SE-RESNET34-based tail bearing fault feature fusion diagnosis system

Abstract

The invention relates to the technical field of intelligent fault diagnosis of mechanical equipment, and discloses a tail bearing fault feature fusion diagnosis system based on SE-RESNET34. According to the system, a sliding window method is adopted to segment an original data stream and perform multi-source feature knitting to form a diagnosis feature data set. After features are fused through an SE-RESNET34 network, two-channel evaluation is carried out, an initial risk value is calculated, and the key point is that time correlation information of historical cases is introduced, and cross-sample collaborative correction is carried out on the risk value. And constructing a fault feature interaction map based on the correction value, identifying key feature nodes and propagation paths through topology analysis, and outputting a visual report and a maintenance list. According to the method, the continuity and the anti-interference capability of risk assessment are improved through time sequence collaborative correction, the interpretability of a fault attribution and propagation mechanism is enhanced by means of atlas analysis, and support is provided for accurate maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology for mechanical equipment, specifically to a tail bearing fault feature fusion diagnosis system based on SE-RESNET34. Background Technology

[0002] Currently, fault diagnosis of critical components in rotating machinery, such as ship stern bearings, mainly relies on the acquisition and analysis of multi-source signals, including vibration and temperature. Mainstream intelligent diagnostic methods are typically based on deep learning models, such as convolutional neural networks or residual networks, to extract features and classify patterns from preprocessed single time-series data samples. These methods treat data within each time window as independent samples, directly outputting fault categories or simple health scores through the model, achieving an end-to-end mapping from raw data to diagnostic conclusions.

[0003] These conventional methods have limitations. The risk assessment process is static and isolated, making judgments based solely on current data snippets, completely severing the continuity of fault evolution over time and the reference value of historical cases. In actual operation, bearing degradation is a gradual process; the risk level of the current state is closely related to the previous state. Isolated assessments are prone to misjudgments due to data noise or fluctuations in operating conditions, failing to reflect the potential development trend of the fault. Existing methods stop at providing classification results or abstract feature vectors, failing to delve into the intrinsic relationships and interaction mechanisms between different fault characteristics. Diagnostic results lack interpretability; maintenance personnel cannot know which key characteristics dominate the current state, or through what paths the fault might affect other parameters, limiting the accuracy of early warnings and the predictability of maintenance decisions. Summary of the Invention

[0004] The purpose of this invention is to provide a tail bearing fault feature fusion diagnostic system based on SE-RESNET34 to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a tail bearing fault feature fusion diagnostic system based on SE-RESNET34, the system comprising: The data preprocessing module acquires the raw operating data stream of the tail bearing, and uses a sliding window method based on a preset time window to segment the raw operating data stream, generating a series of data sample units with time continuity. The feature extraction module performs multi-source data weaving processing on each data sample unit to generate a diagnostic feature dataset containing vibration feature tensors, temperature feature vectors, and acoustic feature vectors. The deep feature fusion module inputs the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion and outputs a primary fault feature vector. The risk assessment module, based on the primary fault feature vector and the preset fault mode knowledge base, performs fault potential assessment and operational health assessment, generates dual-channel assessment results, calculates the preliminary fault risk quantification value through a dynamic weighting strategy based on the dual-channel assessment results, introduces the time correlation information of historical fault cases, and performs cross-sample unit collaborative correction on the preliminary fault risk quantification value to generate the corrected fault risk quantification value. The graph diagnosis module constructs a fault feature interaction graph based on the corrected fault risk quantification value, performs topology analysis on the fault feature interaction graph, identifies key fault feature nodes and their propagation paths, and generates a visual diagnosis report and maintenance priority list.

[0006] Preferably, the specific steps for performing multi-source data weaving processing on each of the data sample units include: The raw operational data stream includes vibration signals, temperature readings, and acoustic waveforms sorted by timestamp; The vibration signal within the data sample unit is resampled and time-aligned to ensure strict correspondence with the sampling time of the temperature reading and acoustic waveform; The frequency domain energy distribution features and time domain statistical features are extracted from the aligned vibration signal and combined to form the vibration feature tensor. The temperature change rate and temperature gradient are calculated from the aligned temperature readings to form the temperature feature vector; The acoustic feature vector is constructed by extracting sound pressure level spectral features and anomalous resonant frequency components from the aligned acoustic waveform; The vibration feature tensor, the temperature feature vector, and the acoustic feature vector are concatenated and normalized along the time dimension to finally output the diagnostic feature dataset.

[0007] Preferably, the process of inputting the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion specifically involves: The vibration feature tensor in the diagnostic feature dataset is input into the backbone convolutional layer of the SE-RESNET34 network to extract spatial hierarchical features. The temperature feature vector and acoustic feature vector in the diagnostic feature dataset are respectively input into independent feature embedding layers and mapped to a high-dimensional feature space; The spatially hierarchical features and the mapped high-dimensional features are concatenated along the channel dimension. The spliced ​​fused features are input into the compression excitation module of the SE-RESNET34 network, and the feature channel weights are adaptively calibrated through the channel attention mechanism. The fused features after weight calibration are input into the residual layer and fully connected layer of the SE-RESNET34 network, and finally the primary fault feature vector is compressed and output.

[0008] Preferably, the execution steps for the fault potential assessment and operational health assessment are as follows: The similarity matching calculation is performed between the primary fault feature vector and the predefined fault feature template in the fault mode knowledge base; Based on the similarity matching calculation results, the fault mode category is determined and a fault confidence score is generated as the fault potential assessment result; Simultaneously, sub-vectors representing operational stability, load fluctuation, and lubrication effectiveness are extracted from the primary fault feature vector; The operational stability subvector, load volatility subvector, and lubrication effectiveness subvector are input into a preset health status scoring model to calculate the operational health score, which is used as the operational health assessment result. The failure potential assessment results and the operational health assessment results together constitute the dual-channel assessment results.

[0009] Preferably, the process of calculating the preliminary fault risk quantification value through a dynamic weighting strategy includes: Extract the fault confidence score and the operational health score from the dual-channel evaluation results; Based on the current operating stage of the tail bearing, query the preset weight allocation table to obtain the fault confidence weight coefficient and the operating health weight coefficient; The fault confidence score is weighted using the fault confidence weighting coefficient, and the operational health score is weighted using the operational health weighting coefficient. The initial risk value is obtained by summing the weighted failure confidence score and the weighted operational health score. The initial risk value is input into a risk value normalizer based on a sigmoid function, and the output value is mapped to a preset quantization range to obtain the preliminary fault risk quantization value.

[0010] Preferably, the process of collaboratively correcting the preliminary fault risk quantification value across sample units specifically includes: Select the current target data sample unit and locate its adjacent preceding and following data sample units in the time series. Preliminary fault risk quantification values ​​are obtained for the preceding data sample unit, the target data sample unit, and the subsequent data sample unit, respectively. Search the historical failure case database to find past cases with similar characteristics to the target data sample unit, and extract the risk evolution sequence of the past cases before the failure occurred; Analyze the local trends of the risk evolution sequence and calculate the trend influence factor; Based on the preliminary fault risk quantification value of the preceding data sample unit, the preliminary fault risk quantification value of the following data sample unit, and the trend influence factor, a three-input collaborative correction function is constructed. The preliminary fault risk quantification value of the target data sample unit is input into the collaborative correction function, and the corrected fault risk quantification value is output after calculation.

[0011] Preferably, the step of constructing a fault feature interaction map based on the corrected fault risk quantification value is as follows: The corrected fault risk quantification value of each data sample unit is used as the basic node attribute of the graph. Calculate the feature similarity, temporal proximity, and risk value correlation between any two data sample units, and use the calculation results as the edge weights connecting the corresponding two nodes in the graph; Based on all nodes and the connecting edges with edge weights, generate an initial global feature association graph; From the global feature association graph, the connecting edges whose edge weights exceed a set threshold are selected, and the nodes they connect are retained to form a densely connected subgraph structure; This densely connected subgraph structure is defined as the fault feature interaction graph.

[0012] Preferably, the step of performing topological analysis on the fault feature interaction map further includes: The community detection algorithm is run on the fault feature interaction graph to identify node communities with tight internal connections and sparse external connections, and each node community is marked as a local fault feature cluster. Calculate the intra-cluster average risk value and risk value variance for each of the local fault feature clusters; According to the preset key node identification rules, nodes with risk values ​​significantly higher than the average risk value within the cluster and a large number of connecting edges are identified in each local fault feature cluster and marked as the key fault feature nodes. The key fault feature node is traced to all paths that reach other nodes through strong connection edges in the fault feature interaction graph, and the paths are organized into the propagation path.

[0013] Preferably, the process of generating a visual diagnostic report includes: The timestamps, original signal segments, and corrected fault risk quantification values ​​of the data sample units corresponding to the key fault feature nodes are integrated. Integrate the data sample unit sequences involved in the propagation path and label the risk value transmission intensity along the path; The integrated information is overlaid and rendered with the topology of the fault feature interaction map to generate a comprehensive diagnostic interface that includes a time axis, risk heat map, feature map and path arrows. In the comprehensive diagnostic interface, all relevant data sample units are sorted from high to low according to the corrected fault risk quantification value; The sorting results are compared with preset maintenance action thresholds to generate the maintenance priority list containing specific maintenance recommendations.

[0014] Preferably, the logic for generating the maintenance priority list is as follows: The maintenance action threshold is set to multiple levels, with each level corresponding to a different risk value range and maintenance response time limit; The corrected fault risk quantification value of the sorted data sample unit is matched with the risk value range to determine the maintenance level of each data sample unit. Based on the timestamp of the data sample unit and the maintenance response time limit, calculate the time window for recommending the execution of maintenance operations; Based on the fault mode descriptions associated with the key fault feature nodes in the fault mode knowledge base, targeted maintenance operation instructions are generated. The final output maintenance priority list includes the data sample unit identifier, its maintenance level, the recommended maintenance time window, and specific maintenance operation instructions.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By incorporating temporal correlation information from historical failure cases, the initial risk quantification value calculated based on a single sample unit is collaboratively corrected across sample units. This mechanism breaks away from the traditional isolated assessment paradigm, enabling risk assessment to integrate contextual information from the time series and historical experience. It improves the robustness and predictability of the risk quantification value, reduces false alarms and missed alarms caused by anomalies in single-point data, and makes the assessment results better reflect the dynamic evolution trend and potential cumulative effects of failures over time, providing a more reliable temporal continuity basis for early warning.

[0016] Based on the corrected risk values, a fault feature interaction graph is constructed, and topological analysis is performed to identify key feature nodes and propagation paths. This method advances fault diagnosis from the traditional "feature extraction-classification" model to the level of "feature relationship mining and networked reasoning." It achieves visualization and deep interpretability of the diagnostic process, not only locating core fault features but also clearly revealing the influence relationships between different feature parameters and the potential propagation chains of the fault. This enables diagnostic reports to directly point out the root causes of the fault and their potential chain reactions, thus providing intuitive graph-based decision support for developing accurate and efficient preventative maintenance strategies and prioritization. Attached Figure Description

[0017] Figure 1 This is a timing diagram of the tail bearing fault feature fusion diagnostic system based on SE-RESNET34 described in this invention; Figure 2 This is a flowchart of the deep feature fusion process; Figure 3 A flowchart for calculating the initial fault risk quantification value for a dynamic weighted strategy; Figure 4 A bar chart showing the quantitative values ​​of node failure risk; Figure 5 A bar chart comparing the risk values ​​of data sample units. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides a tail bearing fault feature fusion diagnostic system based on SE-RESNET34. The system includes: after system startup, a data preprocessing module acquires the raw operating data stream of the tail bearing and segments the raw operating data stream using a sliding window method based on a preset time window, thereby generating a series of data sample units with temporal continuity. A feature extraction module performs multi-source data weaving processing on each data sample unit to generate a diagnostic feature dataset containing vibration feature tensors, temperature feature vectors, and acoustic feature vectors. A deep feature fusion module inputs the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion and outputs a primary fault feature vector. A risk assessment module performs fault potential assessment and operational health assessment based on the primary fault feature vector and a preset fault mode knowledge base to generate a dual-channel assessment result. Based on the dual-channel assessment result, a preliminary fault risk quantification value is calculated using a dynamic weighting strategy. The temporal correlation information of historical fault cases is introduced to collaboratively correct the preliminary fault risk quantification value across sample units, generating a corrected fault risk quantification value. The graph diagnosis module constructs a fault feature interaction graph based on the corrected fault risk quantification value, performs topology analysis on the fault feature interaction graph to identify key fault feature nodes and their propagation paths, and generates a visual diagnosis report and maintenance priority list.

[0020] Example 1: See Figure 2The original running data stream contains vibration signals, temperature readings, and acoustic waveforms sorted by timestamp. Vibration signals within data sample units are resampled and time-aligned to ensure strict correspondence with the sampling times of temperature readings and acoustic waveforms. Frequency domain energy distribution features and time domain statistical features are extracted from the aligned vibration signals and combined to form a vibration feature tensor. Temperature change rate and temperature gradient are calculated from the aligned temperature readings to form a temperature feature vector. Sound pressure level spectrum features and anomalous resonant frequency components are extracted from the aligned acoustic waveforms to construct an acoustic feature vector. The vibration feature tensor, temperature feature vector, and acoustic feature vector are concatenated and normalized along the time dimension to finally output a diagnostic feature dataset. The process of inputting the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion involves inputting the vibration feature tensor from the diagnostic feature dataset into the backbone convolutional layer of the SE-RESNET34 network to extract spatially hierarchical features. The temperature feature vector and acoustic feature vector from the diagnostic feature dataset are input into independent feature embedding layers, mapping them to a high-dimensional feature space. The spatially hierarchical features and the mapped high-dimensional features are concatenated along the channel dimension. The concatenated fused features are then input into the compression activation module of the SE-RESNET34 network, where the feature channel weights are adaptively calibrated using a channel attention mechanism. The weight-calibrated fused features are then input into the residual layer and fully connected layer of the SE-RESNET34 network, ultimately compressing and outputting a primary fault feature vector.

[0021] In specific implementation, the process of performing multi-source data weaving on each data sample unit and inputting the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion is described through a specific example scenario. The original running data stream contains vibration signals, temperature readings, and acoustic waveforms sorted by timestamps. For a data sample unit generated within a preset 10-second time window, the vibration signals within the data sample unit are first resampled and time-aligned to ensure that the time of each sampling point of the vibration signal strictly corresponds to the recording time of the temperature reading and the sampling time of the acoustic waveform. From the aligned vibration signals, the energy values ​​of 10 frequency bands divided within the 0-1000Hz range are extracted as frequency domain energy distribution features using Fast Fourier Transform. Simultaneously, six time-domain statistical features—root mean square, peak factor, kurtosis, skewness, waveform factor, and impulse factor—are calculated. These features are combined into a 10×6 vibration feature tensor. From the aligned temperature readings, the rate of temperature change is calculated based on the temperature values ​​of three consecutive sampling points, and the axial and radial temperature gradients are calculated based on the sensor's spatial location, forming a temperature feature vector containing three elements. From the aligned acoustic waveforms, the sound pressure level spectrum is calculated using an A-weighted network, and sound pressure level values ​​for eight standard octave bands are extracted. Simultaneously, three anomalous resonant frequency components significantly higher than the background noise spectrum are identified through peak detection. These components together construct an acoustic feature vector containing 11 elements. The vibration feature tensor, temperature feature vector, and acoustic feature vector are concatenated along the time dimension to form a multidimensional array. This multidimensional array is then normalized using a minimax normalization method, ultimately outputting a diagnostic feature dataset.

[0022] In some embodiments, the process of inputting the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion is specifically implemented as follows: the normalized vibration feature tensor is input into the backbone convolutional layer of the SE-RESNET34 network. The backbone convolutional layer contains multiple convolutional layers and pooling layers with a kernel size of 3×3, used to extract spatial hierarchical features layer by layer. The temperature feature vector and acoustic feature vector are respectively input into a feature embedding layer composed of two independent fully connected layers. The feature embedding layer maps the temperature feature vector and acoustic feature vector to a high-dimensional feature space, and the mapped dimension is the same as the channel dimension of the spatial hierarchical features. In the channel dimension, the spatial hierarchical features are concatenated with the mapped high-dimensional temperature features and high-dimensional acoustic features to generate fused features. The concatenated fused features are input into the compression activation module of the SE-RESNET34 network. The compression activation module first compresses the two-dimensional features of each channel into a scalar through global average pooling, and then learns the dependencies between channels and generates the weights of each channel through the bottleneck structure composed of two fully connected layers. The weight learning process is described by the following relationship: Where: z represents the vector after global average pooling, and W1 and W2 represent the weight matrices of the two fully connected layers. Represents the ReLU activation function. s represents the Sigmoid activation function, and s represents the final generated channel weight vector. The channel attention mechanism uses the weight vector s to adaptively calibrate the weights of each channel of the fused feature. The fused feature after weight calibration is input into the residual layer of the SE-RESNET34 network. The residual layer contains the superposition of identity mapping and convolutional blocks. Finally, it is connected to a global average pooling layer and a fully connected layer to compress the feature and output it as a fixed-length primary fault feature vector.

[0023] Optionally, the vibration signal resampling process employs a linear interpolation algorithm to ensure that the vibration signal sampling rate and the acoustic waveform sampling rate remain consistent after time alignment. It is understood that the calculation of frequency domain energy distribution characteristics can utilize either Fast Fourier Transform or Short-Time Fourier Transform to obtain joint time-frequency distribution information. Optionally, the temperature gradient calculation relies on readings from multiple temperature sensors installed at different axial and radial positions on the tail bearing housing. It is understood that the identification threshold for abnormal resonant frequency components is dynamically set based on the statistical standard deviation of the background noise spectrum amplitude of the acoustic waveform under fault-free operating conditions. In some embodiments, the high-dimensional feature dimension output by the feature embedding layer is set to 128 dimensions, consistent with the number of channels in the spatially hierarchical features output by the backbone convolutional layer, to facilitate subsequent stitching operations. In a specific implementation, the dimensionality reduction ratio of the first fully connected layer in the compressed excitation module is set to 16 to reduce the number of model parameters and introduce nonlinearity.

[0024] Example 2: The primary fault feature vector is matched with predefined fault feature templates in the fault mode knowledge base to calculate similarity. Based on the similarity matching result, the fault mode category is determined and a fault confidence score is generated as the fault potential assessment result. Simultaneously, sub-vectors representing operational stability, load variability, and lubrication effectiveness are parsed from the primary fault feature vector. These sub-vectors are input into a pre-defined health status scoring model to calculate the operational health score, which is used as the operational health assessment result. The fault potential assessment result and the operational health assessment result together constitute a dual-channel assessment result.

[0025] In practical implementation, the execution steps of fault potential assessment and operational health assessment are illustrated through a specific example. The primary fault feature vector output from the deep feature fusion module is a 256-dimensional vector. The fault mode knowledge base predefines M fault modes, each associated with a 256-dimensional fault feature template vector learned from historical fault data. Similarity matching is performed between the primary fault feature vector corresponding to the current data sample unit and each fault feature template vector in the fault mode knowledge base. The matching calculation uses the cosine similarity method, and its calculation relationship is described as follows:

[0026] Where: V current T represents the primary fault feature vector of the current data sample unit. j C represents the j-th fault feature template vector in the fault mode knowledge base. j This represents the similarity score between the current primary fault feature vector and the j-th template. The symbol · denotes the vector dot product, and ||| denotes the L2 norm of the vector. After calculation, the highest score among the M similarity scores is selected as the final fault confidence score, and the fault mode category corresponding to this score is determined as the most likely fault mode category. This fault confidence score and the determined fault mode category together constitute the fault potential assessment result.

[0027] In some embodiments, the process of parsing sub-vectors representing operational stability, load variability, and lubrication effectiveness from the primary fault feature vector has a specific implementation. The 256 dimensions of the primary fault feature vector are pre-divided into three index groups based on their physical meaning relevance during the preceding network training process. The sub-vectors representing operational stability correspond to index group A, containing feature values ​​from dimension 1 to 80, which are primarily related to the spectral symmetry and harmonic components of the vibration signal. The sub-vectors representing load variability correspond to index group B, containing feature values ​​from dimension 81 to 160, which are related to the signal's energy transients and envelope characteristics. The sub-vectors representing lubrication effectiveness correspond to index group C, containing feature values ​​from dimension 161 to 256, which are related to triboacoustic characteristics and temperature field stability. The operational stability subvector, load volatility subvector, and lubrication effectiveness subvector are input in parallel into a pre-defined health status scoring model. This model is a fully connected neural network with two hidden layers. The first hidden layer maps the concatenated input of the three subvectors to 64 dimensions, and the second hidden layer maps it to 16 dimensions. The final output layer outputs an operational health score ranging from 0 to 100, which serves as the operational health assessment result. The fault potential assessment result, together with the operational health assessment result, constitutes a dual-channel assessment result. The fault potential assessment result includes a fault mode category identifier and a fault confidence score between 0 and 1, while the operational health assessment result is an operational health score between 0 and 100.

[0028] Optionally, the fault feature template vectors in the fault mode knowledge base are generated by clustering a large number of historical fault case primary fault feature vectors using a clustering algorithm, with each cluster center forming a template vector. It can be understood that similarity matching calculations can use not only cosine similarity but also the reciprocal of Euclidean distance or other similarity metrics. Optionally, the training data for the health status scoring model comes from labeled samples of historical normal operating conditions and known performance degradation states, with mean squared error used as the loss function. It can be understood that the dimensional division criteria for the three sub-vectors—operational stability, load variability, and lubrication effectiveness—can be determined by analyzing the regression importance of each dimension of the primary fault feature vector to preset physical indicator labels.

[0029] Example 3: See Figure 3The system extracts fault confidence scores and operational health scores from the dual-channel evaluation results. Based on the current operational stage of the tail bearing, it queries a pre-defined weighting table to obtain fault confidence weighting coefficients and operational health weighting coefficients. The fault confidence score is weighted using the fault confidence weighting coefficients, and the operational health score is weighted using the operational health weighting coefficients. The weighted fault confidence score and the weighted operational health score are summed to obtain an initial risk value. This initial risk value is input into a risk value normalizer based on a sigmoid function, mapping the output value to a pre-defined quantization range to obtain a preliminary fault risk quantification value. The collaborative correction process for the preliminary fault risk quantification value across sample units specifically includes selecting the current target data sample unit and locating its adjacent preceding and following data sample units in the time series. The preliminary fault risk quantification values ​​corresponding to the preceding, target, and following data sample units are obtained respectively. A historical fault case database is searched to find past cases with similar characteristics to the target data sample unit, and the risk evolution sequence of these past cases before the fault occurred is extracted. The local trends of the risk evolution sequence are analyzed, and the trend influence factor is calculated. Based on the preliminary fault risk quantification values ​​of preceding data sample units, the preliminary fault risk quantification values ​​of subsequent data sample units, and the trend influence factor, a three-input collaborative correction function is constructed. The preliminary fault risk quantification value of the target data sample unit is input into the collaborative correction function, and the corrected fault risk quantification value is output after calculation.

[0030] In practical implementation, the process of calculating the initial fault risk quantification value through a dynamic weighting strategy and then collaboratively correcting the initial fault risk quantification value across sample units is described through a specific example scenario. The fault confidence score extracted from the dual-channel evaluation results is 0.72, and the extracted operational health score is 65. The current operating stage of the tail bearing is determined to be the "high-load cruise stage". According to the operating stage, a preset weight allocation table is queried. The weight allocation table defines the values ​​of the fault confidence weight coefficient and the operational health weight coefficient under different operating stages. Under the "high-load cruise stage", the query result of the fault confidence weight coefficient is 0.6, and the query result of the operational health weight coefficient is 0.4. The fault confidence score of 0.72 is weighted using a fault confidence weighting coefficient of 0.6, resulting in a weighted fault confidence score of 0.432. The operational health score of 65 is weighted using an operational health weighting coefficient of 0.4. The operational health score needs to be normalized to the 0-1 range (65 / 100 = 0.65) before weighting, resulting in a weighted operational health score of 0.26. The weighted fault confidence score of 0.432 and the weighted operational health score of 0.26 are summed to obtain an initial risk value of 0.692. This initial risk value is input into a risk value normalizer based on a sigmoid function. The risk value normalizer is implemented as follows:

[0031] Where: x represents the initial risk value, x0 represents the center point parameter of the sigmoid function and is set to 0.5, k represents the kurtosis parameter of the curve and is set to 10, e represents the natural constant, and R... pre The output represents the initial fault risk quantization value mapped to the 0 to 1 quantization range. Substituting the initial risk value of 0.692 into the calculation, the initial fault risk quantization value is approximately 0.912.

[0032] In some embodiments, the process of collaboratively correcting the preliminary fault risk quantification value across sample units is specifically implemented. The current target data sample unit time is selected as T, and its adjacent preceding data sample unit time is T-1 and its subsequent data sample unit time is T+1. The preliminary fault risk quantification values ​​for the data sample unit at time T-1 are obtained as 0.85, for the data sample unit at time T as 0.912, and for the data sample unit at time T+1 as 0.88, respectively. A historical fault case database is searched to find past cases with similar characteristics to the data sample unit at time T. Feature similarity is determined by the Euclidean distance of the primary fault feature vectors being less than a threshold. The risk evolution sequence of past cases within the time window before the fault occurs is extracted. The risk evolution sequence consists of a series of preliminary fault risk quantification values ​​ordered by time. The local trend of the risk evolution sequence is analyzed, and the average slope of risk value change near the same time position is calculated. This average is used as the trend influence factor, assuming the calculated trend influence factor is 0.05. Based on the preliminary fault risk quantification values ​​of the preceding data sample units, the preliminary fault risk quantification values ​​of the subsequent data sample units, and the trend influence factor, a three-input collaborative correction function is constructed. The collaborative correction function is expressed as follows: in: This represents the initial quantification value of the failure risk for the target data sample unit at time T. This represents the initial fault risk quantification value for the preceding data sample unit at time T-1. The initial fault risk quantification value represents the sample unit time T+1 of the subsequent data. α represents the influence coefficient of the nearest sample and is set to 0.1, β represents the influence coefficient of the historical trend and is set to 0.3, γ represents the trend influence factor, and R0 adjThis represents the corrected fault risk quantification value of the output. Substituting the initial fault risk quantification value of 0.912 for the time T data sample unit, the initial fault risk quantification value of 0.85 for the time T-1 data sample unit, the initial fault risk quantification value of 0.88 for the time T+1 data sample unit, and the trend influence factor of 0.05 into the collaborative correction function, the corrected fault risk quantification value is approximately 0.927.

[0033] Optionally, the division criteria for different operating stages in the weight allocation table include the tail bearing speed setting value, load current value, and continuous operating time interval. It is understood that the parameters x0 and k of the risk value normalizer can be adjusted based on prior knowledge of the actual risk value distribution to optimize the distinguishability of risk levels. Optionally, the feature similarity retrieval of cases in the historical fault case library uses a method based on locality-sensitive hashing to improve retrieval efficiency. It is understood that the neighboring sample influence coefficient α and historical trend influence coefficient β in the collaborative correction function can be determined by optimizing on historical data with the goal of minimizing the prediction error between the corrected risk value and the final actual fault state. In some embodiments, the calculation of the trend influence factor considers not only the slope of change but also the variance of the risk sequence within the recent window, and combines the two linearly. In a specific implementation, if the target data sample unit is at the beginning or end of the data stream and lacks preceding or subsequent data sample units, the corresponding term in the collaborative correction function is filled with the preliminary fault risk quantification value of the target data sample unit itself.

[0034] Example 4: The corrected fault risk quantification value of each data sample unit is used as the basic attribute of the nodes in the graph. The feature similarity, temporal proximity, and risk value correlation between any two data sample units are calculated, and the results are used as the edge weights connecting the corresponding two nodes in the graph. An initial global feature association graph is generated based on all nodes and the connecting edges with edge weights. From the global feature association graph, connecting edges with edge weights exceeding a set threshold are selected, and the nodes they connect are retained, forming a densely connected subgraph structure. This densely connected subgraph structure is defined as the fault feature interaction graph. The step of performing topological analysis on the fault feature interaction graph further includes running a community detection algorithm on the fault feature interaction graph to identify node clusters with tightly connected internal structures and sparse external structures, and marking each node cluster as a local fault feature cluster. The intra-cluster average risk value and risk value variance of each local fault feature cluster are calculated. According to the preset key node identification rules, nodes with risk values ​​significantly higher than the intra-cluster average risk value and a large number of connecting edges are identified in each local fault feature cluster and marked as key fault feature nodes. Track all paths from key fault feature nodes to other nodes in the fault feature interaction graph through strong connection edges, and organize these paths into propagation paths.

[0035] In practical implementation, the process of constructing a fault feature interaction graph based on the corrected fault risk quantification value and performing topological analysis on the fault feature interaction graph is described through a specific example containing five data sample units. The five data sample units are labeled as nodes A, B, C, D, and E. The corrected fault risk quantification value for node A is 0.92, for node B it is 0.87, for node C it is 0.45, for node D it is 0.89, and for node E it is 0.50. The feature similarity, temporal proximity, and risk value correlation between any two data sample units are calculated. Feature similarity is calculated using cosine similarity based on the primary fault feature vector; temporal proximity is calculated based on the reciprocal of the timestamp interval between the data sample units; and risk value correlation is calculated using the absolute value of the Pearson correlation coefficient between the corrected fault risk quantification values ​​of the two nodes. The feature similarity, temporal proximity, and risk value correlation are weighted and summed, and the calculated relationship is described as E. ij = w1·Sim ij +w2·Prox ij + w3·Corr ij E ij Sim represents the combined edge weight between node i and node j. ij Prox represents the feature similarity between node i and node j. ij Corr represents the temporal proximity between node i and node j. ij The correlation between the risk values ​​of node i and node j is represented by w1, w2, and w3, which are weighting coefficients set to w1=0.4, w2=0.3, and w3=0.3, respectively. The pairwise calculation results between the five nodes are shown in Table 1. Table 1: Inter-node correlation metrics and combined edge weights Node pairs Feature Similarity (Sim) Time proximity (Prox) Risk Value Correlation (Corr) Overall edge weight (E) AB 0.85 0.90 0.95 0.895 AC 0.30 0.10 0.15 0.195 AD 0.82 0.60 0.88 0.778 AE 0.35 0.20 0.40 0.320 BC 0.25 0.15 0.10 0.175 BD 0.88 0.85 0.90 0.877 BE 0.33 0.25 0.35 0.312 CD 0.28 0.18 0.20 0.226 CE 0.80 0.95 0.10 0.625 DE 0.40 0.30 0.45 0.385 Based on all nodes and the connecting edges calculated from the table above, an initial global feature association graph is generated, containing five nodes and all possible connecting edges. The weight of each connecting edge is the comprehensive edge weight E. Connecting edges with a comprehensive edge weight E exceeding a set threshold of 0.7 are selected from the initial global feature association graph, and the nodes connected to these edges are retained. In this example, connecting edges with weights exceeding 0.7 include AB, AD, and BD, connecting nodes A, B, and D. Nodes A, B, and D, along with the connecting edges between them with weights exceeding 0.7, form a densely connected subgraph structure. This densely connected subgraph structure is defined as the fault feature interaction graph required for the current analysis.

[0036] In some embodiments, the step of performing topological analysis on the fault feature interaction graph is further implemented by running a community detection algorithm on the fault feature interaction graph composed of nodes A, B, and D. In this example, the graph is directly identified as an independent node community, and this node community is marked as a local fault feature cluster Alpha. The intra-cluster average risk value of the local fault feature cluster Alpha is calculated, i.e., (0.92+0.87+0.89) / 3≈0.893, and the variance of the risk value of the local fault feature cluster Alpha is calculated. The preset key node identification rule is defined as follows: the corrected fault risk quantification value of the node is greater than 1.1 times the intra-cluster average risk value, and the degree (number of connecting edges) of the node is greater than or equal to 2. In the local fault feature cluster Alpha, the threshold condition that the risk value of node A, 0.92, is greater than 0.893*1.1≈0.982 is not met, the risk value of node B, 0.87, and the risk value of node D, 0.89, are not met, but the degree of nodes A and B is 2, and the degree of node D is 2. Since this cluster is relatively small, the adjustment rule is to select the node with the highest risk value and a degree of not less than 2 within the cluster. Node A is identified as a critical fault feature node. All paths from critical fault feature node A to other nodes in the fault feature interaction graph through strong connections are traced. These paths include AB and AD, and these two edges are organized into a propagation path originating from node A.

[0037] Optionally, the community detection algorithm can employ the Louvain algorithm or the label propagation algorithm to automatically identify the node community structure. It is understood that in complex graphs with more nodes, the multiple threshold and degree threshold in the key node identification rules can be adjusted based on historical diagnostic experience. Optionally, the calculation of temporal proximity can introduce an exponential decay function, so that nodes with closer time intervals receive higher proximity values. It is understood that, in addition to cosine similarity, the calculation of feature similarity can also employ a Gaussian kernel-based similarity measurement method in specific implementations. In some embodiments, the weight coefficients w1, w2, and w3 in the comprehensive edge weight calculation can be optimized using sample data from known fault propagation paths through supervised learning. In specific implementations, if the filtered subgraph structure is not connected, i.e., there are multiple separate dense subgraphs, each connected subgraph is treated as an independent fault feature interaction graph for subsequent topological analysis.

[0038] See Figure 4This is a bar chart showing the quantitative values ​​of node failure risk. The core content is that node A has the highest risk value (close to 0.9), followed by node D (around 0.9), and node B is slightly lower (around 0.85). This type of chart is a data analysis chart. Nodes C (around 0.45) and E (around 0.5) have significantly lower risk values ​​than the other three. This type of chart is typically used in fault diagnosis and risk assessment scenarios (such as industrial equipment and system operation status analysis) to differentiate the risk levels of different units through quantitative values, helping to identify high-risk objects.

[0039] Example 5: Integrate the timestamps, original signal segments, and corrected fault risk quantification values ​​of data sample units corresponding to key fault feature nodes. Integrate the data sample unit sequences involved in the propagation path and label the risk value transmission intensity along the path. Overlay and render the integrated information with the topology diagram of the fault feature interaction map to generate a comprehensive diagnostic interface including a time axis, risk heatmap, feature map, and path arrows. In the comprehensive diagnostic interface, sort all involved data sample units from high to low according to the corrected fault risk quantification values. Compare the sorting results with preset maintenance action thresholds to generate a maintenance priority list containing specific maintenance recommendations. The specific logic for generating the maintenance priority list is as follows: set the maintenance action thresholds to multiple levels, each level corresponding to a different risk value range and maintenance response time limit. Match the corrected fault risk quantification values ​​of the sorted data sample units with the risk value ranges to determine the maintenance level to which each data sample unit belongs. Calculate the time window for recommended maintenance operations based on the timestamps and maintenance response time limits of the data sample units. Combine the fault mode descriptions associated with key fault feature nodes in the fault mode knowledge base to generate targeted maintenance operation instructions. The final maintenance priority list includes the data sample unit identifier, its maintenance level, the recommended maintenance time window, and specific maintenance operation instructions.

[0040] In practical implementation, the process of generating a visual diagnostic report and maintenance priority list is described through a specific example scenario. This scenario involves five data sample units to be analyzed: nodes A, B, C, D, and E. Integrating the data sample unit information corresponding to key fault feature nodes, node A is identified as a key fault feature node, with a timestamp of "2023-10-26 14:30:00". The original signal segment includes a 10-second vibration waveform and a synchronized acoustic waveform. The corrected fault risk quantification value for node A is 0.92. Node B is also identified as a key fault feature node, with a timestamp of "2023-10-26 14:30:10". The corrected fault risk quantification value for node B is 0.87. Integrating the data sample unit sequences involved in the propagation path, one propagation path is formed from node A to node B, with a risk value transmission intensity labeled as 0.85; another propagation path is formed from node A to node C, with a risk value transmission intensity labeled as 0.60. The integrated timestamp information, original signal segments, corrected fault risk quantification values, propagation path sequences, and transmission intensity information are overlaid and rendered with a topology diagram of the fault feature interaction map. The topology diagram displays nodes A, B, C, D, and E and their connections. A comprehensive diagnostic interface is generated. The left side of the comprehensive diagnostic interface displays a timeline from 14:29:50 to 14:30:40. Below the timeline, a risk heatmap composed of the risk values ​​of the five nodes is displayed. The feature map is displayed in the center of the interface. The size of the nodes in the feature map is proportional to their corrected fault risk quantification values, and the thickness of the lines connecting the nodes is proportional to the risk value transmission intensity. Path arrows indicate the direction of risk propagation. In the comprehensive diagnostic interface, all five data sample units are sorted from highest to lowest according to their corrected fault risk quantification values. The sorting results are: Node A (0.92), Node D (0.89), Node B (0.87), Node E (0.50), and Node C (0.45). The sorting results are compared with preset maintenance action thresholds, which are set at three levels, to generate a maintenance priority list containing specific maintenance recommendations.

[0041] In some embodiments, the logic for generating the maintenance priority list is specifically implemented as follows: the preset maintenance action threshold is set to three levels: level L1 is named "urgent", corresponding to a risk value range greater than 0.8, and the maintenance response time limit is 24 hours; level L2 is named "warning", corresponding to a risk value range of 0.6 to 0.8, and the maintenance response time limit is 72 hours; level L3 is named "normal", corresponding to a risk value range less than 0.6, and the maintenance response time limit is 168 hours. The corrected fault risk quantification values ​​of the sorted data sample units are matched with the risk value ranges. The corrected fault risk quantification value of node A is 0.92, which is greater than 0.8, so the maintenance level of node A is determined to be L1 Emergency; the corrected fault risk quantification value of node D is 0.89, which is greater than 0.8, so the maintenance level of node D is determined to be L1 Emergency; the corrected fault risk quantification value of node B is 0.87, which is greater than 0.8, so the maintenance level of node B is determined to be L1 Emergency; the corrected fault risk quantification value of node E is 0.50, which is less than 0.6, so the maintenance level of node E is determined to be L3 Normal; the corrected fault risk quantification value of node C is 0.45, which is less than 0.6, so the maintenance level of node C is determined to be L3 Normal. Based on the timestamps and maintenance response time limits of the data sample units, the recommended maintenance operation time window is calculated. For node A, the timestamp is "2023-10-26 14:30:00", the maintenance response time limit is 24 hours, and the recommended maintenance time window is calculated from "2023-10-26 14:30:00" to "2023-10-27 14:30:00". The time window calculation method for nodes B and D is similar. Combining the fault mode descriptions associated with the key fault features nodes A and B in the fault mode knowledge base, the fault mode associated with node A is described as "early raceway spalling", and the fault mode associated with node B is described as "temperature rise due to poor lubrication". Targeted maintenance operation instructions are generated. The maintenance operation instruction for node A is "perform high-precision vibration spectrum analysis, check the raceway surface and replenish special grease", and the maintenance operation instruction for node B is "check the lubrication system oil supply pressure and flow, clean the oil circuit and replace the filter element". The final output maintenance priority list is a structured table containing four fields: data sample unit identifier, maintenance level, recommended maintenance time window, and specific maintenance operation instructions.

[0042] Optionally, the risk heatmap in the comprehensive diagnostic interface uses a color gradient for visualization, with risk values ​​ranging from low to high, corresponding to a color gradient from green to red. It can be understood that the calculation of maintenance response time limits can be based on a more complex scheduling model, such as considering the maintenance team's work calendar and task backlog. The calculation relationship of the time window can be described as follows: Wherein: T stamp The timestamp representing the data sample unit, This represents the baseline response time for the corresponding maintenance level. W represents the additional buffer time calculated based on the scheduling model. end This represents the deadline for the recommended maintenance time window. Optionally, the generation of maintenance operation instructions can be linked to a more detailed standard operating procedure database, indexing specific operation steps, required tools, and safety specifications based on fault mode and equipment model. It is understood that the presentation format of the maintenance priority list is not limited to a table; it can also be converted into a card-style list suitable for mobile devices or an interface for work orders integrated into an enterprise asset management system. In some embodiments, for multiple data sample units belonging to the same maintenance level and with overlapping time windows, the maintenance operation instructions in the maintenance priority list can be merged and optimized to generate a comprehensive maintenance task recommendation. In specific implementations, the maintenance priority list can be configured to be automatically pushed to designated maintenance management personnel terminals, with the pushed content including the description of the highest priority fault and the most urgent time window information.

[0043] See Figure 5 This is a bar chart comparing the risk values ​​of different data sample units, used to show the difference between the "initial risk value" and the "corrected risk value" of different units. It belongs to the category of risk assessment analysis charts. The corrected risk value of all units is higher than the initial risk value, with units A, B, and D showing the largest increases. Unit A has the highest corrected risk value among all units, while unit C has the lowest risk values ​​in both categories. This type of chart is typically used in fault diagnosis and risk calibration scenarios (such as industrial equipment condition assessment) to demonstrate the changes in risk values ​​after algorithmic correction, helping to identify the actual risk level of high-risk units.

[0044] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A tail bearing fault feature fusion diagnostic system based on SE-RESNET34, characterized in that, The system includes: The data preprocessing module acquires the raw operating data stream of the tail bearing, and uses a sliding window method based on a preset time window to segment the raw operating data stream, generating a series of data sample units with time continuity. The feature extraction module performs multi-source data weaving processing on each data sample unit to generate a diagnostic feature dataset containing vibration feature tensors, temperature feature vectors, and acoustic feature vectors. The deep feature fusion module inputs the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion and outputs a primary fault feature vector. The risk assessment module, based on the primary fault feature vector and the preset fault mode knowledge base, performs fault potential assessment and operational health assessment, generates dual-channel assessment results, calculates the preliminary fault risk quantification value through a dynamic weighting strategy based on the dual-channel assessment results, introduces the time correlation information of historical fault cases, and performs cross-sample unit collaborative correction on the preliminary fault risk quantification value to generate the corrected fault risk quantification value. The graph diagnosis module constructs a fault feature interaction graph based on the corrected fault risk quantification value, performs topology analysis on the fault feature interaction graph, identifies key fault feature nodes and their propagation paths, and generates a visual diagnosis report and maintenance priority list.

2. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 1, characterized in that, The specific steps for performing multi-source data weaving processing on each data sample unit include: The raw operational data stream includes vibration signals, temperature readings, and acoustic waveforms sorted by timestamp; The vibration signal within the data sample unit is resampled and time-aligned to ensure strict correspondence with the sampling time of the temperature reading and acoustic waveform; The frequency domain energy distribution features and time domain statistical features are extracted from the aligned vibration signal and combined to form the vibration feature tensor. The temperature change rate and temperature gradient are calculated from the aligned temperature readings to form the temperature feature vector; The acoustic feature vector is constructed by extracting sound pressure level spectral features and anomalous resonant frequency components from the aligned acoustic waveform; The vibration feature tensor, the temperature feature vector, and the acoustic feature vector are concatenated and normalized along the time dimension to finally output the diagnostic feature dataset.

3. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 1, characterized in that, The process of inputting the diagnostic feature dataset into the SE-RESNET34 network for deep feature fusion is as follows: The vibration feature tensor in the diagnostic feature dataset is input into the backbone convolutional layer of the SE-RESNET34 network to extract spatial hierarchical features. The temperature feature vector and acoustic feature vector in the diagnostic feature dataset are respectively input into independent feature embedding layers and mapped to a high-dimensional feature space; The spatially hierarchical features and the mapped high-dimensional features are concatenated along the channel dimension. The spliced ​​fused features are input into the compression excitation module of the SE-RESNET34 network, and the feature channel weights are adaptively calibrated through the channel attention mechanism. The fused features after weight calibration are input into the residual layer and fully connected layer of the SE-RESNET34 network, and finally the primary fault feature vector is compressed and output.

4. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 1, characterized in that, The steps for performing the fault potential assessment and operational health assessment are as follows: The similarity matching calculation is performed between the primary fault feature vector and the predefined fault feature template in the fault mode knowledge base; Based on the similarity matching calculation results, the fault mode category is determined and a fault confidence score is generated as the fault potential assessment result; Simultaneously, sub-vectors representing operational stability, load fluctuation, and lubrication effectiveness are extracted from the primary fault feature vector; The operational stability subvector, load volatility subvector, and lubrication effectiveness subvector are input into a preset health status scoring model to calculate the operational health score, which is used as the operational health assessment result. The failure potential assessment results and the operational health assessment results together constitute the dual-channel assessment results.

5. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 4, characterized in that, The process of calculating the preliminary fault risk quantification value through a dynamic weighting strategy includes: Extract the fault confidence score and the operational health score from the dual-channel evaluation results; Based on the current operating stage of the tail bearing, query the preset weight allocation table to obtain the fault confidence weight coefficient and the operating health weight coefficient; The fault confidence score is weighted using the fault confidence weighting coefficient, and the operational health score is weighted using the operational health weighting coefficient. The initial risk value is obtained by summing the weighted failure confidence score and the weighted operational health score. The initial risk value is input into a risk value normalizer based on a sigmoid function, and the output value is mapped to a preset quantization range to obtain the preliminary fault risk quantization value.

6. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 1, characterized in that, The process of collaboratively correcting the preliminary fault risk quantification value across sample units specifically includes: Select the current target data sample unit and locate its adjacent preceding and following data sample units in the time series. Preliminary fault risk quantification values ​​are obtained for the preceding data sample unit, the target data sample unit, and the subsequent data sample unit, respectively. Retrieve the historical failure cases, find past cases with similar characteristics to the target data sample unit, and extract the risk evolution sequence of the past cases before the failure occurred; Analyze the local trends of the risk evolution sequence and calculate the trend influence factor; Based on the preliminary fault risk quantification value of the preceding data sample unit, the preliminary fault risk quantification value of the following data sample unit, and the trend influence factor, a three-input collaborative correction function is constructed. The preliminary fault risk quantification value of the target data sample unit is input into the collaborative correction function, and the corrected fault risk quantification value is output after calculation.

7. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 1, characterized in that, The step of constructing the fault feature interaction map based on the corrected fault risk quantification value is as follows: The corrected fault risk quantification value of each data sample unit is used as the basic node attribute of the graph. Calculate the feature similarity, temporal proximity, and risk value correlation between any two data sample units, and use the calculation results as the edge weights connecting the corresponding two nodes in the graph; Based on all nodes and the connecting edges with edge weights, generate an initial global feature association graph; From the global feature association graph, the connecting edges whose edge weights exceed a set threshold are selected, and the nodes they connect are retained to form a densely connected subgraph structure; This densely connected subgraph structure is defined as the fault feature interaction graph.

8. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 7, characterized in that, The step of performing topological analysis on the fault feature interaction map further includes: The community detection algorithm is run on the fault feature interaction graph to identify node communities with tight internal connections and sparse external connections, and each node community is marked as a local fault feature cluster. Calculate the intra-cluster average risk value and risk value variance for each of the local fault feature clusters; According to the preset key node identification rules, nodes with risk values ​​significantly higher than the average risk value within the cluster and a large number of connecting edges are identified in each local fault feature cluster and marked as the key fault feature nodes. The key fault feature node is traced to all paths that reach other nodes through strong connection edges in the fault feature interaction graph, and the paths are organized into the propagation path.

9. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 1, characterized in that, The process of generating a visual diagnostic report includes: The timestamps, original signal segments, and corrected fault risk quantification values ​​of the data sample units corresponding to the key fault feature nodes are integrated. Integrate the data sample unit sequences involved in the propagation path and label the risk value transmission intensity along the path; The integrated information is overlaid and rendered with the topology of the fault feature interaction map to generate a comprehensive diagnostic interface that includes a time axis, risk heat map, feature map and path arrows. In the comprehensive diagnostic interface, all relevant data sample units are sorted from high to low according to the corrected fault risk quantification value; The sorting results are compared with preset maintenance action thresholds to generate the maintenance priority list containing specific maintenance recommendations.

10. The tail bearing fault feature fusion diagnostic system based on SE-RESNET34 according to claim 9, characterized in that, The specific logic for generating the maintenance priority list is as follows: The maintenance action threshold is set to multiple levels, with each level corresponding to a different risk value range and maintenance response time limit; The corrected fault risk quantification value of the sorted data sample unit is matched with the risk value range to determine the maintenance level of each data sample unit. Based on the timestamp of the data sample unit and the maintenance response time limit, calculate the time window for recommending the execution of maintenance operations; Based on the fault mode descriptions associated with the key fault feature nodes in the fault mode knowledge base, targeted maintenance operation instructions are generated. The final output maintenance priority list includes the data sample unit identifier, its maintenance level, the recommended maintenance time window, and specific maintenance operation instructions.

Images