CN-117252860-B - Low-strain waveform anomaly detection method based on neural network and sawtooth matching

Abstract

The invention relates to a low-strain oscillogram anomaly detection method based on deep neural network and sawtooth point matching, which utilizes an OCR pre-training model, straight line detection and a DB-scan clustering algorithm to obtain a preprocessed image and mapping of a pixel coordinate system and a real length coordinate system; waveform sequence data is obtained through sawtooth point matching, acceleration filtering and one-dimensional linear interpolation, then the waveform data is sent into a deep learning model to generate abnormal prediction, and then data post-processing is carried out to correct the result of the data and convert the format. The method has the advantages that priori knowledge can be utilized to the greatest extent to help improve the accuracy of the method, the waveform diagram range and scale axis digital information are determined through image preprocessing, effective priori in the image is reserved to the greatest extent when the image information is compressed, and the influence of coordinate axes, un-erased texts and vertical dotted lines can be basically eliminated by using the method based on saw-tooth point matching for serialization, so that the method has stronger robustness.

Inventors

• ZHU KEHONG
• XU GUODONG
• XIE QINFANG
• FENG JIANGYU
• XUE XIANGYANG
• LI HOURONG
• LI BIN
• ZHANG DONGFANG
• ZHU YINGLIAN
• NIE LEI
• ZHENG YAOHUA
• LI XIANQING

Assignees

• 中铁二十四局集团有限公司

Dates

Publication Date

20260512

Application Date

20231031

Claims (1)

1. 1. The low-strain waveform abnormality detection method based on neural network and sawtooth matching is used for predicting abnormality in a picture derived by a low-strain machine and is characterized by comprising the following steps of: S1, carrying out picture pretreatment on a picture; S2, carrying out waveform serialization on the preprocessed picture, wherein the waveform serialization refers to finding all sawtooth points in a slice through sawtooth point matching, filtering abnormal points through an acceleration threshold value, obtaining serialization data through linear interpolation, normalizing y-axis coordinates of the sequence according to upper and lower boundaries, and attaching a real physical coordinate system of a corresponding position to correspond to x-axis coordinates; S3, sending the sequence after picture serialization into a low-strain waveform chart abnormality detection model which completes final training, generating a confidence coefficient sequence on the whole sequence according to the sequence by the low-strain waveform chart abnormality detection model, and selecting the type with the maximum probability value as the prediction classification of the point on the sequence; S4, performing post-processing on the prediction classification, removing errors below a standard axis to classify the prediction classification into abnormal-free cases, merging error sequences with continuous lengths larger than 5 pixels, and taking a middle point as a real physical coordinate position corresponding to an error position to output; The picture preprocessing comprises the steps of obtaining all text positions in a picture through trained text detection and recognition models, obtaining approximate positions of coordinate axes through a DB-scan clustering algorithm by taking y-axis coordinates as clustering indexes, detecting all straight lines with the length being more than 50% of the width of the picture through horizontal straight line detection, selecting the nearest straight line to the y-axis coordinates of a clustering center as the coordinate axes, obtaining upper and lower boundaries of a waveform diagram by taking the straight lines as the coordinate axes, determining left and right ranges of the waveform diagram according to a first vertical dotted line and a last vertical dotted line in the picture, obtaining mapping relations between a pixel coordinate system and a real length coordinate system according to a clustering result, erasing the Chinese in the picture according to the text detection result, cutting according to the boundary of the waveform diagram, and binarizing; The final training of the low strain oscillogram anomaly detection model comprises the following steps: A step of preparing data construction, in which error labeling is carried out on the sequence by manual labeling, and the data is divided into a training set and a testing set; Training an anomaly detection model, namely training a preset bidirectional LSTM network by adopting a supervised network based on the training set and the training label; A termination condition judging step, namely entering a detection model training step until a preset termination condition is reached; And outputting, namely taking the bidirectional LSTM network which completes the network training at the moment as the low-strain waveform chart abnormality detection model which completes the final training, wherein the termination condition is that the prediction result output by the detection model training step is a correct label and the confidence is all 1 or the detection model training step is operated for 50 times.

Description

Low-strain waveform anomaly detection method based on neural network and sawtooth matching Technical Field The invention relates to the technical field of image processing, in particular to a low-strain waveform anomaly detection method based on neural network and sawtooth matching. Background In recent years, deep learning technology has been rapidly developed and applied to various actual scenes. However, no method for detecting anomalies in low-strain oscillogram images is currently available on the market. The method mainly comprises the following steps of (1) unfixed curve positions in a low-strain waveform chart, (2) broken lines, coordinate axis scales and character interference exist in an image, (3) the proportion of pixels containing effective information in the low-strain waveform chart to all pixels of the whole image is not high, and (4) the waveform chart is wrong and has no fixed visual mode. Because of the difficulty (4) of using deep learning to fit the distribution of outlier data from among the data, it is not straightforward to detect with conventional expert systems. And the signal-to-noise ratio of the data is too large because of the difficulty (2) (3) that the picture is directly input as a model. Disclosure of Invention According to the defects of the prior art, the invention provides a low-strain waveform abnormality detection method based on neural network and sawtooth matching, and the abnormality in an image derived by a low-strain machine is predicted through the combination of a neural network model and the sawtooth matching. The invention is realized by the following technical scheme: the low-strain oscillogram anomaly detection method based on deep neural network and sawtooth point matching is used for predicting anomalies in images derived by a low-strain machine and is characterized by comprising the following steps of: The picture preprocessing method comprises the specific steps of obtaining all text positions through pre-trained text detection and recognition models, obtaining coordinate axis approximate positions through a DB-scan clustering algorithm by taking y-axis coordinates as clustering indexes, detecting all straight lines with the length being greater than 50% of the picture width through horizontal straight line detection, selecting the nearest straight line to the y-axis coordinates of a clustering center as the coordinate axis, obtaining upper and lower boundaries of a oscillogram according to the upper and lower boundaries, and determining left and right ranges of the oscillogram according to a first vertical dotted line and a last vertical dotted line in the picture. And simultaneously, obtaining the mapping relation between the pixel coordinate system and the real length coordinate system according to the clustering result. And erasing the Chinese characters in the graph according to the text detection result, cutting according to the waveform graph boundary, and finally binarizing. Step 2, carrying out waveform serialization on the preprocessed picture, wherein the specific steps are that all saw-tooth points (including upward and downward) in the picture are found through saw-tooth point matching, and meanwhile, some abnormal points are filtered through an acceleration threshold value. And obtaining the serialized data through linear interpolation, normalizing the y-axis coordinates of the sequence according to the upper and lower boundaries, and attaching a real physical coordinate system corresponding to the corresponding position to the x-axis coordinates. And 3, sending the sequence after picture serialization into a low-strain waveform chart abnormality detection model which completes final training, generating a confidence coefficient sequence on the whole sequence according to the sequence by the low-strain waveform chart abnormality detection model, and selecting the type with the maximum probability value as the prediction classification of the point on the sequence. The final training comprises the following steps of preparing data construction, manually labeling the sequences for error labeling, and dividing the data into a training set and a testing set. An anomaly detection model training step of performing supervised network training on a preset bidirectional LSTM network based on the training data set and a training label, a termination condition judging step of entering a detection model training step until reaching a preset termination condition, an output step of taking the BiLSTM network which completes the network training at this time as the anomaly detection model of the low-strain waveform chart which completes the final training, wherein the termination condition is that the prediction result output by the detection model training step is a correct label and the confidence is all 1 or the detection model adjusting step is operated for 50 times And 4, post-processing the predicted data, namely, carrying out post-processing on the