An Intelligent Fault Diagnosis Method that Combines Operating Condition Attribute Encoding and Multi-scale Cascading Concept

In order to improve the generalization ability of the model under different working conditions and the robustness of intelligent fault diagnosis, and learn a broader feature representation, this paper proposes an intelligent fault diagnosis method that integrates working condition attribute encoding and multi-scale cascade concepts. This method integrates working condition information into vibration data by introducing methods such as working condition attribute coding, multi-scale cascade modules and quadruple losses, and effectively extracts invariant features. this method trains the fault classification model through a three-stage training process. Finally, the objective is to accomplish fault diagnosis in diverse operational scenarios. Experimental results show that this method improves the fault diagnosis accuracy across diverse operational conditions, indicating that the model has good generalization ability.


Introduction
Over the last two to three decades, intelligent fault diagnosis has emerged as a promising avenue for alleviating human labor, garnering considerable attention.Traditional machine fault diagnosis methods [1], [2], [3], [4] are difficult to establish accurate data models.In terms of equipment fault diagnosis applications Very limited.Due to the powerful feature extraction capability of deep learning, it can extract features of deep-level diversity of data [5], [6].The growing focus on the utilization of intelligent computing techniques, particularly deep learning models, in the field of mechanical fault diagnosis can be attributed to the advancements in deep learning model development [7].The working conditions change all the time, and bearing fault diagnosis in diverse operational scenarios is a challenging problem [8], because changing working conditions may lead to changes or fuzzification of the failure mode.It brings some difficulties and dilemmas to the diagnostic process.At present, various fault diagnosis algorithms are currently employed to address the challenges posed by changing working conditions [9].
Fault diagnosis methods based on improved CNN models [10], [11], [12].This type of fault diagnosis method improves the model's ability to extract features by improving the model structure of the neural network.However, there are some shortcomings in solving fault diagnosis based on the improved CNN model.Given that neural network models typically demand a quantity of data, getting and labeling data may pose increased challenges in the context of diagnosing faults under variable working conditions.In addition, improved neural network models often need to be retrained or finetuned for new working conditions and lack good generalization capabilities.
This transfer learning method [13], [14], [15] enhances fault diagnosis performance in novel working conditions.In [16], researchers proposed a new cross-domain fault diagnosis technology based on deep GAN, which performs domain adaptation by synthesizing fake samples.In reference [17], a domain adaptation module was introduced to assist the model in acquiring features that are invariant across domains.Reference [18] employs a three-layer sparse autoencoder to extract features from the original data.However, the model in the source domain may not be able to adapt well to changes in the target domain, resulting in performance degradation.
Based on meta-learning and unsupervised fault diagnosis related methods [19], [20], [21], this type of method helps the model better adapt to new tasks by learning how to learn from past tasks.[19] proposed a method MTLSAM based on a combination of meta transfer learning and self-attention mechanism.[21] proposed a small-sample rolling bearing fault intelligent identification method based on meta-learning.First, a deep convolutional neural network is used to perform deep adaptive extraction of fault features in small sample original signals, and then the improved model method is used to train and optimize the model parameters.Meta-learning and unsupervised fault diagnosis methods face some challenges and shortcomings.First, although meta-learning can help the model adapt to fault diagnosis tasks under different working conditions, in practical applications, the design and training of meta-learning methods may be more complex and time-consuming.
Given the aforementioned challenges, this paper proposes an intelligent fault diagnosis method that integrates working condition attribute encoding and multi-scale cascade ideas.By introducing working condition attribute encoding, multi-scale cascade group attention module and quadruple loss, this method can better integrate working condition information and extract universal invariant features under different changing working conditions.By decoupling operating condition information and effectively utilizing multi-scale features, this method can improve the accuracy and robustness of fault diagnosis.This fusion idea enables the model to better adapt to fault diagnosis tasks under different working conditions, thereby improving the effectiveness and reliability in practical applications.
The main contributions of this study are as follows: 1) This method incorporates the quadruplet loss as a metric learning approach in the fault diagnosis model.The model learns more dis-criminative feature representations.This strategy improves the accuracy and reliability of fault diagnosis.
2) This method adopts a three-stage training mode, allowing the model to learn more generalized feature representations and enhancing the generalization capability and robustness of intelligent fault diagnosis under variable operating conditions.

Methodology
This method is designed to address the challenges of fault diagnosis under different operating conditions.This method improves the feature extraction layer and enhances the feature extraction capability of the model by introducing the working condition attribute encoding layer and the multiscale cascade module.In addition, this method adopts a quadruple metric learning method to guide the model to learn domain-invariant features that are not affected by operating conditions, thereby improving the accuracy and reliability of fault diagnosis.The model architecture diagram is shown in figure 1.

Quadruple Sample Pair Construction
In order to construct quadruple sample pairs, vibration data need to be organized according to certain rules.Each sample pair consists of four time series of vibration data, including 3 categories: Anchor (anchor), Positive (positive sample), Negative1 (negative sample 1) and Negative2 (negative sample 2).Anchor and Positive belong to the same category, while Anchor, Negative1, and Negative2 represent different categories of data respectively.

Feature Extraction Layer
The feature extraction layer consists of two parts, including a module combining Embedding and Condition Coding and a Multi-scale expansion module.The Condition Coding layer includes a fully connected layer and an activation layer, whose function is to encode the working condition information corresponding to the vibration data.The Embedding layer consists of a fully connected layer and an activation layer, which is used to encode and embed the original data of the vibration data at each moment.
Multi-scale expansion module includes Multi-scale segmentation layer and Expansion module layer.The Multi-scale segmentation layer consists of three convolutional layers, activation layers and pooling layers, and is used to perform multi-scale transformation on the data encoded by the module combining Embedding and Condition Coding.Using the sizes of different CNN convolution kernels and pooling kernels, we obtain transformed data of the same scale as the data encoded in the previous step, and two sets of transformed data with half the scale.At the same time, nonlinear transformation of data is achieved through this process.
The Expansion module layer contains a combination of 3 basic convolutional layers.These three groups of convolutional layer modules perform convolution operations on multi-scale data respectively.By concatenating the output of the first convolution group to the input of the second convolution group, and concatenating the output of the first convolution group and the output of the second convolution group to the third convolution group input to achieve the fusion of multi-scale information.Finally, the output results of the three groups are spliced to obtain the final output of the multi-scale cascade group module layer in stage 1.Stage 2 and stage 3 perform the same operation, and finally obtain the feature output of the feature extraction layer.

Feature Extractor and Classifier Tail
The tail of the feature extractor consists of fully connected layers.By introducing the quadruple loss function, similarity learning is used to recombine the features of the sample pairs and calculate the quadruple loss.The tail of the classifier consists of a fully connected layer, which is used to calculate multi-classification losses on features extracted through the feature extraction layer.The classification results calculated through the fully connected layer are used to predict the fault category of the sample.

Experiment
First paragraph.In this study, we use the Case Western Reserve University (CWRU) bearing data set to validate the proposed method.Using data sampled by a single-point drive end at a speed of 48,000 samples/second, the data has four rotational speeds: 1797rpm, 1772rpm, 1750rpm and 1730rpm.We adopted a three-stage training approach when conducting all fault diagnosis experiments.There were 100 training sessions in the first phase, 100 training sessions in the second phase, and 100 training sessions in the third phase.
In the experiment, we used a computer with a 64-bit Ubuntu18.04system.We used NVIDIA TITAN V model graphics cards and used the PyTorch deep learning framework to conduct experiments.The method in this article is the idea of working condition attribute encoding and multi-scale cascade introduced in the basic convolution model in table 1.Therefore, this experiment selected the model structure shown in the table below as the comparison model.In the following, CNN-BASE is used to represent the basic convolution model.From the above comparative experiments, it is evident that the approach presented in this paper exhibits distinct advantages when compared to the CNN-based method in the context of variable working conditions.Since the feature extraction layer structure of CNN-base is consistent with the proposed method, the model proposed in this paper introduces the idea of working condition attribute coding and multi-scale cascade on this basis.Compared with the traditional methods in the references [2][3], the fault diagnosis accuracy of the proposed method is greatly improved.Compared with the improved convolutional neural network [10], the accuracy of fault diagnosis has also been improved.The findings demonstrate the effectiveness of the method presented in this paper in significantly enhancing the accuracy of fault diagnosis in the presence of varying speeds.

Discussion and Conclusion
This paper proposes an intelligent fault diagnosis method that integrates working condition attribute coding and multi-scale cascade ideas.This method integrates working condition information into vibration data by introducing working condition attribute coding, multi-scale cascade modules and quadruple losses, and effectively extracts invariant features under changing working conditions.The design of this fusion approach has certain in-depth implications.The introduction of the working condition attribute coding layer can effectively combine the working condition information into the feature extraction process, thereby improving the accuracy of the model's fault diagnosis.At the same time, the introduction of multi-scale cascade modules can capture feature information at different scales, allowing the model to understand and express data more comprehensively.In addition, by using quadruple loss for metric learning, more discriminative feature representations can be learned, thereby improving the accuracy and reliability of fault diagnosis.In the experiment, we trained the fault classification model through a three-stage training mode.Experiments have shown that the method in this paper can significantly improve the accuracy of fault diagnosis under variable working conditions.In this way, this method enhances the accuracy and robustness of the model in fault diagnosis under varying working conditions, allowing the model to perform fault diagnosis more reliably in complex actual working conditions.

Table 1 .
Convolutional model structure of CNN-BASE.We used the CNN-BASE model and a fault diagnosis classification model trained under specific operating conditions to make predictions on data from different operating conditions.The training dataset consisted of 100 samples.The test results are shown in table 2.

Table 2 .
The experimental results of CNN-BASE model.The proposed method in this paper trains the feature extractor on multiple operating condition conditions and trains the fault classifier on a specific operating condition.The training dataset consists of 100 samples.The test results are shown in table3.

Table 3 .
The experimental results of the method presented in this paper.