore,using acoustic denoising technology to denoise signals is crucial for improving the detection accuracy of dynamic liquid surface depth.In response to the above issues,this article adopts deep learning technology to denoise oil well echoes,and designs algorithms to detect oil well echo peaks to calculate the dynamic liquid level depth of the oil well.The main content is as follows:1.In response to the problem of environmental and mechanical noise interference constraining the detection of oil well dynamic liquid level depth,an algorithm for oil well dynamic liquid level depth detection is proposed,which combines acoustic denoising with audio peak detection.Firstly,a dual branch time-frequency denoising model WMNet is established to denoise the original acoustic signal in both the time and frequency domains;Secondly,by improving the multi period scale wave peak detection algorithm,different frequency peaks are found in pseudo periodic waves,thereby calculating the average sound velocity and dynamic liquid surface depth underground.The experimental results show that the denoising model can effectively remove noise in audio and improve signal-to-noise ratio,and has a significant improvement compared to traditional filtering algorithms;The experimental results on real audio data from oil wells show that the relative error of this method in detecting the depth of dynamic liquid surface does not exceed 0.4%,which is reduced by 7% compared to the non-denoised method.The average relative error is 0.153%,and the root mean square error is 2.82 meters.2.To address the issues of artificial artifacts caused by phase chaos,limited denoising performance,low speech quality,and low signal-to-noise ratio in frequency domain acoustic denoising algorithms,a multi-scale stepped time-frequency converter is proposed to generate adversarial networks.Using the real part,imaginary part,and amplitude spectrum of the audio spectrogram as inputs to the generator,the timefrequency transformer is first used to learn global and local feature dependencies in the time and frequency domains at multiple scales;Secondly,the Mask Decoder branch is used to learn amplitude masks,while the Complex Decoder branch directly learns clean spectrograms.The outputs of the two Decoder branches are fused to obtain the reconstructed audio;Finally,the indicator discriminator is used to distinguish the evaluation indicator scores of speech,and the generator generates high-quality and high signal-to-noise ratio audio through minimax training.The experimental results on the public dataset Voice Bank+Demand show that compared to various current acoustic denoising models,the algorithm ranks at a leading level in multiple subjective and objective speech quality evaluations,with objective speech quality perception evaluation scores and subjective speech noise distortion scores increased by 17% and21%,respectively,compared to Metric GAN.After denoising the audio data of oil wells using this method,the depth detection of oil well dynamic liquid level is carried out.The relative error does not exceed 0.2%,the average relative error is as low as 0.096%,and the root mean square error is as low as 1.43 meters,effectively improving the detection accuracy of dynamic liquid level depth. |
