Mechanical fault diagnosis method based on spatio-temporal feature enhanced transfer learning
By constructing a spatiotemporal feature enhanced transfer learning network based on neural ordinary differential equations and path signatures, the problems of data imbalance and insufficient feature enhancement in rotating machinery fault diagnosis are solved, and high-precision and robust fault diagnosis is achieved, especially showing stronger adaptability in noisy environments.
Patent Information
- Application Number
- CN202510792780.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-09-16
AI Technical Summary
Existing technologies in rotating machinery fault diagnosis, especially rolling bearing fault diagnosis, have problems with data imbalance and unlabeled data, which limits the diagnostic accuracy and generalization of deep learning models, and existing feature enhancement methods are not suitable for vibration signal feature enhancement.
A spatiotemporal feature enhancement transfer learning method based on neural ordinary differential equations and path signatures is adopted to construct a spatiotemporal feature enhancement transfer learning network. Combined with the MMD domain adaptation loss function and the cross entropy classification loss function, adaptive feature extraction is performed through a residual-enhanced convolutional neural network. Feature enhancement is performed using the ODE solver and path signature module to overcome the shortcomings of the convolutional neural network and enhance the temporal and spatial characteristics of the vibration signal.
The accuracy and robustness of rotating machinery fault diagnosis under unbalanced data are improved, high-precision fault diagnosis across working conditions is achieved, and the diagnostic accuracy and generalization ability are improved, especially showing stronger robustness in noisy environments.
Smart Images
Figure CN120653964A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to a mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning. Background Art
[0002] With the rapid development of industry, fault diagnosis plays an increasingly important role in maintaining equipment health and ensuring safe operation. As a vital component of rotating machinery, rolling bearing failures can directly impact machine performance and can even lead to life-threatening situations. Therefore, accurately detecting and diagnosing bearing health is crucial for reducing operating and maintenance costs and ensuring safe operation of rotating machinery systems.
[0003] In recent years, deep learning has been widely used in rotating machinery fault diagnosis, achieving significant success. However, this success relies on sufficient labeled data to train diagnostic models and assumes that the source and target domain data must reside in the same feature space and have the same distribution, which is unrealistic in real-world engineering. First, rolling bearings remain healthy for long periods of time and rarely fail. Once a fault is detected, they require immediate replacement. Therefore, data collected from healthy states is much richer than data from faulty states, and data samples from different health states follow a long-tailed distribution. If unbalanced data is used to train deep learning-based diagnostic models, the model's decision boundary may be forced to shift toward states with fewer samples, thereby reducing diagnostic accuracy. Second, in engineering scenarios, most collected data is unlabeled, and obtaining labeled data for both healthy and faulty machine states is costly. Traditional fault diagnosis methods struggle to achieve satisfactory diagnostic results on unlabeled data under diverse operating conditions and health states. Fortunately, transfer learning can overcome these weaknesses by applying knowledge learned from one or more tasks to other related new tasks, enabling cross-operating and cross-equipment machinery fault diagnosis.
[0004] The goal of transfer learning is to leverage the knowledge contained in the source domain to improve the performance of predictive models in the target domain by reducing the distribution gap between the source and target domains. Transfer methods include instance transfer, feature transfer, parameter transfer, and relation transfer, with feature transfer being the most widely studied. However, feature transfer primarily focuses on extracting features from data or designing complex neural networks for raw vibration data, overlooking the potential of data augmentation and feature enhancement for domain adaptability. Data augmentation techniques can expand the diversity of training data, while feature enhancement aims to enrich features and improve the robustness and transferability of feature extraction. In recent years, several studies have applied data augmentation techniques to mechanical fault diagnosis and achieved fruitful results. However, data augmentation methods rely on large amounts of labeled data and typically require significant computational resources for model training and optimization. Lack of training data under diverse operating conditions can limit the generalization of the model. Existing results using feature enhancement strategies for mechanical fault diagnosis are very limited, and most of these methods are developed based on image data, making them not fully applicable to feature enhancement of vibration signals for mechanical fault diagnosis. Summary of the Invention
[0005] In response to the above-mentioned deficiencies in the prior art, the present invention provides a mechanical fault diagnosis method based on spatiotemporal feature enhancement transfer learning, which solves the problem that the existing methods are not fully applicable to vibration signal feature enhancement in mechanical fault diagnosis.
[0006] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: a mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning, comprising the following steps: S1: collecting the original vibration signal of the bearing on the mechanical equipment, and dividing the original vibration signal of the bearing into a source domain and a target domain; S2: Construct a spatiotemporal feature enhanced transfer learning network based on neural ordinary differential equations and path signatures; S3: Construct a joint loss function based on the MMD domain adaptation loss function and the cross entropy classification loss function; S4: Input the divided source domain and target domain into the spatiotemporal feature enhancement transfer learning network, and use the cross entropy classification loss function of the source domain labeled samples and the MMD domain adaptation loss function between the source domain and the target domain to iteratively train the spatiotemporal feature enhancement transfer learning network until the loss converges, thus obtaining the trained spatiotemporal feature enhancement transfer learning network; S5: Apply the trained spatiotemporal feature enhanced transfer learning network to the cross-operating-condition fault migration diagnosis of bearings under unbalanced data.
[0007] Furthermore, the spatiotemporal feature enhancement transfer learning network in S2 uses a residual-enhanced convolutional neural network as the backbone network for adaptive feature extraction, and combines a neural ordinary differential equation module and a path signature module for feature enhancement; The residual-enhanced convolutional neural network includes a CNN module, a ResCNN module 1 and a ResCNN module 2. The neural ordinary differential equation module includes two fully connected layers and an ODE solver, which is added after the residual-enhanced convolutional neural network. The path signature module performs feature enhancement by solving the iterative integral of the input path sequence and is connected in parallel with the neural ordinary differential equation module. Before adding three fully connected layers, the MMD domain adaptation loss function and the cross entropy classification loss function are added to the last fully connected layer.
[0008] Furthermore, the neural ordinary differential equation uses the local features extracted by the convolutional neural network as its initial value, and uses the ODE solver to solve the dynamic evolution process of the local features over time. The formula is:
[0009]
[0010]
[0011] in, is hidden state, For the moment, For neural networks, are the learnable parameters of the network, for The initial value of the hidden state at all times, for The final evolution value of the hidden state at all times, is the initial moment, For the final moment; The loss function of the hidden state is:
[0012] in, is the loss function, is the ODE solver; The accompanying state is:
[0013]
[0014]
[0015] in, is the accompanying state, the superscript represents the transpose symbol; By calling the ODE solver to solve , and .
[0016] Furthermore, the path signature for:
[0017] in, For path of Order truncated signature, For the start time, To end the moment, Represents an index, , is a positive integer;
[0018] in, are all real-valued paths.
[0019] Furthermore, the MMD domain adaptation loss function in S3 for:
[0020] in, is the maximum mean difference, and In order to retain only the source domain features that have similar samples in the target domain, and In order to retain only the features of samples of the same category in the target domain, is the sample size of the source domain, is the sample size of the target domain, is the mapping function in the regenerated Hilbert space, The source domain samples, The target domain samples;
[0021] in, and is a variable, is a hyperparameter with a value of 1;
[0022]
[0023] in, is the instruction function, is the number of fault categories, is the label of the source domain, is the label of the target domain, is the characteristic of the source domain, is the characteristic of the target domain, For the existence symbol.
[0024] Furthermore, the cross entropy classification loss function in S3 is for:
[0025] in, For the The samples belong to The label of the class, For the The samples belong to The probability of the class.
[0026] The beneficial effects of the present invention are: 1. This paper proposes two feature enhancement modules from the perspectives of temporal and spatial features: a dynamic evolution feature enhancement module based on the Neural Ordinary Differential Equation (NODE-DEFAM) and a spatial feature enhancement module based on path features (PS-SFAM), which enhance the features of vibration signals from both temporal and spatial perspectives.
[0027] 2. To address the problems of small samples and unbalanced fault data, the present invention proposes to embed two feature enhancement modules into the transfer learning framework, constructing an intelligent diagnostic migration framework suitable for variable operating condition fault diagnosis, which can realize bearing fault diagnosis under unbalanced data and improve the accuracy and robustness of existing intelligent migration diagnosis methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 This is a flow chart of a mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning proposed by the present invention.
[0029] Figure 2 This is a structural schematic diagram of a mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning proposed in the present invention.
[0030] Figure 3 This is a schematic diagram of the structure of the enhanced transfer learning network based on spatiotemporal features (ETL-STF) proposed in this invention.
[0031] Figure 4 Flowchart of the module for enhancing the dynamic evolution characteristics based on the Neural Ordinary Differential Equations.
[0032] Figure 5 This is the flow chart of the spatial feature enhancement module based on path features. DETAILED DESCRIPTION
[0033] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0034] like Figure 1 As shown, a mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning includes the following steps: S1: collecting the original vibration signal of the bearing on the mechanical equipment, and dividing the original vibration signal of the bearing into a source domain and a target domain; S2: Construct a spatiotemporal feature enhanced transfer learning network based on neural ordinary differential equations and path signatures; S3: Construct a joint loss function based on the MMD domain adaptation loss function and the cross entropy classification loss function; S4: Input the divided source domain and target domain into the spatiotemporal feature enhancement transfer learning network, and use the cross entropy classification loss function of the source domain labeled samples and the MMD domain adaptation loss function between the source domain and the target domain to iteratively train the spatiotemporal feature enhancement transfer learning network until the loss converges, thus obtaining the trained spatiotemporal feature enhancement transfer learning network; S5: Apply the trained spatiotemporal feature enhanced transfer learning network to the cross-operating-condition fault migration diagnosis of bearings under unbalanced data.
[0035] In S1, the original vibration signals of the bearings on the mechanical equipment are collected, and then the collected data signals are divided into the source domain and the target domain. To simulate the actual situation, the number of vibration signals of the bearings in a healthy state is far greater than the number of vibration signals in a faulty state. Therefore, the data ratio of the source domain is set to healthy state: outer ring fault: inner ring fault: roller fault = 20:9:1:2.
[0036] like Figure 2 、 Figure 3 、 Figure 4 and Figure 5 As shown, the spatiotemporal feature enhancement transfer learning network (ETL-STF) in S2 uses the residual enhanced convolutional neural network (ResCNN) as the backbone network for adaptive feature extraction, and combines the neural ordinary differential equation module (NODE-TFAM) and the path signature module (PS-SFAM) for feature enhancement; The residual enhanced convolutional neural network (ResCNN) consists of three submodules: the CNN module, the ResCNN module 1, and the ResCNN module 2. The CNN module consists of a Conv1D convolution block, a batch normalization layer, a LeakyReLU activation function, and a maximum pooling layer; the ResCNN module 1 consists of two series-connected Conv1D convolution blocks, a batch normalization layer, and a LeakyReLU activation function, and the module uses residual connections for enhancement; the ResCNN module 2 consists of two parallel-connected Conv1D convolution blocks, a batch normalization layer, a LeakyReLU activation function, and a global average pooling layer, and the module uses residual connections for enhancement.
[0037] The Neural Ordinary Differential Equation Module (NODE-TFAM) consists of two fully connected layers and an ODE solver, added after the Residual Enhanced Convolutional Neural Network (ResCNN). The Path Signature Module (PS-SFAM) performs feature enhancement by solving the iterative integral of the input path sequence. It is connected in parallel with the Neural Ordinary Differential Equation Module (NODE-TFAM) and before the three fully connected layers. The MMD domain adaptation loss is added to the last fully connected layer; similarly, the cross-entropy classification loss is also applied to the last fully connected layer.
[0038] The neural ordinary differential equation uses the local features extracted by the convolutional neural network as its initial value, and uses the ODE solver to solve the dynamic evolution of the local features over time. The formula is:
[0039]
[0040]
[0041] in, is hidden state, For the moment, For neural networks, are the learnable parameters of the network, for The initial value of the hidden state at all times, for The final evolution value of the hidden state at all times, is the initial moment, For the final moment; The loss function of the hidden state is:
[0042] in, is the loss function, is the ODE solver; The accompanying state is:
[0043]
[0044]
[0045] in, is the accompanying state, the superscript represents the transpose symbol; By calling the ODE solver to solve , and .
[0046] Neural ODEs integrate neural networks with ODEs and perform backpropagation through the ODE solver to solve the gradients via the adjoint state method.
[0047] Convolutional neural networks, due to continuous downsampling operations, can lose key information and ignore the temporal correlation in the signal. Therefore, the present invention proposes to use a neural ordinary differential equation model, using the local features extracted by the convolutional neural network as its initial value, and using an ODE solver to solve the dynamic evolution of the local features over time, revealing the evolution law of the signal at different time scales, thus compensating for the shortcomings of convolutional neural networks. Among them, the expression of the temporal dynamic evolution feature is:
[0048]
[0049] in, is the feature map, To define a neural network with continuous feature transformation, is an N-dimensional real number space, is the number of channels, that is, the number of data in the channel dimension, is the height, that is, the size of the data in the vertical dimension, is the width, that is, the size of the data in the horizontal dimension; Obtaining the final moment through the ODE solver Feature map , whose expression is:
[0050] The path signature for:
[0051] in, represents a real-valued path, For path of Order truncated signature, For the start time, To end the moment, Represents an index, , is a positive integer;
[0052] in, are all real-valued paths.
[0053] From a spatial perspective, a path signature treats the vibration signal as a mathematically rough path. This embodiment calculates the third-order truncated signature of the vibration signal to obtain interpretable, robust, and inherent characteristics of various fault samples. Therefore, the present invention proposes to enhance the characteristics of the vibration signal from both temporal and spatial perspectives based on the Neural Ordinary Differential Equation and the path signature method. The expression for the third-order truncated signature is:
[0054]
[0055] The MMD domain adaptation loss function in S3 for:
[0056] in, is the maximum mean difference, and In order to retain only the source domain features that have similar samples in the target domain, and In order to retain only the features of samples of the same category in the target domain, is the sample size of the source domain, is the sample size of the target domain, To regenerate the mapping function in the Hilbert space, the Gaussian kernel function is used here. The source domain samples, The target domain samples;
[0057] in, and is a variable, is a hyperparameter with a value of 1;
[0058]
[0059] in, It is a command function that returns 1 when the condition is met, otherwise it returns 0. is the number of fault categories, is the label of the source domain, is the label of the target domain, is the characteristic of the source domain, is the characteristic of the target domain, For the existence symbol.
[0060] The cross entropy classification loss function in S3 for:
[0061] in, For the The samples belong to The label of the class, For the The samples belong to The probability of the class.
[0062] In one embodiment of the present invention, the effectiveness of this method is verified through the following experimental results.
[0063] The detailed parameters of the spatiotemporal feature enhanced transfer learning network (ETL-STF) in this experiment are shown in Table 1.
[0064] Table 1 Detailed parameters of the ETL-STF network
[0065] In Table 1, Conv1, Conv2, Conv22, Conv3, and Conv32 are all convolutional layers, BN1, BN2, and BN3 are all batch normalization layers, LeakyReLU1, LeakyReLU2, and LeakyReLU3 are LeakyReLU activation functions, MaxPool1 is the maximum pooling layer, GlobalAvgPool1D is the global pooling layer, ODEFunc is the proposed NODE-TFAM module, Sig_features is the path signature feature proposed by the proposed PS-SFAM module, and FC1, FC2, and FC3 are fully connected layers.
[0066] Verification Experiment: This experiment collected raw bearing vibration signals from two test benches: the publicly available PU bearing dataset from the University of Paderborn in Germany and the laboratory's in-house STU railway bearing dataset. The PU bearing dataset test bench has four health states: healthy bearing (N), primary outer race fault (OF_1), secondary outer race fault (OF_2), and inner race fault (IF). The radial force used in the experiment was 1000 N, the rotational speed was 1500 rpm, and the acquisition frequency was 64 kHz. Damage severity was categorized into two levels. For primary damage, the assigned percentage ranged from 0-2%, with a bearing limit less than or equal to 2 mm. For secondary damage, the assigned percentage ranged from 2-5%, with a bearing limit greater than 2 mm. As shown in Table 2 for the PU dataset, the load torques for the source and target domain data were 0.7 Nm and 0.1 Nm, respectively.
[0067] The STU bearing dataset testbed conducted vibration signal experiments using a rolling bearing model 197726. The experiment controlled two rotational speeds, 380 rpm and 505 rpm, and collected vibration signals under different health states at a uniform sampling frequency of 25.6 kHz under zero load conditions. As shown in Table 2 for the STU dataset, there are four health states: healthy bearing (N), outer race fault (OF), inner race fault (IF), and rolling element fault (RF).
[0068] Table 2 Detailed parameters of the dataset
[0069] Ablation experiment: In order to prove the effectiveness of the time feature enhancement module based on the neural ordinary differential equation and the space feature enhancement module based on the path signature proposed in this invention, experiments were conducted by removing these two feature enhancement modules respectively. The experimental results are shown in Table 3.
[0070] Table 3 Experimental results of five repeated experiments (%)
[0071] From Table 3, it can be concluded that the time feature enhancement module based on the neural ordinary differential equation and the space feature enhancement module based on the path signature proposed in the present invention can effectively improve the accuracy of the basic model, and the average diagnostic accuracy is increased by 3% to 6%.
[0072] Comparative test: In order to demonstrate the superiority of the cross-working condition bearing migration diagnosis method proposed in this paper based on the spatiotemporal feature enhanced transfer learning network (ETL-STF) based on the neural ordinary differential equation and path signature, the current typical network models (WDCNN, Resdan, MK_RESCNN, MA1DCNN) and the classical distribution alignment (CORAL) method are compared. The experimental results of the PU dataset are shown in Table 4, and the experimental results of the STU dataset are shown in Table 5.
[0073] Table 4 Experimental results of five repeated experiments (%)
[0074] Table 5 Experimental results of five repeated experiments (%)
[0075] Tables 4 and 5 show that the proposed transfer learning network based on neural ordinary differential equations and path signatures for enhanced spatiotemporal features has higher transfer diagnosis accuracy and stronger generalization capabilities. In particular, the average diagnostic accuracy of ETL-STF reached over 93%, an improvement of approximately 6% to 10% compared to other transfer diagnosis methods.
[0076] Robustness experiment: To demonstrate the robustness of the proposed cross-operation bearing transfer diagnosis method based on the neural ordinary differential equation and path signature enhanced spatiotemporal feature transfer learning network (ETL-STF), Gaussian white noise with a mean of 0 and standard deviations of 0.01, 0.05, and 0.1 was added to the PU and STU bearing datasets, respectively. The experimental results for the PU dataset are shown in Table 6, and the experimental results for the STU dataset are shown in Table 7.
[0077] Table 6 Experimental results of five repeated experiments under three noise intensities (%)
[0078] Table 7 Experimental results of five repeated experiments under three noise intensities (%)
[0079] Tables 6 and 7 show that the proposed transfer learning network for spatiotemporal feature enhancement based on neural ordinary differential equations and path signatures exhibits greater robustness. ETL-STF achieves an average diagnostic accuracy of over 88.5% in the presence of white noise with a mean of 0 and a standard deviation of 0.1. In particular, the average accuracy of ETL-STF on the STU dataset exceeds 95%, demonstrating significant robustness compared to other transfer diagnosis methods.
[0080] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the invention.
Claims
1. A mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning, characterized in that: The following steps are involved: S1: collecting the original vibration signal of the bearing on the mechanical equipment, and dividing the original vibration signal of the bearing into a source domain and a target domain; S2: Construct a spatiotemporal feature enhanced transfer learning network based on neural ordinary differential equations and path signatures; S3: Construct a joint loss function based on the MMD domain adaptation loss function and the cross entropy classification loss function; S4: Input the divided source domain and target domain into the spatiotemporal feature enhancement transfer learning network, and use the cross entropy classification loss function of the source domain labeled samples and the MMD domain adaptation loss function between the source domain and the target domain to iteratively train the spatiotemporal feature enhancement transfer learning network until the loss converges, thus obtaining the trained spatiotemporal feature enhancement transfer learning network; S5: Apply the trained spatiotemporal feature enhanced transfer learning network to the cross-operating-condition fault migration diagnosis of bearings under unbalanced data.
2. The mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning according to claim 1 is characterized in that: The spatiotemporal feature enhancement transfer learning network in S2 uses a residual-enhanced convolutional neural network as the backbone network for adaptive feature extraction, and combines the neural ordinary differential equation module and the path signature module for feature enhancement; The residual-enhanced convolutional neural network includes a CNN module, a ResCNN module 1 and a ResCNN module 2. The neural ordinary differential equation module includes two fully connected layers and an ODE solver, which is added after the residual-enhanced convolutional neural network. The path signature module performs feature enhancement by solving the iterative integral of the input path sequence and is connected in parallel with the neural ordinary differential equation module. Before adding three fully connected layers, the MMD domain adaptation loss function and the cross entropy classification loss function are added to the last fully connected layer.
3. The mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning according to claim 2 is characterized in that: The neural ordinary differential equation uses the local features extracted by the convolutional neural network as its initial value, and uses the ODE solver to solve the dynamic evolution of the local features over time. The formula is: in, is hidden state, For the moment, For neural networks, are the learnable parameters of the network, for The initial value of the hidden state at all times, for The final evolution value of the hidden state at all times, is the initial moment, For the final moment; The loss function of the hidden state is: in, is the loss function, is the ODE solver; The accompanying state is: in, is the accompanying state, the superscript represents the transpose symbol; By calling the ODE solver to solve , and .
4. The mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning according to claim 3 is characterized in that: The path signature for: in, For path of Order truncated signature, For the start time, To end the moment, Represents an index, , is a positive integer; in, are all real-valued paths.
5. The mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning according to claim 1, characterized in that: The MMD domain adaptation loss function in S3 for: in, is the maximum mean difference, and In order to retain only the source domain features that have similar samples in the target domain, and In order to retain only the features of samples of the same category in the target domain, is the sample size of the source domain, is the sample size of the target domain, is the mapping function in the regenerated Hilbert space, The source domain samples, The target domain samples; in, and is a variable, is a hyperparameter with a value of 1; in, is the instruction function, is the number of fault categories, is the label of the source domain, is the label of the target domain, is the characteristic of the source domain, is the characteristic of the target domain, For the existence symbol.
6. The mechanical fault diagnosis method based on spatiotemporal feature enhanced transfer learning according to claim 5 is characterized in that: The cross entropy classification loss function in S3 for: in, For the The samples belong to The label of the class, For the The samples belong to The probability of the class.