Dry-type reactor overheat fault detection method and system based on LSTM network

Abstract

The invention discloses a dry-type reactor overheating fault detection method and system based on an LSTM network. The dry-type reactor overheating fault detection method comprises the steps that S1, a single temperature sensor is used for collecting surface temperature data of key hot spot positions of a dry-type reactor; s2, the collected surface temperature data are preprocessed, the preprocessing process comprises normalization processing, and then a sliding window algorithm is used for generating time sequence data; s3, the LSTM network model receives the preprocessed time sequence data, and the LSTM network model outputs the overheating fault probability of the current time point; and S4, checking whether the fault probability output by the LSTM network model exceeds a preset alarm threshold value or not, if the fault probability exceeds the threshold value, sending an alarm signal, and if the fault probability does not exceed the threshold value, continuing monitoring by the system, and returning to the data acquisition step. According to the method, the efficiency and accuracy of fault detection can be improved.

Description

Technical Field

[0001] The present invention relates to the field of power equipment fault detection, and particularly to a method and system for detecting overheating faults of dry-type reactors based on a long short-term memory (LSTM) network. Background Art

[0002] In modern power systems, dry-type reactors, as an important electrical equipment, are widely used in power transmission and distribution networks. Dry-type reactors are mainly used to limit short-circuit current, improve voltage quality, and smooth current fluctuations under different load conditions. Compared with oil-immersed reactors, dry-type reactors have the advantages of no oil pollution, easy maintenance, and strong environmental adaptability, so they are widely used in urban power grids, wind farms, solar power generation systems and other fields. However, a large amount of heat is generated during the operation of dry-type reactors, especially when operating at high load for a long time, the temperature will rise significantly. Excessive temperature will not only accelerate the aging of insulating materials, but also cause damage to the reactor and even trigger serious safety accidents such as fires. Therefore, effective overheating fault detection and temperature monitoring of dry-type reactors have important practical significance.

[0003] Traditional temperature monitoring methods usually rely on the arrangement of multi-point temperature sensors to monitor different parts of the reactor through multiple sensors. Although this method can provide relatively comprehensive temperature information, it also has significant defects. First, the arrangement of multi-point sensors increases the complexity and cost of the system. Second, the installation and maintenance of sensors are more difficult, especially in a high-voltage environment, where the requirements for safe operation are relatively high. In addition, the data acquisition and processing of the sensor network require high real-time performance and stability, otherwise data delay or loss may occur, which will affect the accuracy of fault judgment.

[0004] In recent years, with the development of artificial intelligence and deep learning technologies, data-driven methods have been increasingly applied in the fault diagnosis of power equipment. In particular, the recurrent neural network (RNN) and its variant, the long short-term memory network (LSTM), have been widely used in the fields of prediction and fault detection due to their excellent performance in processing time series data. The LSTM network can capture and remember important information in a long time series through its unique gating mechanism, and is suitable for processing dynamic data during the operation of power equipment. In the detection of overheating faults of dry-type reactors, an innovative solution combining a single temperature sensor with an LSTM network has gradually become a research hotspot. This method installs a single high-precision temperature sensor at the key hot spot of the reactor to collect temperature data in real time. Combining the powerful sequence modeling ability of the LSTM network, it can accurately predict overheating faults based on the data of a single sensor. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: during the monitoring of the reactor using sensors, how to accurately predict the monitoring and fault data using the LSTM network to improve the robustness and adaptability of the system and provide an economical and effective solution for the safe operation of dry-type reactors.

[0006] To solve the above technical problem, the present invention provides a method for detecting overheating faults of dry-type reactors based on the LSTM network, including the following steps:

[0007] S1: Use a single temperature sensor to collect the surface temperature data of the key hot spots of the dry-type reactor;

[0008] S2: Preprocess the collected surface temperature data. The preprocessing process includes normalization processing, and then use the sliding window algorithm to generate time series data; use an array with a length of 7 to store the current collected value and its previous 24 collected data respectively. When the next sampling arrives, shift the array one bit to the right, that is, {x′ -24 = x′ -23 , x′ 1 = x′ 0 , x′ 0 = the current collected value}, and finally generate the sequence: {x′ 6 , x′ 5 , …, x′ 0}. S3: The LSTM network model receives the preprocessed time series data, and the LSTM network model outputs the overheating fault probability at the current time point;

[0009] S4: Check whether the fault probability output by the LSTM network model exceeds the preset alarm threshold. If the fault probability P(t) exceeds the threshold θ, an alarm signal is sent. If the fault probability P(t) does not exceed the threshold θ, the system continues to monitor and returns to the data collection step.

[0010] In the aforementioned method for detecting overheating faults of dry-type reactors based on the LSTM network, in step S1, the single temperature sensor is installed at the preset key hot spot position of the winding or iron core of the dry-type reactor.

[0011] In the aforementioned method for detecting overheating faults of dry-type reactors based on the LSTM network, in step S2, during the normalization process of the surface temperature data, the surface temperature data is scaled to [0, 1], and the formula is:

[0012]

[0013] Among them, x is the original temperature data, x′ is the normalized data, x min is the lowest temperature, and x max is the highest temperature.

[0014] The above-mentioned dry-type reactor overheating fault detection method based on the LSTM network, in the process of generating time series data, adopts the sliding window algorithm to generate time series data. The window length is set to W = 7, and the generated sequence is: {x′ 6 , x′ 5 , …, x′ 0}.

[0015] The above-mentioned dry-type reactor overheating fault detection method based on the LSTM network, in step S3, the LSTM network model includes 1 input layer, the first layer of LSTM units, the second layer of LSTM units and 1 output layer: The input layer is used to receive the preprocessed time series data {x′ 6 , x′ 5 , …, x′ 0};

[0016] The first layer of LSTM units receives the data from the input layer and calculates the hidden state h t and the cell state c t , and the formulas used are:

[0017] The sigmoid activation function σ C , σ C (n) = (1 + e -n ) -1

[0018] Calculate the input gate vector i t , i t = σ C (W i + R i h t-1 + b i )

[0019] Calculate the forget gate vector f t , f t = σ C (W f x t + R f h t-1 + b f )

[0020] Calculate the candidate gate vector g t , g t = σ g (W g x t + R g h t-1 + b g )

[0021] Calculate the output gate vector o t, o t = σ C (W o x t + R o h t-1 + b o )

[0022] Computing unit state c t ,

[0023] Computing hidden state h t , where σ C is the sigmoid activation function, e is the natural constant or Euler's number, n is the input variable of the activation function, σ g is the hyperbolic tangent activation function, denotes element-wise multiplication, i t is the input gate vector, f t is the forget gate vector, g t is the candidate gate vector, o t is the output gate vector, c t is the cell state, c t-1 is the cell state at the previous time step, h t is the hidden state, h t-1 is the hidden state at the previous time step, xt is the input at the current time step, W i , W f , W g , W o are the input weight matrix one, weight matrix two, weight matrix three, and weight matrix four respectively, R i , R f , R g , R o are the regression weight matrix one, regression weight matrix two, regression weight matrix three, and regression weight matrix four respectively, which are used for the calculations of the input gate, forget gate, candidate gate, and output gate, b i , b f , b g , b o are the bias vector one, bias vector two, bias vector three, and bias vector four respectively;

[0024] The second-layer LSTM unit receives the output of the first layer as input;

[0025] The output layer is used to map the hidden state h t at the last time step to the output space, uses sigmoid as the activation function to convert the result of the output layer into a probability value, and calculates the overheating fault probability P(t).

[0026] The above-mentioned dry-type reactor overheating fault detection method based on the LSTM network is characterized in that: in step S2, during the process of generating time series data, the sliding window algorithm is used to generate time series data, and the window length is set to W = 25, then the generated sequence is: {x′ -24 ,x′ -23 ,…,x′ 0}.

[0027] A dry-type reactor overheating fault detection system based on the LSTM network includes the following functional modules: Temperature data acquisition module: Use a single temperature sensor to collect the surface temperature data of the key hot spots of the dry-type reactor and transmit the surface temperature data to the data preprocessing module in real time;

[0028] Data preprocessing module: Normalize the collected temperature data and use the sliding window algorithm to generate time series data as the input data of the LSTM network model;

[0029] LSTM network model module: Receive the time series data processed by the data preprocessing module, and the LSTM network model outputs the overheating fault probability at the current time point;

[0030] Alarm processing module: Check whether the fault probability output by the LSTM network model exceeds the preset alarm threshold. If the fault probability P(t) exceeds the threshold θ, an alarm signal is issued. If the fault probability P(t) does not exceed the threshold θ, the system continues to monitor and returns to data acquisition.

[0031] In the above-mentioned dry-type reactor overheating fault detection system based on the LSTM network, in the temperature data acquisition module, the single temperature sensor is installed at the preset key hot spot position of the winding or iron core of the dry-type reactor.

[0032] In the above-mentioned dry-type reactor overheating fault detection system based on the LSTM network, in the data preprocessing module, during the process of normalizing the surface temperature data, the surface temperature data is scaled to [0,1], and the formula is:

[0033]

[0034] where x is the original temperature data, x′ is the normalized data, x min is the lowest temperature, x max is the highest temperature;

[0035] In the above-mentioned dry-type reactor overheating fault detection system based on the LSTM network, during the process of generating time series data, the sliding window algorithm is used to generate time series data, and the window length is set to W = 7, then the generated sequence is: {x ′ 6 ,xv 5 ,…,x ′ 0 Setting the window length to 7 can achieve a faster response speed while maintaining the same overheating fault prediction accuracy.

[0036] The above-mentioned dry-type reactor overheating fault detection system based on LSTM network, in the LSTM network model module, the LSTM network model includes 1 input layer, the first layer LSTM unit, the second layer LSTM unit and 1 output layer:

[0037] The input layer is used to receive the preprocessed time series data {x ′ 6 ,x ′ 5 ,…,x ′ 0};

[0038] The first layer of LSTM units receives data from the input layer and calculates the hidden state h at each time step. t and the cell state c t , using the formula:

[0039] Sigmoid activation function σ C , σ C (n) = (1 + e -n ) -1

[0040] Calculate the input gate vector i t ,i t =σ C (W i +R i h t-1 +b i )

[0041] Calculate the forget gate vector f t , f t =σ C (W f x t +R f h t-1 +b f )

[0042] Calculate the candidate gate vector g t , g t =σ g (W g x t +R g h t-1 +b g )

[0043] Calculate the output gate vector o t , o t = σ c (W o x t + R o h t-1 + b o )

[0044] Calculate the cell state c t ,

[0045] Calculate the hidden state h t , where σ C is the sigmoid activation function, e is the natural constant or Euler's number, n is the input variable of the activation function, σ g is the hyperbolic tangent activation function, represents element-wise multiplication, i t is the input gate vector, f t is the forget gate vector, g t is the candidate gate vector, o t is the output gate vector, c t is the cell state, c t-1 is the previous cell state, h t is the hidden state, h t-1 is the previous hidden state, xt is the current input, W i , W f , W g , W o are the input weight matrix one, weight matrix two, weight matrix three, and weight matrix four respectively, R i , R f , R g , R o are the regression weight matrix one, regression weight matrix two, regression weight matrix three, and regression weight matrix four respectively, and are used for the calculations of the input gate, forget gate, candidate gate, and output gate. b i , b f , b g , b o are the bias vector one, bias vector two, bias vector three, and bias vector four respectively;

[0046] The second-layer LSTM cell receives the output of the first layer as input;

[0047] The output layer is used to map the hidden state ht of the last time step to the output space, uses sigmoid as the activation function to convert the result of the output layer into a probability value, and calculates the overheating fault probability P(t).

[0048] The output layer performs on the {x′ 6 , x ′ 5 , …, x ′ 0 For the hidden state ht of each collected value, calculate the attention score, normalize the attention score through the softmax function, and calculate the context vector through weighted sum to finally generate the overheating fault probability P(t).

[0049] Beneficial effects achieved by the present invention: By installing a single high-precision temperature sensor at the key hot spots of the dry-type reactor, the present invention realizes real-time monitoring of the temperature during the operation of the equipment. Compared with the multi-point sensor arrangement in the prior art, the complexity of the system is greatly simplified, and the installation and maintenance costs are reduced. At the same time, the technical solution of the present invention utilizes the powerful time series modeling ability of the long short-term memory network (LSTM) to process and analyze the collected temperature data. The LSTM network can effectively capture and learn the patterns of temperature changes, thereby improving the accuracy of fault prediction. The technical solution of the present invention simplifies the installation and operation processes, especially in high-voltage environments, reducing the safety risks of maintenance and operation. Description of the Drawings

[0050] Figure 1 It is a flowchart of a dry-type reactor overheating fault detection method based on the LSTM network in Embodiment 1 of the present invention;

[0051] Figure 2 It is a schematic diagram of the system composition module in Embodiment 1 of the present invention;

[0052] Figure 3 It is a schematic diagram of the LSTM network structure flow block in Embodiment 1 of the present invention;

[0053] Figure 4 It is a schematic diagram of the LSTM memory cell structure in Embodiment 1 of the present invention;

[0054] Figure 5 It is a schematic diagram of the LSTM network result block in Embodiment 1 of the present invention;

[0055] Figure 6 It is a schematic diagram of the LSTM network result block in Embodiment 2 of the present invention. Detailed Embodiments

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the following will clearly and completely describe the technical solutions in the embodiments of this application in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are some, but not all, of the embodiments of this application. In the following description, providing specific details such as specific configurations and components is only to help a comprehensive understanding of the embodiments of this application. Therefore, those skilled in the art should clearly understand that various changes and modifications can be made to the embodiments described here without departing from the scope and spirit of this application. Additionally, descriptions of known functions and configurations are omitted for clarity and conciseness.

[0057] Embodiment 1

[0058] As Figure 1 and Figure 2 shown, this embodiment provides a method for detecting overheating faults of dry-type reactors based on an LSTM network, including the following steps:

[0059] S1: Use a single temperature sensor to collect surface temperature data at the key hot spots of the dry-type reactor.

[0060] S2: Preprocess the collected surface temperature data. The preprocessing process includes normalization processing, and then use the sliding window algorithm to generate time series data.

[0061] S3: The LSTM network model receives the preprocessed time series data, and the LSTM network model outputs the overheating fault probability at the current time point.

[0062] S4: Check whether the fault probability output by the LSTM network model exceeds the preset alarm threshold. If the fault probability P(t) exceeds the threshold θ, an alarm signal is issued. If the fault probability P(t) does not exceed the threshold θ, the system continues to monitor and returns to the data collection step.

[0063] In step S1, the single temperature sensor is installed at the preset key hot spots of the winding or iron core of the dry-type reactor.

[0064] In step S2, during the normalization process of the surface temperature data, the surface temperature data is scaled to [0, 1]. The formula is:

[0065]

[0066] where x is the original temperature data, x′ is the normalized data, x min is the lowest temperature, and x max is the highest temperature.

[0067] During the process of generating time series data, the sliding window algorithm is used to generate time series data. The window length is set to W = 7, and the generated sequence is: {x ′ 6 , x ′ 5 , …, x ′ 0}.

[0068] In step S3, the LSTM network model includes 1 input layer, the first layer of LSTM cells, the second layer of LSTM cells, and 1 output layer:

[0069] The input layer is used to receive the preprocessed time series data {x ′ 6 , x ′ 5 , …, x ′ 0};

[0070] The first layer of LSTM cells receives the data from the input layer and calculates the hidden state h t and the cell state c t , and the formulas used are:

[0071] The sigmoid activation function σ C , σ C (n) = (1 + e -n ) -1

[0072] Calculate the input gate vector i t , i t = σ C (W i + R i h t-1 + b i )

[0073] Calculate the forget gate vector f t , f t = σ C (W f x t + R f h t-1 + b f )

[0074] Calculate the candidate gate vector g t , g t = σ g (W g x t + R g h t-1 + b g )

[0075] Calculate the output gate vector o t , o t = σ C (W o x t + R o h t-1 + b o )

[0076] Calculate the cell state c t ,

[0077] Calculate the hidden state h t ,

[0078] where σ C is the sigmoid activation function, e is the natural constant or Euler's number, n is the input variable of the activation function, σ g is the hyperbolic tangent activation function, represents element-wise multiplication, i t is the input gate vector, f t is the forget gate vector, g t is the candidate gate vector, o t is the output gate vector, c t is the cell state, c t-1 is the previous cell state, h t is the hidden state, h t-1 is the previous hidden state, xt is the current input, W i , W f , W g , W o are the input weight matrix one, weight matrix two, weight matrix three, and weight matrix four respectively, R i , R f , R g , R o are the regression weight matrix one, regression weight matrix two, regression weight matrix three, and regression weight matrix four respectively, and are used for the calculations of the input gate, forget gate, candidate gate, and output gate, b i , b f , b g , b o are the bias vector one, bias vector two, bias vector three, and bias vector four respectively;

[0079] The second-layer LSTM cell receives the output of the first layer as input;

[0080] The output layer is used to output the hidden state h of the last time step tMap it to the output space, use the sigmoid as the activation function to convert the result of the output layer into a probability value, and calculate the overheating fault probability P(t).

[0081] The output layer calculates the hidden state ht of each acquired value {x′ 6 ,x′ 5 ,…,x′ 0}, calculates the attention score, normalizes the attention score through the softmax function, and calculates the context vector through the weighted sum to finally generate the overheating fault probability P(t).

[0082] In step S4, the threshold can be set to 0.75.

[0083] A dry-type reactor overheating fault detection system based on the LSTM network includes the following functional modules: Temperature data acquisition module 101: Use a single temperature sensor to acquire the surface temperature data of the key hot spots of the dry-type reactor; Transmit the surface temperature data to the data preprocessing module 103 in real time;

[0084] The single temperature sensor is installed at the preset key hot spot positions of the winding or iron core of the dry-type reactor;

[0085] The single temperature sensor measures the temperature range from -40°C to +150°C to meet the operation requirements of the dry-type reactor under various environmental conditions.

[0086] The response time of the single temperature sensor is less than 1 second, which can quickly capture the dynamic changes of the temperature and provide real-time temperature data.

[0087] The drift of the single temperature sensor is below ±0.1°C / year, which can ensure the long-term reliability of the data.

[0088] The single temperature sensor is a high-precision and fast-response temperature sensor. The single temperature sensor supports serial data transmission (UART) output, and serial communication is used for data transmission between the single temperature sensor and the data preprocessing module 102 to achieve efficient data transmission and processing.

[0089] Data preprocessing module 102: Normalize the acquired surface temperature data and use the sliding window algorithm to generate time series data as the input data of the LSTM network model 103;

[0090] The data preprocessing module 102 includes a data normalization processing unit and a time series generation unit:

[0091] The data normalization unit is used to normalize the received temperature data, scale the temperature data to [0,1], and the formula is:

[0092]

[0093] where x is the original temperature data, x′ is the normalized data, x min is the lowest temperature, x max is the highest temperature.

[0094] The time series generation unit dynamically adjusts the length W of the sliding window to adapt to different monitoring requirements. The time series generation unit uses the sliding window algorithm to generate time series data. If the window length is set to W = 7, the generated sequence is: {x ′ 6 ,x ′ ′ ,…,x ′ 0}; When the sampling period is 1 second and the window length is set to 7, the response speed can be less than 10 seconds under the condition of maintaining the same overheating fault prediction accuracy.

[0095] The LSTM network model module 103: receives the time series data processed by the data preprocessing module 102, and the LSTM network model outputs the overheating fault probability at the current time point;

[0096] As Figure 3 、 Figure 5 shown, the LSTM network model includes 1 input layer, the first layer of LSTM units, the second layer of LSTM units and 1 output layer:

[0097] The input layer is used to receive the preprocessed time series data {x′ 6 ,x′ 5 ,…,x′ 0};

[0098] As Figure 4 shown, the first layer of LSTM units receives the data from the input layer and calculates the hidden state h t and the cell state c t , and the formula is:

[0099] The sigmoid activation function σ C , σ C (n)=(1 + e -n ) -1

[0100] Calculate the input gate vector i t , i t =σ C (Wi +R i h t-1 +b i )

[0101] Calculate the forget gate vector f t ,f t =σ C (W f x t +R f h t-1 +b f )

[0102] Calculate the candidate gate vector g t ,g t =σ g (W g x t +R g h t-1 +b g )

[0103] Calculate the output gate vector o t ,o t =σ C (W o x t +R o h t-1 +b o )

[0104] Calculate the cell state c t ,

[0105] Calculate the hidden state h t ,

[0106] Among them, σ C is the sigmoid activation function, e is the natural constant or Euler's number, n is the input variable of the activation function, σ h is the hyperbolic tangent activation function, represents element-wise multiplication, i t is the input gate vector, f t is the forget gate vector, g t is the candidate gate vector, o t is the output gate vector, c t is the cell state, c t-1 is the cell state of the previous moment, h t is the hidden state, h t-1 is the hidden state of the previous moment, x t is the input at the current moment, W i 、W f 、W g 、Wo They are the input weight matrix 1, weight matrix 2, weight matrix 3, and weight matrix 4 respectively, R i , R f , R g , R o They are the regression weight matrix 1, regression weight matrix 2, regression weight matrix 3, and regression weight matrix 4 respectively, which are used for the calculations of the input gate, forget gate, candidate gate, and output gate respectively, b i , b f , b g , b o They are the bias vector 1, bias vector 2, bias vector 3, and bias vector 4 respectively;

[0107] The second-layer LSTM unit receives the output of the first layer as input, which is used to enhance the time memory ability and prediction accuracy of the model;

[0108] The output layer is used to map the hidden state ht at the last time step to the output space, uses sigmoid as the activation function to convert the result of the output layer into a probability value, and calculates the overheat fault probability P(t); the LSTM network model is trained with historical temperature data {x ′ 6 , x ′ 5 , …, x ′ 0} with a spatial length of w = 7 to improve the accuracy and robustness of fault detection.

[0109] Alarm processing module 104: Check whether the fault probability output by the LSTM network model 103 exceeds a preset alarm threshold. If the fault probability P(t) exceeds the threshold θ, an alarm message is sent. If the fault probability P(t) does not exceed the threshold θ, the system continues to monitor and returns to the data acquisition step.

[0110] The alarm processing module includes a comparison unit and an alarm unit:

[0111] The comparison unit is used to compare the fault probability P(t) output by the LSTM network model with the preset alarm threshold θ;

[0112] The alarm unit, if the fault probability P(t) > θ, sends an alarm message, and the remote communication module is used to send an alarm signal.

[0113] The alarm module includes an audible and visual alarm device and a remote communication module, which provides local and remote alarms during fault detection.

[0114] Embodiment 2

[0115] This embodiment provides a dry-type reactor overheating fault detection method based on an LSTM network, including the following steps:

[0116] S1: Use a single temperature sensor to collect the surface temperature data of the key hot spots of the dry-type reactor;

[0117] S2: Preprocess the collected surface temperature data. The preprocessing process includes normalization processing, and then use the sliding window algorithm to generate time series data;

[0118] S3: The LSTM network model receives the preprocessed time series data, and the LSTM network model outputs the overheating fault probability at the current time point;

[0119] S4: Check whether the fault probability output by the LSTM network model exceeds the preset alarm threshold. If the fault probability P(t) exceeds the threshold θ, an alarm signal is sent. If the fault probability P(t) does not exceed the threshold θ, the system continues to monitor and returns to the data collection step.

[0120] As Figure 6 shown, in step S2, in the process of generating time series data, the sliding window algorithm is used to generate time series data. Set the window length to W = 25, then the generated sequence is: {x′ -24 ,x′ -23 ,…,x′ 0}; In step S3, in the LSTM network model module, the LSTM network model includes 1 input layer, the first layer of LSTM cells, the second layer of LSTM cells and 1 fully connected layer:

[0121] The input layer is used to receive the preprocessed time series data {x′ -24 ,x′ -23 ,…,x′ 0};

[0122] The first layer of LSTM cells receives the data from the input layer and calculates the hidden state h t and the cell state c t , and the formula used is:

[0123] The sigmoid activation function σ C , σ C (n)=(1 + e -n ) -1

[0124] Calculate the input gate vector i t , i t =σ C (W i +R i h t-1+b i )

[0125] Calculate the forget gate vector f t , f t = σ C (W f x t + R f h t-1 + b f )

[0126] Calculate the candidate gate vector g t , g t = σ g (W g x t + R g h t-1 + b g )

[0127] Calculate the output gate vector o t , o t = σ C (W o x t + R o h t-1 + b o )

[0128] Calculate the cell state c t ,

[0129] Calculate the hidden state h t ,

[0130] where σ C is the sigmoid activation function, e is the natural constant or Euler's number, n is the input variable of the activation function, σ g is the hyperbolic tangent activation function, denotes element-wise multiplication, i t is the input gate vector, f t is the forget gate vector, g t is the candidate gate vector, o t is the output gate vector, c t is the cell state, c t-1 is the cell state of the previous time step, h t is the hidden state, h t-1 is the hidden state of the previous time step, x t is the input at the current time step, W i 、W f 、W g 、W o are the input weight matrix one, weight matrix two, weight matrix three, and weight matrix four respectively, Ri , R f , R g , R o are respectively the first regression weight matrix, the second regression weight matrix, the third regression weight matrix, and the fourth regression weight matrix, which are respectively used for the calculations of the input gate, the forget gate, the candidate gate, and the output gate. b i , b f , b g , b o are respectively the first bias vector, the second bias vector, the third bias vector, and the fourth bias vector;

[0131] The second-layer LSTM unit receives the output of the first layer as input;

[0132] The fully connected layer is used to map the hidden state ht at the last time step to the output space, uses sigmoid as the activation function to convert the result of the output layer into a probability value, and calculates the overheating fault probability P(t).

[0133] In this embodiment, for the time series data, the window length is set to 25. Although the response speed is slightly longer, it can significantly improve the accuracy and robustness of fault detection.

[0134] Through the above implementation steps, compared with the traditional multi-point sensor layout, this embodiment greatly simplifies the complexity of the system, reduces the installation and maintenance costs, and utilizes the powerful time series modeling ability of the long short-term memory network (LSTM) to process and analyze the collected temperature data. The LSTM network can effectively capture and learn the patterns of temperature changes, thereby improving the accuracy of fault prediction.

[0135] The above are only the preferred embodiments of the present invention, and it does not thereby limit the protection scope of the present invention. For those skilled in the art, the present invention can have various changes and modifications; all changes, modifications, substitutions, integrations, and parameter changes made to these embodiments without departing from the principle and spirit of the present invention by means of conventional substitutions or capable of achieving the same functions fall within the protection scope of the present invention.

Claims

1. A dry-type reactor overheating fault detection method based on LSTM network, characterized in that: The following steps are involved: S1: Use a single temperature sensor to collect surface temperature data at key hotspot locations of dry-type reactors; S2: preprocessing the collected surface temperature data, wherein the preprocessing process includes normalization processing, and then using a sliding window algorithm to generate time series data; S3: The LSTM network model receives the preprocessed time series data, and the LSTM network model outputs the overheating failure probability at the current time point; S4: Check whether the failure probability output by the LSTM network model exceeds the preset alarm threshold. Exceeding the threshold An alarm signal is issued if the failure probability Not exceeding the threshold , the system continues to monitor and returns to the data collection step.

2. The dry-type reactor overheating fault detection method based on LSTM network according to claim 1 is characterized in that: In step S1, the single temperature sensor is installed at a preset key hot spot position of the winding or core of the dry-type reactor.

3. According to the method for detecting overheating fault of dry-type reactor based on LSTM network in claim 1, in step S2, in the process of normalizing the surface temperature data, the surface temperature data is scaled to , the formula is: ； in, is the original temperature data, is the normalized data, is the minimum temperature, is the maximum temperature.

4. The dry-type reactor overheating fault detection method based on LSTM network according to claim 3 is characterized in that: In the process of generating time series data, the sliding window algorithm is used to generate time series data, and the window length is set to , use an array of length 7 to store the current acquisition value and the previous 6 acquisition data respectively. When the next sample arrives, shift the array right by one position, that is, , the final generated sequence is: .

5. The dry-type reactor overheating fault detection method based on LSTM network according to claim 4 is characterized in that: In step S3, the LSTM network model includes 1 input layer, a first layer of LSTM units, a second layer of LSTM units and 1 output layer: The input layer is used to receive the preprocessed time series data ; The first layer of LSTM units receives data from the input layer and calculates the hidden state of each time step and cell status , using the formula: Activation Function ; Calculate the input gate vector , ; Calculate the forget gate vector , ; Calculate candidate gate vectors , ; Calculate the output gate vector , ; Compute unit status , ; Calculate hidden state , ; in, for Activation function, is a natural constant or Euler number, Input variables for the activation function, is the hyperbolic tangent activation function, represents element-wise multiplication, is the input gate vector, is the forget gate vector, is the candidate gate vector, is the output gate vector, is the unit state, is the unit state at the previous moment, is the hidden state, is the hidden state at the previous moment, is the input quantity at the current moment, They are input weight matrix one, weight matrix two, weight matrix three, and weight matrix four respectively. They are regression weight matrix one, regression weight matrix two, regression weight matrix three, and regression weight matrix four, which are used for the calculation of input gate, forget gate, candidate gate, and output gate respectively. They are bias vector one, bias vector two, bias vector three, and bias vector four respectively; The second layer of LSTM units receives the output of the first layer as input; The output layer is used to convert the hidden state of the last time step Mapped to the output space, use As an activation function, the result of the output layer is converted into a probability value and the probability of overheating failure is calculated ; The output layer is The hidden state of each collected value , calculate the attention score, through The function normalizes the attention score and finally generates the overheating failure probability by weighted summing the context vector .

6. The dry-type reactor overheating fault detection method based on LSTM network according to claim 3 is characterized in that: In step S2, in the process of generating time series data, a sliding window algorithm is used to generate time series data, and the window length is set to , then the generated sequence is: .

7. A dry-type reactor overheating fault detection system based on LSTM network, characterized in that: Includes the following functional modules: Temperature data acquisition module: Use a single temperature sensor to collect surface temperature data at key hotspots of dry-type reactors, and transmit the surface temperature data to the data preprocessing module in real time; Data preprocessing module: normalizes the collected temperature data and uses the sliding window algorithm to generate time series data as input data for the LSTM network model; LSTM network model module: receives the time series data processed by the data preprocessing module, and the LSTM network model outputs the overheating failure probability at the current time point; Alarm processing module: Checks whether the failure probability output by the LSTM network model exceeds the preset alarm threshold. Exceeding the threshold Send out an alarm signal if the failure probability Not exceeding the threshold , the system continues monitoring and returns to data collection.

8. The dry-type reactor overheating fault detection system based on LSTM network according to claim 1 is characterized in that: In the temperature data acquisition module, the single temperature sensor is installed at a preset key hot spot position of the winding or core of the dry-type reactor.

9. The dry-type reactor overheating fault detection system based on LSTM network according to claim 1 is characterized in that: In the data preprocessing module, during the normalization process of the surface temperature data, the surface temperature data is scaled to , the formula is: ； in, is the original temperature data, is the normalized data, is the minimum temperature, is the maximum temperature.

10. The dry-type reactor overheating fault detection system based on LSTM network according to claim 9, characterized in that: In the process of generating time series data, the sliding window algorithm is used to generate time series data, and the window length is set to , then the generated sequence is: .

11. The dry-type reactor overheating fault detection system based on LSTM network according to claim 10, characterized in that: In the LSTM network model module, the LSTM network model includes 1 input layer, the first layer LSTM unit, the second layer LSTM unit and 1 output layer: The input layer is used to receive the preprocessed time series data ; The first layer of LSTM units receives data from the input layer and calculates the hidden state of each time step and cell status , using the formula: Activation Function ; Calculate the input gate vector , ; Calculate the forget gate vector , ; Calculate candidate gate vectors , ; Calculate the output gate vector , ; Compute unit status , ; Calculate hidden state , ; in, for Activation function, is a natural constant or Euler number, Input variables for the activation function, is the hyperbolic tangent activation function, represents element-wise multiplication, is the input gate vector, is the forget gate vector, is the candidate gate vector, is the output gate vector, is the unit state, is the unit state at the previous moment, is the hidden state, is the hidden state at the previous moment, is the input quantity at the current moment, They are input weight matrix one, weight matrix two, weight matrix three, and weight matrix four respectively. They are regression weight matrix one, regression weight matrix two, regression weight matrix three, and regression weight matrix four, which are used for the calculation of input gate, forget gate, candidate gate, and output gate respectively. They are bias vector one, bias vector two, bias vector three, and bias vector four respectively; The second layer of LSTM units receives the output of the first layer as input; The output layer is used to convert the hidden state of the last time step Mapped to the output space, use As an activation function, the result of the output layer is converted into a probability value and the probability of overheating failure is calculated ; The output layer is The hidden state of each collected value , calculate the attention score, through The function normalizes the attention score and finally generates the overheating failure probability by weighted summing the context vector .

Images