An unmanned platform electromechanical system fault early warning method based on a Transformer-BiGRU cross attention mechanism

Abstract

The application discloses a fault early warning method for an unmanned platform electromechanical system based on a Transformer-BiGRU cross attention mechanism and belongs to the technical field of fault early warning. The method comprises the following steps: acquiring parameters and performing time sequence alignment and key feature screening; a parallelized space-time feature extraction architecture is designed; on one hand, a multilayer Transformer encoder is adopted to capture long time sequence dependency; on the other hand, a bidirectional gated recurrent unit is adopted to realize bidirectional context modeling of a sequence, and an interpretable global attention mechanism is introduced to dynamically distribute feature weights of a BiGRU hidden layer state; then, a cross attention feature fusion module is constructed for the fusion problem of the Transformer and the BIGRU, interactive attention calculation of a space-time feature tensor is performed, and the collaborative representation of time sequence dynamic characteristics and spatial correlation characteristics is realized. The method realizes advanced early warning from diesel engine lubricating oil pressure abnormality to ship power system cascade failure.

Description

TECHNICAL FIELD

[0001] The application relates to a fault early warning method for an unmanned platform electromechanical system based on a Transformer-BiGRU cross attention mechanism, and belongs to the technical field of fault early warning. The algorithm is applied to fault diagnosis and prediction of unmanned platform electromechanical systems, other platform power systems and other large and complex equipment. By monitoring multi-modal sensor data in real time, a deep learning model is used to effectively extract fault features, and high-accuracy fault early warning is realized. The technology can be widely used in automated equipment, industrial control systems and other complex equipment that require high reliability and real-time prediction. BACKGROUND

[0002] With the rapid development of unmanned platforms and intelligent equipment, the operating environment of ship power systems and other complex equipment has become increasingly complex. Traditional fault diagnosis methods rely mainly on experience and rule-based models, but when dealing with large-scale, high-dimensional, multi-source and heterogeneous sensor data, these methods often encounter problems such as low precision and poor stability, especially in complex environments, which cannot accurately identify faults in real time, resulting in high false positive and false negative rates. Therefore, there is an increasing demand for real-time and accurate fault early warning technology.

[0003] In recent years, deep learning methods have made significant breakthroughs in the field of fault diagnosis, especially time series data processing models such as LSTM and GRU, which have been widely used in fault prediction of various equipment. However, existing GRU and LSTM models still have limitations in handling long-range dependencies and multi-modal data, and cannot fully capture local and global features in time series data. The Transformer model excels at capturing global dependencies with its self-attention mechanism, but is weak in handling short-term time series features. SUMMARY

[0004] To solve the problems in the prior art, the application proposes a multi-source fault early warning algorithm based on Transformer-BiGRU cross attention fusion, which can effectively process data from multi-modal and heterogeneous sensors and accurately predict system faults through a deep learning model. The algorithm combines the global dependency capturing ability of Transformer with the local time series modeling ability of BiGRU, and uses the cross attention mechanism to realize dynamic fusion of multi-source data, thereby improving the accuracy and robustness of fault prediction. By fusing global dependency features and local time series features, it becomes an effective way to improve the accuracy and reliability of the fault early warning algorithm.

[0005] The technical solution adopted by the application is: a fault early warning method for an unmanned platform electromechanical system based on a Transformer-BiGRU cross attention mechanism, comprising the following steps:

[0006] S1, heterogeneous data collection and governance analysis of multi-source sensors under complex working conditions;

[0007] S1.1 Real-time collection of multi-modal sensor data of unmanned platform electromechanical system;

[0008] S1.2 Classification of multi-modal sensor data of electromechanical system: according to the different collection positions, the data is classified and processed;

[0009] S1.3 Governance of multi-source sensor heterogeneous data of electromechanical system: outlier elimination, missing value filling, Bayesian wavelet packet denoising and normalization processing are performed on the data;

[0010] S2, extraction of operating characteristic parameters under multi-source heterogeneous data of electromechanical system:

[0011] Based on the calculation results of Pearson correlation, the parameters with Pearson correlation coefficient greater than 0.6 are selected as the characteristic parameters of the unmanned platform electromechanical system fault identification and early warning model;

[0012] S3, build an unmanned platform fault early warning model based on Transformer-BiGRU cross attention mechanism:

[0013] S3.1 Transformer encoder: it maps the input sequence into query Query, key Key and value Value vectors, and calculates the correlation weight between any positions in the sequence through dot product operation:

[0014]

[0015]

[0016] After multi-layer stacking, the final output is:

[0017]

[0018] Where Q, K and V are queries, keys and values, respectively, and dk is the dimension of the key;

[0019] According to the task allocation of the electromechanical system, the input and output feature dimensions of the model are specified, and the attention dimension of the Transformer encoder, the number of encoder layers and the number of multi-head attention heads are allocated to capture the long-term correlation characteristics across time steps;

[0020] S3.2 BiGRU network optimized by fusion of global attention mechanism: the network is composed of forward and backward neural networks, which encodes the input sequence through multi-layer bidirectional GRU, and the state update equation is:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] Among them, z t Update gate, r t To reset the door, Represents element-wise multiplication, BIGRU Output For the output of each BIGRU layer;

[0027] S3.3 Global-local feature prediction fusion analysis under cross-attention mechanism, the calculation process is as follows: the cross-attention mechanism performs linear transformation on the input parameters of electromechanical system through query, key and value, calculates attention weight and generates weighted feature representation; the output of cross-attention is reduced in dimensionality through adaptive average pooling layer, and finally the final prediction result is generated through fully connected layer;

[0028]

[0029]

[0030] Where Q is the output from the Transformer, and K and V are the outputs from the BiGRU;

[0031]

[0032]

[0033]

[0034] Among them, W f Let b be the weight matrix. f For bias;

[0035] S4. Determine the dynamic warning threshold: The formula for the dynamic threshold T is as follows:

[0036]

[0037] Where μΔr and σΔr are the mean and standard deviation of the changes in the predicted residuals of the historical early warning model, respectively, and k is an adjustment factor;

[0038] S5. Fault warning and output: After processing by the model, a fault prediction value is output. If the error between the prediction value and the actual fault occurrence exceeds the set dynamic threshold, it is considered that there is a possibility of a fault occurring in the system.

[0039] Further, in step S1.1, the collected multimodal sensor data includes diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, diesel engine temperature after cooling, DC bus voltage of motor 1, output current of motor 1, output AC voltage of motor 1, output power of motor 1, output speed of motor 1, output torque of motor 1, heat sink temperature of motor 1, temperature of motor 1, DC bus voltage of motor 2, output current of motor 2, output AC voltage of motor 2, output power of motor 2, output speed of motor 2, output torque of motor 2, heat sink temperature of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, seawater pump outlet pressure, cooling water outlet pressure, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

[0040] Furthermore, in step S1.2, the data is categorized as follows:

[0041] The first set of diesel engine monitoring parameters includes: diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, and diesel engine temperature after cooling.

[0042] The second set of monitoring parameters for motor 1 includes: DC bus voltage of motor 1, output current of motor 1, output AC voltage of motor 1, output power of motor 1, output speed of motor 1, output torque of motor 1, heat sink temperature of motor 1, and temperature of motor 1.

[0043] The third set of monitoring parameters for motor 2 includes: DC bus voltage of motor 2, output current of motor 2, output AC voltage of motor 2, output power of motor 2, output speed of motor 2, output torque of motor 2, heat sink temperature of motor 2, and temperature of motor 2.

[0044] The fourth set of ship-wide operating parameters includes: gearbox 1 temperature, gearbox 2 temperature, seawater pump outlet pressure, cooling water outlet pressure, gearbox 1 oil pressure, gearbox 2 oil pressure, total battery voltage, bus voltage, and ship speed.

[0045] Furthermore, in step S1.3, the data undergoes outlier removal, missing value imputation, Bayesian wavelet packet denoising, and normalization. The specific processing steps are as follows:

[0046] An adaptive anomaly detection model is established based on the dynamic sliding window Z-score method. The standardized score of data points is calculated using the window mean µ and standard deviation. When |Zi|>3, it is judged as an outlier and removed.

[0047] An improved KNN weighted imputation algorithm is introduced to address missing data. It selects k nearest neighbor samples by weighted Euclidean distance and generates imputation values ​​according to the weight formula.

[0048] Multi-scale Bayesian wavelet packet decomposition technology is used to perform signal denoising. The signal is decomposed into 3-5 frequency bands by wavelet packet expansion, and the coefficients are shrunk based on the maximum a posteriori probability threshold.

[0049] Multi-source heterogeneous data are mapped to the [0,1] interval using an adaptive range normalization formula.

[0050] Furthermore, in step S2, the Pearson correlation coefficient method is used to compare the correlation between the principal component parameters and other operating parameters in each group; through the data matrix X=(x) m×n To calculate x in the a-th column of the matrix a and column b x b Correlation:

[0051]

[0052] Where m is the length of each column, and the correlation coefficient ranges from 0 to 1; a value of 1 indicates perfect correlation, and a value of 0 indicates no correlation between columns.

[0053] Furthermore, the data preprocessing method evaluates the quality and accuracy of the cleaned data using information entropy, the formula for which information entropy is calculated is:

[0054]

[0055] Where H(X) represents the entropy of the random variable X, p(x) i ) is the probability of the i-th event in the sample, and n is the total number of events.

[0056] Furthermore, the Transformer encoder layer employs a multi-head self-attention mechanism (MHSA) to capture long-range dependency features by dynamically modeling global contextual dependencies of sequences.

[0057] Furthermore, when processing time-series signals, the BiGRU network performs bidirectional encoding through forward and backward neural networks, utilizing information from preceding and following time steps to capture time-series data.

[0058] Furthermore, the cross-attention mechanism calculates attention weights and generates weighted feature representations through linear transformations of queries, keys, and values.

[0059] Furthermore, the unmanned platform fault early warning model also includes an adaptive feature compression layer, which adaptively adjusts the feature dimension according to the noise level of the input features.

[0060] Specifically, the method includes the following steps:

[0061] Step 1: Acquisition and Management Analysis of Heterogeneous Data from Multi-Source Sensors under Complex Working Conditions

[0062] (1) Real-time acquisition of multi-modal sensor data of the electromechanical system of the unmanned platform, including diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, diesel engine temperature after cooling, DC bus voltage of motor 1, output current of motor 1, output AC voltage of motor 1, output power of motor 1, output speed of motor 1, output torque of motor 1, heat sink temperature of motor 1, temperature of motor 1, DC bus voltage of motor 2, output current of motor 2, output AC voltage of motor 2, output power of motor 2, output speed of motor 2, output torque of motor 2, heat sink temperature of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, seawater pump outlet pressure, cooling water outlet pressure, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

[0063] (2) Classification of multi-component sensing parameters of electromechanical system: In view of the problem of multi-component, multi-coupling and high complexity of electromechanical system, the data in (1) are classified to improve the efficiency of data processing. According to the different collection locations, the data are classified and processed. Group 1 (Diesel engine monitoring parameters): diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, and diesel engine temperature after cooling; Group 2 (Motor 1 monitoring parameters): motor 1 DC bus voltage, motor 1 output current, motor 1 output AC voltage, motor 1 output power, motor 1 output speed, motor 1 output torque, motor 1 heat sink temperature, and motor 1 temperature; Group 3 (Motor 2 monitoring parameters): motor 2 DC bus voltage, motor 2 output current, motor 2 output AC voltage, motor 2 output power, motor 2 output speed, motor 2 output torque, motor 2 heat sink temperature, and motor 2 temperature; Group 4 (ship operating parameters): gearbox 1 temperature, gearbox 2 temperature, seawater pump outlet pressure, cooling water outlet pressure, gearbox 1 oil pressure, gearbox 2 oil pressure, battery total voltage, bus voltage, and ship speed.

[0064] (3) Treatment of heterogeneous data from multiple sensor sources in electromechanical systems: Since these data come from multiple heterogeneous sensors and contain noise and missing values, the system uses a data preprocessing module to clean, standardize, and normalize the raw data to ensure the quality of the input data. The goal of this module is to reduce data noise and fill in missing values, thereby providing accurate and consistent input for subsequent deep learning models. First, an adaptive anomaly detection model is established based on the dynamic sliding window Z-score method. The standardized score of data points is calculated using the window mean µ and standard deviation. When |Zi|>3, it is identified as an outlier and removed. Then, an improved KNN weighted imputation algorithm is introduced for missing data. The k nearest neighbor samples are selected by weighted Euclidean distance, and the imputation value is generated according to the weight formula, where the parameter p controls the nearest neighbor attenuation rate. On this basis, multi-scale Bayesian wavelet packet decomposition technology is used to perform signal denoising. The signal is decomposed into 3-5 frequency bands by wavelet packet expansion, and the coefficients are shrunk based on the maximum a posteriori probability threshold to achieve noise suppression. Finally, the multi-source heterogeneous data is mapped to the [0,1] interval by the adaptive range normalization formula to eliminate dimensional differences.

[0065] Finally, information entropy is used as the evaluation metric. This method, based on sample data, can effectively determine the possible range of parameters and measure the confidence level of the estimate, thus quantitatively verifying the uncertainty of the cleaned data. A lower entropy value indicates higher data consistency and certainty. By evaluating the quality, accuracy, and reliability of the cleaned data, it can be ensured that the data cleaning steps have not introduced new problems. After outlier removal, missing value imputation, Bayesian wavelet packet denoising, and normalization, the information entropy is significantly reduced compared to the original data, indicating a significant improvement in data quality. These data preprocessing steps not only enhance the reliability of the data but also improve its consistency and interpretability. The formula for calculating information entropy is:

[0066]

[0067] Where H(X) represents the entropy of random variable X, p(xi) is the probability of the i-th event in the sample, and n is the total number of events.

[0068] Step 2: Extraction of operational characteristic parameters from multi-source heterogeneous data of electromechanical systems. This invention uses the Pearson correlation coefficient method to compare the correlation between principal component parameters and other operational parameters in each group. The correlation between column a (x) and column b (xb) in the research matrix is ​​calculated using the data matrix X = (x)m×n.

[0069]

[0070] Where m is the length of each column, and the correlation coefficient ranges from 0 to 1. A value of 1 indicates perfect correlation, and a value of 0 indicates no correlation between the columns.

[0071] According to a widely accepted definition, parameters with a Pearson correlation coefficient greater than 0.6 are considered to have strong correlation, while those with a Pearson correlation coefficient less than 0.6 are considered to have weak correlation. Therefore, based on the calculation results of Pearson correlation, parameters with a Pearson correlation coefficient greater than 0.6 are selected as the feature parameters of the fault identification and early warning model for the electromechanical system of unmanned platforms.

[0072] Step 3: Construction of the Transformer-BiGRU Early Warning Model. The core model of this invention is a combination of Transformer and BiGRU, which achieves the fusion of multimodal data through a cross-attention mechanism. Specifically, the Transformer part is used to capture long-term global dependencies, while the BiGRU part is good at handling local short-term dependencies. The two complement each other, improving the ability to model complex time-series data.

[0073] (1) Transformer Module: One of the core branches of the Transformer-BiGRU model is the Transformer encoder layer based on a multi-head self-attention mechanism, which achieves efficient and robust temporal feature extraction by dynamically modeling the global contextual dependencies of the sequence. The core of this encoder lies in the self-attention mechanism, which maps the input sequence into query, key, and value vectors, and calculates the relevance weights between any positions in the sequence through dot product operations:

[0074]

[0075]

[0076] After multiple stacking layers, the final output is:

[0077]

[0078] Where Q, K, and V represent the query, key, and value, respectively, and dk is the dimension of the key.

[0079] (2) BiGRU Module: The BiGRU (Bidirectional Gated Cyclic Unit) module is used to capture local temporal information in time-series data. Compared with the traditional GRU, BiGRU effectively captures contextual information by modeling simultaneously in both the forward and backward directions. The update formula for BiGRU is as follows:

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] Here, zt updates the door, and rt resets the door. Represents element-wise multiplication, BIGRU Output This is the output of each BIGRU layer.

[0086] Multi-feature sequence data employs a mechanism that enhances the model's attention to different parts of the input sequence through a BiGRU network optimized based on GlobalAttention. This process handles time-series signals, extracts local contextual features, and enhances the model's sensitivity to short-term changes. It demonstrates high recognition capability for pattern changes in electromechanical systems under multi-task conditions. The added global attention mechanism in the BiGRU model helps the model better focus on the most relevant parts of the input sequence, thereby improving model performance and generalization ability. Within each time step or adjacent sliding window, the global attention mechanism calculates a weight vector representing the model's attention to each part of the input sequence. These weights are then applied to the feature representation output by the BiGRU. By weighting features at all locations, the model can more effectively focus on important temporal features, improving its ability to perceive the temporal features of multi-feature sequences.

[0087] (3) Cross-attention mechanism: In the process of multimodal data fusion, the outputs of the Transformer and BiGRU modules need to be fused. To effectively fuse the outputs of these two models, this invention designs a cross-attention mechanism. This mechanism uses the output of the Transformer as the query and the output of the BiGRU as the key and value, and uses self-attention calculation to weight and fuse information from different data sources. The formula for calculating cross-attention is:

[0088]

[0089] Where Q, K, and V represent the query, key, and value, respectively, and d k The dimension of the key.

[0090] Step 4: Dynamic Determination of Warning Threshold: In this invention, the threshold is defined using the average value plus or minus a certain number of standard deviations, i.e., the threshold range is set to [0, μ+kσ], where k is a coefficient determined based on actual needs and experience to control the sensitivity of anomaly detection. The advantage of this method is that it can adaptively adjust the threshold based on the changing characteristics of historical data, thereby improving the accuracy and reliability of anomaly detection. The formula for the dynamic threshold T is as follows.

[0091]

[0092] Where μΔr and σΔr are the mean and standard deviation of the historical correlation changes, respectively, and k is the adjustment factor.

[0093] Step 5: Fault Warning and Output. After processing by the above model, the algorithm outputs a fault prediction value, which represents the probability of the system's operating condition in the future. The algorithm optimizes the model parameters by calculating the error between the predicted value and the actual fault occurrence. When the deviation between the optimized parameters and the actual value is large, exceeding the threshold set in Step 4, it can be considered that the system has the possibility of a fault occurring.

[0094] The beneficial effects of this invention are as follows: High-precision anomaly state prediction is achieved through multimodal data fusion and spatiotemporal feature collaborative extraction. This includes acquiring time-series electromechanical system operating parameters to construct a multi-source heterogeneous data preprocessing framework, and using sliding window partitioning and Pearson correlation coefficient matrix to perform temporal alignment and key feature screening of the electromechanical system operating parameters. Secondly, a parallelized spatiotemporal feature extraction architecture is designed: on the one hand, a multi-layer Transformer encoder captures long-term temporal dependencies; on the other hand, a bidirectional gated recurrent unit (BiGRU) is used to achieve bidirectional contextual modeling of the sequence, and an interpretable global attention mechanism is introduced to dynamically allocate feature weights to the hidden state of the BiGRU. Furthermore, a cross-attention feature fusion module is constructed to address the Transformer and BIGRU fusion problem, achieving collaborative representation of temporal dynamic characteristics and spatial correlation characteristics through interactive attention calculation of spatiotemporal feature tensors. To verify the effectiveness of the algorithm, a training set was constructed using health monitoring data of the entire life cycle of the electromechanical system, and a verification dataset was constructed based on typical fault modes. An adaptive threshold dynamic optimization strategy based on the distribution of prediction error was proposed, which realized the early warning of cascading faults from abnormal diesel engine lubricating oil pressure to the entire ship's power system. A cascading fault verification chain of "abnormal lubricating oil pressure - bearing wear - generator failure - power system collapse" was established through fault propagation path reverse analysis technology. Attached Figure Description

[0095] Figure 1This is a flowchart of the present invention.

[0096] Figure 2 Distribution diagram of monitoring sensors for electromechanical systems.

[0097] Figure 3 This is a box plot showing outlier displays in an application example of the present invention.

[0098] Figure 4 This is a comparison of information entropy after data governance in an application example of the present invention.

[0099] Figure 5 This is a heatmap showing the correlation of runtime data in an application example of the present invention.

[0100] Figure 6 This example compares the healthy and abnormal data of the diesel engine and auxiliary subsystems.

[0101] Figure 7 This is the early warning residual result curve of the present invention. Detailed Implementation

[0102] Taking a 15-ton unmanned surface vessel as an example, it is equipped with multiple sensors to monitor 30 parameters, including diesel engine speed, motor output power, and sailing speed. This example is used to demonstrate the specific real-time method of the present invention.

[0103] Step 1: Acquisition and Management Analysis of Heterogeneous Data from Multi-Source Sensors under Complex Working Conditions

[0104] (1) such as Figure 2 The real-time data collected from the multi-modal sensors of the unmanned platform's power system, as shown in Table 1, includes diesel engine oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, diesel engine temperature after cooling, DC bus voltage of motor 1, output current of motor 1, output AC voltage of motor 1, output power of motor 1, output speed of motor 1, output torque of motor 1, heat sink temperature of motor 1, temperature of motor 1, DC bus voltage of motor 2, output current of motor 2, output AC voltage of motor 2, output power of motor 2, output speed of motor 2, output torque of motor 2, heat sink temperature of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, seawater pump outlet pressure, cooling water outlet pressure, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

[0105] Table 1. Monitoring Data Measurement Points for Unmanned Surface Vessels

[0106]

[0107] (2) Classification of multi-component sensing parameters of electromechanical system: In view of the problem of multi-component, multi-coupling and high complexity of electromechanical system, the data in (1) are classified to improve the efficiency of data processing. According to the different collection locations, the data are classified and processed. Group 1 (Diesel engine monitoring parameters): diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, and diesel engine temperature after cooling; Group 2 (Motor 1 monitoring parameters): motor 1 DC bus voltage, motor 1 output current, motor 1 output AC voltage, motor 1 output power, motor 1 output speed, motor 1 output torque, motor 1 heat sink temperature, and motor 1 temperature; Group 3 (Motor 2 monitoring parameters): motor 2 DC bus voltage, motor 2 output current, motor 2 output AC voltage, motor 2 output power, motor 2 output speed, motor 2 output torque, motor 2 heat sink temperature, and motor 2 temperature; Group 4 (ship operating parameters): gearbox 1 temperature, gearbox 2 temperature, seawater pump outlet pressure, cooling water outlet pressure, gearbox 1 oil pressure, gearbox 2 oil pressure, battery total voltage, bus voltage, and ship speed.

[0108] (3) Treatment of Heterogeneous Data from Multi-Source Sensors in Electromechanical Systems: Since these data come from multi-source heterogeneous sensors and contain noise and missing values, the system uses a data preprocessing module to clean, standardize, and normalize the raw data to ensure the quality of the input data. The goal of this module is to reduce data noise and fill in missing values, thereby providing accurate and uniform input for subsequent deep learning models. First, an adaptive anomaly detection model is established based on the dynamic sliding window Z-score method. The standardized score of the data points is calculated using the window mean µ and standard deviation. When |Zi|>3, it is judged as an outlier and removed. The box plot of outlier processing is shown below. Figure 3 As shown; then, an improved KNN weighted interpolation algorithm is introduced to address the missing data. The k nearest neighbor samples are selected by weighted Euclidean distance, and the imputation value is generated according to the weight formula, where the parameter p controls the nearest neighbor attenuation rate; on this basis, multi-scale Bayesian wavelet packet decomposition technology is used to implement signal denoising. The signal is decomposed into 3-5 frequency bands by wavelet packet expansion, and the coefficients are shrunk based on the maximum a posteriori probability threshold to achieve noise suppression; finally, the multi-source heterogeneous data is mapped to the [0,1] interval by the adaptive range normalization formula to eliminate dimensional differences.

[0109] Finally, information entropy is used as the evaluation metric. This method, based on sample data, can effectively determine the possible range of parameters and measure the confidence level of the estimate, thereby quantitatively verifying the uncertainty of the cleaned data. The lower the entropy value, the higher the consistency and certainty of the data. By evaluating the quality, accuracy, and reliability of the cleaned data, it can be ensured that the data cleaning process has not introduced new problems. After outlier removal, missing value imputation, Bayesian wavelet packet denoising, and normalization, as shown...Figure 4 As shown, compared to the original data, the information entropy is significantly reduced, indicating a significant improvement in data quality. These data preprocessing steps not only enhance the reliability of the data but also improve its consistency and interpretability. The formula for calculating information entropy is:

[0110]

[0111] Where H(X) represents the entropy of random variable X, p(xi) is the probability of the i-th event in the sample, and n is the total number of events.

[0112] Finally, information entropy is used as the evaluation metric. Based on sample data, this method effectively determines the possible range of parameters and measures the confidence level of the estimates, thereby quantitatively verifying the uncertainty of the cleaned data. A lower entropy value indicates higher data consistency and certainty. By evaluating the quality, accuracy, and reliability of the cleaned data, it can be ensured that the data cleaning process has not introduced new problems. Figure 3 As shown, after outlier removal, missing value imputation, Bayesian wavelet packet denoising, and normalization, the information entropy is significantly reduced compared to the original data, indicating a significant improvement in data quality. These data preprocessing steps not only enhance data reliability but also improve data consistency and interpretability. The formula for calculating information entropy is:

[0113]

[0114] Where H(X) represents the entropy of random variable X, p(xi) is the probability of the i-th event in the sample, and n is the total number of events.

[0115] Step 2: Fault feature parameter extraction. This invention uses the Pearson correlation coefficient method to compare the correlation between principal component parameters and other operating parameters in each group, through the data matrix X=(x) m×n To calculate x in the a-th column of the research matrix a and column b x b Correlation:

[0116]

[0117] Where m is the length of each column, and the correlation coefficient ranges from 0 to 1. A value of 1 indicates perfect correlation, and a value of 0 indicates no correlation between the columns.

[0118] The calculation is performed using the Pearson correlation algorithm, such as Figure 5As shown in Table 2, the Pearson correlation coefficients for each parameter were obtained. By comparing and analyzing the magnitude of the Pearson correlation coefficients, parameters significantly affected by unmanned surface vessel malfunctions can be identified. According to a widely accepted definition, parameters with a Pearson correlation coefficient greater than 0.6 are considered to have a strong correlation, while parameters with a Pearson correlation coefficient less than 0.6 are considered to have a weak correlation. Therefore, based on the calculation results of the Pearson correlation, as shown in Table 2, parameters with a Pearson correlation coefficient greater than 0.6 are selected as the feature parameters of the unmanned platform electromechanical system fault identification and early warning model. Specifically, the following parameters are selected for analysis: diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, and diesel engine cooling temperature; motor 1 output current, motor 1 output AC voltage, motor 1 output power, motor 1 output speed, and motor 1 output torque; motor 2 output current, motor 2 output AC voltage, motor 2 output power, motor 2 output speed, and motor 2 output torque; and battery total voltage, bus voltage, and ship speed are selected as the overall ship operation analysis parameters.

[0119] Table 2. Pearson Correlation Analysis and Feature Selection Table for Unmanned Platform Parameters

[0120]

[0121] Step 3: Construct an early warning model for unmanned surface vessels. Using 80% of the health data, an early warning model for unmanned surface vessels based on the Transformer-BiGRU GlobalAttention-CrossAttention deep learning algorithm was constructed. The remaining 20% ​​of the health data was used as test data to verify the model's predictive ability. Several basic time-series prediction models were selected, including LSTM (Long Short-Term Memory), GRU (Gated Recurrent Neural Network), BILSTM (Bidirectional Long Short-Term Memory), BIGRU (Bidirectional Gated Recurrent Neural Network), and CNN-BIGRU (Convolutional Neural Network-Bidirectional Gated Recurrent Neural Network), to comprehensively evaluate the performance of different models in time-series data prediction tasks. The learning rate of GRU, LSTM, BIGRU, BILSTM, and CNN-BIGRU networks was set to 0.0015. GRU and LSTM were single-layer models, while BIGRU and BILSTM were bidirectional models. The CNN-BIGRU network structure consisted of a single CNN layer and a single BiGRU layer, with a convolutional kernel size of 3x1. Multiple evaluation metrics were calculated for model analysis, including the coefficient of determination (R²), mean absolute percentage error (MAPE), absolute error (MAE), symmetric mean absolute percentage error (SMAPE), median absolute error (MedAE), and explained variance (EVS).

[0122] Table 3 Statistical Table of Prediction Accuracy Indicators for Each Model

[0123]

[0124]

[0125] Table 3 details the specific values ​​for each indicator. The calculation and analysis of these evaluation indicators provide an important basis for a deeper understanding of the model's performance and applicability. The calculation formulas for each evaluation indicator are as follows.

[0126] Coefficient of determination (R²):

[0127]

[0128] Mean Absolute Percentage Error (MAPE):

[0129]

[0130] Relative absolute error (MAE):

[0131]

[0132] Symmetric Mean Absolute Percentage Error (SMAPE):

[0133]

[0134] Median Absolute Error (MedAE):

[0135]

[0136] Explanation of variance (EVS):

[0137]

[0138] Where yi is the actual value. For predicted values, The mean of the actual values ​​is given by , and Var represents the variance.

[0139] The prediction model, trained and tested using multiple sets of health status data, uses a real-world test of an unmanned surface vessel. The test involved a diesel engine driving motor 2 to generate electricity and power the vessel, while motor 1 self-propelled the vessel during low-speed cruising. In this case, the diesel engine stalled due to a control component malfunction, causing the motor to shut down suddenly. This case serves as a fault warning analysis. The comparison between healthy and abnormal states in this data segment is shown below. Figure 6 As shown.

[0140] Step 4: Dynamic Determination of Warning Threshold: In this invention, the threshold is defined using the average value plus or minus a certain number of standard deviations, i.e., the threshold range is set to [0, μ+kσ], where k is a coefficient determined based on actual needs and experience to control the sensitivity of anomaly detection. The advantage of this method is that it can adaptively adjust the threshold based on the changing characteristics of historical data, thereby improving the accuracy and reliability of anomaly detection. The formula for the dynamic threshold T is as follows.

[0141]

[0142] Where μΔr and σΔr are the mean and standard deviation of the historical correlation changes, respectively, and k is an adjustment factor. Based on multiple tests, k=5 is set, therefore MAE is 0.2 as the fault warning threshold.

[0143] Step 5: Based on the analysis results of the early warning model, the diesel engine lubricating oil pressure first showed abnormal fluctuations, which then triggered a series of related abnormal phenomena. Due to insufficient lubricating oil supply, the diesel engine speed and output power subsequently showed significant abnormal changes. Furthermore, since the shaft-driven motor participated in the power generation process, these abnormal fluctuations disrupted the voltage stability of the entire ship's electrical system, resulting in voltage anomalies. Figure 7 As shown, the early warning model first detected the abnormal diesel engine lubricating oil pressure at 370 seconds, which was 300 seconds earlier than the voltage disorder of the entire ship's electrical system. This provided a sufficient time window for the system's emergency response. At the same time, the fault affected the diesel engine cooling temperature, but the fluctuation exceeded the overall early warning threshold, so no alarm was triggered. The ship speed began to decrease due to the fault, and the trend was not much different from the normal decrease time. Therefore, no early warning was issued for the ship speed parameter.

[0144] The above embodiments are only used to illustrate the present invention. Any equivalent transformations and improvements made on the basis of the technical solutions of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. A fault early warning method for electromechanical systems of unmanned platforms based on the Transformer-BiGRU cross-attention mechanism, characterized in that, Includes the following steps: S1. Acquisition, management and analysis of heterogeneous data from multi-source sensors under complex working conditions; S1.1 Real-time acquisition of multi-modal sensor data from the electromechanical system of the unmanned platform; S1.2 Classify the multimodal sensor data of the electromechanical system: classify and process the data according to the different acquisition locations; S1.3 Heterogeneous data management of multi-source sensors in electromechanical systems: outlier removal, missing value imputation, Bayesian wavelet packet denoising, and normalization processing of the data; S2. Extraction of operational characteristic parameters from multi-source heterogeneous data of electromechanical systems: Based on the calculation results of Pearson correlation, parameters with a Pearson correlation coefficient greater than 0.6 are selected as feature parameters of the unmanned platform electromechanical system fault identification and early warning model. S3. Construct a fault early warning model for unmanned platforms based on the Transformer-BiGRU cross-attention mechanism: The S3.1 Transformer encoder maps the input sequence to a vector of query, key, and value, and calculates the relevance weights between any positions within the sequence using dot product operations. ； ； After multiple stacking layers, the final output is: ； Where Q, K, and V represent the query, key, and value, respectively, and dk is the dimension of the key; Based on the task allocation of the electromechanical system, parameters are specified for the input and output feature dimensions of the model. At the same time, attention dimensions, number of encoder layers, and number of multi-head attention heads of the Transformer encoder are allocated to capture long-term correlation characteristics across time steps. S3.2 BiGRU Network Optimized by Global Attention Mechanism: This network consists of forward and backward neural networks, encoding the input sequence through a multi-layer bidirectional GRU. The state update equation is: ； ； ； ； ； Among them, z t Update gate, r t To reset the door, Represents element-wise multiplication, BIGRU Output For the output of each BIGRU layer; S3.3 Global-local feature prediction fusion analysis under cross-attention mechanism, the calculation process is as follows: the cross-attention mechanism performs linear transformation on the input parameters of electromechanical system through query, key and value, calculates attention weight and generates weighted feature representation; the output of cross-attention is reduced in dimensionality through adaptive average pooling layer, and finally the final prediction result is generated through fully connected layer; ； ； Where Q is the output from the Transformer, and K and V are the outputs from the BiGRU; ； ； ； Among them, W f Let b be the weight matrix. f For bias; S4. Determine the dynamic warning threshold: The formula for the dynamic threshold T is as follows: ； Where μΔr and σΔr are the mean and standard deviation of the changes in the predicted residuals of the historical early warning model, respectively, and k is an adjustment factor; S5. Fault warning and output: After processing by the model, a fault prediction value is output. If the error between the prediction value and the actual fault occurrence exceeds the set dynamic threshold, it is considered that there is a possibility of a fault occurring in the system.

2. The method for early warning of faults in the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that, In step S1.1, the multimodal sensor data includes diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, diesel engine temperature after cooling, DC bus voltage of motor 1, output current of motor 1, output AC voltage of motor 1, output power of motor 1, output speed of motor 1, output torque of motor 1, heat sink temperature of motor 1, temperature of motor 1, DC bus voltage of motor 2, output current of motor 2, output AC voltage of motor 2, output power of motor 2, output speed of motor 2, output torque of motor 2, heat sink temperature of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, seawater pump outlet pressure, cooling water outlet pressure, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

3. The method for early warning of faults in the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 2, characterized in that, In step S1.2, the data is categorized as follows: The first set of diesel engine monitoring parameters includes: diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, and diesel engine temperature after cooling. The second set of monitoring parameters for motor 1 includes: DC bus voltage of motor 1, output current of motor 1, output AC voltage of motor 1, output power of motor 1, output speed of motor 1, output torque of motor 1, heat sink temperature of motor 1, and temperature of motor 1. The third set of monitoring parameters for motor 2 includes: DC bus voltage of motor 2, output current of motor 2, output AC voltage of motor 2, output power of motor 2, output speed of motor 2, output torque of motor 2, heat sink temperature of motor 2, and temperature of motor 2. The fourth set of ship-wide operating parameters includes: gearbox 1 temperature, gearbox 2 temperature, seawater pump outlet pressure, cooling water outlet pressure, gearbox 1 oil pressure, gearbox 2 oil pressure, total battery voltage, bus voltage, and ship speed.

4. The method for early warning of faults in the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that, In step S1.3, the data undergoes outlier removal, missing value imputation, Bayesian wavelet packet denoising, and normalization. The specific processing steps are as follows: An adaptive anomaly detection model is established based on the dynamic sliding window Z-score method. The standardized score of data points is calculated using the window mean µ and standard deviation. When |Zi|>3, it is judged as an outlier and removed. An improved KNN weighted imputation algorithm is introduced to address missing data. It selects k nearest neighbor samples by weighted Euclidean distance and generates imputation values ​​according to the weight formula. Multi-scale Bayesian wavelet packet decomposition technology is used to perform signal denoising. The signal is decomposed into 3-5 frequency bands by wavelet packet expansion, and the coefficients are shrunk based on the maximum a posteriori probability threshold. Multi-source heterogeneous data are mapped to the [0,1] interval using an adaptive range normalization formula.

5. The method for early warning of electromechanical system faults in an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: In step S1.3, after processing the heterogeneous data from multiple sensors, the quality and accuracy of the cleaned data are evaluated using information entropy. The formula for calculating information entropy is: ； Where H(X) represents the entropy of the random variable X, p(x) i ) is the probability of the i-th event in the sample, and n is the total number of events.

6. The method for early warning of faults in the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: In step S2, the Pearson correlation coefficient method is used to compare the correlation between the principal component parameters and other operating parameters in each group; this is done through the data matrix X=(x). m×n To calculate x in the a-th column of the matrix a and column b x b Correlation: ； Where m is the length of each column, and the correlation coefficient ranges from 0 to 1; a value of 1 indicates perfect correlation, and a value of 0 indicates no correlation between columns.

7. The method for early warning of electromechanical system faults in an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: The Transformer encoder employs a multi-head self-attention mechanism (MHSA) to capture long-range dependency features by dynamically modeling global contextual dependencies of sequences.

8. The method for early warning of faults in the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: When processing time-series signals, the BiGRU network performs bidirectional encoding through forward and backward neural networks, utilizing information from previous and subsequent time steps to capture time-series data.

9. A fault early warning method for an unmanned platform electromechanical system based on a Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: The cross-attention mechanism calculates attention weights and generates weighted feature representations through linear transformations of queries, keys, and values.

10. A fault early warning method for an unmanned platform electromechanical system based on a Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: The unmanned platform fault early warning model also includes an adaptive feature compression layer, which adaptively adjusts the feature dimension according to the noise level of the input features.

Images