Power distribution cabinet fire warning method and system based on multi-modal image feature fusion
By using a convolutional neural network model that fuses multimodal image features, parameters are adjusted in real time to adapt to environmental changes, solving the environmental interference problem of traditional power distribution cabinet fire detection and achieving efficient fire identification and early warning in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANXI ZHONGSHI ELECTRICITY TECH CO LTD
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-16
AI Technical Summary
Traditional fire detection methods for electrical distribution cabinets are easily affected by environmental interference, leading to false alarms or missed alarms. They also cannot adapt to complex lighting conditions such as day-night cycles and rainy weather, resulting in poor detection stability.
A multimodal image feature fusion method is adopted, which combines environmental data and image data through a convolutional neural network model and adjusts the model parameters in real time to adapt to different environments, thereby achieving fire identification.
It significantly improves the robustness and accuracy of fire identification, enabling it to accurately capture fire characteristics and provide early warnings in complex environments.
Smart Images
Figure CN121884286B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical distribution cabinet safety technology, specifically to a method and system for early warning of electrical distribution cabinet fires based on multimodal image feature fusion. Background Technology
[0002] In the operation and maintenance safety of power systems, distribution cabinets are critical nodes, and the accuracy and real-time performance of their fire early warning directly affect the reliability of power supply. Traditional fire detection in distribution cabinets mainly relies on single sensors, such as smoke detectors or temperature sensors. These methods have obvious technical defects: on the one hand, single sensors are easily affected by environmental interference, and in the high dust, high humidity, or electromagnetic noise environments commonly found in distribution cabinets, they are prone to false alarms or missed alarms; on the other hand, traditional visual monitoring methods often use fixed recognition models, which cannot adapt to changes in complex lighting conditions such as day-night cycles and rainy weather, resulting in poor detection stability in different environments.
[0003] Overcoming the interference of environmental factors on fire detection and improving the model's adaptability in different modal environments are problems we need to solve. To this end, we now provide a method and system for early warning of fire in distribution cabinets based on multimodal image feature fusion. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for early warning of fires in power distribution cabinets based on multimodal image feature fusion.
[0005] The objective of this invention can be achieved through the following technical solution: a method for early warning of fires in power distribution cabinets based on multimodal image feature fusion, comprising:
[0006] Collect sample data of electrical distribution cabinet fires under multimodal environments, construct a convolutional neural network model, and train the convolutional neural network model using the electrical distribution cabinet fire sample data;
[0007] Real-time environmental data and image data of the space where the power distribution cabinet is located are acquired through environmental data acquisition terminals and image data acquisition terminals.
[0008] The environmental data and the image data are preprocessed accordingly to obtain corresponding environmental data sample points and image frame samples.
[0009] The environmental data sample points and the image frame samples are input into the trained convolutional neural network model. The model's environmental modality is obtained through the environmental data sample points. The environmental parameter value sequence of real-time environmental data within a time span is obtained, and the environmental parameter value sequence is matched with a preset environmental type parameter range to obtain the algorithm's environmental modality. The algorithm's environmental modality is compared with the model's environmental modality. If they are inconsistent, the weights of the corresponding convolutional layers in the convolutional neural network model are adjusted according to the differing environmental attribute labels.
[0010] The adjusted convolutional neural network model is used to perform image recognition on the image frame samples and output the fire recognition result.
[0011] Furthermore, the multimodal environment includes environmental categories composed of several environmental attribute tags, and the environmental attribute tags include at least one of night, day, rainy day, sunny day, low temperature, high temperature, and normal temperature;
[0012] The electrical distribution cabinet fire sample data includes image sample data and / or environmental sample data.
[0013] Furthermore, the construction of the convolutional neural network model, and the training of the convolutional neural network model using the fire sample data of the power distribution cabinet, includes:
[0014] The environmental sample data corresponding to different environmental types are summarized, and the summarized environmental sample data is divided into training set, test set and validation set corresponding to the environmental type.
[0015] The convolutional neural network model is trained using a training set, and then tested and validated using a test set and a validation set, thereby enabling the convolutional neural network model to recognize different types of environments.
[0016] Fire labels are set according to fire factors in image sample data, and the fire labels include at least one of fire point labels, smoke labels and overheat labels.
[0017] The image sample data containing different fire labels is aggregated, and the aggregated image sample data is divided into training set, test set and validation set corresponding to the fire labels.
[0018] The convolutional neural network model is trained using the training set corresponding to the fire label, and then tested and validated using the test set and validation set, thereby enabling the convolutional neural network model to recognize different fire labels.
[0019] Furthermore, the real-time acquisition of environmental data and image data of the space where the power distribution cabinet is located through the environmental data acquisition terminal and the image data acquisition terminal includes:
[0020] Corresponding image data acquisition terminals and environmental data acquisition terminals are set up in different locations of the power distribution cabinet;
[0021] The image data acquisition terminal and the environmental data acquisition terminal located in the same direction are associated to obtain the corresponding real-time environmental data and image data. The real-time environmental data includes at least one of the following: ambient temperature, light intensity and air humidity.
[0022] After time alignment of the acquired real-time environmental data and image data, appropriate preprocessing is performed.
[0023] Furthermore, the real-time environmental data will undergo appropriate preprocessing, including:
[0024] Construct a timeline, map the obtained real-time environmental data onto the timeline, and generate corresponding environmental data change curves;
[0025] Select the current time as the reference time and obtain the environmental data corresponding to the environmental data change curve at the reference time;
[0026] The environmental data at each position on the newly generated environmental data change curve is compared with the environmental data corresponding to the baseline time to obtain the environmental data change rate;
[0027] If the rate of change of environmental data exceeds the corresponding rate of change threshold, then each environmental data point at the corresponding time will be used as an environmental data sample point.
[0028] Further, the image data undergoes appropriate preprocessing, including:
[0029] The obtained image data is converted into image frames, and the converted image frames are mapped onto the timeline;
[0030] Based on the generated environmental data sample points, the image frames at the corresponding times of the environmental data sample points are marked as image frame samples, and the obtained image frame samples are used as input for subsequent models.
[0031] Furthermore, the environmental data sample points and the image frame samples are input into the trained convolutional neural network model to obtain the model's environmental modalities through the environmental data sample points, including:
[0032] The environmental data and image frame samples corresponding to the environmental data sample point closest to the current time are input into the completed training convolutional neural network model;
[0033] Based on the environmental data corresponding to the input environmental data sample points, the model environmental modal type of the current location of the power distribution cabinet is obtained.
[0034] Furthermore, the sequence of environmental parameter values in real-time environmental data within a time span is obtained, and the sequence of environmental parameter values is matched with a preset range of environmental type parameters to obtain the algorithm's environmental modal types, including:
[0035] In the environmental data change curves on the time axis corresponding to the real-time environmental data, a corresponding time span is set according to the time corresponding to the environmental data sample point; the middle time of the time span is the time corresponding to the environmental data sample point.
[0036] Based on the environmental data change curves across a time span, an environmental parameter value sequence is obtained through integration, specifically:
[0037] ;
[0038] in, A sequence of environmental parameter values. This represents a curve showing changes in environmental data, where K represents the type of environmental data. This represents the environmental data value corresponding to an environmental data sample point, where Kmid represents an environmental data sample point of type K. Record the starting moment of the time transition. The end time is denoted as ;
[0039] The obtained environmental parameter value sequence The algorithm matches the environment type parameter range corresponding to each environment attribute label and outputs the final algorithm environment modality type based on the matching results.
[0040] Furthermore, the process of performing image recognition on image frame samples using the adjusted convolutional neural network model and outputting fire recognition results is as follows:
[0041] Based on the image frame samples input into the convolutional neural network model, the system identifies whether there are corresponding fire factors within the image frame samples. If fire factors exist, the system generates corresponding fire labels based on the existing fire factors and marks the corresponding areas within the image frame samples with the generated fire labels, while simultaneously generating fire warning information.
[0042] Furthermore, a fire early warning system for distribution cabinets based on multimodal image feature fusion includes:
[0043] The model building and training module is used to collect sample data of power distribution cabinet fires in a multimodal environment, and to build a convolutional neural network model. The convolutional neural network model is trained using the sample data of power distribution cabinet fires to obtain a trained convolutional neural network model.
[0044] The data acquisition module is used to acquire real-time environmental data and image data of the space where the power distribution cabinet is located.
[0045] The data preprocessing module is used to perform corresponding preprocessing on the real-time environmental data and the image data to obtain corresponding environmental data sample points and image frame samples.
[0046] The model adjustment module is used to input the environmental data sample points and image frame samples into the trained convolutional neural network model, obtain the model's environmental modality type through the environmental data sample points; acquire the environmental parameter value sequence of real-time environmental data within a time span, and match the environmental parameter value sequence with a preset environmental type parameter range to obtain the algorithm's environmental modality type; compare the algorithm's environmental modality type with the model's environmental modality type, and if they are inconsistent, adjust the weights of the corresponding convolutional layers in the convolutional neural network model according to the differing environmental attribute labels;
[0047] The fire identification module is used to perform image recognition on image frame samples using an adjusted convolutional neural network model and output the fire identification results.
[0048] Compared with the prior art, the beneficial effects of the present invention are:
[0049] 1. By constructing a convolutional neural network with environmental type recognition capabilities, the model parameters are dynamically adjusted using real-time collected environmental data. When the algorithm's environmental modality is inconsistent with the model output, the weights of the corresponding convolutional layers are adjusted accordingly, enabling the model to adapt to complex environments such as day and night, sunny and cloudy weather, and temperature changes. This significantly improves the robustness and accuracy of fire identification under different working conditions. Compared with adaptive methods that rely solely on internal attention mechanisms, the external environmental perception-guided weight adjustment method of this invention is more interpretable and targeted, and can directly address the model failure problem caused by drastic environmental changes.
[0050] 2. By acquiring images and environmental data from dual terminals and combining time axis alignment and data change rate filtering mechanisms, precise matching of multi-source data in the time dimension is achieved. Based on the refined identification of multiple tags such as fire point, smoke, and overheating, fire characteristics can be accurately captured in the early stage, realizing end-to-end accurate early warning from "environmental perception" to "image recognition". Attached Figure Description
[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0052] Figure 1 This is a flowchart illustrating the fire early warning method for power distribution cabinets based on multimodal image feature fusion according to the present invention. Detailed Implementation
[0053] like Figure 1 As shown, a method for early warning of fires in power distribution cabinets based on multimodal image feature fusion includes:
[0054] Collect sample data of electrical distribution cabinet fires under multimodal environments and construct a convolutional neural network model. Train the constructed convolutional neural network model using the collected sample data of electrical distribution cabinets under multimodal environments.
[0055] Deploy environmental data acquisition terminals and image data acquisition terminals to acquire real-time environmental data and image data of the space where the power distribution cabinet is located;
[0056] The obtained real-time environmental data and image data are preprocessed accordingly to obtain the corresponding environmental data sample points and image frame samples.
[0057] The obtained environmental data sample points and image frame samples are input into the trained convolutional neural network model. The model's environmental modality is obtained through the input environmental data sample points. The environmental parameter value sequence of real-time environmental data within a time span is obtained, and the obtained environmental parameter value sequence is matched with the preset environmental type parameter range to obtain the algorithm's environmental modality. The algorithm's environmental modality is compared with the model's environmental modality. If they are inconsistent, the weights of the corresponding convolutional layers in the convolutional neural network model are adjusted according to the environmental attribute labels that differ.
[0058] The adjusted convolutional neural network model is used to perform image recognition on image frame samples and output fire recognition results.
[0059] It should be further explained that, in the specific implementation process, the multimodal environment includes environmental types composed of several environmental attribute tags, and the environmental attribute tags include at least one of night, day, rainy day, sunny day, low temperature, high temperature, and normal temperature;
[0060] An environment type consisting of several environmental attribute tags can be: night-rainy day-low temperature, night-sunny day-normal temperature, etc., which will not be listed here.
[0061] Then, fire sample data of electrical distribution cabinets under different environmental conditions are collected. The fire sample data of electrical distribution cabinets includes image sample data and / or environmental sample data.
[0062] The process of constructing a convolutional neural network model and training it using sample data of distribution cabinets in multimodal environments includes:
[0063] The environmental sample data corresponding to the same environmental type are summarized, and the summarized environmental sample data is divided into training set, test set and validation set corresponding to the environmental type.
[0064] The convolutional neural network model is trained using a training set, and then tested and validated using a test set and a validation set, thereby enabling the convolutional neural network model to recognize the type of environment.
[0065] The convolutional neural network model is trained in the same way using training sets, test sets, and validation sets corresponding to different environment types, so that the convolutional neural network model has the function of recognizing the corresponding environment types.
[0066] Fire labels are set according to fire factors in image sample data. Fire labels include at least one of fire point labels, smoke labels and overheat labels.
[0067] The image sample data containing different fire labels is aggregated, and the aggregated image sample data is divided into training set, test set and validation set corresponding to the fire label.
[0068] The convolutional neural network model is trained using the training set corresponding to the fire label, and then tested and validated using the test set and validation set, thereby enabling the convolutional neural network model to recognize the fire label.
[0069] The convolutional neural network model was trained in the same way using the training set, test set, and validation set corresponding to different fire labels, so that the convolutional neural network model could identify the corresponding fire label.
[0070] It should be further explained that when the test results of the convolutional neural network model reach the expected accuracy, it means that the training for the corresponding environment type is complete. If it does not reach the expected accuracy, iterative training will be performed until the test results reach the expected accuracy or the number of iterations reaches the preset number.
[0071] It should be further explained that, in the specific implementation process, the deployment of environmental data acquisition terminals and image data acquisition terminals to acquire real-time environmental data and image data of the space where the power distribution cabinet is located includes:
[0072] Set up corresponding image data acquisition terminals and environmental data acquisition terminals at different locations of the power distribution cabinet;
[0073] Image data acquisition terminals and environmental data acquisition terminals located in the same direction are associated to obtain corresponding real-time environmental data and image data. The real-time environmental data includes at least one of ambient temperature, light intensity, and air humidity.
[0074] After time alignment of the obtained real-time environmental data and image data, the obtained real-time environmental data and image data are preprocessed accordingly.
[0075] It should be further explained that, in the specific implementation process, the preprocessing of the acquired real-time environmental data includes:
[0076] Construct a timeline, map the obtained real-time environmental data onto the timeline, and generate corresponding environmental data change curves;
[0077] Select the current time as the reference time and obtain the environmental data corresponding to the environmental data change curve at the reference time;
[0078] The environmental data at each position on the newly generated environmental data change curve is compared with the environmental data corresponding to the baseline time to obtain the environmental data change rate;
[0079] If the rate of change of environmental data exceeds the corresponding rate of change threshold, then each environmental data point at the corresponding time will be used as an environmental data sample point.
[0080] Example:
[0081] Taking ambient temperature as an example, the ambient temperature corresponding to the reference time t1 is recorded as T1, and the ambient temperature at the corresponding position on the newly generated ambient data change curve is recorded as T2.
[0082] The rate of change of environmental data is H1 = |(T2-T1) / T1|;
[0083] The rate of change of environmental data corresponding to ambient temperature is denoted as w;
[0084] If H1≤w, no operation is performed. If H1>w, the corresponding time is recorded as t2, and the ambient temperature, light intensity, and air humidity at time t2 are obtained as environmental data sample points. That is, as long as the rate of change of any environmental data change curve exceeds the corresponding rate of change threshold, environmental data sample points containing all environmental data are generated.
[0085] The rate of change threshold is set according to the environmental data type, for example:
[0086] The threshold for the rate of change of ambient temperature can be set to 5%;
[0087] The threshold for the rate of change of light intensity can be set to 20%;
[0088] The threshold for the rate of change of air humidity can be set to 10%.
[0089] The aforementioned thresholds can be derived by analyzing the normal fluctuation range in historical environmental data, or determined through a limited number of tests based on the actual operating conditions of the environment where the distribution cabinet is located. When the rate of change of a certain environmental data exceeds the corresponding threshold, it is considered that the environment has changed significantly, and sample point collection needs to be triggered.
[0090] It should be further explained that, in the specific implementation process, the preprocessing of the acquired image data includes:
[0091] The obtained image data is converted into image frames, and the converted image frames are mapped onto the timeline;
[0092] Based on the generated environmental data sample points, the image frames at the corresponding times of the environmental data sample points are marked as image frame samples, and the obtained image frame samples are used as input for subsequent models.
[0093] It should be further explained that, in the specific implementation process, the process of inputting the obtained environmental data sample points and image frame samples into the trained convolutional neural network model, and obtaining the model's environmental modalities through the input environmental data sample points, includes:
[0094] The environmental data and image frame samples corresponding to the environmental data sample point closest to the current time are input into the completed training convolutional neural network model;
[0095] Based on the environmental data corresponding to the input environmental data sample points, the model environmental modal type of the current location of the power distribution cabinet is obtained.
[0096] It should be further explained that, in the specific implementation process, the process of adjusting the convolutional neural network model according to the model environment modality includes:
[0097] In the environmental data change curves on the time axis corresponding to the real-time environmental data, a corresponding time span is set according to the time corresponding to the environmental data sample point; the middle time of the time span is the time corresponding to the environmental data sample point.
[0098] Record the start time of the time span as The end time is recorded as ;
[0099] Based on the environmental data change curves across time zones, the corresponding algorithm environmental modal types are obtained;
[0100] Specifically:
[0101] ;
[0102] in, A sequence of environmental parameter values. This represents a curve showing changes in environmental data, where K represents the type of environmental data. This represents the environmental data value corresponding to the environmental data sample point, and Kmid represents the environmental data sample point of type K;
[0103] The obtained environmental parameter value sequence Match the environment type parameter range corresponding to each environment attribute label, and output the final algorithm environment modality type based on the matching results;
[0104] It should be further explained that the length of the time span in the above formula (i.e. It can be set according to the changing characteristics of environmental data.
[0105] As a preferred implementation, the time span can be set between 5 and 30 minutes, with the specific value determined based on the fluctuation frequency of the environmental data change curve: when environmental data changes drastically (e.g., light intensity changes frequently due to rapid cloud movement), a shorter time span (e.g., 5-10 minutes) is preferable to capture instantaneous environmental characteristics; when environmental data changes gradually (e.g., temperature rises and falls slowly), a longer time span (e.g., 20-30 minutes) can be used to obtain a stable environmental modal assessment. The environmental parameter value sequence... Essentially, it is a normalized index of the average change level of environmental data over a time span relative to the instantaneous value of the sample point. Its value range is usually between 0.5 and 1.5. Those skilled in the art can determine the specific threshold that matches each environmental attribute label through a limited number of experiments based on actual working conditions without excessive labor.
[0106] As a specific implementation method, the length of the time span can be dynamically adjusted based on the fluctuation frequency of the environmental data:
[0107] When light intensity fluctuates frequently due to cloud movement, tree obstruction, or other reasons (e.g., changes more than 3 times per minute), a shorter time span, such as 5 to 10 minutes, should be used to capture the instantaneous environmental change characteristics.
[0108] When the ambient temperature changes slowly with the day-night cycle (e.g., the change does not exceed 2°C per hour), it is advisable to use a longer time span, such as 20-30 minutes, to obtain a stable environmental modal assessment.
[0109] For air humidity, a medium span of 15-20 minutes can be used in rainy or humid environments; a longer span of 25-30 minutes can be used in dry and stable environments.
[0110] The obtained algorithm environment modal types are compared with the model environment modal types. If the algorithm environment modal types are the same as the model environment modal types, it means that the convolutional neural network model does not need to be adjusted. Otherwise, if they are different, the weights of the corresponding convolutional layers in the convolutional neural network model are adjusted according to the different environment attribute labels. The adjustment rules are set by technical personnel according to the actual situation.
[0111] As a specific implementation method for matching, the obtained sequence of environmental parameter values is matched with the range of environmental type parameters corresponding to each environmental attribute label, including the following steps:
[0112] Step 1: Calculate the matching degree of a single attribute
[0113] For each environmental attribute label (e.g., daytime, nighttime, rainy day), a pre-defined range of environmental category parameters is established. The currently calculated sequence of environmental parameter values is compared to the parameter range for that label: if the sequence falls entirely within the label's parameter range, the label's match is considered to be at its highest value (1); if it falls outside the range, the match is determined by the degree of deviation from the range boundary—smaller deviation results in a higher match, and larger deviation results in a lower match. The match score range is set between 0 and 1, with values closer to 1 indicating a higher match.
[0114] Step 2: Calculation of Combined Matching Degree
[0115] Since a real-world environmental modality is composed of multiple environmental attribute labels (e.g., "night-rainy day-low temperature"), each candidate environmental modality is composed of a set of environmental attribute labels. The combined matching degree of the candidate environmental modality is the minimum value among the matching degrees of its individual environmental attribute labels, i.e., the "barrel principle"—the least matching label determines the overall matching degree.
[0116] Step 3: Algorithm Environment Modality Decision
[0117] Iterate through all possible candidate environment modalities and select the candidate modality with the highest combination matching degree that exceeds a preset confidence threshold (e.g., 0.8) as the final algorithm environment modality. If the combination matching degree of all candidate modalities is lower than the threshold, the current environment state is determined to be ambiguous. At this time, the automatic adjustment of the model can be paused, or the preset default environment modality can be used to continue running.
[0118] The matching degree calculation rules and confidence thresholds mentioned above can be calibrated by those skilled in the art through a limited number of experiments based on actual working conditions, without the need for excessive labor.
[0119] As an example, the range of environment type parameters corresponding to each environment attribute label can be pre-calibrated experimentally as follows:
[0120] Daytime: Light intensity ≥ 500 lux;
[0121] Nighttime: Light intensity <10 lux;
[0122] Rainy days: light intensity <300 lux and air humidity ≥70%;
[0123] Sunny day: Light intensity ≥ 500 lux and air humidity < 60%;
[0124] Low temperature: Ambient temperature <5℃;
[0125] Room temperature: 5℃ ≤ ambient temperature ≤ 30℃;
[0126] High temperature: Ambient temperature > 30℃.
[0127] It should be noted that the above values are only a preferred embodiment of the present invention. In actual applications, adjustments can be made based on the geographical location of the distribution cabinet, seasonal changes, and historical data statistics. Those skilled in the art can determine the parameter range applicable to specific working conditions through a limited number of experiments without excessive labor.
[0128] The adjustment of the weights of the corresponding convolutional layers based on the differing environmental attribute labels can be achieved using one or a combination of the following methods:
[0129] Method 1 (Weight Scaling): If the model's environment modality and the algorithm's environment modality differ in a certain environmental attribute label (such as "night"), the weights of the convolutional layer corresponding to that environmental attribute label (such as the convolutional kernel responsible for extracting illumination features) are multiplied by a coefficient greater than 1 (such as 1.2) to enhance the layer's response to the current environmental features; conversely, if the difference is that the model overestimates the influence of that environmental attribute, it can be multiplied by a coefficient less than 1 (such as 0.8) for attenuation.
[0130] Method 2 (Bias Adjustment): For convolutional layers corresponding to different environmental attributes, increase or decrease their bias terms by a preset adjustment amount (such as +0.1 or -0.1) to change the output threshold of the activation function of that layer.
[0131] Method 3 (Loading pre-trained weights): Multiple sets of convolutional layer weights are pre-trained for different environmental attribute labels. When an environmental modality difference is detected, the pre-trained weights that match the environmental attribute label are directly loaded to replace the original weights.
[0132] Method 4 (Learning Rate Adjustment): In subsequent online learning or fine-tuning processes, set a higher learning rate for convolutional layers corresponding to different environmental attributes to enable them to adapt to the new environment more quickly.
[0133] The aforementioned adjustment rules can be pre-set by technicians based on actual working conditions and model structure, and their effectiveness can be verified through a limited number of experiments. The adjusted model, used for subsequent fire identification, can significantly improve environmental adaptability.
[0134] It should be further explained that, in the specific implementation process, the process of performing image recognition on image frame samples using the adjusted convolutional neural network model and outputting fire recognition results is as follows:
[0135] Based on the image frame samples input into the convolutional neural network model, the system identifies whether there are corresponding fire factors within the image frame samples. If fire factors exist, the system generates corresponding fire labels based on the existing fire factors and marks the corresponding areas within the image frame samples with the generated fire labels, while simultaneously generating fire warning information.
[0136] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications or equivalent substitutions made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for early warning of fires in power distribution cabinets based on multimodal image feature fusion, characterized in that, include: Collect sample data of electrical distribution cabinet fires under multimodal environments, construct a convolutional neural network model, and train the convolutional neural network model using the electrical distribution cabinet fire sample data; Real-time environmental data and image data of the space where the power distribution cabinet is located are acquired through environmental data acquisition terminals and image data acquisition terminals. The real-time environmental data and the image data are preprocessed accordingly to obtain corresponding environmental data sample points and image frame samples. The environmental data sample points and the image frame samples are input into the trained convolutional neural network model. The model's environmental modality is obtained through the environmental data sample points. The environmental parameter value sequence of real-time environmental data within a time span is obtained, and the environmental parameter value sequence is matched with a preset environmental type parameter range to obtain the algorithm's environmental modality. The algorithm's environmental modality is compared with the model's environmental modality. If they are inconsistent, the weights of the corresponding convolutional layers in the convolutional neural network model are adjusted according to the differing environmental attribute labels. The adjusted convolutional neural network model is used to perform image recognition on the image frame samples and output the fire recognition result.
2. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 1, characterized in that, The multimodal environment includes several environmental attribute tags, which include at least one of the following: night, day, rainy day, sunny day, low temperature, high temperature, and normal temperature. The sample data of the electrical distribution cabinet fire includes image sample data and environmental sample data.
3. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 1, characterized in that, The construction of the convolutional neural network model, which involves training the model using the fire sample data from the distribution cabinet, includes: The environmental sample data corresponding to different environmental types are summarized, and the summarized environmental sample data is divided into training set, test set and validation set corresponding to the environmental type. The convolutional neural network model is trained using a training set, and then tested and validated using a test set and a validation set, thereby enabling the convolutional neural network model to recognize different types of environments. Fire labels are set according to fire factors in image sample data, and the fire labels include at least one of fire point labels, smoke labels and overheat labels. The image sample data containing different fire labels is aggregated, and the aggregated image sample data is divided into training set, test set and validation set corresponding to the fire labels. The convolutional neural network model is trained using the training set corresponding to the fire label, and then tested and validated using the test set and validation set, thereby enabling the convolutional neural network model to recognize different fire labels.
4. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 1, characterized in that, The process of acquiring real-time environmental data and image data of the space where the power distribution cabinet is located through an environmental data acquisition terminal and an image data acquisition terminal includes: Corresponding image data acquisition terminals and environmental data acquisition terminals are set up in different locations of the power distribution cabinet; The image data acquisition terminal and the environmental data acquisition terminal located in the same direction are associated to obtain the corresponding real-time environmental data and image data. The real-time environmental data includes at least one of ambient temperature, light intensity and air humidity. After aligning the real-time environmental data and the image data in time, appropriate preprocessing is performed.
5. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 1, characterized in that, The preprocessing of real-time environmental data includes: Construct a time axis, map the real-time environmental data onto the time axis, and generate a corresponding environmental data change curve; Select the current time as the reference time and obtain the environmental data corresponding to the environmental data change curve at the reference time; The environmental data at each position on the newly generated environmental data change curve is compared with the environmental data corresponding to the baseline time to obtain the environmental data change rate; If the rate of change of environmental data exceeds the corresponding rate of change threshold, then each environmental data point at the corresponding time will be used as an environmental data sample point.
6. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 5, characterized in that, Preprocessing the image data includes: The image data is converted into image frames, and the converted image frames are mapped onto the timeline. Based on the generated environmental data sample points, the image frames at the corresponding times of the environmental data sample points are marked as image frame samples, and the obtained image frame samples are used as input for subsequent models.
7. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 5, characterized in that, The environmental data sample points and the image frame samples are input into the trained convolutional neural network model. The environmental modalities of the model are obtained through the environmental data sample points, including: The environmental data and image frame samples corresponding to the environmental data sample point closest to the current time are input into the completed training convolutional neural network model; Based on the environmental data corresponding to the input environmental data sample points, the model environmental modal type of the current location of the power distribution cabinet is obtained.
8. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 5, characterized in that, Obtain the sequence of environmental parameter values from real-time environmental data within a time span, and match the sequence of environmental parameter values with a preset range of environmental type parameters to obtain the algorithm's environmental modal types, including: In the environmental data change curves on the time axis corresponding to the real-time environmental data, a corresponding time span is set according to the time corresponding to the environmental data sample point; the middle time of the time span is the time corresponding to the environmental data sample point. Based on the environmental data change curves across a time span, an environmental parameter value sequence is obtained through integration, specifically: ； in, A sequence of environmental parameter values. This represents a curve showing changes in environmental data, where K represents the type of environmental data. This represents the environmental data value corresponding to an environmental data sample point, where Kmid represents an environmental data sample point of type K. This marks the start of a time transition. The end time; The obtained environmental parameter value sequence The algorithm matches the environment type parameter range corresponding to each environment attribute label and outputs the final algorithm environment modality type based on the matching results.
9. The method for early warning of fire in distribution cabinets based on multimodal image feature fusion according to claim 1, characterized in that, The process of performing image recognition on image frame samples using an adjusted convolutional neural network model and outputting fire recognition results includes: Based on the image frame samples input into the convolutional neural network model, the system identifies whether there are corresponding fire factors within the image frame samples. If fire factors are present, a corresponding fire label is generated based on the existing fire factors, and the generated fire label is marked in the corresponding area within the image frame sample to generate fire early warning information.
10. A distribution cabinet fire early warning system applied to the distribution cabinet fire early warning method based on multimodal image feature fusion as described in any one of claims 1 to 9, characterized in that, include: The model building and training module is used to collect sample data of power distribution cabinet fires in a multimodal environment, and to build a convolutional neural network model. The convolutional neural network model is trained using the sample data of power distribution cabinet fires to obtain a trained convolutional neural network model. The data acquisition module is used to acquire real-time environmental data and image data of the space where the power distribution cabinet is located. The data preprocessing module is used to preprocess the real-time environmental data and image data respectively to obtain corresponding environmental data sample points and image frame samples. The model adjustment module is used to input the environmental data sample points and the image frame samples into the trained convolutional neural network model, obtain the model's environmental modality type through the environmental data sample points; acquire the environmental parameter value sequence of real-time environmental data within a time span, and match the environmental parameter value sequence with a preset environmental type parameter range to obtain the algorithm's environmental modality type; compare the algorithm's environmental modality type with the model's environmental modality type, and if they are inconsistent, adjust the weights of the corresponding convolutional layers in the convolutional neural network model according to the differing environmental attribute labels; The fire identification module is used to perform image recognition on the image frame samples using an adjusted convolutional neural network model and output the fire identification result.