Generator fault identification method, equipment, product and medium

Abstract

A generator fault identification method, device, product and medium relate to the technical field of generator fault category identification. The method comprises the following steps: acquiring a vibration signal and a noise signal; performing multi-mode decomposition processing on the vibration signal and the noise signal to obtain a first sub-mode component and a second sub-mode component; determining a vibration signal sequence based on each first sub-mode component and the vibration signal; determining a noise signal sequence based on each second sub-mode component and the noise signal; determining a vibration characteristic matrix based on the vibration signal sequence; determining a noise feature matrix based on the noise signal sequence; determining a first recognition result and a first confidence coefficient based on the vibration feature matrix; determining a second recognition result and a second confidence coefficient based on the noise feature matrix; when the identification results are inconsistent, determining a fault identification result according to the confidence coefficient; and when the identification results are consistent, determining the first identification result as a fault identification result. The method is advantaged in that generator fault identification accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of generator fault classification, and particularly relates to a generator fault identification method, device, product and medium. BACKGROUND

[0002] In modern industrial production and social life, as the core equipment for converting other forms of energy into electric energy, generators are widely used in power, transportation, manufacturing and many other fields, and their stable operation is directly related to the safety and reliability of the entire power system. However, due to the complex operating environment, high working intensity and precise internal structure of the generator, various faults are inevitable in the long-term operation process, among which electromagnetic interference and mechanical looseness are two common and significant fault types.

[0003] At present, in the field of generator fault diagnosis, a large-scale hydro-generator unit fault diagnosis method is disclosed in a Chinese patent with the publication number CN112879200A. This method collects vibration signals of the generator under normal and fault conditions, converts the time-domain signals into frequency-domain signals using fast spectral kurtosis, and then combines a stacked sparse autoencoder to classify and identify the fault type.

[0004] However, in actual application, due to the complex working environment of the generator, its vibration signals are easily affected by external factors such as environmental noise and load changes. Therefore, relying solely on a single vibration signal for fault diagnosis can easily lead to low accuracy in generator fault identification. SUMMARY

[0005] The present application provides a generator fault identification method, device, product and medium, which has the effect of improving the accuracy of generator fault identification.

[0006] In a first aspect of the present application, a generator fault identification method is provided, which specifically includes: obtaining a vibration signal and a noise signal of a target generator; performing multi-modal decomposition processing on the vibration signal to obtain a plurality of first sub-modal components, and performing multi-modal decomposition processing on the noise signal to obtain a plurality of second sub-modal components; determining a vibration signal sequence based on each first sub-modal component and the vibration signal, and determining a noise signal sequence based on each second sub-modal component and the noise signal; determining a vibration feature matrix based on the vibration signal sequence, and determining a noise feature matrix based on the noise signal sequence; determining a first identification result and a first confidence based on the vibration feature matrix, and determining a second identification result and a second confidence based on the noise feature matrix; When the first identification result is inconsistent with the second identification result, the fault identification result is determined based on the first confidence level and the second confidence level. When the first identification result matches the second identification result, the first identification result is determined as the fault identification result.

[0007] By adopting the above technical solution, the vibration and noise signals of the target generator are obtained, providing a dual data source for subsequent processing. Multimodal decomposition is performed on the vibration and noise signals respectively to obtain multiple first sub-mode components and multiple second sub-mode components, thereby decomposing the original signal into components with different frequency characteristics. This effectively separates useful information and interference components from the signal. Then, based on each first sub-mode component and the vibration signal, a vibration signal sequence is determined; based on each second sub-mode component and the noise signal, a noise signal sequence is determined. Coarse-grained processing reduces the complexity of the signal while retaining key feature information, reducing the computational burden of subsequent processing. Based on this, a vibration feature matrix is ​​determined based on the vibration signal sequence, and a noise signal sequence is determined based on the noise signal sequence. By determining the noise feature matrix, the time-domain signal is converted into a structured feature representation, which facilitates subsequent fault mode identification and analysis. Then, based on the vibration feature matrix, a first identification result and a first confidence level are determined, and based on the noise feature matrix, a second identification result and a second confidence level are determined, forming independent judgment results based on different signal sources. This provides a multi-dimensional judgment basis for the final decision. Finally, when the first identification result and the second identification result are inconsistent, the fault identification result is determined according to the first confidence level and the second confidence level. When the first identification result and the second identification result are consistent, the first identification result is determined as the fault identification result. The confidence-weighted fusion decision strategy avoids the risk of misjudgment from a single signal source and significantly improves the accuracy of generator fault identification.

[0008] Optionally, the multimodal decomposition processing of the vibration signal to obtain multiple first sub-mode components and the multimodal decomposition processing of the noise signal to obtain multiple second sub-mode components specifically include: Based on the preset number of initial modes, the vibration signal and noise signal are decomposed for the first time to obtain multiple sets of first intrinsic mode function components corresponding to the vibration signal and multiple sets of second intrinsic mode function components corresponding to the noise signal. A first penalty factor for the vibration signal is determined based on the first intrinsic mode function component; a second penalty factor for the noise signal is determined based on the second intrinsic mode function component. The vibration signal is decomposed a second time based on the first penalty factor to obtain multiple sets of first decomposition components; the noise signal is decomposed a second time using the second penalty factor to obtain multiple sets of second decomposition components. Calculate the correlation coefficient of adjacent components in each group of first decomposition components, and obtain the average correlation coefficient of each group of first decomposition components. The number of modes with the smallest average correlation coefficient is taken as the number of first modes of the vibration signal. Calculate the correlation coefficient of adjacent components in each group of second decomposition components to obtain the average correlation coefficient of each group of second decomposition components, and take the number of modes with the smallest average correlation coefficient as the number of second modes of the noise signal. The vibration signal is decomposed a third time based on the first penalty factor and the number of first modes to obtain multiple first sub-mode components; the noise signal is decomposed a third time based on the second penalty factor and the number of second modes to obtain multiple second sub-mode components.

[0009] By adopting the above technical solution, the vibration signal and noise signal are decomposed for the first time based on a preset initial number of modes, resulting in multiple sets of first intrinsic mode function components corresponding to the vibration signal and multiple sets of second intrinsic mode function components corresponding to the noise signal. This provides an initial decomposition basis for subsequent parameter optimization. Then, a first penalty factor for the vibration signal is determined based on the first intrinsic mode function components, and a second penalty factor for the noise signal is determined based on the second intrinsic mode function components. This achieves adaptive determination of the penalty factor for different signal characteristics. Therefore, a second decomposition is performed on the vibration signal based on the first penalty factor to obtain multiple sets of first decomposition components, and a second decomposition is performed on the noise signal using the second penalty factor to obtain multiple sets of second decomposition components. The optimized penalty factor improves the decomposition quality. Subsequently, the correlation coefficient between adjacent components in each set of first decomposition components is calculated to obtain the correlation coefficient of each set of first decomposition components. The average correlation coefficient of the first decomposition component is used, and the number of modes with the smallest average correlation coefficient is taken as the number of first modes of the vibration signal. At the same time, the correlation coefficient of adjacent components in each group of second decomposition components is calculated to obtain the average correlation coefficient of each group of second decomposition components. The number of modes with the smallest average correlation coefficient is taken as the number of second modes of the noise signal. The optimal number of modes parameter is determined through correlation analysis, avoiding the problems of over-decomposition and under-decomposition. Finally, the vibration signal is decomposed a third time based on the first penalty factor and the number of first modes to obtain multiple first sub-mode components. The noise signal is decomposed a third time based on the second penalty factor and the number of second modes to obtain multiple second sub-mode components. It can adaptively determine the penalty factor and number of modes required for decomposition according to the characteristics of the signal itself, avoiding the blindness of manually setting parameters and improving the accuracy and robustness of signal decomposition.

[0010] Optionally, determining the first penalty factor of the vibration signal based on the first intrinsic mode function component and the second penalty factor of the noise signal based on the second intrinsic mode function component specifically includes: Select the high-frequency component in the first natural mode function component, calculate the power spectral entropy of the high-frequency component, and take the penalty factor when the power spectral entropy is maximum as the first penalty factor of the vibration signal. Select the high-frequency components in the second intrinsic mode function components, calculate the power spectral entropy of the high-frequency components, and use the penalty factor when the power spectral entropy is at its maximum as the second penalty factor for the noise signal.

[0011] By adopting the above technical solution, high-frequency components in the first intrinsic mode function are selected and their power spectral entropy is calculated. The penalty factor when the power spectral entropy is maximized is used as the first penalty factor for the vibration signal. At the same time, high-frequency components in the second intrinsic mode function are selected and their power spectral entropy is calculated. The penalty factor when the power spectral entropy is maximized is used as the second penalty factor for the noise signal. Through power spectral entropy analysis of high-frequency components, the penalty factor is accurately optimized. Since high-frequency components often contain rich fault feature information, the power spectral entropy maximization criterion ensures the effective preservation of high-frequency fault information during the decomposition process, avoids the loss of important fault features, and thus improves the sensitivity and accuracy of subsequent fault identification. At the same time, independent optimization is performed for the characteristics of vibration signals and noise signals, so that each signal type can obtain the penalty factor parameters most suitable for its spectral characteristics.

[0012] Optionally, determining the vibration signal sequence based on each first sub-mode component and the vibration signal, and determining the noise signal sequence based on each second sub-mode component and the noise signal, specifically includes: Calculate the mutual information between each first sub-mode component and the vibration signal, select the first sub-mode component with the largest mutual information for reconstruction, and obtain the vibration signal sequence; Calculate the mutual information between each second sub-mode component and the noise signal, select the second sub-mode component with the largest mutual information for reconstruction, and obtain the noise signal sequence.

[0013] By adopting the above technical solution, the mutual information between each first sub-mode component and the vibration signal is calculated, and the first sub-mode component with the largest mutual information is selected for reconstruction to obtain the vibration signal sequence. At the same time, the mutual information between each second sub-mode component and the noise signal is calculated, and the second sub-mode component with the largest mutual information is selected for reconstruction to obtain the noise signal sequence. The sub-mode component selection strategy based on the mutual information criterion can accurately identify the component with the strongest correlation to the original signal, ensuring that the reconstructed signal retains the core information and key features of the original signal, and effectively filtering out noise interference and redundant components. This intelligent component screening mechanism not only realizes the dimensionality reduction of the signal, but also ensures the integrity of the fault feature information, avoiding the information loss problem that may occur in traditional methods, thus providing a high-quality signal foundation for subsequent feature extraction and fault identification.

[0014] Optionally, determining the vibration feature matrix based on the vibration signal sequence and the noise feature matrix based on the noise signal sequence specifically includes: The vibration signal sequence is decomposed into multiple levels to obtain multiple vibration level components; the noise signal sequence is decomposed into multiple levels to obtain multiple noise level components. Sliding window calculations were performed on the vibration hierarchy components at different scales to obtain coarse-grained vibration sequences of the vibration hierarchy components at different scales; sliding window calculations were also performed on the noise hierarchy components at different scales to obtain coarse-grained noise sequences of the noise hierarchy components at different scales. The fuzzy entropy value of the coarse-grained vibration sequence is calculated, and the fuzzy entropy values ​​of multiple vibration level components at different scales are integrated to obtain the vibration feature matrix; the fuzzy entropy value of the coarse-grained noise sequence is calculated, and the fuzzy entropy values ​​of multiple noise level components at different scales are integrated to obtain the noise feature matrix.

[0015] By employing the above technical solutions, multiple vibration level components are obtained by multi-level decomposition of the vibration signal sequence and multiple noise level components by multi-level decomposition of the noise signal sequence. This enables refined analysis of the signal at different frequency levels, further uncovering the signal's intrinsic structural features. Then, sliding window calculations are performed on the vibration level components at different scales to obtain coarse-grained vibration sequences at different scales, and sliding window calculations are performed on the noise level components at different scales to obtain coarse-grained noise sequences at different scales. Multi-scale analysis captures the dynamic variation patterns of the signal at different time scales, enhancing the comprehensiveness and robustness of feature representation. Subsequently, the fuzzy entropy value of the coarse-grained vibration sequence is calculated, and the fuzzy entropy values ​​of multiple vibration level components at different scales are integrated to obtain the vibration feature matrix. Similarly, the fuzzy entropy value of the coarse-grained noise sequence is calculated, and the fuzzy entropy values ​​of multiple noise level components at different scales are integrated to obtain the noise feature matrix. The feature extraction method based on fuzzy entropy can effectively quantify the complexity and irregularity of the signal, forming a multi-dimensional feature representation with rich information and strong discriminative power.

[0016] Optionally, determining the first identification result and the first confidence level based on the vibration feature matrix, and determining the second identification result and the second confidence level based on the noise feature matrix, specifically includes: The vibration feature matrix is ​​input into the pre-trained vector projection extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the vibration signal. The fault type corresponding to the largest probability value is determined as the first identification result, and the largest probability value is determined as the first confidence level. The noise feature matrix is ​​input into the pre-trained maximum a posteriori probability kernel extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the noise signal. The fault type corresponding to the largest probability value is determined as the second identification result, and the largest probability value is determined as the second confidence level.

[0017] By adopting the above technical solution, the vibration feature matrix is ​​input into a pre-trained vector projection extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the vibration signal. The fault type corresponding to the largest probability value is determined as the first recognition result, and the largest probability value is determined as the first confidence level. The vector projection extreme learning machine model has fast learning ability and good generalization performance, and can effectively handle the complex nonlinear characteristics of vibration signals to achieve high-precision fault mode recognition. At the same time, the noise feature matrix is ​​input into a pre-trained maximum a posteriori probability kernel extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the noise signal. The fault type corresponding to the largest probability value is determined as the second recognition result, and the largest probability value is determined as the second confidence level. The maximum a posteriori probability kernel extreme learning machine model, by introducing kernel functions and Bayesian inference mechanisms, can better handle the uncertainty and ambiguity in noise signals and improve the recognition accuracy in strong noise environments.

[0018] Optionally, determining the fault identification result based on the first confidence level and the second confidence level specifically includes: Calculate the difference between the first confidence level and the second confidence level to obtain the confidence level difference value; When the confidence difference value is greater than the preset difference threshold, the first confidence level and the second confidence level are compared. If the first confidence level is greater than the second confidence level, the first identification result is determined as the fault identification result; if the second confidence level is greater than the first confidence level, the second identification result is determined as the fault identification result. When the confidence difference value is less than or equal to the preset difference threshold, the operating condition data of the target generator is obtained; based on the operating condition data of the target generator, the first confidence weight of the first identification result and the second confidence weight of the second identification result are obtained from the preset fault association rule base. Multiply the first confidence level by its weight to obtain the first weighted confidence level; multiply the second confidence level by its weight to obtain the second weighted confidence level; compare the first weighted confidence level with the second weighted confidence level. If the first weighted confidence level is greater than the second weighted confidence level, the first identification result is determined as the fault identification result; If the second weighted confidence level is greater than the first weighted confidence level, the second identification result is determined as the fault identification result.

[0019] By employing the above technical solution, the difference between the first confidence level and the second confidence level is calculated to obtain the confidence difference value. When the confidence difference value is greater than a preset difference threshold, the identification result with higher confidence is selected as the fault identification result by comparing the magnitudes of the first confidence level and the second confidence level. This direct comparison strategy can quickly and accurately select the most reliable result when the identification results of the two models differ significantly. When the confidence difference value is less than or equal to the preset difference threshold, it indicates that the confidence levels of the identification results of the two models are similar. At this time, the operating condition data of the target generator is obtained, and based on the operating condition data, the first confidence weight of the first identification result and the second confidence weight of the second identification result are obtained from the preset fault association rule base. The first confidence score is then multiplied by its weight to obtain the first weighted confidence score, and the second confidence score is multiplied by its weight to obtain the second weighted confidence score. The final fault identification result is determined by comparing the magnitudes of the first and second weighted confidence scores. This adaptive weight adjustment mechanism based on operating condition data fully considers the relative importance of vibration and noise signals of the generator under different operating conditions. By introducing operating condition data to dynamically weight the confidence scores, the decision-making process can adapt to the real-time operating status of the generator, solving the decision-making problem when the classifier confidence scores are similar, thereby improving the accuracy of fault diagnosis under complex and variable operating conditions.

[0020] In a second aspect, this application provides an electronic device for a generator fault identification method, the electronic device comprising: one or more processors and a memory; the memory being coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors calling the computer instructions to cause the electronic device for the generator fault identification method to perform the method as described in the first aspect and any possible implementation thereof.

[0021] Thirdly, this application provides a computer program product containing instructions that, when run on an electronic device of a generator fault identification method, cause the electronic device to perform the method as described in the first aspect and any possible implementation thereof.

[0022] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a generator fault identification device, cause the electronic device to perform the method described in the first aspect and any possible implementation thereof. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the architecture of a generator fault identification system provided in an embodiment of this application; Figure 2This is a flowchart illustrating a generator fault identification method provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the optimization of mode decomposition number in adaptive variational mode decomposition provided in an embodiment of this application; Figure 4 This is an exemplary hardware structure diagram of an electronic device for generator fault identification provided in an embodiment of this application. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0025] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0026] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0027] Figure 1 An exemplary system architecture for a generator fault identification system is shown.

[0028] like Figure 1 As shown, the system architecture may include electronic device 11, network 12, and data acquisition device 13. Network 12 serves as the medium for providing a communication link between electronic device 11 and data acquisition device 13. Network 12 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0029] Maintenance personnel can use electronic device 11 to interact with data acquisition device 13 via network 12 to receive fault diagnosis data such as vibration signals and noise signals. Various fault diagnosis applications, such as vibration monitoring applications and fault analysis applications, can be installed on electronic device 11.

[0030] Electronic device 11 is hardware and can be various electronic devices with a display screen and data processing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0031] The data acquisition device 13 can be a device that provides various signal acquisition services, such as field sensor devices like vibration sensors, acoustic sensors, and electromagnetic sensors that collect generator vibration signals, noise signals, and electromagnetic signals in real time. The data acquisition device transmits the collected raw vibration data, noise data, and electromagnetic data to the electronic device 11 via the network 12. The electronic device 11 performs multimodal decomposition processing, feature extraction, fault identification, and other analysis on the received signal data, and displays the final fault identification results to the maintenance personnel.

[0032] It should be understood that Figure 1 The number of electronic devices, networks, and data acquisition devices shown is merely illustrative. Depending on the implementation requirements, any number of electronic devices, networks, and data acquisition devices can be included. The following detailed explanation uses the electronic device side as an example.

[0033] This embodiment provides a generator fault identification method. Figure 2 This is a flowchart illustrating a generator fault identification method provided in an embodiment of this application, as shown below. Figure 2 As shown, the method includes steps S101 to S107: S101: Acquire the vibration and noise signals of the target generator.

[0034] In this embodiment of the application, the target generator refers to a specific generator device that needs to be fault identified, the vibration signal refers to the electrical signal formed by the physical vibration of the generator due to the movement or fault of the mechanical parts during operation, and the noise signal is used to represent the acoustic or electromagnetic interference signal generated by the generator due to factors such as electromagnetic induction and mechanical friction during operation.

[0035] Specifically, the electronic equipment synchronously collects multi-element vibration signals and multi-element noise signals through multiple sensors pre-installed at different key locations of the target generator. The multi-element vibration signals include independent vibration signal streams from multiple mechanical components such as the generator bearings, stator, and rotor. The multi-element noise signals include independent acoustic and electromagnetic signal streams collected from the stator windings and rotor windings by acoustic and electromagnetic sensors. All sensors maintain the same sampling frequency and are set with synchronized acquisition time windows. The vibration sensors convert mechanical vibrations into corresponding voltage signals as vibration signal outputs, the acoustic sensors convert sound wave vibrations into corresponding electrical signals as part of the noise signal, and the electromagnetic sensors convert electromagnetic field changes into corresponding electrical signals as another part of the noise signal. In this embodiment, the collected acoustic and electromagnetic signals can be merged into a complete noise signal using time-domain splicing or data fusion techniques for subsequent unified processing.

[0036] S102: Perform multimodal decomposition on the vibration signal to obtain multiple first submodal components; perform multimodal decomposition on the noise signal to obtain multiple second submodal components.

[0037] In the embodiments of this application, multimodal decomposition processing refers to a signal processing method that decomposes a single complex signal into multiple sub-signal components with different frequency characteristics.

[0038] Specifically, the electronic device employs variational mode decomposition (VMD) to perform multimodal decomposition processing on vibration and noise signals. First, initial parameters for VMD are set, including the number of modes and a penalty factor. Then, the vibration signal is input into the VMD algorithm, which iteratively optimizes the process to decompose the vibration signal into multiple intrinsic mode functions (EMFs) with different frequency centers. Each EMF is output as a first sub-mode component. Similarly, the noise signal is input into the VMD algorithm, which decomposes it into multiple EMFs with different frequency centers. Each EMF is output as a second sub-mode component. During the decomposition process, the VMD algorithm ensures that each sub-mode component is separated in the frequency domain and can reconstruct the original signal. The electronic device then obtains multiple first sub-mode components corresponding to the vibration signal and multiple second sub-mode components corresponding to the noise signal.

[0039] To more clearly illustrate how the optimal parameters in variational mode decomposition are adaptively determined in the embodiments of this application, please refer to [link to relevant documentation]. Figure 3 . Figure 3 This is a schematic diagram illustrating the optimization of mode decomposition numbers in adaptive variational mode decomposition provided in an embodiment of this application.

[0040] In traditional variational mode decomposition (VMD) applications, the mode decomposition number K is a crucial parameter that needs to be preset, and its value directly affects the signal decomposition effect. If the K value is set too small, signals with different frequency components will be classified into the same mode, resulting in the so-called mode aliasing problem; if the K value is set too large, over-decomposition may occur, generating spurious components with no actual physical meaning or misdecomposing noise into independent modes. Existing techniques usually rely on experience or manual trial and error to select the K value, lacking objectivity and adaptability.

[0041] Therefore, this application proposes an adaptive optimization method to solve this problem. For example... Figure 3 As shown, the process of this method can be summarized as follows: First, within a preset reasonable range (e.g., K ranges from 2 to N), different numbers of candidate modes K are traversed. For each candidate K value, variational mode decomposition is performed once to obtain K intrinsic mode function (IMF) components. Then, the correlation coefficients between adjacent components in each group of decomposed components are calculated (e.g., the correlation coefficients between IMF1 and IMF2, IMF2 and IMF3, etc.), and their arithmetic mean is obtained to obtain an average correlation coefficient corresponding to the current K value.

[0042] By plotting different K values ​​and their corresponding average correlation coefficients, we can obtain the following: Figure 3 The relationship curve shown illustrates this. The physical meaning of this curve is that, ideally, each effective modal component should be as orthogonal or uncorrelated as possible to represent different physical components in the signal. When an inappropriate K value leads to "modal aliasing" or "over-decomposition," the correlation between adjacent decomposed components will increase. Especially when "over-decomposition" occurs, a true mode is decomposed into multiple pseudo-components, resulting in a significant increase in the correlation between adjacent pseudo-components. Therefore, the point where the average correlation coefficient reaches its minimum (valley) signifies that the independence between modal components is strongest at this moment, and the decomposition effect is optimal.

[0043] like Figure 3 The diagram illustrates that the horizontal axis of the curve represents the number of candidate modes, K, and the vertical axis represents the calculated average correlation coefficient. The curve reaches its trough (minimum) at K=5, indicating that when the number of mode decompositions is 5, the overall adjacent correlation of the decomposed modal components is the lowest. Therefore, the method in this application automatically selects K=5 as the optimal number of modes and uses it in the subsequent formal signal decomposition steps. In this way, this application implements a data-driven, automated parameter optimization mechanism, avoiding the subjectivity and blindness of manually selecting the K value, significantly improving the accuracy and reliability of signal decomposition, and laying a solid foundation for subsequent feature extraction and fault identification.

[0044] Based on the above embodiments, as an optional embodiment, the step of performing multimodal decomposition processing on the vibration signal to obtain multiple first submodal components and performing multimodal decomposition processing on the noise signal to obtain multiple second submodal components may include steps S201 to S206: S201: Based on the preset number of initial modes, the vibration signal and noise signal are decomposed for the first time to obtain multiple sets of first intrinsic mode function components corresponding to the vibration signal and multiple sets of second intrinsic mode function components corresponding to the noise signal.

[0045] In this embodiment, the preset initial mode number represents the number of modal function components predetermined before signal decomposition. The first intrinsic mode function component represents the intrinsic function component with specific frequency characteristics obtained after the vibration signal is decomposed for the first time according to the preset initial mode number. The second intrinsic mode function component refers to the intrinsic function component with specific frequency characteristics obtained after the noise signal is decomposed for the first time according to the preset initial mode number. The first intrinsic mode function component and the second intrinsic mode function component are different manifestations of intrinsic mode function components. The intrinsic mode function component represents the single-component amplitude-frequency modulated signal obtained after signal decomposition.

[0046] Specifically, the electronic device sets the decomposition parameters of the signal decomposition algorithm according to the preset number of initial modes, and then inputs the vibration signal into the signal decomposition algorithm for the first decomposition process. The signal decomposition algorithm decomposes the vibration signal into a corresponding number of intrinsic mode function components according to the preset number of initial modes. Each intrinsic mode function component has a different frequency center and bandwidth characteristics and is output as the first intrinsic mode function component. At the same time, the electronic device inputs the noise signal into the same signal decomposition algorithm for the first decomposition process. The signal decomposition algorithm decomposes the noise signal into a corresponding number of intrinsic mode function components based on the same preset number of initial modes. Each intrinsic mode function component is output as the second intrinsic mode function component. During the decomposition process, the signal decomposition algorithm ensures that the linear combination of all first intrinsic mode function components can reconstruct the original vibration signal and the linear combination of all second intrinsic mode function components can reconstruct the original noise signal. After the electronic device completes the first decomposition, it obtains multiple sets of first intrinsic mode function components corresponding to the vibration signal and multiple sets of second intrinsic mode function components corresponding to the noise signal.

[0047] S202: Determine the first penalty factor of the vibration signal based on the first intrinsic mode function component; determine the second penalty factor of the noise signal based on the second intrinsic mode function component.

[0048] In the embodiments of this application, the first penalty factor is used as a control parameter to constrain the bandwidth of each modal function component during the vibration signal decomposition process, and the second penalty factor is used as a control parameter to constrain the bandwidth of each modal function component during the noise signal decomposition process.

[0049] Specifically, the electronic device performs frequency domain characteristic analysis on the first intrinsic mode function (IMF) component of the vibration signal, calculates the frequency interval and bandwidth distribution between each IMF component, and then calculates a first penalty factor suitable for the characteristics of the vibration signal based on the frequency interval and bandwidth distribution. The magnitude of the first penalty factor directly affects the frequency domain separation effect of the IMF component in the subsequent decomposition process. At the same time, the electronic device performs the same frequency domain characteristic analysis on the second intrinsic mode function (IMF) component of the noise signal, calculates the frequency interval and bandwidth distribution between each IMF component, and then calculates a second penalty factor suitable for the characteristics of the noise signal based on the frequency domain characteristics of the noise signal. The second penalty factor is used to optimize the separation effect of the second IMF component in the subsequent noise signal decomposition process. The electronic device achieves adaptive decomposition processing for different signal characteristics by determining the penalty factors for the vibration signal and the noise signal respectively.

[0050] Based on the above embodiments, as an optional embodiment, the step of determining the first penalty factor of the vibration signal based on the first intrinsic mode function component and the second penalty factor of the noise signal based on the second intrinsic mode function component may include steps S301 to S302: S301: Select the high-frequency component in the first natural mode function component, calculate the power spectral entropy of the high-frequency component, and take the penalty factor when the power spectral entropy is maximum as the first penalty factor of the vibration signal.

[0051] In this embodiment, power spectral entropy is a numerical index used to quantify the uniformity of the frequency domain energy distribution of high-frequency components. The larger the value of power spectral entropy, the more uniform the frequency domain energy distribution of high-frequency components.

[0052] Specifically, the electronic device selects components with frequencies higher than a preset high-frequency threshold from the first intrinsic mode function components of the vibration signal as high-frequency components. Then, the electronic device performs a Fourier transform on the high-frequency components to obtain the power spectrum, and normalizes the power spectrum to obtain the power spectral density of the high-frequency components. The power spectral density is substituted into a preset formula for calculating the power spectral entropy to obtain the power spectral entropy. The electronic device sets multiple candidate penalty factors within a predefined numerical range, and for each candidate penalty factor, re-decomposes the vibration signal to obtain new first intrinsic mode function components. High-frequency components are extracted from the new first intrinsic mode function components, and the corresponding power spectral entropy is calculated. The electronic device compares the power spectral entropy values ​​corresponding to all candidate penalty factors and selects the candidate penalty factor with the maximum power spectral entropy as the first penalty factor for the vibration signal. The formula (1) for calculating the power spectral entropy can be expressed as: (1) in For power spectral entropy, denoted as the power spectral density of the high-frequency components.

[0053] S302: Select the high-frequency component in the second intrinsic mode function component, calculate the power spectral entropy of the high-frequency component, and use the penalty factor when the power spectral entropy is maximum as the second penalty factor of the noise signal.

[0054] Specifically, the electronic device sorts all second intrinsic mode function (IIMF) components of the noise signal from high to low according to their center frequencies. From the sorting results, it selects a preset number of IIMF components with the highest frequencies as high-frequency components. Then, the electronic device performs a Fourier transform on the selected high-frequency components to obtain the power spectrum, and then calculates the normalized entropy value of the power spectrum as the power spectrum entropy. The electronic device sets multiple candidate penalty factors within a predefined numerical range, and re-decomposes the noise signal for each candidate penalty factor to obtain new IIMF components. From the new IIMF components, high-frequency components are extracted according to the same standard, and the corresponding power spectrum entropy is calculated. The electronic device compares the power spectrum entropy values ​​corresponding to all candidate penalty factors and selects the candidate penalty factor with the largest power spectrum entropy as the second penalty factor of the noise signal.

[0055] S203: Perform a second decomposition on the vibration signal based on the first penalty factor to obtain multiple sets of first decomposition components; perform a second decomposition on the noise signal using the second penalty factor to obtain multiple sets of second decomposition components.

[0056] Specifically, the electronic device uses a determined first penalty factor as the penalty parameter for the variational mode decomposition algorithm to re-execute the variational mode decomposition process on the vibration signal. This decomposition process, through the optimized penalty factor, can better control the frequency domain overlap between each modal component, thereby obtaining multiple sets of first decomposition components with higher frequency domain separation. Each set of first decomposition components corresponds to the characteristic information of the vibration signal in different frequency ranges. Simultaneously, the electronic device uses a determined second penalty factor as the penalty parameter for the variational mode decomposition algorithm to re-execute the variational mode decomposition process on the noise signal, obtaining multiple sets of second decomposition components with clearer frequency domain characteristics. Each set of second decomposition components corresponds to the characteristic information of the noise signal in different frequency ranges.

[0057] S204: Calculate the correlation coefficient of adjacent components in each group of first decomposition components, obtain the average correlation coefficient of each group of first decomposition components, and take the number of modes with the smallest average correlation coefficient as the number of first modes of vibration signal.

[0058] In the embodiments of this application, the first number of modes refers to the optimal number of modal components for variational mode decomposition of vibration signal determined by optimization analysis; the correlation coefficient ranges from negative one to positive one, the smaller the value, the weaker the correlation between adjacent components, and the better the frequency domain separation effect. When the average correlation coefficient reaches the minimum value, the corresponding number of modes can achieve the optimal decomposition of vibration signal.

[0059] Specifically, the electronic device sets multiple candidate modes and performs variational mode decomposition on each. For each candidate mode, a corresponding number of first decomposition components are obtained. The electronic device sorts the first decomposition components for each mode from low to high frequency, and then calculates the Pearson correlation coefficient between two adjacent components after sorting. The correlation coefficients of all adjacent component pairs are arithmetically averaged to obtain the average correlation coefficient corresponding to the current mode. The electronic device repeats the above process to calculate the average correlation coefficient corresponding to all candidate modes. By comparative analysis, the candidate mode corresponding to the minimum average correlation coefficient is found, and this mode is determined as the first mode of the vibration signal.

[0060] The formula (2) for calculating the Pearson correlation coefficient can be expressed as: (2) in, For adjacent components and The Pearson correlation coefficient between them and : refers to the aj-th and j+1-th first decomposition components obtained after variational mode decomposition and sorting by frequency from low to high, given a certain number of candidate modes. They are time series data representing the signal's variation over time. and : These refer to the first decomposition components respectively and The arithmetic mean of all data points for each component. These two values ​​are used in the denominator of the formula to calculate the dispersion of each component.

[0061] S205: Calculate the correlation coefficient of adjacent components in each group of second decomposition components, obtain the average correlation coefficient of each group of second decomposition components, and take the number of modes with the smallest average correlation coefficient as the number of second modes of the noise signal.

[0062] In this embodiment, the number of second modes refers to the optimal number of mode components for variational mode decomposition of the noise signal determined through optimization analysis; the correlation coefficient of adjacent components in the second decomposition component reflects the degree of linear correlation between adjacent mode components after the noise signal is decomposed. When the average correlation coefficient of the second decomposition component reaches the minimum value, the corresponding number of modes can achieve the optimal decomposition of the noise signal, and at this time, the frequency domain independence between each mode component is the strongest.

[0063] Specifically, the electronic device sets multiple candidate modes and performs variational mode decomposition on the noise signal for each. For each candidate mode, a corresponding number of second decomposition components are obtained. The electronic device sorts the second decomposition components for each mode from low to high frequency, and then calculates the Pearson correlation coefficient between two adjacent components after sorting. The correlation coefficients of all adjacent component pairs are arithmetically averaged to obtain the average correlation coefficient corresponding to the current mode. The electronic device repeats the above process to calculate the average correlation coefficient corresponding to all candidate modes. By comparative analysis, the candidate mode corresponding to the minimum average correlation coefficient is found, and this mode is determined as the second mode of the noise signal.

[0064] S206: The vibration signal is decomposed a third time based on the first penalty factor and the number of first modes to obtain multiple first sub-mode components; the noise signal is decomposed a third time based on the second penalty factor and the number of second modes to obtain multiple second sub-mode components.

[0065] Specifically, the electronic device simultaneously uses the aforementioned determined first penalty factor and first mode number as input parameters for the variational mode decomposition algorithm to re-decompose the vibration signal. Under the frequency domain constraint of the first penalty factor and the component number limitation of the first mode number, the variational mode decomposition algorithm iteratively solves the center frequency and bandwidth of each mode component, and finally outputs a first sub-mode component with the same number as the first mode number. At the same time, the electronic device uses the second penalty factor and second mode number as input parameters for the variational mode decomposition algorithm to re-decompose the noise signal. Under the frequency domain constraint of the second penalty factor and the component number limitation of the second mode number, the variational mode decomposition algorithm iteratively solves the center frequency and bandwidth of each mode component, and finally outputs a second sub-mode component with the same number as the second mode number.

[0066] It should be noted that in this embodiment, the third decomposition process is a constrained optimization process. Specifically, it obtains the optimal sub-mode components by solving a pre-constructed multi-objective optimization function. The core constraints of this objective function are "maximum energy concentration of each sub-mode component" and "lowest correlation between adjacent sub-mode components". This ensures that the final decomposed sub-mode components can focus on fault feature energy at specific frequencies, are independent of each other, and have low information redundancy, thus providing high-quality input for subsequent accurate feature extraction.

[0067] S103: Determine the vibration signal sequence based on each first sub-mode component and the vibration signal; determine the noise signal sequence based on each second sub-mode component and the noise signal.

[0068] Specifically, the electronic device first calculates the mutual information between each first sub-mode component and the vibration signal. A higher mutual information indicates a higher correlation between the first sub-mode component and the vibration signal, and contains richer fault information. Next, the component with the highest mutual information is selected from all first sub-mode components as the component containing the most significant vibration characteristics. Then, the selected highly correlated first sub-mode components undergo signal reconstruction processing to generate a vibration signal sequence. This vibration signal sequence retains the key fault characteristics of the original vibration signal while reducing noise interference. Similarly, the mutual information between each second sub-mode component and the noise signal is calculated. A higher mutual information indicates a higher correlation between the second sub-mode component and the noise signal, and contains richer fault information. Then, the component with the highest mutual information is selected from all second sub-mode components as the component containing the most significant noise characteristics. Finally, the selected highly correlated second sub-mode components undergo signal reconstruction processing to generate a noise signal sequence. This noise signal sequence retains the key fault characteristics of the original noise signal while reducing background interference.

[0069] Based on the above embodiments, as an optional embodiment, the step of determining the vibration signal sequence based on each first sub-mode component and the vibration signal, and determining the noise signal sequence based on each second sub-mode component and the noise signal, may include steps S401 to S402: S401: Calculate the mutual information between each first sub-mode component and the vibration signal, select the first sub-mode component with the largest mutual information for reconstruction, and obtain the vibration signal sequence.

[0070] Specifically, the electronic device divides the vibration signal and each first sub-mode component into time windows or segments to ensure a sufficient sample size for calculating reliable statistics. Then, the electronic device constructs an amplitude histogram for each first sub-mode component and the vibration signal, and obtains the marginal probability distribution function by counting the number of sample points within each amplitude interval and normalizing the results. Next, the electronic device constructs a two-dimensional histogram and counts the joint frequency of the amplitude values ​​of the first sub-mode component and the vibration signal at corresponding time points, normalizing these frequencies to obtain the joint probability distribution function. After obtaining the probability distribution, the electronic device calculates the marginal entropy of the first sub-mode component and the vibration signal by multiplying each probability value in the marginal probability distribution by the negative of its own logarithm and summing the results. Similarly, the electronic device calculates the joint entropy by multiplying each probability value in the joint probability distribution by the negative of its own logarithm and summing the results. After completing the entropy calculation, the electronic device obtains the mutual information by subtracting the joint entropy value from the sum of the two marginal entropy values. Higher mutual information indicates stronger statistical dependence between signals and more shared information. After calculating the mutual information of all first submode components, the electronic device compares these values ​​and selects the component with the highest mutual information for reconstruction. By retaining the component with the highest mutual information and discarding other components, the device effectively extracts components containing key fault information and reduces interference. During the reconstruction process, the electronic device can directly use the selected component as a new vibration signal sequence, or process the selected component using a specific algorithm to generate a sequence. This reconstruction method ensures that the sequence retains the key features of the original signal.

[0071] S402: Calculate the mutual information between each second sub-mode component and the noise signal, select the second sub-mode component with the largest mutual information for reconstruction, and obtain the noise signal sequence.

[0072] Specifically, the electronic device calculates the mutual information between each second sub-mode component and the noise signal. In this process, the electronic device first divides the noise signal and each second sub-mode component into time windows or segments to ensure sufficient sample size. Then, the electronic device constructs amplitude histograms of the noise signal and each second sub-mode component, and normalizes them by counting the number of sample points within the amplitude interval to obtain the marginal probability distribution function. Next, the electronic device constructs a two-dimensional histogram to count the joint frequency of the amplitude values ​​of the noise signal and each second sub-mode component at corresponding time points, and normalizes it to obtain the joint probability distribution function. Then, the electronic device calculates the marginal entropy and joint entropy, and obtains the mutual information by subtracting the joint entropy from the sum of the two marginal entropies. After completing all mutual information calculations, the electronic device compares the mutual information between each second sub-mode component and the noise signal and selects the second sub-mode component with the largest mutual information. This indicates that the component has the strongest statistical dependence on the noise signal and contains the most noise feature information. Finally, the electronic device uses the selected second sub-mode component with the largest mutual information for reconstruction. By retaining the component most closely associated with the noise and discarding other components, the electronic device generates a noise signal sequence. This sequence retains the main features of the original noise signal and can be used for noise identification and suppression in subsequent processing.

[0073] S104: Determine the vibration feature matrix based on the vibration signal sequence; determine the noise feature matrix based on the noise signal sequence.

[0074] In this embodiment, the vibration feature matrix refers to a two-dimensional data structure obtained by feature extraction and mathematical transformation of the vibration signal sequence, used to characterize the feature information of the vibration signal at different time scales; the noise feature matrix refers to a two-dimensional data structure obtained by feature extraction and mathematical transformation of the noise signal sequence, used to characterize the feature information of the noise signal at different time scales.

[0075] Specifically, the electronic device first extracts features from the vibration signal sequence by calculating time-domain features (such as mean, variance, peak value, kurtosis, skewness, etc.), frequency-domain features (such as spectral energy distribution, dominant frequency components, frequency band energy ratio, etc.), and time-frequency features (such as wavelet coefficients, Hilbert transform parameters, etc.). These multi-dimensional features are then arranged and organized according to predetermined rules to form a vibration feature matrix. Each row in the vibration feature matrix can represent a feature set of a time window or signal segment, while each column represents a specific type of feature. Similarly, the electronic device performs a similar feature extraction process on the noise signal sequence, extracting various feature parameters that reflect the characteristics of the noise signal, and systematically organizing these features into a noise feature matrix. The matrix structure is similar to the vibration feature matrix, but its content reflects the characteristics of the noise.

[0076] Based on the above embodiments, as an optional embodiment, the step of determining the vibration feature matrix based on the vibration signal sequence and the step of determining the noise feature matrix based on the noise signal sequence may include steps S501 to S503: S501: Perform multi-level decomposition on the vibration signal sequence to obtain multiple vibration level components; perform multi-level decomposition on the noise signal sequence to obtain multiple noise level components.

[0077] In the embodiments of this application, vibration level components refer to sub-signals of different frequency levels extracted from a vibration signal sequence through multi-level decomposition, and each vibration level component contains vibration information within a specific frequency range; noise level components refer to sub-signals of different frequency levels extracted from a noise signal sequence through multi-level decomposition, and each noise level component contains noise information within a specific frequency range.

[0078] Specifically, the electronic device employs wavelet packet decomposition to perform multi-level decomposition of the vibration signal sequence (multi-level decomposition refers to a signal processing technique that uses multi-resolution analysis to decompose a signal layer by layer at different frequency levels, enabling the decomposition of complex signals into sub-signal components with different frequency ranges). First, a suitable wavelet basis function is selected as the basis for decomposition. Commonly used wavelet basis functions include Daubechies wavelet, Morlet wavelet, and Haar wavelet. The electronic device selects a wavelet basis function with good time-frequency localization characteristics based on the characteristics of the signal. Then, the number of decomposition levels is set, generally between three and six. The more decomposition levels, the higher the frequency resolution but the lower the time resolution. The electronic device performs wavelet packet decomposition layer by layer starting from the highest level. Each decomposition layer decomposes the signal of the previous layer into low-frequency and high-frequency components. The low-frequency component contains the trend information of the signal, and the high-frequency component contains the detailed information of the signal. Through multi-level recursive decomposition, multiple vibration level components with different frequency ranges are finally obtained. Each vibration level component corresponds to the component of the original signal within a specific frequency band. The electronic device reconstructs and filters each vibration level component to remove boundary effects and artifact noise introduced during the decomposition process.

[0079] S502: Perform sliding window calculations on the vibration level components at different scales to obtain coarse-grained vibration sequences of the vibration level components at different scales; perform sliding window calculations on the noise level components at different scales to obtain coarse-grained noise sequences of the noise level components at different scales.

[0080] In the embodiments of this application, the coarse-grained vibration sequence and the coarse-grained noise sequence refer to the downsampled sequences obtained after sliding window calculation processing. These sequences retain the main feature information of the original hierarchical components and have better computational efficiency.

[0081] Specifically, the electronic device first sets a series of scale parameter values ​​for multi-scale analysis. Each scale parameter value determines the window length and coarsening level of the corresponding sliding window. The electronic device uses each vibration level component as input data and sequentially applies all the set scale parameters for sliding window calculation. During the sliding window calculation, the window length is set to the value of the current scale parameter. The window moves continuously backward in unit steps from the starting position of the vibration level component until it reaches the end of the sequence. At each window position, the electronic device calculates the arithmetic mean of all data points within the window as the coarsening result value for the current position. The coarsening result values ​​for all window positions are arranged in chronological order. The permutation constitutes a complete coarse-grained vibration sequence. Due to the limitation of window length, the data length of each coarse-grained vibration sequence will be reduced by the corresponding number of data points compared to the length of the original vibration level component. The electronic device repeatedly executes the above sliding window calculation process for each vibration level component and applies all the set scale parameters to finally obtain multiple coarse-grained vibration sequences at different scales. The electronic device uses the exact same scale parameter settings and sliding window calculation method to process all noise level components to ensure that the coarse-grained noise sequence and the coarse-grained vibration sequence maintain a strict correspondence in processing method and scale settings, and finally obtains multiple coarse-grained noise sequences at different scales.

[0082] S503: Calculate the fuzzy entropy value of the coarse-grained vibration sequence, integrate the fuzzy entropy values ​​of multiple vibration level components at different scales to obtain the vibration feature matrix; calculate the fuzzy entropy value of the coarse-grained noise sequence, integrate the fuzzy entropy values ​​of multiple noise level components at different scales to obtain the noise feature matrix.

[0083] In the embodiments of this application, the fuzzy entropy value refers to the numerical result obtained by quantifying the complexity and irregularity of a time series through the fuzzy entropy algorithm. It can effectively measure the fuzziness of pattern matching and the randomness of the signal in the sequence. The larger the value, the higher the complexity and unpredictability of the sequence.

[0084] Specifically, the electronic device applies a fuzzy entropy algorithm to each coarse-grained vibration sequence for complexity analysis. The fuzzy entropy algorithm first sets the dimension parameter and fuzzy boundary width parameter for pattern matching. The electronic device constructs all possible pattern vectors in the coarse-grained vibration sequence and calculates the maximum absolute distance between different pattern vectors. It then converts the distance values ​​into fuzzy similarity using a fuzzy function and counts the number of pattern matches that satisfy the similarity condition. The electronic device calculates the fuzzy similarity statistics for the case with the dimension parameter and the case with the dimension parameter increased by one. The logarithmic ratio of the statistical results in both cases constitutes the fuzzy entropy value of the current coarse-grained vibration sequence. The electronic device repeats the fuzzy entropy algorithm for all coarse-grained vibration sequences. The algorithm calculation process obtains the fuzzy entropy value corresponding to each vibration level component at each scale. The electronic device arranges all fuzzy entropy values ​​in a matrix according to the combination relationship between the vibration level component number and the scale parameter. The vibration level component number is used as the row index of the matrix and the scale parameter is used as the column index of the matrix. Each fuzzy entropy value is filled into the corresponding row and column position to form a complete vibration feature matrix. The electronic device uses the same fuzzy entropy algorithm parameter settings and calculation process to calculate the fuzzy entropy value of all coarse-grained noise sequences to ensure that the noise feature matrix and the vibration feature matrix maintain strict consistency in algorithm implementation and data structure. Finally, the noise feature matrix is ​​constructed according to the same matrix arrangement method.

[0085] S105: Based on the vibration feature matrix, determine the first identification result and the first confidence level; based on the noise feature matrix, determine the second identification result and the second confidence level.

[0086] In this embodiment of the application, the first identification result refers to the equipment status classification judgment result obtained by analyzing the multi-scale complex feature patterns in the vibration feature matrix. The first identification result reflects the equipment health status assessment conclusion based on the hierarchical components of the vibration signal, including the category identifier of the current operating status of the equipment and the corresponding fault type identification information.

[0087] The second identification result represents the equipment status classification result obtained by analyzing the multi-scale complexity feature patterns in the noise feature matrix. The second identification result reflects the equipment health status assessment conclusion based on the hierarchical components of the noise signal, including the category identifier of the current operating status of the equipment and the corresponding fault type identification information.

[0088] Specifically, the electronic device inputs the vibration feature matrix into a pre-trained vibration recognition model for state classification analysis. The vibration recognition model uses a deep learning network structure to perform nonlinear mapping and pattern extraction on the multi-dimensional features in the vibration feature matrix. During the forward propagation of the vibration recognition model, the electronic device extracts high-level abstract features of the vibration feature matrix layer by layer. The fully connected layers of the model map the extracted feature vectors to the probability distribution of each device state category. The electronic device selects the state category with the highest probability value as the first recognition result, and uses the highest probability value as the initial estimate of the first confidence level. The electronic device further adjusts the first confidence level by calculating the entropy value of the probability distribution and the difference between the highest probability and the second highest probability, ensuring that the first confidence level is maintained. A confidence level can accurately reflect the reliability of the recognition result. The electronic device uses the same processing flow to input the noise feature matrix into the pre-trained noise recognition model for state classification analysis. The noise recognition model has a similar network architecture and training method to the vibration recognition model, but it is specifically optimized for noise signal features. The electronic device performs feature extraction and probability distribution calculation on the noise feature matrix through the noise recognition model, selects the state category corresponding to the highest probability as the second recognition result, and determines the corresponding second confidence level through probability statistical analysis. The electronic device also performs cross-validation and consistency checks on the output results of the two recognition models to ensure that the first recognition result and the second recognition result have a reasonable logical correlation.

[0089] Based on the above embodiments, as an optional embodiment, a first identification result and a first confidence level are determined based on the vibration feature matrix. The step of determining a second identification result and a second confidence level based on the noise feature matrix may include steps S601 to S602: S601: Input the vibration feature matrix into the pre-trained vector projection extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the vibration signal. The fault type corresponding to the largest probability value is determined as the first identification result, and the largest probability value is determined as the first confidence level.

[0090] In the embodiments of this application, the vector projection extreme learning machine model refers to a machine learning classification model constructed by combining vector projection transformation technology and extreme learning machine algorithm. The vector projection extreme learning machine model maps high-dimensional input features to low-dimensional space through vector projection operation and uses the fast learning capability of extreme learning machine to achieve efficient pattern recognition and classification prediction functions.

[0091] Specifically, the electronic device feeds the vibration feature matrix as input data into a pre-trained vector projection extreme learning machine (VLM) model for fault identification. The VLM model first performs a vector projection transformation on the vibration feature matrix. This transformation maps the high-dimensional feature space of the vibration feature matrix to a lower-dimensional feature space with better separability through the learned projection matrix. The electronic device then inputs the transformed feature vectors into the VLM network for classification calculation. The VLM network consists of three layers: an input layer, a hidden layer, and an output layer. The input layer receives the transformed feature vectors, and the hidden layer performs nonlinear transformation on the input features using randomly generated weights and bias parameters. Instead, the output layer maps the activation values ​​of the hidden layer to probability values ​​of each fault type through the analytically solved output weight matrix. The electronic device normalizes the original values ​​of the output layer through a probability distribution function to ensure that the probability values ​​of each fault type meet the mathematical constraints of the probability distribution. The electronic device iterates through all elements of the probability values ​​of each fault type and finds the probability value with the largest value. The fault type identifier corresponding to the largest probability value is identified as the first identification result. At the same time, the value of the largest probability value itself is directly assigned to the first confidence level. The electronic device also evaluates the difference between the largest probability value and the second largest probability value to ensure that the first identification result has sufficient discriminative power and reliability.

[0092] S602: Input the noise feature matrix into the pre-trained maximum a posteriori probability kernel extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the noise signal. The fault type corresponding to the largest probability value is determined as the second identification result, and the largest probability value is determined as the second confidence level.

[0093] In this embodiment, the maximum a posteriori probability kernel extreme learning machine model refers to an advanced machine learning classification model constructed by combining maximum a posteriori probability estimation theory, kernel function mapping technology and extreme learning machine algorithm. The maximum a posteriori probability kernel extreme learning machine model maps input features to a high-dimensional kernel space through kernel function and optimizes parameters using the maximum a posteriori probability criterion to achieve robust fault identification and probabilistic reasoning functions for noisy signals.

[0094] Specifically, the electronic device inputs the noise feature matrix as input data into a pre-trained maximum a posteriori (MAP) kernel-based extreme learning machine (XLM) model for fault identification and analysis. The MAP model first maps the noise feature matrix to a high-dimensional kernel feature space through kernel function transformation. Kernel function transformation effectively handles the nonlinear characteristics and complex patterns of noise signals. The electronic device then constructs an extreme learning machine network structure within the kernel feature space. The hidden layer neurons of the extreme learning machine perform inner product calculations with the input noise feature matrix using the kernel function, generating hidden layer activation values ​​in the kernel space. Finally, the electronic device optimizes the output weights of the extreme learning machine using the maximum a posteriori (MAP) probability estimation method. The system improves the generalization ability and noise resistance of the model by introducing the prior probability distribution of the weight parameters. The electronic device obtains the optimal output weight matrix by analytically solving the maximum a posteriori probability objective function. The output layer uses the optimal weight matrix to map the hidden layer activation values ​​in the kernel space to the probability values ​​of each fault type. The electronic device normalizes the original values ​​of the output layer through the probability distribution function to ensure that the probability values ​​of each fault type constitute an effective probability distribution. The electronic device traverses the normalized probability values ​​of each fault type and identifies the maximum probability value. The fault type corresponding to the maximum probability value is determined as the second identification result, and the value of the maximum probability value is directly used as the quantitative index of the second confidence.

[0095] In this embodiment, different Extreme Learning Machine (ELM) models are used for vibration and noise signals, based on a deep matching of the physical characteristics of the two signals. Specifically, generator vibration signals typically contain deterministic impact components caused by mechanical faults and periodic harmonics. Their fault features exhibit good intra-class clustering and linear separability in high-dimensional space. The Vector Projection ELM model effectively reduces feature dimensionality and eliminates redundancy through projection transformation, focusing on these highly deterministic core features to achieve fast and accurate classification. In contrast, noise signals (especially electromagnetic noise) exhibit stronger randomness, non-stationarity, and a lower signal-to-noise ratio. Their fault features are often submerged in strong background noise, presenting complex nonlinear relationships. The Maximum A posteriori (MAP) kernel ELM model, by introducing a kernel function to handle nonlinear problems and combining MAP estimation within a Bayesian framework, has a natural advantage in handling uncertain data and suppressing noise interference, enabling it to more robustly extract fault modes from strong noise. Therefore, this differentiated classifier configuration is the choice to maximize the use of complementary information from the two signals and optimize the overall system performance.

[0096] S106: When the first identification result is inconsistent with the second identification result, the fault identification result shall be determined according to the first confidence level and the second confidence level.

[0097] Specifically, the electronic device first performs a consistency check on the first and second identification results. The electronic device compares whether the fault type labels corresponding to the two identification results are exactly the same. When the fault type labels are different, it is determined that the identification results are inconsistent. The electronic device then initiates a conflict resolution mechanism based on confidence level. The electronic device obtains the first confidence level value corresponding to the first identification result and the second confidence level value corresponding to the second identification result. The electronic device compares the first confidence level and the second confidence level. When the first confidence level is greater than the second confidence level, it indicates that the reliability of the vibration signal analysis is higher. The electronic device determines the first identification result as the final fault identification result. When the second confidence level is greater than the first confidence level, it indicates that the reliability of the noise signal analysis is higher. The electronic device determines the second identification result as the final fault identification result. When the first confidence level is equal to the second confidence level, the electronic device adopts a preset priority strategy or introduces other auxiliary criteria to make the final decision.

[0098] Based on the above embodiments, as an optional embodiment, the step of determining the fault identification result according to the first confidence level and the second confidence level may include steps S701 to S705: S701: Calculate the difference between the first confidence level and the second confidence level to obtain the confidence level difference value.

[0099] In the embodiments of this application, the confidence difference value reflects the relative difference between vibration signal analysis and noise signal analysis in terms of fault diagnosis accuracy.

[0100] Specifically, the electronic device acquires the values ​​of the first confidence level and the second confidence level. The electronic device performs a subtraction operation to calculate the difference between the first confidence level and the second confidence level. The electronic device obtains the confidence difference value by subtracting the second confidence level from the first confidence level. When the confidence difference value is positive, it means that the first confidence level is higher than the second confidence level, that is, the vibration signal analysis has a higher degree of credibility. When the confidence difference value is negative, it means that the second confidence level is higher than the first confidence level, that is, the noise signal analysis has a higher degree of credibility. When the confidence difference value is close to zero, it means that the credibility of the two recognition channels is basically the same. The electronic device stores the calculated confidence difference value as the basis for confidence comparison analysis and subsequent processing.

[0101] S702: When the confidence difference value is greater than the preset difference threshold, compare the first confidence level with the second confidence level. If the first confidence level is greater than the second confidence level, determine the first identification result as the fault identification result. If the second confidence level is greater than the first confidence level, determine the second identification result as the fault identification result.

[0102] Specifically, the electronic device acquires the confidence difference value and calculates its absolute value, and compares the absolute value of the confidence difference value with a preset difference threshold. When the absolute value of the confidence difference value is greater than the preset difference threshold, it indicates that there is a significant difference in the credibility of the two recognition channels. The electronic device directly compares the values ​​of the first confidence level and the second confidence level. When the first confidence level value is greater than the second confidence level value, the electronic device selects the first recognition result as the final fault recognition result. When the second confidence level value is greater than the first confidence level value, the electronic device selects the second recognition result as the final fault recognition result. The electronic device completes a rapid decision-making process based on significant confidence differences.

[0103] S703: When the confidence difference value is less than or equal to the preset difference threshold, obtain the operating condition data of the target generator; based on the operating condition data of the target generator, obtain the first confidence weight of the first identification result and the second confidence weight of the second identification result from the preset fault association rule base. In this embodiment of the application, the fault association rule base refers to a pre-established knowledge base containing the reliability laws of various fault identification methods under different operating conditions. The fault association rule base forms fault diagnosis confidence adjustment rules for different operating conditions through historical data statistical analysis and expert experience summary.

[0104] Specifically, when the absolute value of the confidence difference value detected by the electronic device is less than or equal to a preset difference threshold, it is determined that the confidence levels of the two recognition channels are relatively close. A weighted fusion strategy adapted to the operating conditions needs to be adopted. The electronic device starts the operating condition data acquisition program of the target generator to obtain key operating parameters such as real-time load rate, speed, stator temperature, bearing temperature, stator current, and terminal voltage through the sensor network. The collected operating condition data is standardized and feature extracted to form an operating condition feature vector. The operating condition feature vector is matched and queried in the fault association rule base to find the historical operating condition record most similar to the current operating condition. The fault association rule base analyzes the similarity between the current operating condition feature vector and the historical operating condition record to calculate a weight allocation scheme suitable for the current operating state. The query results return the first confidence weight corresponding to the first recognition result and the second confidence weight corresponding to the second recognition result as adjustment parameters for subsequent weighted calculations. The acquisition of weight parameters provides a quantitative basis for the operating condition adaptive adjustment of the confidence level. The preset fault association rule base can be constructed offline. For example, a large amount of normal and various fault data of generators under different operating conditions (different loads, speeds, temperatures, etc.) can be collected. Data mining techniques (such as association rule mining) can be used to analyze the relationship between the significance of specific operating conditions and specific fault signals (vibration or noise). This data can then be modified based on the experience of domain experts to ultimately form a mapping rule table of operating conditions, fault types, and signal source confidence weights. One rule could be: when the operating condition is high load and high temperature, the features related to bearing wear in the vibration signal are more significant. Therefore, the first confidence weight of the vibration signal (e.g., 0.7) should be higher than the second confidence weight of the noise signal (e.g., 0.3).

[0105] S704: Multiply the first confidence level by the first confidence level weight to obtain the first weighted confidence level; multiply the second confidence level by the second confidence level weight to obtain the second weighted confidence level.

[0106] In the embodiments of this application, the first weighted confidence level refers to the confidence level value after the first confidence level is adjusted by the first confidence level weight, which reflects the actual reliability of the fault identification result based on vibration signal under specific operating conditions; the second weighted confidence level represents the confidence level value after the second confidence level is adjusted by the second confidence level weight, which reflects the actual reliability of the fault identification result based on noise signal under specific operating conditions.

[0107] Specifically, the electronic device acquires the value of the first confidence level and the corresponding first confidence level weight value and performs a multiplication operation. The multiplication operation multiplies the first confidence level and the first confidence level weight value to obtain the first weighted confidence level, which is used as the credibility of vibration signal fault identification after operating condition adjustment. At the same time, the electronic device acquires the value of the second confidence level and the corresponding second confidence level weight value and performs the same multiplication operation. The multiplication operation multiplies the second confidence level and the second confidence level weight value to obtain the second weighted confidence level, which is used as the credibility of noise signal fault identification after operating condition adjustment. The first weighted confidence level and the second weighted confidence level comprehensively reflect the actual effectiveness of the two fault identification methods under the current operating conditions. The weighting calculation process eliminates the adverse effects of changes in operating conditions on the accuracy of confidence level assessment. The calculated first weighted confidence level and second weighted confidence level are used as the basic data for fusion decision-making and subsequent final fault judgment.

[0108] S705: Compare the first weighted confidence level with the second weighted confidence level. If the first weighted confidence level is greater than the second weighted confidence level, the first identification result is determined as the fault identification result; if the second weighted confidence level is greater than the first weighted confidence level, the second identification result is determined as the fault identification result.

[0109] Specifically, the electronic device performs a numerical comparison operation, comparing the first weighted confidence level with the second weighted confidence level to determine the relative strengths of the two weighted confidence levels. When the comparison result shows that the value of the first weighted confidence level is greater than that of the second weighted confidence level, it indicates that the fault identification method based on vibration signals has a higher degree of reliability after operating condition adjustment. The electronic device then selects the first identification result as the optimal choice and outputs it as the final fault identification result to the user or control system. When the comparison result shows that the value of the second weighted confidence level is greater than that of the first weighted confidence level, it indicates that the fault identification method based on noise signals has a higher degree of reliability after operating condition adjustment. The electronic device then selects the second identification result as the optimal choice and outputs it as the final fault identification result. The decision-making process ensures that the identification channel result with higher confidence is selected as the final diagnostic conclusion. The determination of the fault identification result completes the entire multi-signal fusion fault diagnosis process, providing a reliable basis for generator maintenance decisions.

[0110] S107: When the first identification result is consistent with the second identification result, the first identification result is determined as the fault identification result.

[0111] Specifically, the electronic device compares the fault type identifier of the first identification result with the fault type identifier of the second identification result bit by bit to determine whether the two identification results represent the same fault type. When the comparison process confirms that the fault type identifiers of the first identification result and the second identification result are exactly the same, the electronic device directly selects the first identification result as the final fault identification result.

[0112] The following describes an exemplary electronic device for generator fault identification provided in an embodiment of this application. Figure 4 This is an exemplary hardware structure diagram of an electronic device for generator fault identification provided in an embodiment of this application.

[0113] In some embodiments, the electronic device for generator fault identification is a computer device or includes a computer device in the electronic device for generator fault identification. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0114] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0115] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application 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 scope of the technical solutions of the embodiments of this application.

[0116] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0117] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0118] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A generator fault identification method, characterized in that, The method includes: Acquire vibration and noise signals from the target generator; The vibration signal is subjected to multimodal decomposition to obtain multiple first submodal components; the noise signal is subjected to multimodal decomposition to obtain multiple second submodal components. Based on each first sub-mode component and vibration signal, determine the vibration signal sequence; based on each second sub-mode component and noise signal, determine the noise signal sequence. Based on the vibration signal sequence, determine the vibration feature matrix; based on the noise signal sequence, determine the noise feature matrix. Based on the vibration feature matrix, a first identification result and a first confidence level are determined; based on the noise feature matrix, a second identification result and a second confidence level are determined. When the first identification result is inconsistent with the second identification result, the fault identification result is determined based on the first confidence level and the second confidence level. When the first identification result matches the second identification result, the first identification result is determined as the fault identification result.

2. The generator fault identification method according to claim 1, characterized in that, The vibration signal is subjected to multimodal decomposition processing to obtain multiple first submodal components; The noise signal is subjected to multimodal decomposition to obtain multiple second submodal components, specifically including: Based on the preset number of initial modes, the vibration signal and noise signal are decomposed for the first time to obtain multiple sets of first intrinsic mode function components corresponding to the vibration signal and multiple sets of second intrinsic mode function components corresponding to the noise signal. A first penalty factor for the vibration signal is determined based on the first intrinsic mode function component; a second penalty factor for the noise signal is determined based on the second intrinsic mode function component. The vibration signal is decomposed a second time based on the first penalty factor to obtain multiple sets of first decomposition components; the noise signal is decomposed a second time using the second penalty factor to obtain multiple sets of second decomposition components. Calculate the correlation coefficient of adjacent components in each group of first decomposition components, and obtain the average correlation coefficient of each group of first decomposition components. The number of modes with the smallest average correlation coefficient is taken as the number of first modes of the vibration signal. Calculate the correlation coefficient of adjacent components in each group of second decomposition components to obtain the average correlation coefficient of each group of second decomposition components, and take the number of modes with the smallest average correlation coefficient as the number of second modes of the noise signal. The vibration signal is decomposed a third time based on the first penalty factor and the number of first modes to obtain multiple first sub-mode components; the noise signal is decomposed a third time based on the second penalty factor and the number of second modes to obtain multiple second sub-mode components.

3. The generator fault identification method according to claim 2, characterized in that, The first penalty factor for determining the vibration signal is based on the first intrinsic mode function component; The second penalty factor for the noise signal is determined based on the second intrinsic mode function component, specifically including: Select the high-frequency component in the first natural mode function component, calculate the power spectral entropy of the high-frequency component, and take the penalty factor when the power spectral entropy is maximum as the first penalty factor of the vibration signal. Select the high-frequency components in the second intrinsic mode function components, calculate the power spectral entropy of the high-frequency components, and use the penalty factor when the power spectral entropy is at its maximum as the second penalty factor for the noise signal.

4. The generator fault identification method according to claim 1, characterized in that, The determination of the vibration signal sequence based on each first sub-mode component and the vibration signal, and the determination of the noise signal sequence based on each second sub-mode component and the noise signal, specifically include: Calculate the mutual information between each first sub-mode component and the vibration signal, select the first sub-mode component with the largest mutual information for reconstruction, and obtain the vibration signal sequence; Calculate the mutual information between each second sub-mode component and the noise signal, select the second sub-mode component with the largest mutual information for reconstruction, and obtain the noise signal sequence.

5. The generator fault identification method according to claim 1, characterized in that, The vibration feature matrix is ​​determined based on the vibration signal sequence; Based on the noise signal sequence, the noise feature matrix is ​​determined, specifically including: The vibration signal sequence is decomposed into multiple levels to obtain multiple vibration level components; the noise signal sequence is decomposed into multiple levels to obtain multiple noise level components. Sliding window calculations were performed on the vibration hierarchy components at different scales to obtain coarse-grained vibration sequences of the vibration hierarchy components at different scales; sliding window calculations were also performed on the noise hierarchy components at different scales to obtain coarse-grained noise sequences of the noise hierarchy components at different scales. The fuzzy entropy value of the coarse-grained vibration sequence is calculated, and the fuzzy entropy values ​​of multiple vibration hierarchical components at different scales are integrated to obtain the vibration feature matrix; the fuzzy entropy value of the coarse-grained noise sequence is calculated, and the fuzzy entropy values ​​of multiple noise hierarchical components at different scales are integrated to obtain the noise feature matrix.

6. The generator fault identification method according to claim 1, characterized in that, The first identification result and the first confidence level are determined based on the vibration feature matrix; Based on the noise feature matrix, the second identification result and the second confidence level are determined, specifically including: The vibration feature matrix is ​​input into the pre-trained vector projection extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the vibration signal. The fault type corresponding to the largest probability value is determined as the first identification result, and the largest probability value is determined as the first confidence level. The noise feature matrix is ​​input into the pre-trained maximum a posteriori probability kernel extreme learning machine model to obtain the probability values ​​of each fault type corresponding to the noise signal. The fault type corresponding to the largest probability value is determined as the second identification result, and the largest probability value is determined as the second confidence level.

7. The generator fault identification method according to claim 1, characterized in that, The process of determining the fault identification result based on the first confidence level and the second confidence level specifically includes: Calculate the difference between the first confidence level and the second confidence level to obtain the confidence level difference value; When the confidence difference value is greater than the preset difference threshold, the first confidence level and the second confidence level are compared. If the first confidence level is greater than the second confidence level, the first identification result is determined as the fault identification result; if the second confidence level is greater than the first confidence level, the second identification result is determined as the fault identification result. When the confidence difference value is less than or equal to the preset difference threshold, the operating condition data of the target generator is obtained; based on the operating condition data of the target generator, the first confidence weight of the first identification result and the second confidence weight of the second identification result are obtained from the preset fault association rule base. Multiply the first confidence level by its weight to obtain the first weighted confidence level; multiply the second confidence level by its weight to obtain the second weighted confidence level; compare the first weighted confidence level with the second weighted confidence level. If the first weighted confidence level is greater than the second weighted confidence level, the first identification result is determined as the fault identification result; If the second weighted confidence level is greater than the first weighted confidence level, the second identification result is determined as the fault identification result.

8. An electronic device for generator fault identification, characterized in that, The electronic device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-7.

9. A computer program product containing instructions, characterized in that, When the computer program product is run on an electronic device for generator fault identification, the electronic device causes the electronic device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on an electronic device for generator fault identification, the electronic device causes the electronic device to perform the method as described in any one of claims 1-7.

Images