Machine tool bearing diagnosis method and system based on improved SMOTE

Abstract

The invention relates to the technical field of intelligent diagnosis, and provides a machine tool bearing diagnosis method and system based on improved SMOTE. According to the method, for the problem of data imbalance caused by scarcity of fault samples in bearing vibration signals, the vibration signals are obtained and processed into a feature sample set; identifying minority class fault samples, calculating weight values according to sample distribution of the minority class fault samples in the feature space, adaptively determining a generation area for each sample based on the weight values, and synthesizing new samples to balance data; and finally using the balanced sample set to train a fault diagnosis model for classification. According to the bearing fault diagnosis method, through weight-driven adaptive region adjustment, the classification boundary is effectively protected while the sample diversity is enhanced, the accuracy of bearing fault diagnosis is improved, and the missing report rate of minority class faults is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent diagnosis, more particularly, to a bearing diagnosis method and system for machine tools based on improved SMOTE. BACKGROUND

[0002] With the rapid development of industrial automation technology, modern production processes are increasingly complex, and higher requirements are placed on the accuracy and efficiency of equipment fault diagnosis. However, in actual industrial systems, due to the low frequency of abnormal events of equipment and the need to shut down in time to ensure safety and reduce economic losses when faults occur, most of the data that can be collected are normal operation state samples, and fault samples are extremely scarce. This inherent class imbalance problem (normal data is much more than fault data) seriously restricts the performance of the classification model, causing the model to excessively bias towards the majority class (normal state) during training, resulting in a significantly increased risk of missed detection of the minority class (fault events). In addition, industrial scenarios often involve multiple fault types, and the characteristics of each fault type are different, further exacerbating the difficulty of multi-fault classification under imbalanced data distribution.

[0003] Current methods for addressing data imbalance problems are mainly divided into algorithm level and data level. Algorithm level methods (such as cost-sensitive learning, ensemble learning, etc.) adjust the classification algorithm to focus more on minority class samples, but this type of method highly depends on parameter tuning and has limited generalization ability in different industrial scenarios. Data level methods balance the dataset by adjusting the number of samples, where random undersampling may lose key distribution information, and oversampling methods such as Figure 1 The synthetic minority over-sampling technique (SMOTE) shown in the synthetic minority over-sampling technique (SMOTE) generates new samples by interpolation, which can alleviate overfitting, but its traditional version uniformly synthesizes new samples for all minority class samples, ignoring local data density differences, which can easily introduce noise or generate samples limited to the local neighborhood. Although subsequent improvements such as SMOTE-ENN, ESLEBS, etc. have appeared, it is still difficult to effectively solve the problems of sample overlap and lack of diversity, especially in the generation of boundary samples and cross-regional interpolation.

[0004] As the core component of rotating machinery, the fault diagnosis of bearings is a key link of industrial health management. However, the fault features in the vibration signal often appear in a specific period in the form of pulse sequence or amplitude mutation, and if the original signal is directly analyzed as a whole, the fault features are easy to be covered. In addition, the high-dimensional vibration signal needs to be preprocessed by segmentation, feature extraction (such as 20 kinds of time and frequency domain features) and dimension reduction, which further increases the classification complexity under unbalanced data. Although the existing methods based on SMOTE or generative adversarial network (GAN) have certain effect, the training process of GAN is complex and prone to mode collapse, and the current improved version of SMOTE is still insufficient in adaptive adjustment of sample generation area and safety constraint of boundary samples, so a new scheme that can balance sample diversity and classification boundary optimization is needed. SUMMARY

[0005] The present application aims at the technical problems existing in the prior art, and provides a bearing diagnosis method and system for machine tools based on improved SMOTE, which solves the technical problems of low recognition rate of minority class fault samples and high risk of missed detection caused by unbalanced vibration signal data in machine tool bearing fault diagnosis, and poor quality of samples generated by traditional SMOTE method and easy introduction of boundary overlapping samples.

[0006] According to a first aspect of the present application, a bearing diagnosis method for machine tools based on improved SMOTE is provided, comprising: S1, obtaining the running data of the machine tool bearing and processing it into a feature data original sample set containing a plurality of samples, wherein each sample is labeled with a corresponding state category; S2, identifying the minority class fault samples in the original sample set, and for each minority class fault sample, calculating a weight value according to its sample distribution in the feature space; based on the weight value, adaptively determining a corresponding sample generation area for each minority class fault sample, synthesizing new minority class fault samples in the determined sample generation area, and obtaining a balanced sample set; S3, using the balanced sample set to train a deep learning-based fault diagnosis model to classify the running data of the bearing to be diagnosed.

[0007] According to a second aspect of the present application, a bearing diagnosis system for machine tools based on improved SMOTE is provided, comprising: A data preparation module is configured to obtain the running data of the machine tool bearing and process the running data into a feature data original sample set containing a plurality of samples, wherein each sample is labeled with a corresponding state category; The sample balancing processing module, connected to the data preparation module, is used to identify minority class fault samples in the original sample set. For each minority class fault sample, a weight value is calculated based on its sample distribution in the feature space. Based on the weight value, the corresponding sample generation region is adaptively determined for each minority class fault sample. New minority class fault samples are synthesized in the determined sample generation region to obtain a balanced sample set. The fault diagnosis application module, connected to the sample balancing processing module, includes a deep learning-based fault diagnosis model, which is used to train the fault diagnosis model using the balanced sample set and to classify the state of the input bearing operating data to be diagnosed.

[0008] According to a third aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the processor is configured to implement the steps of the above-described machine tool bearing diagnosis method based on the improved SMOTE when executing a computer management program stored in the memory.

[0009] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer management class program stored thereon, which, when executed by a processor, implements the steps of the above-described machine tool bearing diagnosis method based on the improved SMOTE.

[0010] This invention provides a machine tool bearing diagnosis method, system, electronic device, and storage medium based on an improved SMOTE (Self-Modified Model for Machine Tool Bearings). Based on differentiated processing of the distribution characteristics of minority class samples in the feature space, it calculates the sample distribution weight value for each minority class fault sample and adaptively determines the sample generation region accordingly. Finally, it uses the balanced sample set to train a fault diagnosis model, achieving accurate diagnosis of machine tool bearing faults. Specifically, samples near the classification boundary are assigned higher weights and a relatively conservative sample generation region is used to avoid sample overlap, while internal samples are assigned lower weights and a more relaxed sample generation region is used to enhance diversity. This invention, through weight-driven adaptive region adjustment, improves the generation quality of boundary samples to enhance the clarity of classification boundaries, while ensuring the generation quantity of internal samples to improve data distribution balance and reduce the risk of missed detections. This effectively improves the recognition accuracy and robustness of the bearing fault diagnosis model for samples of each category. Attached Figure Description

[0011] Figure 1 This is a schematic diagram illustrating the principle of the traditional synthetic minority class oversampling technique (SMOTE). Figure 2 A flowchart of a machine tool bearing diagnostic method based on an improved SMOTE is provided for this invention. Figure 3 A schematic diagram illustrating the principle of a machine tool bearing diagnostic method based on an improved SMOTE, provided for one embodiment; Figure 4 This is a schematic diagram illustrating the processing of continuous vibration signals via a sliding window in one embodiment. Figure 5 In the diagram, (a) shows the distance of a sample relative to the centroids of the minority and majority clusters; (b) shows the density of a sample relative to the centroids of the minority and majority clusters. Figure 6 This is a schematic diagram illustrating the degree of geometric region variation under different weight values ​​in a certain embodiment. Figure 7 A schematic diagram illustrating the generation of new samples using the adaptive geometric region adjustment method in one embodiment; Figure 8 The diagram illustrates the average performance of the oversampling method for datasets A, B, and C, which correspond to the results in Tables 3-5. Figure 9 A schematic diagram of the average performance of the oversampling method for datasets D, E, and F, corresponding to the results in Tables 6-8; Figure 10 A block diagram of a machine tool bearing diagnostic system based on an improved SMOTE is provided for this invention. Figure 11 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention; Figure 12 This is a schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. Detailed Implementation

[0012] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] Figure 2 This invention provides a flowchart of a machine tool bearing diagnostic method based on an improved SMOTE. Figure 3 This is a schematic diagram illustrating the principle of a machine tool bearing diagnostic method based on an improved SMOTE, provided for one embodiment. (Combined with...) Figure 2 and Figure 3 As shown, this embodiment provides a machine tool bearing diagnostic method based on an improved SMOTE, including steps S1 to S3: S1, acquire the operating data of the machine tool bearing and process it into a raw sample set of feature data containing multiple samples, wherein the samples are labeled with corresponding state categories.

[0014] This step transforms the original vibration signal into a feature sample set. First, the continuous vibration signal is segmented using a sliding window. For each signal segment, time-domain features (e.g., statistical features such as mean and root mean square) and frequency-domain features (e.g., spectral centroid and frequency variance) are extracted in parallel. These extracted multidimensional features are then combined into a feature vector. Subsequently, techniques such as principal component analysis are used for dimensionality reduction. The cumulative contribution rate of the eigenvalues ​​is calculated to filter out the main feature components, removing redundant information while retaining key features that distinguish different state categories, thus forming a sample set suitable for classification.

[0015] S2, identify minority class fault samples in the original sample set, calculate a weight value for each minority class fault sample based on its sample distribution in the feature space; based on the weight value, adaptively determine the corresponding sample generation region for each minority class fault sample, and synthesize new minority class fault samples in the determined sample generation region to obtain a balanced sample set.

[0016] This step achieves adaptive sample balancing. For example, a clustering algorithm is first used to divide minority fault samples into different sub-clusters. For each sample, four parameters are calculated: distance to the center of its own cluster, distance to the center of samples from other clusters, local intra-cluster density, and inter-cluster density. These parameters are then combined to calculate a weight value reflecting the importance of the sample boundary. Based on the weight value, the nearest neighbors that meet the spatial visibility condition are selected for each sample to avoid generating invalid samples across clusters. Finally, the range of the generated region is dynamically adjusted according to the weight value. A conservative, near-linear generated region is used for high-weight samples, while a hyperspherical expanded region is used for low-weight samples. New samples are then synthesized through linear interpolation.

[0017] S3 uses a balanced sample set to train a deep learning-based fault diagnosis model to classify the operating data of the bearing to be diagnosed.

[0018] This step completes the construction and application of the fault diagnosis model. For example, a deep learning model is trained using a balanced sample set. The data is divided into a training set and a validation set. A deep learning network structure that includes feature extraction and classification decisions is adopted. The model parameters are optimized through iterative training, and the hyperparameters are adjusted based on the performance on the validation set. The trained model can perform end-to-end feature learning and state classification on the input vibration signal, output the health status diagnosis result of the bearing, and achieve accurate fault identification under unbalanced data conditions.

[0019] Understandably, given the deficiencies in the background technology, this invention proposes a machine tool bearing diagnosis method based on an improved SMOTE. This method, through weight-driven adaptive region adjustment, improves the generation quality of boundary samples to enhance the clarity of classification boundaries, while ensuring the generation quantity of internal samples to improve data distribution balance, effectively reducing the risk of missed detections. This significantly improves the bearing fault diagnosis model's accuracy and robustness in identifying samples of each category.

[0020] Based on the above technical solutions, the embodiments of the present invention can be further improved as follows.

[0021] In bearing fault diagnosis, relevant fault characteristics typically appear only briefly within a specific time period in the form of pulse sequences or amplitude abrupt changes. Directly analyzing the original vibration signal as a whole can easily lead to fault characteristics being masked by other information. Furthermore, the high-dimensional nature of vibration signals not only increases computational complexity but also makes establishing fault mapping relationships difficult. Therefore, preprocessing the original vibration signal is necessary in step S1.

[0022] In one possible embodiment, step S1 includes sub-steps S101 to S103.

[0023] S101, acquire the operating data of the machine tool bearing, the operating data including vibration signals; use a sliding window to divide the continuous vibration signals into multiple signal segments, wherein the window length and sliding step size are preset according to the bearing fault characteristic frequency; by Figure 4 For example, Figure 4 The horizontally extending blue waveform represents the acquired raw continuous vibration signal. The "L" above the waveform indicates the window length of the sliding window, and the "S" below the waveform indicates the sliding step size of the sliding window. For example... Figure 4 As shown, a red dashed window of length L captures the first segment of the signal starting from the signal's initial position, and then the window slides forward a step size S (e.g., ...). Figure 4 As shown in the green dashed window, a segment of the signal is captured, and this process is repeated (e.g., ...). Figure 4 (As shown in the orange dashed window in the middle), thus dividing the long sequence of vibration signals into a series of continuous and potentially overlapping short-time signal segments.

[0024] Figure 4The vertically arranged dashed boxes of the same color on the right show the signal segments extracted from each window. This step converts non-stationary vibration signals into multiple analyzable local signal units, which can be used to capture transient impact characteristics caused by local bearing damage (such as pitting and cracks). These characteristics may be masked in the complete signal but become prominent in the segmented signal. The signal segments obtained in step S101 are the direct objects for time-domain and frequency-domain feature extraction in the subsequent step S102, and can be used to construct feature vectors characterizing the bearing state.

[0025] S102, based on the segmented signal segments, extract time-domain features and frequency-domain features respectively. The time-domain features include at least three of the following: mean, root mean square, peak value, kurtosis, and impulse index. The frequency-domain features include at least two of the following: spectral centroid, root mean square frequency, and frequency variance. These features are combined to form a feature vector characterizing the bearing's operating state.

[0026] Table 1 illustrates 20 possible feature extraction formulas for feature extraction of segmented vibration signal segments in this embodiment to assist in fault identification.

[0027] Table 1. Vibration signal segment feature extraction formula

[0028] Table 1 shows the time-domain characteristic formula ( ~ Directly analyze the amplitude sequence of the original vibration signal. Statistical calculations are performed to characterize the amplitude distribution, volatility, and waveform morphology of the signal. Specifically, Mean (μ): Calculates the average amplitude of the signal, reflecting the average level of vibration or DC component. Standard deviation (σ): measures the degree of dispersion of the signal amplitude relative to the mean, characterizing the intensity of vibration fluctuations. Root mean square (rms): Calculates the effective amplitude of the signal and is a key indicator for measuring the magnitude of vibration energy. Peak value: Extracts the maximum absolute value of the signal, and is very sensitive to instantaneous shocks. Peak-to-peak value (pp): Calculates the difference between the maximum and minimum values ​​of the signal, reflecting the overall amplitude range of the vibration. Skewness (ske): Describes the asymmetry of the signal amplitude distribution. A positive value indicates a right-skewed distribution, which may indicate a one-sided impulse. Kurtosis: Describes the sharpness of the signal amplitude distribution. It is extremely sensitive to impact pulses caused by faults such as bearing pitting and cracks, and is a classic indicator for fault diagnosis. Crest factor: The ratio of peak value to root mean square value, used to indicate whether there are prominent impact components in the signal. Impulse factor (imp): The ratio of peak value to absolute average value, also used to highlight the pulse characteristics of a signal. Waveform factor (shape): The ratio of the root mean square value to the absolute mean value, describing the waveform characteristics of the signal. Clearance factor: The ratio of peak value to root square magnitude (the square of the square root of the absolute value), which has a certain ability to detect minor impacts caused by early faults.

[0029] Frequency domain characteristic formula ( ~ First, a spectral analysis of the signal segment is required, followed by statistical analysis in the frequency domain to characterize the frequency distribution characteristics of the vibration energy. Specifically, Frequency mean ( ): Calculates the average value of the frequency components in the spectrum. Frequency centroid ( ): The weighted average frequency of the spectrum is calculated, which reflects the position of the main frequency of the vibration energy. Its deviation is often related to the fault. Root mean square frequency (RMS) ): Calculates the weighted root mean square frequency of the spectrum, which characterizes the degree of concentration of the frequency distribution. Frequency variance ): Measures the degree of dispersion of frequency components relative to the frequency centroid, describing the range of the spectrum. Frequency energy ( ): The summation of the spectral amplitudes represents the total energy of the signal segment. Peak frequency ( Find the frequency point with the largest amplitude in the spectrum, i.e., the dominant frequency. Frequency standard deviation ( ): Describes the skewness characteristics of the spectrum shape from the perspective of cubic moments. Frequency entropy ( ): The complexity or uncertainty of calculating the spectrum based on the information entropy theory; changes in entropy can reflect state anomalies. Frequency skewness ( ): Describes the skew direction of the shape of the spectrum distribution.

[0030] The 20 feature extraction formulas in Table 1 systematically quantify the bearing vibration signal from two dimensions: amplitude statistics (time domain) and energy distribution (frequency domain). Together, they constitute a high-dimensional feature vector that comprehensively characterizes the bearing's operating state (normal, inner race fault, outer race fault, rolling element fault), providing accurate input data for subsequent sample balancing and deep learning model training. Since these 20 feature extraction formulas are existing technologies, their specific parameter definitions are not explained. In practical applications, these 20 feature extraction formulas can be freely combined to obtain various feature extraction schemes.

[0031] For example, in a specific application scenario, for a 1024-point signal segment extracted from a bearing vibration signal, time-domain feature extraction calculates the signal segment's mean (reflecting the average vibration level), root mean square (characterizing vibration energy), peak value (capturing instantaneous impact), kurtosis (sensitively detecting impact faults), and impulse index (highlighting impulse components), thus obtaining a set of time-domain descriptors. Simultaneously, after performing a Fast Fourier Transform on the signal segment to obtain its spectrum, frequency-domain feature extraction calculates its spectral centroid (indicating the dominant frequency position), root mean square frequency (reflecting the concentration of frequency distribution), and frequency variance (describing the degree of spectral dispersion). Finally, these eight extracted values ​​are combined in a predetermined order into an 8-dimensional feature vector. This feature vector comprehensively characterizes the key features of this signal segment in terms of amplitude, waveform, and spectrum, providing a quantitative basis for accurately distinguishing different bearing states such as normal bearing, inner race fault, and outer race fault.

[0032] S103, Principal Component Analysis (PCA) is used to reduce the dimensionality of the feature vectors, removing redundant information and retaining key features that distinguish different state categories, forming an original sample set for bearing fault classification. Specifically, this includes: Calculate the eigenvalues ​​of the covariance matrix of the eigenvectors, sort the corresponding principal components by eigenvalue from largest to smallest, and retain the principal components whose cumulative variance contribution rate exceeds a preset threshold.

[0033] For example, suppose 20 time-domain and frequency-domain features have been extracted from the original vibration signal to form a high-dimensional feature vector. First, the covariance matrix among these 20 feature dimensions is calculated, and eigenvalue decomposition yields 20 eigenvalues ​​and their corresponding principal component directions. These 20 eigenvalues ​​are then sorted from largest to smallest; for example, the cumulative variance contribution rate of the first 5 principal components is relatively large (e.g., possibly reaching 96%), while the contribution rate of the last 15 principal components is smaller (e.g., all less than 1%). If a preset threshold is set to 95%, only the first 5 principal components are retained. The 5 retained new feature dimensions are linear combinations of the original 20 features. For example, the first principal component may mainly integrate vibration energy-related features (such as root mean square and root mean square frequency), while the second principal component may focus on impact characteristic-related features (such as kurtosis and impulse index). This transformation reduces the original 20-dimensional features to 5-dimensional new features. While preserving the distinguishability between different categories such as normal bearing condition, inner ring fault, and outer ring fault, it can effectively eliminate feature redundancy caused by irrelevant factors such as temperature fluctuations, and finally form a low-dimensional original sample set suitable for fault diagnosis / classification models.

[0034] Understandably, this embodiment captures the local transient fault features of bearing vibration signals in segments through a sliding window, constructs a feature vector that comprehensively describes the fault state by combining multi-dimensional feature extraction in the time and frequency domains, and uses principal component analysis to reduce the dimensionality of high-dimensional features. While retaining principal components whose cumulative variance contribution rate exceeds the threshold, redundant information is effectively removed and the distinguishing features of different fault modes are highlighted. This significantly reduces the input dimension and computational complexity of subsequent bearing fault diagnosis / classification models, while enhancing the feature set's ability to represent different fault types such as the bearing inner ring, outer ring, and rolling elements, providing a data foundation for subsequent unbalanced data processing and high-precision fault classification.

[0035] Traditional SMOTE oversampling algorithms randomly generate new samples on the line segment between minority class samples and their nearest neighbors. However, the oversampling effect significantly decreases when majority class samples exist between minority class samples and their nearest neighbors. If new samples are only generated near the central region of minority class samples, their role in determining the classification boundary is greatly diminished. The adaptive geometric SMOTE algorithm proposed in step S2 effectively avoids new samples falling into the majority class region by dynamically adjusting the geometric range of the generated samples, thereby achieving accurate expansion of the sub-cluster boundaries.

[0036] In one possible embodiment, step S2 includes sub-steps S201 to S204.

[0037] S201, Identify the minority fault samples from the original sample set and perform clustering to form multiple sub-clusters corresponding to different fault characteristics.

[0038] Specifically, this includes: identifying the minority class fault samples in the original sample set and performing denoising processing; dividing the denoised minority class fault samples into multiple sub-clusters based on the K-means clustering algorithm, with each sub-cluster corresponding to the local feature distribution of a bearing fault mode.

[0039] Suppose that in a certain scenario, the original sample set contains four types of samples: normal state, inner race fault, outer race fault, and rolling element fault, with the latter three types of fault samples being significantly fewer than normal samples. First, noise reduction is performed on the minority fault samples by editing the nearest neighbor rule: for example, a sample point marked as having an inner race fault is considered noise and removed if most of its K nearest neighbors are normal samples. Then, for the remaining minority fault samples after noise reduction, K-means clustering is used to cluster them into n sub-clusters based on their Euclidean distance in the feature space. For example, after clustering, three sub-clusters are obtained. The samples in the first sub-cluster show prominent characteristics in high-frequency resonance band energy and impact index features, corresponding to the local feature distribution of inner ring failure. The samples in the second sub-cluster cluster are clustered in terms of frequency amplitude features, corresponding to the outer ring failure mode. The samples in the third sub-cluster show a distribution across multiple frequency components, corresponding to the typical features of rolling element failure. The samples within each sub-cluster have highly similar feature patterns, each representing a specific bearing failure mechanism.

[0040] S202, For each faulty sample, calculate a weight value based on its distribution density within its class and its separation from samples of different classes.

[0041] Specifically, this includes: for each faulty sample in a subcluster, calculating a weight value based on its intra-class distance to the subcluster center, inter-class distance to the center of the majority class sample, local intra-class density, and local inter-class density. Among these, boundary neighbor samples with smaller inter-class distances and lower intra-class densities receive higher weights.

[0042] Combination Figure 5 An example is provided to illustrate the calculation of weight values. Figure 5 (a) is a schematic diagram showing the distance of a sample relative to the centroids of the minority and majority clusters. Figure 5 (b) shows the density of the sample relative to the centroids of the minority and majority clusters.

[0043] Figure 5 (a) illustrates the calculation of the distance parameter: taking a minority class fault sample (blue triangle) as an example, its distance from the center of its own sub-cluster ( The Euclidean distance of the red triangle is (Intra-class distance), to the majority class sample center ( The Euclidean distance between the red dots is (Inter-class distance).

[0044] During the calculation, it is assumed that the majority class sample set is ,in, , Represents the majority class sample set The i-th majority class sample in the cluster has the following set of minority class samples: , , , Represents the minority class sample set The i-th minority class sample in the cluster. The centroids of the majority and minority class samples within the cluster are calculated using the following formula:

[0045]

[0046] For each minority class sample ,like Figure 5 As shown in (a), the Euclidean distance from a given element to the center of its own sub-cluster is calculated, which is the intra-cluster distance. ; and, calculate its distance to the majority class sample center. The Euclidean distance, i.e., the inter-class distance. The calculation formula is as follows:

[0047]

[0048] Figure 5 Image (b) illustrates the calculation of the density parameter: the local intra-class density of the same sample. Measured by the average distance to its k nearest samples of the same class (the range of the small white clusters in the figure); local inter-class density It is then measured by the average distance to the k nearest outlier (majority class) samples (the dashed arrows point to the yellow circular cluster).

[0049] During the calculation, for each minority class sample Determine k similar nearest neighbors and k dissimilar nearest neighbors, and calculate the distance between the sample and its similar nearest neighbors. and the distance to the least dissimilar nearest neighbor. .like Figure 5 Figure (b) illustrates the principle of local density, where the local density index characterizes the distribution characteristics of minority samples within a subcluster. Specifically, it refers to local intra-class density (within minority sub-clusters). And local inter-class density (between minority and majority classes) The calculation formula is as follows:

[0050]

[0051] In bearing fault diagnosis scenarios, assuming a sample point representing an "inner race fault," its intra-class distance... The value is smaller (closer to the center of the inner circle faulty sub-cluster), while the inter-class distance is smaller. The value is also very small (very close to the distribution area of ​​normal samples), and its local intra-class density is also low. Larger values ​​(sparse surrounding samples of similar faults), local inter-class density The value is relatively small (surrounded by a large number of normal samples). Combining these four parameters, the weight is calculated using the following formula. :

[0052] In this scenario, the sample will receive a very high weight value. The calculation result of this weight value indicates that this sample is a key sample located at the classification boundary and sparsely distributed. In the subsequent sample generation steps, the system will assign it a very conservative (nearly linear) generation region to ensure that the newly generated sample can accurately strengthen the key boundary without intruding into the normal sample region or generating meaningless noise samples.

[0053] In addition, to ensure the comparability of weights between samples, weights Within the sub-cluster Standardize the process.

[0054] S203 involves selecting similar nearest neighbors for each fault sample that meet the spatial visibility criteria. This specifically includes sub-steps S2031~S2033: S2031, For each fault sample, calculate the K nearest neighbor samples of the fault sample; S2032, For each nearest neighbor sample, determine whether the nearest neighbor sample satisfies the spatial visibility condition. If the direction of the line connecting the current nearest neighbor sample and the current fault sample is not blocked by other nearest neighbor samples, then the current nearest neighbor sample is determined to satisfy the spatial visibility condition. S2033, the nearest neighbor samples that satisfy the spatial visibility condition are taken as the visible nearest neighbor set, that is, the same type of nearest neighbors.

[0055] For example, suppose an inner-circle fault sample point A is located in the feature space. Its nearest neighbors (K=5) include 3 inner-circle fault samples (B1, B2, B3) and 2 outer-circle fault samples. First, calculate the vector relationship between all nearest neighbor samples and point A. For inner-circle fault sample B1, if there exists another nearest neighbor sample X such that the dot product of vector AX and vector B1-X is negative, then inner-circle fault sample B1 is determined to be occluded by nearest neighbor sample X and is invisible. Inner-circle fault sample B2 satisfies that the dot product of vector AX and vector B2-X is ≥0 for all nearest neighbor samples X, then inner-circle fault sample B2 is included in the visible nearest neighbor set. This filtering operation ensures that new samples are generated only in the visible nearest neighbor direction with the same fault mode (such as inner-circle fault), avoiding the generation of invalid samples that confuse inner-circle and outer-circle fault features, thus ensuring the clarity of the physical meaning and diagnostic effectiveness of the generated samples.

[0056] S204, dynamically adjust the sample generation region based on the corresponding sample and its similar nearest neighbors according to the weight value, and synthesize new fault samples in the adjusted generation region to obtain a balanced sample set.

[0057] Understandably, the traditional SMOTE oversampling algorithm randomly generates new samples on the line segment between minority class samples and their nearest neighbors. However, when majority class samples exist between minority class samples and their nearest neighbors, the oversampling effect significantly decreases. If new samples are generated only near the central region of minority class samples, the final sample set's effectiveness in defining the classification boundary is greatly diminished. The adaptive geometric SMOTE algorithm proposed in this embodiment effectively avoids new samples falling into majority class regions by dynamically adjusting the geometric range of the generated samples, thereby achieving accurate expansion of sub-cluster boundaries.

[0058] In one possible implementation, step S204 specifically includes sub-steps S2041 to S2042.

[0059] S2041, Construct a hyperspherical generation region using the current fault sample and its visible nearest neighbor as base points, and perform radial compression adjustment on the hyperspherical generation region according to the weight value: The larger the weight value, the higher the compression ratio, and the closer the generated region is to the linear interval formed by the line connecting the two points. The smaller the weight value, the lower the compression ratio, and the closer the generated region is to the complete hyperspherical space.

[0060] In practice, firstly, the visible nearest neighbors of the current faulty sample are filtered out from the sub-cluster to which it belongs. VNN With the current faulty sample point as the center point, the nearest visible neighbor ( VNN Using points as boundary points, a hyperspherical region is constructed. In the research of neighborhood selection algorithms, VNNA method has been proposed to prevent long-edge connections between samples. The discriminant formula is defined as follows:

[0061] in, and express The K nearest visible neighbors. If satisfy ,but yes The visible nearest neighbor.

[0062] Then, based on the sample weights The hypersphere region is compressed to adjust its geometrically generated region. For example, Figure 6 This demonstrates the degree of change in the geometric region under different weight values. Weight The larger the value, the more significant the deformation, and the smaller the geometric region. When When = 1, the hypersphere is compressed into a spatial straight line, and the generated region of the sample is completely consistent with the original sample; when When =0, the hypersphere remains unchanged.

[0063] different The effect of changes in geometric region, such as Figure 6 As shown, it is presented in two-dimensional form. Figure 6 The three subplots, from left to right, represent the values ​​of the sample weights. The changes in the corresponding generated regions (shown as dashed ellipses) when the values ​​are 0.3, 0.5, and 0.8.

[0064] For example, combining Figure 6 Taking a specific bearing outer ring fault sample point (blue point) and its nearest visible neighbor (yellow point) as an example. From step S202, the weight value is... It is calculated from its sample distribution. When the value is 0.3, it indicates that the sample is likely a safe sample located within the fault category and far from the classification boundary. Therefore, the system assigns it a large, approximately hyperspherical generation region (e.g., ...). Figure 6 The left-hand diagram shows a relatively flattened ellipse, allowing new samples to be synthesized through linear interpolation over a wider range, effectively increasing the diversity of generated samples and enriching the internal feature representation of the fault mode. When When the value is 0.8, it indicates that the sample is a critical sample extremely close to the classification boundary. To prevent the generation of invalid or boundary-confusing samples, the system highly compresses the generated region, making it approximate the line connecting the sample and its two nearest neighbors. Figure 6The right-hand diagram shows a highly compressed narrow ellipse. This conservative strategy ensures that the generated new samples closely follow the original boundary sample distribution, thereby sharpening and solidifying the classification decision boundary. Figure 6 The middle The subgraph corresponding to 0.5 represents an intermediate state.

[0065] This embodiment uses weights Driven adaptive geometry adjustment enables precise protection and reinforcement of classification boundaries while enhancing data diversity, which is key to improving the classification accuracy of imbalanced data.

[0066] S2042, new fault samples are generated within the adjusted hyperspherical generation region using a linear interpolation method, so that the generated samples are evenly distributed within the adjusted hyperspherical generation region, resulting in a balanced sample set; wherein, the number of new fault samples generated is dynamically determined based on the difference in the number of samples of each category in the original sample set.

[0067] like Figure 7 The diagram illustrates the generation of new samples using the adaptive geometric region adjustment method in a specific implementation scenario. Figure 7 As shown, the original minority class of fault samples (Green / blue triangles) are few in number and distributed within their feature space, while most normal samples are... (Yellow circles) are the most numerous. For each green minority class sample, the algorithm uses the adjusted hyperspherical region determined in step S2042 ( Figure 7 Based on the dashed ellipse (illustrated in the middle), new fault samples are generated within this geometric region by performing linear interpolation with its nearest visible neighbor. (See Figure 7 (The red triangle in the middle).

[0068] from Figure 7 It can be seen that the newly generated red triangle sample All samples are strictly located within the dashed elliptical region of their corresponding green sample, ensuring the reliability of the new sample generation process. Furthermore, these red triangular samples exhibit a relatively uniform distribution within the region, rather than clustering at a single point, effectively increasing the diversity of minority class samples. Figure 8 The number of red triangle samples is significantly greater than that of the original green samples. The number of red triangle samples generated is dynamically calculated based on the huge difference in the number of blue majority class and green minority class samples in the original dataset. This allows the total number of red, green and blue triangle samples to approach the number of yellow circle samples, thus ultimately obtaining a class-balanced sample set, providing a data foundation for subsequent training of high-performance fault diagnosis / classification models.

[0069] After obtaining the balanced sample set through step S2, the sample set can be applied to the machine tool bearing diagnosis application scenario through step S3.

[0070] In one possible embodiment, step S3 specifically includes sub-steps S301 to S304.

[0071] S301, the balanced sample set is divided into a training set and a validation set, where the training set is used for model parameter learning and the validation set is used for hyperparameter tuning; S302, Construct a fault diagnosis model based on deep learning. The fault diagnosis model includes a feature extraction module and a classification decision module. The feature extraction module adopts a convolutional neural network structure to adapt to the local correlation features of vibration signals. S303, the fault diagnosis model is iteratively trained using the training set, and the hyperparameters of the fault diagnosis model are dynamically adjusted based on the classification performance on the validation set. S304, The trained fault diagnosis model is applied to the operating data of the bearing to be diagnosed, and the health status classification result of the bearing to be diagnosed is output.

[0072] To comprehensively verify the effectiveness and generalization ability of the technical solution of this invention, four commonly used and representative machine learning classification models were selected for experiments in this implementation scenario, including decision tree (DT), support vector classifier (SVC), random forest (RF), and gradient boosting decision tree (GBDT). The method and related experiments were completed on a Windows 10 system using Python 3.9 and an NVIDIA RTX 2080Ti graphics card.

[0073] To evaluate the effectiveness of the technical solution of this invention in fault diagnosis, tests were conducted on two bearing datasets. The datasets used included bearing data from Case Western Reserve University (CWRU) and Southeast University (SEU). Table 2 summarizes the relevant experimental data from the datasets and details the sample distribution at different proportions.

[0074] Table 2 Characteristics of bearing datasets at different scales

[0075] The CWRU bearing dataset is one of the benchmark datasets in the field of mechanical fault diagnosis. Its corresponding experimental platform includes a motor, drive-end bearings, fan-end bearings, and an adjustable load device. Single-point faults of different sizes are introduced into the inner ring, outer ring, and rolling elements of the bearings using electrical discharge machining (EDM) technology, with fault diameters of 0.007, 0.014, and 0.021 mm, respectively. The experiment selects drive-end vibration data including normal conditions, inner ring faults, outer ring faults, and rolling element faults for analysis, with a sampling frequency of 12 kHz and a power of 0 horsepower (HP). To simulate unbalanced operating conditions, this implementation scenario adjusts the category distribution ratio, changing the number of normal samples and fault samples (including different degrees of damage).

[0076] The SEU bearing dataset is an important data source for the research and development of fault diagnosis algorithms. Its corresponding experimental platform consists of a motor, motor controller, planetary gearbox, reduction gearbox, brake, and brake controller. The experiment selects vibration data including normal operating conditions, inner race faults, outer race faults, and rolling element faults, with a sampling frequency of 5120 Hz, a rotational speed of 1200 rpm, and a load condition of 0 N·m. This implementation scenario aims to evaluate the performance of the fault diagnosis algorithm under unbalanced data conditions; the experiment is configured with a normal sample size and various fault samples with different imbalance ratios.

[0077] In all the following experiments, parameter K is set to 5. Under the same conditions, the method proposed in this invention is compared with six other methods (NONE (no sample balancing), SMOTE (synthetic minority oversampling technique), BorderSMOTE (Borderline-SMOTE, boundary-sensitive SMOTE), DBSMOTE (Density-Based Synthetic Minority Over-sampling Technique), MWMOTE (Majority Weighted Minority Oversampling Technique), and ADASYN (Adaptive Synthetic Sampling)) and 10-fold cross-validation is used.

[0078] In imbalanced learning, traditional accuracy cannot accurately reflect the classification performance of bearing fault diagnosis models on minority class samples. Therefore, this implementation scenario supplements evaluation metrics such as recall, F1 score, and G-means to more comprehensively reflect the model's performance across different classes.

[0079] Experiment 1: To address three different imbalance ratios, this experiment employed the method proposed in this invention to oversample the fault samples in the training set. The balanced dataset obtained after oversampling was then used to train four fault diagnosis / classification models: Decision Tree (DT), Support Vector Classifier (SVC), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT). Subsequently, these four trained fault diagnosis / classification models were applied to classify the imbalanced dataset in the test set to verify the performance of the proposed method in unbalanced bearing fault diagnosis. The results are shown in Tables 3, 4, and 5.

[0080] Table 3 Comparison test results for dataset A

[0081] Table 4 Comparison test results for dataset B

[0082] Table 5 Comparison test results for dataset C

[0083] Figure 8 This provides a more intuitive demonstration of the average performance of the oversampling method for datasets A, B, and C, corresponding to the results in Tables 3-5. Tables 3-5 and... Figure 8 The research results show that, tested with four classification models, the method provided in this invention outperforms other oversampling techniques in most cases. Referring to Table 3, taking dataset A (imbalance ratio 10:1) as an example, the method of this invention achieves accuracies of 97.31%, 98.42%, 98.87%, and 98.02% on the four models, respectively. When using the DT classification model, compared to the baseline dataset without oversampling, this method improves accuracy, recall, F1 score, and mean G by 1.17%, 1.83%, 1.86%, and 0.94%, respectively.

[0084] As shown in Tables 4 and 5, similar results were observed in the experimental results for datasets B and C. In dataset B, the accuracy of the proposed method on the four models reached 97.68%, 99.08%, 98.77%, and 98.38%, respectively. When using the DT classification model, the method improved the accuracy of the four metrics by 0.28%, 0.5%, 0.6%, and 0.37%, respectively. In dataset C, the accuracy of the proposed method on the four models was 97.05%, 98.99%, 98.12%, and 97.08%, respectively. After using the DT classification model, the method improved the accuracy of the four metrics by 0.34%, 0.48%, 0.41%, and 0.48%, respectively. The results show that the proposed method improves the classification accuracy to a certain extent compared with other methods in most cases. This method enhances the classification model's focus on faulty samples and avoids excessive bias towards normal samples. Therefore, it can accurately distinguish all samples, effectively alleviating the problem of misjudging faulty samples caused by data imbalance.

[0085] To comprehensively evaluate the performance of various oversampling methods and mitigate the impact on classification models, the results of different classifiers under the same data conditions were averaged. These average results are as follows: Figure 8 As shown. Figure 9 The results show that the proposed method exhibits superior overall performance compared to other oversampling methods illustrated in the experiments. Based on experimental data, this method demonstrates excellent performance on the CWRU dataset, effectively mitigating data imbalance by implementing appropriate oversampling for fault samples with different imbalance ratios. The empirical results in Experiment 1 demonstrate that this method not only improves the representation quality of a minority of fault samples but also optimizes overall classification performance, thereby enhancing the reliability and accuracy of fault diagnosis for imbalanced distributions in bearing datasets.

[0086] Experiment 2: In Experiment 2, multiple classifiers were used to evaluate the SEU bearing dataset with different sample proportions, and the results are listed in Tables 6, 7, and 8. In dataset D, this experiment again selected four different classification models (DT, SVC, RF, and GBDT) for training and performed classification validation on the test dataset.

[0087] Table 6 Comparison test results for dataset D

[0088] Table 7 Comparison test results for dataset E

[0089] Table 8 Comparison test results for dataset F

[0090] As shown in Table 6, taking dataset D as an example, the accuracy of the proposed method under the four models reached 96.23%, 98.15%, 97.52%, and 96.82%, respectively. When using the DT classification model, compared with the unsampled dataset, the proposed method improved accuracy, recall, F1 score, and G-mean by 1.57%, 3.03%, 3.1%, and 3.43%, respectively. As shown in Table 7, for dataset E, when using the DT classification model, compared with the unsampled dataset, the proposed method improved accuracy, recall, F1 score, and G-mean by 1.1%, 1.55%, 1.54%, and 1.6%, respectively. As shown in Table 8, for dataset F, when using DF as the classification model, compared with the unsampled dataset, the method proposed in this invention improves the accuracy, recall, F1 score and mean G by 0.75%, 0.99%, 0.97% and 1.01% respectively.

[0091] Experiment 2 shows that the proposed method outperforms other methods in most cases. This method effectively avoids excessive bias towards normal samples by enhancing the classification model's attention to faulty samples. To comprehensively evaluate the performance of different oversampling methods and reduce the impact on the classification model, this experiment averaged the results of various classifiers under the same data conditions. Figure 9 This provides a more intuitive demonstration of the average performance of the oversampling method for datasets A, B, and C, corresponding to the results in Tables 6-8. (See Tables 6-8 for details.) Figure 10 The experimental data shown demonstrates that the method of this invention performs excellently on the SEU dataset, effectively mitigating data imbalance by performing precise oversampling on fault samples with varying imbalance ratios. The method of this invention significantly outperforms the other six oversampling methods in overall performance. Empirical results indicate that the method of this invention not only improves the representation quality of a small number of fault samples but also optimizes the overall classification performance, thereby enhancing the reliability and accuracy of fault diagnosis in bearing datasets.

[0092] Experiments 1 and 2 above delve into the problem of fault diagnosis under imbalanced data conditions, aiming to address the shortcomings of classification models that over-rely on majority class samples from a data perspective. Experiments were conducted on the CWRU and SEU bearing datasets, and compared with various traditional oversampling methods. The experimental results show that when fault samples are extremely scarce, compared with other oversampling methods, the method of this invention can generate higher quality samples, significantly improve the diversity of sample distribution, and exhibit superior performance on multiple evaluation metrics.

[0093] Figure 10A structural diagram of a machine tool bearing diagnostic system based on an improved SMOTE is provided for an embodiment of the present invention, as shown below. Figure 11 As shown, a machine tool bearing diagnostic system based on an improved SMOTE includes a data preparation module, a sample balancing processing module, and a fault diagnosis application module, wherein: The data preparation module is used to acquire the operating data of the machine tool bearing and process the operating data into a raw sample set of feature data containing multiple samples, wherein each sample is labeled with a corresponding state category; The sample balancing processing module, connected to the data preparation module, is used to identify minority class fault samples in the original sample set. For each minority class fault sample, a weight value is calculated based on its sample distribution in the feature space. Based on the weight value, the corresponding sample generation region is adaptively determined for each minority class fault sample. New minority class fault samples are synthesized in the determined sample generation region to obtain a balanced sample set. The fault diagnosis application module, connected to the sample balancing processing module, includes a deep learning-based fault diagnosis model, which is used to train the fault diagnosis model using the balanced sample set and to classify the state of the input bearing operating data to be diagnosed.

[0094] It is understood that the machine tool bearing diagnostic system based on improved SMOTE provided by the present invention corresponds to the machine tool bearing diagnostic method based on improved SMOTE provided in the foregoing embodiments. The relevant technical features of the machine tool bearing diagnostic system based on improved SMOTE can be referred to the relevant technical features of the machine tool bearing diagnostic method based on improved SMOTE, and will not be repeated here.

[0095] Please see Figure 11 , Figure 11 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 12 As shown, an embodiment of the present invention provides an electronic device, including a memory 1110, a processor 1120, and a computer program 1111 stored in the memory 1110 and executable on the processor 1120. When the processor 1120 executes the computer program 1111, it performs the following steps: S1, acquire the operating data of the machine tool bearing and process it into a raw sample set of feature data containing multiple samples, wherein the samples are labeled with corresponding state categories; S2, identify minority class fault samples in the original sample set, calculate a weight value for each minority class fault sample based on its sample distribution in the feature space; based on the weight value, adaptively determine the corresponding sample generation region for each minority class fault sample, and synthesize new minority class fault samples in the determined sample generation region to obtain a balanced sample set. S3 uses a balanced sample set to train a deep learning-based fault diagnosis model to classify the operating data of the bearing to be diagnosed.

[0096] Please see Figure 12 , Figure 12 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by the present invention. (See diagram below.) Figure 1 As shown, this embodiment provides a computer-readable storage medium 1200, on which a computer program 1111 is stored. When the computer program 1111 is executed by a processor, it performs the following steps: S1, acquire the operating data of the machine tool bearing and process it into a raw sample set of feature data containing multiple samples, wherein the samples are labeled with corresponding state categories; S2, identify minority class fault samples in the original sample set, calculate a weight value for each minority class fault sample based on its sample distribution in the feature space; based on the weight value, adaptively determine the corresponding sample generation region for each minority class fault sample, and synthesize new minority class fault samples in the determined sample generation region to obtain a balanced sample set. S3 uses a balanced sample set to train a deep learning-based fault diagnosis model to classify the operating data of the bearing to be diagnosed.

[0097] This invention provides a machine tool bearing diagnosis method, system, electronic device, and storage medium based on an improved SMOTE. Addressing the data imbalance problem caused by the scarcity of vibration signal fault samples in machine tool bearing fault diagnosis, the invention first performs sliding window segmentation, time-domain and frequency-domain feature extraction, and dimensionality reduction on the vibration signals to construct a feature sample set. In the sample balancing stage, minority fault samples are divided into sub-clusters through clustering. For each sample, four parameters are calculated to quantify its sample distribution weight: distance to the sub-cluster center, distance to the normal class center, local intra-class density, and inter-class density. Based on these weights, the visible nearest neighbors of each sample are selected, and a generation region is constructed accordingly: boundary samples with high weight values ​​are generated using highly compressed, near-linear generation regions, while internal samples with low weight values ​​are generated using more relaxed regions closer to a hypersphere. Finally, new fault samples are synthesized through linear interpolation within this adaptively adjusted region. The balanced sample set is then used to train a deep learning fault diagnosis model to achieve accurate bearing fault identification.

[0098] This invention adaptively adjusts the sample generation region based on sample distribution weights, achieving focused enhancement of key regions at the classification boundary and supplementation of diversity within the sample distribution region when generating new samples. Comparative experiments on public bearing datasets such as CWRU demonstrate that, compared to traditional methods such as SMOTE and Borderline-SMOTE, this invention significantly improves various classification models, including gradient boosting decision trees, across multiple evaluation metrics. For example, with a 10:1 imbalance ratio, accuracy, recall, and F1 score are all superior to the baseline, effectively reducing the false negative rate of minority class faults and improving the overall performance and robustness of the diagnostic model in real-world imbalanced data scenarios. It is suitable for online or offline bearing fault identification and classification.

[0099] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes ​ The steps of the function specified in one or more boxes.

[0104] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0105] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A machine tool bearing diagnostic method based on an improved SMOTE, characterized in that, include: S1, acquire the operating data of the machine tool bearing and process it into a raw sample set of feature data containing multiple samples, wherein the samples are labeled with corresponding state categories; S2, identify minority class fault samples in the original sample set, calculate a weight value for each minority class fault sample based on its sample distribution in the feature space; based on the weight value, adaptively determine the corresponding sample generation region for each minority class fault sample, and synthesize new minority class fault samples in the determined sample generation region to obtain a balanced sample set. S3 uses a balanced sample set to train a deep learning-based fault diagnosis model to classify the operating data of the bearing to be diagnosed.

2. The machine tool bearing diagnostic method based on improved SMOTE according to claim 1, characterized in that, Step S1 includes: S101, acquire the operating data of the machine tool bearing, the operating data including vibration signals; use a sliding window to divide the continuous vibration signals into multiple signal segments, wherein the window length and sliding step size are preset according to the bearing fault characteristic frequency; S102, based on the segmented signal segments, extract time-domain features and frequency-domain features respectively. The time-domain features include at least three of the following: mean, root mean square, peak value, kurtosis, and impulse index. The frequency-domain features include at least two of the following: spectral centroid, frequency root mean square, and frequency variance. These features are combined to form a feature vector characterizing the bearing's operating state. S103, the feature vector is subjected to dimensionality reduction processing to remove redundant information and retain key features that distinguish different state categories, forming an original sample set for bearing fault classification.

3. The machine tool bearing diagnostic method based on improved SMOTE according to claim 2, characterized in that, In step S103, the dimensionality reduction of the eigenvectors is performed using principal component analysis, including: Calculate the eigenvalues ​​of the covariance matrix of the eigenvectors, sort the corresponding principal components by eigenvalue from largest to smallest, and retain the principal components whose cumulative variance contribution rate exceeds a preset threshold.

4. The machine tool bearing diagnostic method based on improved SMOTE according to claim 1, characterized in that, Step S2 includes: S201, Identify the minority class of fault samples from the original sample set and perform clustering processing to form multiple sub-clusters corresponding to different fault characteristics; S202, For each fault sample, calculate the weight value based on its distribution density within its class and its separation from samples of different classes; S203, for each fault sample, select similar nearest neighbors that meet the spatial visibility condition; S204, dynamically adjust the sample generation region based on the corresponding sample and its similar nearest neighbors according to the weight value, and synthesize new fault samples in the adjusted generation region to obtain a balanced sample set.

5. A machine tool bearing diagnostic method based on improved SMOTE according to claim 4, characterized in that, Step S201 specifically includes: The minority class fault samples in the original sample set are identified and denoised. Based on the K-means clustering algorithm, the denoised minority class fault samples are divided into multiple sub-clusters, and each sub-cluster corresponds to the local feature distribution of a bearing fault mode.

6. The machine tool bearing diagnostic method based on improved SMOTE according to claim 5, characterized in that, Step S202 specifically includes: For each faulty sample in a subcluster, a weight value is calculated based on its intra-class distance to the subcluster center, inter-class distance to the majority class sample center, local intra-class density, and local inter-class density. Among these, the boundary neighbor samples with smaller inter-class distance and lower intra-class density receive higher weights.

7. A machine tool bearing diagnostic method based on improved SMOTE according to claim 4 or 5, characterized in that, Step S203 specifically includes: S2031, For each fault sample, calculate the K nearest neighbor samples of the fault sample; S2032, For each nearest neighbor sample, determine whether the nearest neighbor sample satisfies the spatial visibility condition. If the direction of the line connecting the current nearest neighbor sample and the current fault sample is not blocked by other nearest neighbor samples, then the current nearest neighbor sample is determined to satisfy the spatial visibility condition. S2033, the nearest neighbor samples that satisfy the spatial visibility condition are taken as the visible nearest neighbor set, that is, the same type of nearest neighbors.

8. A machine tool bearing diagnostic method based on improved SMOTE according to claim 7, characterized in that, Step S204 specifically includes: A hyperspherical generation region is constructed using the current fault sample and its nearest visible neighbor as base points. The hyperspherical generation region is then radially compressed and adjusted according to the weight value. The larger the weight value, the higher the compression ratio, and the closer the generated region is to the linear interval formed by the line connecting the two points. The smaller the weight value, the lower the compression ratio, and the closer the generated region is to the complete hyperspherical space; New fault samples are generated using a linear interpolation method within the adjusted hyperspherical generation region, ensuring that the generated samples are evenly distributed within the adjusted hyperspherical generation region, resulting in a balanced sample set. The number of new fault samples generated is dynamically determined based on the differences in the number of samples of each category in the original sample set.

9. A machine tool bearing diagnostic method based on improved SMOTE according to claim 8, characterized in that, Step S3 specifically includes: S301, the balanced sample set is divided into a training set and a validation set, where the training set is used for model parameter learning and the validation set is used for hyperparameter tuning; S302, Construct a fault diagnosis model based on deep learning. The fault diagnosis model includes a feature extraction module and a classification decision module. The feature extraction module adopts a convolutional neural network structure to adapt to the local correlation features of vibration signals. S303, the fault diagnosis model is iteratively trained using the training set, and the hyperparameters of the fault diagnosis model are dynamically adjusted based on the classification performance on the validation set. S304, The trained fault diagnosis model is applied to the operating data of the bearing to be diagnosed, and the health status classification result of the bearing to be diagnosed is output.

10. A machine tool bearing diagnostic system based on an improved SMOTE, characterized in that, include: The data preparation module is used to acquire the operating data of the machine tool bearing and process the operating data into a raw sample set of feature data containing multiple samples, wherein each sample is labeled with a corresponding state category; The sample balancing processing module, connected to the data preparation module, is used to identify minority class fault samples in the original sample set. For each minority class fault sample, a weight value is calculated based on its sample distribution in the feature space. Based on the weight value, the corresponding sample generation region is adaptively determined for each minority class fault sample. New minority class fault samples are synthesized in the determined sample generation region to obtain a balanced sample set. The fault diagnosis application module, connected to the sample balancing processing module, includes a deep learning-based fault diagnosis model, which is used to train the fault diagnosis model using the balanced sample set and to classify the state of the input bearing operating data to be diagnosed.

Images