A substation fire early identification warning system embedded with AI algorithm

Abstract

The application provides a substation fire early identification alarm system embedded with an AI algorithm, and relates to the technical field of data processing, which comprises the following steps: determining a containing state by calculating and analyzing the positional relationship between a grid point and each side of a triangle, calculating a local space weight according to the containing state and the distance relationship between the analysis grid and three fixed space coordinate points, obtaining a global space calibration weight by aggregating the local space weights of all analysis grids, and performing a weighted operation on the fire early risk index by using the global space calibration weight to generate a calibrated fire early risk index; the calibrated fire early risk index is compared with a preset dynamic probability threshold to generate a preliminary fire warning signal; the preliminary fire warning signal is verified, and a final fire early identification alarm instruction is generated and an emergency disposal plan is started after the verification is passed. The application can realize accurate identification and rapid alarm of a substation fire in an early stage, effectively reduce the false alarm rate and timely start emergency disposal.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a substation fire early identification alarm system embedded with an AI algorithm. BACKGROUND

[0002] In the daily monitoring of substations, common fire alarms (such as sensors responding to changes in infrared radiation or ion concentration) are usually deployed in specific locations to monitor single physical parameters such as temperature, smoke, etc. However, in complex scenarios such as near the main transformer, where oil vapor volatilization or transient arc radiation generated by high-voltage switch operation exists, such alarms relying on single phenomenon detection may not be able to fully distinguish between normal operating interference and real early fire characteristics, and their early warning results sometimes do not match the actual risk importance of the device area. For example, the pre-warning for the main transformer area and the pre-warning for the ordinary cable trench are often treated equally in traditional systems, making it difficult for maintenance personnel to quickly determine the priority of disposal and affecting the efficiency of early response. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a substation fire early identification alarm system embedded with an AI algorithm, which can realize accurate identification and rapid alarm of substation fire in early stage, effectively reduce the false alarm rate and timely start emergency disposal.

[0004] To solve the above technical problems, the technical scheme of the present application is as follows:

[0005] In a first aspect, a substation fire early identification alarm system embedded with an AI algorithm comprises:

[0006] A collection module for collecting multi-source heterogeneous real-time monitoring data in a substation;

[0007] A fusion module for pre-processing the multi-source heterogeneous real-time monitoring data, extracting multi-dimensional dynamic features, and combining the multi-dimensional dynamic features into a high-dimensional feature vector;

[0008] A calculation module for inputting the high-dimensional feature vector into a pre-trained embedded AI analysis model, analyzing the high-dimensional feature vector and generating a fire early risk index;

[0009] The calibration module is used for selecting three fixed spatial coordinate points respectively located in a core equipment area, a main transformer area and a high-voltage switch area on a two-dimensional plan view of a substation monitoring area, connecting the three fixed spatial coordinate points to form a triangular convex polygon area, dividing the triangular convex polygon area to generate an analysis grid, calculating a containing state of a grid point and a position relationship of each side of the triangle, calculating a local spatial weight according to the containing state and a distance relationship between the analysis grid and the three fixed spatial coordinate points, aggregating the local spatial weights of all analysis grids to obtain a global spatial calibration weight, and performing weighted operation on the fire early risk index by using the global spatial calibration weight to generate a calibrated fire early risk index;

[0010] The generation module is used for comparing the calibrated fire early risk index with a preset dynamic probability threshold to generate a preliminary fire early warning signal.

[0011] The verification module is used for verifying the preliminary fire early warning signal, generating a final fire early identification alarm instruction and starting an emergency disposal plan after verification.

[0012] In a second aspect, a computing device includes:

[0013] One or more processors;

[0014] A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the system.

[0015] In a third aspect, a computer readable storage medium has a program stored therein, and the program is executed by a processor to implement the system.

[0016] The above scheme of the present application at least has the following beneficial effects:

[0017] By defining the spatial coordinates of the key areas such as the core equipment area and the main transformer area and calibrating the grid weight, the identification sensitivity of the fire risk in the key areas is strengthened, the risk misjudgment caused by the difference in spatial layout is avoided, the risk index is more suitable for the actual equipment distribution and safety priority of the substation, a lightweight embedded AI analysis model is adopted, the requirements of the substation on the equipment response speed, data privacy protection and edge computing deployment are met, and the alarm instruction is generated quickly. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a substation fire early identification alarm system schematic diagram provided by an embodiment of the present application.

[0019] Figure 2The flowchart of the process of verifying the preliminary fire warning signal, generating the final fire early identification alarm instruction and starting the emergency disposal plan after the verification is passed is provided by the embodiment of the present application. DETAILED DESCRIPTION

[0020] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0021] As Figure 1 shown, the embodiment of the present application proposes a substation fire early identification alarm system embedded with an AI algorithm, comprising:

[0022] The acquisition module is configured to acquire multi-source heterogeneous real-time monitoring data in the substation.

[0023] The fusion module is configured to pre-process the multi-source heterogeneous real-time monitoring data, extract multi-dimensional dynamic features, and integrate the multi-dimensional dynamic features into a high-dimensional feature vector.

[0024] The calculation module is configured to input the high-dimensional feature vector into a pre-trained embedded AI analysis model, analyze the high-dimensional feature vector, and generate a fire early risk index.

[0025] The calibration module is configured to select three fixed spatial coordinate points respectively located in a core equipment area, a main transformer area, and a high-voltage switch area on a two-dimensional plan view of the substation monitoring area, connect the three fixed spatial coordinate points to form a triangular convex polygon area, divide the triangular convex polygon area to generate an analysis grid, calculate the position relationship of the analysis grid points and the sides of the triangle to determine a containment state, calculate a local spatial weight according to the containment state and the distance relationship between the analysis grid and the three fixed spatial coordinate points, aggregate the local spatial weights of all analysis grids to obtain a global spatial calibration weight, and perform weighted operation on the fire early risk index using the global spatial calibration weight to generate a calibrated fire early risk index.

[0026] The generation module is configured to compare the calibrated fire early risk index with a preset dynamic probability threshold to generate a preliminary fire warning signal.

[0027] The verification module is configured to verify the preliminary fire warning signal, generate a final fire early identification alarm instruction after the verification is passed, and start an emergency disposal plan.

[0028] In the embodiment of the present application, the identification sensitivity of the fire risk of the key area is strengthened by the spatial coordinate definition and the grid weight calibration of the key areas such as the core equipment area and the main transformer area, the risk misjudgment caused by the difference in spatial layout is avoided, the risk index is more in line with the actual equipment distribution and safety priority of the substation, a lightweight embedded AI analysis model is used to meet the requirements of the substation on the equipment response speed, data privacy protection and edge computing deployment, and the alarm instruction is generated quickly.

[0029] In a preferred embodiment of the present application, multi-source heterogeneous real-time monitoring data in the substation is collected, specifically including: collecting equipment including high-definition cameras, infrared thermal imagers, and temperature, smoke, humidity, and gas sensors, and attaching to risk points in each area during deployment: the high-definition camera is installed above the core equipment control cabinet, on the side of the main transformer oil tank, and other positions, the lens is aimed at the equipment body and the key connection parts, video image data with a resolution of 1,920x1,008 pixels is collected at a frame rate of 30 frames per second, to ensure that early details such as discoloration of the insulation layer and weak smoke are captured; the infrared thermal imager is deployed at the regional commanding heights, covering the main transformer winding joints, high-voltage switch contacts and other heat-prone points, infrared thermal imaging data with a temperature range of -20 to 300 degrees Celsius and a temperature resolution of 0.1 degrees Celsius is collected at a frame rate of 10 frames per second, to accurately capture the small temperature rise of the equipment; various environmental sensors are installed at positions such as equipment vents, oil pillows, and insulation sleeves, and data is collected at a frequency of once per second, wherein the temperature sensor has a range of -40 to 125 degrees Celsius, the smoke sensor has a range of 0 to 10 mg / m3, the humidity sensor has a range of 0% to 100% relative humidity, and the characteristic gas sensor has a range of 0 to 1,000 ppm; after collection, all data is synchronously packaged according to the millisecond-level timestamp, and identification such as data type and equipment number is added to form a unified data stream, which is transmitted to the data processing unit.

[0030] In this embodiment, through multi-source data collection, the blind area problem of single parameter monitoring is solved, and image changes and environmental parameter abnormalities in the early stage of fire are captured.

[0031] In a preferred embodiment of the present application, the multi-source heterogeneous real-time monitoring data is preprocessed, and multi-dimensional dynamic features are extracted, and the multi-dimensional dynamic features are combined into a high-dimensional feature vector, which can include:

[0032] The collected video image data and infrared thermal imaging data are processed to obtain standardized image data, and the standardized image data is analyzed to generate a set of geometric moments, specifically including: processing image data is to use the weighted average method to gray the color video image, the calculation method is to multiply the red channel value by 0.3, add the green channel value multiplied by 0.59, and then add the blue channel value multiplied by 0.11, and the sum is the single channel gray value, so as to highlight the gray difference between the device and the background; the infrared thermal imaging data is mapped to the gray value according to the linear proportion, the calculation method is to subtract the difference between the measured temperature and minus twenty degrees Celsius from three hundred degrees Celsius, and then multiply by two hundred and fifty-five, to obtain the corresponding gray value, which directly reflects the device temperature rise; a five-by-five Gaussian filter is used to remove image noise, the specific calculation process is as follows: first, a five-by-five filter kernel is constructed, the center position of the filter kernel is set as the coordinate origin, and the remaining twenty-four positions are marked as negative two, negative one, zero, positive one and positive two in the horizontal and vertical directions, that is, the coordinates of each position are represented by (horizontal coordinate, vertical coordinate), such as the first position on the right side of the center is (1, 0), the upper left corner position is (-2, -2), etc.; calculate the weight of each position according to the Gaussian function, the calculation method of the Gaussian function is: first, calculate the sum of the square of the horizontal coordinate and the square of the vertical coordinate of the position, then divide the sum by two times the square of the standard deviation (the standard deviation is set to 0.8, that is, two times 0.8 squared is equal to 1.208), take the negative of the result as the exponential of the natural constant, then multiply the exponential result by one divided by the quotient of two times pi times the square of the standard deviation, which is the initial weight of the position; for example, the weight calculation of the center position (0, 0): the square of the horizontal coordinate plus the square of the vertical coordinate is zero, divided by 1.208 is zero, the zero power of the natural constant is one, multiplied by one divided by the quotient of two times 3.1416 times the square of 0.8, the initial weight is about 0.1201; after calculating the initial weights of the twenty-five positions in this way, add all the initial weights to get the total weight, and then divide the initial weight of each position by the total weight to complete the weight normalization (ensure that the sum of all weights is one, to avoid the overall brightness change of the filtered image); after normalization, the weight range of each position is as follows: the weight of the center position (0, 0) is the largest, about 0.1201; the weights of the eight positions adjacent to the center (such as (1, 0) (0, 1) (-1, 0) etc.) are consistent, about 0.1083; the weights of the eight positions one position away from the center (such as (2, 0) (1, 1) (0, 2) etc.) are about 0.0540; the weights of the four diagonal positions ((1, 1) (1, -1) (-1, 1) (-1, -1)) are about 0.0812; the weights of the four corner positions ((2, 2) (2, -2) (-2, 2) (-2, -2)) are the smallest, about 0.0304.In filtering, taking twenty-five pixels in a five-by-five range around each pixel in the image as the center, the gray value of each pixel is multiplied by the normalized weight of the corresponding position of the filter kernel, and the twenty-five product results are added to obtain the gray value of the center pixel after filtering. All pixels are traversed in turn to complete denoising.

[0033] After denoising, the image is scaled to sixty-four by sixty-four pixels by bilinear interpolation, and the specific calculation process is as follows: the scaling ratio is determined, the original image size is one thousand nine hundred and twenty by one thousand and eight pixels, the target size is sixty-four by sixty-four pixels, the horizontal scaling ratio is sixty-four divided by one thousand nine hundred and twenty equal to one hundred and twenty-fifth, and the vertical scaling ratio is sixty-four divided by one thousand and eight equal to zero point zero six three four. For each pixel of the target image, first calculate its corresponding position in the original image, the calculation method is the horizontal coordinate of the target pixel divided by the horizontal scaling ratio to obtain the horizontal floating point coordinate of the original image; the vertical coordinate of the target pixel is divided by the vertical scaling ratio to obtain the vertical floating point coordinate of the original image; for example, the (10, 20) pixel of the target image, the horizontal coordinate of the original image is ten divided by one hundred and twenty-fifth equal to one thousand two hundred, and the vertical coordinate is twenty divided by zero point zero six three four equal to three hundred and seventeen point nine five; take the nearest four integer coordinate pixels of the floating point coordinate in the original image, that is, the horizontal coordinate takes the integer less than and close to the floating point coordinate and the integer greater than and close to the floating point coordinate, and the vertical coordinate is the same, to obtain the coordinates and corresponding gray values of the four pixels, which are recorded as top left (x1, y1, gray value 1), top right (x2, y2, gray value 2), bottom left (x3, y3, gray value 3), and bottom right (x4, y4, gray value 4), wherein x2 is equal to x1 plus one, and y3 is equal to y1 plus one; then calculate the horizontal interpolation weight, which is obtained by subtracting x1 from the horizontal floating point coordinate of the original image to obtain the decimal part a (the value range of a is zero to one), and the horizontal weight is (one minus a) and a, both of which take values in the range of zero to one, and the sum is one; calculate the vertical interpolation weight, which is obtained by subtracting y1 from the vertical floating point coordinate of the original image to obtain the decimal part b (the value range of b is zero to one), and the vertical weight is (one minus b) and b, both of which take values in the range of zero to one, and the sum is one; first, horizontal interpolation is performed, and the gray values of the top left and top right in the floating point horizontal coordinate are calculated, that is, the top left horizontal interpolation gray value is gray value 1 multiplied by (one minus a) plus gray value 2 multiplied by a, and the top right horizontal interpolation gray value is gray value 3 multiplied by (one minus a) plus gray value 4 multiplied by a; then, vertical interpolation is performed, and the gray value of the target pixel is obtained by multiplying the top left horizontal interpolation gray value by (one minus b) and adding the top right horizontal interpolation gray value by b; the gray values of all target pixels are calculated in turn to obtain a sixty-four by sixty-four pixel standardized image.

[0034] The zero to third order geometric moments are calculated for the standardized image, the zero order geometric moment is the sum of all pixel gray values; the first order geometric moment contains two values, which are the sum of the product of all pixel gray values and corresponding x coordinate and the sum of the product of all pixel gray values and corresponding y coordinate; the second order geometric moment contains three values, which are the sum of the product of all pixel gray values and corresponding x coordinate square, the sum of the product of all pixel gray values and corresponding x coordinate and y coordinate, and the sum of the product of all pixel gray values and corresponding y coordinate square; the third order geometric moment contains four values, forming a geometric moment set of ten elements.

[0035] Based on the geometric moment set, shape moment features are calculated and combined; and the collected environmental sensor data is processed to extract numerical features, specifically including: calculating central moments based on geometric moments, first calculating the x direction coordinate mean value as the x related value in the first order geometric moment divided by the zero order geometric moment, and the y direction coordinate mean value as the y related value in the first order geometric moment divided by the zero order geometric moment, then subtracting the x direction mean value from the x coordinate of each pixel and the y direction mean value from the y coordinate, and obtaining the central moments according to the calculation method of corresponding order geometric moments; then calculating the normalized central moments, dividing each central moment by the t power of the zero order geometric moment, where t is the corresponding moment order rounded up to the nearest integer, selecting features with a correlation greater than zero point seven with early fire features, and combining them into twelve-dimensional shape moment features.

[0036] After completing the image feature extraction, the environmental sensor data is first processed according to the three times standard deviation criterion to remove outliers, collects the last one hundred times of sensor data, calculates the mean value as the sum of all data divided by one hundred, calculates the standard deviation as the sum of the square of the difference between each data and the mean value, and then divides by one hundred to get the variance, and the standard deviation is obtained by taking the square root of the variance, the data outside the range of mean value minus three times standard deviation to mean value plus three times standard deviation is determined as an outlier, the outlier is replaced by the mean value of the first five times of data collection, and the mean value of the first five times is the sum of the five times of data divided by five; then the data is standardized to the interval of zero to one, the calculation method is to subtract the minimum value of the sensor measurement range from each data value, and the difference is divided by the difference between the maximum value and the minimum value of the measurement range, for example, the temperature sensor is subtracted by minus forty degrees Celsius, and then divided by one hundred and twenty-five degrees Celsius minus minus forty degrees Celsius; the standardized temperature, smoke concentration, humidity, and characteristic gas concentration are extracted as four-dimensional numerical features.

[0037] The shape moment feature is fused with the numerical feature to generate a multi-dimensional dynamic feature, and the multi-dimensional dynamic feature is integrated into a high-dimensional feature vector, specifically including: performing feature timestamp alignment to ensure that the fused features come from the same monitoring moment; the shape moment feature is a 12-dimensional feature calculated based on a standardized image, each frame of image corresponds to a set of 12-dimensional data, and carries a millisecond-level timestamp of image acquisition; the numerical feature is a 4-dimensional feature (temperature, smoke concentration, humidity, and characteristic gas concentration) obtained after preprocessing of an environmental sensor, a set of 4-dimensional data is collected every second, and also carries a millisecond-level timestamp; the data processing unit matches the 12-dimensional shape moment feature and the 4-dimensional numerical feature at the same moment to associate them into a set of data to be fused, if there is a timestamp deviation (not more than 50 milliseconds at most) between the image frame and the sensor data, a linear interpolation method is used to complete the sensor data to ensure the time consistency of each set of data to be fused.

[0038] The feature standardization secondary calibration is performed to eliminate the dimensional difference of two types of features; although the shape moment features are calculated by the normalized central moment, the numerical range may still deviate from the features standardized by the sensor, and need to be calibrated to the interval of 0 to 1; for the 12-dimensional shape moment features, the data processing unit calls the shape moment feature data under 300 groups of normal working conditions in history to calculate the minimum value and the maximum value of each dimension, and adopts the minimum-maximum normalization method for calibration, and the calculation method is to subtract the historical minimum value of a certain dimension of the current shape moment feature from the value of the dimension, and divide the difference value by the difference between the historical maximum value and the historical minimum value of the dimension. For the 4-dimensional numerical features, the results standardized to the interval of 0 to 1 are directly used, and the standardization calculation is repeatedly performed only for the data completed by interpolation to ensure that the two types of features are in the same numerical scale; the weighted fusion is implemented to generate multi-dimensional dynamic features, and the features with high correlation degree with the fire are highlighted; combined with the experience of the substation fire warning, the three dimensions (corresponding to the second-order cross moment, the third-order x moment and the third-order y moment in the normalized central moment) in the shape moment features reflecting the irregular edge and the two dimensions (smoke concentration and characteristic gas concentration) in the numerical features reflecting the fire precursor are more critical to the fire identification, and the weights need to be increased; the specific weight distribution is that the three key dimensions in the shape moment features are each allocated a weight of 0.08, and the remaining nine dimensions are each allocated a weight of 0.06; the smoke concentration and the characteristic gas concentration in the numerical features are each allocated a weight of 0.1, and the temperature and the humidity are each allocated a weight of 0.07, and the total weight of all dimensions is 1; in the fusion calculation, the calibrated value of each feature dimension is multiplied by the corresponding weight, and the 12 shape moment features after weighting are combined with the 4 numerical features after weighting in sequence to form a 16-dimensional multi-dimensional dynamic feature; the multi-dimensional dynamic features are combined into a high-dimensional feature vector, that is, the data processing unit adopts the feature tiling method, arranges the 16-dimensional multi-dimensional dynamic features in the fixed sequence of shape moment features in front and numerical features behind, wherein the shape moment features are arranged in the order of zero-order to third-order normalized central moment, and the numerical features are arranged in the order of temperature, smoke concentration, humidity and characteristic gas concentration, and finally a one-dimensional high-dimensional feature vector with 16 columns is formed.

[0039] In this embodiment, the targeted preprocessing process eliminates the field interference, and the extracted multi-dimensional features can accurately reflect the early comprehensive signals of the fire.

[0040] In a preferred embodiment of the present application, the high-dimensional feature vector is input into a pre-trained embedded AI analysis model, the high-dimensional feature vector is analyzed, and a fire early risk index is generated, which can include:

[0041] The high-dimensional feature vector is a row of 16 columns of data obtained by fusing the shape matrix and numerical features in the early stage, corresponding to 16 fire-related feature dimensions; the input layer of the embedded AI analysis model is adapted to this dimension, and 16 neurons are set, each of which separately receives a dimension of feature vector data to ensure that the feature transmission is not missed; considering that the subsequent hidden layer adopts a convolutional calculation logic that requires three-dimensional input, the one-dimensional vector of one row and 16 columns is converted into a three-dimensional tensor of one row and 16 columns when the feature tensor is reconstructed, the first dimension represents a single sample of a single monitoring, the second dimension corresponds to 16 feature dimensions, and the third dimension is the default channel number of the image feature. This structure fully adapts to the calculation needs of the hidden layer; the core of standardization scaling is to eliminate the dimensional difference. First, the mean of the 16 elements in the tensor is calculated, and the mean is obtained by dividing the sum of the 16 elements by 16; the standard deviation is calculated by first calculating the difference between each element and the mean, squaring each difference, adding all the squared values, dividing by 16 to get the variance, and then taking the square root of the variance to get the standard deviation; finally, each element is subtracted from the mean, and the difference is divided by the standard deviation to obtain a standardized feature tensor with a distribution in the interval of -1 to 1.

[0042] The standardized feature tensor is input into the hidden layer network, and layer-by-layer nonlinear transformation and high-order feature synthesis are performed in the hidden layer network to generate a high-level comprehensive feature tensor, which specifically includes: considering the complexity of the substation fire early features (such as weak smoke shape, local temperature rise) and the limited computing power of embedded devices, the hidden layer adopts a three-order dimension reduction fully connected layer structure to gradually refine key features; the first fully connected layer is provided with 128 neurons, which receives the standardized feature tensor and first performs linear weighting calculation, i.e., each feature element is multiplied by the weight coefficient of the corresponding neuron (the weight is initialized by Xavier to ensure consistent input and output variance), and then the weighted results are added to the bias item of this layer (the initial value is set to zero point zero one); after linear calculation, the ReLU activation function transformation is performed, the input value greater than zero is directly output, and the value less than or equal to zero is output as zero, which highlights the nonlinear correlation of key features such as temperature anomalies and smoke profiles through this transformation; batch normalization is performed to stabilize the feature distribution, i.e., the mean (sum of all feature values divided by 128) and variance (sum of squares of differences between each feature value and the mean divided by 128) of 128 output features are calculated, then each feature value is divided by the square root of the variance after being subtracted from the mean, and finally multiplied by the scaling coefficient one and added by the offset zero to stabilize the feature value in the interval of -1 to 1.

[0043] The 128 features output by the first layer are transmitted to the second layer full connection layer, which has 64 neurons. The linear weighting, ReLU activation and batch normalization operations are repeated. When linearly weighting, the 128 input features are multiplied by the corresponding weights (also Xavier initialized) and added together, and then a bias is added. The activation and normalization processes are the same as those of the first layer, and a 64-dimensional feature is output, realizing feature dimension reduction and information condensation. The 64-dimensional features output by the second layer are transmitted to the third layer full connection layer, which has 32 neurons. When linearly weighting, the 64 input features are multiplied by the weights and added together, and then a bias is added. Considering that there may be weak negative features (such as local humidity abnormally decreasing) in the early stage of a fire, the activation function uses LeakyReLU, which outputs the input value when the input value is greater than zero, and outputs the input value multiplied by zero point one when the input value is less than or equal to zero, avoiding the complete suppression of negative features by ReLU. Batch normalization is performed, and finally a 32-dimensional high-level comprehensive feature tensor is output.

[0044] The overall model has an input layer, three layers of hidden full connection layers, and an output layer architecture. The output layer has one Sigmoid function unit for outputting the risk probability. The training data is derived from three years of historical monitoring data of the substation, totaling 5,000 groups. 1,000 groups are simulated fire scene data (such as igniting a cotton yarn dipped in oil to simulate cable fire, and heating elements to simulate equipment overheating), and 4,000 groups are normal operation scene data (covering different seasons, device loads, and weather conditions). All data are processed into high-dimensional feature vectors and labeled with tags: fire scene label one and normal scene label zero. The data is divided into training set, validation set, and test set in a ratio of 7:2:1: 3,500 groups of training set are used to update model parameters, 1,000 groups of validation set are used to adjust hyperparameters, and 500 groups of test set are used for final performance evaluation. The training uses a stochastic gradient descent optimizer, with a batch size of 32 (adapted to the computing power of embedded devices), and an initial learning rate of 0.01. The learning rate is multiplied by 0.1 to decay every 50 rounds of training to avoid parameter fluctuations. The loss function uses the cross-entropy loss function. When calculating, for each training data: if the label is one, calculate the natural logarithm of the model prediction probability; if the label is zero, calculate the difference between one and the natural logarithm of the prediction probability; take the negative of the sum of the two results, and then divide the sum of all training set data by the total number of training sets to get the average loss value. After each round of training, the average loss value is calculated using the validation set. When the validation set loss value does not decrease for 10 consecutive rounds (with a decrease of less than 0.0001), the training is stopped, and the model reaches the optimal generalization ability. Overfitting (low training set loss but high validation set loss) or underfitting (high loss for both) is avoided. The weight coefficients and bias terms of each layer at this time are saved to form a pre-trained model adapted to the substation scene.

[0045] The high-level comprehensive feature tensor is transmitted to the Sigmoid function unit of the output layer, linear weighted summation and probability transformation are performed, and a scalar probability value is generated; nonlinear amplification and range mapping are performed on the scalar probability value to generate a fire early risk index, which specifically includes: the high-level comprehensive feature tensor is thirty-two-dimensional, and the Sigmoid function unit of the output layer first performs linear weighted summation on it, thirty-two feature elements are multiplied by the weight coefficients (fixed after training) of the output layer respectively, and all weighted results are added to the bias term (fixed after training) of the output layer to obtain a linear output value; the linear output value is input into the Sigmoid function for probability transformation, and the calculation method is to take the negative linear output value power of one divided by one plus the natural constant (taking the approximate value of two point seven one two), to obtain a scalar probability value between zero and one, the closer the value is to one, the higher the fire risk of the current scene represents; considering that the operation and maintenance personnel of the substation need to intuitively distinguish between normal conditions, slight abnormalities and early fire risks, the scalar probability value is subjected to targeted nonlinear amplification, and the core logic of selecting one point two power operation is to neither distort the risk gradient of the high probability interval too much nor to widen the difference in the low probability interval (early fire in the substation is mostly represented as a low probability slight abnormality, which needs to be accurately identified).

[0046] The specific calculation process is to perform power operation on the scalar probability value as the base and one point two as the exponent, and first calculate the product of the first power (i.e. itself) of the scalar probability value and the second power of zero point two, wherein the second power of zero point two is calculated by natural logarithm conversion, i.e. first taking the natural logarithm of the target probability value, multiplying the result by zero point two, and then calculating the exponent to obtain the result of the second power of zero point two, and then multiplying the result by the first power of the scalar probability value to obtain the one point two power operation result; for example, when the scalar probability value is zero point five, first take the natural logarithm to be about negative zero point six six nine three, multiply by zero point two to get negative zero point one three three eight, the negative zero point one three three eight power of the natural constant is about zero point eight seven three, and then multiply by zero point five to get zero point four three six five (the example value of the original expression is slightly different due to approximate calculation, and the actual accurate operation is based on this step); when the scalar probability value is zero point six, the natural logarithm is about negative zero point five four zero four, multiplied by zero point two to get negative zero point one zero eight one, etc., the negative zero point one zero eight one power of the natural constant is about zero point eight nine two, and multiplied by zero point six to get zero point five six five, the difference between the two is expanded from zero point one to zero point one three three forty-five, effectively distinguishing between slight abnormalities and normal states; finally, the amplified result is multiplied by one hundred to map to the range of zero to one hundred to generate the fire early risk index.

[0047] In this embodiment, the continuous collection, extraction and analysis process and intuitive risk index improve the fire early identification efficiency, clearly define the early warning priority, provide support for the operation and maintenance personnel to quickly handle, and solve the problem of low system response efficiency.

[0048] In a preferred embodiment of the present application, three fixed spatial coordinate points respectively located in the core equipment area, the main transformer area and the high-voltage switch area are selected on the two-dimensional plan view of the substation monitoring area, the three fixed spatial coordinate points are connected to form a triangular convex polygon area, the triangular convex polygon area is divided to generate an analysis grid, the position relationship of the analysis grid points and the edges of the triangle is calculated to determine the inclusion state, the local spatial weight is calculated according to the inclusion state and the distance relationship between the analysis grid and the three fixed spatial coordinate points, the global spatial calibration weight is obtained by aggregating the local spatial weights of all analysis grids, and the calibrated fire early risk index is generated by using the global spatial calibration weight to perform weighted operation on the fire early risk index, which can include:

[0049] Three fixed spatial coordinate points respectively located at the geometric center point of the core equipment area, the geometric center point of the main transformer area and the geometric center point of the high-voltage switch area are selected; the three fixed spatial coordinate points are connected to form a triangular convex polygon area covering the key monitoring objects, specifically including: selecting three core monitoring points on the two-dimensional plan view of the substation, corresponding to the geometric center points of the core equipment area, the main transformer area and the high-voltage switch area, that is, these three areas are fire-prone areas, and the area covering them is the key monitoring range; the center point coordinates are obtained by measuring the boundary vertices of the area by a laser range finder: taking the core equipment area as an example, the horizontal coordinates and the vertical coordinates of the four vertices of the rectangular boundary are measured, the horizontal coordinates of the four vertices are all added and then divided by four to obtain the horizontal coordinates of the center point, and the vertical coordinates of the four vertices are all added and then divided by four to obtain the vertical coordinates of the center point; the main transformer area and the high-voltage switch area use the same method, and the actual measurement determines the coordinates as core equipment area (ten meters, eight meters), main transformer area (twenty-five meters, fifteen meters), and high-voltage switch area (eighteen meters, twenty-eight meters); the three center points are connected in sequence by straight lines to form a triangular convex polygon key monitoring area.

[0050] A uniform grid covering the outer rectangle of the triangular convex polygon region is generated in the planar space of the triangular convex polygon region to obtain a set of analysis grids; according to the vertex coordinates of the analysis grids, the geometric center point coordinates of each analysis grid are calculated, specifically including: for refined area risk assessment, a uniform grid covering the outer rectangle of the region is generated; the upper left corner coordinates of the outer rectangle are the minimum value of the horizontal coordinates of the three center points minus one meter, that is, ten meters minus one meter equals nine meters, and the minimum value of the vertical coordinates minus one meter, that is, eight meters minus one meter equals seven meters; the right lower corner coordinates are the maximum value of the horizontal coordinates plus one meter, that is, twenty-five meters plus one meter equals twenty-six meters, and the maximum value of the vertical coordinates plus one meter, that is, twenty-eight meters plus one meter equals twenty-nine meters; the grid size is set to zero point five meters by zero point five meters, which is suitable for the positioning accuracy of the equipment; each grid is an analysis grid; the geometric center point coordinates of each grid are calculated by adding the horizontal coordinates of the upper left corner of the grid to the horizontal coordinates of the lower right corner, and the sum is divided by two as the horizontal coordinates of the center point; the vertical coordinates of the upper left corner of the grid are added to the vertical coordinates of the lower right corner, and the sum is divided by two as the vertical coordinates of the center point, such as the upper left corner (nine meters, seven meters) and the lower right corner (nine point five meters, seven point five meters) of the grid. The center point is (nine point two five meters, seven point two five meters).

[0051] According to the position relationship between the geometric center point coordinates and the three edges of the triangular convex polygon region, three position judgment values are calculated; according to the consistency of the signs of the three position judgment values, the inclusion state of the analysis grid in the triangular convex polygon region is judged, specifically including: when judging whether the grid is in the triangular region, for each edge of the triangle, the relative position of the grid center point is judged by the difference relationship of the horizontal coordinates and the vertical coordinates. Taking the edge composed of the core equipment area (ten meters, eight meters) and the main transformer area (twenty-five meters, fifteen meters) as an example, first calculate the horizontal length of this edge as twenty-five meters minus ten meters equals fifteen meters, and the vertical length as fifteen meters minus eight meters equals seven meters; then calculate the offset correlation value of the grid center point relative to this edge, specifically by subtracting the horizontal coordinates of the core equipment area from the horizontal coordinates of the center point to obtain the horizontal offset amount, and multiplying the vertical length of this edge; subtract the vertical coordinates of the core equipment area from the vertical coordinates of the center point to obtain the vertical offset amount, and multiply the horizontal length of this edge; subtract the calculation result of the latter from the calculation result of the former to obtain the position correlation value corresponding to the edge.

[0052] For the other two sides of the triangle, repeat the above calculation: first calculate the horizontal length and vertical length of the side, then multiply the horizontal offset of the center point relative to one end point of the side by the vertical length of the side, multiply the vertical offset by the horizontal length of the side, and take the difference as the position correlation value of the corresponding side; when the three position correlation values are all non-negative or all non-positive, it means that the grid center point is on the same side of the three sides, i.e. located in the triangular convex polygon region; otherwise, it is determined to be outside the region, such as the center point (fifteen meters, twelve meters), the position correlation value of the side between the core equipment area and the main transformer area is (fifteen minus ten) times seven minus (twelve minus eight) times fifteen, i.e. five times seven minus four times fifteen equals thirty-five minus sixty equals negative twenty-five, and the position correlation values of the other two sides are calculated. If they are all negative, it is determined that the grid is in the region.

[0053] According to the Euclidean distance between the center point coordinates of the analysis grid with the state of being located in the region and the three fixed spatial coordinates, the local spatial weight of each analysis grid is calculated, which specifically includes: only the local spatial weight of the grid in the region is calculated, and the core logic is that the closer to the core equipment, the higher the risk weight; first, calculate the Euclidean distance between the center point of the grid and the three fixed points, the calculation method is to subtract the horizontal coordinate of the fixed point from the horizontal coordinate of the center point of the grid, and the difference is squared, then subtract the vertical coordinate of the fixed point from the vertical coordinate of the center point of the grid, and the difference is also squared, and then square the sum of the two square values, i.e. taking the center point (fifteen meters, twelve meters) as an example, the distance from the core equipment area (ten meters, eight meters) is (fifteen minus ten) squared plus (twelve minus eight) squared, and then square, i.e. twenty-five plus sixteen equals forty-one, square root is about six point four meters; the distances from the other two fixed points are calculated in the same way, and the sum of the three distance values is obtained, and then divided by three to obtain the average Euclidean distance; take the maximum distance between the three fixed points as the reference, such as the distance from the high-voltage switch area to the core equipment area is about twenty meters, the calculation method of the local spatial weight is to subtract the average Euclidean distance of the grid from the maximum reference distance, and then divide the difference by the maximum reference distance; combined with the layout of the core area of the substation, the average Euclidean distance is usually between three meters and fifteen meters, and the corresponding local spatial weight value range is zero point two five to zero point eight five, if the average Euclidean distance is three meters, the weight is (twenty minus three) divided by twenty, which equals zero point eight five; if the average Euclidean distance is fifteen meters, the weight is (twenty minus fifteen) divided by twenty, which equals zero point two five.

[0054] Based on the local spatial weight, the global spatial calibration weight is calculated; the global spatial calibration weight is used to perform weighted calibration on the fire early risk index to generate the calibrated fire early risk index, specifically including: aggregating the local weights of all regional grids to obtain the global spatial calibration weight, the calculation method is to add all the local weights of all regional grids, and then divide the sum by the total number of regional grids, for example, the total weight of fifty grids is thirty-four point five, and divided by fifty is zero point six nine; because the regional grids are close to the key areas such as the core equipment area and the main transformer area, the local weight is mostly concentrated between zero point six and zero point nine, so the global spatial calibration weight is stable in the range of zero point six to zero point nine; the generated fire early risk index is weighted by the global weight, that is, the fire early risk index is multiplied by the global weight to obtain the calibrated risk index

[0055] In this embodiment, the spatial weight is combined with the risk priority of the core equipment area by defining the value range of the local and global, so that the evaluation is more in line with the actual layout of the substation.

[0056] In a preferred embodiment of the present application, the calibrated fire early risk index is compared with the preset dynamic probability threshold to generate a preliminary fire warning signal, which can include:

[0057] Two-dimensional plane coordinates of three fixed spatial coordinate points are obtained to generate three sets of coordinate data; based on the coordinates of any two points in the three sets of coordinate data, the perpendicular bisector equation parameters of the two edges are calculated respectively, specifically including: the three fixed spatial coordinate points correspond to the core equipment area, the main transformer area and the high-voltage switch area of the substation respectively, these three areas are high-incidence areas of substation fire, and their coordinates are calculated after measuring the region boundary vertices by a laser range finder; taking the core equipment area as an example, the horizontal coordinates and vertical coordinates of the four vertices of the rectangular boundary of the area are measured using a laser range finder, the horizontal coordinate calculation formula of the center of the core equipment area is , wherein , , , is the horizontal coordinate of the four vertices; the vertical coordinate calculation formula of the center is , wherein , , , is the vertical coordinate of the four vertices, that is, the corresponding coordinates of the four vertices are all added and divided by four; the main transformer area and the high-voltage switch area use the same measurement and calculation method, and the center coordinates are calculated by the coordinates of the four vertices of the corresponding area according to the above formula, and finally three sets of coordinate data are generated, which are the core equipment area coordinates, the main transformer area coordinates and the high-voltage switch area coordinates.

[0058] When calculating the parameters of the perpendicular bisector of the line connecting the core equipment area and the main transformer area, the midpoint coordinates of the line are calculated, the horizontal coordinate of the midpoint is calculated by the formula , and the vertical coordinate of the midpoint is calculated by the formula , that is, the corresponding coordinates of the core equipment area and the main transformer area are added and then divided by two; the slope of the line is calculated by the formula , that is, the vertical difference is obtained by subtracting the vertical coordinate of the core equipment area from the vertical coordinate of the main transformer area, the horizontal difference is obtained by subtracting the horizontal coordinate of the core equipment area from the horizontal coordinate of the main transformer area, and the vertical difference is divided by the horizontal difference; then the slope of the perpendicular bisector is calculated by the formula , that is, negative one is divided by the slope of the line; finally, the parameters of the perpendicular bisector equation are determined based on the midpoint coordinates and the slope kvertical of the perpendicular bisector, and the expression of the perpendicular bisector is constructed in the point-slope form , which ensures that the line passes through the midpoint and is perpendicular to the original line.

[0059] When calculating the parameters of the perpendicular bisector of the line connecting the core equipment area and the high-voltage switch area, the same calculation process as above is repeated. First, the midpoint coordinates are calculated, the horizontal coordinate of the midpoint is calculated by the formula , and the vertical coordinate of the midpoint is calculated by the formula , that is, the corresponding coordinates of the core equipment area and the high-voltage switch area are added and then divided by two; then the slope of the line is calculated by the formula , that is, the vertical difference is obtained by subtracting the vertical coordinate of the core equipment area from the vertical coordinate of the high-voltage switch area, the horizontal difference is obtained by subtracting the horizontal coordinate of the core equipment area from the horizontal coordinate of the high-voltage switch area, and the vertical difference is divided by the horizontal difference; the slope of the perpendicular bisector is calculated by the formula , that is, negative one is divided by the slope of the line; finally, the parameters of the perpendicular bisector equation are determined based on the midpoint coordinates and the slope kvertical of the perpendicular bisector, and the expression of the perpendicular bisector is constructed in the point-slope form .

[0060] According to the parameters of the perpendicular bisectors of the two lines, the intersection coordinates of the two perpendicular bisectors are solved, and the center coordinates of the circumscribed circle are obtained, which specifically includes: according to the parameters of the two perpendicular bisectors, the expressions of the two perpendicular bisectors are solved to obtain the intersection coordinates, which are the center coordinates of the circumscribed circle; when solving, first, the equation parameters of the two perpendicular bisectors are arranged into standard binary linear relations, the first relation is the arrangement result of the perpendicular bisector of the line connecting the core equipment area and the main transformer area, and the second relation is the arrangement result of the perpendicular bisector of the line connecting the core equipment area and the high-voltage switch area; taking specific coordinate calculation as an example, the binary linear equations to be solved are , where x represents the horizontal coordinate and y represents the vertical coordinate.

[0061] The second relationship is transformed to get the longitudinal coordinate equal to twenty-three point six minus zero point four times the transverse coordinate. The result is substituted into the first relationship to get fifteen times the transverse coordinate plus seven times (twenty-three point six minus zero point four times the transverse coordinate) equal to three hundred forty-three. After expansion, fifteen times the transverse coordinate plus seven times twenty-three point six minus seven times zero point four times the transverse coordinate equal to three hundred forty-three is obtained. Further calculation gets fifteen times the transverse coordinate plus one hundred sixty-five point two minus two point eight times the transverse coordinate equal to three hundred forty-three. The terms containing the transverse coordinate are combined to get twelve point two times the transverse coordinate equal to the difference of three hundred forty-three minus one hundred sixty-five point two, i.e. twelve point two times the transverse coordinate equal to one hundred seventy-seven point eight. One hundred seventy-seven point eight divided by twelve point two gets the transverse coordinate equal to about fourteen point six meters. The transverse coordinate fourteen point six meters is substituted into the formula of the longitudinal coordinate equal to twenty-three point six minus zero point four times the transverse coordinate to get the result of the longitudinal coordinate equal to twenty-three point six minus zero point four times fourteen point six, i.e. twenty-three point six minus five point eight equal to seventeen point eight meters. Thus, the center coordinates of the circumscribed circle are fourteen point six meters in the transverse direction and seventeen point eight meters in the longitudinal direction.

[0062] According to the center coordinates and the coordinates of any one of the three fixed spatial coordinate points, the Euclidean distance between the two points is calculated to obtain the radius length of the circumscribed circle, specifically including: according to the center coordinates and the coordinates of any one of the three fixed spatial coordinate points, the radius length of the circumscribed circle is calculated, the core device area coordinates are selected for calculation, and the Euclidean distance calculation method is adopted, and the calculation formula is ; in specific calculation, first, the difference between the transverse coordinates of the center and the core device area is calculated, i.e. fourteen point six meters of the transverse coordinate of the center minus ten meters of the transverse coordinate of the core device area to get four point six meters; then, the difference between the longitudinal coordinates of the center and the core device area is calculated, i.e. seventeen point eight meters of the longitudinal coordinate of the center minus eight meters of the longitudinal coordinate of the core device area to get nine point eight meters; four point six meters of the transverse coordinate difference is multiplied by four point six meters to get twenty-one point one six square meters, and nine point eight meters of the longitudinal coordinate difference is multiplied by nine point eight meters to get ninety-six point zero four square meters; the two square results are added to get one hundred seventeen point two square meters; the square root of one hundred seventeen point two square meters is taken to get about ten point eight meters, which is the radius length of the circumscribed circle.

[0063] The preset basic probability threshold is proportionally adjusted according to the radius length to generate a dynamic probability threshold; the calibrated fire early risk index is compared with the dynamic probability threshold to generate a preliminary fire warning signal, and the specific steps include: the preset basic probability threshold is proportionally adjusted according to the radius length to generate a dynamic probability threshold, the preset basic probability threshold is 0.6, and the core logic of the adjustment is that the larger the monitoring range is, the lower the threshold is to avoid missing detection; first, the value range of the radius is determined, and the radius value range is stabilized between 8 meters and 24 meters in combination with the actual size of the key monitoring area of the transformer substation; when the radius is less than 10 meters, it belongs to small-range intensive monitoring, and the threshold remains the basic threshold of 0.6 unchanged; when the radius is greater than 20 meters, it belongs to large-range monitoring, and the threshold is reduced to 0.4 to reduce missing detection; when the radius is between 10 meters and 20 meters, the threshold is adjusted in a linear decay manner, and the coefficient calculation formula is , the dynamic threshold calculation formula is ; taking 10.8 meters as an example, the formula is calculated as , and the dynamic probability threshold is 0.584.

[0064] The calibrated fire early risk index is compared with the dynamic probability threshold to generate a preliminary fire warning signal, first, the calibrated risk index is mapped into a probability value of 0 to 1, the mapping formula is ; taking the calibrated risk index of 42.4 as an example, the formula is calculated as , that is, the mapping probability value is 0.4224; the mapping probability value is compared with the dynamic probability threshold, and three comparison levels are set, when the mapping probability value is greater than the dynamic probability threshold, a high-risk warning is generated; when the mapping probability value is between 0.8 times the dynamic probability threshold and the dynamic probability threshold, a medium-risk warning is generated; when the mapping probability value is less than 0.8 times the dynamic probability threshold, a normal signal is output; taking the dynamic probability threshold of 0.584 as an example, 0.8 times thereof is calculated as ; if the mapping probability value is 0.4224, it is less than 0.447, and a normal signal is output; if the mapping probability value is 0.6, it is greater than 0.584, and a high-risk preliminary warning signal is generated.

[0065] In this embodiment, the boundary vertex is measured by the laser range finder, and the center coordinate is calculated, so that the accuracy of obtaining the three fixed point coordinates is improved.

[0066] As shown in Figure 2 , in another preferred embodiment of the present application, the preliminary fire warning signal is verified, and after the verification is passed, a final fire early identification alarm instruction is generated and an emergency disposal plan is started, which can include:

[0067] According to the preliminary fire warning signal, the multi-source heterogeneous real-time monitoring data corresponding to the warning time is obtained; according to the multi-source heterogeneous real-time monitoring data, video flame smoke feature recognition, infrared abnormal high temperature area analysis and environmental parameter exceeding judgment are carried out, and multi-modal risk judgment result is generated, specifically including: after the preliminary fire warning signal is generated, the system automatically extracts the accurate time stamp of the warning time, takes the time stamp as the center reference, synchronously calls the multi-source heterogeneous real-time monitoring data of the previous and next 30 seconds, and ensures that the data can completely cover the risk germination stage before the warning occurs and the feature appearing stage at the warning time; the data collection is completed by relying on three types of monitoring equipment pre-planned and deployed in the substation, all equipment establishes stable communication link with the core analysis unit through industrial Ethernet, and the real-time and integrity of data transmission are ensured; the video monitoring data is collected by the high-definition network camera deployed above the main transformer area, high-voltage switch area and core equipment area in the substation, 2 cameras are deployed in each core area to form an intersecting shooting angle, and the monitoring blind area is eliminated; the resolution of the camera is set to 1920*1080, and the frame rate is adjusted to 25 frames per second, so that the subtle actions of smoke diffusion and the initial form of flame germination can be clearly captured, and the collected video data is transmitted to the core analysis unit in the form of continuous frame sequence, and each frame of data is attached with a unique time stamp and a collection equipment number; the infrared temperature data is collected by the infrared thermal imager installed on the side of each core area, the temperature measurement range of the thermal imager is set to 0-300 degrees Celsius, the temperature measurement accuracy is controlled to be plus or minus 0.5 degrees Celsius, and the frame rate is set to 10 frames per second, which can ensure the timeliness of the temperature data and avoid data redundancy; the collected infrared data is transmitted in the form of gray scale image containing pixel temperature information, and each pixel gray value of the image has a fixed corresponding relationship with the actual temperature value, and each frame of image is associated with collection time and collection position information.

[0068] The environmental parameter data is collected by three types of special sensors distributed around the three core areas, wherein the smoke sensor is deployed at the ventilation port and the top of the equipment in each area, the characteristic gas sensor is installed close to the sealed connection of the equipment, and the temperature sensor is directly attached to the key parts of the equipment shell through a heat-conducting patch; the collection frequency of the smoke sensor is set to 0.5 seconds, the collection frequency of the characteristic gas sensor is one per second, and the collection frequency of the temperature sensor is two per second; the original data collected by all sensors is subjected to analog-digital conversion processing, and after being converted into standard numerical form, the original data is transmitted to the core analysis unit with collection time stamp and sensor number; after receiving the three types of monitoring data, the core analysis unit processes and analyzes each type of data respectively, generates the risk judgment result of single type of data, cross verifies the three types of results, generates the alarm instruction and starts the corresponding emergency disposal plan after confirming the existence of fire risk.

[0069] The video flame smoke feature recognition is realized by using frame difference method combined with gray scale feature analysis. First, the collected video frame sequence is preprocessed, and each frame of color image is converted into a gray scale image to reduce color interference. Then, the difference between the adjacent two frames of gray scale images is calculated to obtain a frame difference image. The frame difference image is binarized, and the gray scale difference threshold is set to 30. The pixel points with a gray scale difference greater than 30 are marked as white, and the rest are marked as black. The binarized frame difference image is subjected to dilation operation and erosion operation. The dilation operation uses a 3x3 structure element to scan the image, and the black pixel points around the white pixel points are converted into white. The erosion operation also uses a 3x3 structure element to scan the image, and the white pixel points on the edge of the white pixel points are converted into black. Through these two operations, the noise points in the image are eliminated.

[0070] The number of white pixel points in the frame difference image after statistical processing is counted, and the proportion of white pixel points in the total number of pixel points in the whole frame image is calculated. The calculation method is to divide the number of white pixel points by the total number of pixel points in the whole frame image. The total number of pixel points in the whole frame image is the horizontal resolution 1920 multiplied by the vertical resolution 1080, i.e. 2073600 pixel points. When the white pixel point proportion of three consecutive frames of images is greater than 0.05, the gray scale features in the image are further extracted, and the gray scale mean and variance of the suspicious area are calculated. The gray scale mean is the sum of the gray scale values of all pixel points in the suspicious area divided by the number of pixel points. The variance is the square of the difference between each pixel point gray scale value and the mean value added to the number of pixel points. When the gray scale mean is less than 80 and the variance is greater than 100, it is determined that the video flame smoke feature recognition result is at risk. The infrared abnormally high temperature area analysis first converts the infrared gray scale image into temperature. According to the calibration parameters of the thermal imager, the gray scale value of each pixel point in the image is converted into the actual temperature value. The conversion method is that the actual temperature is equal to the gray scale value multiplied by the temperature coefficient plus the reference temperature. The temperature coefficient and the reference temperature are determined by the factory calibration of the thermal imager. After the conversion is completed, all pixel points in the image are traversed, and the pixel points with an actual temperature not lower than 80 degrees Celsius are marked. The marked high temperature pixel points are subjected to connectivity analysis, and adjacent high temperature pixel points are classified into a high temperature area. The adjacent determination standard is that there are other high temperature pixel points in the upper, lower, left and right four directions of a pixel point.

[0071] The actual area of ​​each high-temperature region is calculated by multiplying the number of pixels within the high-temperature region by the actual area corresponding to a single pixel. The actual area corresponding to a single pixel is calculated using the thermal imager's field of view and shooting distance. The field of view is the thermal imager's horizontal field of view divided by its horizontal resolution to obtain the horizontal viewing angle of each pixel. The vertical direction is calculated similarly. The actual horizontal and vertical lengths corresponding to a single pixel are then calculated using the shooting distance, and multiplied together to obtain the actual area of ​​a single pixel. When a high-temperature region appears at the same location in three consecutive frames, and the actual area of ​​that region is not less than 0.1 square meters, the infrared abnormal high-temperature region analysis result is deemed to pose a risk. Environmental parameters exceeding the standard are judged... The smoke concentration and characteristic gas concentration data were analyzed separately. For smoke concentration assessment, smoke concentration data was continuously read over a 5-second period, totaling 10 data points. Each data point was compared to a threshold of 0.3 mg / m³. If all 10 data points were not lower than 0.3 mg / m³, the smoke concentration was considered excessive. Similarly, for characteristic gas concentration assessment, data was continuously read over a 5-second period, totaling 5 data points. Each data point was compared to a threshold of 100 ppm. If all 5 data points were not lower than 100 ppm, the characteristic gas concentration was considered excessive. If either smoke concentration or characteristic gas concentration exceeded the standard, the environmental parameter was deemed to be at risk.

[0072] Based on the multimodal risk assessment results, when the assessment results consistently confirm the existence of a fire risk, a final early fire identification alarm command is generated and the emergency response plan is activated. Specifically, this includes: when the multimodal risk assessment results are integrated, the number of results marked as having risk in the three categories of video flame and smoke feature identification, infrared abnormal high temperature area analysis, and environmental parameter exceeding the standard judgment is counted. When at least two of the three categories of results indicate the existence of risk, the assessment results consistently confirm the existence of a fire risk. Based on the risk level of the preliminary fire warning signal, a final early fire identification alarm command is generated and the corresponding emergency response plan is activated.

[0073] If the preliminary early warning is high risk, a first-level alarm instruction is generated: immediately trigger the sound and light alarm device at the entrance of the main control room and each core equipment area in the substation, set the decibel of the alarm sound to not less than 90 decibels, and continuously prompt the red strobe mode of the alarm light; automatically send a disconnecting control signal to the high-voltage power supply control cabinet corresponding to the early warning area to cut off the high-voltage power supply loop of the area to prevent the production of arc to assist combustion fire; start the carbon dioxide extinguishing device in the early warning area, and the extinguishing device automatically adjusts the direction of the spray head according to the early warning position information, opens the spray head to release the carbon dioxide extinguishing agent; push the alarm information to the mobile terminal of all on-duty operation and maintenance personnel through the operation and maintenance management platform, and the information content includes the early warning area name risk level, and is accompanied by an access link of the real-time video stream on site; if the preliminary early warning is medium risk, a second-level alarm instruction is generated: trigger the main control room and the sound and light alarm device, set the decibel of the alarm sound to not less than 80 decibels, and prompt the alarm light in yellow strobe mode; push the alarm information to the mobile terminal of the operation and maintenance personnel through the operation and maintenance management platform, and the information is accompanied by the video frame screenshot, infrared temperature curve, environmental parameter change trend and other data of the early warning area; arrange the operation and maintenance personnel to carry the portable smoke detector and infrared thermometer to the scene for manual verification, manually start the extinguishing device after confirming the risk, and manually release the alarm instruction through the operation and maintenance management platform after confirming that there is no risk.

[0074] In this embodiment, the multi-level comparison of the calibrated risk index and the dynamic threshold value makes the generation of the early warning signal more targeted, improves the accuracy of early fire identification, and provides effective protection for the safe operation of the substation.

[0075] Embodiments of the present application also provide a computing device, comprising a processor and a memory storing a computer program, wherein the computer program is executed by the processor to implement the system as described above. All implementation manners in the above system embodiments are applicable to this embodiment, and the same technical effects can also be achieved.

[0076] Embodiments of the present application also provide a computer readable storage medium storing instructions, which, when executed on a computer, cause the computer to execute the system as described above. All implementation manners in the above system embodiments are applicable to this embodiment, and the same technical effects can also be achieved.

[0077] The above is the preferred embodiment of the present application. It should be noted that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A substation fire early detection and alarm system embedding AI algorithms, characterized in that, include: The acquisition module is used to collect real-time monitoring data from multiple heterogeneous sources within the substation. The fusion module is used to preprocess multi-source heterogeneous real-time monitoring data, extract multi-dimensional dynamic features, and synthesize the multi-dimensional dynamic features into a high-dimensional feature vector. The calculation module is used to input high-dimensional feature vectors into a pre-trained embedded AI analysis model, analyze the high-dimensional feature vectors, and generate an early fire risk index. The calibration module selects three fixed spatial coordinate points located at the geometric center points of the core equipment area, the main transformer area, and the high-voltage switch area, respectively. It connects these three fixed spatial coordinate points to form a triangular convex polygon region covering the key monitoring objects. The module performs spatial partitioning on the triangular convex polygon region, generating a set of analysis grids. It calculates the coordinates of the center point of each analysis grid and determines the grid's inclusion status within the triangular convex polygon region based on the positional relationship between the center point coordinates and the sides of the triangle. Based on the Euclidean distance between the center point coordinates of the analysis grids within the region and the three fixed spatial coordinate points, it calculates the local spatial weight of each analysis grid. Global spatial calibration weights are calculated and generated based on local spatial weights. The early fire risk index is weighted and calibrated using global spatial calibration weights to generate a calibrated early fire risk index. The generation module is used to compare the calibrated early fire risk index with the preset dynamic probability threshold to generate a preliminary fire warning signal. The verification module is used to verify the preliminary fire warning signal. After the verification is successful, it generates the final early fire identification alarm command and activates the emergency response plan.

2. The substation fire early identification and alarm system with embedded AI algorithm according to claim 1, characterized in that, Multi-source heterogeneous real-time monitoring data is preprocessed to extract multi-dimensional dynamic features, which are then synthesized into a high-dimensional feature vector, including: The acquired video image data and infrared thermal imaging data are processed to obtain standardized image data, and the standardized image data is analyzed to generate a set of geometric moments. Based on the geometric moment set, the shape moment features are calculated and combined; and the collected environmental sensor data are processed to extract numerical features. Shape moment features and numerical features are fused to generate multi-dimensional dynamic features, and these multi-dimensional dynamic features are then synthesized into a high-dimensional feature vector.

3. The substation fire early identification and alarm system with embedded AI algorithm according to claim 2, characterized in that, High-dimensional feature vectors are input into a pre-trained embedded AI analysis model to analyze the feature vectors and generate an early fire risk index, including: The high-dimensional feature vector is received by the input layer of the embedded AI analysis model, and feature tensor reconstruction and normalization scaling operations are performed to obtain a normalized feature tensor with uniform dimensions. The standardized feature tensor is input into the hidden layer network, and layer-by-layer nonlinear transformation and high-order feature synthesis are performed in the hidden layer network to generate a high-level comprehensive feature tensor. The high-level comprehensive feature tensor is passed to the Sigmoid function unit of the output layer, and linear weighted summation and probability transformation are performed to generate scalar probability values. Nonlinear amplification and range mapping are performed on the scalar probability values ​​to generate the early fire risk index.

4. The substation fire early identification and alarm system with embedded AI algorithm according to claim 3, characterized in that, Perform spatial partitioning on the triangular convex polygon region to generate a set of analysis grids; Calculate the center point coordinates of each analysis grid, and determine the inclusion status of the analysis grid within the convex polygon region of the triangle based on the positional relationship between the center point coordinates and the sides of the triangle, including: A uniform grid covering the outer rectangle is generated in the planar space of the triangular convex polygon region to obtain a set of analysis grids; Calculate the coordinates of the geometric center point of each analysis grid based on the vertex coordinates of the analysis grid. Based on the positional relationship between the coordinates of the geometric center point and the three sides of the triangular convex polygon region, three position determination values ​​are calculated; based on the consistency of the signs of the three position determination values, the inclusion status of the raster within the triangular convex polygon region is determined.

5. The substation fire early identification and alarm system with embedded AI algorithm according to claim 4, characterized in that, The calibrated early fire risk index is compared with a preset dynamic probability threshold to generate a preliminary fire warning signal, including: Determine the center coordinates and radius of the circumcircle passing through the three fixed spatial coordinate points using their planar coordinates. The preset base probability threshold is adjusted proportionally based on the radius length to generate a dynamic probability threshold. The calibrated early fire risk index is compared with the dynamic probability threshold to generate a preliminary fire warning signal.

6. The substation fire early identification and alarm system with embedded AI algorithm according to claim 5, characterized in that, Given the planar coordinates of three fixed spatial points, determine the coordinates of the center and the radius of the circumcircle passing through these three points, including: Obtain the two-dimensional plane coordinates of three fixed spatial coordinate points to generate three sets of coordinate data; based on the coordinates of any two points in the three sets of coordinate data, calculate the equation parameters of the perpendicular bisectors of the two sides respectively; Based on the parameters of the equations of the perpendicular bisectors of the two sides, solve the equations simultaneously to find the coordinates of the intersection point of the two perpendicular bisectors, and then obtain the coordinates of the center of the circumcircle. Calculate the Euclidean distance between two points based on the coordinates of the center of the circle and the coordinates of any one of the three fixed spatial coordinate points, and obtain the radius of the circumcircle.

7. The substation fire early identification and alarm system with embedded AI algorithm according to claim 6, characterized in that, The initial fire warning signal is verified. Upon successful verification, a final early fire identification alarm command is generated, and the emergency response plan is activated, including: Based on the initial fire warning signal, acquire multi-source heterogeneous real-time monitoring data corresponding to the warning time; Based on multi-source heterogeneous real-time monitoring data, video flame and smoke feature recognition, infrared abnormal high temperature area analysis, and environmental parameter exceedance judgment are performed to generate multimodal risk assessment results. Based on the multimodal risk assessment results, when the assessment results consistently confirm the existence of a fire risk, a final early fire identification alarm command is generated and the emergency response plan is activated.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 7.

Images