A system and method for monitoring fault points of heating equipment in a power plant

Abstract

The present invention discloses a system and method for monitoring fault points of heating equipment in a power plant, which solves the problem that existing systems are unable to perform deep multi-dimensional mining based on monitoring data associated with fault points, thereby reducing the accuracy of diagnostic evaluation of the heating system. The method comprises: preprocessing historical data to obtain envelope sample information, and pre-building a fault evaluation model based on a support vector machine and a convolutional neural network; collecting heating equipment fault point monitoring data in real time, preprocessing the heating equipment fault point monitoring data to obtain envelope monitoring information, and monitoring and tracking the obtained envelope monitoring information based on the fault evaluation model to obtain fault point monitoring results. The present invention combines a support vector machine, a residual improvement mechanism, and a lightweight detection structure Slim-Neck to achieve deep multi-dimensional mining of monitoring data without the need for a large number of samples, thereby significantly improving the diagnostic evaluation accuracy of the fault point monitoring data.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment monitoring, and in particular relates to a system and method for monitoring fault points of heating equipment in a power plant. Background Art

[0002] Currently, heat tracing equipment in power plants is primarily used for frost protection, heat preservation, and raising the operating temperature of specific equipment, thereby ensuring stable and safe operation of power generation facilities in cold environments. Heat tracing equipment is a crucial component of ensuring production safety and efficiency in power plants. By providing optimal temperature conditions, it not only prevents damage to pipelines and instruments due to freezing, but also ensures the continuity and reliability of the production process. As a crucial component of ensuring production safety and efficiency, the monitoring and maintenance of heat tracing equipment in power plants is particularly critical.

[0003] Chinese patent CN114040527A discloses an intelligent online monitoring heat tracing system, which includes a hardware system and a software system. The hardware system is used to collect the temperature of the pipeline and then connect the temperature data to a temperature controller with data storage and transmission functions. The temperature controller demodulates the optical signal with information to obtain the distributed temperature along the pipeline. The temperature of the entire pipeline is monitored in real time, and the temperature signals of multiple measuring points over long distances are transmitted to the monitoring end. The configuration screen can simultaneously display the signals of multiple measuring points throughout the pipeline. However, the existing system can only monitor and analyze the heat tracing system based on temperature data, and cannot achieve in-depth multi-dimensional mining of the monitoring data associated with the fault point, thereby reducing the diagnostic evaluation accuracy of the heat tracing system. To address the above problems, we propose a fault point monitoring system and method for heat tracing equipment in power plants. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the existing technology and provide a system and method for monitoring fault points of power plant heating equipment. The system solves the problem that the existing system can only monitor and analyze the heating system based on temperature data, but cannot perform in-depth multi-dimensional mining of the monitoring data associated with the fault point, thereby reducing the accuracy of the diagnosis and evaluation of the heating system.

[0005] Existing systems can only monitor and analyze heating systems based on temperature data, but are unable to conduct in-depth, multi-dimensional mining of monitoring data associated with fault points, thereby reducing the diagnostic and evaluation accuracy of heating systems. To address this issue, we propose a method for monitoring fault points in power plant heating equipment. Briefly, the method includes preprocessing historical data to obtain envelope sample information, dividing the envelope sample information into a training set and a test set, and then pre-building a fault assessment model based on a support vector machine and a convolutional neural network. The fault assessment model is iteratively trained using the training set, and then real-time monitoring data of the heating equipment fault points is collected. The obtained envelope monitoring information is monitored and tracked based on the fault assessment model to obtain fault point monitoring results. In an embodiment of the present invention, a fault assessment model based on a support vector machine and a convolutional neural network is constructed to diagnose and evaluate the fault point monitoring data. This method combines the support vector machine, a residual improvement mechanism, and a lightweight detection structure called Slim-Neck. This method can achieve in-depth, multi-dimensional mining of the monitoring data without requiring a large number of samples, significantly improving the diagnostic and evaluation accuracy of the fault point monitoring data.

[0006] The present invention is implemented as follows: a method for monitoring a failure point of a heat tracing device in a power plant, the method comprising:

[0007] Obtain historical data of heating equipment failure points, pre-process the historical data to obtain envelope sample information, and divide the envelope sample information into a training set and a test set;

[0008] Pre-build a fault assessment model based on support vector machines and convolutional neural networks, iteratively train the fault assessment model using a training set, test the model using a test set, verify the diagnostic effectiveness of the fault assessment model using a model optimizer, and optimize the parameters of the fault assessment model to obtain the optimally configured fault assessment model.

[0009] The monitoring data of the heating equipment fault point is collected in real time, and the monitoring data of the heating equipment fault point is preprocessed to obtain envelope monitoring information. The obtained envelope monitoring information is monitored and tracked based on the fault assessment model to obtain the fault point monitoring result.

[0010] In response to the fault point monitoring result, the fault point coordinate data associated with the fault point monitoring result is identified, the fault point is visually marked using a three-dimensional topological model and a visualization tool, and a three-dimensional visualization model is constructed.

[0011] The method of preprocessing historical data to obtain envelope sample information specifically includes:

[0012] Load historical data and use the smoothing noise reduction method to filter and reduce noise on the historical data to obtain the noise reduction average sequence;

[0013] Among them, when performing filtering and denoising, a symmetrical sliding window with an odd length is set based on machine vision. The symmetrical sliding window moves along the time series direction of the historical data. During the movement, the average value of the current window is calculated as the filtering value, and finally the denoised average sequence is obtained;

[0014] The denoised average sequence is obtained, and the time-frequency conversion of the denoised average sequence is performed based on the Hilbert-Huang transform to obtain a multidimensional feature time-frequency map.

[0015] The method for performing time-frequency conversion on the denoised average sequence based on Hilbert-Huang transform specifically includes:

[0016] Load the noise reduction average sequence, and define the empirical mode decomposition function based on the working current, working voltage, heating temperature, and output power of the heating equipment in the noise reduction average sequence;

[0017] Among them, the empirical mode decomposition function is expressed as:

[0018]

[0019] Where f(x) is the empirical mode decomposition function for the denoised sequence signal x in the denoised average sequence i , λ is the scale parameter of the empirical mode decomposition function, which is used to characterize the abnormal proportion of the working current, working voltage, heating temperature, and output power of the heating equipment in the denoised average sequence, and A is a binary mask used to control the spatial transformation range of the denoised sequence signal to be decomposed, and B is the data dimension of the denoised sequence signal;

[0020]

[0021] Where n is the number of samples in the denoised average sequence, x i A ,x i V ,x i T ,x i G are the instantaneous frequency conversion amplitude of working current, working voltage, heating temperature and output power respectively, q A ,q V ,q T ,q G are the current denoising sequence signal x i The weights of working current, working voltage, heating temperature and output power, and q A ,q V ,q T ,q G The sum is set to 1;

[0022] The time-frequency energy distribution coefficient is obtained based on the inner product of the empirical mode decomposition function and the denoised sequence signal;

[0023] The time-frequency energy distribution coefficient is calculated by formula (3):

[0024]

[0025] is the time-frequency energy distribution coefficient, f(x) is the empirical mode decomposition function for the denoised sequence signal x in the denoised average sequence i The parsed value of

[0026] Obtain the time-frequency energy distribution coefficient and plot the time-frequency energy distribution coefficient into a multi-dimensional feature time-frequency graph.

[0027] The method of pre-building a fault assessment model based on a support vector machine and a convolutional neural network and iteratively training the fault assessment model using a training set specifically includes:

[0028] A pre-built fault assessment model based on support vector machines and convolutional neural networks, which includes an input layer, convolution layer, pooling layer, feature mapping layer, fault diagnosis module, and output layer;

[0029] A custom model optimizer is placed after the fault diagnosis module to verify the diagnostic effect of the fault assessment model and tune its parameters;

[0030] Define the kernel function of the fault assessment model based on the attributes of the training set, select the radial basis function as the model kernel function, and set the hyperparameters of the kernel function;

[0031] Introduce the feature extraction network in the pooling layer, and add the SMO algorithm to the feature extraction network. Based on the SMO algorithm, measure the nonlinear separable training set and define the energy function based on the SMO algorithm:

[0032]

[0033] Among them, e(y i ,t i ) is the energy response value based on the SMO algorithm energy function, y i ,t i is the input multidimensional feature time-frequency graph and the convolution feature vector extracted by the convolution layer, G is the scale parameter of the energy function based on the SMO algorithm, E is the step size of the pooling layer, and are the mean of the multidimensional feature time-frequency graph and the convolution feature vector respectively;

[0034] Based on the residual improvement mechanism and the lightweight detection structure Slim-Neck, the fully connected layer of the fault diagnosis module is improved, and a fault diagnosis module with improved fully connected layers and BN layers is built;

[0035] The energy response value is added to the training set, and the fault assessment model is iteratively trained by alternating training of the feature extraction network and the fault diagnosis module until the SMO algorithm energy function converges or reaches Nash equilibrium.

[0036] The fault assessment model is tested using the test set to determine whether the diagnosis result meets the test accuracy. If so, the fault assessment model is output.

[0037] The method of monitoring and tracking the obtained envelope monitoring information based on the fault assessment model to obtain the fault point monitoring result specifically includes:

[0038] Obtain the pre-processed envelope monitoring information, the input layer identifies the multi-dimensional feature time-frequency graph corresponding to the envelope monitoring information, and transmits the recognition result to the convolution layer;

[0039] The convolution layer identifies the multidimensional feature time-frequency map. The convolution layer extracts the local maximum, minimum, and median values in the multidimensional feature time-frequency map based on the morphological reconstruction method, and performs envelope fitting on the local maximum, minimum, and median values respectively to calculate the average envelope value of the multidimensional feature time-frequency map.

[0040] The pooling layer calculates the energy response value corresponding to the average envelope value of the multidimensional feature time-frequency graph based on the SMO algorithm energy function;

[0041] The energy response value is input into the feature mapping layer, which performs dilation convolution on the energy response value and outputs monitoring feature information;

[0042]

[0043] Among them, T(y) is the output monitoring feature information, e(y i ,t i ) is the energy response value based on the SMO algorithm energy function, M is the number of convolution kernels in the feature map layer, ω i To expand the convolution receptive field;

[0044]

[0045] Among them, ω i is the expanded convolution receptive field, M is the number of convolution kernels in the feature map layer, l is the number of feature map layers, and β is the expansion factor of the feature map layer;

[0046] The monitoring feature information is mapped to the fault diagnosis module, which diagnoses abnormal values of the output information based on the residual improvement mechanism and the lightweight detection structure Slim-Neck.

[0047]

[0048] Among them, W i is the information outlier, T is the bias coefficient of the lightweight detection structure Sl im-Neck, u is the residual connection weight of the residual block in the residual improvement mechanism, and T(y+w) is the result of the monitoring feature information after weight transformation;

[0049]

[0050] Among them, T(y+w) is the result of weight transformation of monitoring feature information, u is the residual connection weight of residual block in residual improvement mechanism, and v is the weight transformation coefficient;

[0051] Preset anomaly threshold to determine whether the information anomaly value exceeds the anomaly threshold. If it exceeds the anomaly threshold, the current fault point is determined to be a fault. The anomaly classification parameter is indexed based on the information anomaly value. The information anomaly values of the fault point are summed based on the cosine similarity scoring function to obtain the fault point monitoring result.

[0052]

[0053] is the fault point monitoring result, sim(W i ,W0) is the similarity scoring function between the information outlier and the anomaly classification parameter, and σ is the index coefficient of the similarity scoring function.

[0054] On the other hand, the present invention also provides a power plant heating equipment failure point monitoring system, the power plant heating equipment failure point monitoring system specifically comprising:

[0055] A data processing module is used to obtain historical data of fault points of the heating equipment, pre-process the historical data to obtain envelope sample information, and divide the envelope sample information into a training set and a test set;

[0056] The model training module pre-builds a fault assessment model based on support vector machines and convolutional neural networks, iteratively trains the fault assessment model using a training set, tests the model using a test set, verifies the diagnostic effectiveness of the fault assessment model using a model optimizer, and optimizes the parameters of the fault assessment model to obtain the optimally configured fault assessment model.

[0057] The monitoring and evaluation module collects the monitoring data of the heating equipment fault point in real time, pre-processes the monitoring data of the heating equipment fault point to obtain envelope monitoring information, monitors and tracks the obtained envelope monitoring information based on the fault assessment model, and obtains the fault point monitoring results;

[0058] The visualization module, in response to the fault point monitoring result, identifies the fault point coordinate data associated with the fault point monitoring result, visually marks the fault point using a three-dimensional topology model and a visualization tool, and builds a three-dimensional visualization model.

[0059] The data processing module includes:

[0060] The noise reduction processing unit loads historical data and uses a smoothing noise reduction method to filter and reduce noise on the historical data to obtain a noise reduction average sequence;

[0061] Among them, when performing filtering and denoising, a symmetrical sliding window with an odd length is set based on machine vision. The symmetrical sliding window moves along the time series direction of the historical data. During the movement, the average value of the current window is calculated as the filtering value, and finally the denoised average sequence is obtained;

[0062] The time-frequency graph output unit obtains the denoised average sequence, performs time-frequency conversion on the denoised average sequence based on the Hilbert-Huang transform, and obtains a multi-dimensional feature time-frequency graph.

[0063] The visualization module includes:

[0064] a fault point identification unit, in response to the fault point monitoring result, identifying fault point coordinate data associated with the fault point monitoring result, wherein the fault point coordinate data is a three-dimensional coordinate parameter;

[0065] A topology marking unit, which visually marks the fault point based on a three-dimensional topology model and a visualization tool;

[0066] The visualization presentation unit obtains the visual marking results of the fault point and builds a three-dimensional visualization model.

[0067] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0068] In an embodiment of the present invention, a fault assessment model based on support vector machines and convolutional neural networks is established to perform diagnostic evaluation on fault point monitoring data, thereby combining support vector machines, residual improvement mechanisms, and lightweight detection structures like Sl im-Neck. This allows for deep multi-dimensional mining of monitoring data without the need for a large number of samples, significantly improving the diagnostic evaluation accuracy of fault point monitoring data.

[0069] In an embodiment of the present invention, the use of a smoothing denoising method to perform filtering and denoising on historical data can effectively achieve denoising of the historical data, thereby ensuring that the reconstructed denoised average sequence can better retain local features and is not affected by time and space scales. The time-frequency conversion of the denoised average sequence based on the Hilbert-Huang transform facilitates the fault assessment model to perform fault analysis on the data, ensuring that the fault assessment model can accurately analyze one-dimensional and multi-dimensional denoised average sequences, and avoiding the problem of dimensional frequency aliasing when traditional convolutional neural networks analyze one-dimensional and multi-dimensional denoised average sequences.

[0070] In the embodiment of the present invention, the denoised average sequence is converted into time-frequency form based on the Hilbert-Huang transform, and the time-frequency energy distribution coefficient is obtained based on the analytical value of the empirical mode decomposition function and the inner product of the denoised sequence signal. This can solve the problems of limited analysis ability of non-stationary signals, the contradiction between time resolution and frequency resolution, and cross-term interference in the time-frequency conversion of existing data, thereby ensuring the evaluation and analysis accuracy of the fault assessment model.

[0071] In an embodiment of the present invention, a fault assessment model is provided. The fault assessment model selects a radial basis function as a model kernel function, so that it can identify and classify nonlinearly separable data, and improves the feature extraction network introduced in the pooling layer. The SMO algorithm is added to the feature extraction network, and the nonlinearly separable training set is measured based on the SMO algorithm, so as to expand the data of the nonlinearly separable data, which significantly improves the diagnostic speed and accuracy of the fault assessment model. At the same time, by introducing a model optimizer to test the diagnostic effect of the fault assessment model and tune the parameters of the fault assessment model, the model convergence speed and model lightweight are guaranteed.

[0072] In the embodiment of the present invention, the fault diagnosis module diagnoses the output information anomalies based on the residual improvement mechanism and the lightweight detection structure Sl im-Neck, and introduces the cosine similarity scoring function to sum the information anomalies of the fault point to obtain the fault point monitoring result, thereby ensuring the distinction of various fault categories based on the monitoring feature information. At the same time, it can also present the potential abnormal risk of the fault point based on the fault point monitoring result, effectively integrating feature extraction and intelligent diagnosis to realize "end-to-end" fault diagnosis and improve the efficiency of heating equipment monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 The present invention is a schematic diagram of the implementation flow of the method for monitoring the failure point of heat tracing equipment in a power plant.

[0074] Figure 2 It is a schematic diagram of the implementation flow of the method for preprocessing historical data to obtain envelope sample information provided by the present invention.

[0075] Figure 3 It is a schematic diagram of the implementation flow of the method for performing time-frequency conversion on a denoised average sequence based on Hilbert-Huang transform provided by the present invention.

[0076] Figure 4 This is a schematic diagram of the implementation process of the method for iteratively training the fault assessment model provided by the present invention using a training set, which is based on a pre-built support vector machine and convolutional neural network.

[0077] Figure 5 It is a schematic diagram of the implementation flow of the method provided by the present invention for monitoring and tracking the obtained envelope monitoring information based on the fault assessment model to obtain the fault point monitoring result.

[0078] Figure 6 It is a schematic diagram of the framework of the power plant heating equipment fault point monitoring system provided by the present invention. DETAILED DESCRIPTION

[0079] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0080] Existing systems can only monitor and analyze heating systems based on temperature data, but are unable to conduct in-depth, multi-dimensional mining of monitoring data associated with fault points, thereby reducing the diagnostic and evaluation accuracy of heating systems. To address this issue, we propose a method for monitoring fault points in power plant heating equipment. Briefly, the method includes preprocessing historical data to obtain envelope sample information, dividing the envelope sample information into a training set and a test set, and then pre-building a fault assessment model based on a support vector machine and a convolutional neural network. The fault assessment model is iteratively trained using the training set, and then real-time monitoring data of the heating equipment fault points is collected. The obtained envelope monitoring information is monitored and tracked based on the fault assessment model to obtain fault point monitoring results. In an embodiment of the present invention, a fault assessment model based on a support vector machine and a convolutional neural network is constructed to diagnose and evaluate the fault point monitoring data. This method combines the support vector machine, a residual improvement mechanism, and a lightweight detection structure called Slim-Neck. This method can achieve in-depth, multi-dimensional mining of the monitoring data without requiring a large number of samples, significantly improving the diagnostic and evaluation accuracy of the fault point monitoring data.

[0081] The embodiment of the present invention provides a method for monitoring the failure point of heat tracing equipment in a power plant. Figure 1 The following is a schematic diagram of the implementation process of the method for monitoring the failure point of the heat tracing equipment in a power plant. The method for monitoring the failure point of the heat tracing equipment in a power plant specifically includes:

[0082] Step S10: Obtain historical data of the heating equipment fault point, pre-process the historical data to obtain envelope sample information, and divide the envelope sample information into a training set and a test set. In an embodiment of the present invention, the ratio of the training set to the test set can be 7:3. The historical data of the fault point can be obtained through on-site data collection (direct monitoring, regular testing), heating equipment fault simulation experiments, fault simulation systems, or professional websites. The historical data of the fault point includes the operating current, operating voltage, heating temperature, output power, operating time, speed, load condition, vibration amplitude, frequency, sound frequency, and intensity of the fault point, and also includes a fault label corresponding to the fault point, where the fault label includes a short circuit fault of the heating equipment, a ground fault, a short circuit fault, a switch trip, abnormal heating value, and an abnormal insulation layer.

[0083] It should be noted that heating equipment includes but is not limited to low-temperature self-limiting temperature electric heating tapes, parallel constant power electric heating tapes, heating cables, and steam heating equipment, and equipment failure points can be understood as specific locations or parameters for real-time or regular monitoring and detection of the operating status of power plant pipelines, facilities, and cables.

[0084] Step S20: pre-build a fault assessment model based on a support vector machine and a convolutional neural network, iteratively train the fault assessment model using a training set, test the fault assessment model using a test set, verify the diagnostic effect of the fault assessment model using a model optimizer, and perform parameter tuning on the fault assessment model to obtain an optimally configured fault assessment model.

[0085] Step S30: real-time acquisition of monitoring data of the heating equipment fault point, pre-processing of the monitoring data of the heating equipment fault point to obtain envelope monitoring information, monitoring and tracking of the obtained envelope monitoring information based on the fault assessment model, and obtaining a fault point monitoring result.

[0086] It should be noted that the fault point monitoring data includes but is not limited to the operating current, operating voltage, heating temperature, output power, operating time, speed, load conditions, vibration amplitude, frequency, sound frequency, and intensity of the fault point.

[0087] Step S40 , in response to the fault point monitoring result, identifying the fault point coordinate data associated with the fault point monitoring result, visually marking the fault point using a three-dimensional topology model and a visualization tool, and building a three-dimensional visualization model.

[0088] In an embodiment of the present invention, a fault assessment model based on support vector machines and convolutional neural networks is established to perform diagnostic evaluation on fault point monitoring data, thereby combining support vector machines, residual improvement mechanisms, and lightweight detection structures like Sl im-Neck. This allows for deep multi-dimensional mining of monitoring data without the need for a large number of samples, significantly improving the diagnostic evaluation accuracy of fault point monitoring data.

[0089] The embodiment of the present invention provides a method for preprocessing historical data to obtain envelope sample information. Figure 2 The following is a schematic diagram of a method for preprocessing historical data to obtain envelope sample information. The method for preprocessing historical data to obtain envelope sample information specifically includes:

[0090] Step S101: load historical data, and use a smoothing noise reduction method to filter and reduce noise on the historical data to obtain a noise reduction average sequence;

[0091] Among them, when performing filtering and denoising, a symmetrical sliding window with an odd length is set based on machine vision. The symmetrical sliding window moves along the time series direction of the historical data. During the movement, the average value of the current window is calculated as the filtering value, and finally the denoised average sequence is obtained;

[0092] It should be noted that when calculating the average value of the current window as the filter value, the least squares method is used to determine the sliding window weight coefficient. SG filtering is then used to construct a noise reduction matrix. The noise reduction matrix is subjected to singular value decomposition, and multiple groups of singular values are accumulated and divided by the window size to obtain the average value of the current window. In this embodiment of the present invention, the signal-to-noise ratio is used to evaluate the effect of the filtering noise reduction process. In this embodiment, the signal-to-noise ratio is 100.15, indicating that the noise reduction method described above is effective.

[0093] Step S102: Obtain a denoised average sequence, and perform time-frequency conversion on the denoised average sequence based on Hilbert-Huang transform to obtain a multi-dimensional feature time-frequency graph.

[0094] In an embodiment of the present invention, the use of a smoothing denoising method to perform filtering and denoising on historical data can effectively achieve denoising of the historical data, thereby ensuring that the reconstructed denoised average sequence can better retain local features and is not affected by time and space scales. The time-frequency conversion of the denoised average sequence based on the Hilbert-Huang transform facilitates the fault assessment model to perform fault analysis on the data, ensuring that the fault assessment model can accurately analyze one-dimensional and multi-dimensional denoised average sequences, and avoiding the problem of dimensional frequency aliasing when traditional convolutional neural networks analyze one-dimensional and multi-dimensional denoised average sequences.

[0095] The embodiment of the present invention provides a method for performing time-frequency conversion on a denoised average sequence based on Hilbert-Huang transform. Figure 3 The figure shows a schematic diagram of the implementation process of the method for performing time-frequency conversion on a denoised average sequence based on the Hilbert-Huang transform. The method for performing time-frequency conversion on a denoised average sequence based on the Hilbert-Huang transform specifically includes:

[0096] Step S1021: Load the noise reduction average sequence, and define an empirical mode decomposition function based on the operating current, operating voltage, heating temperature, and output power of the heating equipment in the noise reduction average sequence;

[0097] Among them, the empirical mode decomposition function is expressed as:

[0098]

[0099] Where f(x) is the empirical mode decomposition function for the denoised sequence signal x in the denoised average sequence i , λ is the scale parameter of the empirical mode decomposition function, which is used to characterize the abnormal proportion of the working current, working voltage, heating temperature, and output power of the heating equipment in the denoised average sequence. In this embodiment, the scale parameter of the empirical mode decomposition function is set to 0.1-1, and A is a binary mask used to control the spatial transformation range of the denoised sequence signal to be decomposed, and B is the data dimension of the denoised sequence signal. In this embodiment, the data dimension of the denoised sequence signal is 1-5;

[0100]

[0101] Where n is the number of samples in the denoised average sequence, x i A ,x i V ,x i T ,x i G are the instantaneous frequency conversion amplitude of working current, working voltage, heating temperature and output power respectively, q A ,q V ,q T ,q G are the current denoising sequence signal x i The weights of working current, working voltage, heating temperature and output power, and q A ,q V ,q T ,q G The sum is set to 1;

[0102] Step S1022, obtaining a time-frequency energy distribution coefficient based on the inner product of the empirical mode decomposition function analysis value and the noise reduction sequence signal;

[0103] The time-frequency energy distribution coefficient is calculated by formula (3):

[0104]

[0105] is the time-frequency energy distribution coefficient, f(x) is the empirical mode decomposition function for the denoised sequence signal x in the denoised average sequence i The parsed value of

[0106] Step S1023: Obtain the time-frequency energy distribution coefficients, and plot the time-frequency energy distribution coefficients into a multi-dimensional feature time-frequency diagram.

[0107] In the embodiment of the present invention, the denoised average sequence is converted into time-frequency form based on the Hilbert-Huang transform, and the time-frequency energy distribution coefficient is obtained based on the analytical value of the empirical mode decomposition function and the inner product of the denoised sequence signal. This can solve the problems of limited analysis ability of non-stationary signals, the contradiction between time resolution and frequency resolution, and cross-term interference in the time-frequency conversion of existing data, thereby ensuring the evaluation and analysis accuracy of the fault assessment model.

[0108] The embodiment of the present invention provides a method for pre-building a fault assessment model based on support vector machines and convolutional neural networks, and iteratively training the fault assessment model using a training set. Figure 4 The figure shows a schematic diagram of the implementation process of the method for iteratively training the fault assessment model using a training set by pre-building a fault assessment model based on a support vector machine and a convolutional neural network. The method for iteratively training the fault assessment model using a training set specifically includes:

[0109] Step S201: pre-build a fault assessment model based on support vector machine and convolutional neural network, wherein the fault assessment model includes an input layer, a convolution layer, a pooling layer, a feature mapping layer, a fault diagnosis module and an output layer.

[0110] In an embodiment of the present invention, the convolution layer is provided with three layers of convolution, including a first convolution, a second convolution, and a third convolution. The first convolution includes 32 64*1 convolution kernels, a step size of 32, and 64 channels. The second convolution includes 64 128*1 convolution kernels, a step size of 64, and 128 channels. The third convolution includes 64 3*1 convolution kernels, a step size of 2, and 256 channels. The convolution layer uses an all-pass filter to extract data features. The feature mapping layer is composed of a Concat layer, a VoVGSCSP layer, and a GSConv layer, and the number of channels of the feature mapping layer is 512.

[0111] Step S202: Customize the model optimizer and place it after the fault diagnosis module to verify the diagnostic effect of the fault assessment model and perform parameter tuning on the fault assessment model.

[0112] The model optimizer consists of the Adam optimizer, which uses the Adam optimizer to update the model gradient and perform bias correction. After bias correction, the model hyperparameters are updated based on the iterative learning efficiency. When testing the diagnostic effect of the fault assessment model, the MES mean square error is introduced to calculate the expected value between the diagnostic effect and the preset effect, and to determine whether the expected value is greater than the preset expected threshold. If it is greater than the preset expected threshold, it means that the surface model diagnostic effect is good.

[0113] Step S203: defining a kernel function of a fault assessment model based on the attributes of the training set, selecting a radial basis function as the kernel function of the model, and setting hyperparameters of the kernel function;

[0114] Step S204: introduce a feature extraction network into the pooling layer, add the SMO algorithm to the feature extraction network, measure the nonlinear separable training set based on the SMO algorithm, and define an energy function based on the SMO algorithm:

[0115]

[0116] Among them, e(y i ,t i ) is the energy response value based on the SMO algorithm energy function, y i ,t i is the input multidimensional feature time-frequency graph and the convolution feature vector extracted by the convolution layer, G is the scale parameter of the energy function based on the SMO algorithm. In this embodiment, the scale parameter is set to 0.3-0.8, E is the step size of the pooling layer, and in this embodiment, the step size can be 2-16, and are the mean of the multidimensional feature time-frequency graph and the convolution feature vector respectively;

[0117] Step S205: Based on the residual improvement mechanism and the lightweight detection structure Slim-Neck, the fully connected layer of the fault diagnosis module is improved, and a fault diagnosis module with improved fully connected layer and BN layer is constructed;

[0118] In step S206, the energy response value is added to the training set, and the fault assessment model is iteratively trained by alternating training of the feature extraction network and the fault diagnosis module until the SMO algorithm energy function converges or reaches Nash equilibrium; it should be noted that the number of iterative training times is 100-1000 times, and the learning rate is set to 0.001.

[0119] Step S207: Test the fault assessment model using the test set to determine whether the diagnosis result meets the test accuracy. In this embodiment, the test accuracy can be set to 0.9-0.95.

[0120] Step S208: If the test accuracy is met, the fault assessment model is output; if the test accuracy is not met, the process returns to step S202.

[0121] In an embodiment of the present invention, a fault assessment model is provided. The fault assessment model selects a radial basis function as a model kernel function, so that it can identify and classify nonlinearly separable data, and improves the feature extraction network introduced in the pooling layer. The SMO algorithm is added to the feature extraction network, and the nonlinearly separable training set is measured based on the SMO algorithm, so as to expand the data of the nonlinearly separable data, which significantly improves the diagnostic speed and accuracy of the fault assessment model. At the same time, by introducing a model optimizer to test the diagnostic effect of the fault assessment model and tune the parameters of the fault assessment model, the model convergence speed and model lightweight are guaranteed.

[0122] The embodiment of the present invention provides a method for monitoring and tracking the obtained envelope monitoring information based on the fault assessment model to obtain the fault point monitoring result. Figure 5 A schematic diagram of the implementation flow of the method for monitoring and tracking the obtained envelope monitoring information based on the fault assessment model to obtain the fault point monitoring result is shown. The method for monitoring and tracking the obtained envelope monitoring information based on the fault assessment model to obtain the fault point monitoring result specifically includes:

[0123] Step S301: obtaining pre-processed envelope monitoring information, the input layer identifies the multi-dimensional feature time-frequency graph corresponding to the envelope monitoring information, and transmits the identification result to the convolution layer;

[0124] It should be noted that the envelope monitoring information preprocessing method is similar to preprocessing historical data to obtain envelope sample information. By preprocessing the envelope monitoring information, it can be ensured that the information is uniformly converted into a standard time-frequency diagram.

[0125] Step S302: The convolution layer identifies the multidimensional feature time-frequency graph. The convolution layer extracts the local maximum, minimum, and median values in the multidimensional feature time-frequency graph based on the morphological reconstruction method, and performs envelope fitting on the local maximum, minimum, and median values to calculate the average envelope value of the multidimensional feature time-frequency graph.

[0126] Step S303: The pooling layer calculates the energy response value corresponding to the average envelope value of the multidimensional feature time-frequency graph based on the SMO algorithm energy function;

[0127] Step S304: input the energy response value into the feature mapping layer, and the feature mapping layer performs dilation convolution on the energy response value to output monitoring feature information;

[0128]

[0129] Among them, T(y) is the output monitoring feature information, e(y i ,t i ) is the energy response value based on the SMO algorithm energy function, M is the number of convolution kernels in the feature map layer, ω i To expand the convolution receptive field;

[0130]

[0131] Among them, ω i To expand the convolution receptive field, M is the number of convolution kernels in the feature mapping layer. In this embodiment, the number of convolution kernels in the feature mapping layer is 16-256, l is the number of layers of the feature mapping layer. In this embodiment, the number of layers of the feature mapping layer is 3, β is the expansion factor of the feature mapping layer, and the expansion factor of the feature mapping layer is 1 or 1.5.

[0132] Step S305: Mapping the monitoring feature information to a fault diagnosis module, which diagnoses abnormal values of the output information based on a residual improvement mechanism and a lightweight detection structure Slim-Neck.

[0133]

[0134] Among them, W i is the information outlier, T is the bias coefficient of the lightweight detection structure Sl im-Neck. In this embodiment, the bias coefficient of the lightweight detection structure Sl im-Neck can be 0.2-0.3, u is the residual connection weight of the residual block in the residual improvement mechanism, the residual connection weight can be 0.1-1, and T(y+w) is the result of the monitoring feature information after weight transformation;

[0135]

[0136] Among them, T(y+w) is the result of weight transformation of monitoring feature information, u is the residual connection weight of the residual block in the residual improvement mechanism, v is the weight transformation coefficient, and the weight transformation coefficient can be 0.6-0.9; it should be noted that the information anomaly value is used to characterize the fault point detection status, and the information anomaly value is composed of anomaly vector and fault type label.

[0137] Step S306: Preset an abnormality threshold and determine whether the information abnormal value exceeds the abnormality threshold. If it exceeds the abnormality threshold, the current fault point is determined to be a fault. The abnormal classification parameter is indexed based on the information abnormal value. The information abnormal values of the fault point are summed based on the cosine similarity scoring function to obtain the fault point monitoring result.

[0138]

[0139] is the fault point monitoring result, sim(W i ,W0) is the similarity scoring function between the information outlier and the anomaly classification parameter, σ is the index coefficient of the similarity scoring function. In this embodiment, the index coefficient of the similarity scoring function can be 0.02-0.08.

[0140] In this embodiment, the abnormality threshold can be set to 0.6-0.8. When the abnormality classification parameters are indexed by the information abnormality value, the abnormality classification parameters are indexed with the fault type label. The abnormality classification parameters include but are not limited to short circuit faults of heating equipment, grounding faults, short circuit faults, switch tripping, abnormal heating, abnormal insulation layer, etc.

[0141] In the embodiment of the present invention, the fault diagnosis module diagnoses the output information anomalies based on the residual improvement mechanism and the lightweight detection structure Sl im-Neck, and introduces the cosine similarity scoring function to sum the information anomalies of the fault point to obtain the fault point monitoring result, thereby ensuring the distinction of various fault categories based on the monitoring feature information. At the same time, it can also present the potential abnormal risk of the fault point based on the fault point monitoring result, effectively integrating feature extraction and intelligent diagnosis to realize "end-to-end" fault diagnosis and improve the efficiency of heating equipment monitoring.

[0142] On the other hand, an embodiment of the present invention further provides a power plant heating equipment fault point monitoring system. Figure 6 The schematic diagram of the framework of the power plant heating equipment failure point monitoring system is shown. The power plant heating equipment failure point monitoring system specifically includes:

[0143] The data processing module 100 is used to obtain historical data of the heat tracing equipment failure point, pre-process the historical data to obtain envelope sample information, and divide the envelope sample information into a training set and a test set;

[0144] The model training module 200 pre-builds a fault assessment model based on a support vector machine and a convolutional neural network, iteratively trains the fault assessment model using a training set, tests the fault assessment model using a test set, verifies the diagnostic effect of the fault assessment model using a model optimizer, and optimizes the parameters of the fault assessment model to obtain an optimally configured fault assessment model.

[0145] The monitoring and evaluation module 300 collects the monitoring data of the heating equipment fault point in real time, pre-processes the monitoring data of the heating equipment fault point to obtain envelope monitoring information, monitors and tracks the obtained envelope monitoring information based on the fault assessment model, and obtains the fault point monitoring result;

[0146] The visualization module 400 , in response to the fault point monitoring result, identifies the fault point coordinate data associated with the fault point monitoring result, visually marks the fault point using a three-dimensional topology model and a visualization tool, and builds a three-dimensional visualization model.

[0147] It should be noted that the data processing module 100, the model training module 200, the monitoring and evaluation module 300, and the visualization module 400 are connected via a local area network and Bluetooth communication. The system also includes a background server and a background memory. The processor executes various functional applications and data processing of the server by running non-volatile software programs, instructions and modules stored in the memory, thereby realizing the power plant heating equipment fault point monitoring method of the above method embodiment.

[0148] In this embodiment, the data processing module 100 includes:

[0149] The noise reduction processing unit 110 loads historical data and performs filtering and noise reduction processing on the historical data using a smoothing noise reduction method to obtain a noise reduction average sequence;

[0150] Among them, when performing filtering and denoising, a symmetrical sliding window with an odd length is set based on machine vision. The symmetrical sliding window moves along the time series direction of the historical data. During the movement, the average value of several current windows is used as the filtering value, and finally the denoised average sequence is obtained;

[0151] The time-frequency map output unit 120 obtains a denoised average sequence, performs time-frequency conversion on the denoised average sequence based on Hilbert-Huang transform, and obtains a multi-dimensional feature time-frequency map.

[0152] It should be noted that the visualization module 400 includes:

[0153] The fault point identification unit 410 identifies the fault point coordinate data associated with the fault point monitoring result in response to the fault point monitoring result, wherein the fault point coordinate data is a three-dimensional coordinate parameter;

[0154] A topology marking unit 420 is configured to visually mark the fault point based on a three-dimensional topology model and a visualization tool;

[0155] The visualization presentation unit 430 obtains the visual marking result of the fault point and builds a three-dimensional visualization model.

[0156] It should be noted that the visualization tool can be a three-dimensional modeling tool, which includes but is not limited to 3ds Max, Maya, Cinema 4D, etc., and the three-dimensional topology model can be TorchSummary, Graphviz and Torchviz, Feature Importance. The three-dimensional topology model can be used to perform first-level, second-level, third-level, fourth-level, and fifth-level visual markings on the fault point. The higher the level of the visual marking, the more serious the fault of the fault point, and the first-level and second-level visual markings indicate that there is a potential risk at the current fault point.

[0157] In summary, the present invention provides a system and method for monitoring fault points of heating equipment in a power plant. In an embodiment of the present invention, a fault assessment model based on a support vector machine and a convolutional neural network is established to perform diagnostic evaluation on the fault point monitoring data, thereby combining the support vector machine, the residual improvement mechanism, and the lightweight detection structure Slim-Neck. This allows for deep multi-dimensional mining of the monitoring data without the need for a large number of samples, significantly improving the diagnostic evaluation accuracy of the fault point monitoring data.

Claims

1. A method for monitoring fault points of heat tracing equipment in a power plant, characterized in that: include: Obtain historical data of heating equipment failure points, pre-process the historical data to obtain envelope sample information, and divide the envelope sample information into a training set and a test set; The smoothing noise reduction method is used to filter and reduce noise on historical data to obtain the noise reduction average sequence; Based on the Hilbert-Huang transform, the denoised average sequence is converted into a time-frequency image to obtain a multi-dimensional feature time-frequency image. Pre-build a fault assessment model based on support vector machines and convolutional neural networks, iteratively train the fault assessment model using a training set, test the model using a test set, verify the diagnostic effectiveness of the fault assessment model using a model optimizer, and optimize the parameters of the fault assessment model to obtain the optimally configured fault assessment model. The fault assessment model includes an input layer, a convolutional layer, a pooling layer, a feature mapping layer, a fault diagnosis module, and an output layer. A custom model optimizer is used to introduce a feature extraction network into the pooling layer and add the SMO algorithm to the feature extraction network. The SMO algorithm is used to measure nonlinear separable training sets, and an energy function based on the SMO algorithm is defined. The fully connected layer of the fault diagnosis module is improved based on the residual improvement mechanism and the lightweight detection structure Slim-Neck. Finally, a fault diagnosis module with improved fully connected layers and BN layers is constructed. Real-time collection of heating equipment fault point monitoring data, pre-processing of the heating equipment fault point monitoring data to obtain envelope monitoring information, monitoring and tracking of the obtained envelope monitoring information based on the fault assessment model, and obtaining the fault point monitoring results; When the fault point monitoring results are obtained, the preprocessed envelope monitoring information is obtained, the input layer identifies the multidimensional feature time-frequency graph corresponding to the envelope monitoring information, and calculates the average envelope value of the multidimensional feature time-frequency graph; the pooling layer calculates the energy response value corresponding to the average envelope value of the multidimensional feature time-frequency graph based on the SMO algorithm energy function; and outputs the monitoring feature information.

2. The method for monitoring a fault point of heat tracing equipment in a power plant according to claim 1, wherein: The method further comprises: In response to the fault point monitoring result, the fault point coordinate data associated with the fault point monitoring result is identified, the fault point is visually marked using a three-dimensional topological model and a visualization tool, and a three-dimensional visualization model is constructed.

3. The method for monitoring a fault point of heat tracing equipment in a power plant according to claim 1, wherein: The method of preprocessing historical data to obtain envelope sample information specifically includes: Load historical data; Among them, when performing filtering and denoising, a symmetrical sliding window with an odd length is set based on machine vision. The symmetrical sliding window moves along the time series direction of the historical data. During the movement, the average value of the current window is calculated as the filtering value, and finally the denoised average sequence is obtained; The denoised average sequence is obtained, and the time-frequency conversion of the denoised average sequence is performed based on the Hilbert-Huang transform to obtain a multidimensional feature time-frequency graph.

4. The method for monitoring a fault point of heat tracing equipment in a power plant according to claim 3, wherein: The method for performing time-frequency conversion on the denoised average sequence based on the Hilbert-Huang transform specifically includes: Load the noise reduction average sequence, and define the empirical mode decomposition function based on the working current, working voltage, heating temperature, and output power of the heating equipment in the noise reduction average sequence; Among them, the empirical mode decomposition function is expressed as: （1） in, The empirical mode decomposition function is used to de-noise the denoised sequence signal in the denoised average sequence The analytical value of is the scale parameter of the empirical mode decomposition function, which is used to characterize the abnormal proportion of the working current, working voltage, heating temperature and output power of the heating equipment in the noise reduction average sequence, and is a binary mask used to control the spatial transformation range of the denoised sequence signal that needs to be decomposed, and is the data dimension of the denoised sequence signal; （2） in, is the number of samples in the denoised average sequence, are the instantaneous frequency conversion amplitudes of working current, working voltage, heating temperature and output power respectively, are the current denoising sequence signals respectively The weights of working current, working voltage, heating temperature and output power, and The sum is set to 1; The time-frequency energy distribution coefficient is obtained based on the inner product of the empirical mode decomposition function and the denoised sequence signal; Among them, the time-frequency energy distribution coefficient is calculated by formula (3): （3） is the time-frequency energy distribution coefficient, The empirical mode decomposition function is used to de-noise the denoised sequence signal in the denoised average sequence The parsed value of Obtain the time-frequency energy distribution coefficient and plot the time-frequency energy distribution coefficient into a multi-dimensional feature time-frequency graph.

5. The method for monitoring a fault point of heat tracing equipment in a power plant according to claim 1, wherein: The method of pre-building a fault assessment model based on a support vector machine and a convolutional neural network and iteratively training the fault assessment model using a training set specifically includes: Pre-built fault assessment model based on support vector machine and convolutional neural network; The model optimizer is placed after the fault diagnosis module to test the diagnostic effect of the fault assessment model and tune the parameters of the fault assessment model; Define the kernel function of the fault assessment model based on the attributes of the training set, select the radial basis function as the model kernel function, and set the hyperparameters of the kernel function; （4） in, is the energy response value based on the SMO algorithm energy function, It is the multi-dimensional feature time-frequency map of the input and the convolution feature vector extracted by the convolution layer. is the scale parameter of the energy function based on the SMO algorithm, is the stride of the pooling layer, and are the mean of the multidimensional feature time-frequency graph and the convolution feature vector respectively; The energy response value is added to the training set, and the fault assessment model is iteratively trained by alternating training of the feature extraction network and the fault diagnosis module until the SMO algorithm energy function converges or reaches Nash equilibrium. The fault assessment model is tested using the test set to determine whether the diagnosis result meets the test accuracy. If so, the fault assessment model is output.

6. The method for monitoring a fault point of heat tracing equipment in a power plant according to claim 5, wherein: The method of monitoring and tracking the obtained envelope monitoring information based on the fault assessment model to obtain the fault point monitoring result specifically includes: Obtain the pre-processed envelope monitoring information and transmit the recognition results to the convolution layer; The convolution layer identifies the multidimensional feature time-frequency graph. The convolution layer extracts the local maximum, minimum, and median values in the multidimensional feature time-frequency graph based on the morphological reconstruction method, and performs envelope fitting on the local maximum, minimum, and median values respectively; The energy response value is input into the feature mapping layer, which performs dilation convolution on the energy response value and outputs monitoring feature information; （5） in, is the output monitoring feature information, is the energy response value based on the SMO algorithm energy function, is the number of convolution kernels in the feature map layer, To expand the convolution receptive field; （6） in, To expand the convolution receptive field, is the number of convolution kernels in the feature map layer, is the number of feature mapping layers, is the dilation factor of the feature map layer.

7. The method for monitoring a fault point of heat tracing equipment in a power plant according to claim 6, wherein: The method for monitoring and tracking the obtained envelope monitoring information based on the fault assessment model to obtain the fault point monitoring result specifically further includes: The monitoring feature information is mapped to the fault diagnosis module, which uses the residual improvement mechanism and the lightweight detection structure Slim-Neck to diagnose abnormal values of the output information; （7） in, is an information outlier, is the bias coefficient of the lightweight detection structure Slim-Neck, is the residual connection weight of the residual block in the residual improvement mechanism, It is the result of weight transformation of monitoring feature information; （8） in, To monitor the result of feature information after weight transformation, is the residual connection weight of the residual block in the residual improvement mechanism, is the weight transformation coefficient; Preset anomaly threshold to determine whether the information anomaly value exceeds the anomaly threshold. If it exceeds the anomaly threshold, the current fault point is determined to be a fault. The anomaly classification parameter is indexed based on the information anomaly value. The information anomaly values of the fault point are summed based on the cosine similarity scoring function to obtain the fault point monitoring result. （9） is the fault point monitoring result, is the similarity scoring function between information outliers and anomaly classification parameters, is the index coefficient of the similarity scoring function.

8. A power plant heating equipment failure point monitoring system, configured to implement a power plant heating equipment failure point monitoring method according to any one of claims 1 to 7, characterized in that: The power plant heating equipment fault point monitoring system specifically includes: A data processing module is used to obtain historical data of fault points of the heating equipment, pre-process the historical data to obtain envelope sample information, and divide the envelope sample information into a training set and a test set; The model training module pre-builds a fault assessment model based on support vector machines and convolutional neural networks, iteratively trains the fault assessment model using a training set, tests the model using a test set, verifies the diagnostic effectiveness of the fault assessment model using a model optimizer, and optimizes the parameters of the fault assessment model to obtain the optimally configured fault assessment model. The monitoring and evaluation module collects the monitoring data of the heating equipment fault point in real time, pre-processes the monitoring data of the heating equipment fault point to obtain envelope monitoring information, monitors and tracks the obtained envelope monitoring information based on the fault assessment model, and obtains the fault point monitoring results; The visualization module, in response to the fault point monitoring result, identifies the fault point coordinate data associated with the fault point monitoring result, visually marks the fault point using a three-dimensional topology model and a visualization tool, and builds a three-dimensional visualization model.

9. The power plant heat tracing equipment fault point monitoring system according to claim 8, characterized in that: The data processing module includes: The noise reduction processing unit loads historical data and uses a smoothing noise reduction method to filter and reduce noise on the historical data to obtain a noise reduction average sequence; Among them, when performing filtering and denoising, a symmetrical sliding window with an odd length is set based on machine vision. The symmetrical sliding window moves along the time series direction of the historical data. During the movement, the average value of the current window is calculated as the filtering value, and finally the denoised average sequence is obtained; The time-frequency graph output unit obtains the denoised average sequence, performs time-frequency conversion on the denoised average sequence based on the Hilbert-Huang transform, and obtains a multidimensional feature time-frequency graph.

10. The power plant heat tracing equipment fault point monitoring system according to claim 8, characterized in that: The visualization module includes: a fault point identification unit, in response to the fault point monitoring result, identifying fault point coordinate data associated with the fault point monitoring result, wherein the fault point coordinate data is a three-dimensional coordinate parameter; A topology marking unit, which visually marks the fault point based on a three-dimensional topology model and a visualization tool; The visualization presentation unit obtains the visual marking results of the fault point and builds a three-dimensional visualization model.

Images