Machine Learning Techniques for Noise Attenuation in Geophysical Surveys

Abstract

Techniques are disclosed relating to machine learning in the context of noise filters for sensor data, e.g., as produced by geophysical surveys. In some embodiments, one or more filters are applied to sensor data, such a harsh filter determined to cause a threshold level of distortion in measured reflections, a mild filter determined to leave a threshold level of remaining noise signals, or an acceptable filter. In some embodiments, the system trains a machine learning classifier based on outputs of the filtering procedures and uses the classifier to determine whether other filtered sensor data from the same survey exhibits acceptable filtering. This may improve accuracy or performance in detecting unacceptable filtering, in some embodiments.

What is claimed is: 
     
         1 . A method, comprising:
 accessing, by a computing system, sensor data;   applying one or more filtering procedures to the sensor data, wherein an output of the one or more filtering procedures is determined to have a threshold level of distortion to the sensor data or leave a threshold level of remaining noise signals in the sensor data;   training, by the computing system, a classification engine based on the output of the one or more filtering procedures; and   classifying, by the computing system using the trained classification engine, other filtered sensor data to determine whether the other filtered sensor data exhibits acceptable filtering.   
     
     
         2 . The method of  claim 1 ,
 wherein the one or more filtering procedures include:
 a first filtering procedure, wherein an output of the first filtering procedure is determined to have a threshold level of distortion to the sensor data; 
 a second filtering procedure, wherein an output of the second filtering procedure is determined to leave a threshold level of remaining noise signals in the sensor data; and 
 a third filtering procedure, wherein an output of the third filtering procedure is determined to provide acceptable filtering outputs. 
   
     
     
         3 . The method of  claim 1 ,
 wherein the training comprises:
 generating respective sets of attributes in an attribute space for the one or more filtering procedures based on determined similarity between:
 outputs of the one or more filtering procedures; and 
 differences between the outputs of the one or more filtering procedures and the sensor data; and 
 
 performing feature extraction to generate respective sets of features for the one or more filtering procedures in a feature space; and 
   wherein the classifying generates attributes and features for the other sensor data and determines which portions of the feature space correspond to the generated features.   
     
     
         4 . The method of  claim 3 , wherein the generating the respective sets of attributes include determining one or more of:
 cross-correlation;   mean lambda;   rank correlation; or   Markov correlation.   
     
     
         5 . The method of  claim 4 , wherein the feature space has a smaller number of dimensions than the attribute space. 
     
     
         6 . The method of  claim 1 , further comprising:
 performing additional filtering in response to a classification that indicates that filtering of the other sensor data left noise signals in filtered data.   
     
     
         7 . The method of  claim 1 , further comprising:
 performing another filtering procedure with different parameters in response to a classification that indicates that filtering of the other sensor data caused distortion.   
     
     
         8 . The method of  claim 1 , further comprising:
 storing sensor data that is classified as acceptable on one or more non-transitory computer readable media to generate a geophysical data product.   
     
     
         9 . The method of  claim 1 , wherein the classification engine is a neural network or a support vector machine. 
     
     
         10 . The method of  claim 1 , wherein the sensor data is obtained by a geophysical survey based on sensor measurements of subsurface reflections of signals emitted by one or more survey sources. 
     
     
         11 . The method of  claim 10 , wherein the training and classifying are performed during the same geophysical survey. 
     
     
         12 . A non-transitory computer-readable medium having instructions stored thereon, where the instructions are executable by one or more processors to perform operations comprising:
 accessing sensor data;   applying one or more filtering procedures to the sensor data, wherein an output of the one or more filtering procedures is determined to have a threshold level of distortion to the sensor data or leave a threshold level of remaining noise signals in the sensor data;   training a classification engine based on the output of the one or more filtering procedures; and   classifying, using the trained classification engine, other filtered sensor data to determine whether the other filtered sensor data exhibits acceptable filtering.   
     
     
         13 . The non-transitory computer-readable medium of  claim 12 ,
 wherein the one or more filtering procedures include:
 a first filtering procedure, wherein an output of the first filtering procedure is determined to have a threshold level of distortion to the sensor data; 
 a second filtering procedure, wherein an output of the second filtering procedure is determined to leave a threshold level of remaining noise signals in the sensor data; and 
 a third filtering procedure, wherein an output of the third filtering procedure is determined to provide acceptable filtering outputs. 
   
     
     
         14 . The non-transitory computer-readable medium of  claim 12 ,
 wherein the training comprises:
 generating respective sets of attributes in an attribute space for the one or more filtering procedures based on determined similarity between:
 outputs of the one or more filtering procedures; and 
 differences between the outputs of the one or more filtering procedures and the sensor data; and 
 
 performing feature extraction to generate respective sets of features for the one or more filtering procedures in a feature space; and 
   wherein the classifying generates attributes and features for the other sensor data and determines which portions of the feature space correspond to the generated features.   
     
     
         15 . The non-transitory computer-readable medium of  claim 12 , wherein the operations further comprise:
 performing additional filtering in response to a classification that indicates that filtering of the other sensor data left noise signals in filtered data.   
     
     
         16 . The non-transitory computer-readable medium of  claim 12 , wherein the operations further comprise:
 performing another filtering procedure with different parameters in response to a classification that indicates that filtering of the other sensor data caused distortion.   
     
     
         17 . An apparatus for classifying sensor processed data from a marine geophysical survey, comprising:
 means for accessing sensor data;   means for applying one or more filtering procedures to the sensor data, wherein an output of the one or more filtering procedures is determined to have a threshold level of distortion to the sensor data or leave a threshold level of remaining noise signals in the sensor data;   means for training a classification engine based on the output of the one or more filtering procedures; and   means for classifying, using the trained classification engine, other filtered sensor data to determine whether the other filtered sensor data exhibits acceptable filtering.   
     
     
         18 . The apparatus of  claim 17 , wherein the training includes:
 generating respective sets of attributes in an attribute space for the one or more filtering procedures based on determined similarity between:
 outputs of the one or more filtering procedures; and 
 differences between the outputs of the one or more filtering procedures and the sensor data; and 
   performing feature extraction to generate respective sets of features for the one or more filtering procedures in a feature space; and   wherein the classifying generates attributes and features for the other sensor data and determines which portions of the feature space correspond to the generated features.   
     
     
         19 . In a technological method of filtering survey sensor measurements that includes:
 accessing, by a computing system, sensor data;   
       the specific technological improvement comprising:
 applying one or more filtering procedures to the sensor data, wherein an output of the one or more filtering procedures is determined to have a threshold level of distortion to the sensor data or leave a threshold level of remaining noise signals in the sensor data; 
 training, by the computing system, a classification engine based on the output of the one or more filtering procedures; and 
 classifying, by the computing system using the trained classification engine, other filtered sensor data to determine whether the other filtered sensor data exhibits acceptable filtering, thereby improving automated filtering quality control accuracy for the sensor data. 
 
     
     
         20 . The method of  claim 19 ,
 wherein the training comprises:
 generating respective sets of attributes in an attribute space for the one or more filtering procedures based on determined similarity between:
 outputs of the one or more filtering procedures; and 
 differences between the outputs of the one or more filtering procedures and the sensor data; and 
 
 performing feature extraction to generate respective sets of features for the one or more filtering procedures in a feature space; and 
   wherein the classifying generates attributes and features for the other sensor data and determines which portions of the feature space correspond to the generated features.