Equipment fault diagnosis method based on BO and intelligent model

Abstract

The invention belongs to the technical field of mechanical equipment fault diagnosis, and particularly relates to an equipment fault diagnosis method based on BO and an intelligent model, and the method comprises the following steps: S1, analyzing application scene data; s2, intelligent model selection of anomaly detection and fault identification; s3, optimization parameters are determined; s4, data set division; s5, determining model parameters based on Bayesian optimization; s6, testing an optimal parameter model; s7, determining an output fusion scheme of the anomaly detection and fault recognition model, and performing overall testing; s8, optimal parameter model application; s9, an engineer performs maintenance according to conditions; and S10, updating the model when the data type or scale changes. Test verification shows that the scheme has a good effect on diagnosis of various faults under variable working conditions.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment fault diagnosis technology, specifically relating to a fault diagnosis method for equipment based on BO and intelligent models. Background Technology

[0002] Mechanical equipment is a core force supporting industrialization and modern society, and its development level directly affects the operational efficiency, structural optimization, and quality improvement of the social economy. As mechanical equipment continues to develop towards larger scale, automation, and intelligence, fault diagnosis technology plays an increasingly prominent role in ensuring stable equipment operation and improving operation and maintenance management. Therefore, building an efficient and accurate intelligent fault diagnosis system has become a key link in realizing intelligent operation and maintenance of mechanical equipment.

[0003] In actual operation, mechanical equipment often experiences various malfunctions due to factors such as vibration, impact, wear, and fatigue. These malfunctions not only degrade equipment performance but can also trigger systemic risks, leading to safety hazards and economic losses. Therefore, developing efficient and accurate intelligent fault diagnosis technology has become a key means to improve equipment operation and maintenance efficiency, reduce maintenance costs, and ensure production safety.

[0004] In recent years, fault diagnosis methods based on artificial intelligence technologies such as deep learning have received widespread attention. These methods can automatically extract fault features through data-driven approaches, achieving end-to-end state identification and diagnosis. However, the operating conditions of different types of equipment vary greatly, and the fault types are complex, making model performance susceptible to the influence of hyperparameter settings. Furthermore, traditional hyperparameter tuning methods often rely on empirical settings, making it difficult to balance accuracy and tuning efficiency. Therefore, Bayesian optimization (BO), due to its efficient search capability for black-box functions, has been introduced as an intelligent model hyperparameter tuning tool to improve the performance and adaptability of diagnostic models.

[0005] Therefore, there is an urgent need for a device fault diagnosis method that integrates Bayesian optimization and intelligent diagnostic models. This method should improve the efficiency and robustness of hyperparameter optimization while ensuring the accuracy and generalization ability of the models. In this way, a high-performance intelligent fault diagnosis system can be built to meet the high reliability and high intelligence requirements of modern industrial systems for equipment health management. Summary of the Invention

[0006] The purpose of this invention is to provide a fault diagnosis method for equipment based on BO and intelligent models, which can solve the problem of monitoring and detecting faults in the operation of mechanical equipment. The monitoring model and the fault identification model work in parallel, thereby achieving accurate fault identification during the service of mechanical equipment.

[0007] The specific technical solution adopted by this invention is as follows:

[0008] A device fault diagnosis method based on business logic (BO) and intelligent models includes the following steps:

[0009] S1: Application Scenario Data Analysis: Based on different application scenarios, clarify the data characteristics of the application scenario and select the corresponding diagnostic model or method;

[0010] S2: Intelligent Model Selection: Select the corresponding intelligent model for anomaly detection and fault identification based on the data analysis results of the application scenario;

[0011] S3: Determining Optimization Parameters: Determine optimization parameters by combining data characteristics, model type, and computing resource factors, and construct an optimization objective function based on performance indicators;

[0012] S4: Dataset partitioning: Based on the requirements of parameter optimization and optimal model testing, for the anomaly detection model, 70% of all normal data is used for training, 20% for validation, and 10% for testing; for all anomalous data, 50% is used for validation and 50% for testing. For the fault identification model, the training, validation, and testing data are partitioned in a ratio of 70%:20%:10%.

[0013] S5: Based on the training and validation sets, use the Bayesian optimization algorithm to determine the optimal parameters of the selected model;

[0014] S6: Based on the test set, use intelligent model evaluation metrics to modularly evaluate the optimal parameter model;

[0015] S7: Determine the output fusion scheme for anomaly detection and fault identification, and use the test dataset to obtain the test results of the overall intelligent diagnostic method;

[0016] S8: Apply the best diagnostic model when the model test results are acceptable, and accumulate actual data as needed during the application process;

[0017] S9: Engineers assess the results of testing and diagnosis and perform repairs as needed;

[0018] S10: Update the model when data type or data scale changes: Due to the complexity of the equipment's working environment and the increase in the equipment's service life, new normal data and fault data will be generated, which will lead to a decrease in the correctness and accuracy of the original model's diagnosis; updating the model includes updating the dataset, updating the basic parameters of the model, and updating the overall model, which can prevent missed detections caused by the failure model and enhance the credibility of the diagnostic model in providing guidance for equipment maintenance; considering the dynamic changes in equipment status, when there is a decline in model performance or a significant increase in data volume, the diagnostic model is updated. The model's data processing flow remains the same, maximizing the utilization of information, increasing the applicability of the model, and preventing the equipment fault from continuing to develop as much as possible.

[0019] The technical effects achieved by this invention are as follows:

[0020] 1. In view of the different data conditions of mechanical equipment, an adaptive fault diagnosis model based on Bayesian optimization is proposed, which is conducive to the practical application of the model and promotes the process of accurate equipment fault diagnosis.

[0021] 2. By using Bayesian optimization within the Optuna framework for both continuous and discrete parameters of the model, and combining this with data from the application devices, a more stable and reliable diagnostic model can be obtained.

[0022] 3. For equipment faults with limited data, a GAN model is used to generate fault data. This allows for the use of a high-performance supervised learning model, reducing the impact of limited data and data imbalance on fault diagnosis accuracy and expanding the application scope of intelligent diagnostic models.

[0023] 4. The outputs of unsupervised and supervised learning models are fused to enhance the diagnostic accuracy of the diagnostic model for mechanical equipment in complex states, providing a methodological reference for the widespread application of intelligent models in the future.

[0024] Compared to existing technologies, this method utilizes Bayesian optimization during model building, preventing diagnostic performance limitations caused by empirical parameters. Furthermore, the use of GAN data generation technology enables domains without diagnostic experience to perform intelligent diagnostics. Considering the potential increase in device state data over time, the adaptive model has upgrade capabilities, ensuring the diagnostic model's effectiveness and reliability throughout the device's service lifecycle. Attached Figure Description

[0025] Figure 1 This is a flowchart of a device fault diagnosis method based on BO and intelligent model according to the present invention;

[0026] Figure 2 This is the fault diagnosis process of the present invention;

[0027] Figure 3 This is a flowchart of the mechanical equipment fault diagnosis method based on Bayesian optimization and intelligent models according to the present invention.

[0028] Figure 4 shows the results of modular testing of the model in a practical example of this invention, wherein... Figure 4a yes

[0029] A schematic diagram of AKPCA's detection results for samples in normal and abnormal states; Figure 4b It is the test result of the fault identification module on the normal state and different degrees of rolling element, inner ring and outer ring faults in CWRU. Detailed Implementation

[0030] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0031] like Figure 1 Figure 4 shows a device fault diagnosis method based on BO and intelligent models, which includes the following steps:

[0032] S1: Application Scenario Data Analysis: Based on different application scenarios, clarify the data characteristics of the application scenario and select the corresponding diagnostic model or method;

[0033] S2: Intelligent Model Selection: Select the corresponding intelligent model for anomaly detection and fault identification based on the data analysis results of the application scenario;

[0034] S3: Determining Optimization Parameters: Determine optimization parameters by combining data characteristics, model type, and computing resource factors, and construct an optimization objective function based on performance indicators;

[0035] S4: Dataset partitioning: Based on the requirements of parameter optimization and optimal model testing, for the anomaly detection model, 70% of all normal data is used for training, 20% for validation, and 10% for testing; for all anomalous data, 50% is used for validation and 50% for testing. For the fault identification model, the training, validation, and testing data are partitioned in a ratio of 70%:20%:10%.

[0036] S5: Based on the training and validation sets, use the Bayesian optimization algorithm to determine the optimal parameters of the selected model;

[0037] S6: Based on the test set, use intelligent model evaluation metrics to modularly evaluate the optimal parameter model;

[0038] S7: Determine the output fusion scheme for anomaly detection and fault identification, and use the test dataset to obtain the test results of the overall intelligent diagnostic method;

[0039] S8: Apply the best diagnostic model when the model test results are acceptable, and accumulate actual data as needed during the application process;

[0040] S9: Engineers assess the results of testing and diagnosis and perform repairs as needed;

[0041] S10: Updating the model when data type or data scale changes: Due to the complexity of the equipment's working environment and the increase in the equipment's service life, new normal data and fault data will be generated, leading to a decrease in the accuracy and correctness of the original model's diagnosis. Updating the model includes updating the dataset, updating the basic parameters of the model, and updating the overall model. This can prevent missed detections caused by a failed model and enhance the credibility of the diagnostic model's guidance for equipment maintenance. Considering the dynamic changes in equipment status, when there is a decline in model performance or a significant increase in data volume, the diagnostic model is updated. The model's data processing flow remains the same, maximizing the utilization of information and increasing the model's applicability while preventing the continued development of equipment faults as much as possible. In this invention, step S10, in specific practical processes, for example, occurs when the engineer's experience judgment is inconsistent with the model's diagnostic results, and data analysis and review confirm that the model has misdiagnosed or missed diagnoses.

[0042] In this invention, the operating status of mechanical equipment includes normal operating status and abnormal operating status. Abnormal operating status covers various types and operating conditions of equipment malfunctions. The severity of malfunctions includes minor malfunctions, moderate malfunctions, and severe malfunctions.

[0043] Furthermore, the data characteristics in S1 include signal type, signal sampling frequency, and data accumulation status;

[0044] The operational data accumulation and corresponding processing solutions in S1 include: if the data types are complete and the data volume is large, a balanced fault diagnosis model is established; if the data is unbalanced or the data volume is small, data augmentation or GAN is used to increase the data volume.

[0045] In scenarios with imbalanced or insufficient data, the core process of increasing data volume through data augmentation and generative adversarial networks can be divided into the following steps:

[0046] First: Data Evaluation and Strategy Selection

[0047] Analyze the degree of data imbalance and feature distribution characteristics to identify key features of minority class samples; select augmentation methods for different data types: combine traditional augmentation with generative adversarial networks for image data, generate semantically coherent text through back translation or SeqGAN for text data, and use SMOTE interpolation to generate minority class samples for structured data; if the amount of data is extremely small, prioritize transfer learning to extract features.

[0048] Then, the generative adversarial network is trained.

[0049] Construct an adversarial model between the generator and the discriminator, and optimize the generation capability through alternating training; use Wasserstein GAN or conditional GAN ​​to improve stability and avoid mode collapse; for example, ResBaGAN enhances the reduction of data dependence and generates diverse samples by using residual networks and adaptive discriminators; introduce boundary constraints or hybrid loss functions during training to balance the generation quality.

[0050] Next, generate data quality verification.

[0051] To assess the consistency of the generated data distribution, the difference between the generated data and the real data is calculated using the maximum mean difference or KL divergence test, with a threshold of less than 0.5 considered acceptable. The authenticity is verified using the GAN-test method, where the generated data is used as a test set to check whether the classifier's recognition accuracy has significantly decreased. The feature entropy values ​​of the generated samples are statistically analyzed to ensure diversity coverage.

[0052] Then, data fusion and model training.

[0053] The augmented data is proportionally mixed into the original dataset, and the class weights are adjusted or the Focal Loss loss function is used to balance the loss function. The minority class recall and F1-score are monitored through cross-validation. If the improvement of the metrics is insufficient, the model parameters are iteratively optimized, such as adjusting the discriminator structure or introducing adversarial training to enhance robustness.

[0054] Finally, post-processing and closed-loop optimization.

[0055] Filter out low-quality generated samples and retain data that conforms to physical laws or business logic; continuously monitor model performance after deployment, and if the false positive rate increases, backtrack to adjust the enhancement strategy and retrain the generated model to form a closed-loop process of evaluation, enhancement, verification and iteration, ensuring that the generated data effectively improves the model's generalization ability.

[0056] Furthermore, in S2, when the sampling frequency of the acquired signal is ≥1kHz, the obtained signal contains richer high-frequency information and detailed features, and can more effectively reflect sudden anomalies, impact responses, and minute vibration changes during equipment operation. However, high-frequency sampling also leads to increased data volume, more redundant information, and higher requirements for computing and storage resources. For such complex high-frequency signals, it is advisable to use intelligent models with strong feature extraction and representation capabilities, including but not limited to:

[0057] Deep learning models, such as CNN, 1DCNN, and CNN-Transformer, can automatically learn multi-scale temporal or spatial features from raw signals.

[0058] Frequency domain and deep model fusion methods: such as STFT+CNN, Wavelet+CNN, etc., first retain instantaneous features through time-frequency analysis, and then use deep networks for efficient identification; Feature fusion + deep model strategy: fuse time domain, frequency domain and statistical features and input them into the deep model to enhance the robustness and accuracy of fault identification;

[0059] In S2, when the signal sampling frequency is low (<1kHz), the acquired signal mainly retains low-frequency information, while high-frequency features may be lost. The overall waveform is relatively stable, and the signal-to-noise ratio is high. Although it has certain limitations in reflecting sudden impacts or minor vibrations, it has high processing efficiency and lower computational resource consumption due to its small data volume, low redundancy, and low noise interference. Based on these characteristics, lightweight models with simple structures and high computational efficiency should be preferred, including but not limited to:

[0060] Traditional machine learning algorithms, such as Support Vector Machine (SVM) and Random Forest, are suitable for structured feature inputs.

[0061] Shallow neural network models, such as shallow CNNs or RNNs, have a certain ability to model time sequences and have relatively low computational overhead.

[0062] Statistical feature + classifier combination method: By extracting time domain, frequency domain or empirical feature vectors and combining them with a lightweight classifier for fault identification, it has good generalization ability and engineering practicality.

[0063] In S2, depending on the data accumulation of the signal, if the data accumulation is large and of various types, a balanced supervised learning model is selected; otherwise, a data generation model is selected to assist the supervised learning model.

[0064] In this invention, step S2, in specific practical processes, for example, provides guidance for the initial selection of the model based on the signal type of the applied data. Taking vibration acceleration signals as an example, various analysis methods can be employed: Method 1, extracting time-domain, frequency-domain, and time-frequency-domain features as input to machine learning algorithms, such as support vector machines and random forests; Method 2, directly inputting the original signal into shallow or deep neural networks to automatically complete feature learning and classification. For temperature or pressure signals, due to the low sampling rate and limited data volume, statistical methods such as kernel principal component analysis and machine learning algorithms are typically used for processing and analysis, such as support vector machines and random forests. Furthermore, the amount of data applied in step S2 provides guidance for the initial selection of the intelligent model. For single-state samples less than or equal to 200 and state or condition types less than or equal to 5, machine learning algorithms are recommended; for single-state samples greater than 200 and state or condition types greater than 5, shallow or deep learning algorithms are recommended.

[0065] Furthermore, in S3, the selection of the model parameters to be optimized is initially determined with the selection of the model, and the number of parameters to be optimized is determined according to the computing resources. If the intelligent model is a lightweight, shallow, or traditional ML model, all parameters can be optimized. If the intelligent model is a deep learning model or contains multiple models, the optimization parameters are selected based on the computing resources.

[0066] In this invention, step S3, in specific practical processes, for example: the determination of the parameters to be optimized in step S3 needs to be combined with the characteristics of the data. When the sample size of a single state is less than 200, a coarser optimization granularity should be used; when the sample size of a single state is greater than 200, a finer parameter optimization granularity can be used. The determination of the parameters to be optimized in step S3 also needs to be combined with the specific model type. For example, for CNN, the parameters to be optimized include six parameters: the number of layers in the convolutional neural network, the number of kernels in the first layer, the size of the first layer convolution, the learning rate, the number of layers in the multilayer perceptron, and the dropout value; for SVM, the parameters to be optimized include the kernel function type, the penalty factor C, and the kernel function parameter gamma that affects the model complexity; other models are determined according to the specific situation.

[0067] Furthermore, in S4, the dataset is divided into training, validation, and test sets in a 7:2:1 ratio when the amount of data accumulated is large; if the amount of data accumulated is small, the amount of data should be increased based on the current data using data augmentation, GAN, variational autoencoder, and diffusion model.

[0068] In this invention, step S4, in a specific practical process, for example, involves training data in step S4. The training data X is divided into N segments, each segment being a sample, and each sample has a dimension of m. In step S4, while ensuring that no data is missing, the data in the training, validation, and test sets with high sampling frequencies only undergo detrending processing, while the data in the training, validation, and test sets with low sampling frequencies are not processed. Specifically, the process involves: first, data evaluation and strategy selection; then, generating adversarial network training; next, generating data quality verification; then, data fusion and model training; and finally, post-processing and closed-loop optimization.

[0069] Furthermore, the basic steps of the BO algorithm in S5 include:

[0070] S51 initialization: Randomly sample several parameter points and evaluate their objective function values; the objective function values ​​include model accuracy or F1 score;

[0071] S52 establishes a surrogate model: fits the objective function using a Gaussian Process (GP) or a tree structure; tree structures include TPE.

[0072] S53 Acquisition Function: Based on the surrogate model, selects the next optimal candidate point; the optimal candidate point is the point with the greatest potential among the currently known points.

[0073] S54 Evaluate New Points: Evaluate the true objective function value for new sampling points.

[0074] S55 Update Model: Add new points to the sample set and update the surrogate model.

[0075] S56 Repeat steps S53–S55 until a termination condition is met, such as the number of iterations or a time limit.

[0076] In S5, the objective function is selected as the average of the accuracy and F1 score, taking into account both the accuracy and precision of the model.

[0077] In S5, the number of iterations is set according to the number of parameters. If the number of parameters is less than or equal to 5, the number of iterations is set to 50. If the number of parameters is greater than 5 and less than 10, the number of iterations is set to 100. If the number of parameters is greater than 10, the number of iterations is set to 200.

[0078] In step S5, the parameters of the diagnostic model are adaptively determined according to the actual data requirements, which not only enhances the accuracy and precision of the model, but also improves the model's generalization ability.

[0079] In this invention, step S5, in a specific practical process, for example, uses the automated hyperparameter optimization framework Optuna to implement Bayesian optimization. This framework, based on the TPE algorithm, can flexibly handle continuous parameters (such as the learning rate and dropout of CNNs) and discrete parameters (such as the number of layers and convolutional kernels in CNNs), thereby achieving efficient search for model hyperparameters.

[0080] The Bayesian optimization in step S5 focuses on the following problem: given a black-box objective function f(x), find the x in the input space x∈X that minimizes (or maximizes) f(x). * ,Right now

[0081]

[0082] f(x) cannot be explicitly expressed and is not differentiable.

[0083] The core idea of ​​Bayesian optimization in step S5 lies in the surrogate model and the acquisition function. The surrogate model establishes a probabilistic model of the true function f(x), representing the posterior distribution of f(x). The acquisition function defines a function based on the surrogate model to guide the next point at which to evaluate f(x). The acquisition function balances "exploration" and "exploitation".

[0084] In step S5, the default algorithm for the Bayesian optimization surrogate model implemented using Optuna is not a Gaussian process, but rather the Tree-structured Parzen Estimator (TPE), a Bayesian optimization method based on density estimation. The expectation-maximization objective of the traditional Bayesian optimization objective is improved to...

[0085]

[0086] Where y * It is the current optimal value, and p(y|x) is the posterior distribution of the target value.

[0087] The transformation approach for TPE is to use Bayes' theorem transformation:

[0088]

[0089] TPE uses an estimate of p(x|y) instead of p(y|x) and divides the historical data into two groups based on the target value y:

[0090] Good samples (target value less than threshold y) * ): Construct l(x) = p(x|y) <y * )

[0091] Bad samples (target value greater than y) * Construct g(x) = p(x|y≥y) * )

[0092] Then it is deduced

[0093]

[0094] Therefore, the strategy for choosing the next point is to find a point that makes The largest x.

[0095] Furthermore, the evaluation of the optimal parameter model in S6 is carried out by module testing, namely, testing the anomaly detection model and the fault identification model separately.

[0096] In this invention, step S6, in specific practical processes, for example: when the sample size of data for each state is balanced in step S6, the objective function Object for optimization can be set according to the specific situation. For example, a single objective optimization (maximizing accuracy or F1 score) can be used for simple tasks, while multi-objective optimization can be formulated for complex tasks according to the specific situation; the mean of accuracy and F1 score. In short, the optimization objective is a function related to the evaluation index of model performance, and the optimization objectives of anomaly detection and fault identification models can be different.

[0097] In step S6, when the sample size of each state data is unbalanced, the optimization objective can be the same as when it is balanced, but the impact of the imbalance must be reduced through weighting or other methods.

[0098] If the test results of each model in step S6 meet at least the qualification level evaluation indicators of the relevant standard GB 43555-2023 for intelligent models, the model is considered to have passed the test and can be further used; otherwise, the model parameters need to be further optimized or the model needs to be replaced.

[0099] Furthermore, the output fusion methods of different models in S7 include, but are not limited to, rule-based fusion methods, and the overall test results are mainly accuracy and F1 score.

[0100] Furthermore, in step S8, data is accumulated when a fault is diagnosed and when a misdiagnosis occurs. Specifically, in practical application, after the model test results achieve the expected performance, it is deployed to an actual diagnostic scenario, and a dynamic data accumulation mechanism is established. When the model successfully diagnoses a fault, the system automatically records complete input data, diagnostic conclusions, and maintenance feedback, forming a positive case library. If a misdiagnosis occurs, the deviation between the input features and the true results is highlighted, constructing a negative sample set. Through a real-time data acquisition module, the system continuously acquires information related to equipment operating status and environmental parameters, and performs multi-dimensional correlation analysis with the diagnostic results. For the accumulated fault and misdiagnosis data, an automated cleaning process is used to remove noise interference, feature engineering is used to extract key discriminant factors, and transfer learning is used to fuse new data with the original training set. The model retraining process is triggered periodically, using an incremental learning strategy to update model parameters, while adversarial verification ensures improved model generalization ability. This closed-loop mechanism enables continuous optimization of the diagnostic model in application, forming a reinforced loop of data accumulation, model iteration, and performance improvement.

[0101] For example, the opportunity for data collection in step S8 lies in the occurrence of anomalies / faults. When the model or engineer discovers a fault, data is accumulated and tagged during the equipment's operation while it is malfunctioning, providing a data foundation for subsequent model updates or upgrades.

[0102] Furthermore, in step S9, the engineer judges the results of the detection and diagnosis and maintains the equipment as appropriate. In this invention, step S8, in specific practical processes, for example, involves the detection and diagnosis results in step S9 including the engineer's experienced judgment and the direct results of the model. If the two contradict each other, further data analysis should be conducted to determine if a fault exists. If the fault is minor, it can be ignored or repairs can be planned; if the fault is general, repairs can be planned; if the fault is serious, immediate intervention in the equipment's operating status is necessary to prevent serious consequences. During this process, the engineer accumulates experience or the intelligent model is updated.

[0103] This invention addresses the current situation of varying data conditions in mechanical equipment by proposing an adaptive fault diagnosis model based on Bayesian optimization. This model is beneficial for practical application and for advancing the accuracy of equipment fault diagnosis.

[0104] This invention uses Bayesian optimization within the Optuna framework to optimize both continuous and discrete parameters of the model, and combines this with data from the application devices to obtain a more stable and reliable diagnostic model.

[0105] This invention addresses the issue of limited data for certain equipment faults by employing a GAN model to generate fault data. This allows for the use of a high-performance supervised learning model, reducing the impact of limited data and data imbalance on fault diagnosis accuracy and expanding the application scope of intelligent diagnostic models.

[0106] This invention fuses the outputs of unsupervised and supervised learning models to enhance the diagnostic accuracy of diagnostic models for mechanical equipment in complex states, providing a methodological reference for the widespread application of intelligent models in the future.

[0107] Compared with existing technologies, this method provides a device fault diagnosis approach based on Bayesian optimization and intelligent models. The method utilizes Bayesian optimization during model building to prevent diagnostic performance limitations caused by empirical parameters. The use of GAN data generation technology enables domains without diagnostic experience to perform intelligent diagnoses. Considering that device state data may increase with usage time, the adaptive model has upgrade capabilities, ensuring the diagnostic model is effective and reliable throughout the device's service lifecycle.

[0108] In practical cases, the present invention is described in the following combination. Figure 1 Figure 4 illustrates the specific embodiments of the invention patent in detail. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of this invention patent.

[0109] A mechanical equipment fault diagnosis method based on BO and intelligent models. In this practical case, publicly available data from Case Western Reserve University bearing fault data is used as an example. For ease of explanation, only normal state data from the 12kHz sampling frequency drive end (DE channel) and faults of inner rings, outer rings, and rolling elements at 0.007, 0.014, and 0.021 inches are used. The invention includes the following steps:

[0110] S1: Application Scenario Data Analysis: The current Case Western Reserve University bearing dataset contains normal conditions, different operating conditions, and varying degrees of inner ring, outer ring, and rolling element faults. The sampling frequency of this dataset is 12kHz, and the sample length is set to 1250 points (~0.104s). For ease of presentation and description, the descriptions of each state include: Normal state: Nor; Rolling element faults of varying degrees: B007, B014, B021; Inner ring faults of varying degrees: IR007, IR014, IR021; and Outer ring faults of varying degrees: OR007, OR014, OR021.

[0111] This embodiment presents experimental data on the normal operation of bearings under different operating conditions and the failure of inner ring, outer ring, and rolling elements to varying degrees at Case Western Reserve University.

[0112] S2: Smart Model Selection: Considering the balance of data types and the fact that the data volume is a small to medium-sized dataset, the lightweight smart models KPCA and CNN are selected.

[0113] S3: Optimization Parameter Determination: With limited data and disregarding computational costs, the key parameter for KPCA is the kernel function. Different kernel functions require different optimization parameters. Specifically, the parameters to be optimized for a polynomial kernel are gamma, degree, coef0, and n_components; for a sigmoid kernel, gamma, coef0, and n_components; for a Gaussian kernel (RBF), gamma and n_components; and for a linear kernel, n_components. Key parameters for CNNs include the number of convolutional layers, the number of kernels, the kernel size, the dropout rate, and the learning rate.

[0114] S4: Dataset Partitioning: For AKPCA, 70% of all normal data is used for training, 20% for validation, and 10% for testing; for all anomalous data, 80% is used for validation and 20% for testing. For CNN, training, validation, and testing data are divided in a 70%:20%:10% ratio.

[0115] S5: Based on the training and validation sets, the Boolean optimization (BO) algorithm is used to optimize AKPCA and CNN. To balance experience and learning, the empirical parameters are set as the starting point for BO optimization. If the empirical parameters are optimal, the optimization process will not miss the optimal result. If the empirical parameters are not optimal, the optimization process can find the optimal parameters.

[0116] When optimizing AKPCA, the Gaussian kernel (RBF) is set as the starting point. Based on the Optuna framework, the optimal kernel function for AKPCA is found, and the optimization parameters are shown in Table 1. The initial parameters and optimization ranges of each parameter of the CNN are shown in Table 2. The parameter optimization of the CNN is also implemented based on the Optuna framework.

[0117] Table 1 AKPCA Parameter Optimization Information

[0118]

[0119] Note: '-' means none.

[0120] Table 2 CNN parameter optimization information

[0121] During the optimization process, the F1 score obtained by inputting the validation dataset into the model is used to evaluate the model performance, and the parameter with the highest F1 score is selected.

[0122] S6: The test dataset undergoes the same processing as the training and validation data, then is input into the intelligent model to calculate the F1 score. An F1 score ≥ 0.9 indicates excellent model performance; 0.9 > F1 ≥ 0.8 indicates good model performance; 0.8 > F1 ≥ 0.7 indicates moderate model performance; 0.7 > F1 ≥ 0.6 indicates a satisfactory model; and 0.6 > F1 indicates an unsatisfactory model. The optimal model must reach at least a satisfactory level; otherwise, optimization is considered unsuccessful. In the implementation example, the optimal parameters of AKPCA and CNN, along with their corresponding validation and test accuracies and F1 scores, are recorded, as shown in Table 3. Figure 4 shows the visualized test results of the AKPCA and CNN modules. It can be seen that AKPCA exhibits excellent anomaly detection performance, while CNN demonstrates excellent fault identification capabilities. In summary, firstly, the model's performance on the validation and test data initially indicates that the model has a certain generalization ability. Secondly, referring to standard GB / T43555-2023, the model's validation and test accuracies indicate that the optimized model has reached a good or higher level.

[0123] Table 3 Optimal parameters and modular verification and test results

[0124]

[0125]

[0126] S7: Based on rules, the results of anomaly detection and fault identification models are fused. Specifically, if the detection result of AKPCA is normal and the diagnosis result of CNN is normal, the fused output is normal; if the detection result of AKPCA is abnormal and the detection result of CNN is a specific fault state, the fused result is abnormal, and the fault state is the primary fault to be eliminated; if the outputs of AKPCA and CNN contradict each other, it may be that there is a problem with the model or a new fault may have occurred.

[0127] To compare the advantages and disadvantages of this invention with single anomaly detection or fault identification methods, this embodiment compares the diagnostic results of single AKPCA, single CNN, and rule-based AKPCA fusion with CNN. The results are shown in Table 4. It can be seen that the extracted fault anomaly detection accuracy is high, the fault diagnosis accuracy is high, and the fusion result also completely and accurately diagnoses the fault type.

[0128] Table 4 Comparison of test results using different methods

[0129]

[0130] S8: Since publicly available data is used, the application and data accumulation process is omitted here; we proceed directly to condition-based maintenance. Engineers assess the fusion results, accumulate actual data as needed, and perform maintenance accordingly. This process is not convenient to demonstrate on publicly available online datasets, but it is generally certain that engineers need to pay close attention when the fusion results are contradictory or indicate a specific fault. Contradictory results may be caused by two situations: either a new fault has occurred, or the intelligent model has a problem. In either case, appropriate handling is necessary to avoid missed diagnoses and misdiagnoses. When a new fault occurs, regardless of the amount of data, data needs to be accumulated to lay the foundation for upgrading the model. If the fusion result indicates a specific fault, engineers can perform maintenance as needed.

[0131] S9: Since accumulating data and updating the model through monitoring is a long-term process, the monitoring process is accelerated here by experimentation (without waiting for each fault to occur) and the model is verified. Therefore, in actual use, you only need to operate according to the aforementioned process to ensure the accuracy and sustainability of the condition monitoring and fault identification model.

[0132] S10: Finally, update the model when the data type or data size changes.

[0133] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for equipment fault diagnosis based on BO and intelligent models, characterized in that: Includes the following steps: S1: Application Scenario Data Analysis: Based on different application scenarios, clarify the data characteristics of the application scenario and select the corresponding diagnostic model or method; S2: Intelligent Model Selection: Select the corresponding intelligent model for anomaly detection and fault identification based on the data analysis results of the application scenario; S3: Determining Optimization Parameters: Determine optimization parameters by combining data characteristics, model type, and computing resource factors, and construct an optimization objective function based on performance indicators; S4: Dataset partitioning: Based on the requirements of parameter optimization and optimal model testing, for the anomaly detection model, 70% of all normal data is used for training, 20% for validation, and 10% for testing; and 50% of all anomalous data is used for validation and 50% for testing. For the fault identification model, the training, validation, and test data are divided in a ratio of 70%:20%:10%; S5: Based on the training and validation sets, use the Bayesian optimization algorithm to determine the optimal parameters of the selected model; S6: Based on the test set, use intelligent model evaluation metrics to modularly evaluate the optimal parameter model; S7: Determine the output fusion scheme for anomaly detection and fault identification, and use the test dataset to obtain the test results of the overall intelligent diagnostic method; S8: Apply the best diagnostic model when the model test results are acceptable, and accumulate actual data as needed during the application process; S9: Engineers assess the results of testing and diagnosis and perform repairs as needed; S10: Update the model when the data type or data size changes: Updating the model includes updating the dataset, updating the basic parameters of the model, and updating the overall model.

2. The equipment fault diagnosis method based on BO and intelligent model according to claim 1, characterized in that: The data characteristics in S1 include signal type, signal sampling frequency, and data accumulation status; The operational data accumulation and corresponding processing solutions in S1 include: if the data types are complete and the data volume is large, a balanced fault diagnosis model is established; if the data is unbalanced or the data volume is small, data augmentation or GAN is used to increase the data volume. In scenarios with imbalanced or insufficient data, the core process of increasing data volume through data augmentation and generative adversarial networks can be divided into the following steps: First: Data Evaluation and Strategy Selection Analyze the degree of data imbalance and feature distribution characteristics to identify key features of minority class samples; select augmentation methods for different data types: combine traditional augmentation with generative adversarial networks for image data, generate semantically coherent text through back translation or SeqGAN for text data, and use SMOTE interpolation to generate minority class samples for structured data; if the amount of data is extremely small, prioritize transfer learning to extract features. Then, the generative adversarial network is trained. Construct an adversarial model between the generator and the discriminator, and optimize the generation capability through alternating training; use Wasserstein GAN or conditional GAN ​​to improve stability and avoid mode collapse; introduce boundary constraints or hybrid loss functions during training to balance the generation quality; Next, generate data quality verification. To assess the consistency of the generated data distribution, the difference between the generated data and the real data is calculated using the maximum mean difference or KL divergence test. A threshold of less than 0.5 is considered acceptable. Then, data fusion and model training. The augmented data is proportionally mixed into the original dataset, and the class weights are adjusted or the Focal Loss loss function is used to balance the loss function. The minority class recall and F1-score are monitored through cross-validation. If the improvement of the metrics is insufficient, the generated model parameters are iteratively optimized. Finally, post-processing and closed-loop optimization. Filter out low-quality generated samples and retain data that conforms to physical laws or business logic; After deployment, the model performance is continuously monitored. If the misclassification rate increases, the enhancement strategy is adjusted and the model is retrained to form a closed-loop process of evaluation, enhancement, verification, and iteration.

3. The equipment fault diagnosis method based on BO and intelligent model according to claim 2, characterized in that: In step S2, when the sampling frequency of the acquired signal is ≥1kHz, for such high-frequency complex signals, an intelligent model with strong feature extraction and representation capabilities is adopted, including but not limited to: deep learning models; frequency domain and deep model fusion methods. In S2, when the signal sampling frequency is low (<1kHz), a lightweight model with simple structure and high computational efficiency is selected, including but not limited to: Traditional machine learning algorithms; shallow neural network models; A combination of statistical features and classifiers; In S2, depending on the data accumulation of the signal, if the data accumulation is large and of various types, a balanced supervised learning model is selected; otherwise, a data generation model is selected to assist the supervised learning model.

4. The equipment fault diagnosis method based on BO and intelligent model according to claim 3, characterized in that: In S3, the selection of the model parameters to be optimized is initially determined with the selection of the model, and the number of parameters to be optimized is determined according to the computing resources. If the intelligent model is a lightweight, shallow and traditional ML model, all parameters can be optimized. If the intelligent model is a deep learning model or contains multiple models, the optimization parameters are selected based on the computing resources.

5. The equipment fault diagnosis method based on BO and intelligent model according to claim 4, characterized in that: In S4, the dataset is divided. When the amount of data accumulated is large, the training set, validation set, and test set are directly divided in a 7:2:1 ratio. If the amount of data accumulated is small, the amount of data should be increased based on the current data by using data augmentation, GAN, variational autoencoder, and diffusion model. In the specific process, firstly, data evaluation and strategy selection are performed; then, adversarial network training is carried out. Next, data quality verification is performed; then, data fusion and model training are conducted; finally, post-processing and closed-loop optimization are performed.

6. The equipment fault diagnosis method based on BO and intelligent model according to claim 5, characterized in that: The basic steps of the BO algorithm in S5 include: S51 initialization: Randomly sample several parameter points and evaluate their objective function values; the objective function values ​​include model accuracy or F1 score; S52 Establish a surrogate model: Fit the objective function using a Gaussian process or tree structure; S53 Acquisition Function: Based on the surrogate model, select the next optimal candidate point; the optimal candidate point is the point with the greatest potential among the currently known points. S54 Evaluate New Points: Evaluate the true objective function value at new sampling points; S55 Update Model: Add new points to the sample set and update the surrogate model; S56 Repeat steps S53–S55 until a termination condition is met, such as an iteration limit or time limit.

7. The equipment fault diagnosis method based on BO and intelligent model according to claim 6, characterized in that: The evaluation of the optimal parameter model in S6 is divided into module tests, namely, testing the anomaly detection model and the fault identification model separately. The output fusion methods of different models in S7 include, but are not limited to, rule-based fusion methods. The overall test results are mainly accuracy and F1 score.

8. The equipment fault diagnosis method based on BO and intelligent model according to claim 7, characterized in that: The S8 process accumulates data when a fault is diagnosed and when a misdiagnosis occurs. Specifically, after the model test results achieve the expected performance, it is deployed to the actual diagnostic scenario, and a dynamic data accumulation mechanism is established. When the model successfully diagnoses a fault, the system automatically records complete input data, diagnostic conclusions, and maintenance feedback to form a positive case library. If a misdiagnosis occurs, the deviation between the input features and the true results is highlighted to construct a negative sample set. Through the real-time data acquisition module, the system continuously acquires information related to equipment operating status and environmental parameters, and performs multi-dimensional correlation analysis with the diagnostic results. For the accumulated fault and misdiagnosis data, an automated cleaning process is used to remove noise interference, feature engineering is used to extract key discriminant factors, and transfer learning is used to fuse new data with the original training set. The model retraining process is triggered periodically, and the model parameters are updated using an incremental learning strategy. At the same time, adversarial verification is used to ensure the improvement of the model's generalization ability. This closed-loop mechanism enables the diagnostic model to be continuously optimized in application, forming a reinforcing loop of data accumulation, model iteration, and performance improvement.

9. A device fault diagnosis method based on BO and intelligent model according to claim 8, characterized in that: In S9, the engineer judges the results of the detection and diagnosis and maintains the equipment as needed.

Images