A method and system for cross-load vibration-acoustic fusion fault diagnosis of a spherical roller bearing

Abstract

The application discloses a kind of cross-load sound-vibration fusion fault diagnosis method and system of self-aligning roller bearing, comprising: obtaining the vibration signal and acoustic signal of self-aligning roller bearing;Through time domain coding model, vibration signal is extracted to obtain vibration time domain feature, and acoustic signal is extracted to obtain acoustic frequency domain feature by frequency domain coding model;With vibration time domain feature as query vector, acoustic frequency domain feature as key vector and value vector, asymmetric cross attention is calculated to obtain cross attention feature;Based on vibration time domain feature and acoustic frequency domain feature, generate dynamic gating value, and use it to weight fusion vibration time domain feature and cross attention feature to obtain fusion feature;Based on fusion feature, determine fault diagnosis result.The application realizes high-precision, strong fault-tolerant fault diagnosis under cross-load working condition.

Description

Technical Field

[0001] This invention relates to the fields of mechanical fault diagnosis and artificial intelligence technology, specifically to a method and system for cross-load acoustic-vibration fusion fault diagnosis of self-aligning roller bearings. Background Technology

[0002] Self-aligning roller bearings, as core components of slewing bearings and transmission systems in high-end equipment, operate under complex conditions such as heavy loads, speed changes, and direction changes. Early fault identification is crucial for ensuring safe equipment operation. Existing intelligent fault diagnosis technologies often employ single-mode vibration or acoustic methods, or, while using multi-mode methods, employ homogeneous encoders to process different physical signals. However, vibration signals primarily characterize time-domain impact response, while acoustic signals reflect spectral energy distribution. Homogeneous encoders struggle to fully exploit the complementary physical characteristics of both, limiting feature extraction efficiency. Furthermore, existing fusion strategies often use fixed methods such as direct splicing or summation. When acoustic sensors experience a sharp drop in signal-to-noise ratio or even fail due to environmental noise interference or their own malfunctions, invalid features are forcibly introduced into the classifier, causing a drastic degradation in diagnostic performance and failing to meet the high reliability and fault tolerance requirements of industrial environments. Therefore, a fault diagnosis solution that can adapt to the differences in multi-modal physical characteristics and possess adaptive fault tolerance capabilities for sensor failure is urgently needed. Summary of the Invention

[0003] To address the problems of existing technologies, such as the inability of homogeneous encoders to fully utilize the complementary physical characteristics of acoustic and vibration signals and the severe degradation of diagnostic performance when sensors fail due to fixed fusion methods, this invention provides a cross-load acoustic and vibration fusion fault diagnosis method and system for self-aligning roller bearings, achieving high-precision and highly fault-tolerant fault diagnosis under cross-load conditions.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings, comprising: Step 1: Obtain the vibration and acoustic signals of the self-aligning roller bearing; Step 2: Extract features from the vibration signal using a time-domain coding model to obtain vibration time-domain features; extract features from the acoustic signal using a frequency-domain coding model to obtain acoustic frequency-domain features; Step 3: Using the vibration time-domain features as the query vector and the acoustic frequency-domain features as the key vector and value vector, perform asymmetric cross-attention calculation to obtain cross-attention features; Step 4: Generate a dynamic gating value based on the vibration time-domain features and the acoustic frequency-domain features, and use the dynamic gating value to perform weighted fusion of the vibration time-domain features and the cross-attention features to obtain fused features; Step 5: Determine the fault diagnosis result of the self-aligning roller bearing based on the fusion features; The frequency domain coding model and the network parameters for generating the dynamic gating value are obtained through a training process that includes a random discarding strategy for acoustic features. The training process includes: during the training phase, setting the acoustic frequency domain features to zero with a preset probability to simulate sensor failure in acquiring the acoustic signal.

[0005] The aforementioned method for cross-load acoustic-vibration fusion fault diagnosis of self-aligning roller bearings includes a time-domain coding model that sequentially comprises a convolution module, a sequence modeling module, and a self-attention module. The frequency domain coding model includes a parallel global spectral feature extraction branch and a frequency band sequence dependency extraction branch.

[0006] The aforementioned method for cross-load acoustic-vibration fusion fault diagnosis of self-aligning roller bearings includes a convolutional module with a channel attention mechanism for adaptive channel recalibration of the output feature map; a sequence modeling module employing a long short-term memory network; and a self-attention module employing a multi-head self-attention mechanism. The channel attention mechanism generates channel attention weights through global average pooling and a fully connected layer. The calculation formula is: ; in, For channel attention weights, For the sigmoid function, This represents the channel statistics obtained through global average pooling. , All are fully connected layers, and GELU is the Gaussian error linear unit activation function. The formula for calculating the recalibrated feature map is: ; in, The feature map after recalibration. This is the output feature map of the convolutional module. This indicates element-wise multiplication.

[0007] The aforementioned method for cross-load acoustic-vibration fusion fault diagnosis of self-aligning roller bearings uses a visual Transformer structure for the global spectral feature extraction branch. The spectrum of the acoustic signal is divided into multiple spectral blocks and a classification token and position code are added. Global spectral features are extracted through a multi-layer Transformer encoder. The frequency band sequence dependency extraction branch adopts a bidirectional long short-term memory network structure to reshape the spectrum of the acoustic signal into a frequency band sequence. The sequence dependency features are obtained by splicing and projecting the forward and reverse hidden states. The outputs of the global spectral feature extraction branch and the frequency band sequence dependency extraction branch are dynamically gated and fused through element-wise gated vectors to obtain acoustic frequency domain features, calculated as follows: ; in, Acoustic frequency domain characteristics, For layer normalization operation, For element-wise gated vectors, For global spectral characteristics, It is a sequence-dependent feature. This indicates element-wise multiplication.

[0008] The aforementioned method for diagnosing the acoustic-vibration fusion fault of self-aligning roller bearings under cross-load conditions, specifically includes the acquisition of the fusion features as follows: The vibration time-domain features and acoustic frequency-domain features are spliced ​​together, and a scalar-form dynamic gating value is generated through a fully connected layer. The cross-attention features are scaled based on the dynamic gating value and then residually connected with the vibration time-domain features to obtain fused features.

[0009] The aforementioned method for cross-load acoustic-vibration fusion fault diagnosis of self-aligning roller bearings uses a dual-channel vibration signal of the self-aligning roller bearing. The dual-channel vibration signal is extracted using an independent time-domain coding model to obtain the vibration features of the two channels. Symmetric cross-attention calculations are performed on the vibration characteristics of the two channels to obtain a unified vibration characterization; The unified vibration characterization is used as a vibration time-domain feature to perform the asymmetric cross-attention calculation.

[0010] In the aforementioned method for cross-load acoustic-vibration fusion fault diagnosis of self-aligning roller bearings, the acoustic feature random discarding strategy generates a binary mask with a preset probability of 0.2 during the training phase. When the mask is 0, the entire acoustic frequency domain feature vector is set to a zero vector, while the vibration time domain features remain unchanged.

[0011] The aforementioned method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings further includes the following training process: A file-level cross-load data partitioning strategy is adopted to isolate data files under different load conditions into training sets and test sets; The parameters are updated using a cross-entropy loss function that incorporates class imbalance weighting and label smoothing. The formula for the cross-entropy loss function is as follows: ; in, This is the total loss value. This represents the total number of fault categories. For the first Class weights For label smoothing factor, For uniquely encoded true labels, The model predicts the first Class probability.

[0012] A fault diagnosis system for self-aligning roller bearings under load-integrated acoustic and vibration conditions includes a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the aforementioned fault diagnosis method for self-aligning roller bearings under load-integrated acoustic and vibration conditions.

[0013] The aforementioned self-aligning roller bearing cross-load acoustic-vibration fusion fault diagnosis system also includes: vibration sensors and acoustic sensors associated with the self-aligning roller bearing, as well as a data acquisition module; The vibration sensor is configured to collect the vibration signal of the self-aligning roller bearing; The acoustic sensor is configured to acquire the acoustic signals of the self-aligning roller bearing; The data acquisition module is configured to transmit the vibration signal and the acoustic signal to the processor.

[0014] Compared with existing technologies, this invention constructs a modality-aware heterogeneous encoder framework. It employs a temporal coding model including convolution, sequence modeling, and self-attention for the temporal impact characteristics of vibration signals, and a frequency-domain coding model including parallel dual branches for the spectral energy characteristics of acoustic signals. This ensures precise matching between the encoder structure and the physical characteristics of each mode, overcoming the drawback of homogeneous encoders that struggle to fully exploit the complementary characteristics of acoustic and vibration signals, thus improving the discriminative power of feature representation. Furthermore, by performing asymmetric cross-attention calculation using highly reliable vibration temporal features as the query vector and easily interfered acoustic frequency-domain features as the key and value vectors, and combining this with dynamic gating values ​​for weighted fusion, the network can fully utilize acoustic information to enhance diagnostic accuracy when the acoustic signal quality is good. However, when the acoustic sensor fails or the signal-to-noise ratio is insufficient, the dynamic gating value automatically approaches zero, causing the fused output to smoothly degenerate into a vibration-feature-dominated mode. This achieves sensor failure tolerance without external fault detection, significantly improving diagnostic robustness under complex operating conditions. Furthermore, the symmetrical cross-attention fusion of dual-channel vibration signals provides redundant structural information enhancement capabilities, maintaining feature integrity even when a single vibration sensor fails locally. Combined with a training process that includes a random discarding strategy for acoustic features, the model is forced to learn diagnostic capabilities in acoustically missing conditions during the training phase, further enhancing its fault tolerance performance in practical deployments. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram comparing the ablation experimental results of the fusion method of the present invention; Figure 3 This is a schematic diagram of the confusion matrix for the 15 kN distributed external load diagnostic task of this invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, 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.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] Example 1: This example provides a method for fault diagnosis of self-aligning roller bearings under cross-load acoustic-vibration fusion. This method is mainly applied in the inference stage, aiming to achieve high-precision fault diagnosis with sensor fault tolerance under complex working conditions through heterogeneous feature extraction and adaptive fusion mechanisms. Figure 1 The method includes the following steps: Step 1: Acquire the vibration and acoustic signals of the self-aligning roller bearing. In this step, the vibration signal is typically acquired using an accelerometer mounted on the bearing housing or casing, reflecting the time-domain impact response of the mechanical structure. The acoustic signal is acquired using a non-contact sensor such as a microphone, reflecting the sound pressure fluctuations and spectral energy distribution propagating through the air. Before actual processing, the vibration signal can be z-score standardized, and the vibration signal can be divided into a time-domain window of a preset length (e.g., 2048 points) using a sliding window. The acoustic signal is then subjected to a Fast Fourier Transform (FFT) to obtain the spectrum. After high-pass filtering to remove low-frequency interference, a frequency domain representation of a preset dimension (e.g., 512 dimensions) is obtained. Then, frequency-by-frequency mean-standard deviation standardization is performed to eliminate dimensional differences and adapt to the input requirements of subsequent models. It should be understood that although this embodiment uses acceleration and sound pressure signals as examples, in other embodiments, the vibration signal can also come from a velocity or displacement sensor, and the acoustic signal can also come from a sound intensity probe, as long as it can characterize the bearing's operating state.

[0019] Step 2 involves extracting features from the vibration signal using a time-domain coding model to obtain the vibration time-domain features; and extracting features from the acoustic signal using a frequency-domain coding model to obtain the acoustic frequency-domain features. In this step, since the core fault information of the vibration signal is contained within the transient impacts and trend changes of the time-domain waveform, directly processing the original time-series data using a time-domain coding model effectively preserves local pulse characteristics and long-range time-series dependencies. In contrast, the fault characteristics of the acoustic signal are more manifested as energy modulation and harmonic structures in specific frequency bands; therefore, processing the spectral representation using a frequency-domain coding model can more accurately capture the global spectral envelope and the sequence dependencies between frequency bands. This heterogeneous encoder, designed for differences in physical characteristics, can more fully mine the complementary information of multimodal signals compared to a homogeneous encoder, improving the discriminative power of feature representation.

[0020] Step 3: Using the vibration time-domain features as the query vector and the acoustic frequency-domain features as the key and value vectors, perform asymmetric cross-attention calculation to obtain the cross-attention features. In this step, because vibration sensors are usually rigidly coupled to the equipment body in industrial environments, their signal-to-noise ratio is relatively high and their stability is strong; while acoustic sensors are easily affected by environmental noise, airflow interference, and installation location, resulting in relatively low reliability. Therefore, this embodiment uses the highly reliable vibration time-domain features as the query vector, meaning that the vibration features dominate "what to focus on" and "how to extract relevant information from the acoustic signal"; using the acoustic frequency-domain features as the key and value vectors means that the acoustic signal only serves as a source of information to provide supplementary details. This one-way asymmetric interaction mechanism ensures that the fusion process always uses stable vibration features as a benchmark, avoiding the risk of low-quality acoustic features back-contaminating or misleading the vibration features.

[0021] Step 4: A dynamic gating value is generated based on the vibration time-domain features and acoustic frequency-domain features. This dynamic gating value is then used to weight and fuse the vibration time-domain features and cross-attention features to obtain the fused features. In this step, the dynamic gating value is essentially a real-time confidence assessment of the quality or validity of the current acoustic signal. Specifically, this gating value can be adaptively generated based on the joint distribution of the two modal features. When the acoustic signal is clear and contains rich fault information, the dynamic gating value approaches 1, and the cross-attention features are fully integrated, enhancing diagnostic accuracy. When the acoustic sensor fails, detaches, or is submerged by strong background noise, the dynamic gating value automatically approaches 0, at which point the weighted fusion result smoothly degenerates into pure vibration time-domain features. This mechanism allows the model to achieve adaptive fault tolerance and seamless degradation during end-to-end inference without relying on an external sensor fault detection module, solving the problem of drastic performance degradation in traditional fixed-weight fusion methods when sensors are abnormal.

[0022] Step 5: Determine the fault diagnosis result of the self-aligning roller bearing based on the fused features. In this step, the fused features can be input into the fully connected layer and the classifier (such as Softmax), outputting the probability distribution of various fault states, and the category with the highest probability is taken as the final diagnosis result. Since the fused features have undergone enhancement processing through heterogeneous encoding and asymmetric gating fusion, they contain multimodal complementary information and have robustness under sensor failure, thus supporting high-reliability fault identification under cross-load conditions.

[0023] As a preferred implementation, the time-domain coding model sequentially includes a convolution module, a sequence modeling module, and a self-attention module; the frequency-domain coding model includes parallel global spectral feature extraction branches and frequency band sequence dependency extraction branches. Specifically, fault information in vibration signals often manifests as transient impacts and the resulting structural response attenuation. Therefore, the time-domain coding model adopts a serial cascaded structure of "convolution-sequence-self-attention," aiming to abstract features step-by-step from three dimensions: local waveform morphology, medium-duration temporal evolution, and global long-range dependency. In contrast, after Fourier transform, the fault characteristics of acoustic signals are reflected not only in the energy amplitude of specific frequency components but also in the harmonic correlation and modulation patterns between frequency bands. Therefore, the frequency-domain coding model adopts a parallel dual-branch structure, capturing the global spectral envelope and frequency band sequence dependency respectively, and then integrating the two complementary types of information through a fusion mechanism, thereby overcoming the limitation of a single structure in taking into account both global and local spectral characteristics.

[0024] For the specific implementation of the temporal coding model, the convolutional module adopts a two-layer cascaded convolution-batch normalization-activation-pooling structure. Each layer sequentially includes one-dimensional convolution (Conv1d), batch normalization (BN), ReLU activation, and max pooling operations. The kernel sizes of the two convolutional layers are set to 128 and 64, respectively, to capture low-frequency impulse components and mid-frequency modulation features sequentially; the pooling kernel size is set to 4 to progressively compress the temporal dimension; the number of channels increases from 1 to 32 to 128 to enrich the feature representation capability layer by layer. The output of the convolutional module is regularized by Dropout before entering the channel attention mechanism. The sequence modeling module uses a long short-term memory network; the self-attention module uses a multi-head self-attention mechanism. The channel attention mechanism generates channel attention weights through global average pooling and fully connected layers. The calculation formula is: ; in, For channel attention weights, For the sigmoid function, This represents the channel statistics obtained through global average pooling. , This is a fully connected layer, and GELU is the Gaussian error linear unit activation function. The formula for calculating the recalibrated feature map is: ; in, The feature map after recalibration. This is the output feature map of the convolutional module. This indicates element-wise multiplication.

[0025] In this embodiment, the GELU activation function is introduced instead of the traditional ReLU because GELU has smooth nonlinear characteristics, enabling more precise modeling of complex dependencies between channels and avoiding information loss caused by hard truncation. Meanwhile, and This creates a bottleneck structure, compressing computation while forcing the network to learn compact channel descriptors. Through the aforementioned recalibration operation, the model can automatically suppress redundant channels activated by environmental noise or irrelevant vibration components, enhancing the response strength to fault-sensitive channels.

[0026] Following the channel attention module, the sequence modeling module is implemented as follows: First, it undergoes layer normalization (LayerNorm) processing, then a single-layer unidirectional LSTM (128 hidden dimensions) is used to capture intermediate-length temporal dependencies through its forget gate, input gate, and output gate. The self-attention module is implemented by inputting the LSTM output into a multi-head self-attention module (with 4 attention heads) to capture long-range dependencies between arbitrary time steps. After residual connection and layer normalization, the self-attention output is compressed into a fixed-dimensional vibrational feature vector (e.g., 128-dimensional) through mean pooling with the arithmetic mean along the time dimension. It should be understood that although this embodiment uses LSTM as the sequence modeling module and multi-head self-attention as the self-attention module, in other embodiments, the sequence modeling module can also use a gated recurrent unit (GRU) or a temporal convolutional network (TCN), and the self-attention module can also use linear attention or sparse attention variants, as long as the corresponding temporal dependency capture and long-range correlation modeling functions can be achieved.

[0027] For the specific implementation of the frequency domain coding model, the global spectral feature extraction branch adopts a visual Transformer structure, dividing the spectrum of the acoustic signal into multiple spectral blocks and adding classification tokens and positional codes. Global spectral features are extracted through a multi-layer Transformer encoder. The frequency band sequence dependency extraction branch adopts a bidirectional long short-term memory network structure, reshaping the spectrum of the acoustic signal into a frequency band sequence. Sequence dependency features are obtained by concatenating and projecting forward and reverse hidden states. Specifically, the visual Transformer branch discretizes the continuous spectrum into a block sequence—for example, dividing a 512-dimensional spectrum into 16 spectral blocks of length 32 points, linearly projecting them to 128 dimensions, adding classification tokens (CLStoken) and learnable positional codes, and then inputting them into a multi-layer (e.g., 4-layer) Transformer encoder. It utilizes a self-attention mechanism to establish direct correlations between any two spectral blocks across the entire frequency band, thereby effectively capturing the overall energy distribution pattern and macroscopic envelope features across multiple frequency bands. Finally, global spectral features (e.g., 128 dimensions) are extracted from the classification tokens via a projection head. The bidirectional long short-term memory (LSTM) branch treats the spectrum as an ordered sequence of frequency bands—for example, reshaping a 512-dimensional spectrum into a 64×8 sequence and inputting it into a single-layer bidirectional LSTM (128 hidden dimensions). After concatenating the forward and backward hidden states, the dimension becomes 256, which is then reduced to 128 dimensions through linear projection. Finally, mean pooling aggregates all frequency band information—utilizing its recursive properties to mine local continuity and modulation patterns between adjacent frequency bands. This is particularly crucial for identifying fine structures such as sideband clusters caused by fault impacts. The outputs of the two branches are dynamically gated and fused using element-wise gated vectors to obtain acoustic frequency domain features, calculated using the following formula: ; in, Acoustic frequency domain characteristics, For layer normalization operation, For element-wise gated vectors, For global spectral characteristics, It is a sequence-dependent feature. This indicates element-wise multiplication. It's important to note that this refers to the element-wise gate vector. This is a vector with the same dimension as the feature, and its role is to adaptively balance the contribution ratio of global information and local sequence information at the feature level. For example, in some frequency bands, it may rely more on the global envelope, while in other frequency bands it may rely more on local sequence details. In specific implementation, this gate vector is generated through two fully connected layers: the first fully connected layer applies hyperbolic tangent (tanh) activation to the concatenated vector of ViT features and LSTM features to obtain intermediate latent features; the second fully connected layer maps the intermediate latent features to a vector with the same dimension as the feature and activates it with the sigmoid function, outputting an element-wise gate vector with values ​​in the range (0,1). This is fundamentally different from the scalar dynamic gate value used for acoustic-vibration fusion mentioned earlier in terms of both granularity and function: the former serves multi-scale feature optimization within a single mode, while the latter serves sensor failure tolerance across modes. In addition, the formula... The operation not only stabilizes the gradient distribution during training but also eliminates scale differences in the output features of different branches to a certain extent, making the fusion process smoother. Similarly, the combination of ViT and BiLSTM described above is only a preferred example. In practical applications, global branches can also use variants such as SwingTransformer or Perceiver, sequence branches can use bidirectional GRU or causal convolutional networks, and fusion methods can also use attention weighting or convolution after concatenation. This invention does not impose strict limitations on these methods.

[0028] As a preferred implementation, a dynamic gating value is generated based on the vibration time-domain features and the acoustic frequency-domain features, and the vibration time-domain features and cross-attention features are weighted and fused using the dynamic gating value to obtain fused features. Specifically, this includes: concatenating the vibration time-domain features and the acoustic frequency-domain features, generating a scalar form dynamic gating value through a fully connected layer; scaling the cross-attention features based on the dynamic gating value, and performing a residual connection with the vibration time-domain features to obtain fused features.

[0029] Specifically, the dynamic gating value in this embodiment is a scalar, which differs from the element-wise gating vector used within the frequency domain encoder mentioned earlier. The element-wise gating vector is used to weigh the feature contributions of different sub-branches (such as global spectrum and sequence dependence) within a single mode, and its dimension is the same as the feature dimension. The scalar dynamic gating value here, however, is used to evaluate the overall effectiveness or confidence of the acoustic signal at the cross-modal fusion level, and its value ranges from 0 to 1. In the generation path, the vibration time-domain features and acoustic frequency-domain features are first concatenated along the channel dimension to form a joint feature vector. This vector simultaneously contains the vibration response information of the mechanical structure and the acoustic environment information of air propagation. Subsequently, this joint feature vector is input into a fully connected layer and mapped to a scalar form dynamic gating value using a Sigmoid activation function. Due to the output characteristics of the Sigmoid function, It can smoothly reflect the complementary value of the current acoustic signal relative to the vibration signal: when the acoustic signal is clear and contains rich fault information, Approaching 1; when the acoustic sensor fails, falls off, or is overwhelmed by strong background noise, causing a severe deterioration in signal quality, It automatically approaches 0.

[0030] After obtaining the scalar dynamic gating value, it is used to scale the cross-attention features and then residually connected with the vibration time-domain features to obtain the final fused features. The calculation formula for this process is as follows: ; in, The final output is the fused feature. This is a layer normalization operation used to stabilize feature distribution and accelerate model convergence. For vibration time-domain characteristics, For scalar dynamic gating values, This is the cross-attention feature calculated using vibration time-domain features as the query vector and acoustic frequency-domain features as the key and value vectors. It should be noted that this formula constructs a residual learning framework based on vibration features. From a mathematical perspective, when the acoustic sensor is operating normally and the signal-to-noise ratio is high, Approaching 1, fusion feature It fully absorbs the acoustic complementary information retrieved through vibration characteristics. This improves diagnostic accuracy; however, when the acoustic sensor fails or the signal is interfered with by strong noise, The scaling factor automatically approaches 0, at which point the scaling factor... Approaching the zero vector, fusing features Numerically, it smoothly degenerates into pure vibrational time-domain characteristics. (After layer normalization only). This design enables the model to achieve adaptive elimination and seamless degradation of acoustic modes through end-to-end data-driven methods without adding any external sensor fault detection modules. This effectively avoids the misleading effect of invalid or contaminated acoustic features on diagnostic results and significantly enhances the robustness and fault tolerance of the system in complex industrial environments.

[0031] Furthermore, it should be understood that although this embodiment uses a concatenated fully connected layer followed by a sigmoid function to generate scalar dynamic gating values, other implementations can also employ other network structures capable of outputting scalar values ​​between 0 and 1 based on the joint distribution of bimodal features. Examples include multilayer perceptrons (MLPs), attention pooling mechanisms, or cosine similarity-based metric functions, as long as the function of adaptively adjusting the fusion weights according to the acoustic signal quality can be achieved. Simultaneously, the form of residual connections is not limited to direct addition; gated residual units or adaptive weighted summation variations can also be used, and this invention does not strictly limit these variations.

[0032] As a preferred embodiment, vibration and acoustic signals of the self-aligning roller bearing are acquired. Acquiring the vibration signals includes acquiring dual-channel vibration signals of the self-aligning roller bearing. Specifically, considering that a single vibration sensor in an industrial environment may experience information loss or distortion due to limited installation location, loose coupling, or obstruction of the transmission path by local structures, this embodiment employs a dual-channel redundant acquisition strategy. For example, accelerometers can be arranged in the horizontal and vertical directions of the bearing housing, or sensors can be arranged at different axial positions at the same measuring point to capture multi-dimensional mechanical vibration responses. This physical redundancy design provides a data foundation for subsequent algorithm fault tolerance.

[0033] Based on this, vibration signal features are extracted using a time-domain coding model to obtain vibration time-domain features. Specifically, this includes extracting features from the dual-channel vibration signal using independent time-domain coding models to obtain vibration features for each channel. "Independent" here means that the two time-domain coding models do not interfere with each other during network parameter updates, or although they share some underlying parameters, different regularization constraints are applied during training. For example, a stronger Dropout regularization or attention mask can be applied to the encoder of one channel, forcing the two branches to learn differentiated rather than homogeneous feature representations. In one specific implementation, the encoder of the first vibration channel is subject to stronger regularization: its attention Dropout is set to 0.35 (0.20 for the second vibration channel), and an additional LSTM Dropout of 0.25 is applied (no dropout for the second vibration channel), ensuring that the features extracted by the two channels are sufficiently differentiated. The purpose of this design is to avoid information redundancy caused by high correlation between the two channels, ensuring that even when the signal quality of one channel deteriorates, the other channel can still extract discriminative complementary features, thereby enhancing the model's robustness to local sensor failures.

[0034] Furthermore, symmetrical cross-attention calculation is performed on the vibration features of the two channels to obtain a unified vibration characterization. In asymmetric fusion, vibration features serve as query vectors to dominate information retrieval, reflecting the reliability differences between modes. However, in the dual-vibration channel fusion of this embodiment, since the two vibration sensors are equally important in terms of physical properties and both are prone to failure, a symmetrical interaction mechanism is adopted. That is, the features of the two channels serve as queries, keys, and values ​​for each other, enabling bidirectional information exchange. The specific calculation formula is as follows: ; ; in, This indicates that the first vibration channel focuses on the cross-attention output characteristics of the second vibration channel. This indicates that the second vibration channel focuses on the cross-attention output characteristics of the first vibration channel. For cross-attention function, and These are the channel vibration characteristics output by two independent time-domain coding models. , , These represent the query matrix, key matrix, and value matrix, respectively. Through the bidirectional calculations described above, the features of each channel can dynamically adjust their focus based on the context information of the other channel, achieving equal complementarity and redundancy enhancement of information.

[0035] Subsequently, the features obtained from the two-way interaction are integrated to generate a unified vibration characterization. For example, the features obtained from the two-way interaction can be integrated to generate a unified vibration characterization. and After being stitched along the channel dimension, a unified vibration characterization is obtained through mapping via a fully connected layer. This unified vibration characterization integrates structured information from two channels, preserving the unique perspective of each channel while eliminating local deviations caused by installation differences.

[0036] Finally, the unified vibration representation is used as the vibration time-domain feature for asymmetric cross-attention computation. This means that the vibration representation enhanced by first-level symmetric fusion will serve as a highly reliable reference information source in the second-level acoustic-vibration fusion process. Therefore, this invention not only achieves adaptive degradation when acoustic sensors fail at the algorithm level, but also builds fault tolerance capabilities when vibration sensors fail at the physical perception level, significantly improving the overall reliability and stability of the diagnostic system under complex and harsh operating conditions. It should be understood that although this embodiment uses a dual-channel example, in other embodiments, it can be extended to the acquisition and fusion of vibration signals with three or more channels. As long as the principles of symmetric interaction and independent encoding are followed, similar redundancy enhancement effects can be achieved.

[0037] Furthermore, the frequency domain coding model and the network parameters for generating dynamic gating values ​​in this embodiment are obtained through a training process that includes a random discarding strategy for acoustic features. This training process includes: during the training phase, setting the acoustic frequency domain features to zero with a preset probability to simulate sensor failure in acquiring acoustic signals.

[0038] Specifically, the random discarding strategy of acoustic features is the cornerstone of building the model's fault tolerance. During training, the system generates a binary mask with a preset probability (e.g., 0.2). When the mask is 0, the entire acoustic frequency domain feature vector is set to zero, while the vibration time domain features remain unchanged. Mathematically, this operation is equivalent to forcibly simulating fault scenarios where the acoustic sensor is completely detached, damaged, or submerged by extreme environmental noise. By repeatedly and randomly applying this "acoustic missing" pressure throughout the training cycle, the model is forced to learn a robust vibration-dominated representation: that is, when acoustic information is available, the model can use it to enhance diagnostic accuracy; while when acoustic information is artificially masked, the model must rely solely on vibration features to complete the correct fault classification, otherwise the loss function will not converge. It is this mandatory constraint during the training phase that enables the scalar dynamic gating value described above to... When faced with real acoustic failures during the inference phase, it can naturally approach 0 and smoothly switch to pure vibration mode, thereby achieving seamless end-to-end fault tolerance without the need for additional external sensor status detection modules.

[0039] Furthermore, to ensure stable generalization performance of the model under different load conditions, the training process also includes: employing a file-level cross-load data partitioning strategy to isolate data files under different load conditions into training and test sets. During industrial field data acquisition, consecutive samples within the same data file often exhibit high temporal correlation and similar operating conditions. If traditional sample-level random partitioning is used, it can easily lead to adjacent segments from the same file being included in the training and test sets, causing data leakage and similarity bias. This results in the model performing well in laboratory evaluations but experiencing a sharp performance drop when deployed in real-world cross-load conditions. The file-level partitioning strategy strictly requires that the data files included in the test set be physically independent of the training set. For example, all data files under 15kN, 20kN, and 25kN loads are used for training and validation, while all files under a 30kN load are used as the test set. This partitioning method ensures that the model faces a previously unseen distribution of operating conditions during the testing phase, thus truly reflecting its cross-load generalization ability. It also provides a clean verification environment for the aforementioned acoustic random dropout strategy, avoiding the problem of the model's over-reliance on a single mode being masked by overlapping operating conditions.

[0040] Furthermore, to address the common problems of sample imbalance and label noise in industrial fault diagnosis, the training process also employs a cross-entropy loss function that incorporates class imbalance weighting and label smoothing for parameter updates. The formula for this cross-entropy loss function is as follows: ; in, This is the total loss value. This represents the total number of fault categories. For the first Class weights For label smoothing factor, For uniquely encoded true labels, The model predicts the first Class probability. In this formula, class weights... Label smoothing factor is typically set inversely proportional to the frequency of samples in that class. This compensates for the long-tail distribution problem where normal samples far outnumber faulty samples, forcing the model to focus on learning the features of the minority class faults and avoiding the gradient update direction being dominated by the majority class samples. (For example, a value of 0.15) softens the originally hard labels (either 0 or 1) into a probability distribution, preventing the model from making overconfident predictions about training samples, thereby improving the model's generalization robustness when faced with labeling errors or samples with ambiguous boundaries. These two optimization measures, together with the aforementioned acoustic random discarding and file-level partitioning strategies, jointly create a diagnostic model that can automatically degrade when acoustically fails, remain stable under cross-load conditions, and accurately identify rare faults in imbalanced data.

[0041] Example 2: This example provides a self-aligning roller bearing cross-load acoustic and vibration fusion fault diagnosis system. The system includes a processor and a memory. The memory stores a computer program. When the processor executes the computer program, it implements the self-aligning roller bearing cross-load acoustic and vibration fusion fault diagnosis method as described in the previous example.

[0042] Specifically, the processor and memory constitute the core computing and control unit of the system. The processor can be a general-purpose central processing unit (CPU), graphics processing unit (GPU), field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), or embedded edge computing node (such as NVIDIA Jetson series, Raspberry Pi, etc.). In a preferred embodiment, considering the requirements of real-time performance and low power consumption in industrial settings, the processor uses an embedded edge computing device with integrated AI acceleration unit, which can complete heterogeneous encoder inference and dynamic gating fusion calculation locally without uploading all raw data to the cloud, thereby reducing communication latency and ensuring data security. The memory includes random access memory (RAM) for temporarily storing runtime data and non-volatile memory (such as Flash, SSD, or hard disk) for persistently storing computer programs, model weight parameters, and historical diagnostic records. The computer program referred to here includes not only the operating system and drivers, but more importantly, the parameter files and inference logic code of the trained time-domain coding model, frequency-domain coding model, asymmetric cross-attention module, and dynamic gating network, enabling the processor to load and execute the complete diagnostic process described in Embodiment 1 above. It should be understood that although this embodiment takes edge computing as an example, in other embodiments, the processor may also be a virtual computing instance of a cloud server, receiving data collected from the front end through a high-speed network and performing centralized processing. This invention does not strictly limit this.

[0043] Furthermore, to construct a complete sensing-transmission-processing closed loop, the system also includes: a vibration sensor and an acoustic sensor associated with the self-aligning roller bearing, as well as a data acquisition module. The vibration sensor is configured to acquire the vibration signal of the self-aligning roller bearing; the acoustic sensor is configured to acquire the acoustic signal of the self-aligning roller bearing; and the data acquisition module is configured to transmit the vibration signal and the acoustic signal to the processor.

[0044] Specifically, vibration sensors are used to capture the mechanical structural response of bearings during operation. In practical deployments, vibration sensors can take various forms, such as piezoelectric accelerometers, MEMS accelerometers, velocity sensors, or displacement sensors. Regarding installation methods, while traditional solutions often use rigid bolt coupling or magnetic mounting to ensure high-frequency response fidelity, this invention also supports non-contact measurement methods, such as laser Doppler vibration meters or fiber optic vibration sensors, considering the special constraints of certain high-end equipment (such as wind turbine main shafts and aero-engine accessories) with limited installation space, high temperatures and pressures, or the inability to wire rotating parts. This openness in installation methods ensures that the system can adapt to diverse industrial environments. Correspondingly, acoustic sensors are used to capture airborne sound pressure fluctuations, and can take the form of capacitive microphones, electret microphones, MEMS microphone arrays, or sound intensity probes. Since acoustic sensing is inherently non-contact, its placement is relatively flexible. It can be installed close to the bearing housing to obtain a high signal-to-noise ratio signal, or it can be installed on the inner wall of the cabin or the outside of the protective cover to avoid mechanical interference. This makes it possible to carry out fault diagnosis in scenarios where it is difficult to install vibration sensors.

[0045] The data acquisition module acts as a bridge connecting the physical world and the digital processing unit, responsible for converting the analog signals output by the sensors into digital signals and transmitting them synchronously to the processor. Specifically, the data acquisition module typically includes an analog-to-digital converter (ADC), an anti-aliasing filter, a preamplifier, and a synchronization trigger unit. To ensure the effectiveness of the subsequent acoustic-vibration fusion algorithm, the data acquisition module must have multi-channel synchronous acquisition capabilities, ensuring strict alignment of the vibration and acoustic signals on the time axis, and avoiding incorrect temporal correlation features extracted by the cross-attention mechanism due to sampling time deviations. In terms of data transmission, the data acquisition module and the processor can be connected via wired methods (such as Ethernet, USB, fiber optic, PCIe bus) or wireless methods (such as Wi-Fi, 5G, ZigBee, LoRa). For example, in large rotating machinery monitoring scenarios, a wireless sensor network can be used to aggregate data from dispersed acquisition nodes to an edge gateway, which then forwards the data to the processor, thus avoiding the construction difficulties and signal interference risks associated with long-distance cabling.

[0046] Example 3: To verify the effectiveness of the proposed cross-load acoustic-vibration fusion fault diagnosis method for self-aligning roller bearings under actual complex working conditions, this example constructed a fault diagnosis test bench for steel belt elevators for application verification, as shown in Table 1. It should be understood that although this example uses elevator bearings as a specific application scenario, the technical solution of this invention is also applicable to other high-end equipment using self-aligning roller bearings, such as wind turbine main shafts, mining crushers, and ship propulsion shafting systems; this invention does not strictly limit its application in this regard.

[0047] Table 1

[0048] To comprehensively evaluate the method's performance in terms of accuracy, generalization, and fault tolerance, 14 sets of comparative experiments were set up, as shown in Table 2. IDs 1-7 represent the longitudinal comparison under different channel configurations in this study, used for sensor failure tolerance analysis; IDs 8-11 represent encoder homogeneous ablation; and IDs 12-14 represent fusion-based ablation.

[0049] Table 2

[0050] The experimental results are as follows: 1. Overall performance First, the cross-load performance of the present invention (ID 7) and different channel configurations in the complete sensor state was examined, and the results are shown in Table 3: Table 3

[0051] As can be seen, the method proposed in this invention achieves an average accuracy of 97.81% across four loads, with a standard deviation of only 2.9% across loads. This indicates superior robustness to load variations.

[0052] To evaluate the fault tolerance capability in the event of sensor failure, three typical failure scenarios were simulated: Scenario 1: Vibration sensor acc1 fails, and the model input degrades to ID 6 (acc2 + sound). In this scenario, the average accuracy under four loads is 96.78%, a decrease of 1.03 percentage points from the complete state (97.81%); under a 15 kN distributed external load, the accuracy is 91.77% (93.51% in the complete state), a decrease of 1.74 percentage points. The performance degradation is much smaller than the overall degradation of the single vibration channel acc1 (the average value of acc1 single channel is only 94.03%, 3.78 percentage points lower than the complete state), indicating that the fusion of the remaining vibration and sound effectively compensates for the information loss of the failed channel.

[0053] Scenario 2: Vibration sensor acc2 fails, and the model degenerates to ID 5 (acc1 + sound). In this scenario, the average accuracy across four loads is 95.97%, a decrease of 1.84 percentage points from the complete state, but still higher than the average of the single vibration channel acc1 (94.03%), indicating that the fusion of the sound channel and the remaining vibration channels has substantial diagnostic gain. It is worth noting that the accuracy of acc1 + sound under distributed off-load (15 kN) is only 88.77%, a significant difference from the complete state (a decrease of 4.74 percentage points), reflecting that the acc1 channel faces greater distribution offset in cross-load scenarios.

[0054] Scenario 3: Both vibration sensors fail, and the model relies solely on the audio channel (ID 3). In this scenario, the average accuracy across four loads is 89.21%. Under a 25 kN distributed load, the single audio channel achieves 97.23%, approaching the level of dual-channel fusion (98.09%), demonstrating that the audio signal contains rich fault information under specific operating conditions. However, under a 15 kN distributed load, the accuracy is only 75.63%, revealing the risk of insufficient audio signal-to-noise ratio. Notably, even with the failure of sensor acc2 (acc1 + audio, average 95.97%), the diagnostic performance is still far superior to the scenario where both vibration sensors fail, fully demonstrating the redundancy value of the "vibration + audio" multimodal design.

[0055] The above analysis shows that the dual vibration redundancy design and gated adaptive fusion adopted in the method of this invention make the performance degradation of the model controllable (1.03%-1.84%) when a single sensor fails. Thus, the adaptive fault-tolerant characteristics of fully utilizing multi-source complementary information when each sensor is operating normally and automatically and progressively degrading when a sensor fails have important engineering value for elevator shaft scenarios where frequent maintenance is not possible.

[0056] 2. Ablation test 2.1 Encoder Homogeneity Ablation The results of replacing the heterogeneous encoder of this invention with homogeneous encoders (All-CNN, All-LSTM, All-BiLSTM, All-ViT) are shown in Table 4: Table 4

[0057] As can be seen, compared with this invention, the average accuracy of the four homogeneous encoders in cross-load diagnostic tasks decreased by 0.72, 4.92, 8.61, and 3.40 percentage points, respectively, with the difference being more significant under a 15 kN distributed external load (All-ViT only 85.14%). It can be seen that the homogeneous architecture not only has low accuracy but also lacks customized design for different physical characteristics, cannot adaptively degrade when the sensor fails, and has significantly insufficient robustness.

[0058] 2.2 Fusion Ablation like Figure 2As shown, with a fixed heterogeneous encoder, the gated asymmetric fusion of the present invention is compared with simple concatenation fusion (SimpleConcat), simple addition fusion (SimpleSum), and symmetric attention fusion (SymAttn). The average accuracy of the four methods is similar (97.31%–97.85%), indicating that the choice of fusion strategy has a limited impact on the overall accuracy, provided that the encoder has fully extracted the discriminative features. However, the core advantage of the gated asymmetric fusion of the present invention lies in its sensor failure tolerance: when the sound sensor fails, concatenation fusion and addition fusion cannot adaptively remove the features of the failed channel, which may lead to invalid or highly contaminated features being forcibly introduced into the classifier; while gated asymmetric fusion, through g... s →0 automatically degrades to pure vibration mode, achieving seamless fault tolerance. This explainable degradation behavior is not found in other fusion strategies. Notably, the mean of the dual vibration channels (acc1+acc2, without sound) has reached 97.81%, indicating that in scenarios with good vibration signal quality, dual vibration redundancy can provide reliable diagnostic capabilities, and the contribution of the sound channel is more reflected in its backup role when the vibration sensor fails.

[0059] 2.3 Confusion Matrix Analysis Figure 3 The confusion matrix of this invention under a 15 kN distributed external load diagnostic task is shown. Under this load, the overall accuracy of the model is 93.51%, with 7 types of faults (500 test samples per type). The recall rate analysis of each category is as follows: the recall rate for normal state is 99.0% (495 / 500), indicating that the model rarely misclassifies normal bearings as faults; the recall rates for small outer ring faults, small inner ring faults, and compound fault 1 (small outer ring fault + large inner ring fault) are all 100% (500 / 500), indicating that they are correctly identified; the recall rate for large inner ring faults is 91.4% (457 / 500), and the recall rate for large outer ring faults is 85.4% (427 / 500). The main confusion direction is compound fault 2 (large outer ring fault + small inner ring fault), and the harmonic structures of the two are highly overlapping in the acoustic spectrum; the recall rate for compound fault 2 is only 78.8%, which is the most difficult category to classify under this load, and some of them are misclassified as large outer ring faults and compound fault 1. From a physical mechanism perspective, this is because under low load conditions, the impact energy generated by the composite fault is relatively weak, and the harmonic structure in its acoustic spectrum highly overlaps with that of a single outer ring fault, leading to a reduction in feature distinguishability. This analysis not only demonstrates the authenticity and reproducibility of the technical solution of this invention, but also points the way for subsequent optimization. For example, more fine-grained token-level cross-modal alignment or the introduction of a load-aware adaptive normalization mechanism can be explored to further improve the identification capability of composite faults under low signal-to-noise ratio conditions.

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings, characterized in that, include: Step 1: Obtain the vibration and acoustic signals of the self-aligning roller bearing; Step 2: Extract features from the vibration signal using a time-domain coding model to obtain vibration time-domain features; extract features from the acoustic signal using a frequency-domain coding model to obtain acoustic frequency-domain features; Step 3: Using the vibration time-domain features as the query vector and the acoustic frequency-domain features as the key vector and value vector, perform asymmetric cross-attention calculation to obtain cross-attention features; Step 4: Generate a dynamic gating value based on the vibration time-domain features and the acoustic frequency-domain features, and use the dynamic gating value to perform weighted fusion of the vibration time-domain features and the cross-attention features to obtain fused features; Step 5: Determine the fault diagnosis result of the self-aligning roller bearing based on the fusion features; The frequency domain coding model and the network parameters for generating the dynamic gating value are obtained through a training process that includes a random discarding strategy for acoustic features. The training process includes: during the training phase, setting the acoustic frequency domain features to zero with a preset probability to simulate sensor failure in acquiring the acoustic signal.

2. The method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings according to claim 1, characterized in that, The temporal coding model sequentially includes a convolutional module, a sequence modeling module, and a self-attention module; The frequency domain coding model includes a parallel global spectral feature extraction branch and a frequency band sequence dependency extraction branch.

3. The method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings according to claim 2, characterized in that, The convolutional module includes a channel attention mechanism for adaptive channel recalibration of the output feature map; the sequence modeling module employs a long short-term memory network; and the self-attention module employs a multi-head self-attention mechanism. The channel attention mechanism generates channel attention weights through global average pooling and a fully connected layer. The calculation formula is: ； in, For channel attention weights, For the sigmoid function, This represents the channel statistics obtained through global average pooling. , All are fully connected layers, and GELU is the Gaussian error linear unit activation function. The formula for calculating the recalibrated feature map is: ； in, The feature map after recalibration. This is the output feature map of the convolutional module. This indicates element-wise multiplication.

4. The method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings according to claim 2, characterized in that, The global spectral feature extraction branch adopts a visual Transformer structure, which divides the spectrum of the acoustic signal into multiple spectral blocks and adds classification tokens and position codes, and extracts global spectral features through a multi-layer Transformer encoder. The frequency band sequence dependency extraction branch adopts a bidirectional long short-term memory network structure to reshape the spectrum of the acoustic signal into a frequency band sequence. The sequence dependency features are obtained by splicing and projecting the forward and reverse hidden states. The outputs of the global spectral feature extraction branch and the frequency band sequence dependency extraction branch are dynamically gated and fused through element-wise gated vectors to obtain acoustic frequency domain features, calculated as follows: ； in, Acoustic frequency domain characteristics, For layer normalization operation, For element-wise gated vectors, For global spectral characteristics, It is a sequence-dependent feature. This indicates element-wise multiplication.

5. The method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings according to claim 1, characterized in that, The acquisition of the fusion features specifically includes: The vibration time-domain features and acoustic frequency-domain features are spliced ​​together, and a scalar-form dynamic gating value is generated through a fully connected layer. The cross-attention features are scaled based on the dynamic gating value and then residually connected with the vibration time-domain features to obtain fused features.

6. The method for diagnosing trans-load acoustic-vibration fusion faults of self-aligning roller bearings according to claim 1, characterized in that, The vibration signal is a dual-channel vibration signal of a self-aligning roller bearing. The dual-channel vibration signal is subjected to feature extraction through independent time-domain coding models to obtain the vibration features of the two channels. Symmetric cross-attention calculations are performed on the vibration characteristics of the two channels to obtain a unified vibration characterization; The unified vibration characterization is used as a vibration time-domain feature to perform the asymmetric cross-attention calculation.

7. The method for diagnosing trans-load acoustic-vibration fusion faults in self-aligning roller bearings according to claim 1, characterized in that, In the acoustic feature random discarding strategy, a binary mask is generated with a preset probability of 0.2 during the training phase. When the mask is 0, the entire acoustic frequency domain feature vector is set to a zero vector, while the vibration time domain features remain unchanged.

8. The method for diagnosing trans-load acoustic-vibration fusion faults of self-aligning roller bearings according to claim 7, characterized in that, The training process also includes: A file-level cross-load data partitioning strategy is adopted to isolate data files under different load conditions into training sets and test sets; The parameters are updated using a cross-entropy loss function that incorporates class imbalance weighting and label smoothing. The formula for the cross-entropy loss function is as follows: ； in, This is the total loss value. This represents the total number of fault categories. For the first Class weights For label smoothing factor, For uniquely encoded true labels, The model predicts the first Class probability.

9. A fault diagnosis system for cross-load acoustic and vibration fusion of self-aligning roller bearings, characterized in that, include: The processor and memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the cross-load acoustic-vibration fusion fault diagnosis method for self-aligning roller bearings as described in any one of claims 1-8.

10. The self-aligning roller bearing cross-load acoustic-vibration fusion fault diagnosis system according to claim 9, characterized in that, Also includes: Vibration sensors and acoustic sensors associated with self-aligning roller bearings, as well as a data acquisition module; The vibration sensor is configured to collect the vibration signal of the self-aligning roller bearing; The acoustic sensor is configured to acquire the acoustic signals of the self-aligning roller bearing; The data acquisition module is configured to transmit the vibration signal and the acoustic signal to the processor.

Images