CN-121980330-A - Small sample class increment non-analytic indicator diagram fault diagnosis system based on serialization and contrast learning and diagnosis method thereof

CN121980330ACN 121980330 ACN121980330 ACN 121980330ACN-121980330-A

Abstract

The invention provides a small sample increment non-analytic indicator diagram fault diagnosis method based on serialization and contrast learning, which comprises the following steps of non-analytic indicator diagram serialization, contrast learning pre-training; the invention also provides a small sample increment non-analytic indicator diagram fault diagnosis system based on serialization and contrast learning by applying the method. According to the small sample increment non-analytic indicator diagram fault diagnosis system and the diagnosis method based on serialization and contrast learning, which are disclosed by the invention, the non-analytic indicator diagram information density is improved through a serialization method, and the efficient and stable dynamic fault diagnosis is realized by combining contrast learning pre-training, a lightweight adapter, sample replay and an NME classifier.

Inventors

ZHANG WENJUAN
Su tianyu
CHEN YOUYOU
YANG BINGHUI
YANG BOYU
LI WANJIE

Assignees

宝鸡文理学院

Dates

Publication Date: 20260505
Application Date: 20251222

Claims (9)

1. The small sample increment non-analysis indicator diagram fault diagnosis method based on serialization and contrast learning is characterized by comprising the following steps: (1) The non-analytic indicator diagram serialization is that background noise and redundant pixels are removed from the non-analytic indicator diagram through pretreatment, upper and lower contour points are extracted and ordered according to the abscissa, a one-dimensional sequence is formed, and data compression and effective feature preservation are realized; (2) Pre-training the contrast learning module by using a basic class training set, capturing multi-scale context representation of sequence data through layered contrast learning, and freezing module parameters after pre-training to obtain general feature representation; (3) In the increment stage, introducing a small sample set of a new fault class, combining a lightweight adapter and a sample replay mechanism, fine-tuning adapter parameters, and simultaneously performing fault classification by using an NME classifier so as to relieve catastrophic forgetting and improve diagnosis precision; (4) And (3) fault diagnosis, namely, for an input non-analytic indicator diagram, obtaining a feature vector through a feature extraction link after serialization, and finally outputting fault types by an NME classifier.
2. The small sample incremental non-analytic indicator diagram fault diagnosis method based on serialization and contrast learning, which is characterized by specifically comprising the steps of inputting a preprocessed image, extracting a foreground pixel point set, taking the minimum and maximum ordinate as upper and lower boundary points for each abscissa, sorting the upper boundary points in ascending order of the abscissa, sorting the lower boundary points in descending order, splicing the upper and lower boundary point sequences and normalizing the coordinates, and zero filling to the unified maximum length, outputting a one-dimensional sequence, wherein serialization compresses image data from about 10 to 3, and the data size is reduced by 99.4%.
3. The small sample increment non-analytic indicator diagram fault diagnosis method based on serialization and contrast learning according to claim 1 is characterized in that in contrast learning pre-training, a contrast learning module comprises an input projection layer, an expansion convolution block and an output convolution block, context is generated through random clipping during pre-training, positive and negative sample pairs are set by using context consistency, a loss function comprises index contrast loss and instance contrast loss, and overall loss is the sum of hierarchical contrast loss and is used for learning general characterization of sequence data.
4. The small sample increment non-analytic indicator diagram fault diagnosis method based on serialization and contrast learning of claim 1, wherein the lightweight adapter adopts a convolution module of a reverse pyramid residual error architecture, and comprises an initial convolution layer and a plurality of reverse pyramid residual error blocks; The adapter is trainable in parameters of the basic class task and the incremental task, is used for mapping the general features extracted by the freezing comparison learning module into task specific features, sequentially reduces the number of channels, and finally outputs the features to be classified through global average pooling.
5. The small sample class increment non-analytical indicator diagram fault diagnosis method based on serialization and contrast learning according to claim 1, wherein the sample replay mechanism selects representative examples from old class fault data to store in an example set based on herding priority example selection method; An example selection formula is: p k denotes the currently selected sample, k denotes the number of the current sample, X is the set of training samples of this type, Representing the characteristic vector of the sample, wherein mu is the average value of the characteristic vectors of the training samples; during incremental training, new sample loss and example set loss are simultaneously optimized.
6. The small sample class increment non-analytical indicator diagram fault diagnosis method based on serialization and contrast learning according to claim 1, wherein the NME classifier uses normalized feature vectors, and the classification principle is as follows: Mu y is an example mean vector of each category, namely, the sample feature vector of the example set belonging to the category is averaged, and the NME classifier avoids category conflict of a parameterized classification layer and supports lossless extension.
7. The small sample class increment non-analytical indicator diagram fault diagnosis method based on serialization and contrast learning according to claim 1, wherein the increment stage t is provided with a training set Where n t represents the total number of training set samples at stage t, As a training sample, the corresponding label is And t is not equal to t In learning the class increment stage t, only the training dataset of the current stage can be used Updating the model while testing the set at all known classes The learning ability is evaluated above, if f (x; θ) and Representing the classification model and the classification loss, respectively, the whole class increment learning objective can be expressed as:
8. The small sample increment non-analytical indicator diagram fault diagnosis method based on serialization and contrast learning according to claim 1, wherein the method is applied to oil pumping well non-analytical indicator diagram fault diagnosis, and fault categories comprise various of normal working conditions, liquid impact, gas interference, airlock, slow closing of a traveling valve, pump barrel split, leakage of a traveling valve, oil pipe leakage, pump clamping, insufficient liquid supply, shaft coupling broken rod and plunger separation of a pump barrel.
9. The utility model provides a little sample class increment non-analytical indicator diagram fault diagnosis system based on serialization and contrast study which characterized in that includes: The serialization module is used for converting the non-analytic indicator diagram into a one-dimensional sequence; The contrast learning pre-training module is used for pre-training and freezing parameters on the basic class training set to obtain general feature representation; the adapter module adopts a reverse pyramid residual error architecture and fine-tunes parameters in an incremental task; a replay module for storing and managing the example set, selecting a representative sample based on herding; an NME classifier for classifying based on the distance of the feature vector to the class mean; Wherein the system performs the method of any one of claims 1 to 8.

Description

Small sample class increment non-analytic indicator diagram fault diagnosis system based on serialization and contrast learning and diagnosis method thereof Technical Field The invention belongs to the technical field of industrial fault diagnosis, and particularly relates to a small sample increment non-analytic indicator diagram fault diagnosis model based on serialization and comparison learning, which is particularly suitable for high-efficiency utilization of small samples of a non-analytic indicator diagram and increment learning of dynamic fault categories in working condition fault diagnosis of an oil pumping well. Background The oil pumping well indicator diagram is an important tool for describing the working state of the oil pumping machine, and various typical faults can be effectively diagnosed by analyzing morphological characteristics of the oil pumping machine. However, in actual engineering, a large amount of historical indicator diagram data is stored only in image format, lacking displacement-load raw data, called a non-analytical indicator diagram. Such images contain a large amount of invalid pixels and redundant information, which not only increases the storage and computation burden, but also interferes with efficient extraction of key features by the model. The existing fault diagnosis method of the indicator diagram is mainly based on displacement-load original data or complete image data, and has three limitations when the non-analytic indicator diagram is processed, namely, firstly, the original sequence data cannot be utilized, secondly, the image is directly input to a deep learning network, so that calculation resources are wasted, fitting is easy to cause, and thirdly, the subjectivity of manual feature design is strong, and the information retention degree is low. In addition, the actual fault diagnosis faces two major challenges, namely, on one hand, the acquisition of certain fault type samples is difficult, the labeling cost is high, the training data amount is extremely small, and on the other hand, the fault types are dynamically increased along with the change of working conditions, and the model needs to continuously learn new types on the premise of not forgetting learned knowledge. The traditional deep learning model is difficult to adapt to the scene of data scarcity and category dynamic growth after training on a fixed data set, and is easy to cause 'catastrophic forgetting', namely the recognition performance of the old category is drastically reduced after learning the new category. The existing small sample increment learning method such as TOPIC(TOpology-Preservingknowledge InCrementer)、iCaRL(incremental Classifier and Representation Learning)、DER(Dynamically Expandable Representation) relieves the forgetting problem to a certain extent, but still has the problems of insufficient feature extraction, poor classifier adaptability, parameter redundancy and the like under the scene of a non-analytic indicator diagram, so that the diagnosis precision is not high and the stability is not sufficient. Object of the Invention The invention aims to provide a small sample class increment non-analytic indicator diagram fault diagnosis model based on serialization and contrast learning, which solves the problems of feature redundancy, disastrous forgetting and low diagnosis precision in non-analytic indicator diagram processing and small sample class increment learning of the existing method. Disclosure of Invention In order to achieve the above purpose, the invention provides a small sample increment non-analytic indicator diagram fault diagnosis system and a diagnosis method thereof based on serialization and contrast learning. The invention discloses a small sample increment non-analytic indicator diagram fault diagnosis system based on serialization and contrast learning, which comprises the following steps: The serialization module is used for converting the non-analytic indicator diagram into a one-dimensional sequence; The contrast learning pre-training module is used for pre-training and freezing parameters on the basic class training set to obtain general feature representation; the adapter module adopts a reverse pyramid residual error architecture and fine-tunes parameters in an incremental task; a replay module for storing and managing the example set, selecting a representative sample based on herding; an NME classifier for classifying based on the distance of the feature vector to the class mean; the system executes a small sample increment non-analytic indicator diagram fault diagnosis method based on serialization and contrast learning, and the method comprises the following steps: (1) The non-analytic indicator diagram serialization is that background noise and redundant pixels are removed from the non-analytic indicator diagram through pretreatment, upper and lower contour points are extracted and ordered according to the abscissa, a one-dimensional sequen