A non-stationary gearbox real-time fault diagnosis method and system for transition working conditions

Abstract

This invention provides a real-time fault diagnosis method and system for non-stationary gearboxes under transient operating conditions, comprising: acquiring multi-channel monitoring signals under stable operating conditions to construct an offline labeled sample set; constructing a diagnostic model including a feature extraction module and a fault classification module, introducing an operating condition discrimination module for operating condition robustness training, and retaining the feature extraction module and fault classification module after training; mapping offline samples to a feature space and normalizing them, and calculating the geometric center of various faults as feature prototypes; receiving unlabeled data blocks online, and using the initialized model for feature extraction and fault probability prediction; calculating the geometric similarity between online samples and each prototype, and constructing a probability distribution as a pseudo-supervision signal; constructing a loss function using the pseudo-supervision signal, and updating the model online using an asymmetric learning rate, with the learning rate of the feature extraction module being less than that of the classification module; periodically reprojecting offline samples using the updated feature extraction module to update fault prototypes to maintain reference validity.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and in particular to a real-time fault diagnosis method and system for non-stationary gearboxes under transient operating conditions. Background Technology

[0002] Gearboxes, as typical key components of rotating machinery, are widely used in wind power equipment, rail transportation, metallurgical production lines, and various industrial transmission systems. Their operating status directly affects the safety, stability, and economy of the entire system. During long-term operation, gearboxes are susceptible to load impacts, changes in operating conditions, and environmental factors, leading to various failure modes such as tooth surface wear, pitting, and tooth breakage. To effectively monitor the operating status of gearboxes, industrial sites typically deploy vibration sensors, current sensors, and other multi-source monitoring signals during equipment operation, and then analyze and identify the acquired data.

[0003] With the continuous development of sensing technology, computing power, and data acquisition methods, data-driven fault diagnosis methods based on machine learning and deep learning are gradually being applied to the field of gearbox health monitoring and intelligent operation and maintenance. These methods typically extract features from historical monitoring data and train models to automatically identify different fault types, reducing reliance on human experience to some extent and improving the efficiency and accuracy of fault diagnosis.

[0004] In existing technologies, mainstream data-driven fault diagnosis methods typically employ the following approach: first, a large amount of labeled monitoring data is collected under several typical stable operating conditions; then, a deep learning model is trained based on this offline data; and finally, the trained model is deployed in actual industrial settings for online fault diagnosis. This approach implicitly relies on two fundamental assumptions: first, the training and testing data follow the same or similar distributions, i.e., satisfying the independent and identically distributed (ICD) assumption; and second, sufficient and high-quality labeled samples can be obtained during the offline training phase. However, in real-world industrial applications, these assumptions often fail to hold, leading to the following three problems with existing methods in engineering practice:

[0005] 1) It relies heavily on a large number of high-quality labeled samples. In industrial settings, obtaining fault samples is costly and time-consuming, and it is often difficult to obtain true labels during the online operation phase, which limits the diagnostic performance of existing methods under conditions of scarce labeled samples;

[0006] 2) Difficulty in adapting to the non-stationarity of data distribution caused by transitional operating conditions. Under transitional operating conditions such as start-stop, acceleration / deceleration, or load changes, the characteristics of monitoring signals change continuously over time, making it difficult for offline trained models to maintain stable diagnostic performance;

[0007] 3) The online adaptive process is susceptible to error accumulation due to erroneous predictions. Some methods directly use model predictions as pseudo-labels for updates, which can easily lead to error propagation and reduce the overall reliability of the system under conditions of drastic changes in operating conditions or class imbalance.

[0008] In summary, existing gearbox fault diagnosis technologies have significant shortcomings in terms of reliance on labeled samples, adaptability to transitional operating conditions, and stability of online updates. There is an urgent need for a new method that can achieve stable and reliable real-time fault diagnosis in complex industrial scenarios with limited labeled samples and continuously changing operating conditions. Summary of the Invention

[0009] To address the aforementioned problems in existing technologies, the purpose of this invention is to provide a real-time fault diagnosis method for non-stationary gearboxes under transitional operating conditions. This method overcomes the dependence of existing gearbox fault diagnosis methods on a large number of high-quality labeled samples, as well as their difficulty in adapting to continuous changes in operating conditions and non-stationary data distribution during actual operation.

[0010] The present invention also provides a real-time fault diagnosis system for non-stationary gearboxes oriented towards transitional operating conditions.

[0011] To address the above problems, this application proposes a real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions, comprising the following steps:

[0012] S1. Gearbox detection data acquisition and offline sample construction: Under several stable operating conditions of the gearbox, multi-channel monitoring signals are collected by sensors placed at preset positions on the equipment; the continuous time series monitoring signals are segmented using a fixed-length sliding window method to convert the original signals into discrete samples; and an offline labeled sample set containing fault category labels and operating condition identifiers is constructed.

[0013] S2. Offline diagnostic model initialization and robust training under operating conditions: Based on the offline labeled sample set, a diagnostic model including a feature extraction module and a fault classification module is constructed. During the offline training phase, the diagnostic model introduces an operating condition discrimination module. By minimizing the fault classification loss and maximizing the operating condition discrimination loss, sensitive information about operating conditions in the feature representation is filtered out, while sensitive information about faults is retained. After training, the operating condition discrimination module is discarded, and the feature extraction module and the fault classification module are retained as the initialization model for online diagnosis.

[0014] S3. Static memory and fault prototype construction: Based on the fault category labels of the offline labeled sample set, the samples are divided into different fault category sets; the feature extraction module is used to map each type of sample to the feature space and perform normalization processing; the geometric center of each type of fault in the feature space is calculated as the feature prototype of the corresponding fault category; all fault feature prototypes together constitute the prototype set.

[0015] S4. Online monitoring data input and feature prediction: During the actual operation of the gearbox, unlabeled online samples are continuously received. The initialization model is used to extract features and predict the probability of fault categories of the online samples to form basic diagnostic output.

[0016] S5. Geometric pseudo-supervision based on fault prototypes: Normalize the features of online samples, calculate the geometric similarity between them and each fault feature prototype, and construct a geometric probability distribution reflecting the degree of alignment between the sample and different fault types based on the geometric similarity, which serves as the source of pseudo-supervision signal for online stage model updates.

[0017] S6. Optimization of the online diagnostic model with asymmetric adaptive update: Using geometric probability distribution as pseudo-supervision signal, a loss function is constructed for the online diagnostic stage, and the model parameters are updated online. Different learning rates are used for the feature extraction module and the fault classification module, with the learning rate of the feature extraction module being smaller than that of the fault classification module, so that the model can adapt to changes in transitional working conditions.

[0018] S7. Periodic reprojection update of fault prototypes: During online operation, the updated feature extraction module is used to remap offline labeled samples and periodically update the feature prototypes of each fault category.

[0019] Furthermore, in step S1, the sensor includes a vibration sensor, a speed sensor, a torque sensor, and / or a current sensor; the multi-channel monitoring signal includes a vibration signal, a speed signal, a torque signal, and / or a current signal.

[0020] Furthermore, in step S2, the diagnostic model includes a feature extraction module, a fault classification module, and a working condition discrimination module. During the offline training phase, the parameters of the three modules are jointly iteratively updated with the optimization objective of minimizing the fault classification cross-entropy loss and maximizing the working condition discrimination cross-entropy loss with preset trade-off parameters. The gradient of the working condition discrimination loss is applied to the feature extraction module in the form of a negative value. After training is completed, the working condition discrimination module is removed, while the feature extraction module and the fault classification module are retained.

[0021] Furthermore, the fault feature prototype mentioned in step S3 is calculated as follows: For all offline labeled samples belonging to the k-th type of fault, where k represents the fault category index, the feature extraction module is used to map them to the feature space. The feature vector of each sample is normalized using the L2 norm, and then the arithmetic mean of all normalized feature vectors is calculated as the geometric center of the fault in the feature space.

[0022] Furthermore, in step S5, the geometric similarity is calculated in the following manner:

[0023] The feature vectors of online samples mapped by the feature extraction module are normalized by the L2 norm, and each fault prototype is normalized by the L2 norm. Then, the inner product between the normalized online sample feature vectors and the normalized fault prototypes is calculated to obtain the geometric similarity.

[0024] The geometric probability distribution is constructed by scaling the geometric similarity of each fault category with a preset temperature parameter, and then converting the scaled geometric similarity into a geometric probability distribution through a softmax function. This geometric probability distribution reflects the alignment between online samples and each fault prototype in the feature space.

[0025] Furthermore, in step S6, the loss function for the online diagnosis phase is constructed in the following way:

[0026] For each sample in the online data block, the geometric probability distribution constructed in step S5 is used as the target distribution, and the fault category prediction probability distribution obtained in step S4 is used as the prediction distribution. The cross-entropy loss between the two distributions is calculated, and then the average of the cross-entropy loss of all samples in the online data block is taken as the online loss function of the current data block.

[0027] Furthermore, in step S6, the asymmetric adaptive update specifically includes:

[0028] After each online data block arrives, multiple gradient descent updates are performed on the feature extraction module and the fault classification module respectively. The feature extraction module is updated with a smaller learning rate, while the fault classification module is updated with a larger learning rate, so that the fault classification module can respond to changes in operating conditions.

[0029] Furthermore, in step S7, during the periodic reprojection update, the prototype of the k-th type of fault is updated in the following way: the feature extraction module after the current update is used to remap the features of all offline labeled samples belonging to the k-th type of fault, the L2 norm normalization is performed on each remapped feature vector, and then the arithmetic mean of all normalized feature vectors is calculated as the updated prototype of the k-th type of fault.

[0030] The present invention also provides a real-time fault diagnosis system for non-stationary gearboxes under transient operating conditions, for implementing the method, comprising:

[0031] The data acquisition module is used to acquire multi-channel monitoring signals under several stable operating conditions of the gearbox, and to convert continuous time-series signals into discrete samples through a sliding window method to construct an offline labeled sample set.

[0032] The offline training module is used to construct a diagnostic model containing a feature extraction module and a fault classification module based on the offline labeled sample set. The working condition discrimination module is introduced to carry out working condition robust training. After training, the feature extraction module and the fault classification module are retained.

[0033] The prototype construction module is used to map the samples to the feature space and perform normalization processing based on the fault categories of the offline labeled sample set, calculate the geometric center of each type of fault in the feature space as the feature prototype, and form a prototype set.

[0034] The online diagnostic module is used to continuously receive unlabeled online monitoring data blocks during actual operation, and to perform feature extraction and fault category probability prediction on online samples using an initialization model;

[0035] The pseudo-supervision construction module is used to calculate the geometric similarity between online sample features and each fault prototype, and construct a geometric probability distribution that reflects the degree of alignment between the sample and different fault types as a pseudo-supervision signal;

[0036] The asymmetric update module is used to construct an online loss function using geometric probability distribution as a pseudo-supervisory signal, and to update the feature extraction module and the fault classification module online using different learning rates;

[0037] The prototype reprojection module is used to update the mapping of offline labeled samples using the updated feature extraction module, and periodically update the feature prototypes of each fault category.

[0038] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0039] 1. This invention constructs a diagnostic model with robustness to operating conditions during the offline training phase and introduces an adaptive update mechanism during the online phase. This effectively adapts to the non-stationary monitoring data distribution generated by the gearbox under transitional operating conditions such as start-stop, acceleration-deceleration, and load changes, and significantly reduces the impact of transitional operating conditions on diagnostic performance.

[0040] 2. In the online diagnosis process, this invention utilizes geometric pseudo-supervision signals based on fault prototypes to update the model. A probability distribution is constructed by calculating the geometric similarity between online sample features and each fault prototype, and this distribution serves as the pseudo-supervision signal to drive model adaptation. The method provided by this invention does not require real fault labels for online samples, nor does it require obtaining explicit operating condition number information. It only relies on the geometric relationship between the fault prototypes constructed offline and the online samples to achieve model updates, reducing the dependence on online labeled data and improving the feasibility of the method described in this application in actual industrial scenarios.

[0041] 3. This invention updates the feature extraction module and the fault classification module using different learning rates. This asymmetric adaptive update method enables the fault classification module to quickly respond to changes in operating conditions and adjust the classification decision boundary in a timely manner. At the same time, it limits the update amplitude of the feature extraction module, maintains the overall structural stability of the feature space, and prevents the loss of discriminative information due to drastic changes in the feature space. By separating the update intensity of feature representation learning and classification decision adjustment, it effectively suppresses the pollution of the feature space by erroneous pseudo-labels.

[0042] 4. After a preset number of update steps, the present invention uses the updated feature extraction module to re-extract features from the offline labeled samples, and re-projects and updates the prototypes of various fault categories accordingly, so that the fault prototypes can continuously reflect the structure of the current feature space, ensuring the accuracy and reliability of the pseudo-supervision signal. Attached Figure Description

[0043] Figure 1 This is a graph showing the real-time diagnostic performance changes of tooth surface pitting faults during the online data stream process under test conditions.

[0044] Figure 2 This is a graph showing the real-time diagnostic performance changes of gear wear faults during the online data stream process under test conditions.

[0045] Figure 3 This is a graph showing the real-time diagnostic performance changes of a broken tooth fault during the online data stream process under test conditions.

[0046] Figure 4 This is a flowchart of the fault diagnosis method described in this invention. Detailed Implementation

[0047] This invention provides a real-time fault diagnosis method for gearboxes under transient operating conditions, applicable to non-stationary monitoring data scenarios generated during gearbox operation, including start-stop, acceleration, deceleration, and load adjustment. The method initializes the diagnostic model during offline training using a small amount of labeled monitoring data collected from the gearbox under different stable operating conditions, and constructs normalized prototypes of various gearbox faults in the feature space as static reference anchors. During the online phase, when new unlabeled monitoring data is generated during gearbox operation, a geometric probability distribution is constructed based on the geometric similarity between online sample features and various fault prototypes. This distribution serves as a pseudo-supervisory signal to drive the model for online adaptive updates. By periodically reprojecting and updating the offline-constructed fault prototypes, the prototypes continuously reflect the current gearbox feature space structure, ultimately achieving reliable and real-time fault diagnosis of the gearbox's non-stationary operating data stream. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] like Figure 4 As shown, this application provides a real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions, the method comprising the following steps:

[0050] Step S1: Gearbox monitoring data acquisition and offline sample construction. Under several stable speed and load conditions, multi-channel vibration monitoring signals are acquired using vibration sensors located on the gearbox housing, motor drive end, or intermediate shaft. Optionally, speed, torque, or current signals can also be acquired simultaneously. The acquired continuous time series monitoring signals are segmented using a fixed-length sliding window method, thereby converting the original signals into discrete samples.

[0051] Construct an offline labeled sample set

[0052]

[0053] in, This represents the feature vector formed by splicing multi-channel monitoring signals from the gearbox. This represents the total number of offline samples. This indicates the corresponding fault category of the gearbox (including but not limited to gear wear, tooth surface pitting, broken teeth, and health status). Indicates the stable operating condition number to which the sample belongs. For the feature vector dimension, This represents the total number of fault categories. To stabilize the number of operating conditions.

[0054] Step S2: Initialization and Robust Training of the Offline Gearbox Diagnostic Model. Based on the offline labeled sample set, a gearbox fault diagnosis model is constructed, including a feature extraction module for extracting gearbox fault discrimination features. And a fault classification module for outputting the probability of gearbox fault categories. To reduce the differences in characteristic distribution of the gearbox under different speeds and load conditions, a working condition discrimination module is introduced during offline training. This is to constrain the feature extraction module's insensitivity to changes in operating conditions. , , These are the learnable parameters for the feature extraction module, fault classification module, and operating condition discrimination module, respectively.

[0055] The training objective function during the offline training phase is defined as follows:

[0056]

[0057] in, Let cross-entropy be the loss function. Parameters that balance the fault detection capability of the gearbox with the invariance of operating conditions;

[0058] Represents the mathematical expectation; This indicates random sampling from an offline set of labeled samples.

[0059] After training, the working condition discrimination module is discarded, while the feature extraction module and fault classification module are retained as the initialization model for online gearbox diagnosis.

[0060] Step S3: Gearbox Static Memory and Fault Prototype Construction. Based on the gearbox fault category labels of the offline labeled sample set, the samples are divided into different fault category sets, and a static memory is constructed from them. For each type of gearbox fault sample, the feature extraction module is used to map it to the feature space, and the feature vector is normalized to reduce the impact of gearbox vibration amplitude changes with speed and load.

[0061] Based on this, the first The geometric center of the normalized feature vector corresponding to the gearbox-like fault sample is obtained and used as the feature prototype of this fault category. The calculation method is as follows:

[0062]

[0063] in, Indicates belonging to the first A set of labeled offline samples of gearbox faults; all gearbox fault prototypes together constitute a prototype set, which is used to constrain the geometry of the gearbox feature space in the online phase.

[0064] Step S4: Gearbox Online Monitoring Data Input and Feature Prediction. During the actual operation of the gearbox, online monitoring data arrives at the diagnostic system sequentially in the form of data blocks. Let the first... The data block that arrives at time 1 is

[0065]

[0066] All online samples were unlabeled, and no corresponding gearbox fault labels or explicit operating condition information could be obtained. For online runtime indexing, For the first The number of samples in the online data block arriving at each time point; This is the normalized exponential function, used to convert the output into a probability distribution.

[0067] Any online sample First, its feature representation is obtained through the feature extraction module. Then, the fault classification module outputs the predicted probability distribution of its belonging to each gearbox fault category;

[0068]

[0069] Step S5: Constructing geometric pseudo-supervision based on gearbox fault prototypes. Normalize the features of online samples and calculate their geometric similarity to each gearbox fault prototype.

[0070]

[0071] A geometric probability distribution is constructed based on the above similarity.

[0072]

[0073] in For temperature parameters, Represents the dot product of vectors; The exponential function is represented. The geometric probability distribution reflects the alignment of the online gearbox monitoring samples with respect to each fault prototype in the feature space and serves as a source of pseudo-monitoring signals for the online phase.

[0074] Step S6: Optimize the asymmetric adaptive update gearbox online diagnostic model, using the geometric probability distribution as a pseudo-supervision signal to construct the loss function for the gearbox online diagnostic stage.

[0075]

[0076] The model parameters are then updated online based on this loss function. To improve the model's ability to track changes in feature distribution caused by gearbox transient conditions, multiple inner loop updates are performed after each data block arrives. Let the online phase be... The model parameters at time t are Then in the first After each data block arrives, the feature extraction module and the fault classification module are updated as follows:

[0077]

[0078] log represents the natural logarithm; Update the index for the number of times the inner loop is updated; and These are the learning rates for the feature extraction module and the fault classification module, respectively. Indicates about parameters The gradient; where the gradient is set. This allows the gearbox fault classification module to respond quickly to changes in operating conditions, while limiting the update range of the feature extraction module, thereby maintaining the overall stability of the gearbox feature space.

[0079] Step S7: Periodic reprojection update of gearbox fault prototypes. To ensure the gearbox fault prototypes remain relevant during the online phase, after a preset number of update steps, the updated feature extraction module is used to remap the feature space of the offline labeled samples, and the prototypes for each gearbox fault category are reprojected and updated accordingly. During the next update, the first Gearbox-like fault prototype updated to

[0080]

[0081] To verify the real-time fault diagnosis performance of the method described in this invention under transitional gearbox conditions, the experiment used measured data collected from a two-stage parallel shaft gearbox test bench. The test bench mainly consists of a 2.2 kW three-phase asynchronous motor, a two-stage parallel gearbox, a magnetic powder brake (for applying torque), a measurement and control system, and speed and torque sensors. During gearbox operation, the speed and load can be precisely adjusted to simulate stable and transitional operating conditions that may occur in actual industrial scenarios. During data acquisition, triaxial accelerometers were placed at the motor drive end and the intermediate shaft of the gearbox to collect multi-channel vibration signals during gearbox operation. The sampling frequency of the vibration signals was set to 12.8 kHz to ensure effective capture of gear meshing impacts and early fault characteristics. The experiment focused on the 36-tooth gear on the intermediate shaft of the gearbox and its adjacent support bearings, with corresponding fault types including healthy conditions, gear wear, tooth surface pitting, and broken teeth—typical gearbox fault forms. Related faults were manually introduced through precision machining to ensure the controllability of fault types and severity. In the data preprocessing stage, a sliding window segmentation strategy is used to process the original multi-channel vibration signals. The window length is set to 1024 sampling points, and the sliding step size is 16 sampling points, thereby constructing highly overlapping time-series samples. Each sample is formed by splicing multiple vibration channel data and is used as the input feature of the model.

[0082] The experiment is mainly used to evaluate the online fault diagnosis performance of the gearbox under transient operating conditions caused by load changes.

[0083] The data for the offline training phase were all collected under the condition that the gearbox was running stably at a speed of 1000 rpm. The offline training dataset was constructed by setting different constant load torques. Specifically, during the offline training phase, gearbox vibration monitoring data were collected under three stable load conditions with torques of 10 N·m, 15 N·m, and 20 N·m. The number of samples collected under each condition was 6344, used to simulate the gearbox operation scenario under multiple loads but stable conditions. Only a small number of samples from the above offline data were selected as labeled data for offline model initialization; the remaining samples were not used in offline training.

[0084] Data during the online phase was also collected at a speed of 1000 rpm, but the load conditions varied over time to simulate the load adjustment process experienced by the gearbox during actual operation. The online data stream initially consisted of stable operating data with a torque of 20 N·m; it then transitioned to a load variation phase, gradually increasing from 20 N·m to 15 N·m, with the monitoring data corresponding to this phase defined as transition operating data; finally, the online data stream entered the operating phase where the torque stabilized at 15 N·m. No actual fault labels were provided for any samples during the online phase, and they were continuously input into the diagnostic model in chronological order.

[0085] The following is combined Figures 1 to 3 A detailed analysis of the experimental results of the present invention.

[0086] like Figure 1 The figure shows the real-time diagnostic performance changes of tooth surface pitting faults during the online data stream process under test conditions. In the initial stage, when the rotational speed is maintained at 1000 rpm and the load is stable, all methods maintain high diagnostic accuracy, indicating that the offline initialization model has good discriminative ability under stable conditions. When the load transitions from 20 N·m to 15 N·m, the amplitude and spectral structure of the monitoring signal change significantly, and the diagnostic performance of the baseline method drops rapidly and is difficult to recover in subsequent stages. The naive online self-training strategy shows a brief performance improvement in the early stage of the transition, but because it relies on the model's own prediction results for updates, it generates significant error accumulation during the load change stage, causing the diagnostic results to continuously fail in the subsequent stable conditions. In contrast, although the proposed method exhibits some fluctuations when the transition occurs, it can quickly complete adaptive adjustments and rapidly recover to a diagnostic accuracy close to the initial level after the load stabilizes again, demonstrating its good robustness to tooth surface pitting faults under load change conditions.

[0087] like Figure 2The figure shows the real-time diagnostic performance variation curves of gear wear faults during the online data stream process under test conditions. Compared with tooth surface pitting, gear wear faults typically exhibit more gradual but wider-ranging variations. During the load stabilization phase, the overall performance of each method is relatively similar. When the load enters the transition range from 20 N·m to 15 N·m, the diagnostic accuracy of the baseline method declines significantly, and the recovery speed is slow. Although the naive online update strategy can partially alleviate the performance degradation during the transition condition, its diagnostic curve exhibits significant jitter, reflecting its insufficient stability when facing continuous changes in feature distribution. The proposed method maintains a relatively stable diagnostic trend throughout the transition phase and quickly recovers to a high accuracy after the load stabilizes, indicating that this method can effectively balance adaptive capability and diagnostic stability when dealing with fault types such as gear wear with relatively continuous feature changes.

[0088] like Figure 3 The figure shows the real-time diagnostic performance variation curves of tooth breakage faults during online data flow under test conditions. Tooth breakage faults are usually accompanied by significant shock components, and their characteristics are highly sensitive to load changes, thus making them more prone to significant distribution drift under transitional conditions. As can be observed from the figure, during the load change phase, the diagnostic performance of the baseline method remains at a low level, almost unable to effectively identify the tooth breakage state. The naive online self-training strategy exhibits significant performance fluctuations in the early stages of the transitional condition and fails to fully recover even in the subsequent stable condition, indicating its susceptibility to the long-term effects of erroneous pseudo-labels under strongly non-stationary conditions. In contrast, although the proposed method is also subject to some shocks in the early stages of load changes, the overall diagnostic trend is smoother, and it can gradually recover stable performance in the subsequent stages, demonstrating that the method still possesses strong online adaptability and robustness when facing highly sensitive faults such as tooth breakage.

[0089] Example 2

[0090] This invention also provides a real-time fault diagnosis system for non-stationary gearboxes under transient operating conditions, comprising:

[0091] The data acquisition module is used to acquire multi-channel monitoring signals under several stable operating conditions of the gearbox, and to convert continuous time-series signals into discrete samples through a sliding window method to construct an offline labeled sample set.

[0092] The offline training module is used to construct a diagnostic model containing a feature extraction module and a fault classification module based on the offline labeled sample set. The working condition discrimination module is introduced to carry out working condition robust training. After training, the feature extraction module and the fault classification module are retained.

[0093] The prototype construction module is used to map the samples to the feature space and perform normalization processing based on the fault categories of the offline labeled sample set, calculate the geometric center of each type of fault in the feature space as the feature prototype, and form a prototype set.

[0094] The online diagnostic module is used to continuously receive unlabeled online monitoring data blocks during actual operation, and to perform feature extraction and fault category probability prediction on online samples using an initialization model;

[0095] The pseudo-supervision construction module is used to calculate the geometric similarity between online sample features and each fault prototype, and construct a geometric probability distribution that reflects the degree of alignment between the sample and different fault types as a pseudo-supervision signal;

[0096] An asymmetric update module is used to construct an online loss function using the geometric probability distribution as a pseudo-supervisory signal, and to update the feature extraction module and the fault classification module online using different learning rates;

[0097] The prototype reprojection module is used to update the mapping of offline labeled samples using the updated feature extraction module, and periodically update the feature prototypes of each fault category.

[0098] The data acquisition module can be implemented by vibration sensors arranged in the gearbox housing, motor drive end, or intermediate shaft position; the offline training module, prototype construction module, online diagnosis module, pseudo-supervised construction module, asymmetric update module, and prototype reprojection module can be implemented by computer programs and run on industrial computers or edge computing devices.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 process, method, article, or apparatus.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions, characterized in that, Includes the following steps: S1. Gearbox detection data acquisition and offline sample construction: Under several stable operating conditions of the gearbox, multi-channel monitoring signals are collected by sensors placed at preset positions on the equipment; the continuous time series monitoring signals are segmented using a fixed-length sliding window method to convert the original signals into discrete samples; and an offline labeled sample set containing fault category labels and operating condition identifiers is constructed. S2. Offline diagnostic model initialization and robust training under working conditions: Based on the offline labeled sample set, a diagnostic model including a feature extraction module and a fault classification module is constructed; The diagnostic model introduces a working condition discrimination module during the offline training stage, which filters out working condition sensitive information in the feature representation and retains fault sensitive information by minimizing the fault classification loss and maximizing the working condition discrimination loss. After training, the working condition discrimination module is discarded, and the feature extraction module and fault classification module are retained as the initialization model for online diagnosis. S3. Static memory and fault prototype construction: Based on the fault category labels of the offline labeled sample set, the samples are divided into different fault category sets; the feature extraction module is used to map each type of sample to the feature space and perform normalization processing; the geometric center of each type of fault in the feature space is calculated as the feature prototype of the corresponding fault category; all fault feature prototypes together constitute the prototype set. S4. Online monitoring data input and feature prediction: During the actual operation of the gearbox, unlabeled online samples are continuously received. The initialization model is used to extract features and predict the probability of fault categories of the online samples to form basic diagnostic output. S5. Geometric pseudo-supervision based on fault prototypes: Normalize the features of online samples, calculate the geometric similarity between them and each fault feature prototype, and construct a geometric probability distribution reflecting the degree of alignment between the sample and different fault types based on the geometric similarity, which serves as the source of pseudo-supervision signal for online stage model updates. S6. Optimization of online diagnostic model with asymmetric adaptive update: Using geometric probability distribution as pseudo-supervision signal, construct the loss function in the online diagnostic stage and update the model parameters online. Different learning rates are used for the feature extraction module and the fault classification module. The learning rate of the feature extraction module is smaller than that of the fault classification module, so that the model can adapt to changes in transitional operating conditions. S7. Periodic reprojection update of fault prototypes: During online operation, the updated feature extraction module is used to remap offline labeled samples and periodically update the feature prototypes of each fault category.

2. The real-time fault diagnosis method for non-stationary gearboxes oriented towards transient operating conditions according to claim 1, characterized in that, In step S1, the sensors include a vibration sensor, a speed sensor, a torque sensor, and / or a current sensor; the multi-channel monitoring signals include vibration signals, speed signals, torque signals, and / or current signals.

3. The real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions according to claim 1, characterized in that, In step S2, the diagnostic model includes a feature extraction module, a fault classification module, and a working condition discrimination module. During the offline training phase, the parameters of the three modules are jointly iteratively updated with the optimization objectives of minimizing the cross-entropy loss of fault classification and maximizing the cross-entropy loss of working condition discrimination with preset trade-off parameters. The gradient of the working condition discrimination loss is applied to the feature extraction module in the form of a negative value. After training is completed, the working condition discrimination module is removed, while the feature extraction module and the fault classification module are retained.

4. The real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions according to claim 1, characterized in that, The fault feature prototype described in step S3 is calculated as follows: For all offline labeled samples belonging to the k-th type of fault, where k represents the fault category index, the feature extraction module is used to map them to the feature space. The feature vector of each sample is normalized using the L2 norm, and the arithmetic mean of all normalized feature vectors is calculated as the geometric center of the fault in the feature space.

5. The real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions according to claim 1, characterized in that, In step S5, the geometric similarity is calculated in the following way: The feature vectors of online samples mapped by the feature extraction module are normalized by the L2 norm, and each fault prototype is normalized by the L2 norm. Then, the inner product between the normalized online sample feature vectors and the normalized fault prototypes is calculated to obtain the geometric similarity. The geometric probability distribution is constructed by scaling the geometric similarity of each fault category with a preset temperature parameter, and then converting the scaled geometric similarity into a geometric probability distribution through a softmax function. This geometric probability distribution reflects the alignment between online samples and each fault prototype in the feature space.

6. The real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions according to claim 1, characterized in that, In step S6, the loss function for the online diagnosis phase is constructed in the following way: For each sample in the online data block, the geometric probability distribution constructed in step S5 is used as the target distribution, and the fault category prediction probability distribution obtained in step S4 is used as the prediction distribution. The cross-entropy loss between the two distributions is calculated, and the average of the cross-entropy loss of all samples in the online data block is taken as the online loss function of the current data block.

7. The real-time fault diagnosis method for non-stationary gearboxes under transient operating conditions according to claim 1, characterized in that, In step S6, the asymmetric adaptive update specifically includes: After each online data block arrives, multiple gradient descent updates are performed on the feature extraction module and the fault classification module respectively. The feature extraction module is updated with a smaller learning rate, while the fault classification module is updated with a larger learning rate, so that the fault classification module can respond to changes in operating conditions.

8. The real-time fault diagnosis method for a non-stationary gearbox oriented towards transient operating conditions according to claim 1, characterized in that, In step S7, during the periodic reprojection update, the prototype of the k-th type of fault is updated in the following way: the feature extraction module after the current update is used to remap the features of all offline labeled samples belonging to the k-th type of fault, the L2 norm normalization is performed on each remapped feature vector, and then the arithmetic mean of all normalized feature vectors is calculated as the updated prototype of the k-th type of fault.

9. A real-time fault diagnosis system for non-stationary gearboxes under transient operating conditions, used to implement the method as described in claim 1, characterized in that, include: The data acquisition module is used to acquire multi-channel monitoring signals under several stable operating conditions of the gearbox, and to convert continuous time-series signals into discrete samples through a sliding window method to construct an offline labeled sample set. The offline training module is used to construct a diagnostic model containing a feature extraction module and a fault classification module based on the offline labeled sample set. The working condition discrimination module is introduced to carry out working condition robust training. After training, the feature extraction module and the fault classification module are retained. The prototype construction module is used to map the samples to the feature space and perform normalization processing based on the fault categories of the offline labeled sample set, calculate the geometric center of each type of fault in the feature space as the feature prototype, and form a prototype set. The online diagnostic module is used to continuously receive unlabeled online monitoring data blocks during actual operation, and to perform feature extraction and fault category probability prediction on online samples using an initialization model; The pseudo-supervision construction module is used to calculate the geometric similarity between online sample features and each fault prototype, and construct a geometric probability distribution that reflects the degree of alignment between the sample and different fault types as a pseudo-supervision signal; The asymmetric update module is used to construct an online loss function using geometric probability distribution as a pseudo-supervisory signal, and to update the feature extraction module and the fault classification module online using different learning rates; The prototype reprojection module is used to update the mapping of offline labeled samples using the updated feature extraction module, and periodically update the feature prototypes of each fault category.

Images