A fault detection method for a marine power condenser system based on multi-scale feature fusion

Abstract

A kind of ship power condenser system fault detection method based on multiscale feature fusion, the present application relates to ship power condenser system fault detection method based on multiscale feature fusion, belongs to ship power system fault diagnosis and safety monitoring field.The present application aims to solve the problem of insufficient detection accuracy and high false alarm rate of traditional fault diagnosis method.Process is as follows:step one, obtain the input and output information of condenser system;Step two, construct multiscale feature extraction model;Step three, obtain trained multiscale feature extraction model and parameters;Step four, input the online data of condenser system into trained multiscale feature extraction model, and trained multiscale feature extraction model outputs the predicted value of condenser liquid level and vacuum degree;Residual generator obtains online prediction error signal based on predicted value;Step five, fault diagnosis based on online prediction error.

Description

Technical Field

[0001] This invention relates to a fault detection method for marine power condenser systems based on multi-scale feature fusion, belonging to the field of marine power system fault diagnosis and safety monitoring. Background Technology

[0002] With the rapid development of digital and intelligent technologies, the level of intelligence and automation in marine power systems is constantly improving. As the core system of a ship, the safety and reliability of the power system are directly related to the ship's navigation safety and operational efficiency. Among them, the condenser, as an important auxiliary device of the steam turbine generator set, directly affects the thermal efficiency and stability of the entire power system. However, with the increase in system complexity, the equipment failure rate has risen significantly. In particular, the condenser system, due to its harsh working environment and complex operating conditions, is prone to various failures, such as sensor drift, actuator jamming, and pipeline blockage.

[0003] Existing fault diagnosis technologies can be mainly categorized into model-based, signal processing-based, and data-driven methods. Model-based methods achieve fault detection by establishing physical or mathematical models of the system, but they suffer from high modeling difficulty and computational complexity in complex nonlinear systems. Signal processing-based methods directly extract fault features from sensor signals, but they are susceptible to noise interference and their performance degrades when multiple faults occur concurrently. Although data-driven fault diagnosis methods have developed rapidly, traditional deep learning methods have inherent limitations: recurrent networks such as LSTM are inefficient when processing long sequences and struggle to capture long-term dependencies; CNNs are limited by convolutional kernel parameters, resulting in a small receptive field and difficulty in capturing long-distance features; and single-timescale feature extraction cannot simultaneously handle progressive and sudden faults.

[0004] Traditional fault diagnosis methods often rely on single-time-scale feature extraction and simple linear models, making it difficult to accurately capture the multi-scale dynamic characteristics of complex nonlinear systems. This results in insufficient fault detection accuracy and a high false alarm rate. Condenser systems exhibit typical multivariable, temporal, and multi-scale characteristics, which traditional methods struggle to effectively capture. Summary of the Invention

[0005] The purpose of this invention is to address the problem that traditional fault diagnosis methods rely heavily on single-time-scale feature extraction and simple linear models, making it difficult to accurately capture the multi-scale dynamic characteristics of complex nonlinear systems, resulting in insufficient fault detection accuracy and a high false alarm rate. Therefore, this invention proposes a fault detection method for marine power condenser systems based on multi-scale feature fusion.

[0006] The specific process of a fault detection method for a marine power condenser system based on multi-scale feature fusion is as follows:

[0007] Step 1: Obtain the input and output information of the condenser system;

[0008] The input information of the condenser system at any time is , ;

[0009] The output information of the condenser system at any time is , ;

[0010] express Real-time cooling water flow rate; express Steam supply from the steam turbine at any given time;

[0011] express Continuous condenser liquid level; express Vacuum degree at any moment;

[0012] Superscript This indicates the transpose;

[0013] Step 2: Construct a multi-scale feature extraction model;

[0014] The multi-scale feature extraction model includes a long-term feature extraction network, a short-term feature extraction network, an error feedback feature extraction network, a cross-attention fusion module, a decoder, and a residual generator;

[0015] The long-term feature extraction network is a Long Short-Term Memory (LSTM) network;

[0016] The short-term feature extraction network is an encoder architecture based on Transformer;

[0017] The error feedback feature extraction network is a multilayer perceptron (MLP).

[0018] The decoder is a multilayer perceptron (MLP).

[0019] Step 3, Cooling water flow rate , Steam supply from the steam turbine at any time , condenser liquid level , Vacuum degree at any moment Input the multi-scale feature extraction model, and the multi-scale feature extraction model outputs feature vectors; train the multi-scale feature extraction model to obtain the trained multi-scale feature extraction model and parameters;

[0020] Step 4: Collect online data from the condenser system. This online data includes the condenser system's input information at all times. Output information of the condenser system ;

[0021] The online data from the condenser system is input into the trained multi-scale feature extraction model, which outputs predicted values ​​for the condenser liquid level and vacuum degree. , ;

[0022] The residual generator is based on the predicted value. Obtain online prediction error signal ;

[0023] Step 5: Perform fault diagnosis based on online prediction errors.

[0024] The beneficial effects of this invention are as follows:

[0025] This invention proposes a data-driven fault diagnosis method based on multi-scale deep learning. It captures the long-term dependency features of the system through a long-term LSTM network, extracts local temporal patterns using a short-term Transformer network, and designs a prediction error feedback mechanism to achieve adaptive adjustment. It also adopts a cross-attention fusion strategy to effectively integrate multi-source information and finally achieves accurate fault detection and diagnosis through residual analysis.

[0026] Compared with existing technologies, this invention has the following advantages: strong multi-scale feature capture capability, capable of simultaneously handling progressive faults (such as sensor drift) and sudden faults (such as valve jamming); outstanding adaptive capability, able to adapt to changes in operating conditions through a prediction error feedback mechanism; intelligent fusion strategy, with a cross-attention mechanism that can automatically learn the importance of different feature sources; and high detection accuracy, with a fault detection accuracy rate of 95.2% and a false alarm rate reduced to 2.3%.

[0027] This invention extends multi-scale feature extraction technology to the field of fault diagnosis of ship condenser systems, overcomes the limitations of traditional methods in complex nonlinear systems, and provides effective protection for the safe operation of ship power systems. Attached Figure Description

[0028] Figure 1 This is a diagram illustrating the overall architecture of the multi-scale deep learning fault diagnosis system of this invention.

[0029] Figure 2 This is a schematic diagram of the working principle of a condenser.

[0030] Figure 3 Design of time windows and data flow diagram for multi-scale feature extraction;

[0031] Figure 4 This is a comparative analysis chart of the condenser liquid level time series when a fault occurs;

[0032] Figure 5 This is a time-series comparative analysis of condenser vacuum levels during fault conditions;

[0033] Figure 6 This is a graph showing the residual time series analysis when there is a fault;

[0034] Figure 7 This is a graph showing the results of intelligent fault statistics detection. Detailed Implementation

[0035] Specific implementation method one: Combining Figure 1 This embodiment describes a fault detection method for a marine power condenser system based on multi-scale feature fusion. The specific process is as follows:

[0036] Step 1: Obtain the input and output information of the condenser system;

[0037] The input information of the condenser system at any time is , ;

[0038] The output information of the condenser system at any time is , ;

[0039] express Real-time cooling water flow rate; express Steam supply from the steam turbine at any given time;

[0040] express Continuous condenser liquid level; express Vacuum degree at any moment;

[0041] Superscript This indicates the transpose;

[0042] Step 2: Construct a multi-scale feature extraction model;

[0043] The multi-scale feature extraction model includes a long-term feature extraction network, a short-term feature extraction network, an error feedback feature extraction network, a cross-attention fusion module, a decoder, and a residual generator;

[0044] The long-term feature extraction network is a Long Short-Term Memory (LSTM) network;

[0045] The short-term feature extraction network is an encoder architecture based on Transformer;

[0046] The error feedback feature extraction network is a multilayer perceptron (MLP).

[0047] The decoder is a multilayer perceptron (MLP).

[0048] To address the coupling characteristics between short-term dynamic response and long-term evolution trend in condenser systems, a dual-path feature extraction network with both short-term and long-term paths is designed to capture slow changes in thermodynamic equilibrium and immediate responses to control actions, respectively.

[0049] Step 3, Cooling water flow rate , Steam supply from the steam turbine at any time , condenser liquid level , Vacuum degree at any moment Input the multi-scale feature extraction model, and the multi-scale feature extraction model outputs feature vectors; train the multi-scale feature extraction model to obtain the trained multi-scale feature extraction model and parameters;

[0050] Step 4: Collect online data from the condenser system. This online data includes the condenser system's input information at all times. Output information of the condenser system ;

[0051] The online data from the condenser system is input into the trained multi-scale feature extraction model, which outputs predicted values ​​for the condenser liquid level and vacuum degree. , ;

[0052] The residual generator is based on the predicted value. Obtain online prediction error signal ;

[0053] Step 5: Perform fault diagnosis based on online prediction errors.

[0054] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step three, [the following will be included] Cooling water flow rate , Steam supply from the steam turbine at any time , condenser liquid level , Vacuum degree at any moment Input the multi-scale feature extraction model, and the multi-scale feature extraction model outputs feature vectors; train the multi-scale feature extraction model to obtain the trained multi-scale feature extraction model and parameters;

[0055] The specific process is as follows:

[0056] Step 3: 1. Adjust the cooling water flow rate for a window length of 150. Steam supply capacity of a steam turbine with a window length of 150 Condenser liquid level with a window length of 150 Vacuum degree with a window length of 150 Input a long-term feature extraction network, and the long-term feature extraction network outputs a feature vector; capture the long-term evolution features of progressive failures;

[0057] express Time to Cooling water flow rate at any given time;

[0058] express Time to Steam supply from the turbine at any given time;

[0059] express Time to The condenser liquid level at any given time;

[0060] express Time to The degree of vacuum at any given moment;

[0061] Step 3.2: Adjust the cooling water flow rate for a window length of 50. Steam supply capacity of a steam turbine with a window length of 50 Condenser liquid level with a window length of 50 Vacuum degree with a window length of 50 Input a short-term feature extraction network, and the short-term feature extraction network outputs a feature vector; capture the short-term dynamic features of sudden failures;

[0062] express Time to Cooling water flow rate at any given time;

[0063] express Time to Steam supply from the turbine at any given time;

[0064] express Time to The condenser liquid level at any given time;

[0065] express Time to The degree of vacuum at any given moment;

[0066] Step 3: Calculate the historical prediction error with a window length of 20. An input error feedback feature extraction network is used, and the output feature vector is generated by the error feedback feature extraction network; this enhances the model's adaptability to changes in system operating conditions.

[0067] express Time to Historical prediction error at any given time;

[0068] The residual generator obtains historical prediction error signals. ;

[0069] It contains system behaviors that the model cannot explain using normal operating modes, and is the core information carrier for fault detection;

[0070] express Output information of the condenser system at any given time;

[0071] This indicates that the decoder outputs prediction information;

[0072] Initial historical prediction error =0, ;

[0073] Steps 3 and 4: Input the feature vectors output by the long-term feature extraction network, the short-term feature extraction network, and the error feedback feature extraction network into the cross-attention fusion module. The cross-attention fusion module outputs a fused feature vector.

[0074] Step 35: Fuse the feature vectors into the decoder, and the decoder outputs the prediction information. ;

[0075] Step 36: Repeat steps 31 to 35 until the loss function converges, and obtain the trained multi-scale feature extraction model and parameters.

[0076] The other steps and parameters are the same as in Specific Implementation Method 1.

[0077] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One in that: in step three, the cooling water flow rate for a window length of 150 is... Steam supply capacity of a steam turbine with a window length of 150 Condenser liquid level with a window length of 150 Vacuum degree with a window length of 150 Input a long-term feature extraction network, and the long-term feature extraction network outputs a feature vector; capture the long-term evolution features of progressive failures;

[0078] Represented as:

[0079]

[0080] in,

[0081] This represents the feature vector output by the long-term feature extraction network;

[0082] This represents the input information of the long-term feature extraction network;

[0083] Represents the weights of the long-term feature extraction network;

[0084] This represents the bias of the long-term feature extraction network;

[0085] This represents vector multiplication.

[0086] Other steps and parameters are the same as in specific implementation method one or two.

[0087] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that: in step three-two, the cooling water flow rate with a window length of 50 is... Steam supply capacity of a steam turbine with a window length of 50 Condenser liquid level with a window length of 50 Vacuum degree with a window length of 50 Input a short-term feature extraction network, and the short-term feature extraction network outputs a feature vector; capture the short-term dynamic features of sudden failures;

[0088] Represented as:

[0089]

[0090] in,

[0091] This represents the feature vector output by the short-term feature extraction network;

[0092] This represents the input information of the short-term feature extraction network;

[0093] This represents the weights of the short-term feature extraction network;

[0094] This represents the bias of the short-term feature extraction network.

[0095] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0096] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One through Four in that: in step three, the historical prediction error with a window length of 20 is... An input error feedback feature extraction network is used, and the output feature vector is generated by the error feedback feature extraction network; this enhances the model's adaptability to changes in system operating conditions.

[0097] Represented as:

[0098]

[0099] in,

[0100] This represents the feature vector output by the error feedback feature extraction network;

[0101] This represents the input information of the error feedback feature extraction network;

[0102] This represents the weights of the error feedback feature extraction network;

[0103] This represents the bias of the error feedback feature extraction network.

[0104] The other steps and parameters are the same as those in specific implementation methods one through four.

[0105] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One through Five in that: in steps three and four, the feature vectors output by the long-term feature extraction network, the short-term feature extraction network, and the error feedback feature extraction network are input into the cross-attention fusion module, and the cross-attention fusion module outputs a fused feature vector; represented as:

[0106]

[0107] in,

[0108] This indicates that the fork attention fusion module outputs a fused feature vector;

[0109] This represents the feature vector output by the short-term feature extraction network. The feature vector output by the short-term feature extraction network and the feature vector output by the error feedback feature extraction network ;

[0110] This indicates the weights of the fork attention fusion module;

[0111] This indicates the bias of the fork attention fusion module.

[0112] The other steps and parameters are the same as those in specific implementation methods one through five.

[0113] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that: in step three and five, the feature vector is fused and input into the decoder, and the decoder outputs prediction information. ; indicates as:

[0114]

[0115]

[0116]

[0117]

[0118] in,

[0119] , , , These represent the weights of the decoder;

[0120] , , , These represent the decoder bias;

[0121] , , These represent the intermediate results of the decoder;

[0122] express Activation function;

[0123] These represent predicted values ​​for condenser liquid level and vacuum. ;

[0124] This indicates the predicted value of the condenser liquid level; This represents the predicted value of the condenser vacuum level.

[0125] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0126] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that: in step 36, steps 31 through 35 are repeated until the loss function converges, obtaining the trained multi-scale feature extraction model and parameters; the specific process is as follows:

[0127] The loss function uses the weighted mean squared error loss function. , represented as:

[0128]

[0129] in,

[0130] , These represent the weights respectively;

[0131] Indicates mean square error;

[0132] Represents the true value of the condenser liquid level;

[0133] This indicates the predicted value of the condenser liquid level;

[0134] The true value representing the vacuum level of the condenser;

[0135] This represents the predicted value of the condenser vacuum level.

[0136] The long-term window M=150 steps is used to capture progressive faults, the short-term window N=50 steps is used to capture sudden faults, and the feedback window H=20 steps is used for adaptive adjustment.

[0137] Network structure parameters: LSTM hidden layer dimension is set to 256, Transformer encoder layer is 6, multi-head attention head is 8, and feedforward network dimension is 512.

[0138] Training parameters: learning rate set to 0.001, batch size to 32, number of training epochs to 100, and Adam optimizer used.

[0139] The other steps and parameters are the same as those in specific implementation methods one through seven.

[0140] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that: in step five, fault diagnosis is based on online prediction error; the specific process is as follows:

[0141] Step 51: Analyze online prediction errors Perform outlier detection to obtain online data of the condenser system after removing outliers;

[0142] The specific process is as follows:

[0143] A robust outlier detection method based on the absolute deviation of the median is adopted:

[0144]

[0145]

[0146] Represents variables, ;

[0147] express or ;

[0148] Indicates the predicted value of the condenser liquid level. Online prediction error signal compared to the true value of condenser liquid level;

[0149] Indicates the predicted value of condenser vacuum. Online prediction error signal compared to the true value of condenser vacuum;

[0150] like If the value is greater than 1, delete the online prediction error signal. ;

[0151] like Less than or equal to 1, retain the online prediction error signal. ;

[0152] like If the value is greater than 1, delete the online prediction error signal. ;

[0153] like Less than or equal to 1, retain the online prediction error signal. ;

[0154] in,

[0155] Indicates variable 1;

[0156] Indicates taking the median;

[0157] Indicates taking the absolute value;

[0158] Indicates taking the median;

[0159] express The corresponding absolute deviation of the median;

[0160] Indicates variable 2;

[0161] Indicates taking the median;

[0162] Indicates taking the median;

[0163] express The corresponding absolute deviation of the median;

[0164] Step 52: Perform multi-method fusion smoothing on the online prediction error signal retained in Step 51 to obtain the fused online prediction error signal; represented as:

[0165]

[0166] in,

[0167] This represents the error signal after performing a moving average on the online prediction error signal retained in step 51;

[0168] This represents the error signal after Gaussian filtering of the online prediction error signal retained in step 51;

[0169] This represents the error signal after applying Savitzky-Golay filtering to the online prediction error signal retained in step 51.

[0170] express The weights; express The weights; express The weights;

[0171] This represents the fused online prediction error signal;

[0172] Step 53: The fused online prediction error signal obtained in Step 52 is processed by quantile normalization and adaptive thresholding to obtain the final online prediction error signal. ; indicates as:

[0173]

[0174] when If a fault signal is output, the system is considered to be in a normal state; otherwise, it is considered to be in a normal state.

[0175] in,

[0176] Indicates taking 50th percentile;

[0177] Indicates taking The 25th percentile;

[0178] Indicates taking The 75th percentile;

[0179] Indicates taking the median;

[0180] Represents the L2 norm;

[0181] This represents the threshold.

[0182] The other steps and parameters are the same as those in specific implementation methods one through eight.

[0183] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that: the threshold value... Represented as:

[0184]

[0185] in, This represents the percentile constant.

[0186] The other steps and parameters are the same as those in specific implementation methods one through nine.

[0187] Fault type:

[0188] Sensor failure: manifested as slow drift or sudden change in residual signal. Sensor drift is captured by long-term feature network, while sensor failure is detected by short-term feature network.

[0189] Actuator failure: This manifests as a mismatch between control input and system response. Residual mode analysis can be used to identify actuator abnormalities such as valve jamming and pump failure.

[0190] Systematic failures are characterized by the simultaneous abnormality of residuals of multiple output variables. Systematic failures can be detected by analyzing the correlation of multiple variables through a cross-attention mechanism.

[0191] The beneficial effects of the present invention are verified using the following embodiments:

[0192] Example 1:

[0193] The method proposed in this invention is applied to the fault diagnosis and verification of a ship's condenser system. The experiment uses a condenser system simulation platform, with the input variables being cooling water flow rate and turbine steam supply, and the output variables being condenser liquid level and vacuum degree.

[0194] Simulation parameter settings: data sampling frequency 1Hz, total simulation duration 30,000 sampling points, PID closed-loop control is used to maintain stable operation.

[0195] Experimental results show that the system achieves an identification accuracy of R²=0.997, a fault detection accuracy of 95.2%, a false alarm rate of only 2.3%, and a false negative rate of 4.8%. Ablation experiments demonstrate the effectiveness of each component, and multi-scale feature extraction, feedback mechanism, cross-attention fusion, and adaptive threshold strategy all significantly contribute to the overall performance.

[0196] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A fault detection method for a marine power condenser system based on multi-scale feature fusion, characterized in that: The specific process of the method is as follows: Step 1: Obtain the input and output information of the condenser system; The input information of the condenser system at any time is , ; The output information of the condenser system at any time is , ; express Real-time cooling water flow rate; express Steam supply from the steam turbine at any given time; express Continuous condenser liquid level; express Vacuum degree at any moment; Superscript This indicates the transpose; Step 2: Construct a multi-scale feature extraction model; The multi-scale feature extraction model includes a long-term feature extraction network, a short-term feature extraction network, an error feedback feature extraction network, a cross-attention fusion module, a decoder, and a residual generator; The long-term feature extraction network is a Long Short-Term Memory (LSTM) network; The short-term feature extraction network is an encoder architecture based on Transformer; The error feedback feature extraction network is a multilayer perceptron (MLP). The decoder is a multilayer perceptron (MLP). Step 3, Cooling water flow rate , Steam supply from the steam turbine at any time , condenser liquid level , Vacuum degree at any time Input the multi-scale feature extraction model, and the multi-scale feature extraction model outputs feature vectors; train the multi-scale feature extraction model to obtain the trained multi-scale feature extraction model and parameters; Step 4: Collect online data from the condenser system. This online data includes the condenser system's input information at all times. Output information of the condenser system ; The online data from the condenser system is input into the trained multi-scale feature extraction model, which outputs predicted values ​​for the condenser liquid level and vacuum degree. , ; The residual generator is based on the predicted value. Obtain online prediction error signal ; Step 5: Perform fault diagnosis based on online prediction errors; In step three, Cooling water flow rate , Steam supply from the steam turbine at any time , condenser liquid level , Vacuum degree at any time Input the multi-scale feature extraction model, and the multi-scale feature extraction model outputs feature vectors; train the multi-scale feature extraction model to obtain the trained multi-scale feature extraction model and parameters; The specific process is as follows: Step 3:

1. Adjust the cooling water flow rate for a window length of 150. Steam supply capacity of a steam turbine with a window length of 150 Condenser liquid level with a window length of 150 Vacuum degree with a window length of 150 Input the long-term feature extraction network, and the long-term feature extraction network outputs a feature vector; express Time to Cooling water flow rate at any given time; express Time to Steam supply from the turbine at any given time; express Time to The condenser liquid level at any given time; express Time to The degree of vacuum at any given moment; Step 3.2: Adjust the cooling water flow rate for a window length of 50. Steam supply capacity of a steam turbine with a window length of 50 Condenser liquid level with a window length of 50 Vacuum degree with a window length of 50 Input the short-term feature extraction network, and the short-term feature extraction network outputs a feature vector; express Time to Cooling water flow rate at any given time; express Time to Steam supply from the turbine at any given time; express Time to The condenser liquid level at any given time; express Time to The degree of vacuum at any given moment; Step 3: Calculate the historical prediction error with a window length of 20. The input is an error feedback feature extraction network, and the output is a feature vector. express Time to Historical prediction error at any given time; The residual generator obtains historical prediction error signals. ; express Output information of the condenser system at any given time; This indicates that the decoder outputs prediction information; Steps 3 and 4: Input the feature vectors output by the long-term feature extraction network, the short-term feature extraction network, and the error feedback feature extraction network into the cross-attention fusion module. The cross-attention fusion module outputs a fused feature vector. Step 35: Fuse the feature vectors into the decoder, and the decoder outputs the prediction information. ; Step 36: Repeat steps 31 to 35 until the loss function converges, and obtain the trained multi-scale feature extraction model and parameters.

2. The fault detection method for a marine power condenser system based on multi-scale feature fusion according to claim 1, characterized in that: In step three, the cooling water flow rate for a window length of 150 is... Steam supply capacity of a steam turbine with a window length of 150 Condenser liquid level with a window length of 150 Vacuum degree with a window length of 150 Input the long-term feature extraction network, and the long-term feature extraction network outputs a feature vector; Represented as: in, This represents the feature vector output by the long-term feature extraction network; This represents the input information of the long-term feature extraction network; Represents the weights of the long-term feature extraction network; This represents the bias of the long-term feature extraction network; This represents vector multiplication.

3. The fault detection method for a marine power condenser system based on multi-scale feature fusion according to claim 2, characterized in that: In step three, the cooling water flow rate for a window length of 50 is... Steam supply capacity of a steam turbine with a window length of 50 Condenser liquid level with a window length of 50 Vacuum degree with a window length of 50 Input the short-term feature extraction network, and the short-term feature extraction network outputs a feature vector; Represented as: in, This represents the feature vector output by the short-term feature extraction network; This represents the input information of the short-term feature extraction network; This represents the weights of the short-term feature extraction network; This represents the bias of the short-term feature extraction network.

4. The fault detection method for a marine power condenser system based on multi-scale feature fusion according to claim 3, characterized in that: In step three, the historical prediction error with a window length of 20 is used. The input is an error feedback feature extraction network, and the output is a feature vector. Represented as: in, This represents the feature vector output by the error feedback feature extraction network; This represents the input information of the error feedback feature extraction network; This represents the weights of the error feedback feature extraction network; This represents the bias of the error feedback feature extraction network.

5. The fault detection method for a marine power condenser system based on multi-scale feature fusion according to claim 4, characterized in that: In step three, the feature vectors are fused and input into the decoder, and the decoder outputs prediction information. ; indicates as: in, , , , These represent the weights of the decoder; , , , These represent the decoder bias; , , These represent the intermediate results of the decoder; express Activation function; These represent predicted values ​​for condenser liquid level and vacuum. ; This indicates the predicted value of the condenser liquid level; This represents the predicted value of the condenser vacuum level.

6. The fault detection method for a marine power condenser system based on multi-scale feature fusion according to claim 5, characterized in that: In step 36, steps 31 to 35 are repeated until the loss function converges, resulting in the trained multi-scale feature extraction model and its parameters. The specific process is as follows: The loss function uses the weighted mean squared error loss function. , is represented as: in, , These represent the weights respectively; Indicates mean square error; Represents the true value of the condenser liquid level; This indicates the predicted value of the condenser liquid level; The true value representing the vacuum level of the condenser; This represents the predicted value of the condenser vacuum level.

7. A fault detection method for a marine power condenser system based on multi-scale feature fusion according to claim 6, characterized in that: In step five, fault diagnosis is performed based on online prediction errors; The specific process is as follows: Step 51: Analyze online prediction errors Perform outlier detection to obtain online data of the condenser system after removing outliers; The specific process is as follows: Represents variables, ; express or ; Indicates the predicted value of the condenser liquid level. Online prediction error signal compared to the true value of condenser liquid level; Indicates the predicted value of condenser vacuum. Online prediction error signal compared to the true value of condenser vacuum; like If the value is greater than 1, delete the online prediction error signal. ; like Less than or equal to 1, retain the online prediction error signal. ; like If the value is greater than 1, delete the online prediction error signal. ; like Less than or equal to 1, retain the online prediction error signal. ; in, Indicates variable 1; Indicates taking the median; Indicates taking the absolute value; Indicates taking the median; express The corresponding absolute deviation of the median; Indicates variable 2; Indicates taking the median; Indicates taking the median; express The corresponding absolute deviation of the median; Step 52: Perform multi-method fusion smoothing on the online prediction error signal retained in Step 51 to obtain the fused online prediction error signal; represented as: in, This represents the error signal after performing a moving average on the online prediction error signal retained in step 51; This represents the error signal after Gaussian filtering of the online prediction error signal retained in step 51; This represents the error signal after applying Savitzky-Golay filtering to the online prediction error signal retained in step 51. express The weights; express The weights; express The weights; This represents the fused online prediction error signal; Step 53: The fused online prediction error signal obtained in Step 52 is processed by quantile normalization and adaptive thresholding to obtain the final online prediction error signal. ; indicates as: when If a fault signal is output, the system is considered to be in a normal state; otherwise, it is considered to be in a normal state. in, Indicates taking 50th percentile; Indicates taking The 25th percentile; Indicates taking The 75th percentile; Indicates taking the median; Represents the L2 norm; This represents the threshold.

8. The fault detection method for a marine power condenser system based on multi-scale feature fusion according to claim 7, characterized in that: The threshold Represented as: in, This represents the percentile constant.

Images