Rotating machinery diagnosis method and system based on adaptive wavelet and CNN

Abstract

The present application relates to the field of rotating machinery equipment fault diagnosis, and provides a rotating machinery diagnosis method and system based on adaptive wavelet and CNN, a multi-working-condition finite element simulation model is constructed, simulation torque signals and simulation vibration signals of the rotating machinery equipment are collected, and a mapping relationship between the two kinds of simulation signals is established; measured torque signals and measured vibration signals of the rotating machinery equipment are collected, and a mapping relationship between the two kinds of measured signals is established; adaptive multi-scale DB wavelet transformation and improved threshold denoising are performed on the four kinds of signals respectively, and pretreatment and data mixed enhancement are performed, simulation and measurement features are extracted, a space-channel attention fusion multi-branch CNN model is constructed, simulation pre-training and measured fine-tuning transfer learning training are performed, the trained CNN model performs fault diagnosis on the four kinds of signals, the diagnosis results of the four kinds of signals are cross-validated, when the cross-validation passes, the fault type, severity and confidence are output.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis technology, and in particular to a method and system for diagnosing rotating machinery faults based on adaptive wavelet and CNN (Convolutional Neural Network). Background Technology

[0002] Industrial fans, pumps, gearboxes, and other rotating machinery are core equipment in industrial production, shipping, energy and chemical industries. They operate under high speed, heavy load, and harsh conditions for extended periods, making them prone to rotor bar breakage, air gap eccentricity, bearing wear, and combined faults. Failure to diagnose these faults in a timely manner can lead to serious safety accidents and economic losses.

[0003] Existing fault diagnosis technologies for rotating machinery have the following drawbacks:

[0004] (1) Signal level: The mainstream uses vibration and current signals, which are easily affected by environmental noise and load fluctuations. Although torque signals can directly reflect the mechanical transmission characteristics, existing technologies lack dedicated feature extraction methods for torque signals and have not solved the problem of insufficient reliability of single signal sources.

[0005] (2) Wavelet Transform Level: The DB (DauBechies) wavelet transform is an orthogonal wavelet transform with excellent properties such as tight support, orthogonality, and vanishing moments. It has good time-frequency localization and multi-resolution analysis capabilities, and is a signal processing tool that breaks through the limitations of Fourier analysis. It is widely used in signal compression, image processing, pattern recognition, and numerical analysis. The DB wavelet transform uses a fixed wavelet basis and a constant correlation threshold, which cannot adapt to the time-frequency characteristics of different fault types. Single-scale decomposition is prone to losing subtle fault information, and the traditional threshold function has a constant deviation, which leads to the smoothing of fault mutation characteristics.

[0006] (3) Network model level: Conventional CNNs only use basic convolutional structures and lack the ability to selectively focus on fault features. Key fault areas are easily submerged by background noise. Single-scale convolutional kernels cannot capture global and local fault features at the same time, and can only achieve fault classification, but cannot quantify the severity of faults.

[0007] (4) Sample construction level: Conventional finite element simulation samples have significant domain offset from actual industrial data, and the generalization ability of directly trained models is extremely poor; existing data augmentation methods have limited effect on generating composite fault samples, which can easily lead to model overfitting.

[0008] (5) Working condition adaptation level: Existing methods are mostly designed for fixed working conditions and cannot adaptively adjust model parameters to adapt to complex industrial scenarios with variable speed and load. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by proposing a fault diagnosis method and system for rotating machinery based on adaptive wavelet and CNN. This method utilizes adaptive multi-scale DB wavelet feature extraction and attention-enhanced multi-branch CNN as its core, combined with adaptive calibration in the simulation-measurement domain, to achieve accurate classification and quantitative assessment of the severity of rotating machinery faults under varying operating conditions. It outputs the fault type, severity, and confidence level of the rotating machinery to be diagnosed in real time, providing a quantitative basis for the maintenance of rotating machinery.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a fault diagnosis method for rotating machinery based on adaptive wavelet and CNN, comprising the following steps:

[0012] S1. The processor uses electromagnetic field finite element simulation software to construct several multi-condition finite element simulation models based on the factory parameters and fault types of rotating machinery. These models simulate the operating condition fluctuations of rotating machinery in different industrial scenarios, covering the normal state, single fault state, and compound fault state of rotating machinery. At least three multi-condition finite element simulation models are constructed for each state. Each multi-condition finite element simulation model simulates at least one simulated torque sensor and at least one simulated vibration sensor to collect the simulated torque signal and simulated vibration signal of rotating machinery under multiple conditions, and establishes the mapping relationship between the simulated torque signal and simulated vibration signal of rotating machinery.

[0013] A torque sensor is connected in series between the output shaft of the drive motor and the load input end of the rotating machinery to collect the measured torque signal of the rotating machinery; a vibration sensor patch is installed on the radial surface of the bearing housing at the drive end and non-drive end of the rotating machinery to collect the measured vibration signal of the rotating machinery; and the processor establishes the mapping relationship between the measured torque signal and the measured vibration signal of the rotating machinery.

[0014] S2. The processor performs adaptive multi-scale multi-Bessie DB wavelet transform and improved threshold denoising on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of the rotating machinery, respectively, to generate simulated multi-scale fused fault detail map and measured multi-scale fused fault detail map of the rotating machinery.

[0015] S3 and the processor perform preprocessing and data mixing enhancement Mixup on the simulated multi-scale fused fault detail map and the measured multi-scale fused fault detail map of the rotating machinery respectively, and construct the simulated fault sample library and the measured fault sample library of the rotating machinery respectively.

[0016] S4. The processor establishes multiple parallel branches, and embeds a spatial-channel attention fusion convolutional neural network (CNN) model after each branch to obtain a spatial-channel attention fusion multi-branch CNN model; the processor extracts the features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery respectively; the processor introduces maximum mean difference (MMD) loss, and performs adaptive calibration and transfer learning training on the spatial-channel attention fusion multi-branch CNN model using the extracted features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery to obtain the trained spatial-channel attention fusion multi-branch CNN model;

[0017] S5. The processor preprocesses the simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal of the rotating machinery under the same working condition. The preprocessed simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal are then input into the trained spatial-channel attention fusion multi-branch CNN model. The trained spatial-channel attention fusion multi-branch CNN model performs fault diagnosis on the input simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal, respectively, and obtains the fault diagnosis results of the four signals.

[0018] S6. The processor performs cross-validation on the fault diagnosis results of the four signals: when all four fault diagnosis results are the same, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery to be diagnosed are output in real time, providing a quantitative basis for the maintenance of the rotating machinery; otherwise, the cross-validation is confirmed to be unsuccessful, and the processor uses the weighted fusion method of evidence theory to determine which signal needs to be re-diagnosed based on the fault type, severity, and confidence level of the rotating machinery in the obtained fault diagnosis results of the four signals, and performs subsequent fault diagnosis based on the determination result; the trained spatial-channel attention fusion multi-branch CNN model completes parameter adaptive update.

[0019] In a second aspect, the present invention provides a fault diagnosis system for rotating machinery based on adaptive wavelet and CNN for performing the method described in the first aspect, comprising:

[0020] A torque sensor is installed in series between the output shaft of the drive motor of a rotating machinery and the load input end to collect the measured torque signal of the rotating machinery.

[0021] Vibration sensors are patched and mounted on the radial surface of the bearing housing housing at both the drive end and non-drive end of rotating machinery to collect measured vibration signals from the rotating machinery.

[0022] Processor, including:

[0023] The measured signal mapping relationship establishment module is used to: establish the mapping relationship between the measured torque signal and the measured vibration signal of rotating machinery;

[0024] The finite element simulation model construction module is used to: construct several multi-condition finite element simulation models based on the factory parameters and fault types of rotating machinery using electromagnetic field finite element simulation software. These models simulate the operating condition fluctuations of rotating machinery in different industrial scenarios, covering the normal state, single fault state, and compound fault state of rotating machinery. At least three multi-condition finite element simulation models are constructed for each state. Each multi-condition finite element simulation model simulates at least one simulated torque sensor and at least one simulated vibration sensor to collect the simulated torque signal and simulated vibration signal of rotating machinery under multiple conditions, and to establish the mapping relationship between the simulated torque signal and simulated vibration signal of rotating machinery.

[0025] The multi-scale fusion fault detail map generation module is used to: perform adaptive multi-scale multi-Bessie DB wavelet transform and improved threshold denoising on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of rotating machinery, respectively, to generate simulated multi-scale fusion fault detail maps and measured multi-scale fusion fault detail maps of rotating machinery.

[0026] The module for constructing simulation and measured fault sample libraries is used to: preprocess and perform data mixing and enhancement Mixup on the simulation multi-scale fusion fault detail map and the measured multi-scale fusion fault detail map of rotating machinery respectively, and construct the simulation fault sample library and the measured fault sample library of rotating machinery respectively.

[0027] The spatial-channel attention fusion multi-branch CNN model construction module is used to: establish multiple parallel branches, embed a spatial-channel attention fusion convolutional neural network (CNN) model after each branch, and obtain a spatial-channel attention fusion multi-branch CNN model; each branch includes convolutional layers, normalization layers, modified linear unit (ReLU) activation layers, and max pooling layers; the spatial-channel attention fusion multi-branch CNN model adopts a channel attention mechanism to filter out key feature channels related to faults, and uses a spatial attention mechanism in the key feature channels to focus on the key regions of fault features;

[0028] The adaptive calibration and transfer learning training module is used to: extract the features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery respectively; introduce the maximum mean difference (MMD) loss, and use the extracted features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery to perform adaptive calibration and transfer learning training on the spatial-channel attention fusion multi-branch CNN model to obtain the trained spatial-channel attention fusion multi-branch CNN model;

[0029] The fault diagnosis and adaptive update module is used to: preprocess the simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal of the rotating machinery under the same operating condition; input the preprocessed simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal into a trained spatial-channel attention fusion multi-branch CNN model; the trained spatial-channel attention fusion multi-branch CNN model performs fault diagnosis on the input simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal respectively, obtaining fault diagnosis results for the four signals; and analyze the fault diagnosis results for the four signals. Cross-validation of the fault diagnosis results: When the fault diagnosis results of the four signals are all the same, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery to be diagnosed are output in real time, providing a quantitative basis for the maintenance of the rotating machinery; otherwise, the cross-validation is confirmed to be unsuccessful, and the processor, based on the fault diagnosis results of the four signals, uses a weighted fusion method of evidence theory to determine which signal needs to be re-diagnosed, and performs subsequent fault diagnosis based on the determination result; the trained spatial-channel attention fusion multi-branch CNN model completes adaptive parameter updates;

[0030] The torque sensor and vibration sensor are connected to the measured signal mapping relationship module to establish a data flow. The measured signal mapping relationship module and the finite element simulation model construction module are connected to the multi-scale fusion fault detail map generation module to establish a data flow. The multi-scale fusion fault detail map generation module, the simulation and measured fault sample library construction module, the spatial-channel attention fusion multi-branch CNN model construction module, the adaptive calibration and transfer learning training module, and the fault diagnosis and adaptive update module are connected to each other in sequence to establish a data flow.

[0031] Thirdly, the present invention provides an electronic device comprising: at least one processor, at least one memory, and a communication interface, wherein the processor, memory, and communication interface communicate with each other; the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the method described in the first aspect.

[0032] Fourthly, the present invention provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to perform the method described in the first aspect.

[0033] Compared with existing technical solutions, the beneficial effects of the embodiments of the present invention are reflected in the following aspects:

[0034] (1) Significantly improved feature extraction accuracy: The present invention proposes an adaptive multi-scale DB wavelet transform and an improved exponential threshold denoising method. The adaptive multi-scale DB wavelet improves the accuracy of fault feature extraction, and the improved exponential threshold function retains high fault mutation features. It can solve the problems that fixed DB wavelet basis and constant threshold cannot adapt to different fault types and feature extraction is incomplete.

[0035] (2) Significantly enhanced diagnostic accuracy and generalization ability: The embodiments of this invention design a multi-branch CNN model with spatial-channel dual attention fusion, which can solve the problems of low attention to key fault features and insufficient multi-scale feature fusion of conventional CNNs. The embodiments of this invention propose an adaptive calibration and transfer learning training strategy for simulation-test data. The average diagnostic accuracy of the attention fusion multi-branch CNN model is relatively stable under varying working conditions, and the improvement effect is significant compared with the basic CNN. It can solve the problems of low diagnostic accuracy, poor recognition of composite faults, and poor generalization ability of the model caused by the domain offset between simulation samples and test data in traditional methods.

[0036] (3) Enhance multi-task diagnostic capabilities: The embodiments of the present invention construct a classification-regression parallel multi-task diagnostic framework, which for the first time realizes the parallel output of fault type identification and severity quantification. The average absolute error of severity assessment is small, providing a quantitative basis for predictive maintenance of equipment and solving the problem that existing methods can only identify fault types but cannot quantify severity.

[0037] (4) Balancing Real-Time Performance and Reliability: This embodiment of the invention establishes a cross-validation mechanism for torque and vibration signals under the same operating condition: Judgment is made based on the torque signal, but the result cannot be directly output. The vibration signal under the same operating condition is required. Only if the diagnostic results of the two signals under the same operating condition are identical can the diagnostic result be output. The fundamental frequency and harmonic components of the torque signal are synchronized with the rotational speed and load torque, providing a direct and sensitive response to periodic load anomalies, torque pulsations, and shaft torque fluctuations in rotating machinery. The vibration signal has high sensitivity to local impact faults and can capture high-frequency transient impact characteristics that are difficult for the torque signal to reflect. This embodiment of the invention leverages the complementary advantages of torque and vibration signals, improving the accuracy and robustness of fault diagnosis through cross-validation of the two diagnostic results. The multi-branch CNN model with spatial-channel dual-attention fusion has 70% fewer parameters than the conventional CNN model, resulting in shorter single-sample diagnosis time. The torque-vibration signal cross-validation mechanism lowers the false alarm rate and solves the problem of insufficient reliability of a single torque signal.

[0038] (5) Strong engineering adaptability: The embodiments of the present invention integrate adaptive wavelet transform and attention-enhanced convolutional neural network to perform intelligent diagnosis of rotating machinery faults. It supports complex working conditions with variable speed and load, and is especially suitable for the identification of fault types and the quantitative assessment of severity of rotating machinery under variable load and high noise conditions such as ship shipping, industrial fans, water pumps, and gearboxes. It can be directly applied to online monitoring and fault diagnosis of various rotating machinery such as ship propulsion systems, industrial fans, and water pumps, and has broad industrial application prospects. Attached Figure Description

[0039] Figure 1 This is a flowchart of a fault diagnosis method for rotating machinery based on adaptive wavelet and CNN in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the finite element simulation model of the core transmission component in a rotating mechanical device under normal conditions.

[0041] Figure 3 This is a schematic diagram of a finite element simulation model of a rotating mechanical device in the state of rotor bar breakage in an embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram of a finite element simulation model of a rotating mechanical device in an air gap eccentricity fault state in an embodiment of the present invention.

[0043] Figure 5 This is a schematic diagram of a finite element simulation model of a rotating mechanical device in an embodiment of the present invention, showing a combined fault state in which rotor bar breakage and air gap eccentricity coexist.

[0044] Figure 6This is a structural diagram of the channel attention fusion CNN model in this embodiment of the invention, showing the connection relationships, parameter settings and core functions of each network layer.

[0045] Figure 7 This is an overall architecture diagram of the multi-task CNN model in this embodiment of the invention.

[0046] Figure 8 This is a characteristic diagram of the normal state of rotating mechanical equipment in an embodiment of the present invention.

[0047] Figure 9 This is a characteristic diagram of rotor bar breakage fault in a rotating mechanical device according to an embodiment of the present invention.

[0048] Figure 10 This is a fault feature diagram of air gap eccentricity in a rotating mechanical device according to an embodiment of the present invention.

[0049] Figure 11 This is a composite fault feature diagram of a rotating mechanical device in an embodiment of the present invention, where rotor bar breakage and air gap eccentricity coexist.

[0050] Figure 12 This is a comparison chart of learning rates for different optimizers in this embodiment of the invention, showing the changes in accuracy and loss function of three optimizers, Adam, Rmsprop, and Sdgm, under different learning rates.

[0051] Figure 13 This is a flowchart of graph preprocessing for fault feature details in an embodiment of the present invention, illustrating standardized preprocessing steps of pruning, compression, and data augmentation.

[0052] Figure 14 This is a flowchart of the selection process for the adaptive DB wavelet basis in an embodiment of the present invention.

[0053] Figure 15 This is a structural diagram of the spatial-channel dual attention module in an embodiment of the present invention.

[0054] Figure 16 This is a schematic diagram of the simulation-measurement domain adaptive calibration principle in an embodiment of the present invention.

[0055] Figure 17 This is a structural block diagram of a rotating machinery equipment fault diagnosis system based on adaptive wavelet and CNN in an embodiment of the present invention.

[0056] Figure reference numerals: 1-stator, 2-stator guide bar, 3-rotor, 4-rotor guide bar, 5-broken rotor bar, D1-uneven air gap distance one, D2-uneven air gap distance two. Detailed Implementation

[0057] This invention breaks through the traditional diagnostic framework of wavelets and basic CNNs. It takes adaptive multi-scale DB wavelet feature extraction and attention-enhanced multi-branch CNN as the core, and combines simulation-measurement domain adaptive calibration to achieve accurate classification and quantitative assessment of the severity of faults in rotating machinery under varying operating conditions.

[0058] See Figure 1 As shown, this embodiment of the invention provides a fault diagnosis method for rotating machinery based on adaptive wavelet and CNN, including the following steps:

[0059] S1. The processor uses electromagnetic field finite element simulation software to construct several multi-condition finite element simulation models based on the factory parameters and fault types of rotating machinery. These models simulate the operating condition fluctuations of rotating machinery in different industrial scenarios, covering the normal state, single fault state, and compound fault state of rotating machinery. At least three multi-condition finite element simulation models are constructed for each state. Each multi-condition finite element simulation model simulates at least one simulated torque sensor and at least one simulated vibration sensor to collect the simulated torque signal and simulated vibration signal of rotating machinery under multiple conditions, and establishes the mapping relationship between the simulated torque signal and simulated vibration signal of rotating machinery.

[0060] A torque sensor is connected in series between the output shaft of the drive motor and the load input end of the rotating machinery to collect the measured torque signal of the rotating machinery; a vibration sensor patch is installed on the radial surface of the bearing housing at the drive end and non-drive end of the rotating machinery to collect the measured vibration signal of the rotating machinery; and the processor establishes the mapping relationship between the measured torque signal and the measured vibration signal of the rotating machinery.

[0061] S2. The processor performs adaptive multi-scale multi-Bessie DB wavelet transform and improved threshold denoising on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of the rotating machinery, respectively, to generate simulated multi-scale fused fault detail map and measured multi-scale fused fault detail map of the rotating machinery.

[0062] S3 and the processor perform preprocessing and data mixing enhancement Mixup on the simulated multi-scale fused fault detail map and the measured multi-scale fused fault detail map of the rotating machinery respectively, and construct the simulated fault sample library and the measured fault sample library of the rotating machinery respectively.

[0063] S4. The processor establishes multiple parallel branches, and embeds a spatial-channel attention fusion convolutional neural network (CNN) model after each branch to obtain a spatial-channel attention fusion multi-branch CNN model; the processor extracts the features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery respectively; the processor introduces maximum mean difference (MMD) loss, and performs adaptive calibration and transfer learning training on the spatial-channel attention fusion multi-branch CNN model using the extracted features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery to obtain the trained spatial-channel attention fusion multi-branch CNN model;

[0064] S5. The processor preprocesses the simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal of the rotating machinery under the same working condition. The preprocessed simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal are then input into the trained spatial-channel attention fusion multi-branch CNN model. The trained spatial-channel attention fusion multi-branch CNN model performs fault diagnosis on the input simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal, respectively, and obtains the fault diagnosis results of the four signals.

[0065] S6. The processor performs cross-validation on the fault diagnosis results of the four signals: when all four fault diagnosis results are the same, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery to be diagnosed are output in real time, providing a quantitative basis for the maintenance of the rotating machinery; otherwise, the cross-validation is confirmed to be unsuccessful, and the processor uses the weighted fusion method of evidence theory to determine which signal needs to be re-diagnosed based on the fault type, severity, and confidence level of the rotating machinery in the obtained fault diagnosis results of the four signals, and performs subsequent fault diagnosis based on the determination result; the trained spatial-channel attention fusion multi-branch CNN model completes parameter adaptive update.

[0066] The following provides a detailed explanation of steps S1-S6.

[0067] In step S1, the processor uses electromagnetic field finite element simulation software to construct several multi-condition finite element simulation models based on the factory parameters and fault types of the rotating machinery equipment, including the following steps:

[0068] The processor employs electromagnetic field finite element simulation software. Based on the rated speed ±30%, the light and heavy load ranges of the rated load, and the rated power, rated voltage, and structural dimensions of the transmission components of the rotating machinery, several multi-condition finite element simulation models of different fault severity are constructed for rotor bar breakage and air gap eccentricity fault types. These models cover the normal state, single fault state, and compound fault state of the rotating machinery, simulating the operating condition fluctuations of rotating machinery in actual industrial scenarios and generating gradient fault samples of different types of rotating machinery. In a transient field simulation environment, simulation time is set, and independent simulations are performed on different types of gradient fault samples of the rotating machinery. Simulation torque signals and simulation vibration signals of each multi-condition finite element simulation model are collected to obtain the original datasets of simulation torque signals and simulation vibration signals of the rotating machinery in normal state and various fault states, which are used to characterize the simulation dynamic features of the rotating machinery.

[0069] Using rotating machinery such as industrial fans, pumps, and gearboxes as examples, finite element simulation models are employed to avoid interference from other environmental factors in the data acquisition. Torque signals are the core diagnostic basis, while vibration signals are used for auxiliary verification.

[0070] A torque sensor is connected in series between the output shaft of the drive motor of the rotating machinery and the input end of the load (e.g., propulsion shaft or gearbox). In this embodiment, the torque sensor is a high-precision non-contact dynamic torque sensor. A vibration sensor patch is installed on the radial surface of the bearing housing at both the drive end and the non-drive end of the rotating machinery motor. To cover high-frequency harmonics and low-frequency fault characteristics such as rotor bar breakage, the sampling frequency of the torque and vibration signals is uniformly set to 10kHz. The torque sensor collects the torque signal of the rotating machinery, and the vibration sensor collects the vibration signal of the rotating machinery to obtain the core parameters of the rotating machinery: rated power, rated speed, rated voltage, and structural dimensions of the transmission components.

[0071] Covering the rated speed of rotating machinery within ±30%, light load, heavy load, and other ranges, this simulation model simulates the operating condition fluctuations of rotating machinery in actual industrial scenarios. Based on torque signals collected by torque sensors, vibration signals collected by vibration sensors, and core parameters of the rotating machinery, finite element simulation software such as Ansys Maxwell electromagnetic field finite element simulation software, COMSOL multiphysics simulation software, and EasiMotor (motor simulation design platform) are used to establish multi-condition finite element simulation models of rotating machinery under normal, single fault, and compound fault conditions. The multi-condition finite element simulation models establish the mapping relationship between torque and vibration signals of the rotating machinery.

[0072] For rotor bar breakage and air gap eccentricity faults in rotating machinery, simulation models of different fault severity levels are constructed to generate gradient fault samples for the rotating machinery. To avoid mutual interference between fault states of the rotating machinery, multiple identical basic simulation models are set up to independently simulate different fault types of the rotating machinery. In a transient field simulation environment, the simulation time is set, and signal data of each model are collected to obtain the original signal datasets of normal and various fault states of the rotating machinery, which are used to characterize the dynamic features of the rotating machinery.

[0073] Regarding the gradient modeling of the severity of faults in rotating machinery: For different fault types such as the number of broken rotor bars (1, 2, 3, 4 broken bars) and air gap eccentricity (10%, 20%, 30%, 40% eccentricity), simulation models of rotating machinery with different severity levels are constructed to generate gradient fault samples of rotating machinery.

[0074] Table 1. Quantitative Labeling Table for Different Fault Types in Rotating Machinery Equipment

[0075]

[0076] As shown in Table 1, the quantitative labels for different fault types of rotating machinery are different. For example, the quantitative label for the normal state of rotating machinery is set to 0, the quantitative label for a fault with 1 broken rotor bar is 1, the quantitative label for a fault with 2 broken rotor bars is 2, the quantitative label for a fault with 3 broken rotor bars is 3, the quantitative label for a fault with 4 broken rotor bars is 4, the quantitative label for a fault with 10% air gap eccentricity is 5, the quantitative label for a fault with 20% air gap eccentricity is 6, the quantitative label for a fault with 30% air gap eccentricity is 7, the quantitative label for a fault with 40% air gap eccentricity is 8, and the quantitative label for a combined fault is 9.

[0077] S2. The processor performs adaptive multi-scale multi-Bessie DB wavelet transform and improved threshold denoising on the simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal of the rotating machinery, respectively, to generate simulated multi-scale fused fault detail map and measured multi-scale fused fault detail map of the rotating machinery.

[0078] Step S2 includes the following steps:

[0079] S201 The processor adopts the adaptive Dobessi (DB) wavelet basis selection algorithm to calculate the mutual information entropy between different wavelet bases (n=1~10) and the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of rotating machinery. It automatically selects the wavelet base with the largest mutual information entropy as the optimal base for this calculation, thus solving the problem of poor adaptability of fixed wavelet bases.

[0080] S202. Based on the optimal basis of this calculation, the processor sequentially uses wavelet low-pass filter and wavelet high-pass filter to perform multi-level DB wavelet decomposition on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of the rotating machinery equipment, respectively, to obtain the multi-level detail coefficients and approximation coefficients of the four signals, and capture the fault characteristics of the rotating machinery equipment at different levels.

[0081] The adaptive exponential threshold function used by the processor when performing multi-level DB wavelet decomposition is:

[0082] ,

[0083] in, Let be the adaptive exponential threshold corresponding to the k-th level wavelet decomposition, where k is the number of wavelet decomposition levels, k = 1, 2, 3, 4, ..., 10. Let be the noise standard deviation of the detail coefficients at the k-th layer, and ln be a logarithmic function with the natural constant e as the base. This represents the total number of sampling points for the original signal. The wavelet detail coefficients are obtained from the k-th level wavelet decomposition. It is an adaptive adjustment factor that solves the problems of discontinuity in traditional hard thresholds and constant deviation in soft thresholds.

[0084] , The coefficient for smooth attenuation is denoted by e, where e is the natural constant.

[0085] To smooth out the attenuation coefficient, a value of 0.5 is preferred in engineering. The physical meaning of this rule is that it addresses high-frequency shallow detail (…). (Smaller) Contains more high-frequency fault characteristics of rotating machinery, at this time Larger, higher threshold, for powerful noise reduction; deep low-frequency details ( (Larger) Contains more low-frequency fault characteristics of rotating machinery, at this time Smaller thresholds are used to preserve subtle fault mutation characteristics.

[0086] The processor calculates the Pearson coefficients, which are related to the fault type of rotating machinery, for each level of the four signals and uses them as the weight coefficients for that level. Then, the multi-level detail coefficients of the four signals are weighted and fused to obtain the fused rotating machinery fault characteristic signal.

[0087] S203 The processor converts the fused four signals of the rotating machinery fault feature signal into red, green and blue RGB three-channel images, which serve as the simulated multi-scale fused fault detail image and the measured multi-scale fused fault detail image of the rotating machinery.

[0088] DB wavelet transform is an algorithm based on dual-tree complex wavelet transform. It improves the accuracy and efficiency of the transform by expanding wavelet basis functions in two different directions. This transform is particularly useful in processing image and audio signals, capturing signal features in different directions. In this embodiment of the invention, the DB wavelet is selected as the signal processing kernel function. Leveraging its excellent time-frequency localization, multi-resolution analysis, compact support, orthogonality, and bioorthogonality properties, fault features in the torque signal of rotating machinery are accurately extracted.

[0089] Multi-scale hierarchical decomposition and improved exponential threshold denoising: The acquired signal is decomposed into 4-level DB wavelet decomposition. The signal is then decomposed through wavelet low-pass filter and wavelet high-pass filter to obtain level 1 to 4 detail coefficients and approximation coefficients, capturing the fault characteristics of rotating machinery at different levels. Wavelet low-pass filter and wavelet high-pass filter are the core components of discrete wavelet transform, used to decompose the signal into subbands of different frequencies.

[0090] S3. To reduce the computational load of the model and improve the real-time performance, versatility, and adaptability of fault diagnosis for rotating machinery, the processor performs preprocessing and data mixing enhancement Mixup on the simulated multi-scale fused fault detail map and the measured multi-scale fused fault detail map of the rotating machinery, respectively, to construct the simulated fault sample library and the measured fault sample library of the rotating machinery.

[0091] Step S3 includes the following steps:

[0092] S301 and the processor perform adaptive histogram equalization on the simulated multi-scale fusion fault detail map and the measured multi-scale fusion fault detail map of the rotating machinery equipment, respectively, to enhance the contrast of the fault change region of the rotating machinery equipment, highlight the edge features, and obtain the simulated feature map and measured feature map of the rotating machinery equipment with different fault types and different fault severity.

[0093] S302 and the processor linearly blend the simulation feature maps and measured feature maps of different fault types and severity of rotating machinery equipment to generate simulation feature maps and measured feature maps of composite fault samples of rotating machinery equipment; and respectively trim the blank areas in the simulation feature maps and measured feature maps of composite fault samples of rotating machinery equipment, retaining only the areas containing fault features, to obtain the trimmed simulation fault feature maps and measured fault feature maps of rotating machinery equipment.

[0094] S303 The processor preprocesses the cropped simulated fault feature map and measured fault feature map of the rotating machinery, and compresses the preprocessed simulated fault feature map and measured fault feature map of the rotating machinery into a uniform size while maintaining the same channel color information. This uniform size is a generalized image size that is suitable for the recognition of simulated fault feature maps and measured fault feature maps of various rotating machinery.

[0095] S304. The processor performs random data augmentation Mixup on the simulated fault feature map and the measured fault feature map of the rotating machinery equipment compressed to a uniform size: rotation, grayscale processing, and addition of differential data are performed in sequence; the processor expands the number of simulated fault samples and measured fault samples of the rotating machinery equipment, divides the expanded simulated fault samples and measured fault samples of the rotating machinery equipment into different categories, and constructs a simulated fault sample library and a measured fault sample library of the rotating machinery equipment.

[0096] S4. The processor establishes multiple parallel branches, and embeds a spatial-channel attention fusion convolutional neural network (CNN) model after each branch to obtain a spatial-channel attention fusion multi-branch CNN model; the processor extracts the features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery respectively; the processor introduces maximum mean difference (MMD) loss, and performs adaptive calibration and transfer learning training on the spatial-channel attention fusion multi-branch CNN model using the extracted features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery to obtain the trained spatial-channel attention fusion multi-branch CNN model;

[0097] Step S4 includes the following steps:

[0098] S401. The processor establishes multiple parallel branches, and embeds a spatial-channel attention fusion convolutional neural network (CNN) model after each branch to obtain a spatial-channel attention fusion multi-branch CNN model. Each branch includes a convolutional layer, a normalization layer, a modified linear unit (ReLU) activation layer, and a max-pooling layer. The spatial-channel attention fusion multi-branch CNN model adopts a channel attention mechanism to select key feature channels related to the fault. In the key feature channels, a spatial attention mechanism is used to focus on the key regions of the fault features, suppress background noise interference, and improve the accuracy of feature extraction. The processor uses convolutional kernels of corresponding sizes to extract the local subtle features, mesoscale features, and global contour features of each fault sample in the simulated fault sample library and the measured fault sample library of rotating machinery equipment, respectively.

[0099] To address the issues of conventional CNNs, such as lack of feature selectivity and insufficient single-scale convolutional kernels, the following network structure is designed:

[0100] Three parallel branches are employed, using small-scale convolutional kernels of 1×1, 3×3, and 5×5 respectively, to simultaneously extract local subtle features, mesoscale features, and global contour features from each simulation sample in a general fault simulation sample library for rotating machinery. Each branch contains a convolutional layer, a normalization layer, a ReLU activation layer, and a max-pooling layer, overcoming the limitations of single-scale convolutional kernels. ReLU (Rectified Linear Unit) is one of the most commonly used activation functions in deep learning, primarily used in the hidden layers of neural networks, designed to introduce nonlinear characteristics to address the vanishing gradient problem.

[0101] The CNN model is trained by transfer learning through Matlab simulation pre-training and experimental fine-tuning. Depth-separable convolution is introduced to replace some standard convolutions, realizing a lightweight design of the CNN model and reducing the number of CNN model parameters by more than 70%, which meets the real-time diagnostic needs of industrial sites.

[0102] S402. The processor introduces the maximum mean difference (MMD) loss. By extracting the local subtle features, mesoscale features, and global contour features of each fault sample in the simulated fault sample library and the measured fault sample library of rotating machinery, the spatial-channel attention fusion multi-branch CNN model is subjected to Matlab simulation pre-training and measured fine-tuning transfer learning training to obtain the trained spatial-channel attention fusion multi-branch CNN model.

[0103] S403: The processor establishes a multi-task output including classification and regression branches. The classification branch uses the Softmax activation function to output the probability distribution of various fault types in rotating machinery. The regression branch uses the Sigmoid activation function to output a quantified value (0~N, where N is a positive integer) of the severity of the fault in rotating machinery. The classification and regression branches share the underlying feature extraction network, enabling parameter reuse and improving model efficiency. Softmax is an activation function in deep learning specifically for handling multi-class tasks. It transforms a set of raw scores into a probability distribution that sums to 1, allowing the model to clearly define the probability of each category. In neural networks, it typically acts as the last layer, normalizing the output results so that users can directly select the one with the highest probability as the predicted answer.

[0104] To address the issue of sample domain offset in conventional simulations, this invention introduces Maximum Mean Discrepancy (MMD) to perform adaptive calibration and transfer learning training on simulation-measured data, thereby eliminating the offset between simulation and measured data.

[0105] Step S402 includes the following steps:

[0106] S4021: The processor uses fault samples from the simulated fault sample library and the measured fault sample library of rotating machinery to pre-train the spatial-channel attention fusion multi-branch CNN model to learn common fault features.

[0107] S4022, Processor introduces Maximum Mean Difference (MMD) loss To minimize the distribution difference between simulated and measured samples in the feature space, adaptive calibration is performed on the simulated and measured data.

[0108] MMD loss The calculation formula is: ,

[0109] in, The number of simulation samples, The number of actual samples. Let H be the feature mapping function that maps to the reproducing Hilbert space RKHS, where i is the measured sample, j is the simulated sample, s is the measured dataset, t is the simulated dataset, x is the simulated sample, and H is the reproducing kernel Hilbert space. For the i-th sample, To simulate the j-th sample;

[0110] During training, the processor performs MMD loss. The total loss is calculated by weighting and fusing the fault classification loss and the fault parameter regression loss. for: ,in, Losses due to multi-task collaboration These are the weighting coefficients for the MMD loss. The value range is: dynamically adjusted between 0.1 and 0.5 during annealing, depending on the severity of the offset;

[0111] Multi-task collaborative loss Losses classified by fault Regression loss with fault parameters Weighted composition: , The weighting coefficients for fault classification loss are... These are the weighting coefficients for the regression loss of the fault parameters;

[0112] S4023: The processor uses simulation samples based on finite element and dynamic simulation as the source domain and measured samples of rotating machinery collected through actual ship test benches or industrial field experiments as the target domain, so that the ratio of simulation samples to measured samples is maintained at 10:1.

[0113] The S4024 processor uses real-world samples to fine-tune the top layer of the spatial-channel attention fusion multi-branch CNN model to adapt it to specific industrial scenarios. The processor freezes the weight parameters of the convolutional networks in the bottom feature layers of the spatial-channel attention fusion multi-branch CNN model and only updates the gradients of the high-level networks and fully connected layers that contain the spatial-channel dual attention module CBAM. The processor uses the adaptive moment estimation optimizer Adam to improve the training stability and generalization ability of the spatial-channel attention fusion multi-branch CNN model.

[0114] The Maximum Mean Difference (MMD) loss function is introduced to minimize the distributional difference between simulated and measured samples in the feature space, enabling adaptive calibration of the simulated-measured data. MMD is a statistical tool primarily used to measure the distance between two probability distributions. In machine learning and statistics, it is used to compare the similarity between samples generated by generative models and real data. MMD quantifies the dissimilarity between two distributions by calculating the maximum difference between their means.

[0115] The regenerating kernel Hilbert space (RKHS) is a special function space in which the value of each function at any point can be directly calculated through the inner product, making function operations as convenient as vector operations. It can be imagined as a "function version of a vector space", which not only contains functions, but also specially matched "kernel functions" to realize the regeneration property.

[0116] The value range is: dynamically adjusted between 0.1 and 0.5 through annealing according to the severity of the offset. In spatial-channel attention fusion multi-branch CNN models, The initial value is set to 0.2.

[0117] Sample quantity and ratio: The simulation samples based on finite element and dynamic simulation are used as the source domain, for example, 10,000 sets of simulation samples; a small number of measured samples of rotating machinery equipment collected through actual ship test benches or industrial field experiments are used as the target domain, for example, 1,000 sets of measured samples, so that the ratio of simulation samples to measured samples is maintained at 10:1.

[0118] Fine-tuning stage: Using a small number of real-world samples, the top layer of the spatial-channel attention fusion multi-branch CNN model is fine-tuned to adapt to specific industrial scenarios. To retain the underlying basic time-frequency feature extraction capabilities learned through large-scale simulation pre-training, the weight parameters of the first three convolutional layers of the bottom feature layer of the spatial-channel attention fusion multi-branch CNN model are frozen. Gradient updates are only performed on the high-level network containing the spatial-channel dual attention module CBAM and the fully connected layers to avoid overfitting on a small number of real-world samples. Optimizer strategy: Compared with other optimizers, this embodiment of the invention preferentially selects the Adam (Adaptive Moment Estimation) optimizer to improve the training stability and generalization ability of the spatial-channel attention fusion multi-branch CNN model. The Adam optimizer is one of the most mainstream optimization algorithms in deep learning. Combining the advantages of momentum and adaptive learning rate, it automatically adjusts the learning speed for each parameter, resulting in faster, more stable, and easier convergence.

[0119] S6. The processor performs cross-validation on the fault diagnosis results of the four signals: when all four fault diagnosis results are the same, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery to be diagnosed are output in real time, providing a quantitative basis for the maintenance of the rotating machinery; otherwise, the cross-validation is confirmed to be unsuccessful, and the processor uses the weighted fusion method of evidence theory to determine which signal needs to be re-diagnosed based on the fault type, severity, and confidence level of the rotating machinery in the obtained fault diagnosis results of the four signals, and performs subsequent fault diagnosis based on the determination result; the trained spatial-channel attention fusion multi-branch CNN model completes parameter adaptive update.

[0120] In step S6, the trained spatial-channel attention fusion multi-branch CNN model completes adaptive parameter updates, including the following steps:

[0121] The processor establishes an incremental learning sample library. When new fault samples or changes in operating conditions are collected, the processor uses an incremental learning method to update the parameters of the trained spatial-channel attention fusion multi-branch CNN model without retraining, thus adapting to long-term operating condition drift. When abnormal samples with a confidence level below 0.8 are continuously collected, and after cross-validation or manual annotation, when the cumulative number of newly added difficult samples in the sample library reaches 500, a local network update is triggered, and the trained spatial-channel attention fusion multi-branch CNN model completes online parameter adaptive updates.

[0122] The spatial-channel attention fusion multi-branch CNN model performs online adaptive parameter updates, including the following steps:

[0123] Most parameters of the backbone network (multi-branch convolutional and spatial-channel dual attention module CBAM) are frozen, and only a small number of key parameters are updated online to adapt it to the new sample distribution.

[0124] Statistical analysis of prior experimental datasets revealed that when the Softmax classification output probability is greater than 0.8, the false alarm rate of the spatial-channel attention fusion multi-branch CNN model is less than 1%. In existing technologies, when the confidence level is between 0.5 and 0.8, it is often due to early, weak, complex faults or extreme variable load conditions, which makes it very easy to misjudge.

[0125] Figure 1 This is a flowchart of the rotating machinery diagnostic method based on adaptive wavelet and CNN in an embodiment of the present invention. See also Figure 1 As shown, from data acquisition to online diagnosis: a multi-condition finite element simulation model of rotating machinery is constructed, and the simulated and measured torque and vibration signals of the rotating machinery are collected respectively. The four signals are processed to construct fault samples of rotating machinery. A spatial-channel attention fusion multi-branch CNN model is constructed and trained. The simulation-measured data are adaptively calibrated to diagnose the faults of the four signals of the rotating machinery to be diagnosed. The four diagnostic results are cross-validated.

[0126] Step 1: Establish a multi-condition finite element simulation model of rotating machinery. Collect simulation and measured torque and vibration signals of the rotating machinery to support the solution of scarce measured fault data. Generate gradient samples covering the range of operating conditions and different fault severity levels at low cost through simulation, providing large-scale basic data for subsequent model training.

[0127] The second step involves performing adaptive multi-scale DB transformation and improved threshold denoising on the simulated and measured torque and vibration signals, respectively, to generate a fused fault detail map of the rotating machinery. This innovation at the signal processing level solves the problems of poor adaptability of traditional fixed DB wavelet basis, loss of weak fault features at a single scale, and constant deviation of traditional threshold, demonstrating how to accurately extract fault mutation features at different levels.

[0128] The third step involves standardizing the fused fault detail diagrams of rotating machinery and performing Mixup data augmentation to build a fault simulation sample library for rotating machinery. This library addresses the technical issues of insufficient composite fault samples and model overfitting. Artificial composite fault samples are generated using the Mixup algorithm, while standardization unifies the model input format, reducing computational load and improving real-time diagnostics. Mixup is a data augmentation method that generates virtual training samples through linear interpolation.

[0129] Step 4: Using multiple parallel branches, extract the features of each sample in the general fault simulation sample library of rotating machinery equipment, construct a spatial-channel attention fusion multi-branch convolutional neural network (CNN) model, and perform transfer learning training on the CNN model through Matlab simulation pre-training and experimental fine-tuning. This innovation at the fault diagnosis model level solves the problems of low attention to key fault features and single-scale convolutional kernel limitations of conventional CNNs. At the same time, the transfer learning strategy fully utilizes the value of simulation data and reduces the dependence on experimental data.

[0130] Step 5: Introduce the maximum mean difference to adaptively calibrate the simulation-measured data, eliminate the offset between simulation and measured data, solve the industry pain point of significant domain offset between simulation samples and measured data, and achieve effective transfer from simulation to actual industrial scenarios by minimizing the distribution difference between the two domains in the feature space, greatly improving the model's generalization ability.

[0131] Step 6: Preprocess the simulated and measured torque and vibration signals of the rotating machinery under the same operating condition. Input the preprocessed simulated and measured torque and vibration signals into the trained spatial-channel attention fusion multi-branch CNN model to perform fault diagnosis on these four signals respectively, and cross-validate the fault diagnosis results of the four signals: When the fault diagnosis results of the four signals are consistent, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery under diagnosis are output in real time, providing a quantitative basis for the maintenance of rotating machinery. The spatial-channel attention fusion multi-branch CNN model completes the parameter adaptive update; When the diagnosis results of the four signals are inconsistent, the cross-validation is confirmed to be unsuccessful. Based on the fault type, severity, and confidence level of the rotating machinery in the diagnosis results of the four signals, the weighted fusion method of evidence theory is used to determine which signal needs to be re-diagnosed.

[0132] This invention establishes a cross-verification mechanism for simulated and measured torque and vibration signals under the same working condition: the judgment is made based on the simulated and measured torque signals, but the result cannot be directly output. The simulated and measured vibration signals under the same working condition are required. If the diagnostic results of the simulated and measured torque and vibration signals under the same working condition are the same, then the diagnostic result can be output.

[0133] The fundamental frequency and harmonic components of the torque signal are synchronized with the rotational speed and load torque, providing a direct and sensitive response to periodic load anomalies, torque pulsations, and shaft torque fluctuations in rotating machinery. Vibration signals exhibit high sensitivity to localized impact faults, capturing high-frequency transient impact characteristics that are difficult for torque signals to reflect. This invention leverages the complementary advantages of torque and vibration signals, improving the accuracy and robustness of fault diagnosis through cross-validation of the diagnostic results from simulated and measured torque and vibration signals. The spatial-channel dual-attention fusion multi-branch CNN model has 70% fewer parameters than conventional CNN models, resulting in shorter single-sample diagnosis time. The cross-validation mechanism of the diagnostic results from simulated and measured torque and vibration signals lowers the false alarm rate, addressing the issue of insufficient reliability of a single torque signal. This invention overcomes the limitations of traditional methods that can only identify fault types, achieving for the first time parallel output of qualitative and quantitative fault diagnosis for rotating machinery, providing a quantitative basis for the maintenance of rotating machinery.

[0134] Figure 2 This is a schematic diagram of the finite element simulation model of the core transmission component in a rotating mechanical device under normal conditions. (See attached image) Figure 2 As shown, the core transmission components of rotating machinery, taking an electric motor as an example, include: stator 1, stator guide bars 2, rotor 3, and rotor guide bars 4. Figure 2 It demonstrates the basic physical structure of the core transmission components of rotating machinery, providing a foundation for the subsequent explanation of fault types and mechanisms.

[0135] Figure 3 This is a schematic diagram of a finite element simulation model of a rotating mechanical device in the rotor bar breakage fault state according to an embodiment of the present invention. See [link / reference]. Figure 3 As shown, the broken conductor bars on the rotor structure of the motor are filled with black material, i.e. Figure 3 The broken rotor bar 5 in the image visually demonstrates the physical location and morphology of the broken rotor bar fault, explains the generation mechanism of the broken rotor bar fault, and the physical source of its corresponding periodic impact signal characteristics, establishing the correspondence between physical faults and signal characteristics.

[0136] Figure 4 This is a schematic diagram of the finite element simulation model of the rotating mechanical equipment under air gap eccentricity fault state in an embodiment of the present invention. See [link / reference]. Figure 4 As shown, Figure 4 The text indicates the uneven air gap distances between rotor 3 and stator 1, specifically D1 and D2, demonstrating the state where the center of rotor 3 does not coincide with the center of stator 1. This illustrates the generation mechanism of air gap eccentricity fault, the physical source of its corresponding low-frequency modulation signal characteristics, and the signal strength differences corresponding to different degrees of eccentricity.

[0137] Figure 5This is a schematic diagram of a finite element simulation model of a rotating mechanical device in an embodiment of the present invention, showing a combined fault state with both rotor bar breakage and air gap eccentricity. (See attached diagram.) Figure 5 As shown, Figure 5 The image also marks the broken rotor bar 5 and the uneven air gap distances D1 and D2, demonstrating the physical state of the coexistence of rotor bar breakage and air gap eccentricity faults. This illustrates the more complex physical reasons for the compound fault characteristics, and its signal characteristics are the superposition of the two single fault characteristics.

[0138] Figure 6 This is a structural diagram of the channel attention fusion CNN model in an embodiment of the present invention. Figure 6 It illustrates the connections, parameter settings, and core functions of each network layer. The CNN structure diagram clearly presents the basic feature transfer path from input to output, providing a structural foundation for understanding the spatial-channel attention module and the embedding of multi-task branches, and supporting the explanation of lightweight model design.

[0139] Figure 7 This is an overall architecture diagram of the multi-task CNN model in an embodiment of the present invention. Figure 7 This paper demonstrates a multi-task architecture that combines shared convolutional layers and task-specific fully connected layers, including a classification branch (outputting the probability of fault type) and a regression branch (outputting the quantification value of fault severity). It illustrates the implementation mechanism of multi-task diagnosis, which achieves parameter reuse by sharing the underlying feature extraction network, thereby reducing the number of model parameters while ensuring diagnostic accuracy.

[0140] Figure 8 This is a characteristic diagram of the normal state of rotating mechanical equipment in an embodiment of the present invention. Figure 8 The diagram shows the curves of the changes of four different scale detail components over time after the signal is decomposed into four layers of DB wavelets during normal operation of the motor. Different colors correspond to different decomposition layers. Figure 8 As a benchmark reference diagram, it provides the standard signal characteristic shape under fault-free conditions, providing a quantitative basis for subsequent comparative analysis of various fault characteristics.

[0141] Figure 9 This is a characteristic diagram of rotor bar breakage fault in a rotating mechanical device according to an embodiment of the present invention. Figure 9 The diagram shows the characteristic curves of four scale detail components under a fault condition where one rotor bar is broken. Figure 9 The specific signal characteristics of rotor bar breakage faults are visualized, verifying that the embodiments of the present invention can effectively extract the weak abrupt change characteristics of this type of fault, and provide distinguishable feature basis for model classification.

[0142] Figure 10 This is a fault feature diagram of air gap eccentricity in a rotating mechanical device according to an embodiment of the present invention. Figure 10The characteristic curves of four scale detail components are shown under a fault condition with 20% air gap eccentricity. Figure 10 The low-frequency modulation characteristics visually demonstrate the significant difference in the morphological features between air gap eccentricity faults and bar breakage faults, indicating that the embodiments of the present invention can effectively distinguish between different types of single faults.

[0143] Figure 11 This is a composite fault feature diagram of a rotating mechanical device in an embodiment of the present invention, where rotor bar breakage and air gap eccentricity coexist. Figure 11 The figure shows the characteristic curves of four scale detail components under a combined fault state where rotor bar breakage and air gap eccentricity coexist. Figure 11 The invention presents a superposition of two single fault characteristics. Taking a single rotor bar breakage fault and a 20% air gap eccentricity fault as examples, the invention illustrates the compound fault and verifies that the embodiments of the present invention can effectively extract the superposition characteristics of complex compound faults, thus solving the core problem of low recognition rate of compound faults in traditional methods.

[0144] Figure 12 This is a comparison chart of learning rates for different optimizers in an embodiment of the present invention. Figure 12 This paper demonstrates the accuracy and loss function changes of three optimizers—Adam, Rmsprop, and Sdgm—under different learning rates. Using the same dataset, the training accuracy curves and loss function decline curves of the three optimizers are compared over 500 iterations. Quantitative experimental data validates that the Adam optimizer has the fastest convergence speed and the highest final accuracy in this model, providing objective experimental evidence for the selection of hyperparameters during model training and supporting the credibility of the technical effects.

[0145] Figure 13 This is a flowchart of the graph preprocessing for fault feature details in an embodiment of the present invention. Figure 13 It demonstrates standardized preprocessing steps for cropping, compression, and data augmentation. Figure 13 The comparison between the original detail image size (632×528×3) and the preprocessed standard size (224×224×3), as well as the effective detail component image after cropping and removing blank areas, illustrates that the core operation of data preprocessing is to remove irrelevant background data and unify the input size, providing the model with standardized high-quality input, while reducing the amount of computation and improving the model training efficiency and inference speed.

[0146] Figure 14 This is a flowchart illustrating the selection of the adaptive DB wavelet basis in an embodiment of the present invention. Figure 14This demonstrates a closed-loop process that starts with the input signal, calculates features such as signal energy and mutual information entropy, determines whether a preset threshold is met, and dynamically adjusts the wavelet basis order (DB1~DB10) if not, ultimately outputting the optimal wavelet basis. The innovative feature of supporting adaptive multi-scale DB wavelet transform addresses the problem that a fixed wavelet basis cannot adapt to different fault types and operating conditions, significantly improving the robustness of signal processing and the accuracy of feature extraction.

[0147] Figure 15 This is a structural diagram of the spatial-channel dual attention module in an embodiment of the present invention. Figure 15 The internal structure of the parallel spatial attention submodule and channel attention submodule is shown, including: feature extraction layer, weight generation layer, and final feature fusion process. The core enhancement mechanism of the embodiment of the present invention is explained in detail: filtering fault-related channels through channel attention, locating key fault areas through spatial attention, suppressing background noise interference in all aspects, and improving the feature extraction accuracy of complex composite faults.

[0148] Figure 16 This is a schematic diagram of the simulation-measurement domain adaptive calibration principle in an embodiment of the present invention. Figure 16 It demonstrates the interaction between the simulation domain (simulation model, simulation data) and the measured domain (measured system, measured data), as well as the closed loop of calibration error feedback driven by MMD loss. It clearly presents the implementation mechanism of domain adaptive calibration, which minimizes the distribution difference between the two domains by dynamically adjusting the MMD loss weight, and achieves stable diagnosis of the model under varying operating conditions.

[0149] The innovative aspects of this invention are as follows:

[0150] I. Innovative Signal Acquisition and Diagnosis Architecture: A pioneering cross-validation system for the diagnostic results of simulated and measured torque and vibration signals of rotating machinery under the same operating conditions.

[0151] 1. The fundamental frequency and harmonic components of the torque signal are synchronized with the rotational speed and load torque, providing a direct and sensitive response to periodic load anomalies, torque pulsations, and shaft torque fluctuations in rotating machinery. Vibration signals exhibit high sensitivity to localized impact faults, capturing high-frequency transient impact characteristics that are difficult for torque signals to reflect. This invention leverages the complementary advantages of simulated and measured torque and vibration signals, improving the accuracy and robustness of fault diagnosis through cross-validation of four diagnostic results. The spatial-channel dual-attention fusion multi-branch CNN model has 70% fewer parameters than conventional CNN models, resulting in shorter single-sample diagnosis time. The torque-vibration signal cross-validation mechanism lowers the false alarm rate, addressing the problem of poor anti-interference capability of a single vibration signal at the signal source. This invention overcomes the limitations of traditional methods that can only identify fault types, achieving for the first time parallel output of qualitative and quantitative fault diagnosis for rotating machinery, providing a quantitative basis for the maintenance of rotating machinery.

[0152] 2. Specify that the torque sensor is installed in series at the motor output shaft-load coupling, and the vibration sensor is attached to the radial surface of the bearing housing at the motor drive end / non-drive end, with a uniform sampling frequency of 10kHz.

[0153] II. Innovations in Signal Preprocessing:

[0154] 1. Calculate the mutual information entropy between the DB1~DB10 wavelet basis and the simulated and measured torque and vibration signals, and automatically select the wavelet basis with the highest matching degree, thus solving the problem of poor adaptability of fixed wavelet basis to different fault types.

[0155] 2. An adaptive adjustment factor that dynamically adjusts with the number of wavelet decomposition layers is proposed, overcoming the shortcomings of traditional hard threshold discontinuity and soft threshold constant deviation.

[0156] 3. The details coefficients of each layer and the Pearson correlation coefficient of the fault label are used as weights for fusion. The features of the fused four signals, namely the simulated and measured torque signals and vibration signals, are then mapped to RGB three channels respectively as multi-channel fused fault detail maps. This achieves the standardized conversion from one-dimensional time-frequency features to two-dimensional image texture features, adapting to CNN input requirements.

[0157] III. Network Structure Innovation:

[0158] After convolution, a spatial-channel dual attention module is embedded to first filter key feature channels and then focus on the critical fault region.

[0159] IV. Training Strategy Innovation:

[0160] 1. A three-level training process consisting of simulation pre-training, MMD domain calibration, and freezing the underlying fine-tuning;

[0161] 2. Clear sample matching and parameter rules.

[0162] V. Engineering Implementation Innovation: A quantitative labeling system for fault severity; clearly defining 0 as corresponding to a fully normal operating state, with other labels corresponding to different states.

[0163] See Figure 17 As shown, this embodiment of the invention also provides a fault diagnosis system for rotating machinery based on adaptive wavelet and CNN for performing the above method, comprising:

[0164] A torque sensor is installed in series between the output shaft of the drive motor of a rotating machinery and the load input end to collect the measured torque signal of the rotating machinery.

[0165] Vibration sensors are patched and mounted on the radial surface of the bearing housing housing at both the drive end and non-drive end of rotating machinery to collect measured vibration signals from the rotating machinery.

[0166] Processor, including:

[0167] The measured signal mapping relationship establishment module is used to: establish the mapping relationship between the measured torque signal and the measured vibration signal of rotating machinery;

[0168] The finite element simulation model construction module is used to: construct several multi-condition finite element simulation models based on the factory parameters and fault types of rotating machinery using electromagnetic field finite element simulation software. These models simulate the operating condition fluctuations of rotating machinery in different industrial scenarios, covering the normal state, single fault state, and compound fault state of rotating machinery. At least three multi-condition finite element simulation models are constructed for each state. Each multi-condition finite element simulation model simulates at least one simulated torque sensor and at least one simulated vibration sensor to collect the simulated torque signal and simulated vibration signal of rotating machinery under multiple conditions, and to establish the mapping relationship between the simulated torque signal and simulated vibration signal of rotating machinery.

[0169] The multi-scale fusion fault detail map generation module is used to: perform adaptive multi-scale multi-Bessie DB wavelet transform and improved threshold denoising on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of rotating machinery, respectively, to generate simulated multi-scale fusion fault detail maps and measured multi-scale fusion fault detail maps of rotating machinery.

[0170] The module for constructing simulation and measured fault sample libraries is used to: preprocess and perform data mixing and enhancement Mixup on the simulation multi-scale fusion fault detail map and the measured multi-scale fusion fault detail map of rotating machinery respectively, and construct the simulation fault sample library and the measured fault sample library of rotating machinery respectively.

[0171] The spatial-channel attention fusion multi-branch CNN model construction module is used to: establish multiple parallel branches, embed a spatial-channel attention fusion convolutional neural network (CNN) model after each branch, and obtain a spatial-channel attention fusion multi-branch CNN model; each branch includes convolutional layers, normalization layers, modified linear unit (ReLU) activation layers, and max pooling layers; the spatial-channel attention fusion multi-branch CNN model adopts a channel attention mechanism to filter out key feature channels related to faults, and uses a spatial attention mechanism in the key feature channels to focus on the key regions of fault features;

[0172] The adaptive calibration and transfer learning training module is used to: extract the features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery respectively; introduce the maximum mean difference (MMD) loss, and use the extracted features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery to perform adaptive calibration and transfer learning training on the spatial-channel attention fusion multi-branch CNN model to obtain the trained spatial-channel attention fusion multi-branch CNN model;

[0173] The fault diagnosis and adaptive update module is used to: preprocess the simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal of the rotating machinery under the same operating condition; input the preprocessed simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal into a trained spatial-channel attention fusion multi-branch CNN model; the trained spatial-channel attention fusion multi-branch CNN model performs fault diagnosis on the input simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal respectively, obtaining fault diagnosis results for the four signals; and analyze the fault diagnosis results for the four signals. Cross-validation of the fault diagnosis results: When the fault diagnosis results of the four signals are all the same, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery to be diagnosed are output in real time, providing a quantitative basis for the maintenance of the rotating machinery; otherwise, the cross-validation is confirmed to be unsuccessful, and the processor, based on the fault diagnosis results of the four signals, uses a weighted fusion method of evidence theory to determine which signal needs to be re-diagnosed, and performs subsequent fault diagnosis based on the determination result; the trained spatial-channel attention fusion multi-branch CNN model completes adaptive parameter updates;

[0174] The torque sensor and vibration sensor are connected to the measured signal mapping relationship module to establish a data flow. The measured signal mapping relationship module and the finite element simulation model construction module are connected to the multi-scale fusion fault detail map generation module to establish a data flow. The multi-scale fusion fault detail map generation module, the simulation and measured fault sample library construction module, the spatial-channel attention fusion multi-branch CNN model construction module, the adaptive calibration and transfer learning training module, and the fault diagnosis and adaptive update module are connected to each other in sequence to establish a data flow.

[0175] This invention also provides an electronic device, including: at least one processor, at least one memory, and a communication interface, wherein the processor, memory, and communication interface communicate with each other; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the method in this invention embodiment.

[0176] This invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to execute the methods described in this invention.

[0177] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for rotating machinery based on adaptive wavelet and CNN, characterized in that, Includes the following steps: S1. The processor uses electromagnetic field finite element simulation software to construct several multi-condition finite element simulation models based on the factory parameters and fault types of rotating machinery. These models simulate the operating condition fluctuations of rotating machinery in different industrial scenarios, covering the normal state, single fault state, and compound fault state of rotating machinery. At least three multi-condition finite element simulation models are constructed for each state. Each multi-condition finite element simulation model simulates at least one simulated torque sensor and at least one simulated vibration sensor to collect the simulated torque signal and simulated vibration signal of rotating machinery under multiple conditions, and establishes the mapping relationship between the simulated torque signal and simulated vibration signal of rotating machinery. A torque sensor is connected in series between the output shaft of the drive motor and the load input end of the rotating machinery to collect the measured torque signal of the rotating machinery; a vibration sensor patch is installed on the radial surface of the bearing housing at the drive end and non-drive end of the rotating machinery to collect the measured vibration signal of the rotating machinery; and the processor establishes the mapping relationship between the measured torque signal and the measured vibration signal of the rotating machinery. S2. The processor performs adaptive multi-scale multi-Bessie DB wavelet transform and improved threshold denoising on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of the rotating machinery, respectively, to generate simulated multi-scale fused fault detail map and measured multi-scale fused fault detail map of the rotating machinery. S3 and the processor perform preprocessing and data mixing enhancement Mixup on the simulated multi-scale fused fault detail map and the measured multi-scale fused fault detail map of the rotating machinery respectively, and construct the simulated fault sample library and the measured fault sample library of the rotating machinery respectively. S4. The processor establishes multiple parallel branches, and embeds a spatial-channel attention fusion convolutional neural network (CNN) model after each branch to obtain a spatial-channel attention fusion multi-branch CNN model; the processor extracts the features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery respectively; the processor introduces maximum mean difference (MMD) loss, and performs adaptive calibration and transfer learning training on the spatial-channel attention fusion multi-branch CNN model using the extracted features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery to obtain the trained spatial-channel attention fusion multi-branch CNN model; S5. The processor preprocesses the simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal of the rotating machinery under the same working condition. The preprocessed simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal are then input into the trained spatial-channel attention fusion multi-branch CNN model. The trained spatial-channel attention fusion multi-branch CNN model performs fault diagnosis on the input simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal, respectively, and obtains the fault diagnosis results of the four signals. S6. The processor performs cross-validation on the fault diagnosis results of the four signals: when all four fault diagnosis results are the same, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery to be diagnosed are output in real time, providing a quantitative basis for the maintenance of the rotating machinery; otherwise, the cross-validation is confirmed to be unsuccessful, and the processor uses the weighted fusion method of evidence theory to determine which signal needs to be re-diagnosed based on the fault type, severity, and confidence level of the rotating machinery in the obtained fault diagnosis results of the four signals, and performs subsequent fault diagnosis based on the determination result; the trained spatial-channel attention fusion multi-branch CNN model completes parameter adaptive update.

2. The method for fault diagnosis of rotating machinery based on adaptive wavelet and CNN as described in claim 1, characterized in that, In step S1, the processor uses electromagnetic field finite element simulation software to construct several multi-condition finite element simulation models based on the factory parameters and fault types of the rotating machinery equipment, including the following steps: The processor employs electromagnetic field finite element simulation software. Based on the rated speed ±30%, the light and heavy load ranges of the rated load, and the rated power, rated voltage, and structural dimensions of the transmission components of the rotating machinery, several multi-condition finite element simulation models of different fault severity are constructed for rotor bar breakage and air gap eccentricity fault types. These models cover the normal state, single fault state, and compound fault state of the rotating machinery, simulating the operating condition fluctuations of rotating machinery in actual industrial scenarios and generating gradient fault samples of different types of rotating machinery. In a transient field simulation environment, simulation time is set, and independent simulations are performed on different types of gradient fault samples of the rotating machinery. Simulation torque signals and simulation vibration signals of each multi-condition finite element simulation model are collected to obtain the original datasets of simulation torque signals and simulation vibration signals of the rotating machinery in normal state and various fault states, which are used to characterize the simulation dynamic features of the rotating machinery.

3. The method for fault diagnosis of rotating machinery based on adaptive wavelet and CNN as described in claim 2, characterized in that, Step S2 includes the following steps: S201 The processor adopts the adaptive Dobessi (DB) wavelet basis selection algorithm to calculate the mutual information entropy between different wavelet bases and the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of rotating machinery, and automatically selects the wavelet base with the largest mutual information entropy value as the optimal base for this calculation. S202. Based on the optimal basis of this calculation, the processor sequentially uses wavelet low-pass filter and wavelet high-pass filter to perform multi-level DB wavelet decomposition on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of the rotating machinery equipment, respectively, to obtain the multi-level detail coefficients and approximation coefficients of the four signals, and capture the fault characteristics of the rotating machinery equipment at different levels. The adaptive exponential threshold function used by the processor when performing multi-level DB wavelet decomposition is: ， in, Let be the adaptive exponential threshold corresponding to the k-th level wavelet decomposition, where k is the number of wavelet decomposition levels, k = 1, 2, 3, 4, ..., 10. Let be the noise standard deviation of the detail coefficients at the k-th layer, and ln be a logarithmic function with the natural constant e as the base. This represents the total number of sampling points for the original signal. The wavelet detail coefficients are obtained from the k-th level wavelet decomposition. As an adaptive adjustment factor, , The coefficient for smooth attenuation is denoted by e, where e is the natural constant. The processor calculates the Pearson coefficients, which are related to the fault type of rotating machinery, for each level of the four signals and uses them as the weight coefficients for that level. Then, the multi-level detail coefficients of the four signals are weighted and fused to obtain the fused rotating machinery fault characteristic signal. S203 The processor converts the fused four signals of the rotating machinery fault feature signal into red, green and blue RGB three-channel images, which serve as the simulated multi-scale fused fault detail image and the measured multi-scale fused fault detail image of the rotating machinery.

4. The fault diagnosis method for rotating machinery based on adaptive wavelet and CNN as described in claim 3, characterized in that, Step S3 includes the following steps: S301 and the processor perform adaptive histogram equalization on the simulated multi-scale fusion fault detail map and the measured multi-scale fusion fault detail map of the rotating machinery equipment, respectively, to enhance the contrast of the fault change region of the rotating machinery equipment, highlight the edge features, and obtain the simulated feature map and measured feature map of the rotating machinery equipment with different fault types and different fault severity. S302 and the processor linearly blend the simulation feature maps and measured feature maps of different fault types and severity of rotating machinery equipment to generate simulation feature maps and measured feature maps of composite fault samples of rotating machinery equipment; and respectively trim the blank areas in the simulation feature maps and measured feature maps of composite fault samples of rotating machinery equipment, retaining only the areas containing fault features, to obtain the trimmed simulation fault feature maps and measured fault feature maps of rotating machinery equipment. S303 The processor preprocesses the cropped simulated fault feature map and measured fault feature map of the rotating machinery, and compresses the preprocessed simulated fault feature map and measured fault feature map of the rotating machinery into a uniform size while maintaining the same channel color information. This uniform size is a generalized image size that is suitable for the recognition of simulated fault feature maps and measured fault feature maps of various rotating machinery. S304. The processor performs random data augmentation Mixup on the simulated fault feature map and the measured fault feature map of the rotating machinery equipment compressed to a uniform size: rotation, grayscale processing, and addition of differential data are performed in sequence; the processor expands the number of simulated fault samples and measured fault samples of the rotating machinery equipment, divides the expanded simulated fault samples and measured fault samples of the rotating machinery equipment into different categories, and constructs a simulated fault sample library and a measured fault sample library of the rotating machinery equipment.

5. The fault diagnosis method for rotating machinery based on adaptive wavelet and CNN as described in claim 4, characterized in that, Step S4 includes the following steps: S401. The processor establishes multiple parallel branches, and embeds a spatial-channel attention fusion convolutional neural network (CNN) model after each branch to obtain a spatial-channel attention fusion multi-branch CNN model. Each branch includes a convolutional layer, a normalization layer, a modified linear unit (ReLU) activation layer, and a max-pooling layer. The spatial-channel attention fusion multi-branch CNN model adopts a channel attention mechanism to filter out key feature channels related to the fault, and uses a spatial attention mechanism in the key feature channels to focus on the key regions of the fault features. The processor uses convolutional kernels of corresponding sizes to extract the local subtle features, mesoscale features, and global contour features of each fault sample in the simulated fault sample library and the measured fault sample library of rotating machinery. S402. The processor introduces the maximum mean difference (MMD) loss. By extracting the local subtle features, mesoscale features, and global contour features of each fault sample in the simulated fault sample library and the measured fault sample library of rotating machinery, the spatial-channel attention fusion multi-branch CNN model is subjected to Matlab simulation pre-training and measured fine-tuning transfer learning training to obtain the trained spatial-channel attention fusion multi-branch CNN model. S403 The processor establishes a multi-task output including classification and regression branches. The classification branch uses the softmax function to output the probability distribution of various fault types; the regression branch uses the sigmoid activation function to output the quantified value of the fault severity; the classification and regression branches share the underlying feature extraction network to achieve parameter reuse.

6. The fault diagnosis method for rotating machinery based on adaptive wavelet and CNN as described in claim 5, characterized in that, Step S402 includes the following steps: S4021: The processor uses fault samples from the simulated fault sample library and the measured fault sample library of rotating machinery to pre-train the spatial-channel attention fusion multi-branch CNN model to learn general fault features. S4022, Processor introduces Maximum Mean Difference (MMD) loss To minimize the distribution difference between simulated and measured samples in the feature space, adaptive calibration is performed on the simulated and measured data. MMD loss The calculation formula is: , in, The number of simulation samples, This represents the number of measured samples. Let H be the feature mapping function that maps to the reproducing Hilbert space RKHS, where i is the measured sample, j is the simulated sample, s is the measured dataset, t is the simulated dataset, x is the simulated sample, and H is the reproducing kernel Hilbert space. For the i-th sample, To simulate the j-th sample; During training, the processor performs MMD loss. The total loss is calculated by weighting and fusing the fault classification loss and the fault parameter regression loss. for: ,in, Losses due to multi-task collaboration These are the weighting coefficients for the MMD loss. The value range is: dynamically adjusted between 0.1 and 0.5 during annealing, depending on the severity of the offset; Multi-task collaborative loss Losses classified by fault Regression loss with fault parameters Weighted composition: , The weighting coefficients for fault classification loss are... These are the weighting coefficients for the regression loss of the fault parameters; S4023: The processor uses simulation samples based on finite element and dynamic simulation as the source domain and measured samples of rotating machinery collected through actual ship test benches or industrial field experiments as the target domain, so that the ratio of simulation samples to measured samples is maintained at 10:

1. The S4024 processor uses real-world samples to fine-tune the top layer of the spatial-channel attention fusion multi-branch CNN model to adapt it to specific industrial scenarios. The processor freezes the weight parameters of the convolutional networks in the bottom feature layers of the spatial-channel attention fusion multi-branch CNN model and only updates the gradients of the high-level networks and fully connected layers that contain the spatial-channel dual attention module CBAM. The processor uses the adaptive moment estimation optimizer Adam to improve the training stability and generalization ability of the spatial-channel attention fusion multi-branch CNN model.

7. The fault diagnosis method for rotating machinery based on adaptive wavelet and CNN as described in claim 6, characterized in that, In step S6, the trained spatial-channel attention fusion multi-branch CNN model completes adaptive parameter updates, including the following steps: The processor establishes an incremental learning sample library. When new fault samples or changes in operating conditions are collected, the processor uses an incremental learning method to update the parameters of the trained spatial-channel attention fusion multi-branch CNN model. After continuously collecting abnormal samples with a confidence level below 0.8, and confirming them through cross-validation or manual annotation, when the cumulative number of newly added difficult samples in the sample library reaches 500, a local network update is triggered, and the trained spatial-channel attention fusion multi-branch CNN model completes online parameter adaptive updates.

8. A fault diagnosis system for rotating machinery based on adaptive wavelet and CNN for performing the method of claim 1, characterized in that, include: A torque sensor is installed in series between the output shaft of the drive motor of a rotating machinery and the load input end to collect the measured torque signal of the rotating machinery. Vibration sensors are patched and mounted on the radial surface of the bearing housing housing at both the drive end and non-drive end of rotating machinery to collect measured vibration signals from the rotating machinery. Processor, including: The measured signal mapping relationship establishment module is used to: establish the mapping relationship between the measured torque signal and the measured vibration signal of rotating machinery; The finite element simulation model construction module is used to: construct several multi-condition finite element simulation models based on the factory parameters and fault types of rotating machinery using electromagnetic field finite element simulation software. These models simulate the operating condition fluctuations of rotating machinery in different industrial scenarios, covering the normal state, single fault state, and compound fault state of rotating machinery. At least three multi-condition finite element simulation models are constructed for each state. Each multi-condition finite element simulation model simulates at least one simulated torque sensor and at least one simulated vibration sensor to collect the simulated torque signal and simulated vibration signal of rotating machinery under multiple conditions, and to establish the mapping relationship between the simulated torque signal and simulated vibration signal of rotating machinery. The multi-scale fusion fault detail map generation module is used to: perform adaptive multi-scale multi-Bessie DB wavelet transform and improved threshold denoising on the simulated torque signal, simulated vibration signal, measured torque signal and measured vibration signal of rotating machinery, respectively, to generate simulated multi-scale fusion fault detail maps and measured multi-scale fusion fault detail maps of rotating machinery. The module for constructing simulation and measured fault sample libraries is used to: preprocess and perform data mixing and enhancement Mixup on the simulation multi-scale fusion fault detail map and the measured multi-scale fusion fault detail map of rotating machinery respectively, and construct the simulation fault sample library and the measured fault sample library of rotating machinery respectively. The spatial-channel attention fusion multi-branch CNN model construction module is used to: establish multiple parallel branches, embed a spatial-channel attention fusion convolutional neural network (CNN) model after each branch, and obtain a spatial-channel attention fusion multi-branch CNN model; each branch includes convolutional layers, normalization layers, modified linear unit (ReLU) activation layers, and max pooling layers; the spatial-channel attention fusion multi-branch CNN model adopts a channel attention mechanism to filter out key feature channels related to faults, and uses a spatial attention mechanism in the key feature channels to focus on the key regions of fault features; The adaptive calibration and transfer learning training module is used to: extract the features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery respectively; introduce the maximum mean difference (MMD) loss, and use the extracted features of each fault sample in the simulation fault sample library and the measured fault sample library of rotating machinery to perform adaptive calibration and transfer learning training on the spatial-channel attention fusion multi-branch CNN model to obtain the trained spatial-channel attention fusion multi-branch CNN model; The fault diagnosis and adaptive update module is used to: preprocess the simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal of the rotating machinery under the same operating condition; input the preprocessed simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal into a trained spatial-channel attention fusion multi-branch CNN model; the trained spatial-channel attention fusion multi-branch CNN model performs fault diagnosis on the input simulated torque signal, simulated vibration signal, measured torque signal, and measured vibration signal respectively, obtaining fault diagnosis results for the four signals; and analyze the fault diagnosis results for the four signals. Cross-validation of the fault diagnosis results: When the fault diagnosis results of the four signals are all the same, the cross-validation is confirmed to be successful, and the fault type, severity, and confidence level of the rotating machinery to be diagnosed are output in real time, providing a quantitative basis for the maintenance of the rotating machinery; otherwise, the cross-validation is confirmed to be unsuccessful, and the processor, based on the fault diagnosis results of the four signals, uses a weighted fusion method of evidence theory to determine which signal needs to be re-diagnosed, and performs subsequent fault diagnosis based on the determination result; the trained spatial-channel attention fusion multi-branch CNN model completes adaptive parameter updates; The torque sensor and vibration sensor are connected to the measured signal mapping relationship module to establish a data flow. The measured signal mapping relationship module and the finite element simulation model construction module are connected to the multi-scale fusion fault detail map generation module to establish a data flow. The multi-scale fusion fault detail map generation module, the simulation and measured fault sample library construction module, the spatial-channel attention fusion multi-branch CNN model construction module, the adaptive calibration and transfer learning training module, and the fault diagnosis and adaptive update module are connected to each other in sequence to establish a data flow.

9. An electronic device, characterized in that, include: At least one processor, at least one memory, and a communication interface, wherein the processor, memory, and communication interface communicate with each other; The memory stores program instructions that can be executed by a processor, which invokes the program instructions to perform the method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the method according to any one of claims 1 to 7.

Images