Revolute pair assembly double-domain migration health diagnosis method based on sensitivity cross-modal perception

Abstract

The invention discloses a rotation pair assembly double-domain migration health diagnosis method based on sensitivity cross-modal perception. Comprising the following steps: firstly, extracting features such as a time domain, a frequency domain and variational mode decomposition from a vibration signal and a voiceprint signal of a rotation pair assembly to construct a cross-mode feature set; by considering the modal drift and sensitivity index of the voiceprint signal, the acoustic signal and the mechanical state signal are fused, and the accuracy and robustness of the diagnosis model are further optimized. And then, deep feature learning is carried out by adopting a space-time coupling health characterization network, and hyper-parameters of a health diagnosis model are jointly optimized through an alpha evolutionary algorithm in combination with a sensitivity cross-modal perception mechanism. And finally, performing cross-modal sensing on the working state data of the rotating pair assembly part to be detected through the health diagnosis model to obtain a fault diagnosis result. According to the method, the diagnosis accuracy and robustness under the conditions of small samples and variable working conditions can be improved, and the diagnosis capability of the model is further enhanced after the voiceprint signals are fused.

Description

Technical Field

[0001] This invention belongs to the field of mechanical health diagnosis technology, and in particular relates to a dual-domain migration health diagnosis method for rotating pair assemblies based on sensitivity cross-modal sensing. Background Technology

[0002] The increasing scale and complexity of machinery in today's manufacturing industry necessitates larger, faster, and more precise equipment, such as large high-speed moving beam gantry machining centers, heavy-duty AC swing-beam five-axis gantry machining centers, five-axis high-speed gantry milling centers, hydrogen compressors, and metal pipe production lines. Traditional intelligent fault diagnosis (IFD) methods based on shallow machine learning are no longer sufficient to meet the demands of equipment health diagnosis. Next-generation artificial intelligence technologies, characterized by self-sensing, self-adaptation, self-learning, and self-decision-making, and centered on knowledge engineering, with the development of deep learning, have brought new opportunities to IFD.

[0003] Prognostics and Health Management (PHM) is a novel solution for managing system health status, developed by integrating the latest research findings in modern information technology and artificial intelligence. In-Fault Determination (IFD) is a part of PHM. PHM can predict the time and location of impending failures, forecast the remaining service life of the entire system, improve system reliability and safety, reduce maintenance costs and improve maintenance accuracy, and achieve condition-based maintenance decision (CMD) capabilities. PHM can be divided into eight steps: preliminary analysis, data acquisition, detection, feature engineering, diagnosis, health assessment, prediction, maintenance, and management.

[0004] Rotating assembly generates multimodal signals such as vibration, heat, and electromagnetic fields during operation. These signals contain rich equipment status information. Acoustic fingerprint signals are a refined extraction of information such as operating characteristic frequencies, amplitudes, and modulation modes. Accurate monitoring and spectral analysis of multimodal signals can enable intelligent diagnosis of mechanical faults. Among these methods, vibration signal diagnostic sensors do not need to be in direct contact with the machinery, making them flexible, convenient, and easy to implement without power interruption.

[0005] Rolling bearings are a representative example of rotating assembly components. During long-term operation, rolling bearings are subject to continuous alternating loads, contact fatigue, and environmental corrosion. This leads to progressive performance degradation at the interface between the rolling elements and raceways, manifesting as surface wear, material aging, and fatigue spalling. Bearings in industrial settings often operate in complex environments with coupled multi-source interference, including strong background vibrations, electromagnetic noise, and non-stationary excitations. These factors significantly reduce the signal-to-noise ratio of vibration signals, causing early, subtle fault features to be masked by noise, posing a serious challenge to traditional feature extraction methods. Further complicating matters, bearing vibration signals under varying operating conditions exhibit significant non-stationary characteristics, with their statistical distribution dynamically changing with parameters such as load and speed. This condition-dependent nature causes a sharp decline in the generalization performance of diagnostic models based on static distribution assumptions. Moreover, the scarcity of high-quality fault samples in practical engineering further restricts the effectiveness of data-driven methods, especially under small sample conditions, where the performance of deep learning models is often difficult to guarantee. Therefore, developing intelligent diagnostic algorithms with strong noise immunity and cross-condition adaptability has become a core scientific problem urgently needing to be solved in the field of health diagnosis.

[0006] A US patent (Goodman MA, Bishop W, Mohr G. System for bearing fault detection: US9200979B2[P]. 2015-12-01.) proposes a bearing fault detection method based on ultrasound and spectrum analysis: High-frequency signals from the bearing are acquired using an ultrasonic sensor, demodulated and converted to the desired frequency band, and then a spectrum is generated using Fast Fourier Transform (FFT). This spectrum is dynamically compared with four pre-stored fault feature spectra, and the fault type is finally determined by an amplitude threshold. However, the inherent limitations of FFT make it insufficient for non-stationary signals, and this method relies on a pre-set feature library.

[0007] US Patent (Girondin V, Cassar JP. System and method of detecting defects of a rolling bearing by vibration analysis: US10281438B2[P]. 2019-05-07.) proposes a method for health diagnosis of rolling bearings based on vibration signal analysis. This method collects time-domain vibration signals using an accelerometer and performs dynamic frequency correction and signal decomposition, separating the vibration signals into deterministic and random components. It then extracts the characteristic frequencies of bearing defects by combining frequency domain analysis and time-domain statistical features, and determines the fault type using a preset threshold. However, this method relies on a preset theoretical fundamental frequency and its dispersion range, which may not be applicable to actual working conditions.

[0008] The US patent (Unnikrishnan J, He L, Matthews BA, et al. Wind turbinefault detection using acoustic, vibration, and electrical signals: US10495693B2[P]. 2019-12-03.) first demodulates and preprocesses the signal using high-pass filtering, and then uses Welch power spectrum analysis to extract characteristic frequencies. However, traditional FFT is difficult to effectively solve the spectral ambiguity under wind power variable speed conditions, and is not suitable for capturing transient features. Summary of the Invention

[0009] To address the challenge of high-precision health diagnosis of rotating joint assemblies under complex operating conditions, this invention proposes a dual-domain transfer health diagnosis method for rotating joint assemblies based on sensitivity cross-modal sensing. This invention designs a sensitivity cross-modal sensing mechanism based on the fusion of feature extraction and deep learning. First, this method extracts multimodal features, including time-domain, frequency-domain, spectral kurtosis, and variational mode decomposition components, from the vibration signal of the rotating joint assembly, constructing a cross-modal feature set to comprehensively characterize the signal properties. Subsequently, deep feature learning is achieved based on a spatio-temporal health representation (STHP) architecture: MSSE is used to automatically extract local spatial structural features, and TCCA is used to capture temporal dynamic correlations, thus forming a fused feature representation that combines spatial sensitivity and temporal dependence. Simultaneously, a sensitivity cross-modal sensing-based hyperparameter evolution optimization strategy is introduced to replace the traditional experience-based hyperparameter setting process. This strategy simultaneously measures the differences in representation distribution and diagnostic errors between the original and variable operating condition domains during the validation phase. It also constructs a joint fitness function by combining the modal drift degree and sensitivity indices of mechanical state signals and acoustic signature signals. During iterative evolution, the dual-domain and dual-modal weights are adaptively updated based on the drift gating coefficients, thereby achieving adaptive optimization of key parameters such as the number of hidden layer units, learning rate, and random inactivation rate, while also considering cross-domain generalization and training stability. Ultimately, this method improves the accuracy and robustness of fault identification in small sample and variable operating condition scenarios, providing scalable technical support for the intelligent operation and maintenance of rotating part assemblies.

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

[0011] I. A dual-domain migration health diagnosis method for rotating pair assemblies based on sensitivity cross-modal sensing

[0012] Step 1: Obtain the working state data of the rotating joint assembly components in the original working condition and the variable working condition domains, and record them as the first working condition state data and the second working condition state data. After preprocessing the first working condition state data and the second working condition state data respectively, obtain the original domain dataset and the variable domain dataset. Construct training set and validation set based on the original domain dataset and the variable domain dataset. The training set and validation set are both imbalanced datasets where the number of original working condition samples is greater than the number of variable working condition samples.

[0013] Step 2: Construct a health diagnosis model for rotating joint assemblies. This model comprises a multi-scale spatial representation extraction (MSSE) submodule, a temporal contextual aggregation (TCCA) submodule, and a classification module, all connected sequentially. The rotating joint assembly health diagnosis model is trained using training and validation sets to obtain a trained model. The key hyperparameters of the temporal contextual aggregation submodule are optimized using an alpha evolutionary algorithm based on sensitivity-based cross-modal cognition (SCMC).

[0014] Step 3: Collect the working status data of the rotating joint assembly component under test and perform data preprocessing. Then, input the data into the trained rotating joint assembly health diagnosis model. The model outputs the health diagnosis results of the rotating joint assembly component under test.

[0015] The multi-scale spatial representation extraction submodule includes a first convolutional layer, a first batch of normalized layers, a first activation layer, a second convolutional layer, a second batch of normalized layers, a second activation layer, and a global max pooling layer, which are connected in sequence.

[0016] In step two, the key hyperparameters of the temporal context aggregation submodule include the number of first-layer hidden units and the number of second-layer hidden units, as well as the random inactivation rate and the learning rate.

[0017] In step two, the Alpha Evolutionary Algorithm based on sensitivity cross-modal sensing is obtained by combining the sensitivity cross-modal sensing mechanism with an improved fitness function of the Alpha Evolutionary Algorithm. The improved fitness function satisfies the following formula:

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] in, and These represent the current first fitness and the current second fitness, respectively. and These represent the new first fitness and the new second fitness, respectively. and These represent the current key hyperparameters and the new key hyperparameters, respectively. and These represent the health diagnosis error rates of the samples corresponding to the original operating conditions and the health diagnosis error rates of the samples corresponding to the changed operating conditions in the validation set, respectively. Indicates the dynamic adjustment coefficient; Indicates the number of iterations that have been executed so far; Indicates the maximum number of iterations; and These represent the first and second cost-sensitive weights after the current update; and These represent the first and second initial empirical weights, respectively. and These represent the normalized first and second cost-sensitive weights, respectively; and These represent the current cost sensitivity and the new cost sensitivity, respectively.

[0028] In step one, the preprocessing includes denoising, data segmentation, and feature calculation.

[0029] The features used in the feature calculation include time-domain features, frequency-domain features, spectral kurtosis features, and variational mode decomposition features.

[0030] II. A Dual-Domain Migration Health Diagnosis System for Rotating Pair Assemblies Based on Sensitivity Cross-Modal Sensing

[0031] The data acquisition unit is used to acquire the working status data of the rotating pair assembly components;

[0032] The data preprocessing unit is used to preprocess the working status data;

[0033] Data set construction unit, used to construct training and test sets based on datasets from the original and variable operating condition domains;

[0034] The training parameter optimization unit is used to train the health diagnosis model of the rotating joint assembly based on the training set and the test set. During the training process, the key hyperparameters of the time context aggregation submodule are optimized using the alpha evolution algorithm based on sensitivity cross-modal perception to obtain the optimal key hyperparameters, thereby obtaining the trained health diagnosis model of the rotating joint assembly.

[0035] The diagnostic output unit is used to process the preprocessed data corresponding to the working status data of the rotating joint assembly component under test using the rotating joint assembly health diagnostic model, and obtain the health diagnostic results of the rotating joint assembly component under test.

[0036] III. A computer device

[0037] The device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method for dual-domain migration health diagnosis of rotating pair assemblies based on sensitivity cross-modal sensing.

[0038] IV. A computer-readable storage medium

[0039] The medium stores a computer program, which, when executed by a processor, implements the steps of the method for dual-domain migration health diagnosis of a rotating pair assembly based on sensitivity cross-modal sensing.

[0040] V. A computer program product

[0041] The product includes a computer program / instructions that, when executed by a processor, implement the steps of the method for dual-domain migration health diagnosis of a rotating pair assembly based on sensitivity cross-modal sensing.

[0042] This invention, based on the Alpha Evolution (AE) algorithm, innovatively introduces a Sensitivity-based Cross-Modal Cognition (SCMC) mechanism to construct a health diagnostic model supporting cross-domain transfer. This method efficiently transfers knowledge from a data-rich source domain to a data-sparse target domain by jointly optimizing the task error rates of the source and target domains. Simultaneously, the SCMC mechanism adaptively optimizes key hyperparameters of the temporal context aggregation submodule and dynamically adjusts task weights during the transfer process, constructing the SCMC-Net health diagnostic model. This allows the model to maintain high diagnostic accuracy even under conditions of sufficient original working condition samples and small samples under varying working conditions, without requiring manual post-processing parameter adjustments, significantly enhancing the model's cross-domain generalization ability. Furthermore, by combining human identification results, it can achieve accurate classification of different types of defects, meeting the needs for refined defect diagnosis in actual production processes.

[0043] The present invention has the following beneficial effects:

[0044] 1. This invention employs a fitness function design based on dual-task weighted loss, fundamentally achieving joint optimization of the original working conditions (knowledge source domain) and small samples of varying working conditions (knowledge target domain). This constitutes a powerful transfer learning paradigm, effectively avoiding the overfitting of traditional single-objective optimization methods to a particular working condition under varying data distributions, thus effectively mitigating the negative impact of feature distribution shift and domain drift on model performance. This enables the model to efficiently and robustly transfer the rich knowledge learned from the data-rich original working conditions to scenarios with sparse data and variable working conditions, significantly improving the model's accuracy and robustness in health diagnosis under varying working conditions.

[0045] 2. This invention proposes a sensitivity-based cross-modal sensing mechanism that dynamically adjusts the cost weights between tasks based on real-time error feedback during the evolutionary process. This adaptive optimization method achieves self-regulation through multi-task collaboration, automatically increasing attention to tasks with larger errors and higher diagnostic difficulty. This not only reduces reliance on human experience or prior parameters, ensuring the intelligence and stability of model training, but also, because the fitness function eliminates the need for additional fine-tuning steps, greatly simplifies the training process and saves computational resources and time costs.

[0046] 3. This invention, based on the traditional AE algorithm, introduces the SCMC dynamic adjustment mechanism and combines it with the STHR deep feature extraction network and systematic feature engineering methods to construct a complete health diagnosis system for rotating joint assemblies. This system can achieve high-precision fault identification under original operating conditions, while maintaining good diagnostic robustness and generalization ability in small sample and variable operating condition scenarios. By extracting local spatial features through a multi-scale spatial representation extraction submodule and capturing temporal correlation features through a temporal context aggregation submodule, and then dynamically optimizing parameters and feature selection using the SCMC method, it not only effectively coordinates the relationship between exploration and utilization, avoiding premature convergence, but also alleviates overfitting and underfitting problems during multi-objective optimization. Compared with single algorithms, this invention significantly improves the accuracy, stability, and adaptability of health diagnosis, providing a highly reliable solution for health monitoring of rotating joint assemblies. Attached Figure Description

[0047] Figure 1 This is the overall flowchart of a dual-domain sensitivity-sensing method for cross-domain migration health diagnosis of rotating joint assemblies;

[0048] Figure 2 This is a block diagram of the SCMC-Net health diagnosis model;

[0049] Figure 3 This is the overall flowchart of the SCMC algorithm;

[0050] Figure 4 This is a graph showing the loss values ​​during the training process of the cross-domain transfer health diagnosis model and the comparative model;

[0051] Figure 5 It is an accuracy curve during the training process of the cross-domain transfer health diagnosis model and the comparative model;

[0052] Figure 6 This is a bar chart comparing the test set accuracy of the cross-domain transfer health diagnosis model and the comparative model;

[0053] Figure 7 This is a bar chart comparing the test set accuracy of the cross-domain transfer health diagnosis model and the comparative model;

[0054] Figure 8 This is a bar chart comparing the test set recall rates of the cross-domain transfer health diagnosis model and the comparative model;

[0055] Figure 9 This is a bar chart comparing the mean absolute error of the test sets of the cross-domain transfer health diagnosis model and the comparative model;

[0056] Figure 10 It is the confusion matrix of the test set only for the TCCA model;

[0057] Figure 11 It is the test set confusion matrix for the MSSE model only;

[0058] Figure 12 This is the test set confusion matrix of the STHR hybrid model;

[0059] Figure 13 This is the confusion matrix of the test set for the SCMC-Net model;

[0060] Figure 14 This is a t-SNE aggregation and separation plot for the TCCA model only;

[0061] Figure 15 This is a t-SNE aggregation and deaggregation plot for the MSSE model only;

[0062] Figure 16 This is a t-SNE aggregation and separation plot of the STHR mixture model;

[0063] Figure 17 This is the t-SNE aggregation and decomposition graph of the SCMC-Net model;

[0064] Figure 18 This is a bar chart comparing the accuracy of cross-domain transfer health diagnosis models and comparative models at different signal-to-noise ratios.

[0065] Figure 19 This is a comparison chart of the fitness curves of the algorithm of this invention and different optimization algorithms;

[0066] Figure 20 This is a bar chart comparing the running time of the algorithm of this invention and different optimization algorithms;

[0067] Figure 21 This is a test set accuracy density distribution diagram of the algorithm of this invention and different optimization algorithms;

[0068] Figure 22 This is a bar chart comparing the accuracy of the algorithm of this invention in various cross-domain migration scenarios. Detailed Implementation

[0069] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0070] like Figure 1 As shown, the present invention proposes a dual-domain migration health diagnosis method for rotating pair assemblies based on sensitivity cross-modal sensing, which includes the following steps:

[0071] Step 1: Collect vibration signals from healthy and faulty rolling bearings under actual operating conditions.

[0072] The data used in this embodiment comes from the publicly released rolling bearing experimental dataset from Case Western Reserve University. The test system uses a 2-horsepower electric motor and an SKF6205 bearing. The vibration signal acquisition frequency is 12kHz. The bearing faults were artificially created at different locations using electrical discharge machining (EDM), specifically at the 3, 6, and 12 o'clock positions on the outer ring of the bearing. According to the table, the experimental data can be divided into four subsets, labeled A, B, C, and D, corresponding to motor speeds of 0hp-1797 r / min, 1hp-1772 r / min, 2hp-1750 r / min, and 3hp-1730 r / min, respectively, reflecting the vibration response under four different operating conditions.

[0073] The CWRU dataset used in this embodiment contains four load conditions, three fault types, and three different fault diameters. The fault types include rolling element faults (Ball), inner race faults (Inner Race), outer race faults (Outer Race), and a normal condition, with fault diameters of 0.007 inches, 0.014 inches, and 0.021 inches, respectively. The dataset is first divided into z0, z1, z2, and z3 according to load conditions, and then into ten types (0-9) according to fault types. Taking z0 condition as an example, the long-term bearing vibration signal is denoised and segmented using a sliding window. Each segment contains 1024 sampling points with a sliding step size of 100 to ensure that each segment covers complete bearing rotation cycle information. For each fault type, 200 samples are extracted from both the original and variable load conditions. 100 samples from the original load condition and 10 samples from the variable load condition are selected to form the training sample set, and the remaining samples are used as the corresponding test sets, forming complete training and test sample sets. Table 1 provides an overview of the experimental dataset, and Table 2 shows the composition of the bearing experimental samples.

[0074] Table 1 Experimental Dataset

[0075]

[0076] Table 2 Composition of bearing test samples

[0077]

[0078] For each sample in both the training and testing sets, multidimensional feature extraction is performed, including time-domain features, frequency-domain features, spectral kurtosis features, and variational mode decomposition features, forming a multidimensional feature set. Specifically, this includes: time-domain features (such as mean, root mean square value, standard deviation, peak value, peak-to-peak value, kurtosis factor, sharpness, skewness, signal energy, impulse factor, and waveform factor), frequency-domain features (including average amplitude, frequency center, root mean square frequency, and frequency standard deviation), and time-frequency-domain features (subband energy ratio and energy entropy extracted based on wavelet packet decomposition). Furthermore, to enhance the representation of non-stationary impact signals, spectral kurtosis features (including spectral kurtosis mean, standard deviation, skewness, and kurtosis) and variational mode decomposition features (i.e., mean, root mean square value, kurtosis, peak-to-peak value, and energy of each modal component) are further extracted. The feature set is then normalized using the min-max normalization method, calculated as follows:

[0079]

[0080] Where X is the original feature value, The normalized feature set This represents the minimum value of the original feature set. This represents the maximum value of the original feature set.

[0081] Then, the normalized training sample set is divided into a training set and a validation set in an 8:2 ratio. Both the training set and the validation set are imbalanced datasets where the number of samples under the original working condition is greater than the number of samples under the variable working condition; the ratio of the number of samples under the original working condition to the number of samples under the variable working condition can be from 8:1 to 15:1, and 10:1 is used in this example.

[0082] Step 2: Construct a health diagnosis model for the rotating joint assembly. This model comprises a multi-scale spatial representation extraction submodule, a temporal context aggregation submodule, and a classification module, all connected sequentially. The model is trained and optimized using training and validation sets to obtain a well-trained and optimized health diagnosis model. The key hyperparameters of the temporal context aggregation submodule are optimized using an alpha evolution algorithm based on sensitivity cross-modal perception. The optimization of the key hyperparameters of the temporal context aggregation submodule requires fewer training rounds (5 rounds in this invention), thus optimizing them faster than other training parameters. The hyperparameters of the multi-scale spatial representation extraction submodule, the classification module, and other parameters of the temporal context aggregation submodule (such as weights and biases) are obtained using conventional training methods and do not require additional optimization.

[0083] In one feasible implementation, the multi-scale spatial representation extraction submodule includes a first convolutional layer, a first batch of normalized layers, a first activation layer, a second convolutional layer, a second batch of normalized layers, a second activation layer, and a global max pooling layer, all connected in sequence. The activation function of the activation layer is a modified linear unit activation function. Unlike traditional local pooling methods, this embodiment employs global max pooling, which can extract representative salient features from high-dimensional feature maps, thereby effectively reducing feature dimensionality, reducing redundant information, and significantly improving the robustness and generalization performance of the model under various conditions.

[0084] This invention utilizes the gating mechanism of the temporal context aggregation submodule to model the temporal dependence of vibration signals. By employing modified linear unit activation and cyclic control of the sigmoid activation function, key information is effectively captured, resulting in a more complete representation of temporal features. To further avoid overfitting and improve generalization performance across different operating conditions, this invention introduces a dropout regularization layer based on the SCMC mechanism at the output of the temporal context aggregation submodule, thereby enhancing the model's stability and robustness.

[0085] Finally, the fault features extracted and processed by the aforementioned modules are input into the normalized exponential function classification module. This module maps high-dimensional features to the classification space through a fully connected layer, enhances the nonlinear expressive power of the features using a modified linear unit activation function, and combines the normalized exponential function to achieve probabilistic discrimination of multi-class fault states.

[0086] Key hyperparameters for optimizing the temporal context aggregation submodule include the number of first-layer hidden units and second-layer hidden units, as well as the random inactivation rate and learning rate.

[0087] In one feasible implementation, the Alpha Evolutionary Algorithm based on sensitivity cross-modal sensing is obtained by combining a dual-domain sensitivity sensing mechanism and improving the fitness function of the Alpha Evolutionary Algorithm.

[0088] The following section provides a detailed introduction to the Alpha Evolutionary Algorithm based on sensitivity cross-modal sensing, with the specific process as follows: Figure 3 As shown.

[0089] First, initialization is performed to generate a candidate solution matrix. :

[0090]

[0091] in, The first candidate solution in the matrix The individual Element values ​​in dimension A random number between 0 and 1 The sign for element-wise multiplication is given, where N is the population size and D is the problem dimension. and These are the upper and lower bounds of the d-th dimension parameter, respectively.

[0092] Secondly, construct the evolutionary matrix, that is, generate a matrix with replacement by sampling. Matrix of equal size :

[0093]

[0094]

[0095] in, To obtain from the candidate solution matrix The first one randomly selected from the middle Individual; It is a discrete uniform distribution function; It is the random integer index corresponding to the i-th individual.

[0096] Next, the perturbation term is generated. With decay factor :

[0097]

[0098] in, , It is a random matrix. It is a Boolean mask used for partial dimensional perturbations.

[0099]

[0100] in, For the current number of assessments, This is the preset maximum number of evaluations.

[0101] Then, update the constructed basis vectors according to the corresponding evolution path. Based on whether the random number is less than 0.5, the update methods are divided into diagonal sampling and weighted average.

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] in, To construct a matrix by randomly selecting from the population, To select K individuals from the current population, Weights are based on fitness. For updating using the diagonal sampling update method The basis vectors at time t; For updating using the weighted average method The basis vectors at time t; A random number between 0 and 1; and There are two adaptive weights; For the current moment basis vectors; For the next moment basis vectors; A function to extract the diagonal elements of a vector; The fitness function; For the first The fitness function value of each individual; For the selected K individuals The Middle Individual.

[0109] Next, the following formula is used to sample the current individual from the population. A better individual Compared to worse individuals :

[0110]

[0111]

[0112] Then, the individual is updated using the following formula:

[0113]

[0114]

[0115] in, For local development factors, the two possibilities are the same.

[0116] Then, the half-distance method is used to solve the boundary conditions, which smooths the boundary bounce while maintaining the search jump and avoids getting stuck at the boundary point.

[0117]

[0118] in, for Time-based evolution matrix The i-th individual Element values ​​in the dimension; For better individual Element values ​​in the dimension; For worse individuals The element value in the dimension.

[0119] Existing AE optimization algorithms often employ a greedy selection strategy in parameter updates. While this strategy offers advantages such as fast convergence and simple implementation, it often struggles to explicitly characterize the differences in importance between the source and target domains, as well as the cost structure introduced by domain drift, in cross-domain transfer learning scenarios. This can easily lead to a bias towards the source domain and getting stuck in local optima, thus limiting the model's generalization adaptability under complex operating conditions and noise disturbances. To address this, this embodiment proposes a SCMC optimization algorithm, incorporating a dual-domain drift-aware cross-modal sensitivity collaborative hyperparameter evolution optimization strategy to replace the traditional experience-based hyperparameter setting process. This strategy employs progressive training and evaluation of candidate hyperparameters, and during the validation phase, calculates the classification error in the original operating domain and the difference in representation distribution between the two domains. Simultaneously, it calculates modal drift metrics and parameter perturbation sensitivity indices for mechanical state signals and acoustic signature signals, respectively. Based on these, a joint fitness function comprising "error term - drift penalty term - sensitivity penalty term" is constructed. During the iterative evolution process, the contribution weights of the dual-domain and dual-modal approaches are adaptively adjusted through drift gating coefficients. Key parameters such as the number of hidden layer units, learning rate, and random inactivation rate are also optimized collaboratively. This allows hyperparameter search and transfer weight adjustment to be executed in conjunction within a unified sensitivity decision framework, thereby effectively suppressing premature convergence and overfitting risks while maintaining relatively fast convergence, thus improving model training stability and cross-domain generalization ability. Ultimately, this method maintains high fault identification performance and robustness in small sample sizes and variable operating conditions, providing scalable technical support for intelligent operation and maintenance of rotating joint assemblies.

[0120] The improved fitness function satisfies the following formula:

[0121]

[0122]

[0123]

[0124]

[0125]

[0126]

[0127]

[0128]

[0129]

[0130] in, and These represent the current first fitness and the current second fitness, respectively. and These represent the new first fitness and the new second fitness, respectively. and These represent the current key hyperparameters and the new key hyperparameters, respectively. and These represent the health diagnosis error rates of the samples corresponding to the original operating conditions and the health diagnosis error rates of the samples corresponding to the changed operating conditions in the validation set, respectively. Indicates the dynamic adjustment coefficient; Indicates the number of iterations that have been executed so far; Indicates the maximum number of iterations; and These represent the first and second cost-sensitive weights after the current update; and These represent the first and second initial empirical weights, respectively. and These represent the normalized first and second cost-sensitive weights, respectively; and These represent the current cost sensitivity and the new cost sensitivity, respectively.

[0131] The fitness function of this invention can adaptively and dynamically change according to the relative magnitude of the optimization progress and the verification loss of a specific task, so that it can focus more on the more difficult or important tasks in the search process, and complete the cross-domain migration task in a balanced and efficient manner.

[0132] During algorithm execution, if the current iteration count has not reached a preset threshold, the algorithm returns to the evolutionary matrix construction stage for population evolution. After iteration, the individual with the best fitness in the population is selected as the optimal solution for hyperparameters, ultimately completing the construction of the SCMC-Net health diagnosis model.

[0133] Step 3: Collect the working status data of the bearing in the rotating assembly to be tested and perform data preprocessing. Then, input the extracted features into the trained rotating assembly health diagnosis model, and the model outputs the health diagnosis results of the bearing to be tested.

[0134] This invention selects accuracy (Acc), macro precision (MP), macro recall (MR), and mean absolute error (MAE) as evaluation metrics for the model.

[0135] Ablation experiments were also conducted here to compare the rotating pair assembly health diagnosis model (SCMC-Net) of this invention with three other models: TCCA only, MSSE only, and STHR only, to verify the effectiveness of each module of this invention. Dataset z0 was selected as the main dataset here. Figure 4 , Figure 5 The figures compare the loss curves and accuracy curves for the ablation experiments, with the horizontal axis representing the number of training iterations and the vertical axis representing the numerical values. For the TCCA method alone, the loss function at 300th epoch is 0.323, the average slope of the curve is 0.145, and the number of steps in which the curve slope first reaches 0.01 is 44. For the MSSE method alone, the loss function at 300th epoch is 0.269, the average slope of the curve is 0.163, and the number of steps in which the curve slope first reaches 0.01 is 24. For the STHR method, the loss function at 300th epoch is 0.517, the average slope of the curve is 0.129, and the number of steps in which the curve slope first reaches 0.01 is 31. For the MSSE method alone, the loss function at 300th epoch is 0.002, the average slope of the curve is 0.031, and the number of steps in which the curve slope first reaches 0.01 is 20. As can be seen, the method proposed in this invention can quickly reach and maintain an accuracy rate close to 100%, demonstrating excellent learning ability and stability. In terms of loss function convergence performance, the method of this invention significantly outperforms the other three models, exhibiting a faster convergence speed and a smoother optimization trajectory. Particularly noteworthy is that this method can rapidly compress the loss to near zero in the early stages of training, demonstrating its excellent feature fitting ability and optimization robustness, further validating its stable modeling and generalization advantages under complex data distributions. In contrast, MSSE, TCCA, and their fusion model STHR show significantly slower convergence speeds during training, and their accuracy fluctuates greatly, exhibiting a certain degree of instability. In particular, the STHR model still shows significant oscillations in the mid-to-late training stages, indicating that its convergence and stability in complex feature learning still need improvement. Furthermore, as shown in the two figures here, the method of this invention achieves a rapid decrease in the loss value in the early stages of training, with an accuracy rate approaching 100% at 73 rounds and basically stabilizing at 100% at 187 rounds, demonstrating a faster convergence speed than the comparison models. More importantly, the model demonstrated stability during subsequent training, with no significant fluctuations or rebounds in loss values ​​or accuracy. In contrast, although the STHR model achieved higher accuracy in later stages, its accuracy curve fluctuated considerably, indicating a certain risk of overfitting. Comprehensive analysis shows that the method of this invention not only effectively alleviates the underfitting and overfitting problems in model training but also exhibits excellent convergence characteristics and optimization performance, further validating its advantages in hyperparameter tuning and cross-domain transfer learning for health diagnosis modeling.

[0136] Figure 6, Figure 7 , Figure 8 , Figure 9 The graph compares the evaluation metrics obtained from the four methods, with the horizontal axis representing the model name and the vertical axis representing the evaluation metric value. The comparison data shows that when using the TCCA model alone, its accuracy is only 0.87, precision and recall are 0.37 and 0.35 respectively, and the MAE is as high as 0.67, indicating that pure time-series modeling has significant shortcomings in feature extraction and prediction accuracy. Using the MSSE model alone can improve the accuracy to 0.90, with precision and recall reaching 0.92 and 0.90 respectively, and the MAE significantly decreasing to 0.13, indicating that the network has strong capabilities in spatial feature extraction. Combining TCCA and MSSE to obtain STHR further improves the accuracy to 0.93, with precision and recall reaching 0.94 and 0.93 respectively, and the MAE decreasing to 0.10, demonstrating the complementary advantages of the two methods. Adding the optimization algorithm (SCMC) of this invention further achieves optimal performance across all metrics. Accuracy, precision, and recall all reached 0.99, and MAE was further reduced to 0.03, demonstrating the effectiveness of SCMC in optimizing model parameters and feature selection, and providing a high-precision, low-error solution for health diagnosis systems. Therefore, the method of this invention outperforms the comparative model in multiple dimensions such as accuracy, fit, and error control, verifying its effectiveness and robustness in health diagnosis tasks under complex working conditions.

[0137] Figure 10 , Figure 11 , Figure 12 , Figure 13 These are confusion matrices for four different methods, commonly used to evaluate the performance of classification models. Each column represents the predicted value, and each row represents the true value. The graph also uses percentages, with each cell indicating the probability of a sample being predicted as a particular type. The closer all the values ​​on the diagonal are to 1, the higher the classification accuracy and the better the generalization ability of the model. Figure 9 It can be seen that the TCCA model exhibits significant misclassification in certain categories, particularly categories 2 (0.80), 3 (0.83), 9 (0.87), and 10 (0.90). This indicates that the method has weak discriminatory power for specific types. Category 10 is typically an easily identifiable category, yet this model only achieves an accuracy of 0.90, demonstrating its strong dependence on global features and limited generalization ability. Figure 10 It can be seen that only the MSSE model showed improvement, with the accuracy for class 2 increasing to 0.90, and classes 9 and 10 reaching 0.93 and 1.00 respectively, indicating its advantage in local spatial feature extraction. However, class 3 still only reached 0.83, showing its insufficient expression of temporal features, resulting in insufficient fine-grained class discrimination ability. Figure 11It can be seen that the overall accuracy of the STHR model has further improved, with type 2 reaching 0.97 and type 10 remaining at 1.00. However, type 3 is still only 0.87, and a new problem has emerged where the accuracy of type 6 drops to 0.83. This indicates that although fusion improves the overall performance, there are still unstable situations in the recognition of complex categories. Figure 12 As can be seen, the confusion matrix of the test set of the method of this invention shows that the predicted type is basically consistent with the true type. Only samples with the true type 3 have a 7% probability of being predicted as type 7. Therefore, the method of this invention demonstrates excellent performance in classifying these 10 types of faults, good generalization ability, and a significant improvement in prediction performance for type 3.

[0138] Figure 14 , Figure 15 , Figure 16 , Figure 17 The t-SNE aggregation and separation processes of the four methods are visualized, clearly showing the feature extraction capabilities of the classification models. Each point in the graph represents a sample; the more samples of the same type are aggregated, and the more samples of different types are separated, the better the model's feature extraction capability. Figure 14 It can be seen that while the TCCA model has a good aggregation effect on samples, many types are mixed with other types, and some aggregations have connections between them, indicating poor feature extraction ability; Figure 15 It can be seen that only MSSE's model sample aggregation and separation effects are better than the former, but the aggregations of type 2 and type 3 are closer together and have contamination, and type 5 contains points of different types; Figure 16 It can be seen that the STHR mixture model has better sample separation capabilities, but there are still quite a few Type 3 points mixed with Type 4, and some connect with Type 5, so there are still shortcomings; Figure 17 It can be seen that the method of the present invention effectively aggregates and separates the samples, with only one point of type 3 mixed into type 7, proving the strong feature extraction capability of the method of the present invention.

[0139] To more clearly demonstrate the ability of this invention to separate different fault types, the Euclidean distance between the cluster centers of the above t-SNE clustering can be calculated for a more intuitive demonstration. Where D... i-j This represents the Euclidean distance between category i and category j.

[0140] The calculated value for TCCA alone is: D 1-2 =33.03, D 1-3 =71.02, D 1-4 =46.88, D 1-5 =26.10, D 1-6 =22.78, D 1-7 =56.74, D 1-8=16.20, D 1-9 =41.72, D 1-10 =43.61, D 2-3 =52.87, D 2-4 =26.49, D 2-5 =39.14, D 2-6 =47.47, D 2-7 =22.29, D 2-8 =66.82, D 2-9 =20.50, D 2-10 =35.91, D 3-4 =46.91, D 3-5 =62.96, D 3-6 =36.50, D 3-7 =26.02, D 3-8 =48.42, D 3-9 =31.39, D 3-10 =29.97, D 4-5 =47.38, D 4-6 =62.96, D 4-7 =36.50, D 4-8 =26.02, D 4-9 =48.42, D 4-10 =31.39, D 5-6 =22.29, D 5-7 =66.82, D 5-8 =20.50, D 5-9 =35.91, D 5-10 =29.97, D 6-7 =47.47, D 6-8 =22.29, D 6-9 =66.82, D 6-10 =20.50, D 7-8 =36.50, D 7-9 =26.02, D 7-10 =48.42, D 8-9 =31.39, D 8-10 =29.97, D 9-10 =35.91. The average value is 35.66.

[0141] The calculated value for MSSE alone is: D 1-2 =47.51, D 1-3 =44.61, D 1-4 =51.58, D 1-5 =27.30, D 1-6 =65.37, D 1-7 =57.34, D 1-8 =35.95, D1-9 =23.73, D 1-10 =44.16, D 2-3 =8.43, D 2-4 =62.61, D 2-5 =67.06, D 2-6 =25.78, D 2-7 =32.20, D 2-8 =49.07, D 2-9 =81.34, D 2-10 =22.19, D 3-4 =40.59, D 3-5 =19.79, D 3-6 =63.90, D 3-7 =53.06, D 3-8 =51.35, D 3-9 =31.69, D 3-10 =28.59, D 4-5 =19.71, D 4-6 =40.59, D 4-7 =19.79, D 4-8 =63.90, D 4-9 =53.06, D 4-10 =51.35, D 5-6 =25.78, D 5-7 =32.20, D 5-8 =49.07, D 5-9 =81.34, D 5-10 =22.19, D 6-7 =67.06, D 6-8 =25.78, D 6-9 =32.20, D 6-10 =49.07, D 7-8 =81.34, D 7-9 =22.19, D 7-10 =40.59, D 8-9 =19.79, D 8-10 =63.90, D 9-10 =53.06, with an average value of 41.63.

[0142] The calculated value of SHTR is: D 1-2 =52.03, D 1-3 =49.67, D 1-4 =32.99, D 1-5 =31.43, D 1-6 =47.98, D 1-7 =57.35, D 1-8 =82.68, D 1-9=77.17, D 1-10 =68.32, D 2-3 =22.39, D 2-4 =21.77, D 2-5 =42.65, D 2-6 =43.88, D 2-7 =43.17, D 2-8 =34.31, D 2-9 =58.30, D 2-10 =84.31, D 3-4 =18.29, D 3-5 =18.29, D 3-6 =27.25, D 3-7 =22.27, D 3-8 =62.41, D 3-9 =49.17, D 3-10 =46.26, D 4-5 =22.27, D 4-6 =62.41, D 4-7 =49.17, D 4-8 =46.26, D 4-9 =36.54, D 4-10 =28.60, D 5-6 =43.17, D 5-7 =34.31, D 5-8 =58.30, D5-9=84.31, D 5-10 =62.41, D 6-7 =47.98, D 6-8 =57.35, D 6-9 =82.68, D 6-10 =77.17, D 7-8 =31.43, D 7-9 =47.98, D 7-10 =57.35, D 8-9 =82.68, D 8-10 =77.17, D 9-10 =68.32, with an average of 46.07.

[0143] The calculated value of this invention is: D 1-2 =23.04, D 1-3 =46.16, D 1-4 =39.66, D 1-5 =29.35, D 1-6 =46.11, D 1-7 =53.33, D 1-8 =29.56, D 1-9 =30.47, D 1-10=30.17, D 2-3 =32.62, D 2-4 =53.71, D 2-5 =51.93, D 2-6 =23.80, D 2-7 =43.21, D 2-8 =34.10, D 2-9 =48.84, D 2-10 =55.75, D 3-4 =84.05, D 3-5 =19.72, D 3-6 =58.13, D 3-7 =78.56, D 3-8 =31.76, D 3-9 =46.09, D 3-10 =58.11, D 4-5 =46.09, D 4-6 =58.11, D 4-7 =78.56, D 4-8 =31.76, D 4-9 =46.09, D 4-10 =58.11, D 5-6 =23.80, D 5-7 =43.21, D 5-8 =34.10, D 5-9 =48.84, D 5-10 =55.75, D 6-7 =53.71, D 6-8 =23.80, D 6-9 =43.21, D 6-10 =34.10, D7-8=48.84, D 7-9 =55.75, D 7-10 =84.05, D 8-9 =19.72, D 8-10 =58.13, D 9-10 =78.56, with an average value of 48.62.

[0144] Based on the statistical values ​​of the Euclidean distance between the category centers, the method of this invention improves upon the comparative method by 2.55%-12.96%, particularly in critical fault category combinations (such as D). 3-4 The maximum improvement was 65.76%. This verifies that the method of the present invention significantly improves the consistency and stability of feature expression while maintaining high class separation, and has a stronger ability to distinguish multiple fault types. Under multiple working conditions and multiple fault conditions, it can effectively alleviate feature distribution overlap.

[0145] To verify the noise resistance of the classification model, a signal-to-noise ratio experiment was set up. Figure 18 A bar chart comparing the accuracy of four methods under different signal-to-noise ratios (SNRs) is presented, with five SNR options for each method (-12dB, -8dB, -4dB, 0dB, 4dB). A lower SNR indicates higher noise levels. The vertical axis represents accuracy. The chart shows significant differences in health diagnosis accuracy among the models under different SNR conditions. Under strong noise interference (-12dB), the TCCA model alone achieves only 10.00% accuracy, while the MSSE model reaches 47.67%, the STHR hybrid model improves to 79.33%, and the method of this invention significantly outperforms the other models with an accuracy of 92.67%. As the SNR increases, the performance of all models improves, but the method of this invention consistently performs best, especially in the -4dB to 4dB range, where its accuracy remains stable at 99.33%–100%, higher than STHR (91.00%–97.67%) and the single model. This comparison verifies the robustness of the method of the present invention in noisy environments, showing that its adaptive feature optimization capability can effectively suppress noise interference, thereby achieving more reliable intelligent health diagnosis in high-noise industrial scenarios.

[0146] To assess the performance of various intelligent optimization algorithms in the hyperparameter optimization task of the TCCA model, the algorithm of this invention was systematically evaluated against three mainstream optimization algorithms—Arithmetic Optimization Algorithm (AOA), Grey Wolf Optimization Algorithm (GWO), and Sine-Cosine Algorithm (SCA)—in terms of convergence accuracy and running efficiency. Figure 19 This is a comparison chart of the fitness curves of the algorithm of this invention and different optimization algorithms. Figure 20 This is a bar chart comparing the runtime of the algorithm of this invention and different optimization algorithms. This experiment was conducted on a computing platform equipped with an NVIDIA GeForce RTX 4060 Ti (32GB GDDR6 memory) GPU and a 12th generation Intel Core i5-12400F six-core processor (base frequency 2.50 GHz, maximum turbo frequency 4.40 GHz), with 32GB of DDR4 memory. All algorithms were implemented in C language, and the GPU Parallel Computing Toolbox was enabled to utilize CUDA cores for accelerated computation. To ensure fairness and comparability in the comparative experiments, initial parameter settings were standardized across different algorithms. The dimension of the optimization problem was set to 4, corresponding to the four key hyperparameters of the temporal context aggregation submodule: the number of neurons in the two hidden layers, the learning rate, and the random inactivation rate. The search ranges for each parameter are defined as a lower bound of [10, 10, 0.0001, 0.1] and an upper bound of [200, 200, 0.01, 0.5], respectively, to balance model expressiveness and training stability. The algorithm population size is set to 10, and the maximum number of function evaluations is set to 200. Figure 19As can be seen, the algorithm of this invention exhibits optimal convergence characteristics throughout the entire iteration process. Its fitness value decreases rapidly in a short period of time and tends to stabilize after about 120 iterations, eventually reaching the minimum fitness, which is superior to mainstream optimization algorithms such as GWO, SCA, and AOA. This indicates that the algorithm of this invention has stronger global search capabilities and higher convergence accuracy, and can effectively avoid getting trapped in local optima. Figure 20 As can be seen, the algorithm of this invention is highly efficient in terms of optimization, with a total running time of only 239.7 seconds, which is better than other algorithms. The running times of AOA, GWO, and SCA are 2002.9 seconds, 1948.7 seconds, and 1908.1 seconds, respectively. This shows that the algorithm of this invention not only has a significant optimization effect, but also has a greater advantage in terms of computational resource consumption.

[0147] Figure 21 This is a density distribution of the accuracy of the algorithm and different optimization algorithms on the test set. Each point represents a single test data point. The vertical axis represents the accuracy, and the horizontal axis represents the name of each optimization algorithm. A wider width indicates that the accuracy of the model is mostly concentrated in that area, while a shorter length indicates that the model is more stable. Overall, all four algorithms operate at a high accuracy level, but there are significant differences in stability and worst-case performance. Among them, the SCMC method performs best, maintaining a high level of average accuracy and having a concentrated distribution, indicating its consistency and robustness under different experimental conditions. In contrast, the accuracy distribution of the AOA and GWO algorithms is relatively dispersed, with a certain number of low-performance points, indicating that their optimization effect is greatly affected by initial conditions or parameter sensitivity. In particular, the SCA algorithm has the widest distribution range, with the lowest point close to 0.75, showing that it suffers from severe performance degradation in some experiments. Therefore, in terms of accuracy performance and its fluctuation range, the SCMC method demonstrates performance stability and generalization ability in the training of object detection models, further confirming the effectiveness of SCMC in introducing a deep evolution strategy in dynamic optimization, which can effectively avoid local optima and continuously explore the high-quality solution space.

[0148] Figure 22This is a bar chart comparing the accuracy of the proposed algorithm and the original AE algorithm in cross-domain migration scenarios. To verify the migration performance of the algorithm under different operating condition combinations, six cross-domain migration experimental combinations were selected (z0-z1, z0-z2, z0-z3, z1-z2, z1-z3, z2-z3). Here, z0-z1 represents migration from operating condition z0 to operating condition z1, i.e., using the large sample data of operating condition z0 as the source domain, and employing a dual-task learning mechanism to assist in fault diagnosis using the small sample target domain of operating condition z1. The diagnostic accuracy of the proposed improved algorithm (SCMC) is compared with that of the original AE algorithm. The experimental results are shown in the table. Overall, the accuracy of the proposed algorithm is superior to that of the original AE algorithm in all operating condition combinations. In the combinations of operating conditions 0-2 and 2-3, SCMC achieved accuracy rates of 92.93% and 98.32%, respectively, representing improvements of 10.47 and 1.57 percentage points compared to AE's 82.46% and 96.75%, demonstrating stronger feature extraction and generalization capabilities under complex operating condition transitions. Under the most challenging 0-1 operating condition, SCMC also achieved an accuracy rate of 87.12%, an improvement of 9.48 percentage points compared to AE's 77.64%, further validating the diagnostic advantages of the method in early minor faults and complex background noise. Furthermore, under the two relatively similar operating conditions z1-z2 and z1-z3, SCMC accuracy rates reached 96.91% and 96.18%, respectively, both higher than AE's 90.63% and 85.24%, showing that SCMC also possesses superior discrimination capabilities under changing operating conditions. As can be seen, SCMC achieved superior classification performance to AE under all test conditions, effectively mitigating feature distribution shifts caused by changes in conditions, enhancing the model's adaptability to domain drift and diagnostic robustness, improving the generalization performance of cross-domain health diagnosis, and verifying its effectiveness as a sensitivity-aware-driven deep evolutionary diagnostic framework in cross-domain health diagnosis tasks.

[0149] It should be noted that in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising..." or "including..." does not exclude the presence of additional elements in the process, method, article, or terminal device that includes said element. Additionally, in this invention, "greater than," "less than," "exceeding," etc., are understood to exclude the stated number; "above," "below," "within," etc., are understood to include the stated number.

[0150] Although the above embodiments have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A dual-domain migration health diagnosis method for rotating pair assemblies based on sensitivity cross-modal sensing, characterized in that, Includes the following steps: Step 1: Obtain the working state data of the rotating joint assembly components in the original working condition and the variable working condition domains, and record them as the first working condition state data and the second working condition state data. After preprocessing the first working condition state data and the second working condition state data respectively, obtain the original domain dataset and the variable domain dataset. Construct training set and validation set based on the original domain dataset and the variable domain dataset. The training set and validation set are both imbalanced datasets where the number of original working condition samples is greater than the number of variable working condition samples. Step 2: Construct a health diagnosis model for rotating joint assemblies. The rotating joint assembly health diagnosis model includes a multi-scale spatial representation extraction submodule, a temporal context aggregation submodule, and a classification module connected in sequence. The rotating joint assembly health diagnosis model is trained based on the training set and the validation set to obtain a trained rotating joint assembly health diagnosis model. The key hyperparameters of the temporal context aggregation submodule are obtained by optimization using an alpha evolution algorithm based on sensitivity cross-modal perception. Step 3: Collect the working status data of the rotating joint assembly component under test and perform data preprocessing. Then, input the data into the trained rotating joint assembly health diagnosis model. The model outputs the health diagnosis results of the rotating joint assembly component under test.

2. The method for dual-domain migration health diagnosis of rotating pair assemblies based on sensitivity cross-modal sensing according to claim 1, characterized in that, The multi-scale spatial representation extraction submodule includes a first convolutional layer, a first batch of normalized layers, a first activation layer, a second convolutional layer, a second batch of normalized layers, a second activation layer, and a global max pooling layer, which are connected in sequence.

3. The method for dual-domain migration health diagnosis of rotating pair assemblies based on sensitivity cross-modal sensing according to claim 1, characterized in that, In step two, the key hyperparameters of the temporal context aggregation submodule include the number of first-layer hidden units and the number of second-layer hidden units, as well as the random inactivation rate and the learning rate.

4. The method for dual-domain migration health diagnosis of rotating pair assemblies based on sensitivity cross-modal sensing according to claim 1, characterized in that, In step two, the Alpha Evolutionary Algorithm based on sensitivity cross-modal sensing is obtained by combining the sensitivity cross-modal sensing mechanism with an improved fitness function of the Alpha Evolutionary Algorithm. The improved fitness function satisfies the following formula: in, and These represent the current first fitness and the current second fitness, respectively. and These represent the new first fitness and the new second fitness, respectively. and These represent the current key hyperparameters and the new key hyperparameters, respectively. and These represent the health diagnosis error rates of the samples corresponding to the original operating conditions and the health diagnosis error rates of the samples corresponding to the changed operating conditions in the validation set, respectively. Indicates the dynamic adjustment coefficient; Indicates the number of iterations that have been executed so far; Indicates the maximum number of iterations; and These represent the first and second cost-sensitive weights after the current update; and These represent the first and second initial empirical weights, respectively. and These represent the normalized first and second cost-sensitive weights, respectively; and These represent the current cost sensitivity and the new cost sensitivity, respectively.

5. The method for dual-domain migration health diagnosis of rotating pair assemblies based on sensitivity cross-modal sensing according to claim 1, characterized in that, In step one, the preprocessing includes denoising, data segmentation, and feature calculation.

6. The method for dual-domain migration health diagnosis of rotating pair assemblies based on sensitivity cross-modal sensing according to claim 5, characterized in that, The features used in the feature calculation include time-domain features, frequency-domain features, spectral kurtosis features, and variational mode decomposition features.

7. A dual-domain migration health diagnostic system for rotating pair assemblies based on sensitivity cross-modal sensing, characterized in that, include: The data acquisition unit is used to acquire the working status data of the rotating pair assembly components; The data preprocessing unit is used to preprocess the working status data; Data set construction unit, used to construct training and test sets based on datasets from the original and variable operating condition domains; The training parameter optimization unit is used to train the health diagnosis model of the rotating joint assembly based on the training set and the test set. During training, the key hyperparameters of the temporal context aggregation submodule are optimized using the alpha evolution algorithm based on sensitivity cross-modal perception to obtain the optimal key hyperparameters, thereby obtaining a well-trained health diagnosis model of the rotating joint assembly. The diagnostic output unit is used to process the preprocessed data corresponding to the working status data of the rotating joint assembly component under test using the rotating joint assembly health diagnostic model, and obtain the health diagnostic results of the rotating joint assembly component under test.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the dual-domain migration health diagnosis method for rotating pair assemblies based on sensitivity cross-modal sensing as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the dual-domain migration health diagnosis method for a rotating pair assembly based on sensitivity cross-modal sensing as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the dual-domain migration health diagnosis method for a rotating pair assembly based on sensitivity cross-modal sensing as described in any one of claims 1 to 7.

Images