Intelligent diagnosis method and system for running state of centrifugal fan based on deep learning

Abstract

The invention discloses a centrifugal fan running state intelligent diagnosis method and system based on deep learning, and relates to the technical field of centrifugal fan fault diagnosis and health management. According to the intelligent diagnosis method for the running state of the centrifugal fan based on deep learning, multi-working-condition vibration and multi-source data are collected and preprocessed, and weak fault features are enhanced through self-adaptive multi-scale time-frequency analysis; a parallel network is constructed, vibration deep features are extracted through physical prior attention, multi-source time sequence features are extracted through Transform, and the features are decomposed into fault sharing and working condition related features through decoupling loss and then fused; according to the method, field local feature distribution is aligned through local maximum mean value difference, and the data quality is improved through multi-source data synchronous acquisition and refined preprocessing. And by combining adaptive signal enhancement, physical prior guided feature extraction, feature decoupling fusion and local domain adaptation, working condition and domain difference interference is weakened, and accurate and stable diagnosis of the running state of the fan in a complex scene is realized.

Description

Technical Field

[0001] This invention relates to the field of centrifugal fan fault diagnosis and health management technology, specifically to a method and system for intelligent diagnosis of centrifugal fan operating status based on deep learning. Background Technology

[0002] Centrifugal fans, as key rotating machinery widely used in industrial fields, directly affect the continuity and safety of production lines. In industries such as power, metallurgy, chemical, and mining, centrifugal fans typically undertake important tasks such as ventilation, induced draft, and forced draft. A malfunction can range from affecting production efficiency to causing safety accidents and significant economic losses. Therefore, timely and accurate monitoring and fault diagnosis of centrifugal fan operation are crucial for ensuring safe equipment operation and reducing maintenance costs.

[0003] With the rapid development of sensor technology and data acquisition and processing technology, centrifugal fan fault diagnosis methods based on vibration signals have been widely applied. Vibration signals contain rich information about the equipment's operating status and can reflect various abnormal changes in the centrifugal fan during operation. Traditional vibration signal analysis methods mainly include time-domain analysis, frequency-domain analysis, and time-frequency-domain analysis. These methods extract parameters such as peak value, RMS value, kurtosis index, and spectral characteristics of the vibration signal, and combine them with threshold judgments or expert experience for fault identification. These methods can meet the diagnostic needs under steady-state operating conditions to a certain extent, but in complex industrial environments, they are often affected by background noise, operating condition fluctuations, and other factors, making it difficult to guarantee diagnostic accuracy.

[0004] In recent years, deep learning technology has demonstrated powerful feature extraction and pattern recognition capabilities in the field of mechanical equipment fault diagnosis. Convolutional neural networks, recurrent neural networks, and their variants have been widely applied to feature learning and fault classification of vibration signals, achieving better diagnostic results than traditional methods. Meanwhile, with the popularization of industrial IoT technology, multi-source monitoring data during centrifugal fan operation, such as parameters like temperature, current, pressure, and speed, can be collected and stored in real time, providing a richer information foundation for data-driven intelligent diagnosis.

[0005] The limitations of existing technologies include at least the following problems: When performing intelligent diagnosis of centrifugal fans, features are often directly extracted from the collected vibration signals for analysis. This approach has significant limitations when facing complex operating conditions. When the fan operates under variable speed and load conditions, the signal exhibits obvious non-stationary characteristics. Traditional methods rely on fixed time-frequency resolution to complete signal processing, making it difficult to simultaneously capture transient impact features and steady-state harmonic components. This can easily lead to the loss of early, weak fault information during speed changes. When existing deep learning diagnostic models migrate from laboratory environments to field conditions, they typically only align the global feature distribution, ignoring the local discriminative features corresponding to different fault types. This causes different faults that could originally be separated to overlap in the feature space and triggers a negative migration effect. In actual operation, centrifugal fans often experience multiple coupled faults simultaneously. A single shared feature extractor cannot effectively untangle the features of different fault sources. The features of different fault types modulate each other in the frequency domain and become entangled in the feature space, making the diagnostic results of composite faults unclear and difficult to accurately identify the specific fault type and its severity. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a deep learning-based intelligent diagnostic method and system for the operating status of centrifugal fans. This method solves the problems of existing technologies when facing complex signals under varying operating conditions, such as loss of early weak fault features due to fixed time-frequency resolution, aliasing of fault features during domain migration due to neglect of local discriminative features, and unclear diagnostic results due to the difficulty of decoupling complex faults by a single shared feature extractor.

[0007] To achieve the above objectives, this invention provides the following technical solution: a deep learning-based intelligent diagnostic method for the operating status of centrifugal fans, comprising the following steps: collecting vibration signals and multi-source monitoring data of centrifugal fans under multiple operating conditions, and preprocessing the collected data; performing adaptive multi-scale time-frequency analysis on the preprocessed vibration signals to enhance weak fault features in the signals, obtaining enhanced fault feature representations; constructing a parallel feature extraction network, wherein the vibration feature extraction branch takes the enhanced fault feature representations as input and uses a physical prior-guided attention mechanism to extract deep features related to faults in the vibration signals; and multi-source time-series feature extraction branch... The system takes multi-source monitoring data as input and uses the Transformer model to extract its temporal dependency features. A feature decoupling loss function is introduced to decompose the features extracted from the two branches into fault-shared features and operating condition-related features. The fault-shared features of the two branches are fused through an attention mechanism to obtain decoupled fused features. The decoupled fused features are mapped to the source domain and the target domain respectively. By minimizing the local maximum mean difference of the local discriminative features corresponding to different fault types between the source domain and the target domain, the local feature distribution between the domains is aligned to obtain domain-adaptive features. The domain-adaptive features are input into the classifier to output the diagnostic results of the centrifugal fan's operating status.

[0008] Further, the specific steps to obtain the enhanced fault feature representation are as follows: Adaptive complete ensemble empirical mode decomposition is applied to the preprocessed vibration signal to obtain several intrinsic mode function components; the mutual information entropy between each intrinsic mode function component and the original vibration signal is calculated, and intrinsic mode function components with mutual information entropy values ​​greater than a preset threshold are selected as sensitive components; morphological filtering is performed on the selected sensitive components to remove residual noise components; and the filtered sensitive components are reconstructed to obtain the enhanced fault feature representation.

[0009] Furthermore, the specific steps for extracting fault-related deep features from vibration signals using a physical prior-guided attention mechanism are as follows: The enhanced fault feature representation is input into a convolutional neural network to extract multi-layer convolutional feature maps; a physical prior weight matrix is ​​calculated based on the centrifugal fan's rotational frequency and the bearing fault feature frequency; the physical prior weight matrix is ​​multiplied element-wise with the feature map output from the last layer of the convolutional neural network to obtain a weighted feature map; global average pooling is performed on the weighted feature map to obtain the fault-related deep features from the vibration signal.

[0010] Furthermore, the specific steps for introducing a feature decoupling loss function to decompose the features extracted by the two branches into fault-shared features and operating condition-related features are as follows: The deep features extracted by the vibration feature extraction branch and the time-dependent features extracted by the multi-source temporal feature extraction branch are jointly input into the feature decoupling network; the feature decoupling network consists of two parallel fully connected layers, which output fault-shared features and operating condition-related features respectively; a feature decoupling loss function is constructed, which includes fault classification loss, operating condition classification loss, and feature orthogonality constraint loss; the feature decoupling loss function is jointly optimized, and the parameters of the feature decoupling network are iteratively updated until the network converges, yielding the decoupled fault-shared features and operating condition-related features.

[0011] Furthermore, the specific composition of the feature decoupling loss function is as follows: the feature decoupling loss function is composed of a weighted sum of fault classification loss, operating condition classification loss and feature orthogonal constraint loss; both fault classification loss and operating condition classification loss adopt cross-entropy loss; the feature orthogonal constraint loss is obtained by calculating the covariance matrix of fault-shared features and operating condition-related features.

[0012] Further, the specific steps to obtain the decoupled fused features are as follows: The fault-shared features obtained by decoupling the vibration feature extraction branch and the fault-shared features obtained by decoupling the multi-source temporal feature extraction branch are concatenated to obtain the features to be fused; the features to be fused are input into the attention mechanism module, and the attention weights corresponding to the fault-shared features of the two branches are calculated through a fully connected layer and a Softmax function; the fault-shared features of the two branches are multiplied by their corresponding attention weights to obtain the weighted branch features; the weighted features of the two branches are added element-wise to obtain the decoupled fused features.

[0013] Further, the specific steps to obtain the domain-adaptive features are as follows: The decoupled fused features are mapped to the high-dimensional regeneration kernel Hilbert spaces of the source and target domains, respectively; local discriminative features belonging to the same fault category are extracted from the source and target domains, forming multiple local feature subsets; for each fault category's corresponding local feature subset, the maximum mean difference between the source and target domain features in the regeneration kernel Hilbert space is calculated to obtain the local domain difference for that category; the local domain differences of all fault categories are weighted and summed to obtain the total loss of the local maximum mean difference; the total loss of the local maximum mean difference is minimized, and the feature mapping network parameters are iteratively updated until the network converges, thus obtaining the domain-adaptive features.

[0014] Furthermore, the specific steps for outputting the diagnostic results of the centrifugal fan's operating status are as follows: input the domain adaptive features into a multi-classifier consisting of a fully connected layer and a Softmax function; the multi-classifier calculates the probability distribution of the domain adaptive features belonging to each operating status category; and selects the operating status category with the highest probability value as the diagnostic result of the current centrifugal fan.

[0015] Furthermore, the specific steps for collecting vibration signals and multi-source monitoring data of the centrifugal fan under multiple operating conditions and preprocessing the collected data are as follows: Vibration signals are collected from the bearing housing of the centrifugal fan, along with temperature data, motor current data, outlet air pressure data, and speed data at the bearing housing as multi-source monitoring data; the collected vibration signals are sequentially processed by mean removal, trend removal, and bandpass filtering to obtain preprocessed vibration signals; outlier removal is performed on the collected temperature data, motor current data, outlet air pressure data, and speed data; normalization is performed on each of the outlier-removed data, mapping the numerical range of each data point to the same interval; the normalized data are then time-sequentially aligned according to timestamps to obtain preprocessed multi-source monitoring data.

[0016] A deep learning-based intelligent diagnostic system for centrifugal fan operation status includes: a data acquisition and preprocessing module for acquiring vibration signals and multi-source monitoring data of the centrifugal fan under multiple operating conditions and preprocessing the acquired data; a signal enhancement module for performing adaptive multi-scale time-frequency analysis on the preprocessed vibration signals to enhance weak fault features in the signals and obtain enhanced fault feature representations; and a parallel feature extraction network module, which includes a vibration feature extraction branch and a multi-source time-series feature extraction branch. The vibration feature extraction branch takes the enhanced fault feature representations as input and uses a physical prior-guided attention mechanism to extract deep fault-related features from the vibration signals; and multiple... The source temporal feature extraction branch takes multi-source monitoring data as input and uses the Transformer model to extract its temporal dependency features. The feature decoupling and fusion module introduces a feature decoupling loss function to decompose the features extracted by the two branches into fault-shared features and operating condition-related features, and fuses the fault-shared features of the two branches through an attention mechanism to obtain the decoupled fused features. The domain adaptation module maps the decoupled fused features to the source domain and the target domain respectively, and aligns the local feature distribution between the domains by minimizing the local maximum mean difference to obtain the domain-adaptive features. The diagnostic output module inputs the domain-adaptive features into the classifier and outputs the diagnostic results of the centrifugal fan's operating status.

[0017] The present invention has the following beneficial effects: (1) The intelligent diagnostic method for the operating status of centrifugal fan based on deep learning decomposes the non-stationary signal into multiple intrinsic mode function components by adaptive complete set empirical mode decomposition of the collected vibration signal. Then, by calculating the mutual information entropy between each component and the original signal, sensitive components containing fault information are selected. Then, morphological filtering is performed on the sensitive components to remove residual noise. Finally, the processed components are reconstructed into an enhanced fault feature representation. This changes the limitation of the traditional fixed time-frequency resolution analysis method in losing transient impact features when facing speed changes and load fluctuations. It enables the effective enhancement and extraction of weak fault features such as early bearing micro-damage and slight blade wear hidden in strong background noise, providing higher quality signal input for subsequent diagnosis. This avoids the equipment from operating with faults due to failure to detect faults in time, which leads to the gradual expansion of faults.

[0018] (2) The intelligent diagnostic method for centrifugal fan operation status based on deep learning takes the enhanced fault feature representation as input through the vibration feature extraction branch, uses the convolutional neural network to extract multi-layer feature maps, and calculates the physical prior weight matrix according to the centrifugal fan's rotation frequency and bearing fault feature frequency. The physical prior weight matrix is ​​multiplied element by element with the feature map so that the network focuses on the frequency band region related to the fault. The multi-source time-series feature extraction branch takes multi-source monitoring data such as temperature, current, and pressure as input, and uses the Transformer model to extract its time-series dependent features. The features extracted by the two branches are input into the feature decoupling network. Through the fault classification loss, operating condition classification loss and feature orthogonal constraint loss in the feature decoupling loss function, the features are decomposed into fault-shared features and operating condition-related features. Then, the fault-shared features of the two branches are fused through the attention mechanism, thereby effectively solving the problem of different fault types modulating each other in the frequency domain and entangled in the feature space. This allows the features of various faults in the composite fault to be clearly separated and identified, and can accurately output the specific fault type and its severity.

[0019] (3) The intelligent diagnostic method for the operating status of centrifugal fans based on deep learning maps the decoupled fused features to the high-dimensional regeneration kernel Hilbert space of the source domain and the target domain respectively. It extracts local discriminative features belonging to the same fault category in the source domain and the target domain respectively to form multiple local feature subsets. For each local feature subset corresponding to the fault category, it calculates the maximum mean difference between the source domain features and the target domain features in the regeneration kernel Hilbert space to obtain the local neighborhood difference of the category. It weights and sums the local neighborhood differences of all fault categories to obtain the total loss of the maximum mean difference of the local difference. By minimizing this loss function, iteratively updating the feature mapping network parameters, the local feature distribution of the same fault category in the source domain and the target domain is aligned. This changes the limitation of the traditional domain adaptive method that only aligns the global feature distribution and ignores the local discriminative features corresponding to different fault types. It effectively avoids the negative migration effect of the neighborhood caused by the aliasing of different fault types in the feature space, so that the model can still maintain stable diagnostic performance when migrating between different working conditions and different equipment.

[0020] (4) The intelligent diagnostic system for centrifugal fan operation status based on deep learning completes the acquisition and preprocessing of vibration signals and multi-source monitoring data through the data acquisition and preprocessing module. Then, the signal enhancement module performs adaptive multi-scale time-frequency analysis on the vibration signal to enhance weak fault features. Next, the parallel feature extraction network module extracts deep features and time-dependent features from the enhanced vibration signal and multi-source monitoring data respectively. The feature decoupling and fusion module decomposes the extracted features into fault-shared features and operating condition-related features and fuses them through an attention mechanism. The domain adaptation module aligns the local feature distribution between domains of the fused features. Finally, the diagnostic output module inputs the domain adaptation features into the classifier to obtain the diagnostic results. The functions of each module are clearly divided and the data is processed in sequence according to the diagnostic process, which is convenient for deployment and application in actual industrial sites.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of a deep learning-based intelligent diagnostic method for the operating status of centrifugal fans according to the present invention.

[0023] Figure 2 This is a flowchart illustrating the specific steps involved in extracting fault-related deep features from vibration signals using a physical prior-guided attention mechanism in a deep learning-based intelligent diagnostic method for centrifugal fan operation status.

[0024] Figure 3 This is a block diagram of a deep learning-based intelligent diagnostic system for the operating status of centrifugal fans according to the present invention. Detailed Implementation

[0025] Please see Figure 1 This invention provides a technical solution: a deep learning-based intelligent diagnostic method for the operating status of centrifugal fans, comprising the following steps: collecting vibration signals and multi-source monitoring data of centrifugal fans under multiple operating conditions, and preprocessing the collected data; performing adaptive multi-scale time-frequency analysis on the preprocessed vibration signals to enhance weak fault features in the signals, obtaining enhanced fault feature representations; constructing a parallel feature extraction network, wherein the vibration feature extraction branch takes the enhanced fault feature representations as input and uses a physical prior-guided attention mechanism to extract deep features related to faults in the vibration signals; and the multi-source time-series feature extraction branch uses multi-source... The monitoring data is used as input, and its temporal dependency features are extracted using the Transformer model. A feature decoupling loss function is introduced to decompose the features extracted from the two branches into fault-shared features and operating condition-related features. The fault-shared features of the two branches are fused through an attention mechanism to obtain the decoupled fused features. The decoupled fused features are mapped to the source domain and the target domain respectively. By minimizing the local maximum mean difference of the local discriminative features corresponding to different fault types between the source domain and the target domain, the local feature distribution between the domains is aligned to obtain the domain adaptive features. The domain adaptive features are input into the classifier to output the diagnostic results of the centrifugal fan's operating status.

[0026] Among them, multiple operating conditions include the operating conditions of centrifugal fans at different speeds and under different loads; For example, the speed range is 1000r / min-3000r / min, and the load range is 50%-100% of the rated load, covering the main operating conditions of centrifugal fans in actual industrial operation; Multi-source monitoring data and vibration signals are collected synchronously, with the sampling frequency uniformly set to 10kHz-50kHz to avoid feature misalignment caused by asynchronous sampling.

[0027] Specifically, the steps to obtain the enhanced fault feature representation are as follows: Adaptive complete ensemble empirical mode decomposition is applied to the preprocessed vibration signal to obtain several eigenmode function components, specifically: The vibration signal X(t) is decomposed layer by layer using Adaptive Complete Ensemble Empirical Mode Decomposition (ACEEMD). Adaptive Gaussian white noise is added during the decomposition process to suppress mode aliasing, resulting in K eigenmode function components arranged sequentially from high frequency to low frequency. (The value of K ranges from 5 to 15, and can be adjusted according to the complexity of the vibration signal, for example, K=8). Each intrinsic mode function component satisfies the following core conditions: The number of extreme points is equal to or differs from the number of zero-crossing points by no more than 1. Second, the upper and lower envelopes are locally symmetrical about the time axis; After decomposition, the remaining signal components that are difficult to decompose further are called residual components res(t), i.e.: ; The residual components do not contain fault features and will not be used for feature extraction in the future. The mutual information entropy between each intrinsic mode function component and the original vibration signal is calculated, and intrinsic mode function components with mutual information entropy values ​​greater than a preset threshold are selected as sensitive components. Specifically: Mutual information entropy is used to characterize the degree of correlation between two random variables. The larger the value, the stronger the correlation. The information entropy H(x) of the original vibration signal X(t) is calculated using the following formula: ; in, Let be the probability of the i-th value after discretization of X(t), and n be the number of values ​​after discretization; Calculate each intrinsic mode function component Mutual information entropy with X(t) The calculation formula is: in, for Information entropy For X(t) and The joint information entropy; Set the filter threshold , Calibrate based on experimental data; For example , retain satisfaction The intrinsic mode function components are used as fault-sensitive components; For example, filtering Three sensitive components, which contain the richest range of fault characteristics; The selected sensitive components are subjected to morphological filtering to remove residual noise components, specifically as follows: A combination of morphological opening and closing operations is performed using flat structural elements. The opening operation is used to remove positive impulse noise from the sensitive components, and the closing operation is used to remove negative impulse noise. The length L of the structural element is matched with the fault impact characteristic scale and is determined based on the vibration signal sampling frequency and the fault characteristic frequency. For example, when the sampling frequency is 10kHz and the bearing fault characteristic frequency is 100Hz, the length of the structural element L=50 (corresponding to a time of 5ms). The filtering process is as follows: For sensitive components Perform an opening operation to obtain In this formula This represents a single sensitive component that has been selected. It is the standard symbol for morphological opening operation, and B is a flat structuring element; Then perform a closing operation on the result of the opening operation to obtain... ,in, It is the standard symbol for morphological closing operations. This is the intermediate signal after the opening operation. This is the sensitive component obtained after removing negative impulse noise through a closing operation; The filtered sensitive components are then reconstructed to obtain the enhanced fault feature representation, which is as follows: All sensitive components after noise filtering are linearly superimposed in the order of decomposition, and the reconstruction formula is: ; Where S is the index set of sensitive components (S corresponds to the sensitive components selected above, such as...) (the set of indices); Reconstructed enhanced vibration signal In this process, the amplitude of weak fault characteristics is improved (e.g., the fault impact amplitude is increased from 0.1V to 0.3V), and the amplitude of noise components is suppressed (e.g., the noise amplitude is reduced from 0.05V to 0.01V), which can more clearly reflect the impact characteristics and periodic characteristics caused by the fault, providing high-quality signals for subsequent feature extraction.

[0028] In this implementation scheme, by performing adaptive complete set empirical mode decomposition on the vibration signal, mode aliasing can be effectively suppressed under complex working conditions, making the signal decomposition results more stable and reliable. By combining mutual information entropy to complete sensitive component screening, signal components with a higher degree of correlation with the fault can be retained, redundant information can be eliminated, and subsequent processing can be more targeted. By using a combination of morphological opening and closing operations, positive and negative impulse noise can be removed respectively, improving signal quality while retaining the fault impact characteristics. By completing signal reconstruction according to the sensitive component index set, fault information that was originally hidden in the noise can be highlighted, making the impact and periodic changes caused by the fault clearer.

[0029] Specifically, such as Figure 2 As shown, the specific steps for extracting fault-related deep features from vibration signals using a physics-prior-guided attention mechanism are as follows: The enhanced fault feature representation is input into a convolutional neural network to extract multi-layer convolutional feature maps, specifically as follows: Representing the enhanced fault characteristics Converted into a two-dimensional time-frequency graph using Short Time Fourier Transform (STFT). , where H is the height of the time-frequency plot (corresponding to the frequency dimension) and W is the width of the time-frequency plot (corresponding to the time dimension). For example, H=256, W=256; The two-dimensional time-frequency graph is input into a convolutional neural network (CNN). The CNN contains 4 convolutional layers, 4 batch normalization layers, 4 ReLU activation layers and 2 max pooling layers. The convolutional kernel sizes are 3×3, 3×3, 5×5 and 5×5 respectively, and the convolutional stride is 1. The pooling kernel size is 2×2 and the pooling stride is 2. Feature maps are extracted from different receptive fields through multiple convolutional, activation, and pooling operations. The size of the feature map output by the last convolutional layer is denoted as H×W×C, where H=32, W=32, and C=64 (C is the number of channels). ; The physical prior weight matrix is ​​calculated based on the centrifugal fan's rotational frequency and the bearing failure characteristic frequency, specifically as follows: Obtain the core parameters of the centrifugal fan; For example: bearing model 6205, number of rolling elements Z=9, rolling element diameter d=7.94mm, bearing pitch circle diameter D=30mm, shaft speed n=1500r / min (rotation frequency) ); Calculate the typical fault characteristic frequencies based on bearing geometric parameters: Inner ring failure frequency ; Outer ring failure frequency ; Rolling element failure frequency ; in, The contact angle between the rolling elements of the bearing and the inner and outer rings, for example... The calculation yields: ; ; Based on these fault characteristic frequencies in the time-frequency diagram At the corresponding positions in the matrix, construct a physical prior weight matrix with the same size as the last layer feature map F. The weight values ​​for the locations corresponding to fault features are set to 0.8-1.0 (e.g., 0.9), and the weight values ​​for non-fault areas are set to 0.1-0.3 (e.g., 0.2), so that the weight matrix can guide the network to focus on the fault feature areas. The physical prior weight matrix is ​​multiplied element-wise with the feature map output from the last layer of the convolutional neural network to obtain the weighted feature map, which is as follows: The physical prior weight matrix M is fused with the feature map F output from the last layer of the convolutional neural network using element-wise multiplication (Hadamard product). The calculation formula is as follows: ; in, This represents element-wise multiplication (this notation is the standard representation of the Hadamard product), and the resulting fusion yields a weighted feature map. ; This operation can amplify the signal of the fault-related feature region and suppress irrelevant background features. For example, the feature value of the fault feature region is increased from 0.5 to 0.9, and the feature value of the non-fault region is decreased from 0.4 to 0.1, making the network pay more attention to the feature information related to the fault. Global average pooling is performed on the weighted feature map to obtain the deep features related to the fault in the vibration signal, specifically: Weighted feature map Perform a global average pooling operation on the spatial dimension (H×W), averaging all elements of each channel, using the following formula: ; in, ; Finally, a one-dimensional deep feature vector with a length consistent with the number of channels C is obtained. ; For example, when C=64 It is a 64-dimensional vector; This vector integrates prior physical knowledge with data-driven features, effectively distinguishing different types of faults and possessing stronger fault detection capabilities.

[0030] In this implementation scheme, the enhanced fault features are converted into a two-dimensional time-frequency map and fed into a multi-layer convolutional neural network. This allows the network to extract spatial and frequency information from the signal at different scales, enabling it to fully learn the distribution characteristics of the fault in the time-frequency domain. By combining the actual operating parameters of the wind turbine with the bearing structure, the fault feature frequency is calculated, thereby constructing a physical prior weight matrix. This integrates real equipment operation knowledge into the network, avoiding biases caused by relying solely on data-driven approaches. By multiplying the weight matrix and feature map element by element, the features of the fault-corresponding region can be automatically amplified while suppressing irrelevant background interference, allowing the network to focus more on effective information. Then, global average pooling is used to obtain one-dimensional deep features, thereby compressing the spatial distribution information into a compact and discriminative feature vector.

[0031] Specifically, the steps for introducing a feature decoupling loss function to decompose the features extracted from the two branches into fault-shared features and operating condition-related features are as follows: The deep features extracted by the vibration feature extraction branch and the temporal-dependent features extracted by the multi-source temporal feature extraction branch are jointly input into the feature decoupling network, specifically as follows: The deep features output by the vibration feature extraction branch are (For example (corresponding to the CNN output channel number C=64 mentioned earlier), the temporal dependency features output by the multi-source temporal feature extraction branch are: (For example ); The two feature vectors are concatenated along their feature dimensions to form a joint feature vector. The splicing formula is: (T represents matrix transpose, converting column vectors to row vectors and then concatenating them); the joint eigenvectors The input is fed into the feature decoupling network, serving as the network's input features. The input dimension of the feature decoupling network is... ; The feature decoupling network consists of two parallel fully connected layers, which output fault-sharing features and operating condition-related features, respectively. The feature decoupling network consists of two parallel fully connected layer branches, with the joint feature vector serving as the input to both branches. ; The first fully connected layer (fault-sharing feature branch) contains two fully connected layers. The first layer has an output dimension of 64, and the second layer has an output dimension of 32, ultimately outputting the fault-sharing features. This characteristic remains stable under different operating conditions (such as different speeds and different loads) and is strongly correlated with the fault type, and is not affected by changes in operating conditions. The second fully connected layer (the working condition-related feature branch) also contains two fully connected layers. The first layer has an output dimension of 32, and the second layer has an output dimension of 16, ultimately outputting the working condition-related features. This characteristic changes with operating conditions such as speed and load, and is unrelated to the type of fault; it only reflects the operating condition. A feature decoupling loss function is constructed, which includes fault classification loss, operating condition classification loss, and feature orthogonality constraint loss, specifically as follows: The feature decoupling loss function consists of three weighted parts, used to simultaneously achieve fault identification, operating condition identification, and feature separation. The specific expression is as follows: ; in, Set the weight coefficients for each loss component (the sum of the weight coefficients does not have to be 1, and can be adjusted according to actual needs), and set them according to actual diagnostic requirements; For example , , ; The fault classification loss is used to supervise fault-shared features to achieve accurate fault category differentiation. The operating condition classification loss is used to supervise operating condition-related features to achieve accurate distinction between operating condition categories. The orthogonal constraint loss is used to constrain fault-sharing features and condition-related features to be orthogonal and uncorrelated, ensuring the decoupling effect of features. The joint optimization of the feature decoupling loss function is used to iteratively update the feature decoupling network parameters until the network converges, resulting in the decoupled fault-sharing features and operating condition-related features, which are as follows: To minimize the feature decoupling loss function To optimize the objective, an adaptive moment estimation optimizer (Adam) is used to iteratively update the parameters of the fully connected layers of the feature decoupling network. The learning rate of the optimizer is set to... attenuation coefficient , The weight decay coefficient is ; During the iteration process, the validation set loss is calculated every 10 epochs (i.e., traversing all training samples once). The validation set loss is calculated when the change in the validation set loss is less than a certain value over 5 consecutive epochs. At that time, the network is determined to have converged; After network convergence, output the decoupled fault-sharing features. Operating condition related characteristics At this point, the correlation between the two types of features is less than 0.1, achieving effective decoupling.

[0032] The specific structure of the feature decoupling loss function is as follows: The feature decoupling loss function is a weighted average of the fault classification loss, the operating condition classification loss, and the feature orthogonality constraint loss, specifically as follows: The feature decoupling loss function is a weighted sum of three losses, each constraining a different objective of feature decoupling, collectively ensuring complete separation between fault-shared features and condition-related features. The specific expression is as follows: ; in, The weighting coefficients are used to balance the contributions of the three losses. They are adjusted according to the priority of the actual diagnostic task to ensure that the decoupled network simultaneously takes into account the accuracy of fault identification, the accuracy of working condition identification, and the feature separation effect. Both fault classification loss and operating condition classification loss use cross-entropy loss, specifically as follows: Fault Classification Loss Multi-class cross-entropy loss is used to supervise fault-shared features. To achieve accurate fault category differentiation, the calculation formula is as follows: ; Where N is the sample size. The number of fault categories (e.g., 6 categories: normal, unbalanced, misaligned, bearing, loose, and combined faults). The fault category label for the i-th sample is (one-hot encoded, i.e., the correct category is 1 and the other categories are 0). Let be the predicted probability that the i-th sample belongs to the c-th type of fault; Losses classified by operating condition The multi-class cross-entropy loss is also used to supervise the relevant features of the working conditions. To achieve accurate differentiation of operating conditions, the calculation formula is as follows: ; in, The number of operating conditions (e.g., 4 categories: 1000r / min light load, 1500r / min medium load, 2000r / min heavy load, 3000r / min full load). The working condition category label for the i-th sample (one-hot encoded, consistent with the fault category label encoding method). Let be the predicted probability that the i-th sample belongs to the k-th working condition; The orthogonal constraint loss is obtained by calculating the covariance matrix of fault-shared features and operating condition-related features. The weight coefficients of each loss component are set according to actual diagnostic needs to achieve effective decoupling between fault-shared features and operating condition-related features. Specifically: Feature Orthogonal Constraint Loss By calculating the fault-sharing feature vector Feature vectors related to operating conditions The Frobenius norm of the covariance matrix between the two variables is obtained, and the calculation formula is as follows: ; in , For the fault-shared features of the i-th sample, The mean of the shared fault features across all samples. For the i-th sample, the relevant features of the working conditions are... The mean of the relevant features for all sample operating conditions. This represents the Frobenius norm (i.e., the square root of the sum of the squares of all elements of a matrix). By setting up appropriately The value of allows the decoupled network to simultaneously handle the three tasks of fault identification, operating condition identification, and feature separation, ensuring that the fault-sharing features after decoupling are not affected by operating conditions, and that the operating condition-related features do not contain fault information.

[0033] The specific steps to obtain the decoupled fused features are as follows: The fault-shared features obtained by decoupling the vibration feature extraction branch and the fault-shared features obtained by decoupling the multi-source time-series feature extraction branch are concatenated to obtain the features to be fused, which are as follows: The fault-sharing features of the vibration feature extraction branch decoupling output are as follows: (Originated from the decoupling of deep vibration features, with dimensions consistent with the previous text) Consistent), the fault-sharing feature output of the multi-source temporal feature extraction branch decoupling is (Originated from decoupling of multi-source temporal dependency features, with the same dimension as...) Consistency is ensured to guarantee the feasibility of splicing. The two fault-shared feature vectors are concatenated along their feature dimensions to form a higher-dimensional feature vector to be fused. The splicing formula is: ; The feature to be fused contains fault-related information from both vibration signals and multi-source monitoring signals, which can overcome the limitations of a single signal feature. For example, vibration signal fault features reflect bearing impact, while multi-source signal fault features reflect abnormal motor current. After splicing, the fault state can be comprehensively characterized. The features to be fused are input into the attention mechanism module, and the attention weights corresponding to the fault-shared features of the two branches are calculated through a fully connected layer and a softmax function. Specifically: Features to be fused The input attention mechanism module performs feature transformation through two fully connected layers. The first fully connected layer has an input dimension of 64 and an output dimension of 32, with ReLU as the activation function. The second fully connected layer has an input dimension of 32 and an output dimension of 2 (corresponding to two branches: vibration feature extraction branch and multi-source temporal feature extraction branch). The output is normalized using the Softmax function to obtain the attention weights corresponding to the fault-shared features of the two branches. and The calculation formula is: ; ; in, This is the output value of the second fully connected layer, and (The weights sum to 1 after normalization, ensuring a reasonable weight allocation.) The weight reflects the importance of the branch's characteristics to fault diagnosis; For example, when a bearing fails. , The characteristics of the vibration branch contribute more; In case of imbalance fault, , Multi-source temporal branch features contribute more; The fault-shared features of the two branches are weighted and multiplied by their corresponding attention weights to obtain the weighted branch features, which are as follows: Shared characteristics of vibration branch faults Multiply by its corresponding attention weight The weighted vibration branch characteristics were obtained. ; Sharing characteristics of multi-source time-series branch faults Multiply by its corresponding attention weight The weighted multi-source temporal branch features were obtained. ; This operation can highlight the feature contribution of important branches, suppress the interference of secondary branches, and make the fused features more targeted. The weighted features from the two branches are summed element-wise to obtain the decoupled fused features, specifically: The features after weighting the two paths and Element-wise addition along the same dimension yields the final decoupled fusion feature: , (The dimension is consistent with the single-branch fault sharing characteristics to ensure compatibility in subsequent processing). This feature integrates multi-source information and eliminates operating condition interference, while highlighting features that are more important for fault diagnosis. It has stronger robustness and discriminative power, and can effectively improve the accuracy of fault diagnosis under complex operating conditions.

[0034] In this implementation scheme, by concatenating vibration features and temporal features and feeding them into a feature decoupling network, multi-source information can be processed simultaneously. This allows the network to distinguish between fault-related and operating condition-related content from the joint features. By employing a combined loss function that includes fault classification loss, operating condition classification loss, and feature orthogonal constraint loss, fault identification, operating condition differentiation, and feature separation can be completed simultaneously during training. This ensures that fault-related features are not affected by operating conditions. Orthogonal constraints allow the two types of features to be independent of each other, avoiding information mixing and improving the purity of feature representation. After decoupling, an attention mechanism is used to perform weighted fusion of fault features from the two branches, automatically highlighting information more valuable for diagnosis and weakening interference components. The element-wise weighted fusion method can maintain the consistency of feature dimensions and ensure smooth subsequent processing.

[0035] Specifically, the steps to obtain domain-adaptive features are as follows: The decoupled fused features are mapped to the high-dimensional reproducing kernel Hilbert spaces of the source and target domains, respectively, as follows: source domain ,in Let be the decoupling and fusion feature of the i-th sample in the source domain. For the corresponding fault label, The number of source domain samples (e.g., 1000) is the labeled data collected in a laboratory environment (i.e., data containing fault labels). target domain ,in For the decoupling and fusion features of the j-th sample in the target domain, The target domain sample size is 800. The target domain data consists of unlabeled data collected in an industrial field environment (i.e., data that does not contain fault labels). Through nonlinear mapping function Source domain fusion features Features fused with the target domain Mapping to the high-dimensional reproducing kernel Hilbert space H respectively, we obtain and ; The mapping function is implemented using a Gaussian kernel function, the expression of which is: ; in, For example, kernel width ; After mapping, the feature distributions of the source and target domains, which were originally linearly inseparable, present a distinguishable distribution structure in the high-dimensional space, which facilitates subsequent domain alignment operations. Local discriminative features belonging to the same fault category are extracted from both the source and target domains, forming multiple subsets of local features, specifically: Based on the fault category, the source and target domain features are divided into multiple local feature subsets. Assume the number of fault categories is... (For example, 6 types), then the local feature subset of the c-th type of fault is: ; in, The true fault label for the j-th sample in the target domain (used only for loss calculation, not for training, used for domain alignment calculation, and not involved in the model training process); Each local feature subset contains only the features of the same fault type in the source and target domains. This is used to achieve domain alignment according to fault type and avoid mutual interference between features of different fault types. For example, the local feature subset of type 1 (normal state) contains the normal sample features of the source domain and the normal sample features of the target domain, and the local feature subset of type 2 (imbalanced fault) contains the imbalanced fault sample features of the source domain and the imbalanced fault sample features of the target domain. For each fault category, the maximum mean difference between the source domain features and the target domain features in the regeneration kernel Hilbert space is calculated to obtain the local neighborhood difference for that category. Specifically: Local feature subsets for each type of fault The maximum mean difference (MMD, a commonly used metric for measuring the difference between two distributions) between source domain feature samples and target domain feature samples in high-dimensional space H is calculated as the local neighborhood difference for that category. The calculation formula is as follows: ; in, For the feature subset of the c-th type of fault in the source domain, Its sample size, For the feature subset of the c-th type of fault in the target domain, Its sample size, Let H be the norm in the high-dimensional space H; A larger MMD value indicates a greater difference in the characteristic distribution of the same fault between the source and target domains, and vice versa. For example, in a normal state... bearing failure This indicates that the distribution of bearing faults differs more significantly between the source and target domains; The total loss, which is the weighted sum of the local neighborhood differences across all fault categories, is obtained by summing the differences in local maximum mean differences. Specifically: The local maximum mean difference corresponding to each fault category According to the preset weight By performing a weighted summation, we obtain the total loss of the global and local maximum mean difference. The calculation formula is: ; in, The weights for type c faults (the sum of the weights is 1 to ensure a reasonable scale for the total loss) are adjusted based on the number of samples and the differences in distribution for each type of fault. The smaller the number of samples and the greater the differences in distribution, the greater the weight. For example (Normal state) (Bearing failures) Ensure that failure categories with large distribution differences are more fully aligned with the domains; Minimize the total loss due to the local maximum mean difference, iteratively update the feature mapping network parameters until the network converges, and obtain the domain-adaptive features, specifically as follows: To minimize the total loss of local maximum mean difference To achieve the goal, iteratively update the feature mapping network (i.e., implement nonlinear mapping). The network parameters are set, the optimizer is the Adam optimizer, and the learning rate is set to... ; During the iteration process, the calculation is performed once every 20 epochs. Value, when The value tends to stabilize and the change over five consecutive epochs is less than 1. At that time, the network is determined to have converged; After the network converges, the target domain features are fused. Through mapping function The obtained features are used as domain-adaptive features. ; At this point, the characteristic distribution of the same fault in the source and target domains tends to be consistent, and the difference between domains is significantly reduced, for example, in bearing faults. The value was reduced from 0.25 to 0.08 to ensure that the model can maintain high diagnostic accuracy under field conditions.

[0036] In this implementation scheme, the decoupled fused features are mapped to a high-dimensional regenerative kernel Hilbert space, which breaks the original linear inseparability of features between the source and target domains. This makes the two types of features more easily distinguishable in the high-dimensional space. By dividing the local feature subsets according to the fault category, domain alignment can be performed separately by category, avoiding mutual interference between different fault features. By calculating the local maximum mean difference of each category, the distribution difference of the same fault in different domains can be accurately measured. The total loss is obtained by weighted summation, which can focus on fault categories with large distribution differences to ensure balanced alignment. By iteratively updating the network parameters by minimizing the total loss, the feature distribution of the same fault in the source and target domains can be made more consistent, significantly reducing domain differences. This effectively solves the problem of distribution deviation between laboratory labeled data and unlabeled data in industrial fields, ensuring that the model can still play a stable diagnostic role under field conditions.

[0037] Specifically, the steps for inputting domain-adaptive features into the classifier and outputting the diagnostic results of the centrifugal fan's operating status are as follows: The domain-adaptive features are input into a multi-classifier consisting of fully connected layers and a softmax function, specifically as follows: Domain-adaptive features after domain adaptation and feature decoupling (For example, d=64) Input a multi-classifier, which consists of two fully connected layers and a Softmax activation function; The first fully connected layer has an input dimension of d=64, an output dimension of 32, and an activation function of ReLU, which is used to map the domain adaptive features to a low-dimensional feature space. The second fully connected layer has an input dimension of 32 and an output dimension of [missing value]. (Corresponding to 6 operating state categories), used to map features to fault categories; the Softmax activation function is connected after the second fully connected layer to convert the output into a probability distribution; The multi-classifier calculates the probability distribution of adaptive features belonging to each operational state category, which is as follows: The multi-classifier adapts to features from the domain through two fully connected layers. Perform feature transformation to obtain the original output vector. Then, the original output vector is mapped to probability values ​​between 0 and 1 using the Softmax function. The calculation formula is as follows: ; in, Let be the probability that the domain adaptive feature belongs to the c-th class of running states. This is the c-th element of the original output vector; The sum of the probabilities of all categories is 1, that is... ; For example, the probability distribution of a sample is This indicates that the probability of this sample belonging to category 4 (bearing failure) is 92%. The operating state category with the highest probability value is selected as the current diagnostic result for the centrifugal fan. The operating state categories include normal state, imbalance fault, misalignment fault, bearing fault, looseness fault, and combined fault, specifically: Compare the outputs of multiple classifiers We calculate the probability of each category, find the category with the highest probability value, and denote it as... ( To find the function that maximizes the probability of the category index, The corresponding operating status category is used as the diagnostic result for the current centrifugal fan; At the same time, set a probability threshold. (For example When the maximum probability When necessary, output "Diagnosis uncertain, please conduct further testing" to avoid misdiagnosis; The specific definitions of various operating states are as follows: Under normal conditions, all components of the fan operate normally, with no abnormal vibration or abnormal current. Imbalance faults are caused by uneven mass distribution in the impeller, which leads to periodic vibrations. The misalignment fault is caused by a deviation in the coaxiality of the motor and the fan shaft, which leads to abnormal impact. Bearing failure is caused by bearing wear and fatigue, which leads to high-frequency impacts. The loosening fault is caused by a loose fan base or connecting parts, which leads to low-frequency vibration. A compound fault is the simultaneous presence of two or more single faults (such as imbalance + misalignment fault).

[0038] In this implementation scheme, the domain-adaptive features are input into a multi-classifier consisting of a fully connected layer and a Softmax function. Through feature transformation of two fully connected layers, high-dimensional features can be mapped to a low-dimensional space suitable for fault classification, simplifying the calculation while retaining key information for fault differentiation. The Softmax function can convert the original output of the fully connected layer into a reasonable probability distribution, which facilitates intuitive judgment of the fault category. The category with the highest probability is selected as the diagnostic result, which can quickly locate the current operating status of the wind turbine. At the same time, setting a probability threshold can effectively avoid misjudgment caused by excessively low probability and improve the reliability of the diagnostic results.

[0039] Specifically, the steps for collecting vibration signals and multi-source monitoring data of the centrifugal fan under multiple operating conditions and preprocessing the collected data are as follows: Vibration signals were collected at the centrifugal fan bearing housing, along with temperature data, motor current data, outlet air pressure data, and speed data at the bearing housing, serving as multi-source monitoring data. Specifically: Vibration signals at the bearing housing of the centrifugal fan are collected using piezoelectric accelerometers. The sensors are installed in the horizontal and vertical directions of the bearing housing (one each). The sampling frequency is set to 10kHz-50kHz (e.g., 20kHz), and the sampling duration is set to 10s-60s (e.g., 30s). 10-20 sets of samples are collected for each operating condition. A PT100 temperature sensor was used to collect temperature data at the bearing housing. The sampling frequency was 1Hz, the measurement range was -50℃ to 150℃, and the accuracy was ±0.1℃. The three-phase current data of the motor is collected using a current sensor with a sampling frequency of 1kHz, a measurement range of 0-50A, and an accuracy of ±0.01A. The fan outlet air pressure data is collected using a pressure sensor with a sampling frequency of 1Hz, a measurement range of 0-1.0MPa, and an accuracy of ±0.001MPa. A photoelectric encoder is used to collect fan speed data. The sampling frequency is 10Hz, the measurement range is 0-3000r / min, and the accuracy is ±1r / min. All sensors are connected to the data acquisition card to achieve synchronous data acquisition and ensure the time consistency of each data point. The collected vibration signal is subjected to mean removal, trend removal, and bandpass filtering in sequence to obtain the preprocessed vibration signal, as follows: Mean removal is used to eliminate DC component interference. The calculation formula is as follows: ; in, N represents the number of vibration signal sampling points, and X(t) represents the original vibration signal. The vibration signal after removing the mean; Detrending term processing is used to eliminate linear trends in the signal (such as trends caused by sensor drift). The least squares method is used to fit the linear trend term. Where a is the slope and b is the intercept, the detrended signal is ; Bandpass filtering is used to retain the signal within the fault characteristic frequency range while filtering out low-frequency interference (such as base vibration) and high-frequency noise (such as electromagnetic interference). The filtering range is determined according to the fault characteristic frequency, for example, 50Hz-1000Hz. A Butterworth filter is used, with a filter order of 4th order, a passband ripple of less than 1dB, and a stopband attenuation of greater than 40dB. The filtered signal is the pre-processed vibration signal. ( (This serves as the input signal for subsequent adaptive multi-scale time-frequency analysis). Outlier removal was performed on the collected temperature data, motor current data, outlet air pressure data, and speed data. Specifically, the 3σ criterion was used for outlier removal (the 3σ criterion is a commonly used method in statistics to determine outliers, i.e., data deviating from the mean by more than three standard deviations is considered outlier). For each type of multi-source data... (M is the number of data points), calculate the mean. and standard deviation When data points satisfy When this happens, the data point is determined to be an outlier; For example, the average temperature data Standard deviation Data points with temperatures greater than 60℃ or less than 30℃ are considered outliers. After outlier removal, missing data points are filled using linear interpolation to ensure data continuity. The interpolation formula is as follows: ; in, The value is the interpolated value. These are the normal data points before and after the outlier, where i is the index of the outlier. After outlier removal, each data point is normalized to map its numerical range to the same interval. Specifically: The min-max normalization method (a linear normalization method that maps data to a specified interval) is used to map various multi-source data to the [0,1] interval. The calculation formula is as follows: ; in, Here, y represents the normalized data, and y represents the original data after outlier removal. The minimum value of this type of data. The maximum value of this type of data; For example, temperature data , Then, the normalized value of 45℃ is 0.5; Normalization can eliminate the influence of different dimensions on feature extraction, ensuring that vibration signals and multi-source monitoring data are fused at the same scale; The normalized data are aligned according to their timestamps to obtain the preprocessed multi-source monitoring data, which is as follows: All data includes timestamp information (accurate to milliseconds). Based on the timestamp of the vibration signal, the timestamps of temperature, motor current, outlet air pressure, and speed data are matched with the timestamp of the vibration signal. For data points with non-overlapping timestamps, linear interpolation is used to supplement the data with the corresponding timestamp. After alignment, each vibration signal sampling point corresponds to a set of synchronized multi-source monitoring data, forming a time-aligned multi-source dataset. Where T(t) is temperature data, I(t) is motor current data, P(t) is outlet air pressure data, and n(t) is speed data, this ensures that the time-dependent feature extraction of the subsequent multi-source time-series feature extraction branch is accurate and guarantees time-series consistency.

[0040] In this implementation plan, the vibration signal of the centrifugal fan and various monitoring data are collected simultaneously, which can comprehensively capture various status information during equipment operation, avoid missing key abnormal clues from a single signal, and provide rich data support for subsequent diagnosis. The vibration signal is processed by removing the mean, removing the trend term, and bandpass filtering, which can effectively eliminate DC interference, trend deviation and irrelevant noise, and retain the signal components related to the fault. Appropriate methods are used to remove outliers from multi-source data and fill in missing points, which can ensure the continuity and reliability of data and avoid abnormal data from affecting the diagnostic results. Normalization processing eliminates the dimensional differences of different data, ensuring that various types of data are fused and analyzed on the same scale. Time sequence alignment is completed according to timestamps, so that the vibration signal corresponds synchronously with the multi-source monitoring data.

[0041] Please see Figure 3This invention provides a technical solution: a deep learning-based intelligent diagnostic system for the operating status of a centrifugal fan, comprising: a data acquisition and preprocessing module for acquiring vibration signals and multi-source monitoring data of the centrifugal fan under multiple operating conditions, and preprocessing the acquired data; a signal enhancement module for performing adaptive multi-scale time-frequency analysis on the preprocessed vibration signals to enhance weak fault features in the signals, thereby obtaining enhanced fault feature representations; and a parallel feature extraction network module, which includes a vibration feature extraction branch and a multi-source time-series feature extraction branch; the vibration feature extraction branch takes the enhanced fault feature representations as input and uses a physics-prior-guided attention mechanism to extract the fault-related features from the vibration signals. The system comprises the following modules: a deep feature extraction branch that takes multi-source monitoring data as input and uses the Transformer model to extract its temporal dependency features; a feature decoupling and fusion module that introduces a feature decoupling loss function to decompose the features extracted by the two branches into fault-shared features and operating condition-related features, and fuses the fault-shared features of the two branches through an attention mechanism to obtain the decoupled fused features; a domain adaptation module that maps the decoupled fused features to the source and target domains respectively, and aligns the local feature distributions between the domains by minimizing the local maximum mean difference to obtain the domain-adaptive features; and a diagnostic output module that inputs the domain-adaptive features into the classifier and outputs the diagnostic results of the centrifugal fan's operating status.

[0042] The data acquisition and preprocessing module includes a sensor unit and a data preprocessing unit. The sensor unit consists of a piezoelectric accelerometer, a PT100 temperature sensor, a current sensor, a pressure sensor, and a photoelectric encoder, which are used to acquire vibration, temperature, current, wind pressure, and rotational speed data, respectively. All sensors are connected to the main control unit through a data acquisition card to achieve synchronous data acquisition. The data preprocessing unit integrates functions such as mean removal, trend removal, bandpass filtering, outlier removal, normalization, and time series alignment, respectively addressing the preprocessing requirements of vibration signals and multi-source monitoring data. It can automatically complete the preprocessing operations for various types of data and output preprocessed vibration signals and multi-source monitoring data that meet the requirements of subsequent feature extraction, ensuring data quality and time series consistency.

[0043] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0044] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent diagnosis of the operating status of a centrifugal fan based on deep learning, characterized in that, Includes the following steps: Vibration signals and multi-source monitoring data of centrifugal fans under multiple operating conditions are collected, and the collected data are preprocessed. Adaptive multi-scale time-frequency analysis is performed on the preprocessed vibration signal to enhance the weak fault features in the signal, resulting in an enhanced fault feature representation. A parallel feature extraction network is constructed, in which the vibration feature extraction branch takes the enhanced fault feature representation as input and uses a physical prior-guided attention mechanism to extract deep features related to the fault in the vibration signal. The multi-source temporal feature extraction branch takes multi-source monitoring data as input and uses the Transformer model to extract its temporal dependency features. A feature decoupling loss function is introduced to decompose the features extracted from the two branches into fault-shared features and operating condition-related features; By fusing the fault-sharing features of the two branches through an attention mechanism, the decoupled fused features are obtained. The decoupled fused features are mapped to the source and target domains respectively. By minimizing the local maximum mean difference of the local discriminative features corresponding to different fault types between the source and target domains, the local feature distributions between the domains are aligned to obtain the domain adaptive features. The domain-adaptive features are input into the classifier, and the diagnostic results of the centrifugal fan's operating status are output.

2. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 1, characterized in that, The specific steps to obtain the enhanced fault feature representation are as follows: An adaptive complete ensemble empirical mode decomposition is applied to the preprocessed vibration signal to obtain several intrinsic mode function components. Calculate the mutual information entropy between each intrinsic mode function component and the original vibration signal, and select intrinsic mode function components with mutual information entropy values ​​greater than a preset threshold as sensitive components; Morphological filtering is performed on the selected sensitive components to remove residual noise components; The filtered sensitive components are reconstructed to obtain the enhanced fault feature representation.

3. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 1, characterized in that, The specific steps for extracting fault-related deep features from vibration signals using a physics-prior-guided attention mechanism are as follows: The enhanced fault feature representation is input into a convolutional neural network to extract multi-layer convolutional feature maps. The physical prior weight matrix is ​​calculated based on the rotational frequency of the centrifugal fan and the characteristic frequency of bearing failure. The physical prior weight matrix is ​​multiplied element-wise with the feature map output from the last layer of the convolutional neural network to obtain the weighted feature map. Global average pooling is performed on the weighted feature map to obtain the deep features related to the fault in the vibration signal.

4. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 1, characterized in that, The specific steps for introducing a feature decoupling loss function to decompose the features extracted from the two branches into fault-shared features and operating condition-related features are as follows: The deep features extracted by the vibration feature extraction branch and the temporal dependent features extracted by the multi-source temporal feature extraction branch are jointly input into the feature decoupling network; The feature decoupling network consists of two parallel fully connected layers, which output fault-sharing features and operating condition-related features, respectively. Construct a feature decoupling loss function, which includes fault classification loss, operating condition classification loss and feature orthogonal constraint loss; The feature decoupling loss function is jointly optimized, and the feature decoupling network parameters are iteratively updated until the network converges, resulting in the decoupled fault-sharing features and operating condition-related features.

5. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 4, characterized in that, The specific structure of the feature decoupling loss function is as follows: The feature decoupling loss function is composed of a weighted average of the fault classification loss, the operating condition classification loss, and the feature orthogonal constraint loss. Both fault classification loss and operating condition classification loss use cross-entropy loss; The orthogonal constraint loss is obtained by calculating the covariance matrix of fault-sharing features and condition-related features.

6. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 4, characterized in that, The specific steps to obtain the decoupled fused features are as follows: The fault-sharing features obtained by decoupling the vibration feature extraction branch and the fault-sharing features obtained by decoupling the multi-source time-series feature extraction branch are concatenated to obtain the features to be fused. The features to be fused are input into the attention mechanism module, and the attention weights corresponding to the fault-shared features of the two branches are calculated through a fully connected layer and a Softmax function. The fault-shared features of the two branches are multiplied by their corresponding attention weights to obtain the weighted branch features. The weighted features of the two branches are added element by element to obtain the decoupled fused features.

7. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 1, characterized in that, The specific steps to obtain domain-adaptive features are as follows: The decoupled fused features are mapped to the high-dimensional regenerative kernel Hilbert spaces of the source and target domains, respectively. Local discriminative features belonging to the same fault category are extracted from the source domain and the target domain respectively, forming multiple local feature subsets; For each fault category, the maximum mean difference between the source domain features and the target domain features in the regeneration kernel Hilbert space is calculated to obtain the local neighborhood difference of that category. The total loss is obtained by weighting and summing the local domain differences of all fault categories to obtain the local maximum mean difference. Minimize the total loss of the local maximum mean difference, iteratively update the feature map network parameters until the network converges, and obtain the domain-adaptive features.

8. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 1, characterized in that, The specific steps for outputting the diagnostic results of the centrifugal fan's operating status are as follows: The domain-adaptive features are input into a multi-classifier consisting of fully connected layers and a Softmax function; Multi-classifiers calculate the probability distribution of adaptive features belonging to each operational state category; The operating state category with the highest probability value is selected as the diagnostic result for the current centrifugal fan.

9. The intelligent diagnostic method for centrifugal fan operating status based on deep learning according to claim 1, characterized in that, The specific steps for collecting vibration signals and multi-source monitoring data of centrifugal fans under multiple operating conditions and preprocessing the collected data are as follows: Vibration signals at the bearing housing of the centrifugal fan were collected, along with temperature data, motor current data, outlet air pressure data, and speed data at the bearing housing as multi-source monitoring data. The collected vibration signals are sequentially processed by mean removal, trend removal, and bandpass filtering to obtain the preprocessed vibration signals. Outlier removal was performed on the collected temperature data, motor current data, outlet air pressure data, and speed data. After removing outliers, each data point is normalized to map the numerical range of each data point to the same interval. The normalized data are aligned according to their timestamps to obtain preprocessed multi-source monitoring data.

10. A deep learning-based intelligent diagnostic system for the operating status of centrifugal fans, employing the deep learning-based intelligent diagnostic method for the operating status of centrifugal fans as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to collect vibration signals and multi-source monitoring data of the centrifugal fan under multiple operating conditions, and to preprocess the collected data. The signal enhancement module is used to perform adaptive multi-scale time-frequency analysis on the preprocessed vibration signal to enhance the weak fault features in the signal and obtain the enhanced fault feature representation. The parallel feature extraction network module includes a vibration feature extraction branch and a multi-source temporal feature extraction branch. The vibration feature extraction branch takes the enhanced fault feature representation as input and uses a physical prior-guided attention mechanism to extract deep features related to the fault in the vibration signal; the multi-source temporal feature extraction branch takes multi-source monitoring data as input and uses the Transformer model to extract its temporal dependency features. The feature decoupling and fusion module is used to introduce a feature decoupling loss function to decompose the features extracted from the two branches into fault-shared features and operating condition-related features, and to fuse the fault-shared features of the two branches through an attention mechanism to obtain the decoupled fused features. The domain adaptation module is used to map the decoupled fused features to the source and target domains respectively. By minimizing the local maximum mean difference, it aligns the local feature distributions between the domains to obtain the domain-adaptive features. The diagnostic output module is used to input the domain-adaptive features into the classifier and output the diagnostic results of the centrifugal fan's operating status.

Images