Unmanned platform electromechanical system fault early warning method based on Transform-BiGRU cross attention mechanism

Abstract

The invention discloses an unmanned platform electromechanical system fault early warning method based on a Transform-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; secondly, designing a parallelization spatio-temporal feature extraction framework: on one hand, capturing a long time sequence dependency relationship through a multi-layer Transform encoder, on the other hand, realizing bidirectional context modeling of a sequence by adopting a bidirectional gating circulation unit, and introducing an interpretable global attention mechanism to perform feature weight dynamic allocation on a BiGRU hidden layer state; and a cross attention feature fusion module is constructed for a Transform and BIGRU fusion problem, and cooperative characterization of a time sequence dynamic feature and a space correlation feature is realized through interactive attention calculation of a spatial-temporal feature tensor. By means of the method, advanced early warning from diesel engine lubricating oil pressure abnormity to whole ship electric power system cascade faults is achieved.

Description

Technical Field

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

[0002] With the rapid development of unmanned platforms and intelligent devices, the operating environment of ship power systems and other complex equipment has become increasingly complex. Traditional fault diagnosis methods mostly rely 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 accuracy and poor stability, especially in complex environments where faults cannot be accurately identified in real time, resulting in high false alarm and missed alarm rates. Therefore, the demand for real-time and accurate fault warning technology is increasing.

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

[0004] To solve the problems existing in the prior art, the present invention proposes a multi-source fault 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. This algorithm combines the global dependence capture ability of Transformer with the local time series modeling ability of BiGRU, and uses the cross-attention mechanism to realize the dynamic fusion of multi-source data, thereby improving the accuracy and robustness of fault prediction. By fusing global dependence features and local time series features, it becomes an effective way to improve the accuracy and reliability of the fault warning algorithm.

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

[0006] S1. Acquisition and governance analysis of heterogeneous data from multi-source sensors under complex working conditions;

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

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

[0009] S1.3 Governance of heterogeneous data from multi-source sensors of the electromechanical system: Outlier removal, 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 the electromechanical system:

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

[0012] S3. Construct a fault early warning model for the unmanned platform based on the cross-attention mechanism of Transformer-BiGRU:

[0013] S3.1 Transformer encoder: It maps the input sequence into query Query, key Key, and value Value vectors, and calculates the correlation weights between any positions within the sequence through dot product operations:

[0014]

[0015]

[0016] After multiple layers of stacking, the final output is:

[0017]

[0018] Among them, Q, K, and V are the query, key, and value respectively, and dk is the dimension of the key;

[0019] According to the task assignment situation of the electromechanical system, parameter regulations are made for the input and output feature dimensions of the model, and at the same time, the attention dimension, encoder layer number, and number of multi-head attention heads of the Transformer encoder are allocated to capture the long-term correlation characteristics across time steps;

[0020] S3.2 BiGRU network optimized by fusing the global attention mechanism: This network consists of a forward and a backward neural network, and encodes the input sequence through multiple layers of bidirectional GRUs. The state update equation is:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] where z t is the update gate, r t is the reset gate, denotes element-wise multiplication, BIGRU Output is the output of BIGRU for each layer;

[0027] S3.3 Global-local feature prediction fusion analysis under the cross-attention mechanism, and its calculation process is as follows: The cross-attention mechanism performs linear transformation on the input parameters of the electromechanical system through queries, keys, and values, calculates the attention weights, and generates weighted feature representations; the output of the cross-attention is dimension-reduced through an adaptive average pooling layer, and finally the final prediction result is generated through a 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] where W f is the weight matrix, b f is the 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 change in the historical warning model prediction residuals respectively, and k is the adjustment factor;

[0038] S5. Fault warning and output: After being processed 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 the step S1.1, the collected multi-modal sensor data includes diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, temperature after cooling in the diesel engine, 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, temperature of the heat dissipation plate 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, temperature of the heat dissipation plate of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, outlet pressure of the seawater pump, outlet pressure of the cooling water, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

[0040] Further, in the step S1.2, the data is classified as follows:

[0041] The first group of diesel engine monitoring parameters: diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, temperature after cooling in the diesel engine;

[0042] The second group of motor 1 monitoring parameters: 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, temperature of the heat dissipation plate of motor 1, temperature of motor 1;

[0043] The third group of motor 2 monitoring parameters: 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, temperature of the heat dissipation plate of motor 2, temperature of motor 2;

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

[0045] Further, in the step S1.3, the data is subjected to outlier rejection, missing value filling, Bayesian wavelet packet denoising, and normalization processing. The specific processing process is as follows:

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

[0047] An improved KNN weighted imputation algorithm is introduced for missing data. k nearest neighbor samples are selected through weighted Euclidean distance, and the filling value is generated according to the weight formula.

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

[0049] The multi-source heterogeneous data is mapped to the [0,1] interval through the adaptive range normalization formula.

[0050] Further, in step S2, the Pearson correlation coefficient method is adopted 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 the correlation between the a-th column x a and the b-th column x b in the matrix:

[0051]

[0052] where m is the length of each column, and the value range of the correlation coefficient is from 0 to 1; the value 1 indicates complete correlation, and the value 0 indicates no correlation between columns.

[0053] Further, the data preprocessing method evaluates the quality and accuracy of the cleaned data through information entropy. The calculation formula of information entropy 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] Further, the Transformer encoder layer adopts the multi-head self-attention mechanism MHSA to capture long-range dependence features by dynamically modeling the global context dependence of the sequence.

[0057] Further, when the BiGRU network processes time series signals, it performs bidirectional encoding through forward and backward neural networks, and uses the information from the previous and subsequent time steps to capture time series data.

[0058] Further, the cross-attention mechanism calculates the attention weight and generates a weighted feature representation through the linear transformation of query, key, and value.

[0059] Further, the unmanned platform fault 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 governance analysis of heterogeneous data from multi-source sensors under complex working conditions

[0062] (1) Real-time acquisition of multimodal 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, temperature after cooling in the diesel engine, 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, temperature of the heat dissipation plate 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, temperature of the heat dissipation plate of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, outlet pressure of seawater pump, outlet pressure of cooling water, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

[0063] (2) Classification of sensing parameters of multiple components of the electromechanical system: Aiming at the problems of multiple components, multiple couplings, and high complexity of the electromechanical system, the data in (1) is classified to improve the efficiency of data processing. According to different acquisition positions, the data is classified and processed. The first group (diesel engine monitoring parameters): diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, temperature after cooling in the diesel engine; The second group (monitoring parameters of motor 1): 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, temperature of the heat dissipation plate of motor 1, temperature of motor 1; The third group (monitoring parameters of motor 2): 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, temperature of the heat dissipation plate of motor 2, temperature of motor 2; The fourth group (operating parameters of the whole ship): temperature of gearbox 1, temperature of gearbox 2, outlet pressure of seawater pump, outlet pressure of cooling water, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

[0064] (3)Multi-source Sensor Heterogeneous Data Governance for Electromechanical Systems: Since this data comes from multi-source heterogeneous sensors and there are noise and missing values, the system uses a data preprocessing module to clean, standardize, and normalize the original data to ensure the quality of the input data. The goal of this module is to reduce data noise and fill in missing values, thus providing accurate and unified 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 a data point is calculated using the window mean µ and standard deviation. When |Zi|>3, it is determined as an outlier and removed. Subsequently, an improved KNN weighted imputation algorithm is introduced for missing data. k nearest neighbor samples are selected through weighted Euclidean distance, and the filling value is generated according to the weight formula, where the parameter p controls the near neighbor attenuation rate. On this basis, a multi-scale Bayesian wavelet packet decomposition technique is used to implement signal denoising. The signal is decomposed into 3-5 layer frequency bands through the wavelet packet expansion formula, 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 through the adaptive range normalization formula to eliminate the dimension difference.

[0065] Finally, information entropy is used as an evaluation index. This method is based on sample data and can effectively determine the possible range of parameters and measure the confidence 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 no new problems are introduced in the data cleaning step. After removing outliers, filling in missing values, Bayesian wavelet packet denoising, and normalization processing of the data, compared with the original data, the information entropy is significantly reduced, indicating that the data quality has been significantly improved. These data preprocessing steps not only enhance the reliability of the data but also improve the consistency and interpretability of the data. The calculation formula of information entropy is:

[0066]

[0067] Among them, H(X) represents the entropy of the 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 Operating Characteristic Parameters under Multi-source Heterogeneous Data of Electromechanical Systems. The present invention uses the Pearson correlation coefficient method to compare the correlation between the principal component parameters and other operating parameters in each group. Through the data matrix X=(x)m×n, the correlation between the a-th column xa and the b-th column xb in the research matrix is calculated:

[0069]

[0070] Among them, m is the length of each column, and the value range of the correlation coefficient is from 0 to 1. The value 1 represents complete correlation, and the value 0 represents no correlation between columns.

[0071] According to a widely recognized definition, parameters with a Pearson coefficient greater than 0.6 are considered to have a strong correlation, while parameters with a Pearson coefficient less than 0.6 are considered to have a 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 characteristic parameters of the fault identification and early warning model for the electro-mechanical system of the unmanned platform.

[0072] Step 3: Construction of the Transformer-BiGRU early warning model. The core model of the present invention is a combination of Transformer and BiGRU, which realizes the fusion of multi-modal data through the cross-attention mechanism. Specifically, the Transformer part is used to capture the long-term global dependence relationship, while the BiGRU part is good at dealing with local short-term dependence relationships. The two complement each other and improve the modeling ability for 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 the multi-head self-attention mechanism, which realizes efficient and robust time-series feature extraction by dynamically modeling the global context dependence relationship 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 correlation weights between any positions within the sequence through dot product operations:

[0074]

[0075]

[0076] After being stacked in multiple layers, the final output is:

[0077]

[0078] Among them, Q, K, and V are the query, key, and value respectively, and dk is the dimension of the key.

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

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] Among them, zt is the update gate, and rt is the reset gate. denotes element-wise multiplication, BIGRU Output is the output of each layer of BIGRU.

[0086] The multi-feature sequence data passes through a BiGRU network optimized based on GlobalAttention to enhance the model's attention to different parts of the input sequence. It processes time series signals, extracts local context features, and enhances the model's sensitivity to short-term changes. For the pattern changes of electromechanical systems under multiple tasks, it has a high recognition ability. In the BiGRU model, the added global attention mechanism can help the model better focus on the most relevant parts of the input sequence, thereby improving the model's performance and generalization ability. At 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, and then applies these weights to the feature representation output by BiGRU. By weighting the features at all positions, the model can more specifically focus on important time-domain features, improving the model's perception ability of the time-domain features of multi-feature sequences.

[0087] (3) Cross-attention mechanism: During the multi-modal data fusion process, the outputs of the Transformer and BiGRU modules need to be fused. To effectively fuse the outputs of these two models, the present invention designs a cross-attention mechanism. This mechanism takes the output of the Transformer as the query (Query), and the output of the BiGRU as the key (Key) and value (Value), and weights and fuses the information from different data sources through self-attention calculation. The calculation formula for cross-attention is:

[0088]

[0089] Among them, Q, K, and V are the query, key, and value respectively, and d k is the dimension of the key.

[0090] Step 4: Dynamic determination of early warning threshold: In the present invention, the mean value plus or minus several times the standard deviation is used to define the threshold, that is, the threshold range is set as [0, μ + kσ], where k is a coefficient determined according to actual requirements and experience, and is used to control the sensitivity of anomaly detection. The advantage of this method is that it can adaptively adjust the threshold according to the change 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 value and standard deviation of the historical correlation change amount respectively, and k is the adjustment factor.

[0093] Step 5: Fault early warning and output. After being processed by the above model, the algorithm outputs a fault prediction value, which represents the possibility of the system's operating condition in the future period of time. The algorithm optimizes the model parameters by calculating the error between the prediction value and the actual fault occurrence situation. When the optimized parameters deviate greatly from the actual value and exceed the threshold set in Step 4, it can be considered that there is a possibility of a fault occurring in the system.

[0094] The beneficial effects of the present invention are as follows: High-precision abnormal state prediction is realized through multi-modal data fusion and spatio-temporal feature collaborative extraction, including obtaining the operating parameters of the electromechanical system with time sequence to construct a multi-source heterogeneous data preprocessing framework, and using a sliding window division and Pearson correlation coefficient matrix to perform time sequence alignment and key feature screening on the operating parameters of the electromechanical system; Secondly, a parallel spatio-temporal feature extraction architecture is designed: on the one hand, a multi-layer Transformer encoder is used to capture long time sequence dependencies, and on the other hand, a bidirectional gated recurrent unit (BiGRU) is used to realize bidirectional context modeling of the sequence, and an interpretable global attention mechanism (Global Attention) is introduced to dynamically allocate feature weights to the hidden layer state of BiGRU; Furthermore, a cross-attention feature fusion module (Cross-Attention Fusion Module) is constructed for the fusion problem of Transformer and BIGRU, and through the interactive attention calculation of spatio-temporal feature tensors, the collaborative representation of time sequence dynamic characteristics and spatial correlation characteristics is realized. To verify the effectiveness of the algorithm, a training set is constructed using the health monitoring data of the entire life cycle of the electromechanical system, a verification data set is constructed based on typical fault modes, and an adaptive threshold dynamic optimization strategy based on the prediction error distribution is proposed, realizing the early warning of the cascade fault from the abnormal lubricating oil pressure of the diesel engine to the ship's power system. A cascade fault verification chain of "abnormal lubricating oil pressure - bearing wear - generator set failure - power system collapse" is established through the reverse analysis technology of the fault propagation path. Description of the Drawings

[0095] Figure 1This is the flow chart of the present invention.

[0096] Figure 2 Distribution map of electromechanical system monitoring sensors.

[0097] Figure 3 This is the box plot of outliers in the application example of the present invention.

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

[0099] Figure 5 This is the heat map of the correlation of operation data in the application example of the present invention.

[0100] Figure 6 This is the comparison of data health and abnormal data of the diesel engine and auxiliary subsystems in the example.

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

[0102] Taking a 15t-class unmanned surface vessel as an example, it is equipped with multiple sensors for monitoring 30 parameters including diesel engine speed, motor output power, navigation speed, etc., and taking this as an example to demonstrate the specific real-time method of the present invention.

[0103] Step 1: Acquisition and governance analysis of heterogeneous data from multi-source sensors under complex working conditions

[0104] (1) As Figure 2 shown, the multi-modal sensor data of the unmanned platform power system is collected in real time. The collected data shown in Table 1 includes diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, temperature after cooling in the diesel engine, 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, temperature of the heat dissipation plate 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, temperature of the heat dissipation plate of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, outlet pressure of seawater pump, outlet pressure of cooling water, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

[0105] Table 1 Measuring point table of monitoring data for unmanned surface vessels

[0106]

[0107] (2)Classification of multi-component sensing parameters of the electro-mechanical system: Aiming at the problems of multi-components, multi-couplings and high complexity in the electro-mechanical system, the data in (1) are classified to improve the efficiency of data processing. According to the different acquisition positions, the data are classified and processed. The first group (diesel engine monitoring parameters): diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, temperature after cooling in the diesel engine; The second group (monitoring parameters of motor 1): 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, temperature of the heat sink of motor 1, temperature of motor 1; The third group (monitoring parameters of motor 2): 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, temperature of the heat sink of motor 2, temperature of motor 2; The fourth group (parameters of the whole ship operation): temperature of gearbox 1, temperature of gearbox 2, outlet pressure of seawater pump, outlet pressure of cooling water, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, ship speed.

[0108] (3)Governance of heterogeneous data from multi-source sensors in the electro-mechanical system: Since this data comes from multi-source heterogeneous sensors and there are noise and missing values, the system uses a data preprocessing module to clean, standardize and normalize the original data to ensure the quality of the input data. The goal of this module is to reduce data noise and fill in missing values, so as to provide accurate and unified input for the subsequent deep learning model. First, an adaptive anomaly detection model is established based on the dynamic sliding window Z-score method. The standardized score of the data point is calculated using the window mean µ and standard deviation. When |Zi|>3, it is determined as an outlier and removed. The box plot for outlier processing is as shown in Figure 3 Figure [not provided in the original]. Subsequently, an improved KNN weighted imputation algorithm is introduced for missing data. k nearest neighbor samples are selected through weighted Euclidean distance, and the filling value is generated according to the weight formula, where the parameter p controls the near neighbor attenuation rate. On this basis, a multi-scale Bayesian wavelet packet decomposition technology is used to implement signal denoising. The signal is decomposed into 3-5 layer frequency bands through the wavelet packet expansion formula, and the coefficients are shrunk based on the maximum a posteriori probability threshold to achieve noise suppression. Finally, the multi-source heterogeneous data are mapped to the [0,1] interval through the adaptive range normalization formula to eliminate the dimension difference.

[0109] Finally, information entropy is used as an evaluation index. This method is based on sample data, can effectively determine the possible range of parameters, and measure the confidence of the estimation, so as to quantitatively verify 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 no new problems are introduced in the data cleaning step. After removing outliers, filling in missing values, denoising with Bayesian wavelet packets, and normalizing the data, as shown inFigure 4 Compared with the original data, the information entropy is significantly reduced, indicating that the data quality has been significantly improved. These data preprocessing steps not only enhance the reliability of the data, but also improve the consistency and interpretability of the data. The calculation formula of information entropy is:

[0110]

[0111] 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.

[0112] Finally, the information entropy is used as an evaluation index. This method is based on sample data and can effectively determine the possible range of parameters and measure the confidence of the estimation, so as to quantitatively verify 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 no new problems are introduced in the data cleaning steps. As Figure 3 shown, after removing outliers, filling missing values, denoising with Bayesian wavelet packet, and normalizing the data, compared with the original data, the information entropy is significantly reduced, indicating that the data quality has been significantly improved. These data preprocessing steps not only enhance the reliability of the data, but also improve the consistency and interpretability of the data. The calculation formula of information entropy is:

[0113]

[0114] 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.

[0115] Step 2: Fault feature parameter extraction. The present invention uses the Pearson correlation coefficient method 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 the correlation between the \(a\)-th column \(x\) a and the \(b\)-th column \(x\) b in the research matrix:

[0116]

[0117] where \(m\) is the length of each column, and the value range of the correlation coefficient is from 0 to 1. The value 1 indicates perfect correlation, and the value 0 indicates no correlation between columns.

[0118] By using the Pearson correlation algorithm for calculation, as Figure 5As shown, the Pearson correlation coefficient results of each parameter are obtained. By comparing and analyzing the magnitudes of the Pearson correlation coefficients, the parameters significantly affected by the faults of the unmanned surface vessel can be screened out. According to the widely recognized definition, parameters with a Pearson coefficient greater than 0.6 are identified as having strong correlations, while parameters with a Pearson coefficient less than 0.6 are considered to have weak correlations. 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 characteristic parameters of the fault identification and early warning model for the electromechanical system of the unmanned platform. Specifically, the diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, and the temperature after cooling in the diesel engine are selected as the characteristic parameters for diesel engine operation analysis; the output current of motor 1, the output AC voltage of motor 1, the output power of motor 1, the output speed of motor 1, and the output torque of motor 1 are selected as the characteristic parameters for motor 1 operation analysis; the output current of motor 2, the output AC voltage of motor 2, the output power of motor 2, the output speed of motor 2, and the output torque of motor 2 are selected as the characteristic parameters for motor 2 operation analysis; the total battery voltage, bus voltage, and ship speed are selected as the characteristic parameters for the overall ship operation analysis.

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

[0120]

[0121] Step 3: Construct a fault early warning model for the unmanned surface vessel group. By using 80% of the healthy data, an early fault warning model for the unmanned surface vessel based on the Transformer - BiGRU GlobalAttention - CrossAttention deep learning algorithm is constructed. The remaining 20% of the healthy data is used as test data to verify the prediction ability of the model. And a variety of basic time - series prediction models are selected, including LSTM (Long Short - Term Memory Network), GRU (Gated Recurrent Unit), BILSTM (Bidirectional Long Short - Term Memory Network), BIGRU (Bidirectional Gated Recurrent Unit), and CNN - BIGRU (Convolutional Neural Network - Bidirectional Gated Recurrent Unit), to comprehensively evaluate the performance of different models in time - series data prediction tasks. Among them, the learning rates of GRU, LSTM, BIGRU, BILSTM, and CNN - BIGRU are all set to 0.0015. GRU and LSTM are single - layer models, BIGRU and BILSTM are bidirectional models, and the network structure of CNN - BIGRU includes a single CNN layer and a single BiGRU layer, with a convolutional kernel of 3×1. And multiple evaluation metrics are calculated for model analysis, including calculating the coefficient of determination (R²), mean absolute percentage error (MAPE), mean absolute error (MAE), symmetric mean absolute percentage error (SMAPE), median absolute error (MedAE), and explained variance score (EVS).

[0122] Table 3 Statistical Table of Prediction Accuracy Indexes of Each Model

[0123]

[0124]

[0125] The specific values of each index are listed in detail in Table 3. The calculation and analysis of these evaluation indexes provide an important basis for deeply understanding the performance and applicability of the model. The calculation formulas of each evaluation index are as follows.

[0126] Coefficient of determination (R2):

[0127]

[0128] Mean absolute percentage error (MAPE):

[0129]

[0130] Mean absolute error (MAE):

[0131]

[0132] Symmetric mean absolute percentage error (SMAPE):

[0133]

[0134] Median absolute error (MedAE):

[0135]

[0136] Explained variance score (EVS):

[0137]

[0138] Among them, yi is the actual value, is the predicted value, is the mean value of the actual values, and Var represents the variance.

[0139] Taking the prediction model trained and tested with multiple groups of health state data as an example, during the actual ship test of the unmanned surface vessel, when the diesel engine drives the generator 2 to generate electricity and drive, and the motor 1 self-drives the ship to cruise at low speed, the diesel engine stalls due to the damage of the control component, resulting in the output of the diesel engine stall and the emergency shutdown of the motor. The comparison of the health state and abnormal state of this section of data is as Figure 6 shown.

[0140] Step 4: Dynamically determine the warning threshold: In the present invention, the average value plus or minus several times the standard deviation is used to define the threshold, that is, the threshold range is set as [0, μ + kσ], where k is a coefficient determined according to actual requirements and experience to control the sensitivity of anomaly detection. The advantage of this method is that it can adaptively adjust the threshold according to the change 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 average value and standard deviation of the historical correlation change amount respectively, and k is an adjustment factor. According to multiple tests, k = 5 is set, so the MAE of 0.2 is used as the fault warning threshold.

[0143] Step 5: According to the analysis results of the warning model, the lubricating oil pressure rate of the diesel engine first shows abnormal fluctuations, which in turn trigger a series of related abnormal phenomena. Due to insufficient lubricating oil supply, the rotational speed and output power of the diesel engine immediately show significant abnormal changes. Further, since the shaft generator is involved in the power generation process, these abnormal fluctuations cause the voltage stability of the ship's power system to be damaged, resulting in abnormal voltage phenomena. As Figure 7 shown, the warning model first detects the abnormal lubricating oil pressure of the diesel engine at 370 seconds, and realizes abnormal reporting 300 seconds earlier than the voltage disorder of the ship's power system, providing an adequate time window for the system's emergency response. At the same time, this fault affects the cooling temperature of the diesel engine, but the fluctuation is below the overall warning threshold, so no alarm is given. The ship speed starts to decrease abnormally due to the fault, and the trend is not much different from the normal speed reduction moment, so no warning is given for the ship speed parameter.

[0144] The above embodiments are only used to illustrate the present invention, and any equivalent transformation and improvement based on the technical solution of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. A fault warning method for the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism, characterized in that, It includes the following steps: S1. Acquisition and governance analysis of heterogeneous data from multi-source sensors under complex working conditions; S1.1 Real-time acquisition of multi-modal sensor data of the electromechanical system of the unmanned platform; S1.2 Classification of multi-modal sensor data of the electromechanical system: The data is classified according to different acquisition positions; S1.3 Governance of heterogeneous data from multi-source sensors of the electromechanical system: Outlier removal, missing value filling, Bayesian wavelet packet denoising, and normalization processing are performed on the data; S2. Extraction of operating characteristic parameters under multi-source heterogeneous data of the electromechanical system: Based on the calculation results of Pearson correlation, select the parameters with a Pearson correlation coefficient greater than 0.6 as the characteristic parameters of the fault identification and early warning model for the electromechanical system of the unmanned platform; S3. Construct a fault early warning model for the unmanned platform based on the Transformer-BiGRU cross-attention mechanism: S3.1 Transformer encoder: It maps the input sequence to query Query, key Key, and value Value vectors, and calculates the correlation weights between any positions within the sequence through dot product operations: ； ； After multiple layers of stacking, the final output is: ； Among them, Q, K, and V are the query, key, and value respectively, and dk is the dimension of the key; According to the task assignment situation of the electromechanical system, the parameter specifications for the input and output feature dimensions of the model are determined, and at the same time, the attention dimension, encoder layer number, and number of multi-head attention heads of the Transformer encoder are allocated to capture the long-term association characteristics across time steps; S3.2 BiGRU network optimized by integrating the global attention mechanism: This network consists of a forward and a backward neural network, and encodes the input sequence through multiple layers of bidirectional GRUs. The state update equation is: ； ； ； ； ； where z t is the update gate, r t is the reset gate, denotes element-wise multiplication, and BIGRU Output is the output of each layer of BIGRU; S3.3 Global-local feature prediction fusion analysis under the cross-attention mechanism. Its calculation process is: The cross-attention mechanism performs a linear transformation on the input parameters of the electromechanical system through the query, key, and value, calculates the attention weights, and generates a weighted feature representation; The output of the cross-attention is dimension-reduced through an adaptive average pooling layer, and finally the final prediction result is generated through a fully connected layer; ； ； Among them, Q is the output from the Transformer, and K and V are the outputs from the BiGRU; ； ； ； Among them, W f is the weight matrix, and b f is the bias; S4. Determine the dynamic early warning threshold: The formula for the dynamic threshold T is as follows: ； Among them, μΔr and σΔr are the average value and standard deviation of the change in the prediction residuals of the historical early warning model respectively, and k is the adjustment factor; S5. Fault early warning and output. After being processed by the model, a fault prediction value is output. If the error between the prediction value and the actual fault occurrence situation exceeds the set dynamic threshold, it is considered that there is a possibility of system failure.

2. The fault warning method for the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, wherein, In the step S1.1, the multi-modal sensor data includes diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, temperature after cooling in the diesel engine, 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, temperature of the heat dissipation plate 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, temperature of the heat dissipation plate of motor 2, temperature of motor 2, temperature of gearbox 1, temperature of gearbox 2, outlet pressure of seawater pump, outlet pressure of cooling water, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

3. A fault warning method for an electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 2, characterized in that, In the step S1.2, the data is classified as follows: The first group of diesel engine monitoring parameters: diesel engine lubricating oil pressure, diesel engine power, diesel engine speed, diesel engine fuel consumption rate, temperature after cooling in the diesel engine; The second group of motor 1 monitoring parameters: 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, temperature of the heat dissipation plate of motor 1, temperature of motor 1; The third group of motor 2 monitoring parameters: 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, temperature of the heat dissipation plate of motor 2, temperature of motor 2; The fourth group of ship-wide operation parameters: temperature of gearbox 1, temperature of gearbox 2, outlet pressure of seawater pump, outlet pressure of cooling water, oil pressure of gearbox 1, oil pressure of gearbox 2, total battery voltage, bus voltage, and ship speed.

4. A fault warning method for the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that, In the step S1.3, the data is subjected to outlier removal, missing value filling, Bayesian wavelet packet denoising, and normalization. The specific processing process is as follows: An adaptive anomaly detection model is established based on the dynamic sliding window Z-score method. The standardized score of the data point is calculated using the window mean µ and standard deviation. When |Zi|>3, it is determined as an outlier and removed; An improved KNN weighted imputation algorithm is introduced for missing data. k nearest neighbor samples are selected through weighted Euclidean distance, and the filling value is generated according to the weight formula; The multi-scale Bayesian wavelet packet decomposition technology is used to implement signal denoising. The signal is decomposed into 3-5 layer frequency bands through the wavelet packet expansion formula, and the coefficients are shrunk based on the maximum a posteriori probability threshold; The multi-source heterogeneous data is mapped to the [0,1] interval through the adaptive range normalization formula.

5. A fault warning method for the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: In the step S1.3, after the governance of multi-source sensor heterogeneous data, the quality and accuracy of the cleaned data are evaluated through information entropy. The calculation formula of information entropy is: ； where \(H(X)\) represents the entropy of the random variable \(X\), and \(p(x i )\) is the probability of the \(i\)-th event in the sample, and \(n\) is the total number of events.

6. The fault warning method for the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: In the step S2, the Pearson correlation coefficient method is adopted 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 the correlation between the ath column x a and the bth column x b in the matrix: ； where m is the length of each column, and the value range of the correlation coefficient is from 0 to 1; the value 1 indicates complete correlation, and the value 0 indicates no correlation between columns.

7. A fault warning method for the electromechanical system of an unmanned platform based on the Transformer-BiGRU cross-attention mechanism according to claim 1, characterized in that: The Transformer encoder adopts the multi-head self-attention mechanism MHSA, and captures long-range dependence features by dynamically modeling the global context dependence relationship of the sequence.

8. A fault warning method for 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, utilizes information from previous and subsequent time steps, and captures time series data.

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

Images