Intelligent health monitoring method for ship power equipment based on multi-source data fusion

Abstract

A ship power equipment intelligent health monitoring method based on multi-source data fusion comprises the following steps: respectively constructing monitoring indexes for each measuring point based on an auto-encoder model, and enhancing multi-measuring-point data relevance by introducing a consistency criterion; a multi-source data fusion diagnosis model is constructed, and direct mapping between the monitoring data and the equipment operation state is achieved; in order to solve the problem of insufficient data under the condition of few samples, a data set is expanded by adopting a random quantization enhancement technology, a Transform encoder network and a channel attention mechanism are introduced, and the identification precision and robustness of a model to a fault mode are enhanced through weighted fusion of multi-source signal features; according to the method, multi-measuring-point information is fully utilized, the integrity of fault feature extraction and the diagnosis precision are improved, and efficient and accurate monitoring and evaluation of the health state of the ship power equipment under the complex working condition are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of health state monitoring and evaluation of ship power equipment, and particularly relates to an intelligent health monitoring method for ship power equipment with multi-source data fusion. Background Technique

[0002] With the rapid development of the shipbuilding industry, the complexity and integration of power equipment such as ship propellers and diesel engines have been continuously improved, and these equipment play a crucial role in ship operation. However, under complex working conditions such as high rotational speed and large load, various faults will inevitably occur in key components such as bearings, shafts, and gears of these equipment. If the health state of key components cannot be identified in time and effective preventive maintenance measures are not taken, it will not only accelerate the degradation of equipment performance and reduce operation efficiency, but may also cause serious economic losses and even affect the navigation safety of ships. Therefore, researching efficient and intelligent fault monitoring and health state evaluation methods has important engineering value and practical significance for improving the reliability of ship power equipment, reducing maintenance costs, and ensuring navigation safety.

[0003] At present, certain progress has been made in the field of intelligent ships in terms of equipment status monitoring and performance evaluation. Signal processing and deep learning technologies are widely used in the fault diagnosis and health assessment of ship power equipment (Zhang P, Gao Z, Cao L, et al. "Marine systems and equipment prognostics and health management: a systematic review from health condition monitoring to maintenance strategy," Machines, vol. 10, pp: 72, 2022). However, the existing health status monitoring and evaluation methods based on traditional signal processing still have limitations in terms of automation, real-time performance, and large-scale data processing capabilities, and are difficult to meet the engineering application requirements under complex working conditions. In addition, the existing health status evaluation methods based on deep learning mainly rely on a single vibration signal source for performance evaluation. However, the single-point signal has the following deficiencies in the state monitoring of complex ship mechanical systems: (a) The ship operating environment is complex and changeable, and the fault modes often show multi-dimensional vibration characteristic changes. The signal of a single measurement point is difficult to comprehensively characterize the complex fault mechanism of the equipment, easily leading to the omission of key information, thus affecting the accuracy and reliability of the evaluation. (b) The single-point signal is easily affected by local noise, measurement errors, and environmental interference. Especially in a high-noise environment or when the working conditions change rapidly, the fault characteristics may be masked by noise, reducing the generalization ability and diagnostic accuracy of the model. (c) The results of fault monitoring and evaluation highly depend on the arrangement position of the sensors. If the measurement points fail to cover the key fault areas, the fault characteristic signals may be weakened or even unable to be effectively detected, thus affecting the stability and accuracy of the evaluation.

[0004] In summary, there are many defects in the existing fault monitoring and health status evaluation based on signal processing and deep learning. It is necessary to explore a more intelligent and efficient multi-source data fusion method to improve the robustness and reliability of fault monitoring and health evaluation, so as to better meet the state monitoring requirements of ship equipment in complex environments. Summary of the Invention

[0005] In order to overcome the above defects of the existing technologies, the purpose of the present invention is to propose an intelligent health monitoring method for ship power equipment based on multi-source data fusion, which makes full use of multi-measurement point information, improves the integrity of fault feature extraction and the accuracy of diagnosis, and realizes the efficient and accurate monitoring and evaluation of the health status of ship power equipment under complex working conditions.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] An intelligent health monitoring method for ship power equipment based on multi-source data fusion constructs monitoring indicators for each measuring point based on an autoencoder model, and enhances the correlation of multi-measurement point data by introducing a consistency criterion; furthermore, a multi-source data fusion diagnosis model is constructed to achieve a direct mapping between the monitoring data and the equipment operation state; meanwhile, to solve the problem of insufficient data under few-sample working conditions, a random quantization enhancement technology is adopted to expand the data set, and a Transformer encoder network and a channel attention mechanism are introduced to enhance the identification accuracy and robustness of the model to fault modes by weighted fusion of multi-source signal features.

[0008] An intelligent health monitoring method for ship power equipment based on multi-source data fusion includes the following steps:

[0009] Step 1: Data collection and preprocessing; obtain a multi-source monitoring data set of ship power equipment where x i is the monitoring data of the i-th measuring point, N is the length of each measuring point sample, and m is the number of measuring points; extract the time-domain features of the multi-source monitoring data to obtain a time-domain feature data set For the data of each measuring point, divide the training set and the test set according to a ratio;

[0010] Step 2: Establish an intelligent construction model for monitoring indicators, and input the training set into the intelligent construction model for monitoring indicators, and calculate its feature mapping result and reconstruction result where K is the length of each training sample, and Z is the length of the feature h;

[0011] Step 3: Calculate the loss function of the intelligent construction model for monitoring indicators. The loss function consists of two parts: reconstruction loss and consistency loss. The specific calculation formula is as follows:

[0012]

[0013] In the formula: N1 is the total number of training samples; L RMSE,i is the reconstruction error of the i-th measuring point; x' j (k) is the j-th training sample; is the reconstruction result of the model for the j-th training sample; m is the number of measuring points; L RMSE is the reconstruction error loss term;

[0014]

[0015] In the formula: Cor(h i , h j ) is the correlation between the monitoring indicators h i and h j ; and They are the means of h i and h j respectively; L CE is the consistency loss term;

[0016] L AE = ω1L RMSE + ω2L CE (5)

[0017] In the formula, L AE is the loss function of the intelligent construction model of the monitoring index, and ω1, ω2 are the proportionality coefficients for balancing the reconstruction loss and the consistency loss; taking the loss L AE as the optimization objective in the training stage, the gradient descent method is used to update the model parameters θ AE :

[0018]

[0019] In the formula, α represents the learning rate;

[0020] Step 4: Repeat Step 2 - Step 3 to continuously iterate and optimize the intelligent construction model of the monitoring index until the maximum number of iterations is reached;

[0021] Step 5: Input the training set samples and test set samples into the trained intelligent construction model of the monitoring index in chronological order to obtain the monitoring index h i corresponding to the i-th measuring point = [h1, h2,..., h Z ; perform maximum - minimum normalization on the monitoring index values generated for each measuring point and further take the average to obtain the global average monitoring index The specific calculation formula is as follows:

[0022]

[0023] In the formula, h′ i represents the normalized monitoring index of the i-th measuring point, min(·) represents the minimum value, and max(·) represents the maximum value;

[0024] Step 6: According to the amplitude distribution characteristics of the global average monitoring index, divide the data into samples of C different states, and use the division result to assign corresponding label information to the monitoring data of each measuring point;

[0025] Step 7: Divide the quantization interval; for the original input sequence h′ i = [h0, h1, h2,..., h Z-1 of the given measuring point i, where Z is the length of the input sequence h i ′, C hrepresents the number of channels of the input sequence, and statistically calculates the amplitude distribution range (min(h′ i ), max(h′ i )) for each channel; perform a - 1 random samplings within this interval and arrange the sampling values in ascending order. Divide the amplitude interval (min(h i ′), max(h i ′)) of the original sequence into a non - overlapping intervals. The expression is as follows:

[0026]

[0027] where: P = [p1, p2,..., p a-1 is the regional percentile of the endpoints of each interval; U(·) represents random uniform distribution sampling within the interval. The calculation formula for the interval endpoint b is as follows:

[0028] b = P×((max(h i ′)-min(h′ i )))+min(h′ i ) (10)

[0029] The quantized interval after division is S j = [b j , b j+1 );

[0030] Step 8: Random quantization enhancement; perform random sampling within each interval and map all values in this interval to the random sampling value u j , and the expression is as follows:

[0031]

[0032] where h q,i is the quantized sequence of the training sample at the i - th measurement point; q(·) - - quantization operation;

[0033]

[0034] Step 9: Establish an intelligent evaluation model for the performance of multi - source fusion equipment, and use the enhanced training sample set to optimize the intelligent evaluation model for the performance of multi - source fusion equipment. The parameter to be optimized is θ FD . During the optimization process, update the model parameter θ by minimizing the following objective function FD :

[0035]

[0036] In the formula, L FD is the optimization function of the intelligent evaluation model for the performance of multi - source fusion equipment, y is the true label value, is the predicted result, and α is the learning rate;

[0037] Step 10: Repeat Steps 7 - 9 to continuously iterate and optimize the intelligent evaluation model for the performance of the multi-source fusion device until the maximum number of iterations is reached;

[0038] Step 11: Test the model. Input the test set into the trained intelligent construction model for monitoring indicators and the intelligent evaluation model for the performance of the multi-source fusion device in sequence to obtain the evaluation results of the test set.

[0039] The intelligent construction model for monitoring indicators in Step 2 consists of two parts: a feature extraction module and a data reconstruction module. Among them, the feature extraction module is composed of three stacked convolutional blocks, and each convolutional block contains a convolutional layer, a batch normalization layer, and a pooling layer, which are used to extract the potential high-level features of the input signal. The data reconstruction module is composed of two stacked transposed convolutional blocks, and each transposed convolutional block contains a transposed convolutional layer and a batch normalization layer, which are used to reconstruct the extracted features into the original signal.

[0040] The intelligent evaluation model for the performance of the multi-source fusion device in Step 9 consists of three parts: a representation learning network, a feature fusion module, and a health status evaluation network. The representation learning network is composed of three stacked Transformer encoder blocks, and the self-attention mechanism is used to extract key information from the input sequence. The feature fusion module introduces a channel attention mechanism to achieve effective fusion of data from different measurement points. The health status evaluation network contains two fully connected layers to further remove the redundant information irrelevant to the device performance in the fused features.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] The present invention proposes an intelligent health monitoring method for ship power equipment based on multi-source data fusion. By establishing an intelligent performance evaluation framework based on deep learning, it realizes the construction of monitoring indicators from the original multi-source data of ship power equipment, automatically extracts equipment fault information and maps it to the health status of the equipment. This method makes full use of the multi-measurement point data of ship power equipment to improve the model performance evaluation ability, and introduces a consistency index and a channel attention mechanism to enhance the correlation of multi-measurement point data. At the same time, combining the stochastic quantization enhancement technology and the Transformer model improves the model's feature extraction ability and enhances the recognition accuracy and robustness for different fault modes. Description of the Drawings

[0043] Figure 1 is the flowchart of the embodiment of the present invention.

[0044] Figure 2 is the result of constructing the monitoring indicators for the ship diesel engine equipment in the embodiment of the present invention. Detailed Embodiments

[0045] The present invention will be further described in detail below in conjunction with embodiments and the accompanying drawings.

[0046] As Figure 1 shown, an intelligent health monitoring method for ship power equipment with multi-source data fusion includes the following steps:

[0047] Step 1: Data collection and preprocessing; obtaining a multi-source monitoring data set of ship power equipment where x i is the monitoring data of the i-th measurement point, N is the length of each measurement point sample, and m is the number of measurement points; extracting time-domain features from the multi-source monitoring data to obtain a time-domain feature data set For the data of each measurement point, the training set and the test set are divided in a ratio of 5:5. The calculation formulas of each time-domain feature are shown in the following table:

[0048] Table 2 Calculation formulas of time-domain feature indicators

[0049]

[0050]

[0051] The original vibration data of the ship contains a large amount of noise and redundant information. Directly inputting it into the model is likely to lead to high computational complexity and affect the model reconstruction effect. Therefore, first perform feature extraction on the original data, filter the noise, reduce the data dimension, and at the same time retain important features to improve the computational efficiency and robustness of the model;

[0052] Step 2: Establish an intelligent construction model for monitoring indicators, and input the training set into the intelligent construction model for monitoring indicators to calculate its feature mapping result and reconstruction result where K is the length of each training sample and Z is the length of the feature h;

[0053] The intelligent construction model for monitoring indicators includes two parts: a feature extraction module and a data reconstruction module; among them, the feature extraction module is composed of three convolutional blocks stacked together. Each convolutional block contains a convolutional layer, a batch normalization layer, and a pooling layer, which are used to extract the potential high-level features of the input signal; the data reconstruction module is composed of two transposed convolutional blocks stacked together. Each transposed convolutional block contains a transposed convolutional layer and a batch normalization layer, which are used to reconstruct the extracted features into the original signal;

[0054] Step 3: Calculate the loss function of the intelligent construction of monitoring indicators. The loss function consists of two parts: reconstruction loss and consistency loss. The specific calculation formula is as follows:

[0055]

[0056] where: N1 is the total number of training samples; L RMSE,i is the reconstruction error of the i-th measurement point; x′ j (k) is the j-th training sample; is the reconstruction result of the model for the j-th training sample; m is the number of measurement points; L RMSE is the reconstruction error loss term;

[0057]

[0058] where: Cor(h i , h j ) is the correlation between the monitoring indicators h i and h j ; and are the mean values of h i and h j respectively; L CE is the consistency loss term;

[0059] L AE = ω1L RMSE + ω2L CE (5) where, L AE is the loss function of the intelligent construction model of the monitoring indicator, ω1, ω2 are the proportionality coefficients for balancing the reconstruction loss and the consistency loss; taking the loss L AE as the optimization objective in the training stage, and using the gradient descent method to update the model parameters θ AE :

[0060]

[0061] where, α represents the learning rate;

[0062] The reconstruction loss is used to measure the reconstruction ability of the model for the input data, ensuring that the model can accurately extract features from the input data and reconstruct the original input; the consistency loss is used to measure the correlation between different monitoring indicators, ensuring that the change trends of the monitoring indicators at each measurement point are consistent; by jointly optimizing the reconstruction loss and the consistency loss, the reconstruction ability of the model for the input data and the consistency of the monitoring indicators at different measurement points can be improved, thereby achieving a more accurate health status representation;

[0063] Step 4: Repeat Step 2 - Step 3 to continuously iterate and optimize the intelligent construction model of the monitoring indicator until the maximum number of iterations is reached;

[0064] Step 5: Input the training set samples and test set samples into the trained intelligent construction model of the monitoring indicator in chronological order to obtain the monitoring indicator h corresponding to the i-th measurement point i=[h1, h2,..., h Z ; The maximum-minimum normalization process is performed on the monitoring index values generated for each measurement point to eliminate the dimensional differences of the monitoring indexes at different measurement points and ensure the balanced contribution of each measurement point; further averaging it to obtain the global average monitoring index The specific calculation formula is as follows:

[0065]

[0066] In the formula, h′ i represents the normalized monitoring index of the i-th measurement point, min(·) represents the minimum value, and max(·) represents the maximum value;

[0067] Step 6: According to the amplitude distribution characteristics of the global average monitoring index, divide the data into samples of C different states, and use the division result to assign corresponding label information to the monitoring data of each measurement point;

[0068] Step 7: Divide the quantization interval; for the original input sequence h′ i =[h0, h1, h2,..., h Z-1 of the given measurement point i, where Z represents the length of the original input sequence h′ i , C h represents the number of channels of the input sequence, and statistically calculate the amplitude distribution range (min(h i ′), max(h i ′)) of each channel; perform a - 1 random samplings within this interval and arrange the sampling values in ascending order, and divide the amplitude interval (min(h i ′), max(h i ′)) of the original sequence into a non-overlapping intervals, and the expression is as follows:

[0069]

[0070] where: P = [p1, p2,..., p a-1 is the regional percentile of the endpoints of each interval; U(·) represents random uniform distribution sampling within the interval, and the calculation formula for the interval endpoint b is as follows:

[0071] b = P×((max(h i ′)-min(h i ′)))+min(h′ i ) (10)

[0072] The divided quantization interval is S j =[b j , b j+1 );

[0073] Step 8: Random quantization enhancement; perform random sampling within each interval and map all values in the interval to the random sampling value u j , the expression is as follows:

[0074]

[0075] where h q,i is the quantized sequence of the training samples at the i-th measurement point; q(·) —— quantization operation;

[0076]

[0077] The purpose of random quantization enhancement is to expand the sample diversity without changing the overall feature distribution of the data by compressing the information within each divided interval while retaining the differences between different intervals;

[0078] Step 9: Establish an intelligent evaluation model for the performance of multi-source fusion devices, and use the enhanced training sample set to optimize the intelligent evaluation model for the performance of multi-source fusion devices. The parameter to be optimized is θ FD , and during the optimization process, the model parameter θ is updated by minimizing the following objective function FD :

[0079]

[0080] In the formula, L FD is the optimization function of the intelligent evaluation model for the performance of multi-source fusion devices, y is the true label value, is the predicted result, and α is the learning rate;

[0081] The intelligent evaluation model for the performance of multi-source fusion devices consists of three parts: a representation learning network, a feature fusion module, and a health status evaluation network; the representation learning network is composed of three stacked Transformer encoder blocks, and uses the self-attention mechanism to extract key information from the input sequence; the feature fusion module introduces a channel attention mechanism to achieve effective fusion of data from different measurement points; the health status evaluation network contains two fully connected layers to further remove redundant information irrelevant to the device performance in the fused features;

[0082] Step 10: Repeat Step 7 - Step 9 to continuously iterate and optimize the intelligent evaluation model for the performance of multi-source fusion devices until the maximum number of iterations is reached;

[0083] Step 11: Test the model, and sequentially input the test set into the trained intelligent construction model for monitoring indicators and the intelligent evaluation model for the performance of multi-source fusion devices to obtain the evaluation results of the test set.

[0084] Example: Taking the intelligent fault monitoring and performance evaluation of the diesel generator components of a certain ship as an example, the effectiveness of the method of the present invention is verified.

[0085] Multi-source monitoring data of multiple diesel generator sets are obtained from the ship data acquisition system. Figure 2 Subfigure (a) of shows the multi-source monitoring data of the diesel engine equipment DG1, which includes a total of 4 measuring points. Among them, vibration monitoring data are collected at measuring points 1 to 3, and energy consumption monitoring data, specifically the fuel consumption rate of the diesel engine, are collected at measuring point 4. Time-domain features are extracted from the collected vibration data, the data is sampled using a sliding time window, and the data is divided into a training set and a test set according to a ratio of 5:5. Through overlapping sampling, 333 training samples and 954 test samples are obtained. Each sample contains 42 features, and the length of each feature sequence is 30. The training set is input into the intelligent construction model of monitoring indicators for model pre-training. The intelligent construction model of monitoring indicators includes three convolutional layers and two deconvolutional layers. The output dimensions of the convolutional layers are 256, 256, and 128 respectively, and the output dimensions of the deconvolution are 128 and 256. The model pre-training parameter configuration is shown in Table 3:

[0086] Table 3 Pre-training parameters of the intelligent construction model of monitoring indicators

[0087]

[0088] After the pre-training of the intelligent construction model of monitoring indicators is completed, the training set and test set data are input into the intelligent construction model of monitoring indicators in chronological order, and the high-level features output by the last convolutional layer of the model are obtained and used as the monitoring indicators corresponding to each measuring point. Figure 2 Subfigure (b) of shows the monitoring indicators corresponding to each measuring point of the diesel engine equipment DG1. It can be seen from the figure that when the amplitude of the original monitoring data increases (decreases), the monitoring indicator values of each measuring point decrease (increase), indicating that the monitoring indicators of each measuring point can effectively reflect the changes in the equipment working conditions or health status. The consistency criterion is used to evaluate the consistency of the monitoring indicators corresponding to each measuring point. The results are shown in Table 4. It can be seen from the data in the table that the consistency indicators of the diesel generators (DG1 to DG4) are all relatively high, with the lowest being 0.8864 (DG3) and the highest being 0.9141 (DG4). The overall consistency remains at a relatively high level, indicating that the method of the present invention can better capture the correlation between different measuring points of the diesel generator, and the constructed monitoring indicators of each measuring point of the diesel generator have strong consistency.

[0089] Table 4 Evaluation results of the consistency of monitoring indicators for different measuring points of diesel generator equipment

[0090]

[0091] The monitoring indicators of different measurement points above are subjected to maximum-minimum normalization processing, and then averaged to obtain the evaluation monitoring indicators. After that, according to the amplitude distribution characteristics of the average monitoring indicators, the data is divided into two types of samples in different states, and each type of sample corresponds to an operating state or health state of a device. The corresponding label information is assigned to the monitoring data of each measurement point by using the division result. Then, the training and test data sets are divided, and the random quantization enhancement algorithm is used to expand the data of the few-sample category.

[0092] The maximum mean discrepancy (MMD) index is used to quantitatively evaluate the quality of the data generated by random quantization enhancement. The maximum mean discrepancy is usually used to measure the distance between two distributions. The smaller the MMD value, the more similar the distribution of the generated data is considered to be to the real data. On the contrary, the lower the distribution similarity of the two data sets is considered. The calculation formula of MMD is as follows:

[0093]

[0094] In the formula, H is the reproducing kernel Hilbert space, and φ(·) is the non-linear mapping in H;

[0095] Table 5 Distribution differences between the data generated by random quantization and the real data

[0096]

[0097] It can be seen from Table 5 that the minimum MMD value is 0.036 and the maximum is only 0.058, indicating that the distribution differences between the data generated by random quantization of different devices and the real data are small. The data after random quantization enhancement is input into the multi-source fusion device performance intelligent evaluation model for training. The hyperparameter settings of the model during the training process are shown in Table 6.

[0098] Table 6 Training parameters of the multi-source data fusion performance evaluation model

[0099]

[0100] Table 7 shows the results of the method of the present invention in the intelligent evaluation of the performance of diesel engine equipment. It can be seen from the table that the constructed multi-source fusion device performance intelligent evaluation model can effectively capture the operating characteristics of the diesel engine in different states, accurately distinguish each state, and the recognition accuracy for all states reaches more than 87.8%, indicating that the method based on multi-source data fusion has high stability and accuracy in analyzing the states of complex devices.

[0101] Table 7 Intelligent evaluation results of device performance

[0102]

Claims

1. A method for intelligent health monitoring of ship power equipment based on multi-source data fusion, characterized in that: Based on the autoencoder model, monitoring indicators are constructed for each measuring point, and the correlation of data from multiple measuring points is enhanced by introducing the consistency criterion. Then, a multi-source data fusion diagnosis model is constructed to achieve direct mapping between monitoring data and equipment operating status. The random quantization enhancement technology is used to expand the data set, and the Transformer encoder network and channel attention mechanism are introduced to enhance the model's recognition accuracy and robustness for fault modes by weighted fusion of multi-source signal features.

2. According to claim 1, a method for intelligent health monitoring of ship power equipment based on multi-source data fusion is characterized in that: The following steps are involved: Step 1: Data collection and preprocessing; Obtain multi-source monitoring data sets for ship power equipment Among them, x i is the monitoring data of the i-th measuring point, N is the length of each measuring point sample, and m is the number of measuring points; extract the time domain features of the multi-source monitoring data to obtain the time domain feature data set For the data of each measuring point, divide the training set and the test set according to the proportion; Step 2: Establish an intelligent model for monitoring indicators and convert the training set Input into the monitoring indicator intelligent construction model to calculate its feature mapping results and the reconstruction results Where K is the length of each training sample, and Z is the length of feature h; Step 3: Calculate the loss function of the monitoring indicator intelligent construction model. The loss function consists of two parts: reconstruction loss and consistency loss. The specific calculation formula is as follows: Where: N1 is the total number of training samples; L RMSE,i is the reconstruction error of the ith measurement point; x′ j (k) is the jth training sample; is the reconstruction result of the model for the jth training sample; m is the number of measurement points; L RMSE is the reconstruction error loss term; Where: Cor(h i ,h j ) is the monitoring indicator h i and h j relevance; and They are h i and h j The mean value of L CE is the consistency loss term; L AE =ω1L RMSE +ω2L CE (5) Where, L AE The loss function of the intelligent model is constructed for monitoring indicators. ω1 and ω2 are the proportional coefficients of balanced reconstruction loss and consistency loss. AE As the optimization target in the training phase, the model parameters θ are updated using the gradient descent method. AE : In the formula, α represents the learning rate; Step 4: Repeat steps 2 and 3 to continuously iterate and optimize the intelligent construction model of monitoring indicators until the maximum number of iterations is reached; Step 5: Input the training set samples and test set samples into the trained monitoring indicator intelligent construction model in chronological order to obtain the monitoring indicator h corresponding to the i-th measuring point i =[h1,h2,...,h Z ]; The monitoring index values ​​generated by each measuring point are normalized to the maximum and minimum values ​​and then averaged to obtain the global average monitoring index The specific calculation formula is as follows: In the formula, h′ i represents the normalized monitoring index of the ith measuring point, min(·) represents the minimum value, and max(·) represents the maximum value; Step 6: According to the amplitude distribution characteristics of the global average monitoring index, the data is divided into C types of samples with different states, and the corresponding label information is assigned to the monitoring data of each measuring point using the division results; Step 7: Divide the quantization interval; for the original input sequence h′ of a given measurement point i i =[h0,h1,h2,...,h Z-1 ], Where Z represents the original input sequence h′ i Length, C h Indicates the number of channels of the input sequence, and counts the amplitude distribution range of each channel (min(h′ i ),max(h′ i )); perform a-1 random sampling in this interval and arrange the sampled values ​​in ascending order, and convert the amplitude interval of the original sequence (min(h′ i ),max(h′ i )) is divided into a non-overlapping intervals, and the expression is as follows: Where: P = [p1, p2, ..., p a-1 ] is the regional percentile of each interval endpoint; U(·) represents random uniform distribution sampling within the interval, and the calculation formula of the interval endpoint b is as follows: b=P×((max(h′ i )-min(h′ i )))+min(h′ i ) (10) The quantization interval after division is S j =[b j ,b j+1 ); Step 8: Random quantization enhancement; perform random sampling in each interval and map all values ​​in the interval to random sampling values ​​u j , the expression is as follows: Among them, h q,i is the quantized sequence of the i-th test point training sample; q(·)——quantization operation; Step 9: Establish a multi-source fusion equipment performance intelligent evaluation model using the enhanced training sample set Optimize the intelligent evaluation model of multi-source fusion equipment performance, and the parameter to be optimized is θ FD , the model parameters θ are updated by minimizing the following objective function during the optimization process FD : Where, L FD is the optimization function of the multi-source fusion device performance intelligent evaluation model, y is the real label value, is the predicted result, α is the learning rate; Step 10: Repeat steps 7 to 9 to continuously iteratively optimize the multi-source fusion device performance intelligent evaluation model until the maximum number of iterations is reached; Step 11: Test the model. Input the test set into the trained monitoring indicator intelligent construction model and the multi-source fusion equipment performance intelligent evaluation model in turn to obtain the evaluation result of the test set.

3. The method according to claim 2, characterized in that: The intelligent construction model of monitoring indicators in step 2 includes two parts: feature extraction module and data reconstruction module; the feature extraction module is composed of three stacked convolution blocks, each of which contains a convolution layer, a batch normalization layer and a pooling layer, which is used to extract potential high-level features of the input signal; the data reconstruction module is composed of two stacked deconvolution blocks, each of which contains a deconvolution layer and a batch normalization layer, which is used to reconstruct the extracted features into the original signal.

4. The method according to claim 2, characterized in that: The multi-source fusion equipment performance intelligent evaluation model in step 9 includes three parts: representation learning network, feature fusion module and health status evaluation network; the representation learning network is composed of three layers of Transformer encoder blocks, and uses a self-attention mechanism to extract key information from the input sequence; The feature fusion module introduces a channel attention mechanism to achieve effective fusion of data from different measurement points; the health status assessment network contains two fully connected layers to further remove redundant information in the fused features that is not related to device performance.

Images