A method for detecting the concentration of thermally ionized ions in a substation

Abstract

A method for detecting the concentration of thermally released ions in a substation, which includes collecting an air sample in the substation environment and releasing it into a cloud chamber; irradiating the air sample in the cloud chamber with a light wave emitter, and by changing the wavelength λ and the incident angle T of the incident light wave, using sensors in the cloud chamber to collect data on the object distance u, the interference fringe frequency F, and the fringe circle diameter φ; constructing an RBM model to reduce the amount of data while retaining the characteristics of all the data; constructing a prediction model based on a BP neural network to output the predicted concentration of thermally released ions; outputting a comparative analysis of the predicted concentration of thermally released ions and the true value of the concentration of thermally released ions; the model completes model training, and the trained RBM model and BP neural network model are applied for predicting the concentration of thermally released ions. The present invention jointly uses an RBM model and a BP neural network model, greatly reducing the amount of data for calculating the concentration of thermally released ions, reducing the operation time, and at the same time improving the calculation accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of power equipment monitoring, and particularly relates to a method for detecting the concentration of thermally released ions in a substation. Background Art

[0002] In recent years, the occurrence of large-scale power grid fire accidents has caused huge economic losses and negative social impacts. Once an electrical equipment catches fire, it spreads quickly and the fire is fierce and difficult to extinguish. It will not only directly damage a large number of primary and secondary equipment, but also affect nearby equipment due to the fire and explosion of the equipment, resulting in a long power outage for the substation to repair, seriously affecting the operation safety of the power grid. At present, there is no very effective method for fire monitoring in outdoor areas of substations in China.

[0003] When a substance is heated, it decomposes into thermally released particles. Thermally released particles are the smallest substance components that can exist in a free state, and their generation time is earlier than that of smoke. In the fire alarm industry, this phenomenon can be used to detect early symptoms of electrical fires at the scene. The current mainstream detection method is the optical particle counting method. The concentration of thermally released ions is monitored through a power grid fire hazard monitoring and early warning system, and the combustion stage is judged by comparing with a threshold value, and then the early fire alarm function is realized. However, the existing optical particle counting method can only judge the concentration according to the spot area generated by the irradiation of particles by a single-wavelength spectrum. When there are interfering gas components in the indoor environment, spot overlap will occur, resulting in serious deviation of the measurement results. Therefore, it is urgent to develop a detection method that can accurately detect the concentration of thermally released ions in the current substation environment in real time to improve the detection ability of extremely early fires in substations. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for detecting the concentration of thermally released ions in a substation, which has good accuracy in measuring thermally released ions and can meet the requirements of practical applications.

[0005] The technical solution of the present invention is as follows:

[0006] A method for detecting the concentration of thermally released ions in a substation, the specific steps include:

[0007] Step 1: Collect an air sample in the substation environment and release it into a cloud chamber; use a light wave emitter to irradiate the air sample in the cloud chamber, and by changing the wavelength λ and the incident angle τ of the incident light wave, use the sensor in the cloud chamber to collect data on the object distance u, the interference fringe frequency F, and the fringe circle diameter Φ;

[0008] Step 2: Construct an RBM model. Input the wavelength λ of the incident light wave, the incident angle τ, the object distance u collected by the sensor in the cloud chamber, the interference fringe frequency F, and the fringe circle diameter Φ of the collected data for training. While preserving the characteristics of all the data, reduce the amount of data and extract the data incident angle τ', the system object distance u', the interference fringe frequency F', the fringe circle diameter Φ', and the light wave wavelength λ'.

[0009] Step 3: Construct a prediction model based on the BP neural network. Input the data incident angle τ', the system object distance u', the interference fringe frequency F', the fringe circle diameter Φ', the light wave wavelength λ', and the true value of the concentration of thermally released ions, and output the predicted concentration of thermally released ions.

[0010] Step 4: Output a comparative analysis of the predicted concentration of thermally released ions and the true value of the concentration of thermally released ions to obtain the model prediction error. When the model prediction error meets the set model error requirement, complete the model training. When the model prediction error does not meet the set model error requirement, update the connection weights of the RBM model, the biases of the visible layer neurons, and the biases of the hidden layer neurons, and adjust the weights and thresholds of the BP neural network for correction, then return to Step 2.

[0011] Step 5: When the model prediction error meets the set model error requirement or reaches the number of iterations, complete the model training, and apply the trained RBM model and BP neural network model to predict the concentration of thermally released ions.

[0012] Further optimization: In Step 2, through pre-training the RBM model, perform data reduction with feature retention on the input collected data. The specific process is as follows:

[0013] (1) The input training sample data is used to train the first hidden layer output vector through the weight matrix and bias vector. The first hidden layer units are driven by the data vector to obtain random binary states, and the values of the hidden layer units have only two states: 0 and 1. The probability that the hidden layer unit is set to 1 is:

[0014]

[0015] where θ = {W R , a R , b R} represents the parameter set of the restricted Boltzmann machine, a R represents the bias of the visible layer neurons, b Rj represents the bias of the hidden layer neurons, W R,i,j represents the connection weights between the visible layer and hidden layer nodes, v i represents the state value of the i-th node in the visible layer, h j represents the state of the hidden layer neurons. When the value of the hidden layer neuron is 1, it represents the on state, otherwise it is set to 0, representing the off state.

[0016] (2) The output vector of the first hidden layer is input into the visible layer of the second RBM, and step (1) is continued to be repeated to train and obtain the output vector of the second hidden layer, and the reduced data with retained features is extracted.

[0017] Further preferably, a prediction model based on a BP neural network is constructed, and the specific process is as follows:

[0018] (1) According to the system input sequences x1, x2,..., x c , where x is the data after being reduced by the RBM model, x c = {τ c ’, u c ’, F c ’, Φ c ’, λ c ’}; and the true pyroelectric ion concentration value Y is input; among them, the number of input nodes n of the BP neural network is 5, the number of output nodes is m = 1, and the number of hidden layer nodes L. The calculation formula for the number of neurons L in the hidden layer is:

[0019] L = 2n + 1;

[0020] (2) Selection of the transfer function of the BP neural network

[0021] The transfer function between the input layer and the hidden layer of the BP neural network adopts the S function:

[0022]

[0023] The transfer function between the hidden layer and the output layer of the BP neural network adopts the linear function:

[0024] y = purelin(s) = s

[0025] (3) Calculate the hidden layer

[0026] Initialize the connection weights W ij between the input layer and the hidden layer, and initialize the hidden layer threshold a; among them, the connection weights W ij and the hidden layer threshold a are randomly set and continuously adjusted in the subsequent training process along the direction of decreasing error;

[0027] According to the input sequence, the connection weights W ij and the threshold a, calculate the output H of the hidden layer nodes as:

[0028]

[0029] In the formula, f is the hidden layer activation function, and there are various expressions for this function.

[0030] (4) Calculate the output of the output node

[0031] Initialize the connection weights W between the hidden layer and the output layer jk , and initialize the output layer threshold b; among them, the connection weights W jk and the output layer threshold b are randomly set and continuously adjusted in the subsequent training process along the direction of error reduction;

[0032] According to the output H of the hidden layer nodes in step (3), combine the connection weights w jk and the threshold b to calculate the predicted output O as:

[0033]

[0034] Further preferably, the model prediction error e is obtained by calculating according to the model predicted output O and the expected output Y, and the formula is:

[0035] e k = Y k - O k , k = 1, 2,..., m

[0036] Among them, the unit of the model prediction error e is ten thousand · cm -3 .

[0037] Further preferably, update the connection weights of the RBM model, the biases of the visible layer neurons, and the biases of the hidden layer neurons, specifically as follows:

[0038] Reconstruct the visible layer, set the reconstructed visible layer as v*, the hidden layer obtained from v* is h*, the learning rate is ε, and the connection weights W R , the biases a of the visible layer neurons R , and the biases b of the hidden layer neurons R The update formulas are:

[0039]

[0040] a R ← a R + ε × (v - v * )

[0041]

[0042] Further preferably, the weight adjustment and threshold correction of the BP neural network are specifically as follows:

[0043] According to the BP neural network model prediction error e, adjust the network connection weights W ij and W jk between each layer:

[0044]

[0045] w jk = w jk + ηH j e k , k = 1, 2, ..., m

[0046] According to the prediction error e of the BP neural network model, the thresholds a and b of the network nodes are corrected:

[0047]

[0048] b k = b k + e k , k = 1, 2, ..., m

[0049] Where η is the learning rate, 1 > η > 0.

[0050] The present invention adopts the RBM model, which greatly reduces the amount of data to be processed and significantly reduces the calculation time. At the same time, by combining the RBM model with the BP neural network, the calculation accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a schematic diagram of the cloud chamber structure used in the present invention;

[0052] Figure 2 is the overall training process diagram of the RBM model of the present invention;

[0053] Figure 3 is the BP neural network model diagram of the present invention;

[0054] Figure 4 is the flow chart of the BP neural network of the present invention;

[0055] Figure 5 is the comparison diagram of the thermally released ion measurement value and the true value results of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the following clearly and completely describes the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] Embodiment

[0058] The present invention provides a method for detecting the concentration of thermally released ions in a substation, and the specific steps include:

[0059] Step 1: Use sensors to collect air sample data in the substation environment

[0060] Use a sampling tube to sample the air in the substation environment and release it into the cloud chamber; as Figure 1 shown, use a light wave emitter to irradiate the air sample in the cloud chamber. By changing the wavelength λ of the incident light wave and the incident angle τ, use the sensor in the cloud chamber to collect data on the object distance u, the interference fringe frequency F, and the fringe circle diameter Φ, and then detect the concentration of thermally released ions; among them, the data collected by the sensor includes data on thermally released ions, smoke particles, and dust particles.

[0061] Step 2: Construct an RBM model

[0062] Through pre-training of the RBM model, perform data reduction with feature retention on the collected data. The specific process is as follows:

[0063] (1) As Figure 2 shown, the input training sample data includes the incident angle τ, the object distance u, the interference fringe frequency F, the fringe circle diameter Φ, and the wavelength λ of the incident light wave. The output vector of the first hidden layer is obtained through training with the weight matrix and the bias vector; the first hidden layer units are driven by the data vector to obtain a random binary state, and the values of the hidden layer units have only two states, 0 and 1; the probability that the hidden layer unit is set to 1 is:

[0064]

[0065] where θ = {W R , a R , b R} represents the parameter set of the restricted Boltzmann machine, a R represents the visible layer neuron bias, b Rj represents the hidden layer neuron bias, W R,i,j represents the connection weight between the visible layer and the hidden layer nodes, v i represents the state value of the i-th node in the visible layer, h j the state of the hidden layer neuron. When the value of the hidden layer neuron is taken as 1, it represents the on state, otherwise it is set to 0, representing the off state;

[0066] (2) The output vector of the first hidden layer is input into the visible layer of the second RBM, and step (1) is continued to train to obtain the output vector of the second hidden layer, and the data incident angle τ’, the system object distance u’, the interference fringe frequency F’, the fringe circle diameter Φ’, and the light wave wavelength λ’ are extracted; by training the RBM neural network, a large amount of input data can be reduced in quantity without destroying its characteristics, facilitating subsequent calculations;

[0067] Step 3: Construct a prediction model based on the BP neural network

[0068] After pre-training the measured particles through the RBM neural network, the incident angle τ', system object distance u', interference fringe frequency F', fringe circle diameter Φ', and light wave wavelength λ' of the extracted data are used as the inputs of the input layer of the BP neural network to detect the concentration of pyroelectric ions by the BP neural network. The specific process is as follows:

[0069] (1) According to the system input sequence x1, x2,..., x c , where x is the data processed by the RBM network reduction, and x c = {τ c ', u c ', F c ', Φ c ', λ c '}; and the input pyroelectric ion concentration Y, where Y is the true pyroelectric ion concentration value in the input data;

[0070] Among them, the number of input nodes of the BP neural network n = 5, the number of output nodes is m = 1, and the number of hidden layer nodes L. The calculation formula for the number of neurons L in the hidden layer is:

[0071] L = 2n + 1;

[0072] In the formula, n represents the number of neurons in the input layer, that is, n = 5;

[0073] Then the number of neurons in the hidden layer l = 2n + 1 = 11; The structure of the BP neural network is as Figure 3 shown;

[0074] (2) Selection of the transfer function of the BP neural network

[0075] The transfer function between the input layer and the hidden layer of the BP neural network adopts the S function:

[0076]

[0077] The transfer function between the hidden layer and the output layer of the BP neural network adopts the linear function:

[0078] y = purelin(s) = s

[0079] (3) Calculate the hidden layer

[0080] Initialize the connection weights W ij between the input layer and the hidden layer, and initialize the hidden layer threshold a; Among them, the connection weights W ij , the hidden layer threshold a are randomly set and continuously adjusted in the subsequent training process along the direction of error reduction;

[0081] According to the input sequence, connection weight W ij and threshold a, calculate the output H of the hidden layer nodes as follows:

[0082]

[0083] In the formula, f is the activation function of the hidden layer, and this function has multiple expressions.

[0084] (4) Calculate the output of the output nodes

[0085] Initialize the connection weight W between the hidden layer and the output layer jk , and initialize the threshold b of the output layer; among them, the connection weight W jk , the threshold b of the output layer are randomly set and continuously adjusted along the direction of error reduction during the subsequent training process;

[0086] According to the output H of the hidden layer nodes in step (3), combined with the connection weight w jk and threshold b, calculate the predicted output O as follows:

[0087]

[0088] (5) Calculate the error of the output nodes

[0089] According to the model predicted output O and the expected output Y calculated above, calculate the model prediction error e as:

[0090] e k = Y k - O k , k = 1, 2,..., m

[0091] Among them, the unit of the model prediction error e is ten thousand · cm -3 ;

[0092] When the model prediction error reaches the set model error requirement, the model training is completed; when the model prediction error does not meet the set model error requirement, update the weight matrix of the RBM neural network, the bias vector of the visible layer, the bias vector of the hidden layer, adjust the weights of the BP neural network, and correct the threshold, and return to step two;

[0093] Step four: Update the connection weights of the RBM neural network, the biases of the visible layer neurons, the biases of the hidden layer neurons, adjust the weights of the BP neural network, and correct the threshold

[0094] (1) Update the connection weight W of the RBM neural network R , the bias a of the visible layer neurons R , the bias b of the hidden layer neurons R ;

[0095] Reconstruct the visual layer, set the reconstructed visual layer as v*, the hidden layer obtained from v* as h*, the learning rate as ε, and the connection weight as W R , the bias a of the visual layer neurons R , the bias b of the hidden layer neurons R The update formula is as follows:

[0096]

[0097] a R ← a R + ε × (v - v * )

[0098]

[0099] (2) Adjust the weights and correct the thresholds of the BP neural network

[0100] According to the prediction error e of the BP neural network model, adjust the network connection weights W ij and W jk as follows:

[0101]

[0102] w jk = w jk + ηH j e k , k = 1, 2,..., m

[0103] According to the prediction error e of the BP neural network model, correct the network node thresholds a, b:

[0104]

[0105] b k = b k + e k , k = 1, 2,..., m

[0106] where η is the learning rate, 1 > η > 0;

[0107] Step Five: Iterate according to Steps Two to Four. The criterion for judging whether the iteration is complete is: the model-set error ≤ 0.05 or the number of cut-off iterations is 500;

[0108] Step Six: Apply the trained RBM model and BP neural network model to predict the pyroelectric ion concentration.

[0109] Prediction results and data analysis of pyroelectric ion concentration

[0110] In order to better verify the accuracy of the RBM model and the BP neural network model algorithm in detecting the concentration of pyroelectric ions, ten groups of controls were set up in this experiment. The results detected by the RBM model and the BP neural network model algorithm and the true value of the pyroelectric ion concentration are shown in Table 1. It can be seen that the difference between the true value of the pyroelectric ions and the measured values detected by the RBM model and the BP neural network model is very small, and the measurement requirements are relatively well completed.

[0111] The curve graph of the pyroelectric ion concentration obtained from on-site detection and the broken line graph of the actual pyroelectric ion concentration are as Figure 5 .

[0112] Table 1 Pyroelectric ion concentration data detected in on-site application

[0113]

[0114] As can be seen from Table 1, the difference between the ten detection data and the true value is very small, and the accuracy rate can reach 99.8%. The results show that the pyroelectric ion measurement proposed in the present invention has good accuracy and can meet the requirements of practical applications.

[0115] The above are only specific embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting the concentration of thermally released ions in a substation, characterized in that: The specific steps include: Step 1: Collect an air sample in the substation environment and release it into a cloud chamber; irradiate the air sample in the cloud chamber with a light wave emitter, and by changing the wavelength λ and the incident angle τ of the incident light wave, use the sensors in the cloud chamber to collect data on the object distance u, the interference fringe frequency F, and the fringe circle diameter Φ; Step 2: Construct an RBM model, input the wavelength λ of the incident light wave, the incident angle τ, and the object distance u, the interference fringe frequency F, and the fringe circle diameter Φ collected by the sensors in the cloud chamber for training. While retaining the characteristics of all the data, reduce the amount of data, and extract the data incident angle τ’, the system object distance u’, the interference fringe frequency F’, the fringe circle diameter Φ’, and the light wave wavelength λ’; Step 3: Construct a prediction model based on a BP neural network, input the data incident angle τ’, the system object distance u’, the interference fringe frequency F’, the fringe circle diameter Φ’, the light wave wavelength λ’, and the true value of the thermally released ion concentration, and output the predicted concentration of thermally released ions; Step 4: Output a comparative analysis of the predicted concentration of thermally released ions and the true value of the thermally released ion concentration to obtain the model prediction error; when the model prediction error meets the set model error requirement, complete the model training; when the model prediction error does not meet the set model error requirement, update the connection weights of the RBM model, the biases of the visible layer neurons, and the biases of the hidden layer neurons, and adjust the weights and thresholds of the BP neural network for correction, and return to Step 2; Step 5: When the model prediction error meets the set model error requirement or reaches the number of iterations, complete the model training, and apply the trained RBM model and BP neural network model to predict the concentration of thermally released ions.

2. The method for detecting the concentration of thermally released ions in a substation according to claim 1, characterized in that: In Step 2, by pre-training the RBM model, data reduction with feature retention for the input collected data is carried out, and the specific process is as follows: (1) The input training sample data is used to obtain the output vector of the first hidden layer through training with the weight matrix and the bias vector; the first hidden layer units are driven by the data vector to obtain random binary states, and the values of the hidden layer units have only two states, 0 and 1; the probability that the value of the hidden layer unit is 1 is: where θ = {W R , a R , b R} represents the parameter set of the restricted Boltzmann machine, a R represents the visible layer neuron bias, b Rj represents the hidden layer neuron bias, W R,i,j represents the connection weight between the visible layer and the hidden layer nodes, v i represents the state value of the i-th node in the visible layer, h j the state of the hidden layer neuron; when the value of the hidden layer neuron is taken as 1, it represents the on state, otherwise the value is 0, representing the off state; (2) The output vector of the first hidden layer is input to the visible layer of the second-layer RBM, and Step (1) is continued to be repeated for training to obtain the output vector of the second hidden layer, and the reduced data with feature retention is extracted.

3. The method for detecting the concentration of thermally released ions in a substation according to claim 1, characterized in that: Construct a prediction model based on a BP neural network, and the specific process is as follows: (1) According to the system input sequence x1, x2,..., x c , where x is the data after being reduced by the RBM model, x c = {τ c ’, u c ’, F c ’, Φ c ’, λ c ’}; and the input true pyroelectric ion concentration value Y; Among them, the number of input nodes n of the BP neural network is 5, the number of output nodes is m = 1, and the number of hidden layer nodes is L. The calculation formula for the number of neurons L in the hidden layer is: L = 2n + 1; (2) Selection of the transfer function of the BP neural network The transfer function between the input layer and the hidden layer of the BP neural network adopts the S function: The transfer function between the hidden layer and the output layer of the BP neural network adopts the linear function: y = purelin(s) = s (3) Calculate the hidden layer Initialize the connection weights W between the input layer and the hidden layer ij , and initialize the threshold a of the hidden layer; among them, the connection weights W ij , the threshold a of the hidden layer are randomly set and continuously adjusted along the direction of decreasing error during the subsequent training process; According to the input sequence, connection weight W ij and threshold a, calculate the output H of the hidden layer nodes as follows: In the formula, f is the activation function of the hidden layer, and there are various expressions for this function; (4) Calculate the output of the output node Initialize the connection weights W between the hidden layer and the output layer jk , and initialize the output layer threshold b; among them, the connection weights W jk , the output layer threshold b are randomly set and continuously adjusted along the direction of error reduction during subsequent training; According to the output H of the hidden layer nodes in step (3), combined with the connection weights w jk and the threshold b, calculate the predicted output O as follows:

4. The detection method for the concentration of thermally released ions in a substation according to claim 1, wherein: The model prediction error e is obtained by calculating based on the model prediction output O and the expected output Y, and the formula is: e k = Y k - O k , k = 1, 2, ..., m Among them, the unit of the model prediction error e is ten thousand · cm -3 .

5. The detection method of the thermally ionized ion concentration in a substation according to claim 1, characterized in that: Update the connection weights of the RBM model, the biases of the visible layer neurons, and the biases of the hidden layer neurons, specifically as follows: Reconstruct the visual layer, set the reconstructed visual layer as v*, the hidden layer obtained from v* as h*, the learning rate as ε, and the connection weight as W R , the bias a of the visual layer neurons R , the bias b of the hidden layer neurons R The update formulas are as follows: a R ←a R +ε×(v - v * ) 6. The detection method for the concentration of thermally released ions in a substation according to claim 1, wherein: The weight adjustment and threshold correction of the BP neural network are specifically as follows: Adjust the network connection weights W between each layer according to the prediction error e of the BP neural network model ij and W jk as follows: w jk = w jk + ηH j e k , k = 1, 2, ..., m Correct the network node thresholds a and b according to the model prediction error e of the BP neural network: b k = b k + e k , k = 1, 2, ..., m where η is the learning rate, and 1 > η > 0.

Images