Power plant early warning method and system based on neural network of migration learning

Abstract

The application discloses a power plant early warning method and system based on a neural network of transfer learning. The method comprises the following steps: acquiring source domain samples and target domain samples; training a neural network through a transfer learning method according to the source domain samples and the target domain samples, and acquiring the trained neural network; collecting video images around dangerous equipment of a power plant in real time and preprocessing the video images; inputting the preprocessed video images into the trained neural network, obtaining an early warning result and feeding back the early warning result to a terminal to give a prompt. The application adopts a transfer learning method with low inter-domain distribution difference as a core neural network, effectively reduces the distribution difference between source data and target data of a power plant alarm system, and improves the performance of a power plant early warning system.

Description

Technical Field

[0001] This invention relates to the field of thermal power plant safety technology, and more specifically, to a method and system for early warning of thermal power plants based on transfer learning neural networks. Background Technology

[0002] A thermal power plant, or coal-fired power plant for short, is a factory that uses combustible materials as fuel to produce electricity. Its basic production process is as follows: when fuel is burned, it heats water to generate steam, converting the chemical energy of the fuel into heat energy. The steam pressure drives the turbine to rotate, converting the heat energy into mechanical energy. Then, the turbine drives the generator to rotate, converting the mechanical energy into electrical energy.

[0003] In existing technologies, thermal power plants set up safety fences around important hazardous equipment (high pressure, high temperature, rotation, etc.) for isolation and warning. However, safety fences only have a simple isolation function for dangerous areas. When someone approaches or enters the area, they cannot provide timely warnings or stop the operation of the hazardous equipment in the thermal power plant in time, which poses a great safety hazard.

[0004] Due to the unique nature of thermal power plants, it is difficult to collect sufficient sample data. Therefore, the recently popular transfer learning method can be used to design early warning systems for thermal power plants. However, the complex working environment of thermal power plant alarm systems causes the assumption that the samples in the source and target domains of transfer learning follow the same distribution to no longer hold, which seriously affects the performance of thermal power plant alarm systems.

[0005] Therefore, it is necessary to develop an early warning method and system for thermal power plants based on transfer learning neural networks.

[0006] The information disclosed in the background section of this invention is intended only to enhance the understanding of the general background of this invention, and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] This invention proposes a method and system for early warning of thermal power plants based on transfer learning neural networks. It adopts a transfer learning method with low inter-domain distribution differences as the core neural network, which effectively reduces the distribution differences between source data and target data in the alarm system of thermal power plants and improves the performance of the early warning system.

[0008] In a first aspect, embodiments of this disclosure provide a method for early warning of thermal power plants based on transfer learning neural networks, including:

[0009] Obtain source domain samples and target domain samples;

[0010] The neural network is trained using transfer learning based on the source domain samples and the target domain samples to obtain the trained neural network.

[0011] Real-time acquisition and preprocessing of video images around hazardous equipment in thermal power plants;

[0012] The preprocessed video image is input into the trained neural network to obtain an early warning result, which is then fed back to the terminal for prompting.

[0013] Preferably, the neural network includes an encoder, a decoder, and a classifier.

[0014] Preferably, training the neural network using transfer learning based on the source domain samples and the target domain samples includes:

[0015] The source domain samples and the target domain samples are input into the neural network for iterative training. Each training iteration yields the corresponding sample reconstruction error, warning error, and sample distribution deviation.

[0016] The total error is calculated based on the sample reconstruction error, the warning error, and the sample distribution deviation.

[0017] The parameters of the current neural network are adjusted based on the total error until the total error is less than a set threshold, at which point training is complete.

[0018] Preferably, obtaining the corresponding sample reconstruction error and sample distribution deviation includes:

[0019] The source domain sample and the target domain sample are input into the encoder to obtain the corresponding source data features and target data features;

[0020] The source data features and the target data features are input into the decoder to obtain reconstructed source data and reconstructed target data;

[0021] Calculate the sample reconstruction error and the sample distribution deviation.

[0022] Preferably, the sample reconstruction error is calculated using formula (1):

[0023]

[0024] Among them, L Conv S represents the sample reconstruction error. N and T N Let represent the number of samples in the source domain and the target domain, respectively. and Let i represent the i-th sample in the source domain and its reconstructed sample, respectively. and Let i and represent the i-th sample of the target source and its reconstructed sample, respectively.

[0025] Preferably, the sample distribution deviation is calculated using formula (2):

[0026]

[0027] Among them, L Sim The sample distribution bias is represented by l, which represents the encoder E. Conv (·) and decoder D Conv The number of layers (·) and These represent the source and target domain samples in encoder E, respectively. Conv (·) Output features of the l-th layer and These represent the source and target domain samples in the decoder D, respectively. Conv (·) The features reconstructed at the l-th layer, where Sim(A,B) represents the similarity between A and B.

[0028] Preferably, the warning error is calculated using formula (3):

[0029]

[0030] Among them, L Cla Indicates the warning error, y i,j Let y′ represent the label of the j-th class of the i-th target domain sample. i,j Let c represent the output of the classifier for the j-th class of the i-th target domain sample, where c represents the class and c∈(1,2).

[0031] Preferably, the total error is calculated using formula (4):

[0032] L = L Cla +L Sim +L Conv (4)

[0033] Where L represents the total error, L Cla L represents the warning error. Sim L represents the sample distribution bias. Conv This represents the sample reconstruction error.

[0034] Preferably, the source domain samples include the INRIA Person Dataset, the Caltech Pedestrian Dataset, and the PETA Dataset.

[0035] Secondly, embodiments of this disclosure provide a thermal power plant early warning system based on transfer learning neural networks, comprising:

[0036] The neural network learning platform is used to train a transfer learning-based neural network based on source data and target data, obtain the trained neural network, and transmit it to the early warning analysis module.

[0037] The video module is used to acquire real-time video images of the area surrounding hazardous equipment in thermal power plants.

[0038] An image processing module is used to preprocess the video images and send them to the early warning analysis module;

[0039] The early warning analysis and prompting module is used to input the processed video image into the trained neural network, obtain the early warning result, and feed it back to the terminal for prompting.

[0040] The neural network includes an encoder, a decoder, and a classifier.

[0041] Preferably, the early warning analysis and prompting module includes:

[0042] The voice prompt module is used to issue voice prompts to warn personnel near dangerous equipment in thermal power plants to leave in time.

[0043] Preferably, the early warning analysis and prompting module includes:

[0044] The alarm threshold module is used to determine the time that personnel are close to the dangerous equipment in the thermal power plant. If the time that personnel are close to the dangerous equipment in the thermal power plant exceeds the preset value, the audible and visual alarm module will be activated.

[0045] The audible and visual alarm module is used to issue audible and visual alarms.

[0046] Preferably, the early warning analysis and prompting module includes:

[0047] The danger threshold module is used to determine the dangerous situation of personnel near dangerous equipment in a thermal power plant. If the dangerous situation exceeds the preset value, the power control module will be activated.

[0048] The power control module is used to automatically control the start-up and shutdown status of hazardous equipment in thermal power plants.

[0049] The methods and apparatus of the present invention have other features and advantages that will be apparent from or will be set forth in detail in the accompanying drawings and following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0050] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same parts.

[0051] Figure 1A flowchart illustrating the steps of a power plant early warning method based on a transfer learning neural network according to an embodiment of the present invention is shown.

[0052] Figure 2 A schematic diagram of a transfer learning method with low inter-domain distributional differences according to an embodiment of the present invention is shown.

[0053] Figure 3 A block diagram of a power plant early warning system based on a transfer learning neural network according to an embodiment of the present invention is shown. Detailed Implementation

[0054] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0055] To facilitate understanding of the solutions and effects of the embodiments of the present invention, two specific application examples are given below. Those skilled in the art should understand that these examples are merely for the purpose of understanding the present invention, and any specific details therein are not intended to limit the present invention in any way.

[0056] Example 1

[0057] Figure 1 A flowchart illustrating the steps of the early warning method for thermal power plants based on transfer learning neural networks according to the present invention is shown.

[0058] like Figure 1 As shown, the transfer learning method includes: Step 101, acquiring source domain samples and target domain samples; Step 102, training the neural network using the transfer learning method based on the source domain samples and target domain samples to obtain the trained neural network; Step 103, acquiring video images of the vicinity of hazardous equipment in the thermal power plant in real time and performing preprocessing; Step 104, inputting the preprocessed video images into the trained neural network to obtain early warning results and feeding them back to the terminal for prompting.

[0059] In one example, a neural network includes an encoder, a decoder, and a classifier.

[0060] In one example, training a neural network using transfer learning based on source domain samples and target domain samples includes:

[0061] The source domain samples and target domain samples are input into the neural network for iterative training. Each training iteration yields the corresponding sample reconstruction error, warning error and sample distribution deviation.

[0062] Calculate the total error based on the sample reconstruction error, warning error, and sample distribution deviation;

[0063] The parameters of the current neural network are adjusted based on the total error until the total error is less than a set threshold, at which point training is complete.

[0064] In one example, the corresponding sample reconstruction error and sample distribution deviation are obtained as follows:

[0065] Input the source domain samples and target domain samples into the encoder to obtain the corresponding source data features and target data features;

[0066] The source data features and target data features are input into the decoder to obtain the reconstructed source data and reconstructed target data.

[0067] Calculate the sample reconstruction error and sample distribution deviation.

[0068] In one example, the sample reconstruction error is calculated using formula (1):

[0069]

[0070] Among them, L Conv S represents the sample reconstruction error. N and T N Let represent the number of samples in the source domain and the target domain, respectively. and Let i represent the i-th sample in the source domain and its reconstructed sample, respectively. and Let i and represent the i-th sample of the target source and its reconstructed sample, respectively.

[0071] In one example, the sample distribution bias is calculated using formula (2):

[0072]

[0073] Among them, L Sim The sample distribution bias is represented by l, which represents the encoder E. Conv (·) and decoder D Conv The number of layers (·) and These represent the source and target domain samples in encoder E, respectively. Conv (·) Output features of the l-th layer and These represent the source and target domain samples in the decoder D, respectively. Conv (·) The features reconstructed at the l-th layer, where Sim(A,B) represents the similarity between A and B.

[0074] In one example, the warning error is calculated using formula (3):

[0075]

[0076] Among them, L Cla Indicates the warning error, y i,j Let y′ represent the label of the j-th class of the i-th target domain sample. i,j Let c represent the output of the classifier for the j-th class of the i-th target domain sample, where c represents the class and c∈(1,2).

[0077] In one example, the total error is calculated using formula (4):

[0078] L = L Cla +L Sim +L Conv (4)

[0079] Where L represents the total error, L Cla L represents the warning error. Sim L represents the sample distribution bias. Conv This represents the sample reconstruction error.

[0080] In one example, the source domain samples include the INRIA Person Dataset, the Caltech Pedestrian Dataset, and the PETA Dataset.

[0081] Figure 2 A schematic diagram of a transfer learning method with low inter-domain distributional differences according to an embodiment of the present invention is shown.

[0082] like Figure 2 As shown, specifically, source domain samples and target domain samples are acquired. Video images of the vicinity of hazardous equipment in thermal power plants are collected using cameras as the target domain, including historical image data from when there are clear warnings. Since there is relatively little video data of the vicinity of hazardous equipment stored during the actual operation of thermal power plants that can be used as clear warning images, the number of samples that can be processed as the target domain is limited. Therefore, a transfer learning method is used to select some publicly available datasets as the source domain for the thermal power plant early warning system, including the INRIA Person Dataset, the Caltech Pedestrian Dataset, and the PETA (Pedestrian Attribute) dataset.

[0083] The three datasets are explained in detail below:

[0084] The INRIA Person Dataset is a static pedestrian database that includes training and testing sets. This dataset features complex pedestrian backgrounds, various environmental changes (lighting), and diverse human poses, making it a dataset that closely resembles real-world scenarios.

[0085] The Caltech Pedestrian Dataset consists of approximately 10 hours of 640x480 video footage captured from real-world traffic. It is currently one of the largest pedestrian databases available, although it does contain images of poor quality.

[0086] PETA (Pedestrian Attribute) dataset: This dataset contains pedestrian attributes (gender and clothing style) and is a dataset for recognizing distant human images (i.e., images containing full-body information); the dataset consists of 19,000 pedestrian images with 65 pedestrian attributes, including a total of 8,705 people.

[0087] A convolutional autoencoder is used to extract features from the source and target domains. It contains an encoder E. Conv (·) and a decoder D Conv (·), training the neural network using transfer learning based on source domain samples and target domain samples to obtain the trained neural network; training the neural network using transfer learning based on source domain samples and target domain samples includes:

[0088] The source domain samples and target domain samples are input into the neural network for iterative training. Each training iteration yields the corresponding sample reconstruction error, warning error, and sample distribution bias. The source domain samples and target domain samples are then input into the encoder to obtain the corresponding source data features and target data features.

[0089]

[0090] The source data features and target data features are input into the decoder to obtain the reconstructed source data and reconstructed target data:

[0091]

[0092] The sample reconstruction error is calculated using formula (1), and L is minimized. Conv Features of the samples can be extracted. However, due to the complex working environment of the alarm system in thermal power plants, the distribution of samples in the target domain and the source domain deviates, which seriously reduces the performance of the alarm system. When the distribution of the target domain and the source domain deviates, the similarity between the sample features of the target domain and the sample features of the source domain is reduced. To solve this problem, the sample distribution deviation is calculated using formula (2). Minimize L Sim It can increase the similarity between the sample features of the target domain and the sample features of the source domain, which effectively solves the problem of the distribution deviation between the target domain and the source domain.

[0093] Classifier C Mlp(·) The classification of early warnings, i.e., whether to trigger an alarm or not, is a binary classification problem. Therefore, C Mlp (·) An MLP structure is adopted, which contains 3 layers with 1024-512-2 neurons respectively. The warning error is calculated by formula (3), where c represents the category, c∈(1,2), 1 represents warning, and 2 represents no warning. By optimizing L Cla It can achieve the classification of early warnings.

[0094] This method not only addresses the discrepancy between the sample distributions of the target domain and the source domain, but also enables classification for early warning. Based on the sample reconstruction error, early warning error, and sample distribution deviation, the total error is calculated using formula (4), where L... Cla To achieve early warning classification, L Sim To address the discrepancy between the sample distributions of the target domain and the source domain, L Conv Implement encoder E Conv (·) and decoder D Conv Optimization of (·). By optimizing L, the distribution difference between samples in the source and target domains of the thermal power plant alarm system can be reduced, effectively improving the performance of the thermal power plant early warning system; the parameters of the current neural network are adjusted according to the total error until the total error is less than the set threshold, and the training is completed.

[0095] The system collects and preprocesses video images of the area surrounding hazardous equipment in thermal power plants in real time; the preprocessed video images are then input into a trained neural network to obtain early warning results, which are then fed back to the terminal for alerts.

[0096] Example 2

[0097] Figure 3 A block diagram of a power plant early warning system based on a transfer learning neural network according to an embodiment of the present invention is shown.

[0098] like Figure 3 As shown, the early warning system for thermal power plants based on transfer learning neural networks includes:

[0099] The neural network learning platform is used to train a transfer learning-based neural network based on source data and target data, obtain the trained neural network, and transmit it to the early warning analysis module.

[0100] The video module is used to acquire real-time video images of the area surrounding hazardous equipment in thermal power plants.

[0101] The image processing module is used to preprocess video images and send them to the early warning analysis module;

[0102] The early warning analysis and prompting module is used to input the processed video images into the trained neural network, obtain the early warning results, and feed them back to the terminal for prompting.

[0103] Neural networks include encoders, decoders, and classifiers.

[0104] In one example, the early warning analysis and prompting module includes:

[0105] The voice prompt module is used to issue voice prompts to warn personnel near dangerous equipment in thermal power plants to leave in time.

[0106] In one example, the early warning analysis and prompting module includes:

[0107] The alarm threshold module is used to determine the time that personnel are close to the dangerous equipment in the thermal power plant. If the time that personnel are close to the dangerous equipment in the thermal power plant exceeds the preset value, the audible and visual alarm module will be activated.

[0108] The audible and visual alarm module is used to issue audible and visual alarms.

[0109] In one example, the early warning analysis and prompting module includes:

[0110] The danger threshold module is used to determine the dangerous situation of personnel near dangerous equipment in a thermal power plant. If the dangerous situation exceeds the preset value, the power control module will be activated.

[0111] The power control module is used to automatically control the start-up and shutdown status of hazardous equipment in thermal power plants.

[0112] Those skilled in the art should understand that the above description of the embodiments of the present invention is only intended to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any of the examples given.

[0113] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A power plant early warning method based on a neural network of migration learning, characterized in that, The method comprises the following steps: obtaining source domain samples and target domain samples; training a neural network through a transfer learning method according to the source domain samples and the target domain samples, and obtaining a trained neural network; real-time collection of video images around dangerous equipment in a thermal power plant and pre-processing of the video images; inputting the pre-processed video images into the trained neural network to obtain a warning result and feeding back the warning result to a terminal for prompting; wherein the neural network comprises an encoder, a decoder and a classifier; wherein the training of the neural network through the transfer learning method according to the source domain samples and the target domain samples comprises: inputting the source domain samples and the target domain samples into the neural network for iterative training, and obtaining corresponding sample reconstruction errors, warning errors and sample distribution deviations in each training; calculating a total error according to the sample reconstruction errors, the warning errors and the sample distribution deviations; adjusting parameters of the current neural network according to the total error until the total error is less than a set threshold, and completing the training; wherein the obtaining of the corresponding sample reconstruction errors and the sample distribution deviations comprises: inputting the source domain samples and the target domain samples into the encoder to obtain corresponding source data features and target data features; inputting the source data features and the target data features into the decoder to obtain reconstructed source data and reconstructed target data; calculating the sample reconstruction errors and the sample distribution deviations; wherein the sample reconstruction errors are calculated through formula (1): （1） wherein, denotes a sample reconstruction error, and denote the number of samples of the source domain and the target domain, respectively, and denote the source domain's th sample and its reconstructed sample, respectively, and denote the target source's th sample and its reconstructed sample, respectively; wherein the sample distribution deviations are calculated through formula (2): （2） wherein, represents a sample distribution bias, represents an encoder and a decoder , and represent source domain and target domain samples at the output of the encoder first layer, and represent source domain and target domain samples reconstructed at the output of the decoder first layer, represents a similarity and between ; wherein the warning errors are calculated through formula (3): （3） in, Indicates the warning error. Indicates the first The first target domain sample Class tags, This represents the classifier's response to the i-th target domain sample. The output corresponding to the class, where c represents the class, and c∈(1, 2); wherein the total error is calculated through formula (4): （4） where L denotes the total error, denotes the early warning error, denotes the sample distribution deviation, denotes the sample reconstruction error.

2. The neural network based on transfer learning method for power plant early warning according to claim 1, wherein, The source domain samples comprise INRIA Person Dataset, Caltech Pedestrian Dataset and PETA.

3. A neural network based power plant early warning system based on transfer learning, characterized in that, The method comprises the following steps: a neural network learning platform is used to train a neural network through a transfer learning method according to source domain samples and target domain samples, and a trained neural network is obtained; a video module is used to real-time collection of video images around dangerous equipment in a thermal power plant; an image processing module is used to pre-process the video images and send the pre-processed video images to a warning analysis module; a warning analysis and prompting module is used to input the pre-processed video images into the trained neural network to obtain a warning result and feed back the warning result to a terminal for prompting; wherein the neural network comprises an encoder, a decoder and a classifier; wherein the training of the neural network through the transfer learning method according to the source domain samples and the target domain samples comprises: inputting the source domain samples and the target domain samples into the neural network for iterative training, and obtaining corresponding sample reconstruction errors, warning errors and sample distribution deviations in each training; calculating a total error according to the sample reconstruction errors, the warning errors and the sample distribution deviations; adjusting parameters of the current neural network according to the total error until the total error is less than a set threshold, and completing the training; wherein the obtaining of the corresponding sample reconstruction errors and the sample distribution deviations comprises: inputting the source domain sample and the target domain sample into the encoder to obtain corresponding source data features and target data features; inputting the source data features and the target data features into the decoder to obtain reconstructed source data and reconstructed target data; calculating the sample reconstruction error and the sample distribution deviation; wherein the sample reconstruction error is calculated by formula (1): （1） in, Indicates the sample reconstruction error. and Let represent the number of samples in the source domain and the target domain, respectively. and Representing the source domain, respectively One sample and its reconstructed sample, and They represent the target source number respectively. One sample and its reconstructed sample; wherein the sample distribution deviation is calculated by formula (2): （2） wherein, represents a sample distribution bias, represents an encoder and a decoder number of layers, and represent source domain and target domain samples at the encoder first layer output features, and represent source domain and target domain samples at the decoder first layer reconstructed features, represents a similarity between and ; wherein the early warning error is calculated by formula (3): （3） in, Indicates the warning error. Indicates the first The first target domain sample Class tags, This represents the classifier's response to the i-th target domain sample. The output corresponding to the class, where c represents the class, and c∈(1, 2); wherein the total error is calculated by formula (4): （4） where L denotes the total error, denotes the early warning error, denotes the sample distribution deviation, denotes the sample reconstruction error.

Images