Multi-condition fault diagnosis methods, systems, media, and products based on feature similarity measurement and cross-domain equilibrium

Abstract

This invention belongs to the field of mechanical vibration fault diagnosis, and specifically discloses a multi-condition fault diagnosis method, system, medium, and product based on feature similarity measurement and cross-domain equilibrium. It trains a neural network model using fault signals collected under different operating conditions. During training, a similarity measurement module and a cross-domain equilibrium module are designed to enhance the model's classification ability against domain interference and its ability to extract domain-invariant features. The similarity measurement module constructs a prototype for each fault category and uses prototype contrast loss to narrow the distance between samples of the same category under different operating conditions, improving the similarity between similar samples. The cross-domain equilibrium module achieves a balance in the model's performance across different operating condition datasets by minimizing the model's maximum explained variance under different operating conditions, thus improving robustness. This invention can not only identify fault types under preset operating conditions but also accurately identify faults under unknown operating conditions in actual production environments.

Description

Technical Field

[0001] This invention belongs to the field of mechanical vibration fault diagnosis, and more specifically, relates to a multi-condition fault diagnosis method, system, medium and product based on feature similarity measurement and cross-domain equilibrium. Background Technology

[0002] With the increasing complexity and automation of industrial equipment, the potential impact of malfunctions, if not addressed promptly, is becoming increasingly severe. Malfunctions not only hinder normal equipment operation, reducing production efficiency and capacity, but also easily lead to product quality problems and increase defect rates. More seriously, some malfunctions may cause safety issues, such as worker injuries or hazardous working environments, exacerbating accident risks. Therefore, employing advanced fault diagnosis technologies to accurately and promptly identify equipment failure modes helps improve the reliability, efficiency, and safety of industrial production.

[0003] Currently, data-driven algorithms are widely used in fault diagnosis technology. These algorithms can learn relevant knowledge from massive amounts of historical data and can model faults without requiring precise process information. With the rapid development of IoT technology and computing hardware, deep learning technology has received widespread attention and has been introduced into fault diagnosis systems. Deep learning algorithms automatically learn and extract data features through multi-layer neural networks to achieve end-to-end fault diagnosis, ensuring high-precision and high-efficiency diagnostic results.

[0004] However, in actual production, the operating environment of equipment is constantly changing. Different motor speeds, workloads, and other factors can alter the data distribution of equipment fault modes, leading to changes in the distribution of collected vibration signal data. This causes domain drift, affecting the diagnostic performance of the trained model in practical applications. In other words, in real-world applications, fault modes with unknown data distributions may occur, making it difficult for models based on training data to perform accurate fault diagnosis. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a multi-condition fault diagnosis method, system, medium and product based on feature similarity measurement and cross-domain balance. Its purpose is to achieve cross-domain fault diagnosis with high accuracy and adaptability to different data distributions.

[0006] To achieve the above objectives, according to a first aspect of the present invention, a multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium is proposed, comprising the following steps:

[0007] Model training phase:

[0008] The neural network model is trained using a training set, and the trained neural network model is used as a fault diagnosis model.

[0009] The training set includes several source domain datasets obtained under different working conditions. Each source domain dataset includes multiple data points, and each data point includes a fault signal sample and a true label representing its corresponding fault category.

[0010] The neural network model includes a feature extraction module and a classifier. The feature extraction module is used to extract feature vectors from the fault signal, and the classifier is used to predict the probability that the fault signal belongs to each fault category based on the feature vectors.

[0011] Loss function used in training as follows:

[0012]

[0013] in, This is the loss function for the feature extraction module, used to measure the difference between the predicted values ​​output by the model and the true label values; This is the loss function for the similarity measurement module, used to make samples of the same category as close as possible to their corresponding class's prototype. λ1 and λ2 are the cross-domain balancing module loss functions, used to minimize the variance of the model's classification error for samples from different source domains; λ1 and λ2 are the weight hyperparameters; x and y are the sample and true labels, respectively, and θ is the model parameter;

[0014] Model application phase:

[0015] The fault signal to be diagnosed is input into the fault diagnosis model, and the fault diagnosis model outputs the probability of it belonging to each fault category, thereby determining the fault category corresponding to the fault signal.

[0016] As a further preferred option, the similarity measurement module loss function Specifically:

[0017]

[0018] in, This represents all samples in class k of the training set. S represents all samples in the training set except for class k; kj For the j-th sample x j Prototype compared to category k The similarity between samples is given by γ, which is the boundary threshold; α and β represent hyperparameters, and M represents the total number of samples.

[0019] As a further preferred embodiment, the method for determining the boundary threshold γ is as follows:

[0020] Pre-define the corresponding comparison prototype for each fault category; for each category, take the feature vector of the sample farthest from the comparison prototype of that category as the category boundary sample, and use the distance between the category boundary sample and the comparison prototype as the boundary threshold γ.

[0021] As a further preferred method, cosine similarity is used to calculate sample x. j Compared to the prototype Similarity S between kj .

[0022] As a further preferred option, the cross-domain equilibrium module loss function Specifically:

[0023]

[0024] Where n represents the total number of source domains, S represents the i-th source domain. i The cross-entropy loss function of the sample, where c represents the region belonging to the source domain S. i The sample, and Let f(·) represent the feature vector and true label value of sample c, respectively, and let f(·) represent the predicted value output by the neural network model. Indicates all The average value.

[0025] As a further preferred option, the loss function of the feature extraction module Specifically:

[0026]

[0027] As a further preferred embodiment, the neural network model employs the 1DResNet-18 network.

[0028] According to a second aspect of the present invention, a multi-condition fault diagnosis system based on feature similarity measurement and cross-domain equilibrium is provided, including a processor, the processor being used to execute the above-described multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium.

[0029] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium.

[0030] According to a fourth aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the above-described multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium.

[0031] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0032] 1. This invention introduces a similarity measurement module and a cross-domain equalization module into the loss function when training the fault diagnosis model to enhance the model's classification ability against domain interference and its ability to extract domain-invariant features. When performing diagnosis, the resulting fault diagnosis model can not only identify fault types under preset operating conditions but also accurately identify faults under unknown operating conditions in actual production environments. This invention eliminates the need for complex signal transformation and feature extraction preprocessing steps on the original signal, demonstrating strong adaptability to changes in operating conditions and environment. It can effectively reduce labor costs and improve the automation and intelligence level of equipment quality inspection processes.

[0033] 2. The similarity measurement module designed in this invention constructs a category prototype for each fault category. By using the designed prototype contrast loss, it narrows the distance between samples of the same category under different working conditions and improves the similarity between samples of the same category. The cross-domain equilibrium module is inspired by the theory of minimizing invariant risk. By minimizing the maximum explanatory variance of the model under different working conditions, it achieves the balance of the model's performance under different working condition datasets and improves robustness. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the function of the similarity measurement module in an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram illustrating the function of the cross-domain balancing module in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the fault diagnosis model structure according to an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the cross-operating condition fault diagnosis method according to an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0039] This invention provides a multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium, comprising the following steps:

[0040] (1) Model training and testing phase:

[0041] A training set and a neural network model are pre-built. The training set includes several source domain datasets acquired under different working conditions. Each source domain dataset includes multiple samples, each containing a fault signal and its corresponding fault category label. The fault signal can be a mechanical vibration signal, an electrical signal, or an acoustic signal, etc. In this embodiment, the vibration signal of rotating machinery is used. The neural network model includes a feature extraction module and a classifier. After the fault signal is input, the feature extraction module extracts the fault signal features, and the classifier outputs the probability that the extracted feature signal belongs to each fault category. The neural network model preferably uses the 1DResNet-18 network.

[0042] The neural network model was trained using a training set, and a similarity measurement module and a cross-domain equalization module were designed to enhance the model network's classification ability against domain interference and its ability to extract domain-invariant features during the training process; the trained neural network model was then used as a fault diagnosis model.

[0043] Specifically:

[0044] (1-1) Similarity Measurement Module

[0045] In the training set, source domain data from different operating conditions exhibit significant distribution differences, leading to potentially large distances between samples belonging to the same category but originating from different source domains within the feature space. Therefore, this invention designs a similarity measurement module to align the distribution of samples of the same category from different data distributions based on the similarity of sample features (e.g., ...). Figure 1 As shown in the diagram, this module enables the model to learn strong generalization features applicable to different source domain data distributions. It combines the ideas of prototype learning and comparative learning.

[0046] In prototype learning, each category is represented by one or more prototypes, which can be the mean of data points, the center point, or a representative sample. New samples are classified into the corresponding category by comparing their similarity to these prototypes. For the prototype O of the k-th class... k Euclidean distance or cosine similarity is used to measure the similarity between sample x and prototype O. k Distance between:

[0047]

[0048] d d (Θ(x),O k )=Θ(x)·O k #(2)

[0049] Where Θ(x) represents the feature vector of sample x, and m represents the dimension of the feature vector of sample x. e and d d Let Euclidean distance and cosine similarity be represented, respectively.

[0050] The prototype loss can be expressed as:

[0051]

[0052] Where, loss pro P represents the prototype loss of sample x. y Let Θ(x) represent the prototype of the class to which sample x belongs, and let Θ(x) represent the feature map obtained after sample x passes through the feature extraction module. This loss can make samples of each class as close as possible to the prototype of that class, thereby reducing the intra-class distance.

[0053] The core idea of ​​contrastive learning is to bring the representations of similar samples (positive samples) closer together and push the representations of dissimilar samples (negative samples) further apart, thereby enabling the model to learn more discriminative feature representations. In each training sample, three corresponding samples are selected: anchor sample A, positive sample P, and negative sample K, forming a triplet (A, P, K). Here, P represents a sample similar to anchor sample A, usually another sample of the same class as A, and K represents a sample dissimilar to anchor sample A, usually a sample of a different class than A. The goal of triplet loss is to minimize the distance (similarity) between anchor sample A and positive sample P, while maximizing the distance (similarity) between anchor sample A and negative sample K. This objective is achieved through the following loss function:

[0054]

[0055] Here, α is a predefined edge threshold to prevent the model from learning the same positive and negative sample features. That is, if no threshold α is set, as long as d(A,P)=d(A,K), the loss function value will be 0, and the model will be difficult to train.

[0056] The triplet approach considers contrastive loss at the sample level, ignoring other samples between the positive sample P and the negative sample K. These samples still contain a significant amount of information to guide the model, which can negatively impact model performance. This invention, combining prototype learning, considers contrastive loss at the category level, rather than the sample level. It takes all difficult samples into account during training, greatly increasing the information content of the samples during training, as detailed below:

[0057] First, a comparison prototype was designed for each category. (N is the total number of categories in the training set). For category k, during model iteration, each time the model is compared to the prototype of that category. The most recent samples from other categories (negative samples K), and the prototype of the comparison of this category. The most distant sample of the same class (positive sample P) and all samples between these two samples are optimized simultaneously. Through optimization formula (5), samples of the same class in the training set can be made as close as possible to their corresponding contrast prototypes, improving the similarity between samples and ensuring better intra-class compactness of samples from different source domains at the class level, thus obtaining a better decision boundary. The loss function corresponding to the similarity measurement module is:

[0058]

[0059] in, This represents all samples in the training set under fault category k. S represents all samples in the training set except for class k; M represents the total number of samples; S kj For sample x j Prototype compared to category k The similarity between samples is calculated using cosine similarity in this embodiment, as shown in formula (6). γ is the boundary threshold to prevent the model from learning the same positive and negative sample features. α and β represent the hyperparameters of positive and negative sample pairs, respectively. In this embodiment, α = 2 and β = 50.

[0060]

[0061] The similarity measurement module considers all difficult samples during training, but it considers the contrast loss from the category level rather than the sample level. Therefore, the algorithm's computation speed is slightly slower than that of triple loss, but the information that the samples can provide during training is greatly increased.

[0062] (1-2) Cross-domain load balancing module

[0063] When a bearing operates, it rotates at a certain speed. The rolling elements pass over the faulty part of the bearing at specific time intervals, generating periodic frequency characteristics. Therefore, for each fault mode, its vibration theoretically manifests as a vibration at a specific frequency based on the bearing's operating frequency, and the collected vibration signal should have a distinct theoretical frequency. Inner ring fault f i Outer ring fault f o and rolling element failure f b The theoretical failure frequencies are as follows:

[0064]

[0065] In equations (7)-(9), n represents the number of rolling elements contained in the bearing, f represents the inner ring speed of the bearing, d represents the diameter of a single rolling element of the bearing, D represents the pitch circle diameter of a single rolling element of the bearing, and α represents the contact angle.

[0066] However, in actual operation, bearings do not operate as standalone parts; they participate in the operation of the equipment as one of its components. Therefore, the collected vibration signals are not only caused by the vibration of the bearing itself, but also contain a large number of environmental signals. That is, the vibration signals collected by the three-dimensional vibration sensor consist of two parts: fault frequency characteristics and environmental interference characteristics.

[0067] The failure frequency characteristics of a bearing are only affected by its parameters and failure modes, and are independent of the environment and operating conditions in which the bearing operates. Therefore, for datasets collected under different operating conditions, samples of the same category will exhibit the same failure frequency characteristics, but the environmental interference characteristics will differ.

[0068] Inspired by the idea of ​​minimizing invariant risk, this invention proposes a cross-domain balancing module, aiming to find a predictor that achieves Bayesian optimality under all operating conditions. The core idea of ​​the cross-domain balancing module is to balance the model's performance under different source domain data, while learning the minimum error of fault classification and the invariance under different source domain distributions. That is, to make the features extracted by the model as similar as possible to the fault frequency features, while ignoring the influence of environmental interference features.

[0069] Given a dataset collected under various working conditions as the training set S train ={S1,S2,S3,……,S n The dataset collected under each operating condition is called a source domain S. i Each has an independent and identically distributed sample. For source domain S i The sample feature vector below, For source domain S i The sample labels are given below. Introducing the definition of invariant risk minimization, the risk of any source domain is defined as:

[0070]

[0071] In equation (10), L represents the loss function, reflecting the predicted value. and the true value The differences between them; For source domain S i The risk function is determined by the cross-domain balancing module, which aims to find a mapping method that ensures the classifier performs identically across all source domain mappings. In statistical terms, this means minimizing the variance of the model's classification error across different source domains. Therefore, the loss function corresponding to the cross-domain balancing module is... for:

[0072]

[0073] in, For training batches, the source domain S i The cross-entropy loss function value of the corresponding sample. The goal is to minimize the variance of the cross-entropy loss values ​​across different source domains; c represents the number of training batches belonging to source domain S. i The sample, and Let represent the feature vector and the true label of sample c, respectively.

[0074] By minimizing The model exhibits minimal differences in classification performance across different source domains, meaning the extracted features demonstrate consistent classification performance across any known source domain. This ensures that the model focuses on domain-invariant features such as fault frequency. Figure 2 As shown, the cross-domain balancing module can correct the sample distribution, making the class decision boundaries under different source domain distributions approach the same.

[0075] (1-3) Model Training and Testing

[0076] Taking into account the losses of each module, such as Figure 3 As shown, for a neural network model, its overall loss function during training... as follows:

[0077]

[0078] Where λ1 and λ2 are and The weight hyperparameter. This is the cross-entropy loss function of the model in the feature extraction module, used to measure and optimize the difference between the probability distribution of the model output and the true label. Combined with equation (12), The calculation formula is:

[0079]

[0080] Specifically, the neural network model training and testing process is as follows: Figure 4 As shown:

[0081] During the training phase, the N source domain datasets collected from different working conditions are first divided. For each source domain dataset, 80% of the samples are randomly selected to form the training set, and the remaining 20% ​​of the samples form the validation set. During model training, one of the source domain datasets is selected as the model's validation set, and the remaining N-1 source domain datasets participate in the model training.

[0082] During the validation phase, the 20% validation set defined above is input into the finally trained model. The distance between the high-discriminative feature vector extracted by the model for each sample and the prototype corresponding to its category in the similarity measurement module is calculated. For each category, the feature vector of the sample farthest from the category prototype is taken as the category boundary sample, and the distance between the boundary sample and the category prototype is taken as the category boundary threshold γ. This threshold can determine whether a sample belongs to that category.

[0083] During the testing phase, test samples under unknown working conditions are input into the trained model to obtain the predicted class labels output by the model, and the distance between the high discriminative feature vector of the sample and the prototype corresponding to its predicted label is calculated. If distance If the value is less than the category boundary threshold γ corresponding to the predicted label, it indicates that the model belongs to the predicted failure mode. This is used to determine the effectiveness of the trained model.

[0084] The neural network model that has undergone the above training, verification, and testing will be used as the final fault diagnosis model.

[0085] (2) Model application stage:

[0086] The fault signal to be diagnosed under unknown operating conditions is input into the fault diagnosis model. The fault diagnosis model outputs the probability that it belongs to each fault category, thereby determining the fault category corresponding to the fault signal.

[0087] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium, characterized in that, Includes the following steps: Model training phase: The neural network model is trained using a training set, and the trained neural network model is used as a fault diagnosis model. The training set includes several source domain datasets obtained under different working conditions. Each source domain dataset includes multiple data points, and each data point includes a fault signal sample and a true label representing its corresponding fault category. The neural network model includes a feature extraction module and a classifier. The feature extraction module is used to extract feature vectors from the fault signal, and the classifier is used to predict the probability that the fault signal belongs to each fault category based on the feature vectors. Loss function used in training as follows: in, This is the loss function for the feature extraction module, used to measure the difference between the predicted values ​​output by the model and the true label values; This is the loss function for the similarity measurement module, used to make samples of the same category as close as possible to their corresponding class's prototype. This is the cross-domain balancing module loss function, used to minimize the variance of the model's classification error for samples from different source domains; , These are weight hyperparameters; , These are the sample and the real label, respectively. These are model parameters; Similarity measurement module loss function Specifically: in, Indicates the categories in the training set All samples below, In addition to the categories, the training set also includes... All samples under other categories; For the first j Sample With category Contrast Prototype Similarity between them For boundary thresholds; , Indicates hyperparameters, Indicates the total number of samples; Cross-domain equilibrium module loss function Specifically: in, This represents the total number of source domains. Indicates the first i Source domain The cross-entropy loss function of the samples. Indicates belonging to the source domain The sample, and Representing samples respectively The feature vectors and true label values, This represents the predicted value output by the neural network model. Indicates all The average value; Loss function of feature extraction module Specifically: Model application phase: The fault signal to be diagnosed is input into the fault diagnosis model, and the fault diagnosis model outputs the probability of it belonging to each fault category, thereby determining the fault category corresponding to the fault signal.

2. The multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium as described in claim 1, characterized in that, The boundary threshold The method for determining it is as follows: Pre-defined comparison prototypes for each fault category are used; for each category, the feature vector of the sample farthest from the comparison prototype is used as the category boundary sample, and the distance between the category boundary sample and the comparison prototype is used as the boundary threshold. .

3. The multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium as described in claim 1, characterized in that, Using cosine similarity to calculate samples Compared to the prototype Similarity between .

4. The multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium as described in any one of claims 1-3, characterized in that, The neural network model uses the 1DResNet-18 network.

5. A multi-condition fault diagnosis system based on feature similarity measurement and cross-domain equilibrium, characterized in that, Includes a processor, the processor being configured to execute the multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium as described in any one of claims 1-4.

7. A computer program product, characterized in that, The method includes a computer program that, when executed by a processor, implements the multi-condition fault diagnosis method based on feature similarity measurement and cross-domain equilibrium as described in any one of claims 1-4.

Images