An adaptive-based bearing fault detection method and apparatus

Abstract

The application discloses a bearing fault detection method and device based on self-adaption, and belongs to the field of bearing data processing. The method comprises the following steps: obtaining target bearing monitoring data in the same working condition interval, extracting vibration spectrum features, and obtaining a target feature vector; obtaining a first similarity with each state category based on the clustering distribution of each state category in the historical reference data set of the working condition interval; when the first similarity is lower than a similarity threshold, determining that the target feature vector represents a new unknown state which is not recorded in the first historical reference data set; performing cross-working condition clustering distribution correlation analysis on the target feature vector based on the clustering distribution of each state category in the historical reference data set of the adjacent working condition interval, obtaining a second similarity with each state category in the adjacent working condition interval, and determining whether the bearing has a fault in the working condition interval. The method can adapt to working condition changes, effectively identify known faults with feature migration, and autonomously find new unknown states.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an adaptive bearing fault detection method and device. Background Technology

[0002] As a core component of rotating machinery, the health of bearings directly affects the safety, stability, and operational efficiency of the entire industrial equipment system. Therefore, developing efficient and accurate bearing fault detection methods is of great significance for predictive maintenance, avoiding unplanned downtime, and preventing major safety accidents.

[0003] Currently, mainstream methods typically collect historical data under fixed operating conditions to build a fault feature database or train a network classification model, and then compare or classify real-time monitoring data to achieve condition identification. For example, current intelligent diagnostic methods based on vibration signal analysis usually establish a correspondence between fault features and conditions under fixed operating conditions for condition identification.

[0004] However, in real-world industrial scenarios, equipment operating conditions (such as speed and load) often change over time and with production demands, causing systematic changes in bearing characteristics. This presents existing technologies with the following core challenges: First, the constructed general bearing fault detection models are difficult to apply to different operating conditions, exhibiting significant performance degradation—a condition sensitivity problem—leading to numerous false alarms or missed alarms. Second, when equipment experiences entirely new and unknown fault modes, existing systems lack effective reasoning and discovery mechanisms—a weak adaptability to unknown states. They either cannot make a judgment or suffer from misjudgments due to forced matching, failing to achieve adaptive knowledge expansion and state discovery. Summary of the Invention

[0005] To address the aforementioned problems, this application proposes an adaptive bearing fault detection method, comprising:

[0006] Acquire monitoring data of target bearings within the same operating condition range; the time interval between acquisitions of target bearing monitoring data is less than a preset interval threshold.

[0007] Vibration spectrum features are extracted from the monitoring data of the target bearing to obtain the target feature vector;

[0008] In the benchmark operating condition interval knowledge base, the first historical benchmark dataset of the operating condition interval is retrieved; the first historical benchmark dataset stores historical feature vectors that have been clustered by state category.

[0009] Based on the clustering distribution of each known state category in the first historical benchmark dataset, the metric distance from the target feature vector to the cluster center of each category is calculated to obtain the first similarity between the target feature vector and each known state category;

[0010] When the first similarity between the target feature vector and each known state category is lower than a preset similarity threshold, the target feature vector is determined to represent a new unknown state not recorded in the first historical benchmark dataset.

[0011] In the knowledge base of the benchmark working condition interval, retrieve the second historical benchmark dataset of the adjacent working condition intervals of the working condition interval; the adjacent working condition interval refers to the working condition interval with the smallest Euclidean distance to the center point of the working condition interval in the working condition parameter space.

[0012] Based on the clustering distribution of each known state category in the second historical benchmark dataset, cross-working condition clustering distribution correlation analysis is performed on the target feature vector to obtain the second similarity between the target feature vector and each known state category in adjacent working condition intervals.

[0013] Based on the second similarity with each known state category in adjacent operating condition intervals, pattern recognition is performed on the target feature vector to determine whether the bearing has a fault in the operating condition interval.

[0014] On the other hand, embodiments of this application provide an adaptive bearing fault detection device, including:

[0015] At least one processor; and,

[0016] A memory communicatively connected to the at least one processor; wherein,

[0017] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform any of the aforementioned adaptive bearing fault detection methods.

[0018] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0019] This approach fundamentally solves the problem of operating condition dependence, achieving true adaptive diagnostics. Traditional methods often assume that equipment operates under constant conditions or ignore the operating context during feature analysis, leading to drift in fault characteristics with changes in speed and load, resulting in false alarms or missed alarms. This application obtains target monitoring data corresponding to the same operating condition range through core premises, strictly anchoring each analysis to an operating state within a specific physical context. This eliminates interference caused by differences in the operating conditions themselves, laying a solid foundation for subsequent accurate comparisons.

[0020] A two-layer intelligent diagnostic mechanism, prioritizing the current operating condition and providing a fallback across different operating conditions, has been established, significantly improving the coverage and reliability of state identification. First, within the current operating condition interval, matching is performed by calculating the metric distance to known state cluster centers (first-layer diagnosis). Cross-operating condition correlation analysis is triggered only when a complete match is not found in the current knowledge base (second-layer diagnosis). This mechanism has dual advantages: First, when there is sufficient experience with the current operating condition, identification can be completed quickly and efficiently, with optimal computational resources. Second, when current operating condition data is sparse or a completely new pattern is encountered, the system does not fail but can instead infer by retrieving historical experience from adjacent operating condition intervals and utilizing the continuity of the fault's physical mechanism, significantly enhancing its diagnostic survivability in unknown or marginal situations.

[0021] Cross-operational condition knowledge transfer is achieved through a physically interpretable quantitative model, ensuring the rationality and accuracy of the correlation analysis. The core challenge of cross-operational condition analysis lies in establishing comparability between features under different operating conditions. This approach avoids using uninterpretable black-box mappings, instead defining adjacent operating condition intervals as those with the smallest Euclidean distance in the operating condition parameter space, and using this as the basis for correlation analysis. The smallest Euclidean distance implies that the operating conditions (such as speed and load combinations) of the two conditions are most similar, and the changes in vibration response should theoretically follow the most continuous and predictable patterns. This makes cross-operational condition clustering and correlation analysis not a blind data comparison, but a precise knowledge transfer guided by strong physical mechanisms, greatly improving the credibility of the cross-operational condition identification results. Attached Figure Description

[0022] To more clearly illustrate the technical solution of this application, some embodiments of this application will be described in detail below with reference to the accompanying drawings, in which:

[0023] Figure 1 A flowchart illustrating an adaptive bearing fault detection method provided in this application embodiment;

[0024] Figure 2 This is a schematic diagram of the structure of an adaptive bearing fault detection device provided in an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Some embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0027] Figure 1 This is a flowchart illustrating an adaptive bearing fault detection method provided in an embodiment of this application. The process can be executed by a computing device in the relevant field, and certain input parameters or intermediate results in the process can be manually adjusted to help improve accuracy.

[0028] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server, and this application does not impose any special limitations on it. For ease of understanding and description, the following embodiments are all described in detail using a server as an example.

[0029] It should be noted that the server can be a single device or a system composed of multiple devices, i.e., a distributed server. This application does not make any specific limitations in this regard.

[0030] Figure 1 The process includes the following steps:

[0031] S101: Acquire target bearing monitoring data within the same operating condition range; the time interval between acquisitions of target bearing monitoring data is less than a preset interval threshold.

[0032] The target monitoring data is a time-series signal recorded at a fixed sampling interval. The time interval between the acquisitions of the target bearing monitoring data is less than the preset interval threshold, thereby avoiding significant interruptions in data acquisition or crossing different operating condition ranges. This ensures the integrity and continuity of the information, and ensures that the micro-dynamic characteristics of the bearing condition can be captured and characterized, such as the periodic impact components and modulation sidebands caused by local defects (pitting, spalling), which are key to identifying early failures.

[0033] Based on historical equipment operating data and process knowledge, multiple discrete and stable operating state regions are pre-divided within the operating condition parameter space (usually with speed and load as the core dimensions). For example: Range 1: Speed ​​1480-1520 RPM, Load 65-75%; Range 2: Speed ​​980-1020 RPM, Load 45-55%. Each range represents a specific and repeatable steady-state operating mode of the equipment.

[0034] Based on this, bearing monitoring signals and operating parameter signals (such as speed and load values ​​from the PLC) are received synchronously. Through correlation analysis, continuous monitoring data segments that consistently fall within the boundaries of the same predefined operating condition interval are identified and extracted. If a segment of bearing monitoring data spans different operating condition intervals, it will be divided into multiple sub-segments and processed separately at the operating condition switching point.

[0035] In summary, this is the cornerstone of achieving adaptive diagnostics based on operating conditions. The vibration characteristics of bearings (such as the characteristic frequencies, amplitudes, and modulation modes of fault modes) are strongly correlated with operating conditions. Strictly limiting the analysis to the same operating condition range ensures that comparisons between subsequently extracted feature vectors and historical benchmark datasets are performed under the same operational physical context. This eliminates the interference caused by changes in operating conditions themselves on bearing characteristics, greatly improving the accuracy and reliability of condition identification and fault diagnosis.

[0036] S102: Extract vibration spectrum features from the monitoring data of the target bearing to obtain the target feature vector.

[0037] A fast Fourier transform is performed on the target bearing monitoring data (time-domain vibration signal) to obtain the frequency-amplitude spectrum.

[0038] For the frequency-amplitude spectrum, the amplitudes corresponding to all frequency points within the range of a preset start frequency to a preset cutoff frequency (e.g., 0Hz to the bearing Nyquist frequency) are arranged in ascending order of frequency to obtain the target feature vector. The bearing Nyquist frequency refers to the highest analyzable frequency determined by the sampling frequency when sampling the bearing vibration signal. For example, if the sensor collects bearing vibration signals at 12800 data points per second (12.8kHz), the Nyquist frequency is: Fn = 12800Hz / 2 = 6.4kHz.

[0039] In summary, the vibration signal is transformed into a unified, digital state indicator. The generation of the target feature vector is the foundation for all subsequent functions, such as adaptive matching of operating conditions, discovery of unknown states, and intelligent diagnosis.

[0040] S103: Retrieve the first historical benchmark dataset of the operating condition interval from the benchmark operating condition interval knowledge base; the first historical benchmark dataset stores historical feature vectors that have been clustered according to state categories.

[0041] In this example, a baseline operating condition interval knowledge base is pre-built. Within the baseline operating condition interval knowledge base, independent datasets are created for different operating condition intervals such as high speed-low load, medium speed-medium load, and low speed-high load.

[0042] Using the operating condition interval as the query key, a search is performed in a pre-built benchmark operating condition interval knowledge base to obtain the first historical benchmark dataset matching the operating condition interval.

[0043] It should be noted that the historical benchmark dataset of the benchmark operating condition interval knowledge base includes a set of feature vectors collected historically within the same operating condition interval, and whose states (such as health, outer race fatigue, poor lubrication, etc.) have been accurately labeled. The historical benchmark dataset does not store all samples haphazardly. During the offline learning or initialization phase, clustering algorithms (such as K-means, Gaussian mixture models) have been used to inductively learn the samples for each type of state, forming a data cluster distribution. Each known fault state and health state corresponds to one or more data clusters, and each cluster can be characterized by its cluster center (the mean of the feature vectors) and distribution shape (such as the covariance matrix).

[0044] It should be noted that the historical benchmark dataset corresponding to the benchmark operating condition interval knowledge base to which the bearing belongs is based on historical data collection of the bearing under test. Different bearings may correspond to different benchmark operating condition interval knowledge bases, and bearings within the same benchmark operating condition interval knowledge base have the same sampling frequency.

[0045] S104: Based on the clustering distribution of each known state category in the first historical benchmark dataset, calculate the metric distance from the target feature vector to the cluster center of each category to obtain the first similarity between the target feature vector and each known state category.

[0046] Clustering distributions represent typical patterns of various known states under specific operating conditions. They summarize discrete historical feature vectors into statistically significant probabilistic models, serving as a benchmark for quantitative comparison.

[0047] It should be noted that commonly used distance measurement methods include Euclidean distance, Mahalanobis distance, etc.

[0048] Euclidean distance calculates the straight-line distance between vectors. Mahalanobis distance considers not only the distance from a vector to its cluster center but also normalizes it using the covariance matrix of the cluster distribution. This means that it allows for larger deviations in directions where the data distribution is dispersed, while being more sensitive to deviations in directions where the data distribution is concentrated.

[0049] Based on this, a set of quantified distance values ​​can be obtained, each representing the degree of similarity (or difference) between the current state and a known state category. The smaller the value, the higher the similarity.

[0050] For example, based on the cluster distribution of each known state category in the first historical benchmark dataset, the Euclidean distance from the target feature vector to the cluster center of each category is calculated. The Euclidean distance is used as the distance metric. Then, the Euclidean distance is mapped to obtain the first similarity between the target feature vector and each known state category.

[0051] The first similarity can be calculated by inputting the metric distance into a preset monotonically decreasing function; the monotonically decreasing function is configured such that the output value decreases as the input metric distance increases. Alternatively, a matching can be performed based on a preset similarity lookup table to obtain the corresponding first similarity; the similarity lookup table defines the mapping relationship between different metric distance ranges and the first similarity.

[0052] For example, a monotonically decreasing function can be used to define a mapping relationship. A monotonically decreasing function is as follows:

[0053]

[0054] in, For similarity, A sensitivity coefficient used to adjust the conversion relationship between distance and similarity metrics (e.g., (1.2) To measure distance.

[0055] S105: When the first similarity between the target feature vector and each known state category is lower than a preset similarity threshold, the target feature vector is determined to represent a new unknown state that is not recorded in the first historical benchmark dataset.

[0056] When the first similarity between the target feature vector and each known state category is lower than the preset similarity threshold, it indicates that the target feature vector is sufficiently dissimilar to the typical patterns of each known state category. "Unrecorded" means that the state pattern of the target feature vector is not within the current baseline operating condition interval knowledge base (i.e., the historical baseline dataset under the corresponding operating condition interval). A new unknown state indicates that a completely new pattern has been identified.

[0057] S106: In the knowledge base of the benchmark working condition interval, retrieve the second historical benchmark dataset of the adjacent working condition intervals of the working condition interval; the adjacent working condition interval refers to the working condition interval with the smallest Euclidean distance to the center point of the working condition interval in the working condition parameter space.

[0058] Based on the continuity of the fault mechanism, the bearing condition exhibits local smoothing characteristics in the operating parameter space. Therefore, when the current operating condition interval cannot be identified, the system automatically retrieves the historical benchmark dataset of the nearest operating condition interval for analysis. This nearest neighbor matching strategy effectively balances computational complexity and diagnostic reliability, utilizing existing historical experience knowledge in the most efficient way.

[0059] It should be noted that the operating condition parameter space refers to a mathematical space defined using one or more equipment operating parameters (such as speed and load) that have the most significant impact on the bearing vibration state as coordinate axes. Each specific operating condition corresponds to a point in the mathematical space. An operating condition interval corresponds to a sub-region (such as a multi-dimensional cube) in the mathematical space. Calculating the Euclidean distance between the center points of two operating condition intervals is a geometric measurement in the mathematical space; the closer the distance, the more similar the operating conditions.

[0060] S107: Based on the clustering distribution of each known state category in the second historical benchmark dataset, perform cross-working condition clustering distribution association analysis on the target feature vector to obtain the second similarity between the target feature vector and each known state category in adjacent working condition intervals.

[0061] In this example, under the current operating condition space, there is a lack of direct knowledge of the target feature vector state. However, under adjacent operating conditions with similar physical conditions, accumulated historical experience may include related pattern clues. This is because the dynamic behavior and fault characterization of bearings are usually continuously and regularly correlated with operating parameters such as speed and load. For example, for the same fault category, the corresponding characteristic frequency will change proportionally at different speeds, or the vibration energy level will show a trend change with the load. Therefore, it is important to analyze the similarity between an unknown state and the known states of adjacent operating conditions.

[0062] Cross-working-condition clustering correlation analysis can directly calculate the metric distance between the target feature vector and the cluster centers of each known state category in the second historical benchmark dataset. This reveals the proximity relationship between unknown states and various known patterns under adjacent working conditions. Alternatively, the target feature vector can be further transformed by working condition to achieve working condition alignment, thereby calculating the metric distance from the aligned target feature vector to the cluster centers of each category in adjacent working-condition intervals, thus improving accuracy.

[0063] Furthermore, the distance metric is matched to obtain the corresponding second similarity.

[0064] It should be noted that mapping relationships can be implemented in various ways. For example, since a larger distance usually means a lower second similarity, a monotonically decreasing function is often used to define the mapping relationship. The monotonically decreasing function is related to S104.

[0065] Alternatively, the possible distance range can be divided into several intervals, and a similarity value can be directly assigned to each interval, forming a discrete similarity lookup table. It should be noted that the similarity lookup table establishes a one-to-one correspondence between distance metrics and similarity scores.

[0066] S108: Based on the second similarity with each known state category in adjacent operating condition intervals, perform pattern recognition on the target feature vector to determine whether the bearing has a fault in the operating condition interval.

[0067] In this example, the higher the second similarity to a known state category, the stronger the correlation between the target feature vector and the pattern of that known state category. Fault pattern recognition can be approached in two ways:

[0068] One approach is to determine the state category corresponding to the target feature vector as the highest second similarity level when it exceeds a similarity threshold (e.g., 0.85). When the highest second similarity level is below or equal to the similarity threshold, the abnormal state is reported to a human. Management users can identify the pattern state of the target feature vector through other operations. For example, the target feature vector can be input into a fault diagnosis neural network model for learning, thereby identifying the pattern state. Alternatively, the pattern state of the target feature vector can be analyzed by on-site personnel.

[0069] One approach is to select target state categories with a second similarity score higher than a similarity threshold, and then assess the score by considering the trend of the second similarity score with the target state category as the operating conditions change. For example, a weighted sum is calculated of the second similarity score with the target state category and the trend of the second similarity score changing with relevant operating parameters to obtain a comprehensive score. When the highest comprehensive score is greater than a scoring threshold (e.g., 0.75), the target state category with the highest comprehensive score is identified as the pattern state corresponding to the target feature vector. When the highest comprehensive score is lower than or equal to the scoring threshold, the abnormal state is reported to a human.

[0070] It should be noted that when a new failure mode is identified, the cluster distribution of the historical benchmark dataset for the operating condition range is updated to establish a cluster for the new failure mode.

[0071] It should be noted that, although the embodiments in this application are based on... Figure 1 Steps S101 to S108 will be described sequentially, but this does not mean that steps S101 to S108 must be performed in a strict order. The reason this embodiment follows this order is... Figure 1 The order in which steps S101 to S108 are described is provided to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, in the embodiments of this application, the order of steps S101 to S108 can be appropriately adjusted according to actual needs.

[0072] pass Figure 1The proposed method, by constructing a hierarchical, adaptive fault detection process with knowledge transfer capabilities, achieves highly robust and accurate identification of bearing conditions under complex industrial environments. Specifically, it is as follows:

[0073] This approach fundamentally solves the problem of operating condition dependence, achieving true adaptive diagnostics. Traditional methods often assume that equipment operates under constant conditions or ignore the operating context during feature analysis, leading to drift in fault characteristics with changes in speed and load, resulting in false alarms or missed alarms. This application obtains monitoring data corresponding to the same predefined operating condition range through a core premise, strictly anchoring each analysis to an operating state within a specific physical context. This ensures that the extracted target feature vector is fully comparable to the historical feature vector in the first historical benchmark dataset, eliminating interference caused by differences in the operating conditions themselves and laying a solid foundation for subsequent accurate comparisons.

[0074] A two-layer intelligent diagnostic mechanism, prioritizing the current operating condition and providing a fallback across different operating conditions, has been established, significantly improving the coverage and reliability of state identification. First, within the current operating condition interval, matching is performed by calculating the metric distance to the cluster centers of known state categories (first-layer diagnosis). Cross-operating condition clustering distribution association analysis is only triggered when a complete match is not found in the baseline operating condition interval knowledge base (second-layer diagnosis). This mechanism has two advantages: First, when there is sufficient experience with the current operating condition, identification can be completed quickly and efficiently, with optimal computational resources. Second, when current operating condition data is sparse or a completely new pattern is encountered, the system does not fail but can instead infer by retrieving historical experience from adjacent operating condition intervals and utilizing the continuity of the fault's physical mechanism, significantly enhancing its diagnostic survivability in unknown or marginal situations.

[0075] Cross-operational condition knowledge transfer is achieved through a physically interpretable quantitative model, ensuring the rationality and accuracy of the correlation analysis. The core challenge of cross-operational condition analysis lies in establishing comparability between features under different operating conditions. This approach avoids using uninterpretable black-box mappings, instead defining adjacent operating condition intervals as those with the smallest Euclidean distance in the operating condition parameter space, and using this as the basis for correlation analysis. The smallest Euclidean distance implies that the operating conditions (such as speed and load combinations) of the two conditions are most similar, and the changes in vibration response should theoretically follow the most continuous and predictable patterns. This makes cross-operational condition clustering and correlation analysis not a blind data comparison, but a precise knowledge transfer guided by strong physical mechanisms, greatly improving the credibility of the cross-operational condition identification results.

[0076] This system does not rely on the collection of massive fault samples across the entire operating range. Instead, it leverages limited historical data within each discrete operating condition interval and clear physical rules to build powerful adaptive diagnostic capabilities. This significantly reduces the stringent requirements for the completeness of training data, and in particular, it requires fewer computational resources compared to deploying neural network models. This makes the solution easy to deploy and apply in the early stages of data accumulation or in real-world industrial scenarios with diverse operating conditions, demonstrating high practical value and potential for widespread adoption.

[0077] In summary, by integrating operating condition consistency constraints, hierarchical diagnostic logic, and cross-operating condition correlation based on physical space, a bearing fault detection solution has been formed that can adapt to complex operating conditions, ensure accurate and reliable diagnosis, and has a strong ability to cope with unknown states.

[0078] based on Figure 1 In addition to the method described herein, this application also provides some specific implementation schemes and extension schemes of the method, which will be further explained below.

[0079] In one example, when the first similarity between the target feature vector and each known state category is at least not lower than a similarity threshold, the known state category with the highest first similarity is determined as the pattern state represented by the target feature vector.

[0080] Based on this, when the target feature vector is sufficiently close to at least one known state category, it can be directly classified within the current working condition range without the need for cross-working condition retrieval, thus improving efficiency.

[0081] In one example, acquiring monitoring data for target bearings within the same operating condition range includes the following steps:

[0082] Step 1: Obtain bearing monitoring data and equipment operating condition data of the bearing from the industrial equipment monitoring nodes.

[0083] In this example, an industrial equipment monitoring node refers to an edge computing gateway unit deployed in an industrial site, possessing data acquisition, processing, and communication capabilities. Bearing monitoring data refers to raw signals collected by physical sensors (such as accelerometers, temperature sensors, and current sensors) that directly act on the bearing or supporting structure, or signals that have undergone preliminary processing by the industrial equipment monitoring node. Equipment operating condition data refers to parameters describing the overall operating status and working conditions of the equipment, such as real-time speed and load information. Equipment operating condition data typically originates from the industrial equipment's control system (such as a PLC or frequency converter), encoder, or power meter. While these data are not physical quantities that directly measure the bearing's condition, they provide crucial context for understanding the bearing monitoring data.

[0084] Step 2: Perform time axis analysis on the bearing monitoring data and equipment operating condition data to obtain the bearing monitoring data sequence and the equipment operating condition data sequence.

[0085] It should be noted that timeline parsing refers to arranging the original, timestamped data packets in chronological order to form a time series that can be computed.

[0086] Step 3: Based on the preset master clock reference source and the maximum allowable clock drift threshold, perform drift calibration and alignment on the bearing monitoring data sequence and the equipment operating condition data sequence.

[0087] In this example, minor errors (such as drift) exist between the clocks of different devices. Using a high-precision master clock (such as a GPS or network synchronization clock) as a reference, the deviation of the timestamps for each data sequence is checked. If the deviation exceeds the maximum allowable threshold (e.g., ±50ms), a linear or non-linear correction is performed based on the master clock reference source to unify all data to the same time base. For example, from the clock deviation mapping table, it is found that the vibration sensor clock is 2 seconds slower than the master clock per day. Based on the master clock reference source, all timestamps of the vibration data are compensated for by speeding up.

[0088] Step 4: Establish the correlation between bearing monitoring data and equipment operating condition data in terms of acquisition time.

[0089] By establishing a precise time mapping relationship, the system can query the operating condition data corresponding to the collection time for any bearing monitoring data point, thereby establishing a synchronous correlation.

[0090] For example, a timestamp alignment algorithm can be used to find the operating condition data corresponding to the bearing monitoring data points.

[0091] For example, for a vibration point collected at time t, the effective time interval [t1, t2] to which t belongs can be found in the time series of equipment operating condition data.

[0092] If the operating conditions change frequently (for example, within the effective time interval, the speed change exceeds 5% of the preset speed change value or the load change exceeds 10% of the preset load change value between two adjacent operating condition sampling points), linear interpolation can be performed between the operating condition values ​​at t1 and t2 to estimate the operating condition at time t. The logic of linear interpolation is: assuming that the operating condition data changes at a constant rate within this effective time interval.

[0093] If the operating conditions do not change frequently, within the effective time interval, the operating condition data closest to time t is used as the operating condition data corresponding to the vibration point collected at time t.

[0094] Step 5: Determine the data segment whose corresponding equipment operating condition data continuously falls within the same operating condition range as the target bearing monitoring data.

[0095] Based on this, any feature observed in bearing monitoring data (such as vibration spectrum) can be clearly correlated with the specific operating state of the equipment at the time of acquisition (such as a specific combination of speed and load). This ensures that subsequent analysis can accurately establish the correlation between specific state patterns under specific operating conditions, avoiding misjudgments caused by data asynchrony.

[0096] In one example, since the feature space of the first historical benchmark dataset may include a large number of noisy features that are irrelevant to the fault mode corresponding to the current target feature vector, direct distance calculation is easily affected by irrelevant features, leading to decreased fault identification sensitivity and increased false positive rate. Furthermore, different fault modes have different degrees of dependence on various features, and certain key fault features (such as vibration energy in a specific frequency band) should have higher weights in distance calculation.

[0097] Based on this, and based on the clustering distribution of each known state category in the first historical benchmark dataset, the metric distance from the target feature vector to the cluster center of each category is calculated, including the following steps:

[0098] Step 1: For each known state category in the first historical benchmark dataset, the target feature vector is projected onto a low-dimensional sensitive subspace defined by the physical pattern most relevant to the current state category using a sensitive projection model for category learning.

[0099] The sensitive projection model is learned based on the vibration spectrum features in all historical feature vectors of the category.

[0100] It should be noted that a sensitive projection model is pre-learned for each known state category (such as outer ring fatigue and inner ring spalling). The design goal is mainly to preserve the vibration spectrum characteristics (such as target characteristic frequency and sidebands) directly related to the physical mechanism of the fault category, while suppressing irrelevant components. For example, the sensitive projection model mainly preserves the narrowband energy, harmonic components, and modulation sideband characteristics centered on the target characteristic frequency, while projecting the frequency band energy corresponding to broadband background noise and other fault characteristics to near zero.

[0101] Based on this, multiple narrow-band frequency windows of the sensitive projection model are used to extract features for each narrow-band frequency window from the target feature vector. The features include spectral peak value, total window energy, and spectral kurtosis.

[0102] The peak-to-peak value refers to the maximum value among all spectral amplitudes within the frequency window. The spectral kurtosis is obtained by calculating the fourth-order statistical moments of all spectral amplitude sequences within the window. The total energy of the window refers to the sum of the squares of the amplitudes at all frequency points within the window.

[0103] The formula for calculating spectral kurtosis is as follows:

[0104]

[0105] in, For spectral kurtosis, The amplitude at the k-th frequency point within the narrowband frequency window. The mean of the amplitude within the narrowband frequency window is given, and M is the number of frequency points within the narrowband frequency window. Subtracting 3 is to make the spectral kurtosis of the Gaussian signal zero. The numerator represents the fourth central moment of the calculated amplitude (reflecting the sharpness of the distribution). The denominator represents the square of the amplitude variance.

[0106] In this context, a narrowband frequency window is defined with each selected target characteristic frequency, harmonic, and sideband as the center (e.g., target characteristic frequency ± 5 spectral resolutions, harmonic ± 5 spectral resolutions, sideband ± 5 spectral resolutions).

[0107] It should be noted that the peak value of the spectrum directly reflects the strongest impact intensity caused by the fault. For example, pitting faults on the outer ring of a bearing will generate periodic impacts at its characteristic frequency, manifested as a significant increase in amplitude at that frequency. The total energy of the window characterizes the overall vibration energy level of that frequency band. This characteristic changes sensitively when the fault causes energy to spread or intensify in a specific frequency band (such as broadband energy increase caused by poor lubrication). The spectral kurtosis feature is extremely sensitive to transient impact components in the signal, effectively capturing early, weak fault impact characteristics, and is a key indicator for identifying fault impact modes. The higher the spectral kurtosis value, the more significant the impact component within that narrow frequency window.

[0108] Based on this, the peak value reflects the impact intensity, the energy reflects the overall activity level, and the spectral kurtosis reflects the transient characteristics of the impact. Combining these three factors can distinguish different types of failure modes (such as a single strong impact, broadband energy diffusion, weak early impact, etc.) and avoid misjudgments that may be caused by a single feature.

[0109] Finally, the features of all narrowband frequency windows are concatenated to obtain the projected target feature vector.

[0110] It should be noted that the splicing can be based on an order of frequency from low to high. This aligns with the natural order of the vibration spectrum, ensuring that each dimension of the feature vector has a stable and interpretable physical correspondence.

[0111] Based on this, the projected target feature vector has been fully focused on the physical frequency bands related to the fault state category.

[0112] Step 2: Using the feature saliency weight vector learned for the category, calculate the weighted Euclidean distance between the projected target feature vector and the pre-projected category cluster centers.

[0113] It should be noted that the feature saliency weight vector is used to reflect the importance of vibration feature dimensions in identifying specific fault modes. A feature saliency weight vector is learned for each category, and each element of the feature saliency weight vector corresponds to the weight of one dimension in the low-dimensional sensitive subspace. Weight learning is based on two criteria: first, intra-class compactness: the more clustered (less dispersed) all historical feature vectors within a single dimension are within the same category, the higher the weight. Second, inter-class separability: the greater the difference between the class center and other class centers for a single dimension, the higher the weight.

[0114] Weighted Euclidean distance calculation refers to weighting the projected target feature vector and the category cluster center according to the feature significance weight vector in each dimension, and then calculating the weighted Euclidean distance.

[0115] It should be noted that the cluster centers of the categories have been pre-calculated and stored.

[0116] Step 3: Based on the weighted Euclidean distance between the projected target feature vector and the projected cluster centers of each category, obtain the metric distance from the target feature vector to the cluster centers of each category.

[0117] One approach is to directly use the weighted Euclidean distance as the distance metric. Alternatively, one can normalize the weighted Euclidean distance by utilizing the data distribution statistics of all historical feature vectors of each category in the low-dimensional sensitive subspace to the cluster centers, thus obtaining the distance metric from the target feature vector to the cluster centers of each category.

[0118] It should be noted that data distribution statistics are used to reflect the degree of dispersion of all historical feature vectors of a category within a low-dimensional sensitive subspace. Examples include the mean and standard deviation.

[0119] The steps to normalize the weighted Euclidean distance include dividing the weighted Euclidean distance by the data distribution statistic.

[0120] Therefore, distance metrics are made statistically meaningful and can adapt to the dispersion of different categories of data. For categories with widely distributed data, larger absolute distances are permissible. For categories with tightly distributed data, even small absolute distances may be considered significant deviations.

[0121] For example, suppose the data on poor lubrication is widely distributed (large standard deviation) in the low-dimensional sensitive subspace, while the data on inner-circle peeling is concentrated (small standard deviation). For the same sample, the weighted Euclidean distance to both (the cluster centers of poor lubrication and inner-circle peeling) is calculated to be 5.0. After normalization, the distance to the poor lubrication category might become 1.2 (within the normal fluctuation range), while the distance to the inner-circle peeling category might become 4.5 (significantly exceeding the normal range). This more accurately reflects the greater difference from the latter.

[0122] In summary, by introducing sensitive projection and feature saliency weighting, the distance calculation is transformed from global averaging to local focusing, enhancing the ability to distinguish specific fault modes. Furthermore, adaptive distribution normalization ensures statistical comparability of the distance metric, improving the stability and reliability of state discrimination under different operating conditions and fault modes. These three elements work together to define a refined, adaptive, and physically interpretable bearing state measurement system.

[0123] In one example, in data-driven intelligent diagnosis of bearing faults, bearing monitoring data, after feature extraction, often resides in a high-dimensional space. This high-dimensional space contains a large number of redundant features unrelated to specific fault modes (such as broadband background noise and vibration responses of other components) as well as highly correlated feature dimensions. Directly performing distance calculations or pattern matching in this space would result in a heavy computational burden, and fault-sensitive features would easily be overwhelmed, thereby reducing the accuracy and real-time performance of the diagnosis.

[0124] Based on this, a feature space compression and focusing method oriented towards physical mechanisms is proposed. The core of this method lies in adaptively learning a dedicated sensitive projection model for each known bearing state category (such as outer ring fatigue or inner ring spalling). This method automatically constructs a low-dimensional sensitive subspace by combining physical prior guidance with data-driven learning. This subspace retains only the feature components most directly and relevant to the physical essence of the current state category, thereby achieving purification and focusing of high-dimensional feature vectors.

[0125] Based on this, the process of learning various types of sensitive projection models includes the following steps:

[0126] Step 1: Perform aggregate analysis on the vibration spectrum features in all historical feature vectors of the category to identify frequency points that appear stably in the category and whose energy is significantly higher than the background noise, as candidate feature frequency sets; the significance condition is that the mean amplitude of the frequency point is greater than K times the preset noise baseline threshold, where K is a constant greater than 1; the stability condition is that the standard deviation of the amplitude of the frequency point is less than the preset absolute fluctuation threshold (e.g., 0.2).

[0127] It should be noted that aggregate analysis refers to calculating the mean and standard deviation of the amplitude at each frequency point.

[0128] Step 2: Based on the preset bearing physical model corresponding to the category, select the target feature frequency that is directly related to the current state category mode in terms of physical mechanism from the candidate feature frequency set, and determine the harmonics and sidebands of the target feature frequency.

[0129] Harmonics refer to integer multiples (2 times, 3 times, etc.) of the target characteristic frequency, generated due to the nonlinear characteristics of the impulse signal. Sidebands refer to the frequency components generated around the characteristic frequency due to the modulation effect. This value is usually the target characteristic frequency ± the revolving frequency.

[0130] It should be noted that stable sidebands generated by fault modulation effects typically do not exist in a healthy state. Therefore, the set of sidebands can be defined as an empty set.

[0131] In this example, the calculation formula for the physical model of outer ring bearing failure may include the following:

[0132]

[0133] The physical model calculation formula for inner ring bearing failure is as follows:

[0134]

[0135] The calculation formula for the physical model of rolling element bearing failure is as follows:

[0136]

[0137] The physical model calculation formula for cage bearing failure is as follows:

[0138]

[0139] in, This represents the theoretical characteristic frequency of an outer ring fault. This represents the theoretical characteristic frequency of an inner-circle fault. The theoretical characteristic frequency of rolling element failure. To determine the theoretical characteristic frequency of cage failure, The number of balls, For frequency conversion, D is the ball diameter, and D is the pitch diameter. The contact angle is given. Rotational frequency is obtained by dividing the rotational speed by 60. Pitch diameter refers to the diameter of the circle defined by the centers of the balls in the bearing; that is, the balls in the bearing are arranged in a ring, and the diameter of this circle formed by connecting the center points of each ball is the pitch diameter. The contact angle is the angle between the normal direction of the contact point between the ball and the raceway and the radial plane of the bearing.

[0140] It should be noted that the number of balls, pitch diameter, and contact angle mentioned above can all be obtained from a bearing parameter database.

[0141] The screening process may include: for each fault condition category, calculating the theoretical characteristic frequency corresponding to the condition category based on the bearing model parameters (such as the number of balls, pitch diameter, and contact angle mentioned above) and the average speed in the operating range of the category.

[0142] For the health condition category, the vibration spectrum does not contain impact characteristics caused by local defects, but it does contain stable frequency components generated by normal equipment operation. Therefore, the theoretical characteristic frequency of the health condition category is the rotational frequency.

[0143] Then, the frequencies in the candidate feature frequency set are matched with the theoretical feature frequencies. The frequencies that successfully match are determined as the target feature frequencies.

[0144] It should be noted that the difference between the theoretical feature frequency and the frequency in the candidate feature frequency set is calculated, and the ratio of the difference to the theoretical feature frequency is obtained. If the ratio is less than a preset ratio threshold (e.g., 2%), the match is considered successful. Otherwise, the match fails.

[0145] For example, taking the outer ring fatigue failure category of an industrial motor drive end bearing as an example, assuming the bearing parameters are known and the rotational frequency is 10Hz.

[0146] The vibration spectrum (0-2000Hz) of 100 historical outer-ring fatigue eigenvectors was analyzed. The mean and mean difference of 100 amplitude values ​​were calculated for each frequency point. Statistical analysis revealed stable and significant (exceeding 3 standard deviations above the noise baseline and with standard deviations less than the fluctuation threshold) frequencies at 87.5Hz, 175Hz, 262.5Hz, 77.5Hz, and 97.5Hz. These frequencies were included in the candidate feature frequency set.

[0147] Based on the physical model of the outer ring bearing failure, the theoretical characteristic frequency of the outer ring was calculated to be 87.5Hz. Comparison with the candidate characteristic frequency set showed that 87.5Hz precisely matched the outer ring characteristic frequency, confirming it as the target characteristic frequency. Based on this target characteristic frequency, the second harmonic (175Hz), third harmonic (262.5Hz), and 87.5Hz ± 10Hz were derived and can be used as modulation sidebands. These also coincide with the candidate characteristic frequency set.

[0148] Centered at 87.5Hz, a ±5Hz window (82.5-92.5Hz) is defined. For this window, the following are extracted from the spectrum of each historical feature vector: 1) peak-to-peak value; 2) total energy within the window; and 3) spectral kurtosis within the window.

[0149] In addition, the same operation was repeated for 175Hz, 262.5Hz, 77.5Hz, and 97.5Hz.

[0150] Furthermore, for the sideband pair (77.5Hz, 97.5Hz) window, the energy ratio between the two windows can be extracted as a modulation intensity feature. That is, the ratio between the total energy of the window corresponding to 77.5Hz and the total energy of the window corresponding to 97.5Hz.

[0151] Step 3: Determine the corresponding narrowband frequency window in the frequency domain, centered on the target characteristic frequency, harmonics, and sidebands.

[0152] In one example, common, stable spectral features are extracted from group data of similar faults through statistical aggregation analysis.

[0153] Based on this, an aggregate analysis of the vibration spectrum features in all historical feature vectors of the category is performed, including the following steps:

[0154] Step 1: Extract the vibration spectrum features of each historical feature vector to obtain the spectrum data of each historical feature vector; the spectrum data of each historical feature vector includes multiple frequency points and corresponding amplitudes.

[0155] The spectral data of each historical feature vector is essentially a series of pairs of (frequency points, amplitudes). For example, a spectrum covering 0-1000Hz with a resolution of 1Hz will include 1000 frequency points (0Hz, 1Hz, 2Hz, ..., 999Hz) and their corresponding amplitudes (such as the spectral density or amplitude of vibrational acceleration).

[0156] Based on this, we focus on the frequency domain representation that is directly related to the fault frequency, providing input data in a unified format for subsequent statistical analysis.

[0157] Step 2: For all historical feature vectors under the category, at the same frequency point, calculate the mean amplitude and standard deviation of the amplitude of all spectral data at the frequency point.

[0158] The mean reflects the average intensity of vibrational energy at that frequency point for this type of fault. A higher mean indicates that the frequency component is more significant in the population. The standard deviation reflects the dispersion of the amplitudes of different historical eigenvectors at that frequency point. A smaller standard deviation indicates that the amplitudes of all historical eigenvectors at that frequency point are closer and more stable. A larger standard deviation indicates that the amplitude fluctuates greatly at that frequency point, which may be due to random interference or a weak correlation with the fault.

[0159] Based on this, the significance (by the mean) and repeatability / stability (by the standard deviation) of each potential frequency component were quantified using the two key statistics: mean and standard deviation.

[0160] Step 3: Select the frequency points whose mean amplitude is greater than K times the preset noise baseline threshold and whose standard deviation of amplitude is less than the preset absolute fluctuation threshold as significant peak frequencies.

[0161] For each fault state category, a baseline value is determined by analyzing the energy of frequency bands far from suspected peaks in the background vibration spectrum or fault spectrum under healthy conditions. This baseline value represents the system's noise floor level, i.e., the noise baseline threshold. K times is a safety factor greater than 1 (e.g., K=3) used to ensure that the selected frequency components are sufficiently prominent, far exceeding the background noise.

[0162] It should be noted that the noise baseline threshold for each condition category is an adaptively determined or preset empirical value based on specific equipment, sensors, installation conditions, and historical health data. It should also be noted that the noise baseline thresholds for different condition categories can be the same or different, and the noise baseline thresholds for the same condition category can be the same or different across different operating conditions.

[0163] For example, the amplitudes of a large number of frequency points (e.g., frequency bands after removing obvious peaks) in the initial vibration spectrum of the bearing under known healthy conditions are selected, and the higher percentile of the amplitudes at all frequency points (e.g., 95%, 98%, or 99%) is taken as the noise baseline threshold. This represents the upper limit of normal fluctuations in healthy background noise. At this point, during bearing operation, the amplitudes of the characteristic frequencies of the healthy state may be higher than the noise baseline threshold, but usually will not exhibit the drastic, several-fold higher peaks than the background noise seen in a fault state. Therefore, the K for the healthy state category is relatively small, for example, 1.2. The K for the outer race fault state category is 3.

[0164] The absolute fluctuation threshold refers to a preset upper limit for amplitude fluctuation (e.g., a proportion of the median standard deviation of the overall amplitude of the category). This ensures that the frequency component appears consistently across all similar samples. If the standard deviation is large, it means that some historical eigenvectors have peaks at this point, while others have none or are very weak, thus it cannot be used as a reliable indicator of the pattern. This condition excludes frequency components that occur accidentally, intermittently, or are affected by other random factors.

[0165] Step 4: The set of all identified significant peak frequencies is used as the candidate feature frequency set.

[0166] In one example, after obtaining a low-dimensional sensitive subspace through a sensitive projection model, the contributions of each dimension within this subspace to fault identification are not equal. Some dimensions may be highly specialized, clearly distinguishing this type of fault from other states. Other dimensions may include more intra-class variations or general information shared with other classes. Treating all dimensions equally in distance calculation weakens the discriminative power of key features and introduces noise.

[0167] Therefore, a dimension-weighted adaptive learning method that integrates statistical properties and physical priors is proposed, which can generate a feature saliency weight vector for each state category. This vector is obtained by comprehensively quantifying and evaluating the intra-class compactness, inter-class separability, and physical correlation of each dimension of the low-dimensional sensitive subspace, thereby automatically highlighting key discriminative features and suppressing secondary or interfering features in the subsequent weighted distance calculation.

[0168] Based on this, the process of learning the feature saliency weight vector of a category includes the following steps:

[0169] Step 1: Calculate the intra-class scatter of all projected historical feature vectors of each category in each dimension, and the inter-class separability of the projected cluster centers of each category with respect to the projected cluster centers of other categories in each dimension.

[0170] In this example, intra-class dispersion measures the degree of dispersion of all projected historical feature vectors of the same class along a given dimension. For a specific fault class, the values ​​of all projected historical feature vectors along each dimension are treated as a sample set, and the standard deviation or variance is calculated as a measure of intra-class dispersion. The smaller the dispersion, the more concentrated the samples of that class are along that dimension, and the more stable that dimension is.

[0171] Inter-class separability measures the ability of a feature dimension to distinguish the current class from other classes. It is calculated by taking the absolute difference in a dimension between the projected cluster centers of the current class and the projected cluster centers of every other known class, and then averaging or maximizing these absolute differences over all other classes. A higher value indicates a stronger ability of that dimension to distinguish between different classes.

[0172] It should be noted that each feature in each window is an independent dimension. For example, in the projection history feature vector, the peak value of the target feature frequency window is one dimension, while the peak value of the harmonic window is another dimension. That is, each feature value in the projection history feature vector represents a dimension.

[0173] Step 2: For each dimension, use the ratio between intra-class dispersion and inter-class separability as the initial significance score for that dimension.

[0174] It should be noted that the initial significance score reflects the relative magnitude of inter-category differences and intra-category fluctuations. The larger the ratio, the more discriminative that dimension is.

[0175] Step 3: Determine the dimensions related to the target characteristic frequency and harmonics as the target dimensions.

[0176] That is, the peak-to-peak value, total energy, and spectral kurtosis of the target characteristic frequency window, as well as the peak-to-peak value, total energy, and spectral kurtosis of the harmonic window, are used as the target dimensions.

[0177] Step 4: According to the preset gain coefficient mapping relationship, increase the initial saliency score of the target dimension and decrease the initial saliency score of the non-target dimension to obtain the saliency score of the target dimension and the saliency score of the non-target dimension.

[0178] In this example, a gain coefficient mapping table can be pre-set based on the physical knowledge of bearing faults. The fundamental frequency and harmonics are the most direct and stable manifestations of the fault; therefore, their corresponding feature dimensions are the most critical. They should be given special attention in pattern recognition, allowing fault identification to focus more on features with clearly defined physical mechanisms.

[0179] The gain coefficients between target dimensions can be the same or different. The gain coefficients between non-target dimensions can also be the same or different.

[0180] For example, for a clear pitting fault on the outer ring of a bearing, the vibration signal will produce stable impact peaks at the fundamental frequency and the second and third harmonics. From the macroscopic goal of enhancing fault characteristics and suppressing noise, giving equal weight to the characteristics (peak value, energy) of these core frequency bands is an effective and simple strategy.

[0181] When the fault mechanism is complex, or when different features exhibit significant differences in their specificity, stability, or sensitivity to the fault, the fundamental frequency component of the fault is usually the most direct and stable manifestation of the fault, and is less susceptible to interference. Higher harmonics, on the other hand, have weaker energy and may be more easily masked by background noise, or may not be apparent in certain fault stages (such as the early stages). Therefore, a higher gain coefficient is set for the relevant dimensions of the fundamental frequency window. Furthermore, for impulse faults (such as stripping), spectral kurtosis is extremely sensitive to transient impulses and is an excellent indicator of early faults, potentially more important than a simple peak amplitude. Therefore, the gain coefficient for spectral kurtosis characteristics within the fundamental frequency window is set higher than that for the peak amplitude.

[0182] When identifying compound or multi-site faults, the frequency importance varies depending on the fault source. For example, when both inner and outer ring faults exist simultaneously, it may be necessary to distinguish between primary and secondary faults by using the different gain coefficients between the target characteristic frequency and harmonics.

[0183] Improving the initial significance score of the target dimension can be achieved by multiplying by a gain coefficient greater than 1, while reducing the initial significance score of non-target dimensions can be achieved by multiplying by a gain coefficient less than 1.

[0184] Step 5: Normalize the saliency scores of the target dimension and the non-target dimension to make the modulus 1, thus obtaining the feature saliency weight vector for this category of learning.

[0185] In summary, adaptive learning is achieved by automatically adjusting weights based on the data distribution of different state categories. Combining prior knowledge of fault mechanisms enhances key physical feature dimensions, thus achieving physical interpretability. Simultaneously, intra-class compactness and inter-class separability are considered to balance overfitting and discriminative power. Important dimensions are highlighted in distance calculation to suppress noise and improve fault identification accuracy. It is applicable to complex industrial scenarios with multiple operating conditions and fault modes, exhibiting strong generalization. Based on this, the weighted Euclidean distance calculation ensures that the dimension with the strongest discriminative power for the current fault category is highlighted, thereby improving the accuracy and robustness of fault detection.

[0186] For example, we are learning a feature saliency weight vector for an outer raceway pitting fault state category. To simplify the explanation, we select three dimensions of data as an example to demonstrate the weight learning process. Dimension 1: Peak-to-peak value of the spectrum corresponding to the target feature frequency. Dimension 2 represents the peak-to-peak value of the second harmonic, and Dimension 3 represents the peak-to-peak value of the sideband. Two other known state categories are known: inner raceway fault and poor lubrication fault.

[0187] The five projected historical feature vectors of the outer ring fault category are shown in Table 1.

[0188] Table 1 Projection History Feature Vectors

[0189]

[0190] Based on this, the mean and standard deviation of each dimension are calculated. The mean of dimension 1 is 2.10 and the standard deviation is 0.0946; the mean of dimension 2 is 1.26 and the standard deviation is 0.0667; and the mean of dimension 3 is 0.328 and the standard deviation is 0.0383. Therefore, the within-class scatter vector is [0.0946, 0.0667, 0.0383].

[0191] The projected cluster centers for each category are known, as shown in Table 2.

[0192] Table 2 Cluster Centers for Each Category

[0193]

[0194] Based on this, for each dimension, the absolute difference between the outer fault category center and other category centers is calculated, and then the average of the absolute differences is taken.

[0195] For Dimension 1: The difference between Dimension 1 and inner ring faults is 1.58, and the difference between Dimension 1 and poor lubrication faults is 0.45, with an average of 1.615. For Dimension 2: The difference between Dimension 2 and inner ring faults is 0.36, and the difference between Dimension 2 and poor lubrication faults is 0.38, with an average of 0.37. For Dimension 3: The difference between Dimension 3 and inner ring faults is 0.084, and the difference between Dimension 3 and poor lubrication faults is 0.176, with an average of 0.13. Therefore, the within-class scatter vector is [1.615, 0.37, 0.13].

[0196] Based on this, the initial significance score for dimension 1 is 17.07, the initial significance score for dimension 2 is 5.55, and the initial significance score for dimension 3 is 3.39. Therefore, the initial significance score vector is [17.07, 5.55, 3.39].

[0197] Based on the physical knowledge of bearing failure, a preset gain coefficient mapping relationship table is shown in Table 3.

[0198] Table 3 Gain Coefficient Mapping Relationship Table

[0199]

[0200] Given the physical characteristics of the outer ring fault categories, after applying a gain coefficient of 1.8 to dimension 1, the significance score is 30.73; after applying a gain coefficient of 1.5 to dimension 2, the significance score is 8.33; and after applying a gain coefficient of 0.6 to dimension 3, the significance score is 2.03. Based on this, the post-gain score vector is [30.73, 8.33, 2.03].

[0201] The magnitude of the score vector after gain calculation is 31.9. The weights are calculated by normalization based on the ratio between the significance score and the magnitude: 0.963 for dimension 1, 0.261 for dimension 2, and 0.064 for dimension 3. Therefore, the feature significance weight vector of the outer ring fault is [0.963, 0.261, 0.064].

[0202] Based on this, dimension 1 has the highest weight, dimension 2 has a medium weight, and dimension 3 has the lowest weight. Thus, the relevant dimensions that reflect the target feature frequency of outer ring faults dominate the distance calculation, suppressing noise.

[0203] In one example, the process of performing cross-condition clustering distribution association analysis on the target feature vector includes the following steps:

[0204] Step 1: Based on the cluster distribution of each known state category in the second historical benchmark dataset, calculate the metric distance from the target feature vector to the cluster center of each category.

[0205] It should be noted that the method for calculating the distance metric is based on the clustering distribution of each known state category in the first historical benchmark dataset, and the process of calculating the distance metric from the target feature vector to the cluster center of each category is used.

[0206] Step 2: Map the metric distance from the target feature vector to the cluster center of each category to obtain the second similarity between the target feature vector and each known state category under adjacent working conditions.

[0207] The second similarity is calculated by inputting the metric distance into a preset monotonically decreasing function; the monotonically decreasing function is configured such that the output value decreases as the input metric distance increases. Alternatively, a matching based on a preset similarity lookup table is performed to obtain the corresponding second similarity.

[0208] In summary, directly using distance metrics for minimum distance classification is susceptible to extreme values ​​or noise. If the data for a particular fault category is particularly sparsely distributed (with generally large distances), its distance value may always be greater than that of more compactly distributed categories, causing the system to be insensitive to that fault.

[0209] Using similarity helps to balance the system's sensitivity to different state categories, improving the detection rate of various faults (especially those with large characteristic variations) and reducing the risk of missed detections. It also provides a standardized, dimensionless confidence index for diagnostic decision-making.

[0210] Setting an intuitive threshold (e.g., similarity > 0.85 is considered a fault) greatly simplifies system parameter tuning and application.

[0211] In summary, when the target operating condition lacks sufficient historical fault samples, abundant historical benchmark data from adjacent operating conditions can be used for state matching and diagnosis. This significantly reduces the dependence on a full-condition coverage dataset and improves the system's availability under data-sparse conditions. By calculating the second similarity with various known states under adjacent operating conditions, it is possible to effectively identify and associate unknown or unrecorded state patterns under the current operating condition, thereby improving the ability to detect and identify novel faults or unknown operating states.

[0212] Based on this, the problems of operating condition dependence and data sparsity in bearing fault diagnosis are effectively solved, significantly improving the cross-operating condition adaptability, identification reliability and engineering practical value of the diagnostic system. It is particularly suitable for industrial equipment with frequent changes in operating conditions, such as wind turbines, CNC machine tools and compressors, and can significantly improve the practicality and reliability of bearing fault detection systems.

[0213] In one example, typically when monitoring the same equipment, the sampling frequency of the data acquisition system is fixed regardless of changes in rotational speed. However, because bearing vibration characteristics vary with operating parameters such as rotational speed and load, the same fault may appear at different locations in the spectrum under different operating conditions. Therefore, cross-operating condition feature alignment can be used to match the unknown state under the current operating condition with the fault modes in adjacent operating condition intervals, thereby improving accuracy.

[0214] Based on this, before calculating the metric distance from the target feature vector to the cluster centers of each category, the following steps are included:

[0215] Step 1: Extract the dominant difference parameter ratio between the operating condition interval and the adjacent operating condition interval.

[0216] In this example, the dominant difference parameter refers to the operating condition parameter that has the most significant impact on vibration characteristics, typically rotational speed or load. The parameter ratio is used to represent the relative change of the dominant difference parameter between adjacent operating condition intervals; for example, the parameter ratio is the ratio between the parameters of adjacent operating condition intervals and the parameters of the operating condition interval.

[0217] Different operating condition intervals correspond to different dominant difference parameters. The dominant difference parameters matching the operating condition intervals with adjacent operating condition intervals are retrieved from the dominant difference parameter mapping table.

[0218] Step 2: Based on the parameter ratio relationship and the pre-built feature parameter alignment mapping relationship, construct the feature parameter alignment mapping function for vibration spectrum features.

[0219] In this example, different characteristic parameters have different mapping relationships. For instance, the bearing frequency changes linearly with the rotational speed, and the vibration energy is related to the square of the rotational speed. Since load changes also affect the vibration amplitude, the amplitude characteristic is calculated by multiplying the square of the rotational speed by the load ratio.

[0220] For example, frequency is directly proportional to rotational speed, a physical law determined by bearing geometry and kinematics. Vibration amplitude is related to both rotational speed and load. Generally, vibration energy is proportional to the square of the rotational speed, and is also affected by load variations.

[0221] For example, the frequency characteristics are: The amplitude characteristics are: ,in, For the frequency characteristics after alignment, The frequency characteristics of the operating range, This represents the speed ratio between adjacent operating ranges and operating ranges. For the amplitude characteristics after alignment, The amplitude characteristics of the operating condition range, This represents the speed ratio between adjacent operating ranges and operating ranges. This represents the load ratio between adjacent operating condition intervals and between operating condition intervals.

[0222] Based on this, the corresponding parameter ratio relationship is substituted into the pre-constructed mapping relationship to obtain the corresponding feature parameter alignment mapping function.

[0223] Step 3: Perform a forward transformation on the vibration spectrum features in the target feature vector according to the feature parameter alignment mapping function to obtain the target feature vector aligned with the adjacent working condition interval.

[0224] In this example, vibration spectrum features are extracted from the target feature vector, and then a mapping is applied feature by feature. For example, each frequency feature is multiplied by the rotational speed ratio. Each amplitude feature is multiplied by the square of the rotational speed ratio and then by the load ratio.

[0225] In summary, by constructing a feature parameter alignment mapping function based on physical mechanisms, the monitoring data under the current operating condition is effectively mapped to the feature space of adjacent operating conditions. This solves the problem that the same mode exhibits different behaviors under different operating conditions such as speed and load, and achieves stable and reliable fault identification across operating conditions.

[0226] It can automatically identify differences in dominant operating conditions and perform feature alignment accordingly, enabling the diagnostic system to adapt to natural changes in equipment operating conditions and enhancing its overall robustness and practicality in industrial environments with varying operating conditions.

[0227] For example, the equipment is an industrial motor bearing. The current operating condition is low speed operation, 900 RPM, with a load of 30%. The adjacent operating condition is normal speed operation, 1800 RPM, with a load of 60%.

[0228] The dominant difference parameter ratio is the ratio of the current speed to the target speed, which is 0.5. The frequency characteristic parameter alignment mapping function is... 2. The amplitude characteristic parameter alignment mapping function is: .

[0229] In one example, the vibration spectrum features in the target feature vector are positively transformed according to the feature parameter alignment mapping function, including the following steps:

[0230] Step 1: Align the mapping function with the frequency feature parameters and perform frequency axis scaling transformation on the target feature vector.

[0231] Based on this, the frequency feature values ​​of the target feature vector are substituted into the frequency feature parameter alignment mapping function. For example, if the speed is higher in adjacent operating conditions, the spectral feature will be stretched towards higher frequencies, and vice versa.

[0232] Step 2: Map and align the target feature vector after scaling the frequency axis to the frequency grid corresponding to the adjacent operating condition interval.

[0233] The target feature vector, after frequency scaling, is resampled onto the frequency grid used in the adjacent operating condition range through interpolation methods (such as linear interpolation or spline interpolation).

[0234] It should be noted that since the scaled frequency points may not fall on the fixed, discrete frequency grid of adjacent operating conditions, the estimated values ​​on the grid points are obtained through calculation (interpolation).

[0235] For example, the linear interpolation formula can be as follows:

[0236] y

[0237] in, , The frequency values ​​of two adjacent points are known. Less than , The frequency value at the target frequency point. , for The amplitude of the point, for Amplitude of a point y represents the amplitude at the target frequency point.

[0238] For example, we might need to know the amplitude at a frequency point (e.g., 200Hz). The scaled frequency grid might not contain a point exactly at 200Hz (e.g., only 193.75Hz and 206.25Hz). In this case, we need to use an interpolation algorithm (such as linear interpolation) to estimate the amplitude at 200Hz using the amplitudes at 193.75Hz and 206.25Hz.

[0239] Step 3: Perform amplitude scaling transformation on the mapped and aligned target feature vector according to the amplitude feature parameter alignment function.

[0240] Based on this, the magnitude of the target feature vector is substituted into the magnitude feature parameter alignment mapping function.

[0241] Step 4: Use the target feature vector after amplitude scaling transformation as the target feature vector aligned with the adjacent operating condition interval.

[0242] After the aforementioned physical transformations of frequency and amplitude, the target feature vector has been transformed into a virtual feature vector that appears to have been acquired under adjacent operating conditions. This new target feature vector is situated in a physical context that is as similar as possible to the historical feature vectors in adjacent operating condition intervals. This significantly improves the operating condition adaptability of the diagnostic system.

[0243] In summary, scaling transformation based on physical mechanisms further maps vibration characteristics under different working conditions to a unified space, solving the problem of feature distribution shift caused by changes in working conditions.

[0244] In one example, fault mode identification of the target feature vector includes the following steps:

[0245] For the second similarity with each known state category in adjacent working condition intervals, if the highest second similarity is higher than the preset similarity threshold, the target state category corresponding to the highest second similarity is determined as the mode state represented by the target feature vector.

[0246] For example, the preset similarity threshold is 0.85. A similarity value higher than the preset threshold indicates that the current target feature vector matches a known state category sufficiently well to make a deterministic diagnosis.

[0247] If the highest second similarity is lower than or equal to the preset similarity threshold, the target feature vector is determined to be a pattern state that cannot be associated with the known state.

[0248] Among these features, marking the target feature vector as an unknown state and issuing an abnormal state warning can trigger manual inspection or further analysis.

[0249] It should be noted that the preset similarity threshold is a configurable empirical parameter, typically determined through historical data analysis, experimental verification, or the experience of domain experts. For example, the threshold is higher for safety-critical equipment, while the threshold for general monitoring equipment can be appropriately lowered.

[0250] In this example, a confidence level can be set when the similarity exceeds a preset threshold. For instance, the higher the difference between the second similarity and the similarity threshold, the higher the corresponding confidence level. Different confidence levels provide different evaluation suggestions. For example, at a medium confidence level, it is recommended to increase the monitoring frequency and schedule an inspection within one week.

[0251] In one example, after completing the initial cross-condition similarity calculation, a physical trend verification mechanism is introduced to perform secondary confirmation and refined scoring of the similarity results, thereby significantly improving the accuracy, reliability and physical interpretability of fault mode recognition.

[0252] Based on this, fault mode identification of the target feature vector may also include the following steps:

[0253] Step 1: Filter target state categories with a second similarity higher than the preset similarity threshold.

[0254] From all known state categories (such as healthy, outer ring fault, inner ring fault, poor lubrication, etc.), candidate categories that may be related to the current unknown state are initially screened. A preset similarity threshold is used to filter out obviously irrelevant states, reducing subsequent computation and improving efficiency.

[0255] Step 2: Retrieve the sensitivity factors for the target state category from the sensitivity factor mapping table.

[0256] The sensitivity factor mapping table is used to reflect frequencies that characterize the nature of faults and are sensitive to changes in operating conditions. The sensitivity factor mapping table is obtained through a sensitivity projection model. The sensitivity factors for the target state category include at least one of the following: target characteristic frequency, harmonics, sidebands, peak-to-peak value of the corresponding narrowband frequency window, total window energy, and spectral kurtosis.

[0257] Different state categories have different target feature frequencies, harmonics, and sidebands, and a sensitivity factor mapping table is constructed.

[0258] For example, outer ring fault: amplitude at the characteristic frequency of the outer ring target or narrowband energy near that frequency; inner ring fault: amplitude at the characteristic frequency of the inner ring target or narrowband energy of the modulation sideband.

[0259] Step 3: Obtain the first feature value of the sensitivity factor in the target feature vector under the current operating condition range.

[0260] Specifically, the values ​​of the aforementioned sensitive factors under the current operating conditions are extracted from the target feature vector to be diagnosed.

[0261] Step 4: Extract the second feature value corresponding to the sensitivity factor from the historical feature vector of the cluster center of the target state category.

[0262] Step 5: Calculate the relative change and direction of change between the second feature value and the first feature value.

[0263] In this example, the formula for calculating the relative change is as follows:

[0264]

[0265] in, It is a relative change. The second eigenvalue, It is the first eigenvalue. When the value is greater than 0, the direction of change is positive. When the value is less than 0, the direction of change is negative.

[0266] Step 6: Compare the stated direction of change with the expected direction of change of the sensitive factor with the operating conditions, and compare the stated relative change with the expected change of the sensitive factor with the operating conditions to obtain the degree of conformity of the trend of the sensitive factor with the operating conditions.

[0267] In this example, the expected direction of change refers to how the sensitivity factor should theoretically change based on the fault physics model and changes in operating parameters.

[0268] The expected change refers to how the sensitive factor should theoretically change based on the physical model of the fault and changes in operating parameters.

[0269] In the change mapping table, the parameters of the working condition interval are matched with the parameters of the adjacent working condition intervals to obtain the expected change direction and the expected change amount.

[0270] Based on this, the ratio between the relative change and the expected change is used to obtain the trend conformity value. When the direction of change is consistent with the expected direction of change, the direction conformity is 1; when the direction of change is inconsistent with the expected direction of change, the direction conformity is -1. The trend conformity degree is obtained by multiplying the trend conformity value and the direction conformity degree.

[0271] Step 7: Perform a weighted summation of the second similarity and the trend conformity to obtain the comprehensive score corresponding to the target state category.

[0272] For example, for the outer ring fault category, the second similarity weight is 0.6, and the trend conformity is 0.4.

[0273] It should be noted that if there are trend conformity scores for multiple sensitive factors, the trend conformity scores of these multiple sensitive factors are weighted and summed to obtain the total trend conformity score. The weights can be obtained based on a pre-set sensitive factor weight mapping table.

[0274] The second similarity score and the overall trend conformity score are weighted and summed to obtain the comprehensive score corresponding to the target state category.

[0275] Step 8: If the highest comprehensive score is greater than the preset scoring threshold, the target state category corresponding to the highest comprehensive score is determined as the fault mode state represented by the target feature vector.

[0276] For example, the preset rating threshold is 0.75.

[0277] If the highest comprehensive score is lower than or equal to the preset scoring threshold, the target feature vector is determined to be a pattern state that cannot be associated with the known state.

[0278] Based on the same idea, some embodiments of this application also provide devices and non-volatile computer storage media corresponding to the above methods.

[0279] Figure 2 A schematic diagram of an adaptive bearing fault detection device provided in this application embodiment includes:

[0280] At least one processor; and,

[0281] A memory communicatively connected to the at least one processor; wherein,

[0282] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform any of the aforementioned adaptive bearing fault detection methods.

[0283] Some embodiments of this application provide an adaptive bearing fault detection non-volatile computer storage medium storing computer-executable instructions capable of executing any of the above-described adaptive bearing fault detection methods.

[0284] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0285] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0286] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0287] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the technical principles of this application should fall within the protection scope of this application.

Claims

1. An adaptive bearing fault detection method, characterized in that, The method includes: Acquire monitoring data of target bearings within the same operating condition range; the time interval between acquisitions of target bearing monitoring data is less than a preset interval threshold. Vibration spectrum features are extracted from the monitoring data of the target bearing to obtain the target feature vector; In the benchmark operating condition interval knowledge base, the first historical benchmark dataset of the operating condition interval is retrieved; the first historical benchmark dataset stores historical feature vectors that have been clustered by state category. Based on the clustering distribution of each known state category in the first historical benchmark dataset, the metric distance from the target feature vector to the cluster center of each category is calculated to obtain the first similarity between the target feature vector and each known state category; When the first similarity between the target feature vector and each known state category is lower than a preset similarity threshold, the target feature vector is determined to represent a new unknown state that is not recorded in the first historical benchmark dataset. In the knowledge base of the benchmark working condition interval, retrieve the second historical benchmark dataset of the adjacent working condition intervals of the working condition interval; the adjacent working condition interval refers to the working condition interval with the smallest Euclidean distance to the center point of the working condition interval in the working condition parameter space. Based on the clustering distribution of each known state category in the second historical benchmark dataset, cross-working condition clustering distribution correlation analysis is performed on the target feature vector to obtain the second similarity between the target feature vector and each known state category in adjacent working condition intervals. Based on the second similarity with each known state category in adjacent operating condition intervals, pattern recognition is performed on the target feature vector to determine whether the bearing has a fault in the operating condition interval. Based on the clustering distribution of each known state category in the first historical benchmark dataset, the metric distance from the target feature vector to the cluster center of each category is calculated, specifically including: For each known state category in the first historical benchmark dataset, the target feature vector is projected onto a low-dimensional sensitive subspace defined by the physical pattern most relevant to the current state category using a sensitive projection model learned for the category; the sensitive projection model is learned based on the vibration spectrum features in all historical feature vectors of the category. Using the feature saliency weight vector learned for the category, the weighted Euclidean distance between the projected target feature vector and the pre-projected category cluster centers is calculated; the feature saliency weight vector is used to reflect the importance of the vibration feature dimension to the identification of a specific pattern; The metric distance from the target feature vector to each category cluster center is obtained based on the weighted Euclidean distance between the projected target feature vector and the projected cluster centers of each category. Using a sensitive projection model for category learning, the target feature vector is projected onto a low-dimensional sensitive subspace defined by the physical pattern most relevant to the current state category. Specifically, this includes: Using multiple narrowband frequency windows of the sensitive projection model, features of each narrowband frequency window are extracted from the target feature vector; the features include spectral peak value, total window energy, and spectral kurtosis. The features of all narrowband frequency windows are concatenated to obtain the projected target feature vector; The learning process of the sensitive projection model specifically includes: Aggregate analysis is performed on the vibration spectrum features in all historical feature vectors of the category to identify frequency points that appear stably in the category and whose energy is significantly higher than the background noise, as candidate feature frequency sets; the significance condition is that the mean amplitude of the frequency point is greater than K times the preset noise baseline threshold, where K is a constant greater than 1; the stability condition is that the standard deviation of the amplitude of the frequency point is less than the preset absolute fluctuation threshold. Based on the preset bearing physical model corresponding to the category, target feature frequencies that are directly related to the current state category mode in terms of physical mechanism are selected from the candidate feature frequency set, and the harmonics and sidebands of the target feature frequencies are determined. Using the target characteristic frequency, harmonics, and sidebands as centers, a corresponding narrowband frequency window is determined in the frequency domain; The method further includes: Calculate the intra-class scatter of all projected historical feature vectors of each category in each dimension, and the inter-class separability of the projected cluster centers of each category with respect to the projected cluster centers of other categories in each dimension; For each dimension, the ratio between intra-class dispersion and inter-class separability is used as the initial significance score for that dimension. The dimensions related to the target characteristic frequency and harmonics are defined as the target dimensions. According to the preset gain coefficient mapping relationship, the initial saliency score of the target dimension is increased, and the initial saliency score of the non-target dimension is decreased, so as to obtain the saliency score of the target dimension and the saliency score of the non-target dimension. The saliency scores of the target dimension and the non-target dimension are normalized to make the magnitude length 1, thus obtaining the feature saliency weight vector for this category of learning.

2. The method according to claim 1, characterized in that, The method further includes: When the first similarity between the target feature vector and each known state category is at least not lower than the similarity threshold; The known state category with the highest similarity is determined as the pattern state represented by the target feature vector.

3. The method according to claim 1, characterized in that, The aggregation analysis of vibration spectrum features in all historical feature vectors of the category specifically includes: The vibration spectrum features of each historical feature vector are extracted to obtain the spectrum data of each historical feature vector; the spectrum data of each historical feature vector includes multiple frequency points and the amplitude corresponding to each frequency point; For all historical feature vectors under the aforementioned category, at the same frequency point, calculate the mean amplitude and standard deviation of the amplitude of all spectral data at the frequency point; The frequency points whose mean amplitude is greater than K times the preset noise baseline threshold and whose standard deviation of amplitude is less than the preset absolute fluctuation threshold are defined as significant peak frequencies. The set of all identified significant peak frequencies is used as the candidate feature frequency set.

4. The method according to claim 1, characterized in that, Based on the clustering distribution of each known state category in the second historical benchmark dataset, a cross-condition clustering distribution correlation analysis is performed on the target feature vector, specifically including: Based on the clustering distribution of each known state category in the second historical benchmark dataset, the metric distance from the target feature vector to the cluster center of each category is calculated; The target feature vector is mapped to the metric distance of each category cluster center to obtain the second similarity between the target feature vector and each known state category under adjacent working condition intervals.

5. The method according to claim 4, characterized in that, Mapping the target feature vector to the cluster centers of each category yields a second similarity between the target feature vector and each known state category in adjacent working condition intervals, specifically including: The metric distance is input into a preset monotonically decreasing function to calculate the corresponding second similarity; wherein, the monotonically decreasing function is configured such that the output value decreases as the input metric distance increases; or, Matching is performed based on a preset similarity lookup table to obtain the corresponding second similarity; wherein, the similarity lookup table defines the mapping relationship between different metric distance ranges and the second similarity.

6. The method according to claim 1, characterized in that, Based on the second similarity with each known state category in adjacent operating condition intervals, pattern recognition is performed on the target feature vector, specifically including: For the second similarity with each known state category in adjacent working condition intervals, if the highest second similarity is higher than a preset similarity threshold, the state category corresponding to the highest second similarity is determined as the mode state represented by the target feature vector; If the highest second similarity is lower than or equal to the preset similarity threshold, the target feature vector is determined to be a pattern state that cannot be associated with the known state.

7. An adaptive bearing fault detection device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform an adaptive bearing fault detection method according to any one of claims 1-6.

Images