Cross-domain bearing fault diagnosis method, device and equipment

Abstract

The application relates to a bearing fault diagnosis method, device and equipment across domains, which solves the frequency drift problem in the cross-domain scene by performing spectrum normalization processing on an original vibration signal, mapping discrete drift spectrum peaks to unified physical semantic coordinates, introduces double-domain collaborative feature extraction, realizes stable and accurate extraction even in strong noise and weak fault, and finally uses cross-reconstruction to constrain adaptive fusion of the normalized spectrum and the original vibration signal, reduces semantic deviation and information loss caused by manual fusion, thereby solving the problem that accurate bearing features are difficult to extract in the cross-domain scene, and thus the fault diagnosis accuracy is low.

Description

Technical Field

[0001] This application relates to the field of bearing testing technology, and in particular to a cross-domain bearing fault diagnosis method, apparatus and equipment. Background Technology

[0002] As a critical and easily failed component in rotating machinery, the operating condition of rolling bearings directly affects the safety and reliability of equipment. In high-reliability scenarios such as rail transportation, aerospace, and heavy-duty machinery, bearings typically operate in complex environments with significant variations in rotational speed, load fluctuations, and structural differences. The differences between different operating conditions can lead to significant changes in the characteristic distribution of vibration signals, particularly manifested as a drift in fault characteristic frequencies, thus affecting the fault diagnosis of bearings.

[0003] In existing technologies, bearing fault detection mainly employs two methods: one is a bearing fault diagnosis method based on envelope spectrum analysis. This method, with its well-defined physical mechanisms, possesses strong interpretability; however, it typically requires manual selection of filtering parameters and interpretation of characteristic spectral lines, making it difficult to adapt to complex and variable cross-domain operating conditions. The other is a fault diagnosis method based on deep learning. While this method possesses end-to-end feature learning capabilities, the lack of physical prior constraints makes it prone to learning domain features unrelated to the fault. Therefore, neither envelope spectrum analysis nor deep learning methods can achieve high-precision bearing feature extraction in cross-domain scenarios, resulting in low fault diagnosis accuracy. Summary of the Invention

[0004] This application provides a cross-domain bearing fault diagnosis method, apparatus, and device, which aims to solve the problem that existing bearing fault diagnosis schemes are difficult to extract accurate bearing features in cross-domain scenarios, resulting in low fault diagnosis accuracy.

[0005] The first aspect of this application provides a cross-domain bearing fault diagnosis method, the method comprising: acquiring the original vibration signal of the target bearing; performing envelope spectrum extraction and coordinate alignment processing on the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates; performing time-domain and frequency-domain collaborative feature extraction on the original vibration signal to generate dual-domain encoded features; adaptively fusing the normalized spectrum and the dual-domain encoded features through cross-reconstruction constraints to generate fused features; and performing fault diagnosis based on the fused features to obtain the fault diagnosis result of the target bearing.

[0006] In conjunction with the first aspect, in one possible implementation of the first aspect, the step of extracting the envelope spectrum and aligning the coordinates of the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates includes: extracting the envelope spectrum and aligning the coordinates of the original vibration signal according to the theoretical characteristic frequencies corresponding to the structural parameters and fault types of the target bearing to generate a normalized spectrum with unified physical semantic coordinates.

[0007] In conjunction with the first aspect, in one possible implementation of the first aspect, the step of extracting the envelope spectrum and performing coordinate alignment processing on the original vibration signal based on the structural parameters of the target bearing and the theoretical characteristic frequencies corresponding to the fault type, to generate a normalized spectrum with unified physical semantic coordinates, includes: processing the original vibration signal... Perform bandpass filtering to obtain the filtered signal. ; for the filtered signal Perform Hilbert transform and calculate the envelope to obtain the envelope signal. ; for the envelope signal After removing the DC component and performing a windowed Fast Fourier Transform, the envelope spectrum amplitude is obtained. ,in Based on the geometric parameters, fault type, and signal sampling parameters of the target bearing, the fault characteristic frequency is calculated within the envelope spectrum amplitude. The corresponding theoretical frequency position Based on the theoretical frequency point location For the envelope spectrum amplitude By performing truncation and resampling, a normalized spectrum with unified physical semantic coordinates is obtained. .

[0008] In conjunction with the first aspect, in one possible implementation of the first aspect, the step of performing time-domain and frequency-domain collaborative feature extraction on the original vibration signal to generate dual-domain coded features includes: extracting time-domain and frequency-domain features from the original vibration signal... Feature extraction is performed by inputting the data into a dual-domain Kolmogorov-Arnold network to obtain dual-domain encoded features. The dual-domain Kolmogorov-Arnold network includes a time-domain branch network and a frequency-domain branch network. The time-domain branch network is used to extract the time-domain impact features of the original vibration signal, and the frequency-domain branch network is used to extract the frequency-domain periodic structure features of the original vibration signal.

[0009] In conjunction with the first aspect, in one possible implementation of the first aspect, the original vibration signal is... Feature extraction is performed by inputting the data into a dual-domain Kolmogorov-Arnold network to obtain dual-domain encoded features. This includes: the original vibration signal Dimensional adjustment is performed to obtain time-domain input features and frequency-domain input features; the time-domain input features are then input into the time-domain branch network for processing to obtain time-domain features. The frequency domain input features are transformed to the frequency domain using a Fast Fourier Transform to obtain the frequency domain complex features. ; respectively the frequency domain complex features real part and the virtual part The input is processed by the frequency domain branch network to obtain the processed real part. and the virtual part and the real part and the imaginary part Combinations with complex characteristics ; for the complex number features Perform an inverse fast Fourier transform to obtain the frequency domain features. ;

[0010] The frequency domain features With the aforementioned time-domain features Adding them together yields the dual-domain encoded features. .

[0011] In conjunction with the first aspect, in one possible implementation of the first aspect, the adaptive fusion of the normalized spectrum and the dual-domain encoded features through cross-reconstruction constraints to generate fused features includes: [the following is a partial translation of the first part, but the full text is incomplete and requires further context.] and the dual-domain coding features Input a dynamic feature fusion network based on cross-reconstruction, and fuse features through cross-reconstruction loss constraints to obtain fused features.

[0012] In conjunction with the first aspect, in one possible implementation of the first aspect, the dynamic feature fusion network based on cross-reconstruction includes a vibration signal encoder, a normalized spectrum encoder, and a feature fusion module; the normalized spectrum... and the dual-domain coding features Inputting a dynamic feature fusion network based on cross-reconstruction, and fusing features through cross-reconstruction loss constraints to obtain fused features, including: encoding the dual-domain features... Input the vibration signal encoder to obtain vibration signal characteristics. The normalized spectrum Input the normalized spectral encoder to obtain normalized spectral features. ,in Indicates the encoder's hierarchical index; and the vibration signal characteristics. and the normalized spectral features The input feature fusion module performs a convolution operation to obtain fused features.

[0013] In conjunction with the first aspect, in one possible implementation of the first aspect, the dynamic feature fusion network based on cross-reconstruction further includes a cross-task interaction module, a decoupled encoder, and a reconstruction decoder; in the process of fusing the vibration signal features... and the normalized spectral features After the feature fusion module performs a concatenation and convolution operation to obtain the fused features, the method further includes: processing the vibration signal features... and the normalized spectral features The input is processed by the multi-task fusion module to perform cross-task interaction and feature fusion, resulting in fused task-related features. semantically related features The vibration signal characteristics and the normalized spectral features By inputting the decoupled encoder, the decoupling characteristics of the vibration signal are obtained. Decoupling characteristics of normalized spectra The vibration signal characteristics The normalized spectral features The decoupling characteristics of the vibration signal Decoupling characteristics of the normalized spectrum The first vibration signal is reconstructed by inputting it into the reconstruction decoder. Second vibration signal First normalized spectrum Second normalized spectrum Based on the first vibration signal The second vibration signal The first normalized spectrum and the second normalized spectrum Calculate the original vibration signal and the normalized spectrum The reconstruction loss between them is used to reshape the fused features based on the reconstruction loss.

[0014] A second aspect of the present invention provides a cross-domain bearing fault diagnosis device, the device comprising: an acquisition module for acquiring the original vibration signal of a target bearing; a normalization module for extracting the envelope spectrum and aligning the coordinates of the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates; an extraction module for extracting time-domain and frequency-domain collaborative features of the original vibration signal to generate dual-domain encoded features; a fusion module for adaptively fusing the normalized spectrum and the dual-domain encoded features through cross-reconstruction constraints to generate fused features; and a diagnosis module for performing fault diagnosis based on the fused features to obtain the fault diagnosis result of the target bearing.

[0015] A third aspect of the present invention provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the above-described cross-domain bearing fault diagnosis method.

[0016] The technical solution provided in this application may include the following beneficial effects: This application discloses a cross-domain bearing fault diagnosis method, apparatus, and device. The method includes acquiring the original vibration signal of the target bearing; extracting the envelope spectrum and aligning the coordinates of the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates; extracting time-domain and frequency-domain collaborative features of the original vibration signal to generate dual-domain encoded features; adaptively fusing the normalized spectrum and dual-domain encoded features through cross-reconstruction constraints to generate fused features; and performing fault diagnosis based on the fused features to obtain the fault diagnosis result of the target bearing.

[0017] This approach solves the frequency drift problem in cross-domain scenarios by normalizing the spectrum of the original vibration signal and mapping the discrete drifting spectral peaks to unified physical semantic coordinates. Furthermore, it introduces dual-domain collaborative feature extraction, enabling stable and accurate extraction even under strong noise and weak faults. Finally, it utilizes cross-reconstruction constraints to adaptively fuse the normalized spectrum and the original vibration signal, reducing semantic bias and information loss caused by manual fusion. This addresses the problem of difficulty in extracting accurate bearing features in cross-domain scenarios, leading to low fault diagnosis accuracy.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] The above and other objects, features and advantages of this application will become more apparent from the following description of exemplary embodiments of this application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components.

[0020] Figure 1 This is a schematic flowchart illustrating a cross-domain bearing fault diagnosis method according to an embodiment of this application; Figure 2 This is another schematic flowchart illustrating the cross-domain bearing fault diagnosis method shown in the embodiments of this application; Figure 3 This is a schematic diagram of PKGDNet, a cross-domain bearing fault diagnosis architecture based on prior diagnostic knowledge, as shown in an embodiment of this application. Figure 4 This is a structural diagram of DDKAN shown in the embodiments of this application; Figure 5 This is a schematic diagram of the cross-domain bearing fault diagnosis device shown in the embodiments of this application; Figure 6 This is a schematic diagram of one embodiment of the electronic device in this invention. Detailed Implementation

[0021] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0022] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0023] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0024] This invention is primarily applied to rolling bearing fault diagnosis scenarios under complex operating conditions such as rail transportation and aerospace. In these scenarios, bearings operate for extended periods in environments with fluctuating speeds, loads, and strong noise, causing the characteristic frequencies of their vibration signals to drift. Traditional methods based on envelope spectra or pure deep learning struggle to achieve stable and accurate cross-domain diagnosis. This application aims to improve the diagnostic accuracy and robustness of the model in cross-domain scenarios through physical prior guidance and adaptive feature fusion.

[0025] To address the aforementioned issues, this application provides a cross-domain bearing fault diagnosis method, apparatus, and device. Through envelope spectrum extraction and coordinate alignment processing, it maps fault feature frequencies under different operating conditions to a unified physical semantic coordinate system based on the structural parameters and theoretical characteristic frequencies of the target bearing, generating a normalized spectrum. This process eliminates spectral peak position drift caused by differences in rotational speed, load, or structure, enabling automatic alignment of fault features from different domains at the physical level without manual intervention in filtering parameters, significantly improving the method's adaptability to complex cross-domain operating conditions. Addressing the problem of purely data-driven deep learning methods easily learning irrelevant domain features and experiencing decreased generalization performance in cross-domain scenarios, this invention not only introduces physical prior constraints through the normalized spectrum but also collaboratively extracts time-domain impact features and frequency-domain periodic structural features from the original vibration signal using a dual-domain Kolmogorov-Arnold network, generating dual-domain encoded features. Furthermore, this invention adaptively fuses the normalized spectrum and the dual-domain encoded features through cross-reconstruction constraints, forcing the fused features to simultaneously include physically aligned frequency-domain semantic information and fine-grained dynamic information from the original signal. This fusion mechanism effectively suppresses the tendency of the model to learn features from fault-independent domains, ensuring the discriminativeness and stability of feature extraction in cross-domain scenarios. In summary, this invention overcomes the inherent defects of existing technologies in cross-domain scenarios through the synergistic effect of physical semantic coordinate alignment and adaptive feature fusion, thereby achieving high-precision, highly robust, and interpretable cross-domain bearing fault diagnosis. The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0026] Figure 1 This is a schematic flowchart illustrating a cross-domain bearing fault diagnosis method according to an embodiment of this application.

[0027] See Figure 1 A cross-domain bearing fault diagnosis method, comprising: S110: Acquire the raw vibration signal of the target bearing.

[0028] Understandably, the acquisition of this raw vibration signal can be achieved in various ways. For example, an accelerometer can be directly mounted on the bearing housing, and continuous sampling can be performed using a data acquisition device to obtain vibration data over a period of time. Alternatively, data can be collected at specific time points using a handheld vibration analyzer through periodic inspections. In some applications, embedded sensor systems integrated into the device can also monitor and record vibration signals in real time.

[0029] S120: Extract the envelope spectrum and align the coordinates of the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates.

[0030] In this embodiment, the envelope spectrum extraction is specifically achieved by demodulating the original vibration signal to highlight the impact component in the signal, thereby revealing the periodic impact characteristics caused by internal bearing faults.

[0031] The coordinate alignment process is primarily used to adjust the extracted envelope spectrum, making it comparable across different operating conditions or equipment. This process aims to eliminate characteristic frequency drift caused by changes in operating conditions, enabling spectral lines from different sources to be compared and analyzed within a unified physical semantic coordinate system.

[0032] The output normalized spectrum with unified physical semantic coordinates is spectral data with clearly defined physical meaning on its frequency axes, obtained through envelope spectrum extraction and coordinate alignment, and directly comparable between different samples. This normalized spectrum effectively reflects the bearing's fault characteristics and reduces the impact of cross-domain data differences.

[0033] In practical applications, the original vibration signal can first be bandpass filtered, selecting a fixed narrow frequency range to filter out irrelevant noise and highlight the impact component. Then, a Hilbert transform is performed on the filtered signal to obtain its envelope, and a Fourier transform is performed on the envelope signal to obtain the envelope spectrum. To achieve coordinate alignment, the obtained envelope spectrum can be linearly interpolated or resampled to unify its frequency axis to a preset fixed length and range. For example, all envelope spectra can be scaled to a frequency range of 0 to 100 Hz and sampled into 256 frequency points. In this way, the envelope spectra acquired under different operating conditions have a unified representation on the frequency axis, facilitating subsequent comparison and analysis.

[0034] S130: Perform time-domain and frequency-domain collaborative feature extraction on the original vibration signal to generate dual-domain coded features.

[0035] In this step, the collaborative feature extraction in the time and frequency domains actually involves simultaneously extracting features from both the time and frequency domains of the original vibration signal, and then integrating these features through a certain mechanism to obtain dual-domain coded features. The extracted time-domain features reflect the transient impact characteristics of the signal, while the extracted frequency-domain features reflect the periodic structure of the signal.

[0036] The final output dual-domain encoded features can be a compact representation that can simultaneously characterize the time-domain and frequency-domain characteristics of the original vibration signal, generated by a collaborative feature extraction method in the time and frequency domains.

[0037] In practical applications, the extraction of time-domain features can include calculating statistical quantities such as the root mean square value, peak value, kurtosis, and margin factor of the signal. These statistical quantities can reflect the impulsiveness and energy distribution of the signal.

[0038] Frequency domain feature extraction can include performing a fast Fourier transform on the original vibration signal to obtain its power spectral density, and extracting features such as spectral centroid, spectral variance, and frequency band energy ratio from it.

[0039] To achieve synergy, these time-domain and frequency-domain features can be simply concatenated to form a high-dimensional feature vector. For example, 10 time-domain statistical features and 10 frequency-domain statistical features of the original vibration signal can be calculated, and then these 20 features can be combined into a dual-domain encoded feature vector. This combination method aims to capture fault information in the signal from different perspectives.

[0040] S140: Through cross-reconstruction constraints, normalized spectrum and dual-domain encoded features are adaptively fused to generate fused features.

[0041] In this step, the cross-reconstruction constraint refers to a mechanism introduced in the feature fusion process. By requiring the fused features to reconstruct the original input features, the fusion process is ensured to retain the key information of the original features and to promote information interaction and complementarity between different features.

[0042] In one implementation, a multilayer perceptron network can be constructed, taking normalized spectra and dual-domain encoded features as input. The first layer of this network processes the two types of features separately, generating their respective intermediate representations. These intermediate representations are then concatenated and further processed by subsequent layers to obtain the fused features. To introduce cross-reconstruction constraints, two reconstruction branches can be set at the network's output: one branch attempts to reconstruct the original normalized spectra from the fused features, and the other branch attempts to reconstruct the original dual-domain encoded features. During training, minimizing the reconstruction error guides the learning of the fusion network, thereby ensuring that the fused features retain key information from both original features. For example, a loss function can be defined that includes the classification loss of the fused features and the mean squared error loss of the two reconstruction branches to balance the discriminative power and information preservation ability of the features.

[0043] S150: Based on the fusion characteristics, perform fault diagnosis to obtain the fault diagnosis results of the target bearing.

[0044] In this embodiment, the process of determining whether a bearing has a fault, the type of fault, and the severity of the fault based on the extracted features during fault diagnosis typically involves pattern recognition or classification algorithms.

[0045] Specifically, after obtaining the fused features, they can be input into a classifier for fault type identification. For example, a support vector machine (SVM), decision tree, or a simple fully connected neural network can be used as the classifier. This classifier learns the mapping relationship between the fused features and different fault types, thereby outputting the bearing fault diagnosis result, such as "normal," "inner race fault," "outer race fault," or "rolling element fault."

[0046] This embodiment constructs a bearing fault diagnosis system that can effectively address cross-domain challenges by introducing envelope spectrum coordinate alignment processing, collaborative feature extraction in the time and frequency domains, and an adaptive fusion mechanism based on cross-reconstruction constraints. This method can achieve high-precision bearing feature extraction and fault diagnosis under complex and variable operating conditions, providing strong technical support for the reliable operation of industrial equipment.

[0047] like Figure 2 and 3 As shown, this application proposes a cross-domain bearing fault diagnosis architecture, PKGDNet, based on prior diagnostic knowledge, to achieve fault diagnosis. Figure 3 As shown. This embodiment explicitly incorporates the "observable laws of characteristic frequencies and their harmonics / sidebands in the envelope spectrum" from the bearing failure mechanism into the training and inference process, achieving stable and interpretable cross-domain fault identification. The flowchart of the cross-domain bearing fault diagnosis method shown in this embodiment is available in [link to relevant documentation]. Figure 2 The method specifically includes the following steps: S210: Acquire the raw vibration signal of the target bearing.

[0048] In this step, the original vibration signal of the bearing is continuously collected by an accelerometer installed on the bearing housing.

[0049] S220: Based on the structural parameters of the target bearing and the theoretical characteristic frequencies corresponding to the fault type, the envelope spectrum of the original vibration signal is extracted and the coordinates are aligned to generate a normalized spectrum with unified physical semantic coordinates.

[0050] This step first involves bandpass filtering the original vibration signal, for example, selecting a filter with a fixed bandwidth to highlight the impact component that may be caused by bearing failure. Next, a Hilbert transform is performed on the filtered signal to obtain its envelope, and a Fast Fourier Transform is then performed on the envelope signal to obtain the envelope spectrum. To handle cross-domain characteristics, the frequency axis of this envelope spectrum is uniformly resampled to a preset fixed length; for example, all envelope spectra are adjusted to contain 256 frequency points, with a uniform frequency range of 0-500Hz. Therefore, even under different operating conditions, the generated normalized spectra are directly comparable in the frequency dimension, thereby reducing the impact of operating condition changes on feature recognition.

[0051] It should be noted that the structural parameters of the target bearing refer to a series of values ​​describing the bearing's geometry and physical structure, such as the bearing's inner diameter, outer diameter, roller diameter, number of rollers, and contact angle. These structural parameters can be obtained by actually measuring the bearing's physical dimensions or stored in a database within a digital design or maintenance system.

[0052] The theoretical characteristic frequency corresponding to this fault type refers to the vibration frequency that should theoretically occur when a bearing experiences a specific fault (such as inner ring fault, outer ring fault, rolling element fault, cage fault, etc.), calculated using kinematic formulas based on the bearing's structural parameters and operating speed. These theoretical characteristic frequencies can be derived using standard bearing fault frequency calculation formulas, such as the inner ring fault frequency (BPFI), outer ring fault frequency (BPFO), rolling element fault frequency (BSF), and cage fault frequency (FTF). These formulas typically involve structural parameters such as the bearing's pitch diameter, rolling element diameter, number of rolling elements, and contact angle, as well as the bearing's rotational speed. Alternatively, a database containing common bearing models and their corresponding fault characteristic frequencies can be pre-established for direct lookup during diagnosis, or these theoretical frequencies can be automatically calculated and provided using specialized bearing fault diagnosis software tools.

[0053] Specifically, before extracting the envelope spectrum, the theoretical characteristic frequencies that may occur under each fault mode are accurately calculated using the structural parameters of the target bearing and the preset fault types. For example, an appropriate bandpass filter range can be selected based on these frequencies to highlight the modulation information related to the fault. Subsequently, when aligning the extracted envelope spectrum, these theoretical characteristic frequencies are used as a reference to resample or normalize the frequency axis of the spectrum. This alignment method based on physical semantics ensures that the envelope spectra extracted from different bearings and under different operating conditions have a unified physical meaning in the frequency dimension, that is, a specific frequency position always represents a specific fault mode or its harmonics.

[0054] In practical applications, when diagnosing a deep groove ball bearing fault, the first step is to obtain the bearing's structural parameters, such as inner diameter, outer diameter, rolling element diameter, number of rolling elements, and contact angle. Simultaneously, the possible fault types are determined, such as inner ring fault, outer ring fault, and rolling element fault. Based on these structural parameters and fault type, the corresponding theoretical characteristic frequency can be calculated. For example, if the bearing speed is 1500 rpm, the theoretical characteristic frequency of an inner ring fault, calculated using the standard formula, is f. BPFI The theoretical characteristic frequency of an outer ring fault is f. BPFO The theoretical characteristic frequency of rolling element failure is f. BSF When extracting the envelope spectrum of the original vibration signal, the center frequency of the bandpass filter can be set near these theoretical frequencies, or near their harmonic frequencies, to enhance the manifestation of fault characteristics. After obtaining the preliminary envelope spectrum, the frequency axis of the envelope spectrum is normalized using these theoretical characteristic frequencies as reference points. For example, the theoretical characteristic frequency f can be normalized. BPFI The corresponding spectral line position is mapped to a specific coordinate point in the normalized spectrum, so that no matter how the bearing speed changes, this coordinate point always represents the characteristics of an inner ring fault.

[0055] In another embodiment, the step of extracting the envelope spectrum and aligning the coordinates of the original vibration signal based on the structural parameters of the target bearing and the theoretical characteristic frequencies corresponding to the fault type, to generate a normalized spectrum with unified physical semantic coordinates, includes: For the original vibration signal Perform bandpass filtering to obtain the filtered signal. ; for the filtered signal Perform Hilbert transform and calculate the envelope to obtain the envelope signal. ; for the envelope signal After removing the DC component and performing a windowed Fast Fourier Transform, the envelope spectrum amplitude is obtained. ,in Based on the geometric parameters, fault type, and signal sampling parameters of the target bearing, the fault characteristic frequency is calculated within the envelope spectrum amplitude. The corresponding theoretical frequency position Based on the theoretical frequency point location For the envelope spectrum amplitude By performing truncation and resampling, a normalized spectrum with unified physical semantic coordinates is obtained. Wherein, the theoretical frequency point position Calculated using the following formula: ,in, The characteristic frequency constant is determined by the geometric parameters and fault type of the target bearing. This refers to the number of sampling points per revolution when performing equal-angle sampling on the original vibration signal.

[0056] For example, firstly, the original vibration signal can be... A fourth-order Butterworth bandpass filter is applied, the passband frequency range of which can be selected according to the bearing speed and the expected fault frequency range (e.g., set to 500 Hz to 5000 Hz) to obtain the filtered signal. .

[0057] Subsequently, the filtered signal Perform a discrete Hilbert transform and calculate the magnitude of its analytic signal to obtain the envelope signal e[n].

[0058] To further process the signal, the mean of the envelope signal e[n] is removed to eliminate the DC component. Then, a Hanning window function is applied, and a Fast Fourier Transform (FFT) is performed to obtain the envelope spectrum amplitude E[k]. As a specific implementation, assume the bearing experiences an outer race failure. Based on the bearing's pitch circle diameter, roller diameter, number of rollers, and contact angle, as well as the signal sampling frequency and the number of FFT points, the theoretical characteristic frequency (BPFO) of the outer race failure can be accurately calculated and mapped to a specific frequency position L in the envelope spectrum amplitude E[k]. Based on this theoretical frequency position L, the envelope spectrum amplitude E[k] can be truncated; for example, frequencies within a four-harmonic range from zero frequency to the theoretical frequency position L can be truncated to form a frequency sequence.

[0059] Finally, the frequency sequence is resampled using cubic spline interpolation to uniformly adjust it to a preset fixed length (e.g., 1024 points), thereby generating a normalized spectrum with unified physical semantic coordinates. .

[0060] Furthermore, the position based on the theoretical frequency point For the envelope spectrum amplitude By performing truncation and resampling, a normalized spectrum with unified physical semantic coordinates is obtained. ,include: With the theoretical frequency position Based on this, the envelope spectrum amplitude is extracted. The former A frequency point sequence is generated by identifying several frequency points; the frequency point sequence is then resampled using cubic spline interpolation to generate a sequence of length [length missing]. Normalized spectrum ,in, This is a preset fixed value.

[0061] Suppose that during fault diagnosis of a target bearing, the theoretical frequency position L calculated based on its structural parameters and outer ring fault type is 150. At this point, the system will extract the first four values ​​from the original envelope spectrum amplitude E[k]. 150 = 600 frequency points, forming a preliminary frequency point sequence. To ensure that the envelope spectrum of all samples has a uniform length, a preset fixed value is used. The value is 128. Therefore, the frequency sequence of length 600 above will be resampled using cubic spline interpolation to finally generate a sequence of length 4. Normalized spectrum of 128=512 .

[0062] The following is combined Figure 3 The framework in the document provides a detailed explanation of the principles for generating normalized spectra with unified physical semantic coordinates.

[0063] Figure 3 The PKGDNet framework provided in the paper mainly consists of three parts: SPNM, DDKAN, and CRDFM. Among them, SPNM normalizes the spectrum into prior diagnostic knowledge. First, it uses the original vibration signal... Construct the envelope spectrum, and based on the theoretical characteristic frequencies, given the bearing type. The fault-related peaks were aligned to a unified spectral axis position L, and then the normalized spectrum was obtained through truncation, interpolation, and resampling. .because The coordinate system has a clear physical semantics: when a corresponding fault exists, the characteristic peaks and their overtones / sidebands will stably appear near fixed positions; when no corresponding fault exists, no significant peaks will appear near these positions. Therefore, SPNM can not only significantly alleviate peak position drift and scale differences under cross-domain conditions, but also constrain the model discrimination to the vicinity of interpretable prior frequency bands, enabling diagnostic conclusions to be directly physically interpreted from the peak modes on the normalized spectrum.

[0064] However, spectral normalization involves truncation, resampling, and amplitude normalization, which may compress or weaken some fault-related details. To compensate for this potential information loss, this application further introduces more comprehensive dual-domain representation learning from the original signal, using the proposed DDKAN... Modeling is performed in both the time and frequency domains to output dual-domain encoded features. Among them, the time-domain branch focuses on preserving the original dynamic details such as impact, amplitude modulation, and non-stationary abrupt changes, and is more sensitive to early faults and strong noise scenarios; while the frequency-domain branch emphasizes periodic structure and spectral distribution characteristics, providing complementary information for time-domain characterization.

[0065] At the level of cross-domain alignment and information complementarity, relying solely on Stable prior alignment coordinates can be obtained, but it is difficult to fully preserve the fine-grained dynamic information in the original signal; while relying solely on While it retains rich details, it may be affected by domain differences. Therefore, this application designs CRDFM to... and As input, the fused features, guided by cross-reconstruction and decoupling constraints, simultaneously contain key information from two modalities: on the one hand, inheriting... The resulting cross-domain peak alignment and physical interpretability, on the other hand, from It compensates for the time / frequency domain details that may be lost during normalization, thereby improving the robustness and generalization ability of diagnosis in complex working conditions and cross-domain scenarios.

[0066] In summary, the PKGDNet framework provided in this application takes prior diagnostic knowledge as its core, utilizes SPNM to provide a cross-domain stable and interpretable alignment space, DDKAN to provide dual-domain feature representation of the original vibration signal, and CRDFM to achieve complementary fusion of the two through cross-reconstruction. The model is trained end-to-end using a joint optimization objective of classification loss and cross-reconstruction correlation loss, calculated as shown in formula (1): (1) in, This represents the cross-entropy loss, used to supervise the fault category prediction results output by the classifier. Vibration signals reconstructed from decoupled vibration characteristics are constrained. Compared with the original vibration signal Consistency Normalized spectrum reconstructed from decoupled normalized spectral features With normalized spectrum Consistency. Constraints are obtained from the reconstruction of fused features and Consistent Constraints are obtained from the reconstruction of fused features and Consistent.

[0067] In this embodiment, SPNM maps inputs from different domains to a representation space with consistent physical meaning and consistent coordinate alignment, enabling the domain generalization network to learn representations that are both discriminative and domain invariant more efficiently. The calculation method is shown in algorithm (1).

[0068] First, an envelope spectrum is constructed from the original vibration signal to highlight the amplitude modulation characteristics caused by local defects in the rolling bearing. Given a discrete vibration signal... Bandpass filtering is performed to obtain The analytic signal is then obtained through Hilbert transform, and the calculation method is shown in formula (2):

[0069] in This is the envelope. To suppress the influence of the DC component, the envelope can be averaged and then... Then multiply by the window function Then, FFT is performed to obtain the envelope spectrum amplitude, and the calculation method is shown in formula (3).

[0070]

[0071] Envelope spectrum can more clearly present the fault characteristic frequencies and their harmonics / sidebands, laying the foundation for subsequent normalization alignment.

[0072] When N is fixed (e.g., N=4096), L is determined only by c and P. Therefore, as long as the bearing type and fault type are determined, the envelope spectrum under different operating conditions can be extracted into a coordinate system with a unified semantics (if motor-end data is used, P can be corrected by multiplying by the transmission ratio). Therefore, for the envelope spectrum... Extract the first 4L points to form a subsequence {E[1], E[2], E[4L]}, and resampled using cubic spline interpolation to uniformly transform it into a vector of length 4L′. The resampling operator is The calculation method is shown in formula (5).

[0073]

[0074] Therefore, the characteristic peak position k of the original envelope spectrum Point L′ (where k=1,2,3) will be aligned to respectively k Around position 100, a consistent peak pattern is formed across the domain. When the assumptions do not match, the above fixed positions will not show significant peaks, thus providing a basis for subsequent fault detection. The spectrum normalized... It can be used as part of the input to a domain generalization network, where the normalized normal and fault spectra are... .

[0075] S230: Input the original vibration signal into the dual-domain Kolmogorov-Arnold network for feature extraction to obtain dual-domain encoded features.

[0076] In this step, the dual-domain Kolmogorov-Arnold network includes a time-domain branch network and a frequency-domain branch network. The time-domain branch network is used to extract the time-domain impact features of the original vibration signal, and the frequency-domain branch network is used to extract the frequency-domain periodic structure features of the original vibration signal. It should be noted that this dual-domain Kolmogorov-Arnold network is a novel neural network structure based on the Kolmogorov-Arnold representation theorem. Its core idea is to decompose a multivariable function into a combination of univariate functions. The "dual-domain" aspect refers to the network's ability to simultaneously process signal information from both the time and frequency domains, performing deep learning and encoding of features from different domains through parallel or interleaved methods. Its implementation can include, but is not limited to: constructing two independent Kolmogorov-Arnold networks to process time-domain and frequency-domain data respectively, or designing a network structure sharing some layers, which includes branches specifically for time-domain and frequency-domain processing.

[0077] Specifically, the original vibration signal When using a dual-domain Kolmogorov-Arnold network for feature extraction, the original vibration signal can be processed first. Preprocessing, such as framing or normalization, is performed. The preprocessed signal is then directly input into a time-domain branch network, which can consist of multiple stacked Kolmogorov-Arnold layers, each containing basis function transforms and B-spline function transforms to learn the temporal impulse characteristics of the signal. Simultaneously, the preprocessed signal is converted to a frequency domain representation using a Fast Fourier Transform (FFT), and this frequency domain representation is then input into a frequency-domain branch network. This frequency-domain branch network can also consist of multiple Kolmogorov-Arnold layers to capture periodic structural features in the spectrum. For example, the frequency-domain branch network can process the real and imaginary parts of the frequency domain signal separately to more comprehensively encode frequency domain information. Finally, the features output by the time-domain and frequency-domain branch networks can be concatenated or weighted summed to form the final dual-domain encoded features. .

[0078] In another embodiment, the original vibration signal is... Feature extraction is performed by inputting the data into a dual-domain Kolmogorov-Arnold network to obtain dual-domain encoded features. ,include: For the original vibration signal Dimensional adjustment is performed to obtain time-domain input features and frequency-domain input features; the time-domain input features are then input into the time-domain branch network for processing to obtain time-domain features. The frequency domain input features are transformed to the frequency domain using a Fast Fourier Transform to obtain the frequency domain complex features. ; respectively the frequency domain complex features real part and the virtual part The input is processed by the frequency domain branch network to obtain the processed real part. and the virtual part and the real part and the imaginary part Combinations with complex characteristics ; for the complex number features Perform an inverse fast Fourier transform to obtain the frequency domain features. ; the frequency domain features With the aforementioned time-domain features Adding them together yields the dual-domain encoded features. .

[0079] Specifically, in generating dual-domain encoded features At that time, the original vibration signal can be analyzed first. Dimensional adjustments can be made. For example, the original vibration signal can be... It is directly used as a time-domain input feature, and zero-padding is applied to it to achieve a preset length. Simultaneously, the original vibration signal... After performing a Fast Fourier Transform (FFT), the amplitude and phase spectra are taken as frequency domain input features. Next, the dimension-adjusted time-domain input features are fed into a time-domain branch network. This time-domain branch network can be composed of multiple stacked one-dimensional convolutional layers, batch normalization layers, and activation function layers, used to extract the impact component in the signal, such as transient pulses caused by damage to bearing rollers or inner / outer races, thus obtaining the time-domain features. Simultaneously, the frequency domain input features are converted into frequency domain complex features using a Fast Fourier Transform (FFT). Then, the complex features in the frequency domain are... real part and the virtual part The inputs are fed into a frequency domain branching network. This network can consist of multiple fully connected layers, each followed by an activation function to learn and extract periodic harmonic components from the signal, such as characteristic frequencies and their harmonics caused by faults, thereby obtaining the processed real part. and the virtual part Subsequently, and Combining into complex features Next, we will discuss the characteristics of complex numbers. Perform an inverse fast Fourier transform (IFFT) to convert it back to the time domain representation, thus obtaining the frequency domain features. Finally, the frequency domain features With time domain characteristics Element-wise addition is performed to obtain the final dual-domain encoded features. .

[0080] Furthermore, the function transformations of both the time-domain branch network and the frequency-domain branch network include basis function transformation and B-spline function transformation; The formula for calculating the basis function transformation is as follows: ,in This is the weight matrix. For activation functions; The formula for calculating the B-spline function transformation is as follows: ,in For learnable weights, for The B-order spline basis function is recursively defined as: , , in, To define in the interval Uniform grid points on the surface.

[0081] like Figure 4 The diagram shown is a detailed structural diagram of DDKAN. The KAN component in DDKAN is primarily based on the Kolmogorov-Arnold theorem, which states that any multivariate continuous function can represent a combination of single-variable continuous functions. KAN mainly employs basis function transformations and B-spline function transformations, thereby capturing complex relationships while maintaining local plasticity, allowing the model to adapt to new inputs without overwriting previously learned information. The basis transformation captures the main features in the data through linear mapping and nonlinear activation functions, and its calculation method is as follows: .

[0082] B-spline function transformation performs balanced interpolation between data points, thus enabling flexible modeling of complex feature representations. It utilizes a uniformly spaced grid. The resolution of spline interpolation is controlled, where s represents the number of interpolation points and k is the spline order. The vectors are uniformly distributed on the interval [-1, 1]. Therefore, the B-spline transformation is calculated as follows: , in Let i represent the spline of degree k. Finally, KAN will combine... .

[0083] DDKAN first applies a Dimension Ajustment Strategy (DAS) to the vibration signal at each time step t. Dimension adjustments were performed. Different modifications were made to the frequency domain branch and the time minute. In the frequency domain branch, Multiply by a learnable weight vector This generates a hidden feature representation of frequency features. Where d is set to 128, for the time branch, the original vibration input This remains unchanged, thus ensuring efficient processing in both the time and frequency domains. Finally, DDKAN will respectively... and Input frequency domain branch and time domain branch.

[0084] Frequency domain branching first used FFT to... Frequency domain features are obtained by converting from the time domain to the frequency domain. Then, the real part is processed using a two-layer KAN. and the virtual part The real part is obtained respectively. and the virtual part Using the same KAN network for both the real and imaginary parts ensures consistent feature learning and parameter sharing for meaningful signal reconstruction. Furthermore, KAN allows for more efficient fusion of local and imaginary parts. The real and imaginary values ​​are then added and denoted as... The final complex number representation is formed. Finally, IFFT is used to map the frequency domain data to the time domain. Therefore, the calculation method of the frequency domain branch is as shown in formula (8).

[0085] (8) Where v is the integral variable and f is the frequency variable. The imaginary unit, express The real part, denoted as express The imaginary part is denoted as ,matrix and Let represent the weight matrices of the basis functions and the B-spline function, respectively, where s represents the grid size and k represents the spline order.

[0086] In the time domain branch, DDKAN captures the inherent temporal characteristics and dependencies of time series data through KAN, and the calculation method is shown in formula (9).

[0087]

[0088] in This refers to the data after the KAN's time-domain branch has been operated on along the channel dimension. and These are the learnable parameters. Finally, to combine the information from the frequency domain branch and the time branch, we summed their results and denoted them as... .

[0089] S240: Input the dual-domain encoded features into the vibration signal encoder to obtain the vibration signal features.

[0090] S250: Input the normalized spectrum into the normalized spectrum encoder to obtain the normalized spectrum features.

[0091] S260: Input the vibration signal features and normalized spectrum features into the feature fusion module for concatenation and convolution to obtain fused features.

[0092] In this embodiment, the dynamic feature fusion network based on cross-reconstruction includes a vibration signal encoder, a normalized spectrum encoder, and a feature fusion module; the normalized spectrum Ib and the dual-domain encoded feature Ks are input into the dynamic feature fusion network based on cross-reconstruction, and fused through cross-reconstruction loss constraints to obtain fused features.

[0093] It should be noted that the network can contain two independent encoder branches, one for processing the normalized spectrum. Another one is used to process dual-domain encoded features. Each encoder branch can consist of multiple convolutional layers, pooling layers, and fully connected layers to extract deep representations of its respective features. After the encoder output, a fusion layer can be added, for example, by concatenating the output features of the two branches, followed by further feature transformation and fusion through one or more fully connected layers. To achieve cross-reconstruction, the network can also include two decoder branches. The first decoder branch receives the fused features as input and attempts to reconstruct the original normalized spectrum. The second decoder branch also receives the fused features as input and attempts to reconstruct the original dual-domain encoded features. During training, the reconstructed normalized spectrum and the original normalized spectrum can be calculated. The mean squared error loss between the two domains, and the difference between the reconstructed dual-domain encoded features and the original dual-domain encoded features. The mean squared error between the two loss functions constitutes the cross-reconstruction loss constraint, used for backpropagation to update the network weights, thereby enabling the network to learn an effective feature fusion strategy. For example, in the early stages of training, the network may tend to simply pass information, but as training progresses, the cross-reconstruction loss will prompt the network to learn how to recover the key information of the two original features from the fused features, thus ensuring the comprehensiveness and effectiveness of the fused features.

[0094] In this embodiment, dual-domain coding features The signal is first input to a vibration signal encoder for processing. This encoder focuses on extracting deep, abstract vibration signal features from the combined features of the time and frequency domains. Meanwhile, normalized spectrum The data is fed into a normalized spectral encoder, which focuses on extracting normalized spectral features with physical semantics from the frequency spectrum data. These two independent encoders allow for optimized processing of data tailored to the characteristics of different modalities, ensuring that key information for each mode is effectively encoded. Subsequently, the characteristics of these encoded vibration signals are... and normalized spectral characteristics The data is fed into the feature fusion module. In this module, the two features are first concatenated, combining them along the feature dimension so that the network can simultaneously perceive information from both modalities. Next, a convolutional operation is performed on the concatenated features. The convolutional layer learns the complex nonlinear relationships and interactions between features from different modalities, thereby generating a unified fused feature that contains complementary information from both modalities.

[0095] For example, the vibration signal encoder can be composed of a neural network containing three fully connected layers, each followed by a ReLU activation function to introduce a nonlinear transformation. The normalized spectrum encoder can be a one-dimensional convolutional neural network containing two convolutional layers, each followed by a batch normalization layer and a ReLU activation function, and finally a global average pooling layer to unify the feature dimensions. In the feature fusion module, when the vibration signal features... and normalized spectral characteristics After being generated, they are first concatenated along the channel dimension. For example, if The dimensions are (Batch_size, Channels_s, Length_s) and The dimensions are (Batch_size, Channels_b, Length_b). Before concatenation, additional fully connected or convolutional layers may be needed to adjust their Length dimension to be consistent. Then, they are concatenated along the channel dimension to form a feature map with dimensions (Batch_size, Channels_s + Channels_b, Length). Subsequently, this concatenated feature map is fed into a convolutional block, which may contain a 1x1 convolutional layer for initial integration of feature channels, followed by a batch normalization layer and a ReLU activation function, and finally a 3x3 convolutional layer for further extraction of the fused deep features, thus obtaining the final fused feature.

[0096] Specifically, the vibration signal characteristics and the normalized spectral features The input to the feature fusion module is subjected to a convolutional operation to obtain fused features, including: The vibration signal characteristics With the normalized spectral features The features are concatenated to obtain concatenated features; then, a convolution operation is performed on the concatenated features to obtain intermediate features. ; for the intermediate features Two independent Convolution operations yield the first weight map. Second weighted graph The vibration signal characteristics With the first weighted graph Perform element-wise multiplication to normalize the spectral features. With the second weight graph Perform element-wise multiplication; concatenate the weighted features from the two paths again and perform a convolution operation to obtain the fused features. .

[0097] In another embodiment, the dynamic feature fusion network based on cross-reconstruction further includes a cross-task interaction module, a decoupled encoder, and a reconstruction decoder; The vibration signal characteristics and the normalized spectral features After the feature fusion module performs a concatenation and convolution operation to obtain the fused features, the following steps are also included: The vibration signal characteristics and the normalized spectral features The input is processed by the multi-task fusion module to perform cross-task interaction and feature fusion, resulting in fused task-related features. semantically related features The vibration signal characteristics and the normalized spectral features By inputting the decoupled encoder, the decoupling characteristics of the vibration signal are obtained. Decoupling characteristics of normalized spectra The vibration signal characteristics The normalized spectral features The decoupling characteristics of the vibration signal Decoupling characteristics of the normalized spectrum The first vibration signal is reconstructed by inputting it into the reconstruction decoder. Second vibration signal First normalized spectrum Second normalized spectrum Based on the first vibration signal The second vibration signal The first normalized spectrum and the second normalized spectrum Calculate the original vibration signal and the normalized spectrum The reconstruction loss between them is used to reshape the fused features based on the reconstruction loss.

[0098] Specifically, in vibration signal characteristics and normalized spectral characteristics After being concatenated and convolved by the feature fusion module, these features are then fed into the cross-task interaction module and the decoupled encoder. The cross-task interaction module is responsible for facilitating... and Effective information exchange and alignment related to fault diagnosis tasks between the components, thereby generating fused task-related features. semantically related features This helps extract and enhance information crucial for diagnosis. Simultaneously, the decoupling encoder... and Processing is performed to obtain the decoupling characteristics of the vibration signal. Decoupling characteristics of normalized spectra This step aims to isolate domain-specific variations or noise from the features that are irrelevant to the diagnostic task, thereby improving the purity and cross-domain generalization ability of the features. Subsequently, the reconstruction decoder receives the original features. , and their decoupling characteristics , As input, and attempt to reconstruct the original vibration signal. and normalized spectrum Generate the first vibration signal Second vibration signal First normalized spectrum Second normalized spectrum By calculating these reconstruction results and comparing them with the original input... and The reconstruction loss between the two features is used for backpropagation to dynamically adjust and reshape the previously generated fused features. This mechanism forces the network to learn feature representations that effectively preserve the original information and achieve good decoupling, ensuring that the fused features not only contain rich diagnostic information but also have stronger robustness and interpretability. This overcomes the problems of feature entanglement and insufficient generalization ability that may result from simple feature splicing and fusion.

[0099] Wherein, the total loss function The calculation formula is: , in, The cross-entropy loss is the output of the classifier; The vibration signal reconstructed by the first reconstruction decoder Compared with the original vibration signal The reconstruction losses between them; The normalized spectrum reconstructed by the first reconstruction decoder Normalized spectrum of input The reconstruction losses between them; The vibration signal reconstructed by the second reconstruction decoder Compared with the original vibration signal The reconstruction losses between them; Normalization for reconstruction by the second reconstruction decoder Normalized spectrum of input The reconstruction losses between them.

[0100] Furthermore, the reconstruction loss , , , Both methods employ a combination of L1 norm loss and structural similarity loss, calculated using the following formula:

[0101]

[0102] in, This represents the structural similarity index.

[0103] It should be noted that in CRDFM, the original vibration signal is used... With normalized spectrum The algorithm learns the inverse mapping and forces the fused original vibration signal and normalized spectrum to contain key features of both modes. This compensates for information loss that may occur during normalization while maintaining the advantages of cross-domain alignment, thereby improving the robustness and generalization ability of fault detection under complex conditions. First, given the input, the output of DDKAN... and normalized spectrum Input vibration signal encoders respectively and normalized spectrum encoder Extracting vibration signal features and normalized spectral characteristics Where j=1,2,3 represents the number of encoders. and It contains a 3×3 convolutional layer and a residual layer. The residual layer consists of two 3×3 convolutional layers and skip connections. express Features encoded using DDKAN. Then, the output of each encoder... and The inputs are respectively fed into the reconstruction decoder. Reconstructing vibration signals and normalized spectrum .in It includes a 3×3 convolutional layer, residual blocks, and upsampling, with upsampling consisting of bilinear interpolation and a 1×1 convolutional layer.

[0104] Furthermore, by and Input to the Multi-Task Fusion Module (MIFM) to obtain the fused relevant features. and the characteristics of vibration signals and normalized spectral characteristics Input to decoupled encoder Decoupling characteristics of generating vibration signals Decoupling characteristics of normalized spectra .in Includes and . By and The input is fed into two 1×1 convolutions to obtain the enhancement coefficients, and then... and Element-wise multiplication is performed to obtain the emphasized features, and finally... Characteristic obtained by element-wise subtraction and .

[0105] Finally, output each decoupled encoder. and The inputs are respectively fed into the reconstruction decoder. Reconstructing vibration signals and normalized spectrum .in and The structures are the same.

[0106] S270: Perform fault diagnosis based on the fusion characteristics to obtain the fault diagnosis result of the target bearing.

[0107] In this embodiment, fault-related peaks are aligned to unified physical semantic coordinates through spectrum normalization, mitigating domain shift caused by peak drift and scale differences, and significantly improving the interpretability of diagnostic results. The dual-domain Kolmogorov-Arnold network preserves both temporal impact / transient details and frequency domain periodic structure information, enhancing discrimination stability under strong noise and early weak fault conditions. The dynamic feature fusion network based on cross-reconstruction uses reversible cross-reconstruction as a constraint to adaptively fuse normalized spectrum and vibration signal modal features, complementing the frequency domain semantics with the dynamic features of the original vibration signal, thereby avoiding semantic inconsistencies and information omissions caused by manually setting fusion rules or fusion losses.

[0108] The cross-domain bearing fault diagnosis method in the embodiments of the present invention has been described above. The cross-domain bearing fault diagnosis device in the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 5 One embodiment of the cross-domain bearing fault diagnosis device of the present invention includes: The acquisition module 510 is used to acquire the original vibration signal of the target bearing; The normalization module 520 is used to extract the envelope spectrum and perform coordinate alignment processing on the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates. The extraction module 530 is used to perform time-domain and frequency-domain collaborative feature extraction on the original vibration signal to generate dual-domain coded features; The fusion module 540 is used to adaptively fuse the normalized spectrum and the dual-domain encoded features through cross-reconstruction constraints to generate fused features; The diagnostic module 550 is used to perform fault diagnosis based on the fusion features to obtain the fault diagnosis result of the target bearing.

[0109] The method provided in this embodiment obtains the original vibration signal, generates a normalized spectrum and dual-domain encoded features with unified physical semantic coordinates, generates fused features through adaptive fusion, and performs fault diagnosis, thus solving the problem of unstable fault feature extraction under cross-domain operating conditions.

[0110] The aforementioned cross-domain bearing fault diagnosis device can be implemented as a computer program, which can, for example... Figure 6 The electronic device shown operates on the device. The electronic device includes a processor 600 and a memory 601, the memory 601 storing machine-executable instructions that can be executed by the processor 600 to implement the aforementioned cross-domain bearing fault diagnosis method.

[0111] Furthermore, Figure 6 The electronic device shown also includes a bus 602 and a communication interface 603. The processor 600, the communication interface 603 and the memory 601 are connected via the bus 602.

[0112] The memory 601 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 603 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 602 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0113] The processor 600 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 600 or by instructions in software form. The processor 600 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 601. Processor 600 reads the information in memory 601 and, in conjunction with its hardware, completes the method steps of the aforementioned embodiment.

[0114] The present invention also provides an electronic device, the computer device including a memory and a processor, the memory storing computer-readable instructions, which, when executed by the processor, cause the processor to perform the steps of the cross-domain bearing fault diagnosis method provided in the above embodiments.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0116] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-domain bearing fault diagnosis method, characterized in that, The method includes: Obtain the original vibration signal of the target bearing; The original vibration signal is subjected to envelope spectrum extraction and coordinate alignment processing to generate a normalized spectrum with unified physical semantic coordinates; The original vibration signal is subjected to time-domain and frequency-domain collaborative feature extraction to generate dual-domain coded features; By using cross-reconstruction constraints, the normalized spectrum and the dual-domain encoded features are adaptively fused to generate fused features; Based on the fusion characteristics, fault diagnosis is performed to obtain the fault diagnosis results of the target bearing.

2. The cross-domain bearing fault diagnosis method according to claim 1, characterized in that, The process of extracting the envelope spectrum and aligning the coordinates of the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates includes: Based on the structural parameters of the target bearing and the theoretical characteristic frequencies corresponding to the fault types, the original vibration signal is subjected to envelope spectrum extraction and coordinate alignment processing to generate a normalized spectrum with unified physical semantic coordinates.

3. The cross-domain bearing fault diagnosis method according to claim 2, characterized in that, The step of extracting the envelope spectrum and aligning the coordinates of the original vibration signal based on the structural parameters and theoretical characteristic frequencies corresponding to the fault type of the target bearing to generate a normalized spectrum with unified physical semantic coordinates includes: For the original vibration signal Perform bandpass filtering to obtain the filtered signal. ; For the filtered signal Perform Hilbert transform and calculate the envelope to obtain the envelope signal. ; For the envelope signal After removing the DC component and performing a windowed Fast Fourier Transform, the envelope spectrum amplitude is obtained. ,in ; Based on the geometric parameters, fault type, and signal sampling parameters of the target bearing, the fault characteristic frequency is calculated at the envelope spectrum amplitude. The corresponding theoretical frequency position ; Based on the theoretical frequency point position For the envelope spectrum amplitude By performing truncation and resampling, a normalized spectrum with unified physical semantic coordinates is obtained. .

4. The cross-domain bearing fault diagnosis method according to claim 1, characterized in that, The step of performing time-domain and frequency-domain collaborative feature extraction on the original vibration signal to generate dual-domain coded features includes: The original vibration signal Feature extraction is performed by inputting the data into a dual-domain Kolmogorov-Arnold network to obtain dual-domain encoded features. The dual-domain Kolmogorov-Arnold network includes a time-domain branch network and a frequency-domain branch network. The time-domain branch network is used to extract the time-domain impact features of the original vibration signal, and the frequency-domain branch network is used to extract the frequency-domain periodic structure features of the original vibration signal.

5. The cross-domain bearing fault diagnosis method according to claim 4, characterized in that, The original vibration signal Feature extraction is performed by inputting the data into a dual-domain Kolmogorov-Arnold network to obtain dual-domain encoded features. ,include: For the original vibration signal Dimensional adjustment is performed to obtain time-domain input features and frequency-domain input features; The time-domain input features are input into the time-domain branch network for processing to obtain time-domain features. ; The frequency domain input features are transformed to the frequency domain using a Fast Fourier Transform to obtain the frequency domain complex features. ; The frequency domain complex features are respectively real part and the virtual part The input is processed by the frequency domain branch network to obtain the processed real part. and the virtual part and the real part and the imaginary part Combinations with complex characteristics ; For the complex features Perform an inverse fast Fourier transform to obtain the frequency domain features. ; The frequency domain features With the aforementioned time-domain features Adding them together yields the dual-domain encoded features. .

6. The cross-domain bearing fault diagnosis method according to claim 1, characterized in that, The step of adaptively fusing the normalized spectrum and the dual-domain encoded features through cross-reconstruction constraints to generate fused features includes: The normalized spectrum and the dual-domain coding features Input a dynamic feature fusion network based on cross-reconstruction, and fuse features through cross-reconstruction loss constraints to obtain fused features.

7. The cross-domain bearing fault diagnosis method according to claim 6, characterized in that, The dynamic feature fusion network based on cross-reconstruction includes a vibration signal encoder, a normalized spectrum encoder, and a feature fusion module; The normalized spectrum and the dual-domain coding features The input is a dynamic feature fusion network based on cross-reconstruction. The fused features are obtained through cross-reconstruction loss constraints, including: The dual-domain encoded features Input the vibration signal encoder to obtain vibration signal characteristics. ; The normalized spectrum Input the normalized spectral encoder to obtain normalized spectral features. ,in Indicates the encoder's level index; The vibration signal characteristics and the normalized spectral features The input feature fusion module performs a convolution operation to obtain fused features.

8. The cross-domain bearing fault diagnosis method according to claim 7, characterized in that, The dynamic feature fusion network based on cross-reconstruction also includes a cross-task interaction module, a decoupled encoder, and a reconstruction decoder; The vibration signal characteristics and the normalized spectral features After the feature fusion module performs a concatenation and convolution operation to obtain the fused features, the following steps are also included: The vibration signal characteristics and the normalized spectral features The input is processed by the multi-task fusion module to perform cross-task interaction and feature fusion, resulting in fused task-related features. semantically related features ; The vibration signal characteristics and the normalized spectral features By inputting the decoupled encoder, the decoupling characteristics of the vibration signal are obtained. Decoupling characteristics of normalized spectra ; The vibration signal characteristics The normalized spectral features The decoupling characteristics of the vibration signal Decoupling characteristics of the normalized spectrum The first vibration signal is reconstructed by inputting it into the reconstruction decoder. Second vibration signal First normalized spectrum Second normalized spectrum ; Based on the first vibration signal The second vibration signal The first normalized spectrum and the second normalized spectrum Calculate the original vibration signal and the normalized spectrum The reconstruction loss between them is used to reshape the fused features based on the reconstruction loss.

9. A cross-domain bearing fault diagnosis device, characterized in that, The device includes: The acquisition module is used to acquire the raw vibration signal of the target bearing; The normalization module is used to extract the envelope spectrum and align the coordinates of the original vibration signal to generate a normalized spectrum with unified physical semantic coordinates. The extraction module is used to perform time-domain and frequency-domain collaborative feature extraction on the original vibration signal to generate dual-domain coded features; The fusion module is used to adaptively fuse the normalized spectrum and the dual-domain encoded features through cross-reconstruction constraints to generate fused features; The diagnostic module is used to perform fault diagnosis based on the fusion features to obtain the fault diagnosis result of the target bearing.

10. An electronic device, characterized in that, The electronic device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the electronic device to perform the cross-domain bearing fault diagnosis method as described in any one of claims 1-8.

Images