Ship propulsion shafting fault diagnosis method based on multi-modal attention fusion

Abstract

The invention discloses a ship propulsion shafting fault diagnosis method based on multi-modal attention fusion, and relates to the technical field of ship fault diagnosis. According to the method, multi-modal data such as vibration parameters, lubricating oil parameters and cooling water parameters of key parts of a ship propulsion shafting are synchronously collected and preprocessed; a modal specific feature extraction strategy is adopted, a multi-modal attention fusion mechanism including intra-modal self-attention, inter-modal cross attention and adaptive dynamic weight distribution is designed, and heterogeneous modal features are effectively integrated. Compared with a single-mode method and a traditional feature splicing method, the method has the advantage that the diagnosis accuracy is remarkably improved. The method has high robustness, and can still maintain high diagnosis accuracy even under the condition that part of sensors fail.

Description

Technical Field

[0001] The present invention relates to the technical field of ship fault diagnosis, and in particular to a ship propulsion shafting fault diagnosis method based on multimodal attention fusion. Background Art

[0002] As a critical component in ship power transmission, the health of a ship's propulsion shafting system is directly related to its safe operation. Traditional propulsion shafting fault diagnosis methods rely primarily on a single information source (such as vibration signals), which cannot fully reflect the complex fault characteristics of the shafting system. This results in limited diagnostic accuracy and frequent false positives and missed positives. With the advancement of sensing technology, the types of data acquired from propulsion shafting systems are becoming increasingly diverse. The key to improving fault diagnosis performance is the effective integration of this multi-source data.

[0003] Existing multi-source data fusion methods suffer from the following shortcomings: most rely on simple data splicing, failing to fully exploit the correlation and complementarity between different modal data; the physical meanings of different modal data vary significantly, creating a semantic gap that complicates information fusion; and the uneven quality of different modal data and variations in sensor signal-to-noise ratios make diagnostic results unstable and unreliable under complex operating conditions. Therefore, a method for ship propulsion shafting fault diagnosis is urgently needed that can effectively integrate multimodal sensor data and intelligently adjust the importance of different modalities to improve diagnostic accuracy and reliability. Summary of the Invention

[0004] To address the above-mentioned issues and technical needs, the inventors have proposed a method for diagnosing ship propulsion shafting faults based on multimodal attention fusion. This method significantly improves the accuracy and reliability of shafting fault diagnosis by synchronously collecting multi-source data such as vibration parameters, lubricating oil parameters, and cooling water parameters, and adaptively integrating heterogeneous modal features using a multimodal attention fusion mechanism. The technical solutions of the present invention are as follows:

[0005] A ship propulsion shafting fault diagnosis method based on multimodal attention fusion includes the following steps:

[0006] (1) Multi-source data collection is performed on key parts of the ship's propulsion shaft system, and data quality is optimized through data preprocessing. The multi-source data includes vibration parameters, lubricating oil parameters, and cooling water parameters.

[0007] (2) A modal-specific feature extraction strategy is used to obtain the corresponding modal features for different modal parameters, including the multidimensional features of vibration parameters and the time domain feature quantification of lubricating oil and cooling water parameters.

[0008] (3) The features of each modality are fused with weighted multimodal features based on the multimodal attention fusion mechanism.

[0009] (4) The fused features are input into the multi-layer perceptron classifier to classify the shaft system fault type.

[0010] A further technical solution is that the step (1) specifically comprises:

[0011] A multi-channel sensor network is constructed to collect multi-source data from key parts of the ship's propulsion shafting system, including vibration acceleration and displacement signals at key locations of the bearing seat and shafting system, temperature, flow, and pressure signals of the lubricating oil, and temperature, flow, and pressure signals of the cooling water. All collected signal data and their timestamps are stored in a time series database. To improve the accuracy of signal analysis, a corresponding noise reduction algorithm is selected based on the characteristics of each modal data. To address the time alignment problem of multimodal data, interpolation technology is used to align data of different sampling frequencies to a unified time base. At the same time, based on the distribution characteristics of different modal data, standardization methods such as Z-score and Min-Max are used for normalization processing to ensure the comparability of cross-modal data.

[0012] A further technical solution is that the step (2) specifically includes:

[0013] For vibration parameters, time domain features, spectral features, time-frequency features, and nonlinear features are extracted. For lubricating oil and cooling water parameters, time domain feature vectors of temperature (such as temperature gradient), pressure (such as pressure fluctuation), and flow rate changes are extracted.

[0014] Its further technical solution is that the step (3) specifically includes:

[0015] The multimodal attention fusion mechanism achieves effective weighted fusion of multimodal features through the collaborative work of three key components: intra-modal self-attention, inter-modal cross-attention, and adaptive dynamic weight allocation. The multidimensional features of vibration parameters, the time-domain features of lubricating oil parameters, and the time-domain features of cooling water parameters are each self-attentioned through an intra-modal self-attention mechanism, automatically identifying key features within the modality, suppressing redundant information, and improving the quality of feature representation. The inter-modal cross-attention mechanism evaluates the correlation and complementarity between any two modal key features, quantifying the contribution of each modality to fault diagnosis. The adaptive dynamic weight allocation mechanism comprehensively considers the current operating parameters and data quality scores, assigning dynamic weights to each modality and performing weighted fusion to generate fused features. This allows the system to adapt to changing operating conditions, automatically adjust dependency strategies when sensor data quality fluctuates, and maintain diagnostic performance stability.

[0016] A further technical solution is that the step (4) specifically includes:

[0017] The fused features are used to classify shafting fault types through a multi-layer perceptron network and a softmax activation function. To address the imbalance in shafting fault data, a Focal loss function is used as the network optimization objective, combined with an adaptive gradient descent optimization algorithm for iterative parameter updates. This effectively enhances the model's recognition accuracy and robustness for rare fault samples.

[0018] The beneficial technical effects of the present invention are:

[0019] This invention achieves a deep fusion of vibration characteristics, lubricating oil parameters, and cooling water parameters through a combination of intra-modal self-attention and inter-modal cross-attention mechanisms. This effectively overcomes the problems of incomplete information and significant differences in physical meaning from a single modal, significantly improving the accuracy of shaft fault diagnosis. An adaptive dynamic weight allocation mechanism based on operating condition perception and data quality assessment enables the system to dynamically adjust modal dependencies based on operating parameters and sensor status, maintaining high diagnostic performance even in the event of partial sensor failure, significantly enhancing the system's robustness under complex operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a ship propulsion shafting fault diagnosis method based on multimodal attention fusion provided by this application;

[0021] Figure 2 This is a schematic diagram of the ship propulsion shafting test bench provided in this application;

[0022] Figure 3 This is the structure diagram of the multimodal attention fusion mechanism provided by this application;

[0023] Figure 4 This is the computational flow chart of the intra-modal self-attention mechanism provided by this application;

[0024] Figure 5 This is the computational flow chart of the inter-modal cross-attention mechanism provided by this application;

[0025] Figure 6 This is a schematic diagram of the adaptive dynamic weight allocation mechanism provided by this application. DETAILED DESCRIPTION

[0026] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0027] An embodiment of the present application provides a ship propulsion shaft fault diagnosis method based on multimodal attention fusion, please refer to Figure 1 As shown, the method includes the following parts:

[0028] (1) Multi-source data collection and preprocessing.

[0029] This embodiment deploys a multi-type sensor array at key nodes of the ship propulsion shafting test bench to collect multi-source data of key locations. Figure 2 As shown in the figure, the ship propulsion shafting test rig primarily consists of a propulsion motor, a highly elastic coupling, a thrust bearing, a stern shaft, fore and aft bearings, a hydraulic oil system, and a cooling water system. Triaxial accelerometers are installed between the propulsion motor and the highly elastic coupling, between the highly elastic coupling and the thrust bearing, and on the thrust bearing seat, stern shaft, and fore and aft bearing seats to collect vibration acceleration and displacement signals. The sampling frequency is set at 10kHz, the sensitivity is 100mV / g, and the measurement range is ±50g. A temperature sensor (measuring range -20°C to 120°C, accuracy ±0.5°C), a pressure sensor (measuring range 0 to 1MPa, accuracy ±0.5%), and a flow sensor (measuring range 0 to 100L / min, accuracy ±1%) are installed at the inlet and outlet of the hydraulic oil system to collect lubricating oil temperature, flow rate, and pressure signals. Temperature sensors (measuring range -10℃~100℃, accuracy ±0.5℃), pressure sensors (measuring range 0~0.6MPa, accuracy ±0.5%) and flow sensors (measuring range 0~200L / min, accuracy ±1%) are installed at the inlet and outlet of the cooling water system to collect the temperature, flow and pressure signals of the cooling water. At the same time, a temperature sensor is installed on the test bench, and a speed sensor, torque sensor and power sensor are installed between the propulsion motor and the high-elastic coupling to collect operating parameters such as the current speed, load and ambient temperature. All collected sensor data are transmitted to the central data processing unit via industrial Ethernet and stored in a time series database that supports high-frequency data writing and query. Each data record contains a unique sensor identifier, measurement value, timestamp (accuracy up to microseconds) and tag information, where the tag information records the fault type of a single experiment.

[0030] To address the heterogeneous modal data characteristics, a differentiated data preprocessing strategy was implemented. First, the vibration signal was denoised using wavelet thresholding. In one possible implementation, the db4 wavelet basis was selected, with a 5-level decomposition. Signal reconstruction was then performed after processing the wavelet coefficients using a soft thresholding function. Lubricating oil and cooling water parameters were filtered using a combination of median filtering (with a window size of 5) and sliding average filtering (with a window size of 3) to remove outliers and high-frequency noise. In one possible implementation, the parameters were first subjected to median filtering. This method effectively removes impulse noise and salt-and-pepper noise by sorting a set of data from smallest to largest and taking the median value as the output. This method then applied a sliding average filter to the median-filtered parameters. This method smoothes the signal by calculating the average of the data set, effectively removing random noise. Second, a cubic spline interpolation algorithm was used to align signals of different sampling frequencies, ensuring temporal consistency of the multimodal data. Finally, the vibration signal was demeaned and variance normalized using Z-score normalization (μ = 0, σ = 1). Lubricating oil and cooling water parameters were normalized to the range [0, 1] using Min-Max normalization.

[0031] (2) Multimodal feature extraction

[0032] For different modal parameters, a modal-specific feature extraction strategy is adopted to fully explore the corresponding modal characteristics of various signals. Among them, vibration signal feature extraction covers four dimensions, among which the time domain statistical features (13 dimensions) include mean, root mean square value, peak value, kurtosis, skewness, form factor, impulse factor, margin factor, crest factor, variance, standard deviation, maximum value and minimum value. Spectral features (13 dimensions) use FFT transformation to extract spectral energy distribution characteristics, including centroid frequency, root mean square frequency, frequency variance, main frequency amplitude and its corresponding frequency, frequency band energy ratio (divided into 6 frequency bands), spectral entropy and spectral kurtosis. Time-frequency features (10 dimensions) use the variational mode decomposition (VMD) method, setting the decomposition mode number K to 5 and the penalty factor α to 2000 to extract the energy ratio and center frequency of each intrinsic mode component. The nonlinear dynamic characteristics (8 dimensions) are used to characterize the complexity and nonlinear dynamic characteristics of vibration signals by calculating approximate entropy, sample entropy, permutation entropy, fuzzy entropy, Lyapunov exponent, correlation dimension, fractal dimension and Hurst exponent.

[0033] For lubricating oil parameters, the following three types of time-domain features are extracted (9 dimensions for each parameter, for a total of 27 dimensions): temperature features include mean, standard deviation, rate of change, maximum, minimum, peak-to-peak value, temperature rise rate, temperature drop rate, and temperature stability index; pressure and flow features each include mean, standard deviation, rate of change, maximum, minimum, peak-to-peak value, fluctuation index, rate of decrease, and stability index. The time-domain feature extraction for cooling water parameters is consistent with that for lubricating oil parameters, including temperature, pressure, and flow features, with 9 dimensions for each parameter, for a total of 27 dimensions.

[0034] (3) Multimodal Attention Fusion Mechanism

[0035] Each modal feature is weightedly fused based on the multimodal attention fusion mechanism, which consists of the intra-modal self-attention mechanism, the inter-modal cross-attention mechanism and the adaptive dynamic weight allocation mechanism. Figure 3 As shown, in one possible implementation, the intra-modality self-attention mechanism and the inter-modality cross-attention mechanism are stacked N=3 times each to enhance feature learning performance. The working principles of each mechanism are described below:

[0036] Intra-modal self-attention mechanism such as Figure 4 As shown, for each modal feature Generate query matrix through linear transformation Bond Matrix Value Matrix in is the learnable weight matrix, n is the number of samples, d m d k They represent the input feature dimensions of the self-attention within the modality and the projection dimensions of the query, key, and value vectors in the attention mechanism. The self-attention weights are calculated by the query matrix Q. m and bond matrix K m The dot product is calculated and scaled and softmax normalized to finally be combined with the value matrix V m Multiply to get enhanced feature Z m , as the key internal feature of the modality, is expressed as follows:

[0037]

[0038] like Figure 5 As shown, the inter-modal cross attention mechanism is implemented by the following steps: For any two key features within the modality and Calculate the cross attention weight by querying the matrix and bond matrix The dot product is calculated and scaled and normalized by softmax, and finally combined with the value matrix Multiply to get the cross feature C ij , which is expressed as follows:

[0039]

[0040] in is the learnable weight matrix, d p d q They represent the input feature dimensions of the inter-modal cross-attention and the feature dimensions of the cross-attention output, respectively.

[0041] For each modality, aggregate the cross features C from all other modalities ij , get the modal feature C after cross attention i , which is expressed as follows:

[0042] C i =∑ j≠i α ij C ij

[0043] where α ij is the correlation coefficient between the modes, which is calculated through learnable parameters.

[0044] Adaptive dynamic weight allocation mechanism such as Figure 6 As shown in Figure 1, it is used to dynamically adjust the importance of different modes based on the current operating parameters and sensor data quality. The operating condition perception weight is calculated based on operating parameters such as current speed, load (torque and power), and ambient temperature, reflecting the relative importance of each mode under different operating conditions. Data quality assessment evaluates the quality of each modal signal through signal-to-noise ratio, information entropy, and signal consistency indicators. The adaptive dynamic weight allocation mechanism is implemented through the following steps:

[0045] First, the working condition perception network and data quality assessment network are constructed based on the multi-layer perceptron network MLP. The vectors of the current working condition operating parameters (such as speed, load, etc.) are spliced ​​and input into the working condition perception network to obtain the working condition embedding vector H. op , which is expressed as follows:

[0046] H op =MLP(P op )

[0047] Among them, P op is the working condition parameter splicing vector, H op is the working condition embedding vector.

[0048] The collected multi-source data are input into the data quality assessment network to evaluate the quality of each modality data, which is expressed as follows:

[0049] Q i =QualityNet(X i )

[0050] where Q i is the original data X of the i-th mode i The quality score of is in the range of [0,1].

[0051] For each mode, according to the current working condition parameters and the modal features C output by the cross-attention mechanism between modes i And the original data quality score Q corresponding to the modalityi , calculate its dynamic weight coefficient:

[0052] W i =softmax(f(H op ,C i ,Q i ))

[0053] Among them, f is a multi-layer perceptron network that comprehensively considers working condition information, feature representation and signal quality.

[0054] For all modal features C i Perform weighted fusion to generate fusion features:

[0055]

[0056] Wherein, M is the number of modes. In this embodiment, M=3, corresponding to vibration, oil, and cooling water, respectively.

[0057] (4) Shafting fault classification and model optimization

[0058] The fusion feature F fusion The input is used to classify the shafting fault type using a constructed multilayer perceptron classifier. The multilayer perceptron classifier uses two fully connected layers with hidden dimensions of 128 and 64, respectively. It uses the ReLU activation function and adds a batch normalization layer after each fully connected layer to accelerate training convergence and improve generalization. Finally, a softmax function is used to output the probability distribution of each fault type.

[0059] Aiming at the common class imbalance problem in ship shafting fault data, the Focal loss function is used to train the multi-layer perceptron classifier. The function is expressed as:

[0060] FL(p t )=-α t (1-p t ) γ log(p t )

[0061] where p t is the predicted probability of the classifier model for the true category; α t is the category weight, which is inversely proportional to the category frequency and is used to compensate for sample imbalance; γ is the focusing parameter, which is used to adjust the weight ratio of simple samples to difficult samples and enhance the model's learning ability for difficult-to-classify samples.

[0062] (5) Evaluation of multimodal fusion diagnostic effect

[0063] This embodiment constructs a multi-source data set based on the actual operating data of the ship propulsion shafting test bench, and collects a total of 50 hours of valid data, covering normal working conditions and multiple types of fault conditions. The fault types in the experiment include bearing inner ring fault, outer ring fault, rolling element fault, shaft bending, shaft misalignment, poor lubrication and cooling system abnormality, etc. The working condition combination covers 4 speed levels (600rpm, 900rpm, 1200rpm and 1500rpm) and 3 load levels (25%, 50% and 75%). The model training adopts a batch size of 64, a maximum iteration round of 200 (with early stopping strategy), and the optimizer uses Adam optimizer (initial learning rate 0.001), Focal loss function parameter γ = 2.0, category weight α t Dynamically calculated based on sample distribution. The dataset is divided into training set, validation set, and test set in a ratio of 7:2:1.

[0064] Table 1 shows the performance comparison of the single-modal model and the multi-modal fusion model. Although the vibration mode has achieved an accuracy of 85.72% when applied alone, the accuracy is improved by 9.15% after combining the lubricating oil and cooling water modal information, which shows that there is obvious complementarity between the three modalities and that they can capture fault feature information from different physical dimensions. There is a performance gap of 5.93% between the traditional feature splicing method (88.94%) and the multi-modal attention fusion mechanism of the present invention (94.87%), which fully verifies the superiority of the deep attention fusion mechanism. Further analysis shows that this is mainly due to the fact that simple splicing cannot effectively handle the correlation and redundant information between modalities, while the multi-modal attention fusion mechanism can adaptively capture the interaction relationship between modalities.

[0065] Table 1 Performance comparison of single-modal model and multi-modal fusion model

[0066]

[0067] Table 2 shows the diagnostic performance indicators for various fault types. The model's ability to identify different fault types varies. Normal conditions and bearing inner race faults have the highest recognition rates (97.85% and 95.42%, respectively). Rolling element faults and poor lubrication have relatively low recognition rates (93.65% and 93.85%, respectively). This performance difference is primarily due to differences in the significance of fault features and the uneven distribution of training samples.

[0068] Table 2 Various fault diagnosis performance indicators

[0069]

[0070]

[0071] Table 3 shows the diagnostic performance under sensor failure conditions. Despite the reduced data quality score caused by sensor failure, the model exhibits nonlinear performance degradation, maintaining stable performance at low failure rates. Even with 30% sensor failure, the proposed multimodal attention fusion-based diagnostic model maintains an accuracy of 88.92%. When a sensor modality fails, the model automatically increases the contribution weights of other modalities through an adaptive dynamic weight allocation mechanism, thereby achieving robust fault diagnosis.

[0072] Table 3 Diagnostic performance indicators in the case of sensor failure

[0073]

[0074] The above description is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiment. It is understood that other improvements and variations directly derived or imagined by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the scope of protection of the present invention.

Claims

1. A ship propulsion shaft fault diagnosis method based on multimodal attention fusion, characterized in that: The method comprises: Perform multi-source data collection and data preprocessing on key parts of the ship's propulsion shaft system, including vibration parameters, lubricating oil parameters, and cooling water parameters; A modal-specific feature extraction strategy is used for different modal parameters to obtain the corresponding modal features, including the multi-dimensional features of vibration parameters and the time domain features of lubricating oil and cooling water parameters; Each modal feature is fused with weighted multimodal features based on the multimodal attention fusion mechanism; The fused features are input into the multi-layer perceptron classifier to classify the shaft fault type.

2. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 1 is characterized in that: The above-mentioned various modal features are weightedly fused based on the multimodal attention fusion mechanism, including: The multimodal attention fusion mechanism achieves weighted fusion of multimodal features through the collaborative work of three components: intra-modal self-attention, inter-modal cross-attention, and adaptive dynamic weight allocation, where: The multi-dimensional features of the vibration parameters, the time domain features of the lubricating oil parameters, and the time domain features of the cooling water parameters are respectively automatically identified through an intra-modal self-attention mechanism to automatically identify key features within the modality and suppress redundant information; The key features within any two modalities are used to evaluate the correlation and complementarity between different modalities through an inter-modal cross-attention mechanism, and the contribution of each modality to fault diagnosis is quantified; The adaptive dynamic weight allocation mechanism comprehensively considers the current working condition parameters and data quality score, assigns dynamic weights to each mode and performs weighted fusion to generate the fusion feature.

3. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 2 is characterized in that: The implementation process of the intra-modality self-attention mechanism includes: For each modal feature Generate query matrix through linear transformation Bond Matrix Value Matrix The self-attention weight is obtained by the query matrix Q m and the bond matrix K m The dot product is calculated and scaled and normalized by softmax, and finally the value matrix V m Multiply to get enhanced feature Z m , as the key internal feature of the modality; in, is the learnable weight matrix, n is the number of samples, d m d k denote the input feature dimensions of the intra-modal self-attention and the projection dimensions of the query, key, and value vectors in the attention mechanism, respectively.

4. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 2 is characterized in that: The implementation process of the inter-modal cross-attention mechanism includes: For any two modal internal key features and Calculate the cross attention weight by querying the matrix and bond matrix The dot product is calculated and scaled and normalized by softmax, and finally combined with the value matrix Multiply to get the cross feature C ij ; For each modality, aggregate the cross features C from all other modalities ij , get the modal feature C after cross attention i ; in, is the learnable weight matrix, n is the number of samples, d p d q They represent the input feature dimensions of the inter-modal cross-attention and the feature dimensions of the cross-attention output, respectively.

5. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 2 is characterized in that: The implementation process of the adaptive dynamic weight allocation mechanism includes: For each mode, according to the current working condition parameters and the modal features C output by the cross-attention mechanism between the modes i And the original data quality score Q corresponding to the modality i , calculate its dynamic weight coefficient: W i =softmax(f(H op ,C i ,Q i )) Among them, f is a multi-layer perceptron network, H op is a working condition embedding vector calculated based on the current working condition parameters; For all modal features C i Perform weighted fusion to generate the fusion feature: Where M is the number of modes.

6. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 5 is characterized in that: The implementation process of the adaptive dynamic weight allocation mechanism also includes: Build a working condition perception network and data quality assessment network based on the multi-layer perceptron network MLP; The current working condition parameter vectors are spliced ​​and input into the working condition perception network to obtain the working condition embedding vector H op ; The collected multi-source data are input into the data quality assessment network to obtain the original data quality score Q of the i-th modality i .

7. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 1 is characterized in that: The data preprocessing process includes: Select the corresponding noise reduction algorithm for processing according to the characteristics of each modal data; Interpolation technology is used to align data with different sampling frequencies to a unified time base; Based on the distribution characteristics of different modal data, the corresponding standardization method is selected for normalization processing.

8. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 7 is characterized in that: Selecting a noise reduction algorithm, as well as a normalization method, includes: The vibration parameters are processed using a wavelet threshold noise reduction algorithm, and Z-score standardization is used to achieve mean removal and variance normalization; For the lubricating oil parameters and the cooling water parameters, a noise reduction algorithm combining median filtering and sliding average filtering is used to remove outliers and high-frequency noise, and Min-Max normalization is used to the interval [0, 1].

9. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 1 is characterized in that: The method of obtaining corresponding modal features by adopting a modal-specific feature extraction strategy for different modal parameters includes: For the vibration parameters, extracting time domain features, spectrum features, time-frequency features and nonlinear features; For the lubricating oil parameters and the cooling water parameters, the time domain characteristics of temperature, pressure and flow are extracted.

10. The ship propulsion shafting fault diagnosis method based on multimodal attention fusion according to claim 1, characterized in that: The method further comprises: The Focal loss function is used to train the multi-layer perceptron classifier to address the imbalance problem of shafting fault categories and improve the ability to identify rare faults.

Images