Substation fire early-stage identification alarm system with embedded AI algorithm

Abstract

The invention provides a transformer substation fire early identification alarm system embedded with an AI algorithm, and relates to the technical field of data processing, and the system comprises the steps: calculating and analyzing the position relation between grid points and each side of a triangle to determine the inclusion state, calculating the local space weight according to the inclusion state and the distance relation between the analysis grid and three fixed space coordinate points, and calculating the local space weight according to the local space weight; aggregating the local space weights of all the analysis grids to obtain a global space calibration weight, and performing weighted operation on the fire early-stage risk index by using the global space calibration weight to generate a calibrated fire early-stage risk index; the comparison module is used for comparing the calibrated fire early-stage risk index with a preset dynamic probability threshold value and generating a preliminary fire early-warning signal; and the verification module is used for verifying the initial fire early warning signal, generating a final fire early recognition alarm instruction after verification is passed, and starting an emergency disposal plan. According to the invention, early-stage accurate identification and rapid alarm of the transformer substation fire can be realized, the false alarm rate is effectively reduced, and emergency disposal is started in time.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an early fire identification and alarm system for substations that incorporates AI algorithms. Background Technology

[0002] In the daily monitoring of substations, common fire alarms (such as sensors that respond to changes in infrared radiation or ion concentration) are usually deployed in specific locations to monitor single physical parameters such as temperature and smoke. However, in complex scenarios such as oil vapor evaporation near the main transformer or instantaneous arc radiation generated by high-voltage switch operation, these alarms that rely on the detection of a single phenomenon may not be able to adequately distinguish between normal operating condition disturbances and real early fire characteristics. Their warning results may sometimes not match the actual risk importance of the equipment area. For example, warnings for the main transformer area and warnings for ordinary cable trenches are often treated the same in traditional systems, making it difficult for maintenance personnel to quickly determine the priority of handling and affecting the efficiency of early response. Summary of the Invention

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

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] Firstly, a substation fire early detection and alarm system embedded with AI algorithms includes:

[0006] The acquisition module is used to collect real-time monitoring data from multiple heterogeneous sources within the substation.

[0007] 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.

[0008] 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.

[0009] The calibration module is used to select three fixed spatial coordinate points on the two-dimensional plan view of the substation monitoring area, located in the core equipment area, the main transformer area, and the high-voltage switch area, respectively. 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 positional relationship between the analysis grid points and each side of the triangle is calculated to determine the inclusion state. Based on the inclusion state and the distance relationship between the analysis grid and the three fixed spatial coordinate points, the local spatial weight is calculated. The local spatial weights of all analysis grids are aggregated to obtain the global spatial calibration weight. The global spatial calibration weight is used to perform a weighted calculation on the early fire risk index to generate the calibrated early fire risk index.

[0010] 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.

[0011] 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.

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

[0013] One or more processors;

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

[0015] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0016] The above-described solution of the present invention has at least the following beneficial effects:

[0017] By defining the spatial coordinates and calibrating the rasterized weights of key areas such as the core equipment area and the main transformer area, the sensitivity of fire risk identification in key areas is enhanced, avoiding misjudgment of risks due to differences in spatial layout, and making the risk index more consistent with the actual equipment distribution and safety priorities of the substation. A lightweight embedded AI analysis model is adopted to meet the requirements of the substation for equipment response speed, data privacy protection and edge computing deployment, ensuring that alarm commands are generated quickly. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an early fire identification and alarm system for substations embedded with AI algorithms, provided by an embodiment of the present invention.

[0019] Figure 2This is a flowchart illustrating the process of verifying a preliminary fire warning signal, generating a final early fire identification alarm command, and activating an emergency response plan after successful verification, provided by an embodiment of the present invention. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0021] like Figure 1 As shown, an embodiment of the present invention proposes an early fire identification and alarm system for substations embedded with an AI algorithm, comprising:

[0022] The acquisition module is used to collect real-time monitoring data from multiple heterogeneous sources within the substation.

[0023] 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.

[0024] 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.

[0025] The calibration module is used to select three fixed spatial coordinate points on the two-dimensional plan view of the substation monitoring area, located in the core equipment area, the main transformer area, and the high-voltage switch area, respectively. 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 positional relationship between the analysis grid points and each side of the triangle is calculated to determine the inclusion state. Based on the inclusion state and the distance relationship between the analysis grid and the three fixed spatial coordinate points, the local spatial weight is calculated. The local spatial weights of all analysis grids are aggregated to obtain the global spatial calibration weight. The global spatial calibration weight is used to perform a weighted calculation on the early fire risk index to generate the calibrated early fire risk index.

[0026] 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.

[0027] 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.

[0028] In this embodiment of the invention, by defining the spatial coordinates and calibrating the rasterized weights of key areas such as the core equipment area and the main transformer area, the sensitivity of fire risk identification in key areas is enhanced, avoiding misjudgment of risks due to differences in spatial layout, and making the risk index more consistent with the actual equipment distribution and safety priority of the substation; a lightweight embedded AI analysis model is adopted to meet the requirements of the substation for equipment response speed, data privacy protection and edge computing deployment, ensuring that alarm commands are generated quickly.

[0029] In a preferred embodiment of the present invention, the collection of multi-source heterogeneous real-time monitoring data within the substation specifically includes: the collection equipment comprising a high-definition camera, an infrared thermal imager, and temperature, smoke, humidity, and gas sensors, deployed to be located at risk points in each area: the high-definition camera is mounted above the core equipment control cabinet, on the side of the main transformer tank, etc., with the lens aimed at the equipment body and key connection parts, collecting video image data with a resolution of 1920 x 1008 pixels at a frame rate of 30 frames per second, ensuring the capture of early details such as insulation layer discoloration and faint smoke; the infrared thermal imager is deployed at a high point in the area, covering heat-prone points such as the main transformer winding joints and high-voltage switch contacts, collecting data at a frame rate of 10 frames per second. Infrared thermal imaging data with a temperature range of -20°C to 300°C and a temperature resolution of 0.1°C accurately captures minute temperature rises in equipment. Various environmental sensors are installed at equipment vents, oil tanks, insulating sleeves, etc., collecting data once per second. Among them, the temperature sensor has a range of -40°C to 125°C, the smoke sensor has a range of 0 mg / m³ 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 ppm to 1000 ppm. After collection, all data are synchronously encapsulated with millisecond-level timestamps, and data type, equipment number, and other identifiers are added to form a unified data stream, which is then transmitted to the data processing unit.

[0030] This embodiment solves the blind spot problem of single parameter monitoring by acquiring data from multiple sources, and captures early-stage image changes and environmental parameter anomalies in fires.

[0031] In a preferred embodiment of the present invention, preprocessing the multi-source heterogeneous real-time monitoring data to extract multi-dimensional dynamic features and synthesizing the multi-dimensional dynamic features into a high-dimensional feature vector may include:

[0032] The acquired video image data and infrared thermal imaging data are processed to obtain standardized image data. This standardized image data is then analyzed to generate a set of geometric moments. Specifically, image data processing involves converting color video images to grayscale using a weighted average method. The calculation method is to multiply the red channel value by 0.3, add the green channel value multiplied by 0.59, and add the blue channel value multiplied by 0.11, summing these three values ​​to obtain a single-channel grayscale value, thus highlighting the grayscale difference between the device and the background. Infrared thermal imaging data is mapped to grayscale values ​​using a linear ratio. The calculation method is to subtract -20 degrees Celsius from the measured temperature, divide by the difference between 300 degrees Celsius and -20 degrees Celsius, and then multiply by 255 to obtain the corresponding grayscale value. Grayscale values ​​directly reflect the temperature rise of the equipment. A 5x5 Gaussian filter is used to remove image noise. The specific calculation process is as follows: First, a 5x5 filter kernel is constructed, with the center of the kernel set as the origin. The remaining twenty-four positions are labeled as -2, -1, zero, +1, and +2 horizontally and vertically, respectively. That is, the coordinates of each position are represented by (horizontal coordinate, vertical coordinate), such as the first position to the right of the center being (1, zero), and the upper left corner being (-2, -2), etc. The weight of each position is calculated using a Gaussian function. The Gaussian function is calculated by first summing the squares of the horizontal and vertical coordinates of the position, then dividing this sum by two and multiplying by the square of the standard deviation (set to 0.8, i.e., two multiplied by zero). The square of 8 equals 1.28. Taking the negative result, we use it as the exponent of the natural constant. Then, we multiply this exponent by 1 and divide by (2 multiplied by the square of pi multiplied by the standard deviation) to get the initial weight for that position. For example, the weight calculation for the center position (zero, zero): the square of the horizontal coordinate plus the square of the vertical coordinate equals zero. Dividing this by 1.28 gives zero. The zeroth power of the natural constant equals 1. Multiplying this by 1 and dividing by (2 multiplied by 3.1416 multiplied by the square of 0.8) gives an initial weight of approximately 0.1201. After calculating the initial weights for all twenty-five positions in this way, we sum all the initial weights to get the total weight. Then, we divide the initial weight of each position by the total weight to complete weight normalization (ensuring the sum of all weights is equal). To avoid changes in the overall brightness of the image after filtering, the normalized weights for each position are as follows: the center position (zero, zero) has the largest weight, approximately 0.1201; the eight adjacent positions around the center (e.g., (one, zero), (zero, one), (negative one, zero) etc.) have the same weight, approximately 0.1083; the eight positions separated from the center by one position (e.g., (two, zero), (one, one), (zero, two) etc.) have a weight of approximately 0.0540; the four diagonal positions ((one, one), (one, negative one), (negative one, one), (negative one, negative one)) have a weight of approximately 0.0812; the four outermost diagonal positions ((two, two), (two, negative two), (negative two, two), (negative two, negative two)) have the smallest weight, approximately 0.0304.During filtering, each pixel in the image is taken as the center, and 25 pixels within a 5x5 range around it are selected. The gray value of each pixel is multiplied by the normalization weight of the corresponding position of the filtering kernel. The 25 product results are added together to obtain the gray value of the center pixel after filtering. This process is repeated for all pixels to complete the noise reduction.

[0033] After denoising, the image is scaled to 64 x 64 pixels using bilinear interpolation. The specific calculation process is as follows: To determine the scaling ratio, the original image size is 1,920 x 1,008 pixels, the target size is 64 x 64 pixels, the horizontal scaling ratio is 64 divided by 1,920 equals 1 / 125, and the vertical scaling ratio is 64 divided by 1,008 approximately equals 0.0634. For each pixel in the target image, first calculate its corresponding position in the original image. This is done by dividing the target pixel's horizontal coordinate by the horizontal scaling factor to obtain the original image's horizontal floating-point coordinate; and dividing the target pixel's vertical coordinate by the vertical scaling factor to obtain the original image's vertical floating-point coordinate. For example, for the (10, 20) pixel in the target image, the original image's horizontal coordinate is 10 divided by 1 / 125, which equals 1200, and its vertical coordinate is 20 divided by 0.0634, approximately 317.95. Then, take the four nearest integer coordinate pixels around this floating-point coordinate in the original image. That is, the horizontal coordinate is taken as an integer less than and close to the floating-point coordinate, and the vertical coordinate is taken similarly. This yields the coordinates and corresponding grayscale values ​​of four pixels, denoted as top-left (x1, y1, grayscale value 1), top-right (x2, y2, grayscale value 2), bottom-left (x3, y3, grayscale value 3), and bottom-right (x4, y4, grayscale value 4), where x2 equals x1 plus one, and y3 equals y1 plus one. Then... Calculate the horizontal interpolation weights by subtracting x1 from the horizontal floating-point coordinates of the original image to obtain the decimal part 'a' (the value of 'a' ranges from zero to one). The horizontal weights are (1 - a) and 'a', both ranging from zero to one, and their sum is one. Calculate the vertical interpolation weights by subtracting y1 from the vertical floating-point coordinates of the original image to obtain the decimal part 'b' (the value of 'b' ranges from zero to one). The vertical weights are (1 - b) and 'b', both ranging from zero to one, and their sum is one. First, perform horizontal interpolation, calculating the sum of the values ​​at the top left corner. The grayscale value of the top right corner at the floating-point horizontal coordinate is calculated as follows: the grayscale value of the top left corner is grayscale value 1 multiplied by (- - a) plus grayscale value 2 multiplied by a; the grayscale value of the top right corner is grayscale value 3 multiplied by (- - a) plus grayscale value 4 multiplied by a. Then, vertical interpolation is performed by multiplying the grayscale value of the top left corner by (- - b) and adding the grayscale value of the top right corner by b to obtain the grayscale value of the target pixel. The grayscale values ​​of all target pixels are calculated sequentially to obtain a standardized image of 64 by 64 pixels.

[0034] For a standardized image, calculate the zeroth to third order geometric moments. The zeroth order geometric moment is the sum of all pixel gray values. The first order geometric moment contains two values: the sum of the products of all pixel gray values ​​and their corresponding x-coordinates, and the sum of the products of all pixel gray values ​​and their corresponding y-coordinates. The second order geometric moment contains three values: the sum of the products of all pixel gray values ​​and the square of their corresponding x-coordinates, the sum of the products of all pixel gray values ​​and their corresponding x-coordinates multiplied by their y-coordinates, and the sum of the products of all pixel gray values ​​and the square of their corresponding y-coordinates. The third order geometric moment contains four values, forming a set of ten geometric moments.

[0035] Based on the geometric moment set, shape moment features are calculated and combined. The collected environmental sensor data is then processed to extract numerical features, specifically including: calculating the central moment based on the geometric moments. First, the mean x-axis coordinate is calculated as the x-related value of the first-order geometric moment divided by the zero-order geometric moment; the mean y-axis coordinate is calculated as the y-related value of the first-order geometric moment divided by the zero-order geometric moment. Then, the mean x-axis coordinate of each pixel is subtracted from the mean x-axis coordinate, and the mean y-axis coordinate is subtracted from the mean y-axis coordinate, according to the calculation method of the corresponding order of geometric moments, to obtain the central moment. Next, the normalized central moment is calculated by dividing each central moment by the power of t of the zero-order geometric moment, where t is half of the corresponding moment order rounded up. Features with a correlation greater than 0.7 with early fire characteristics are selected and combined into twelve-dimensional shape moment features.

[0036] After image feature extraction, the environmental sensor data processing involves first removing outliers according to the three-standard-deviation criterion. The most recent 100 sensor data points are collected, and the mean is calculated by summing all data points and dividing by 100. The standard deviation is calculated by summing the squares of the differences between each data point and the mean, then dividing by 100 to obtain the variance. The square root of the variance is then used to obtain the standard deviation. Data points exceeding the range of mean minus three standard deviations to mean plus three standard deviations are considered outliers. These outliers are replaced with the mean of the previous five data points, calculated by summing these five data points and dividing by five. Next, the data is standardized to a range of zero to one. This is calculated by subtracting the minimum value of the sensor's measurement range from each data point, then dividing the difference by the difference between the maximum and minimum values ​​within the measurement range. For example, for a temperature sensor, the difference is the difference between the collected temperature and -40 degrees Celsius, then divided by 125 degrees Celsius minus -40 degrees Celsius. The standardized temperature, smoke concentration, humidity, and characteristic gas concentration are then extracted as four-dimensional numerical features.

[0037] The shape moment features and numerical features are fused to generate multi-dimensional dynamic features, which are then synthesized into a high-dimensional feature vector. Specifically, this involves: aligning feature timestamps to ensure that the fused features originate from the same monitoring time; the shape moment features are 12-dimensional features calculated based on standardized images, with each image frame corresponding to a set of 12-dimensional data, carrying a millisecond-level timestamp of image acquisition; the numerical features are 4-dimensional features (temperature, smoke concentration, humidity, and characteristic gas concentration) obtained after preprocessing from environmental sensors, with a set of 4-dimensional data collected every second, also carrying a millisecond-level timestamp; the data processing unit uses timestamp matching to associate the 12-dimensional shape moment features and 4-dimensional numerical features at the same time into a set of data to be fused. If there is a timestamp discrepancy between the image frame and the sensor data (maximum not exceeding 50 milliseconds), linear interpolation is used to complete the sensor data to ensure the temporal consistency of each set of data to be fused.

[0038] The feature standardization secondary calibration is performed to eliminate the dimensional differences between the two types of features. Although the shape moment feature has been calculated with normalized central moments, its numerical range may still deviate from the sensor's standardized features, and it needs to be uniformly calibrated to the 0 to 1 range. For the 12-dimensional shape moment feature, the data processing unit calls up the shape moment feature data of 300 sets of historical normal operating conditions, calculates the minimum and maximum values ​​of each dimension, and uses the min-max normalization method for calibration. The calculation method is to subtract the historical minimum value of a certain dimension of the current shape moment feature from the current dimension value, and divide the difference by the difference between the historical maximum value and the historical minimum value of that dimension. For 4-dimensional numerical features, the results already standardized to the 0-1 range are directly used. Standardization calculations are only repeated for the interpolated data to ensure both types of features are on the same numerical scale. Weighted fusion is implemented to generate multi-dimensional dynamic features, highlighting the weights of features highly correlated with fire. Based on substation fire early warning experience, the three dimensions reflecting irregular edges in the shape moment features (corresponding to the second-order cross moment, third-order x-moment, and third-order y-moment in the normalized central moment) and the two dimensions reflecting fire precursors in the numerical features (smoke concentration and characteristic gas concentration) are more critical for fire identification and their weights need to be increased. Specifically, the weight allocation is 0.08 for each of the three key dimensions in the shape moment features and 0.06 for each of the remaining nine dimensions. In the numerical features, smoke concentration, ... Each characteristic gas concentration is assigned a weight of 0.1, and temperature and humidity are each assigned a weight of 0.07, with the total weight of all dimensions being 1. During the fusion calculation, each calibrated characteristic dimension value is multiplied by its corresponding weight. The 12 shape moment features are weighted and combined sequentially with the 4 numerical features to form a 16-dimensional multi-dimensional dynamic feature. The multi-dimensional dynamic features are then synthesized into a high-dimensional feature vector. That is, the data processing unit adopts a feature tiling method, arranging the 16-dimensional multi-dimensional dynamic features in a fixed order with shape moment features first and numerical features last. The shape moment features are arranged in the order of zero-order to third-order normalized central moments, and the numerical features are arranged in the order of temperature, smoke concentration, humidity, and characteristic gas concentration, ultimately forming a 1-row, 16-column one-dimensional high-dimensional feature vector.

[0039] In this embodiment, the targeted preprocessing eliminates on-site interference, and the extracted multidimensional features can accurately reflect the early comprehensive signals of a fire.

[0040] In a preferred embodiment of the present invention, inputting a high-dimensional feature vector into a pre-trained embedded AI analysis model to analyze the high-dimensional feature vector and generate an early fire risk index may include:

[0041] The high-dimensional feature vector is a row of sixteen columns of data obtained by fusing shape moments and numerical features in the early stage, corresponding to sixteen fire-related feature dimensions. The input layer of the embedded AI analysis model is adapted to this dimension, with sixteen neurons. Each neuron receives one dimension of the feature vector data separately to ensure that no feature is missed. Considering that the subsequent hidden layer uses convolutional computing logic that requires three-dimensional input, the one-dimensional vector of a row of sixteen columns is converted into a three-dimensional tensor of a row of sixteen columns and a row of sixteen columns during feature tensor reconstruction. The first dimension represents a single sample from a single monitoring session, and the second dimension corresponds to the sixteen feature dimensions. The third dimension is the default number of channels for image-type features. This structure is perfectly suited to the computational needs of hidden layers. The core of normalization scaling is to eliminate dimensional differences. First, the mean of the sixteen elements in the tensor is calculated. The mean is obtained by adding all sixteen elements together and dividing by sixteen. The standard deviation is calculated by first calculating the difference between each element and the mean, squaring each difference, adding all the squares together and dividing by sixteen to get the variance, and then taking the square root of the variance to get the standard deviation. Finally, the mean is subtracted from each element, and the difference is divided by the standard deviation to obtain a normalized feature tensor with elements distributed in the interval from -1 to 1.

[0042] The standardized feature tensor is input into the hidden layer network, where layer-by-layer nonlinear transformations and high-order feature synthesis are performed to generate a high-level comprehensive feature tensor. Specifically, considering the complexity of early-stage features of substation fires (such as the shape of faint smoke and local temperature rise) and the limited computing power of embedded devices, the hidden layer adopts a third-order dimensionality-reduced fully connected layer structure to gradually extract key features. The first fully connected layer has 128 neurons. After receiving the standardized feature tensor, it first performs linear weighted calculation, that is, each feature element is multiplied by the weight coefficient of the corresponding neuron (the weights are initialized using Xavier to ensure consistent input and output variance). All weighted results are then combined. After adding the bias term for this layer (initial value set to 0.01), perform ReLU activation function transformation after linear calculation. When the input value is greater than zero, output the value directly; when it is less than or equal to zero, output zero. This transformation highlights the nonlinear correlation of key features such as temperature anomalies and smoke contours. Perform batch normalization to stabilize the feature distribution. First, calculate the mean (sum of all feature values ​​and divide by 128) and variance (sum of the squares of the differences between each feature value and the mean and divide by 128) of the 128 output features. Then, subtract the mean from each feature value and divide by the square root of the variance. Finally, multiply by the scaling factor of 1 and add the offset of zero to stabilize the feature value in the range of -1 to 1.

[0043] The 128 features output from the first layer are fed into the second fully connected layer, which has 64 neurons. This layer repeats the linear weighting, ReLU activation, and batch normalization operations. Specifically, during linear weighting, each of the 128 input features is multiplied by its corresponding weight (also initialized by Xavier), summed, and then a bias is added. The activation and normalization processes are identical to the first layer, outputting 64-dimensional features, achieving feature dimensionality reduction and information condensation. The 64-dimensional features output from the second layer are fed into the third fully connected layer, which has 32 neurons. During linear weighting, the 64 input features are multiplied by their weights, summed, and then a bias is added. Considering the possibility of weak negative features in the early stages of a fire (such as abnormally low local humidity), the activation function is LeakyReLU. When the input value is greater than zero, the output value is the same; when it is less than or equal to zero, the output value is multiplied by 0.01, avoiding complete suppression of negative features by ReLU. Batch normalization is then performed, finally outputting a 32-dimensional high-level comprehensive feature tensor.

[0044] The model consists of an input layer, three fully connected hidden layers, and an output layer. The output layer includes a Sigmoid function unit to output the risk probability. Training data is derived from three years of historical monitoring data from a substation, totaling 5,000 sets. One thousand sets represent simulated fire scenarios (e.g., igniting oil-soaked cotton yarn to simulate cable fires, heating elements to simulate equipment overheating), and four thousand sets represent normal operation scenarios (covering different seasons, equipment loads, and weather conditions). All data is processed into high-dimensional feature vectors and labeled: fire scenarios are labeled 1, and normal scenarios are labeled 0. Data is partitioned into training, validation, and test sets in a 7:2:1 ratio: 3,500 training sets are used to update model parameters, one thousand validation sets are used to adjust hyperparameters, and five hundred test sets are used for final performance evaluation. Training employs a stochastic gradient descent optimizer with a batch size of 32 (adapting to the computing power of embedded devices), and an initial learning rate of [missing information]. The learning rate is set to 0.01, meaning that due to the need for parameter fine-tuning in the later stages of training, the learning rate is multiplied by 0.1 every fifty training epochs to reduce parameter fluctuations. The cross-entropy loss function is used. During calculation, for each training data set: if the label is one, calculate one multiplied by the natural logarithm of the model's predicted probability; if the label is zero, calculate the difference between one and one multiplied by one minus the natural logarithm of the predicted probability. The two results are added together and the negative number is taken. This result is added together for all training data and then divided by the total number of training sets to obtain the average loss value. During training, the average loss value is calculated using the validation set after each epoch. Training stops when the validation set loss value no longer decreases (the decrease is less than 0.0001) for ten consecutive epochs, at which point the model reaches its optimal generalization ability. Overfitting (low loss on the training set but high loss on the validation set) or underfitting (high loss on both sets) is avoided. The weight coefficients and bias terms of each layer are saved at this point to form a pre-trained model adapted to the substation scenario.

[0045] The high-rise comprehensive feature tensor is passed to the Sigmoid function unit of the output layer, where linear weighted summation and probability transformation are performed to generate scalar probability values. Nonlinear amplification and range mapping are then applied to these scalar probability values ​​to generate an early fire risk index. Specifically, the high-rise comprehensive feature tensor is 32 dimensions. The Sigmoid function unit of the output layer first performs a linear weighted sum on it, multiplying each of the 32 feature elements by the weight coefficients of the output layer (fixed after training). All weighted results are summed, and then the output layer bias term (fixed after training) is added to obtain a linear output value. This linear output value is then input into the Sigmoid function unit. The id function performs a probability transformation by dividing one by the negative linear output value of one plus the natural constant (approximately 2.7128), resulting in a scalar probability value between zero and one. The closer this value is to one, the higher the fire risk in the current scenario. Considering that substation maintenance personnel need to intuitively distinguish between normal operating conditions, minor anomalies, and early fire risks, a targeted nonlinear amplification is performed on the scalar probability value. The core logic of using a one-point quadratic operation is to avoid excessively distorting the risk gradient in the high-probability range while also widening the difference in the low-probability range (early fires in substations often manifest as low-probability minor anomalies, which require accurate identification).

[0046] The specific calculation process involves using the scalar probability value as the base and 1.2 as the exponent to perform a power operation. The operation first calculates the product of the first power of the scalar probability value (i.e., itself) and the zero-squared power. The zero-squared power is calculated using the natural logarithm: first, take the natural logarithm of the scalar probability value, multiply the result by 0.2, then calculate the exponent using the natural constant as the base to obtain the zero-squared result. Finally, multiply this result by the first power of the scalar probability value to obtain the 1.2 result. For example, when the scalar probability value is 0.5, the natural logarithm is approximately -0.6693, multiplied by 0.2 to obtain -0.1338, and the natural constant is multiplied by the exponent. 0.1338 raised to the power of 0.873, multiplied by 0.5, yields 0.4365 (the example value in the original description may differ slightly due to approximate calculation; the actual precise calculation should follow this step). When the scalar probability value is 0.6, the natural logarithm is approximately -0.5404, multiplied by 0.2, yielding -0.1081, etc. The natural constant raised to the power of -0.1081 is approximately 0.892, multiplied by 0.6, yielding 0.565. The difference between the two increases from 0.1 to 0.1335, effectively distinguishing between minor anomalies and normal states. Finally, the amplified result is multiplied by 100, mapped to the range of 0 to 100, generating the early fire risk index.

[0047] This embodiment, with its coherent data collection, extraction, and analysis process and intuitive risk index, improves the efficiency of early fire identification, clarifies early warning priorities, provides support for maintenance personnel to respond quickly, and solves the problem of low system response efficiency.

[0048] In a preferred embodiment of the present invention, three fixed spatial coordinate points are selected on a two-dimensional plan view of the substation monitoring area, respectively located in the core equipment area, the main transformer area, and the high-voltage switch area. These three fixed spatial coordinate points are connected to form a triangular convex polygon region. The triangular convex polygon region is divided to generate an analysis grid. The positional relationship between the analysis grid points and the sides of the triangle is calculated to determine the inclusion state. Based on the inclusion state and the distance relationship between the analysis grid and the three fixed spatial coordinate points, local spatial weights are calculated. The local spatial weights of all analysis grids are aggregated to obtain global spatial calibration weights. The global spatial calibration weights are used to perform a weighted calculation on the early fire risk index to generate a calibrated early fire risk index, which may include:

[0049] Three fixed spatial coordinate points are selected, located at the geometric center points of the core equipment area, the main transformer area, and the high-voltage switch area, respectively. These three fixed spatial coordinate points are connected to form a triangular convex polygon area covering the key monitoring objects. Specifically, three core monitoring points are selected on the two-dimensional plan of the substation, corresponding to the geometric center points of the core equipment area, the main transformer area, and the high-voltage switch area, respectively. These three areas are high-risk fire zones, and the area covering them is the key monitoring range. The coordinates of the center points are obtained by measuring the vertices of the area boundaries using a laser rangefinder: taking the core equipment area as an example, the horizontal and vertical coordinates of the four vertices of its rectangular boundary are measured. The horizontal coordinates of the four vertices are added together and divided by four to obtain the horizontal coordinates of the center point; the vertical coordinates of the four vertices are added together and divided by four to obtain the vertical coordinates of the center point. The same method is used for the main transformer area and the high-voltage switch area. After actual measurement, the coordinates are determined as follows: core equipment area (10 meters, 8 meters), main transformer area (25 meters, 15 meters), and high-voltage switch area (18 meters, 28 meters). The three center points are connected sequentially with straight lines to form a triangular convex polygon key monitoring area.

[0050] Within the planar space of the triangular convex polygon region, a uniform grid covering the outer rectangle is generated, resulting in a set of analysis grids. Based on the vertex coordinates of the analysis grids, the coordinates of the geometric center point of each analysis grid are calculated. Specifically, to refine the risk assessment within the region, a uniform grid covering the outer rectangle of the region is generated: the coordinates of the upper left corner of the outer rectangle are the minimum horizontal coordinate of the three center points minus one meter (i.e., ten meters minus one meter equals nine meters), and the minimum vertical coordinate minus one meter (i.e., eight meters minus one meter equals seven meters); the coordinates of the lower right corner are the maximum horizontal coordinate plus one meter (i.e., twenty-five meters plus one meter equals twenty-six meters), and the maximum vertical coordinate plus one meter... One meter equals twenty-eight meters plus one meter equals twenty-nine meters; the grid size is set to 0.5 meters by 0.5 meters to adapt to the positioning accuracy of the equipment, and each grid is an analysis grid; the geometric center point coordinates of each grid are calculated by adding the horizontal coordinate of the upper left corner of the grid to the horizontal coordinate of the lower right corner, and then dividing the sum by two to obtain the horizontal coordinate of the center point; the vertical coordinate of the grid is calculated by adding the vertical coordinate of the upper left corner of the grid to the vertical coordinate of the lower right corner, and then dividing the sum by two to obtain the vertical coordinate of the center point. For example, for the grid with the upper left corner (nine meters, seven meters) and the lower right corner (nine point five meters, seven point five meters), the center point is (nine point two five meters, seven point two five meters).

[0051] Based on the positional relationship between the geometric center point coordinates 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 grid within the triangular convex polygon region is determined. Specifically, when determining whether the grid is within the triangular region, for each side of the triangle, the relative position of the grid center point is determined by the difference between the horizontal and vertical coordinates. Taking the side formed by the core equipment area (10 meters, 8 meters) and the main transformer area (25 meters, 15 meters) as an example, first calculate the horizontal length of this side as 25 meters minus 10 meters equals 15 meters, and the vertical length as 15 meters minus 8 meters equals 7 meters. Then calculate the offset correlation value of the grid center point relative to this side. Specifically, subtract the horizontal coordinate of the core equipment area from the horizontal coordinate of the center point, and multiply the resulting horizontal offset by the vertical length of the side; subtract the vertical coordinate of the core equipment area from the vertical coordinate of the center point, and multiply the resulting vertical offset by the horizontal length of the side; subtract the result of the latter calculation from the former calculation to obtain the position correlation value corresponding to the side.

[0052] For the other two sides of the triangle, repeat the above calculations: first calculate the horizontal and vertical lengths of the sides, then multiply the horizontal offset of the center point relative to one endpoint of the side by the vertical length of the side, and multiply the vertical offset by the horizontal length of the side, taking the difference as the position correlation value of the corresponding side; when all three position correlation values ​​are non-negative or all are non-positive, it means that the grid center point is on the same side of the three sides, that is, it is located within the convex polygon area of ​​the triangle; otherwise, it is determined to be outside the area. For example, for the center point (15 meters, 12 meters), calculate the position correlation value of the core equipment area and the main transformer area as (15 minus 10) multiplied by 7 minus (12 minus 8) multiplied by 15, that is, 5 multiplied by 7 minus 4 multiplied by 15 equals 35 minus 60 equals -25. Then calculate the position correlation values ​​of the other two sides. If they are all negative, it is determined that the grid is within the area.

[0053] Based on the Euclidean distances between the center point coordinates of the analysis grids within the included area and three fixed spatial coordinate points, the local spatial weight of each analysis grid is calculated. Specifically, this involves: calculating the local spatial weight only for grids within the included area; the core logic is that the closer to the core equipment, the higher the risk weight. First, the Euclidean distances between the grid center point and the three fixed points are calculated. This is done by subtracting the horizontal coordinates of the fixed points from the horizontal coordinates of the grid center point, squaring the difference, and then subtracting the vertical coordinates of the fixed points from the vertical coordinates of the grid center point, squaring the difference as well. The two squared values ​​are then added together and the square root is taken. For example, taking the grid center point (15 meters, 12 meters) as an example, the distance to the core equipment area (10 meters, 8 meters) is the square of (15 minus 10) plus the square of (12 minus 8), and then the square root of the sum. That is, 25 plus 16 equals 41, and the square root is approximately 6.4 meters. Calculate the distance to the other two fixed points using the same method, add all three distance values ​​together, and then divide by 3 to obtain the average Euclidean distance. Take the maximum distance between any two of the three fixed points as the benchmark. For example, the distance from the high-voltage switch area to the core equipment area is approximately 20 meters. The local spatial weight is calculated by subtracting the average Euclidean distance of the grid from the maximum benchmark distance, and then dividing the difference by the maximum benchmark distance. Considering the layout of the core area of ​​the substation, the average Euclidean distance is usually between 3 and 15 meters, corresponding to a local spatial weight range of 0.25 to 0.85. If the average Euclidean distance is 3 meters, the weight is (20 minus 3) divided by 20 equals 0.85; if the average Euclidean distance is 15 meters, the weight is (20 minus 15) divided by 20 equals 0.25.

[0054] Based on local spatial weights, a global spatial calibration weight is calculated and generated. This global spatial calibration weight is then used to perform a weighted calibration of the early fire risk index, generating a calibrated early fire risk index. Specifically, this involves aggregating the local weights of all grid cells within a region to obtain the global spatial calibration weight. The calculation method is to add the local weights of all grid cells within the region, and then divide the sum by the total number of grid cells in the region. For example, the sum of the weights of fifty grid cells is 34.5, which, when divided by 50, yields 0.69. Because the grid cells within the region are all close to key areas such as core equipment areas and main transformer areas, the local weights are mostly concentrated between 0.6 and 0.9. Therefore, the global spatial calibration weight's value range is stable between 0.6 and 0.9. This global weight is then used to weight the generated early fire risk index, i.e., the early fire risk index is multiplied by the global weight to obtain the calibrated risk index.

[0055] In this embodiment, spatial weights, by clearly defining the value ranges of local and global values, accurately combine the risk priority of the core equipment area, making the assessment more in line with the actual layout of the substation.

[0056] In a preferred embodiment of the present invention, comparing the calibrated early fire risk index with a preset dynamic probability threshold to generate a preliminary fire warning signal may include:

[0057] 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 perpendicular bisector equation parameters of the two sides respectively. Specifically, the three fixed spatial coordinate points correspond to the core equipment area, main transformer area, and high-voltage switch area of ​​the substation. These three areas are high-risk areas for substation fires, and their coordinates are calculated by measuring the boundary vertices of the areas using a laser rangefinder. Taking the core equipment area as an example, this area has a rectangular structure. The horizontal and vertical coordinates of the four vertices of its rectangular boundary are measured using a laser rangefinder. The formula for calculating the horizontal coordinate of the center of the core equipment area is as follows: ,in , , , The horizontal coordinates of the four vertices are given; the formula for calculating the vertical coordinate of the center is... ,in , , , The vertical coordinates of the four vertices are calculated by adding the corresponding coordinates of the four vertices together and then dividing by four. The same measurement and calculation method is used for the main transformer area and the high-voltage switch area. The center coordinates are calculated using the coordinates of the four vertices of the corresponding area according to the above formula. Finally, three sets of coordinate data are generated, namely the coordinates of the core equipment area, the coordinates of the main transformer area, and the coordinates of the high-voltage switch area.

[0058] When calculating the perpendicular bisector parameters of the line connecting the core equipment area and the main transformer area, the coordinates of the midpoint of this side are calculated. The formula for calculating the transverse coordinates of the midpoint is as follows: The formula for calculating the longitudinal coordinate of the midpoint is: That is, add the corresponding coordinates of the core equipment area and the main transformer area and divide by two; the formula for calculating the slope of this side is: That is, subtract the longitudinal coordinates of the core equipment area from the longitudinal coordinates of the main transformer area to obtain the longitudinal difference, and subtract the lateral coordinates of the core equipment area from the lateral coordinates of the main transformer area to obtain the lateral difference. Divide the longitudinal difference by the lateral difference; then calculate the slope of the perpendicular bisector using the following formula: That is, divide the negative one by the slope of the side; finally, determine the parameters of the perpendicular bisector equation, and construct the perpendicular bisector expression according to the point-slope form, based on the midpoint coordinates and the perpendicularity of the perpendicular bisector slope k. Ensure that the line passes through the midpoint and is perpendicular to the original side.

[0059] When calculating the perpendicular bisector parameters of the line connecting the core equipment area and the high-voltage switch area, repeat the same calculation process described above. First, calculate the midpoint coordinates. The formula for calculating the midpoint's lateral coordinates is: The formula for calculating the longitudinal coordinate of the midpoint is: That is, add the corresponding coordinates of the core equipment area and the high-voltage switch area together and divide by two; then calculate the slope of the side using the following formula: That is, subtract the longitudinal coordinates of the core equipment area from the longitudinal coordinates of the high-voltage switch area to obtain the longitudinal difference, and subtract the lateral coordinates of the core equipment area from the lateral coordinates of the high-voltage switch area to obtain the lateral difference. Divide the longitudinal difference by the lateral difference; calculate the slope of the perpendicular bisector using the following formula: That is, divide the negative one by the slope of the side; finally, based on the midpoint coordinates and the perpendicular bisector slope k, apply the point-slope form. Determine the parameters of the perpendicular bisector equation.

[0060] Based on the equation parameters of the perpendicular bisectors of the two sides, solve the system of equations to find the coordinates of the intersection point of the two perpendicular bisectors, thus obtaining the coordinates of the center of the circumcircle. Specifically, this involves: using the equation parameters of the two perpendicular bisectors, solving the system of equations for the expressions of the two perpendicular bisectors to find the coordinates of the intersection point, which is the coordinate of the center of the circumcircle. When solving the system of equations, first rearrange the equation parameters of the two perpendicular bisectors into standard linear equations in two variables. The first equation is the rearranged result of the perpendicular bisector of the line connecting the core equipment area and the main transformer area, and the second equation is the rearranged 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 system of linear equations in two variables is as follows: , where x represents the horizontal coordinate and y represents the vertical coordinate.

[0061] The substitution and elimination method is used to solve the simultaneous equations. The second equation, when transformed, yields the vertical coordinate as 23.6 minus 0.4 times the horizontal coordinate. Substituting this result into the first equation, we get 15 times the horizontal coordinate plus 7 times the horizontal coordinate within parentheses (23.6 minus 0.4) equals 343. Expanding this, we get 15 times the horizontal coordinate plus 7 times 23.6 minus 7 times 0.4 equals 343. Further calculation yields 15 times the horizontal coordinate plus 165.2 minus 2.8 times equals 343. This process is repeated for the equation containing the horizontal coordinate. Combining the terms, we get 12.2 times the horizontal coordinate, which equals 343 minus 165.2, or 177.8. Dividing 177.8 by 12.2 gives a horizontal coordinate of approximately 14.6 meters. Substituting the horizontal coordinate of 14.6 meters into the formula for the vertical coordinate, which equals 23.6 minus 0.4 times the horizontal coordinate, we get a vertical coordinate of 23.6 minus 0.4 times 14.6, or 23.6 minus 5.8 equals 17.8 meters. Thus, the coordinates of the center of the circumcircle are 14.6 meters horizontally and 17.8 meters vertically.

[0062] The radius of the circumcircle is obtained by calculating the Euclidean distance between any two points, given the center coordinates and the coordinates of any one of the three fixed spatial coordinate points. Specifically, this involves calculating the radius of the circumcircle using the coordinates of the core equipment area, employing the Euclidean distance calculation method. The formula is as follows: In the specific calculation, first calculate the difference in horizontal coordinates between the center of the circle and the core equipment area. Subtract the horizontal coordinate of the core equipment area from the horizontal coordinate of the center (14.6 meters) to get 4.6 meters. Then calculate the difference in vertical coordinates between the center of the circle and the core equipment area. Subtract the vertical coordinate of the core equipment area from the vertical coordinate of the center (17.8 meters) to get 9.8 meters. Multiply the horizontal coordinate difference of 4.6 meters by 4.6 meters to get 21.16 square meters, and multiply the vertical coordinate difference of 9.8 meters by 9.8 meters to get 96.04 square meters. Add the two squared results to get 21.16 plus 96.04, which equals 117.2 square meters. Take the square root of 117.2 square meters to get approximately 10.8 meters, which is the radius of the circumscribed circle.

[0063] A dynamic probability threshold is generated by proportionally adjusting a preset base probability threshold based on the radius length. The calibrated early fire risk index is then compared with the dynamic probability threshold to generate a preliminary fire warning signal. Specifically, this involves: proportionally adjusting the preset base probability threshold based on the radius length to generate the dynamic probability threshold. The preset base probability threshold is 0.6, and the core logic of the adjustment is that the larger the monitoring range, the lower the threshold to avoid missed detections. First, the radius range is determined, taking into account the actual scale of the key monitoring area of ​​the substation; the radius range is stabilized between eight and twenty-four meters. When the radius is less than ten meters, it belongs to small-scale intensive monitoring, requiring high risk identification accuracy, and the threshold remains unchanged at the base threshold of 0.6. When the radius is greater than twenty meters, it belongs to large-scale monitoring, and the threshold is lowered to 0.4 to reduce missed detections. When the radius is between ten and twenty meters, it is adjusted using a linear attenuation method, with the coefficient calculation formula as follows: The formula for calculating the dynamic threshold is: Taking a radius of 10.8 meters as an example, substituting into the formula, we can calculate: The dynamic probability threshold is 0.584.

[0064] The calibrated early fire risk index is compared with a dynamic probability threshold to generate a preliminary fire warning signal. First, the calibrated risk index is mapped to a probability value between zero and one, using the following formula: Taking the calibrated risk index of 42.4 as an example, substituting it into the formula yields... The mapped probability value is 0.424. The mapped probability value is compared with a dynamic probability threshold, with three comparison levels: a high-risk warning is generated when the mapped probability value is greater than the dynamic probability threshold; a medium-risk warning is generated when the mapped probability value is between 0.8 times and the dynamic probability threshold; and a normal signal is output when the mapped probability value is less than 0.8 times the dynamic probability threshold. For example, with a dynamic probability threshold of 0.584, 0.8 times is calculated as... If the mapping probability value is 0.424, which is less than 0.447, a normal signal is output; if the mapping probability value is 0.6, which is greater than 0.584, a high-risk preliminary warning signal is generated.

[0065] This embodiment improves the accuracy of obtaining the coordinates of the three fixed points by measuring the boundary vertices and calculating the center coordinates using a laser rangefinder.

[0066] like Figure 2 As shown, in another preferred embodiment of the present invention, verifying the preliminary fire warning signal, and generating a final early fire identification alarm command and activating the emergency response plan after successful verification, may include:

[0067] Based on the initial fire warning signal, multi-source heterogeneous real-time monitoring data corresponding to the warning time is acquired. Based on this data, video flame and smoke feature recognition, infrared abnormal high-temperature area analysis, and environmental parameter exceedance judgment are performed to generate a multi-modal risk assessment result. Specifically, after the initial fire warning signal is generated, the system automatically extracts the precise timestamp of the warning time. Using this timestamp as the central reference, multi-source heterogeneous real-time monitoring data from 30 seconds before and after the warning is synchronously retrieved to ensure that the data fully covers the risk budding stage before the warning and the characteristic manifestation stage when the warning occurs. Data acquisition relies on three types of monitoring equipment pre-planned and deployed within the substation. All equipment establishes a stable communication link with the core analysis unit via industrial-grade Ethernet to ensure the real-time performance and integrity of data transmission. Video monitoring data is collected by high-definition network cameras deployed above the main transformer area and high-voltage switch area in the core equipment area. Two cameras are deployed at each location to create a cross-angle shooting angle, eliminating blind spots in monitoring. The camera resolution is set to 1920×1080, and the frame rate is adjusted to 25 frames per second to ensure clear capture of subtle movements in the spread of smoke and the initial form of flames. The collected video data is transmitted to the core analysis unit in real time in the form of a continuous frame sequence, with each frame accompanied by a unique timestamp and acquisition device number. Infrared temperature data is collected by infrared thermal imagers installed on the sides of each core area. The temperature measurement range of the thermal imagers is set to 0 degrees Celsius to 300 degrees Celsius, with a temperature measurement accuracy controlled to ±0.5 degrees Celsius. The acquisition frame rate is set to 10 frames per second to ensure the timeliness of temperature data while avoiding data redundancy. The collected infrared data is transmitted in the form of grayscale images containing pixel temperature information. The grayscale value of each pixel in the image has a fixed correspondence with the actual temperature value, and each frame of the image is associated with the acquisition time and acquisition location information.

[0068] Environmental parameter data is collected by three types of dedicated sensors distributed around the three core areas. Smoke sensors are deployed at ventilation openings and on top of equipment in each area, characteristic gas sensors are installed close to the sealed connections of the equipment, and temperature sensors are directly attached to key parts of the equipment casing via thermally conductive patches. The smoke sensors are set to collect data once every 0.5 seconds, the characteristic gas sensors once per second, and the temperature sensors twice per second. All raw data collected by the sensors undergoes analog-to-digital conversion, is converted into standard numerical form, and then transmitted to the core analysis unit along with the collection timestamp and sensor number. After receiving the three types of monitoring data, the core analysis unit performs targeted processing and analysis on each type of data, generates risk assessment results for each type of data, and then cross-verifies the results of the three types of data. If a fire risk is confirmed, an alarm command is generated and the corresponding emergency response plan is activated.

[0069] Video flame and smoke feature recognition is achieved using a frame difference method combined with grayscale feature analysis. First, the acquired video frame sequence is preprocessed by converting each color image to a grayscale image to reduce color interference. Then, the difference between two adjacent grayscale images is calculated to obtain a frame difference image. This frame difference image is binarized, with a grayscale difference threshold of 30. Pixels with a grayscale difference greater than 30 are marked as white, and the rest as black. Dilation and erosion operations are then performed on the binarized frame difference image. The dilation operation uses a 3x3 structuring element to scan the image, converting black pixels surrounding white pixels to white. Similarly, the erosion operation uses a 3x3 structuring element to scan the image, converting white pixels at their edges to black. These two steps eliminate noise points in the image.

[0070] The number of white pixels in the frame difference image after statistical processing is calculated, and the proportion of white pixels to the total number of pixels in the whole frame image is calculated by dividing the number of white pixels by the total number of pixels in the whole frame image. The total number of pixels in the whole frame image is the horizontal resolution of 1920 multiplied by the vertical resolution of 1080, which is 2,073,600 pixels. When the proportion of white pixels in three consecutive frames is greater than 0.05, grayscale features in the image are further extracted, and the mean and variance of grayscale in the suspicious area are calculated. The mean grayscale is the sum of the grayscale values ​​of all pixels in the suspicious area divided by the number of pixels. The variance is the sum of the squares of the differences between the grayscale value of each pixel and the mean, divided by the number of pixels. When the mean grayscale is less than 80 and the variance is greater than 100, the video flame and smoke feature identification result is determined to be risky. For the infrared abnormal high temperature area analysis, the infrared grayscale image is first converted to a temperature value. According to the calibration parameters of the thermal imager, the grayscale value of each pixel in the image is converted to the actual temperature value. The conversion method is that the actual temperature is equal to the grayscale 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 pixels in the image are traversed, and pixels with an actual temperature of not less than 80 degrees Celsius are marked. Connectivity analysis is performed on the marked high-temperature pixels, and adjacent high-temperature pixels are grouped into a high-temperature region. The adjacent determination criterion is that there are other high-temperature pixels in the four directions above, below, left, and right of a pixel.

[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 initial warning indicates a high risk, a Level 1 alarm command is generated: Immediately trigger the audible and visual alarm devices at the main control room and entrances to all core equipment areas within the substation, setting the alarm sound to no less than 90 decibels, while simultaneously displaying a continuous red flashing light; automatically send a trip control signal to the high-voltage power supply control cabinet corresponding to the warning area, cutting off the high-voltage power supply circuit in that area to prevent arcing and subsequent combustion; activate the carbon dioxide fire extinguishing devices within the warning area, automatically adjusting the nozzle direction based on the warning location information and releasing carbon dioxide extinguishing agent; and push alarm information to the mobile terminals of all on-duty maintenance personnel via the operation and maintenance management platform, including the following information: The warning includes the name and risk level of the warning area, along with a link to access the real-time video stream. If the initial warning is for medium risk, a level two alarm command is generated: triggering the main control room and audible and visual alarm devices, with the alarm sound set to no less than 80 decibels and the alarm lights flashing yellow; pushing alarm information to the mobile terminals of maintenance personnel through the maintenance management platform, including video frame screenshots of the warning area, infrared temperature curves, environmental parameter change trends, and other data; dispatching maintenance personnel with portable smoke detectors and infrared thermometers to the site for manual verification, manually activating fire extinguishing devices after confirming the presence of risk, and manually canceling the alarm command through the maintenance management platform after confirming no risk.

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

[0075] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0076] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0077] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

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 is used to select three fixed spatial coordinate points on the two-dimensional plan view of the substation monitoring area, located in the core equipment area, the main transformer area, and the high-voltage switch area, respectively. 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 positional relationship between the analysis grid points and each side of the triangle is calculated to determine the inclusion state. Based on the inclusion state and the distance relationship between the analysis grid and the three fixed spatial coordinate points, the local spatial weight is calculated. The local spatial weights of all analysis grids are aggregated to obtain the global spatial calibration weight. The global spatial calibration weight is used to perform a weighted calculation on the early fire risk index to generate the 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, Three fixed spatial coordinate points are selected on the two-dimensional plan view of the substation monitoring area, located in the core equipment area, the main transformer area, and the high-voltage switch area, respectively. Connecting these three fixed spatial coordinate points forms a triangular convex polygon region. This triangular convex polygon region is divided to generate an analysis grid. The positional relationship between the grid points and the sides of the triangle is calculated to determine the inclusion state. Based on the inclusion state and the distance relationship between the analysis grid and the three fixed spatial coordinate points, local spatial weights are calculated. The local spatial weights of all analysis grids are aggregated to obtain the global spatial calibration weight. The global spatial calibration weight is used to perform a weighted calculation on the early fire risk index to generate a calibrated early fire risk index, including: Select 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; connect the three fixed spatial coordinate points to form a triangular convex polygon area covering the key monitoring objects; 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 triangular convex polygon region based on the positional relationship between the center point coordinates and the sides of the triangle; The local spatial weight of each analysis grid is calculated based on the Euclidean distance between the center point coordinates of the analysis grid whose inclusion state is within the region and three fixed spatial coordinate points. Based on local spatial weights, global spatial calibration weights are calculated and generated; the early fire risk index is then weighted and calibrated using the global spatial calibration weights to generate the calibrated early fire risk index.

5. The substation fire early identification and alarm system with embedded AI algorithm according to claim 4, 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.

6. The substation fire early identification and alarm system with embedded AI algorithm according to claim 5, 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.

7. The substation fire early identification and alarm system with embedded AI algorithm according to claim 6, 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.

8. The substation fire early identification and alarm system with embedded AI algorithm according to claim 7, 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.

9. 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 8.

10. 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 8.

Images