Electric energy metering box with intelligent monitoring and fault early warning functions

Abstract

The invention discloses an electric energy metering box with intelligent monitoring and fault early warning functions, and relates to the technical field of intelligent power equipment. An intelligent monitoring module obtains monitoring data of the electric energy metering box in real time; the data processing module integrates a machine learning algorithm to establish an electric energy metering box operation state model and a multi-category target detection model, and performs deep analysis and prediction on monitoring data to obtain processed data; the data transmission module uploads the monitoring data and the processed data to a remote monitoring center in real time; and the fault early warning module sets an early warning threshold according to the data processing result, and triggers an early warning mechanism when the monitoring data exceed the threshold. According to the electric energy metering box, parameters such as temperature, current and humidity of key parts in the box are monitored in real time, deep analysis and prediction of element aging degree, line contact condition and humidity related hidden dangers are realized, and fault early warning is sent out in time, so that the operation reliability of the electric energy metering box is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent power equipment, more particularly to an electric energy metering box with intelligent monitoring and fault early warning functions. BACKGROUND

[0002] At present, in the operation system of the power system, the electric energy metering box as the core equipment of the power metering and distribution link undertakes the key responsibilities of accurate metering of electric energy consumption and reasonable distribution of electric power resources, and the stability and reliability of its operation state are directly related to the efficient operation of the entire power system and the power experience of users.

[0003] In practical application, the traditional electric energy metering box lacks effective monitoring means for the running state of the equipment in the box, and has many obvious drawbacks. On the one hand, for the aging problem of the electrical components in the box, since there is no real-time monitoring mechanism, the performance degradation of the components cannot be detected in time, and often the components are found to be completely failed; on the other hand, in the face of line overload or short circuit, the traditional metering box is difficult to quickly and accurately identify abnormal current changes, and cannot take measures to intervene at the initial stage of the fault.

[0004] The above problems cause the traditional electric energy metering box to be usually repaired after the fault occurs, and this post-maintenance mode not only causes the interruption of power supply, seriously affects the normal power use of users, but also greatly increases the maintenance cost and time, and also brings potential threat to the safe and stable operation of the power system.

[0005] Therefore, how to provide an electric energy metering box with intelligent monitoring and fault early warning functions, which can monitor the running state in the box in real time and give fault early warning, is a problem that those skilled in the art need to solve. SUMMARY

[0006] Therefore, the present application provides an electric energy metering box with intelligent monitoring and fault early warning functions, which monitors the temperature, current and humidity of the key parts in the box in real time, uses advanced data processing technology and machine learning algorithm to realize deep analysis and prediction of the component aging degree, line contact condition and humidity related hidden dangers, and timely gives fault early warning, so as to improve the reliability of the operation of the electric energy metering box, reduce the fault rate, ensure the stable operation of the power system, and meet the needs of efficient, stable and safe operation of modern power system.

[0007] In order to achieve the above purpose, the present application adopts the following technical scheme: an electric energy metering box with intelligent monitoring and fault early warning functions, comprising:

[0008] a box body and an intelligent monitoring and fault early warning device, the intelligent monitoring and fault early warning device is integrated on the box body;

[0009] The intelligent monitoring and fault early warning device includes an intelligent monitoring module, a data processing module, a data transmission module, and a fault early warning module;

[0010] The intelligent monitoring module is used to acquire monitoring data from the power metering box in real time;

[0011] The data processing module is used to integrate machine learning algorithms to establish an operating status model of the power metering box and a multi-category target detection model. Based on the operating status model of the power metering box and the multi-category target detection model, the monitoring data is analyzed and predicted in depth to obtain the processed data.

[0012] The data transmission module is used to upload the monitoring data and the processed data to the remote monitoring center in real time.

[0013] The fault early warning module is used to set early warning thresholds based on data processing results. When the monitored data exceeds the threshold, the early warning mechanism is triggered to remind maintenance personnel and send early warning information to the remote monitoring center.

[0014] Preferably, the intelligent monitoring module includes: a temperature sensor, a current sensor, a humidity sensor, and an image sensor, wherein the temperature sensor, current sensor, and humidity sensor respectively collect temperature, current, and humidity data of the power metering box;

[0015] The image sensor acquires image information of key areas of the metering box door, lock, and outer shell.

[0016] Preferably, the machine learning algorithm adopts the Stacking ensemble learning framework, which includes a base learning layer, a meta-learning layer, and a parameter optimization module;

[0017] The base learning layer includes Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Lightweight Gradient Boosting Machine (LightGBM), which extract features from time-series data of temperature, current, and humidity, and output component aging index, circuit health, and humidity risk value, respectively.

[0018] The meta-learning layer uses a linear regression model to fuse the output of the base learning layer with the original monitoring data, establishing a multi-dimensional evaluation model. The calculation formula is as follows:

[0019] Y = β0 + β1X temp +β2X curr +β3X hum +ξ;

[0020] Among them, X temp X is the temperature feature vector. curr Let X be the current eigenvector. hum Let β0-β3 be the humidity feature vector, β0-β3 be the regression coefficients, and ξ be the random error term;

[0021] The parameter optimization module optimizes the model parameters using k-fold cross-validation method.

[0022] Preferably, training the electric energy metering box operation state model comprises:

[0023] Integrate historical monitoring data and fault records (including component aging, line fault, moisture, etc. Label) to form a labeled sample set;

[0024] Take the temperature peak and temperature rise rate as input, and output the component aging index through decision tree integration;

[0025] Take the current fluctuation amplitude and current-time integral feature as input, and iteratively train the output line contact risk score;

[0026] Take the humidity cumulative change and moisture risk index as input, use histogram algorithm to accelerate training, and output humidity risk level;

[0027] The meta-learning layer fuses the base learning layer output result and the original feature, trains a linear regression model through k-fold cross-validation, and the objective function is to minimize the mean square error (MSE) of the predicted value and the true label.

[0028] Preferably, based on the trained electric energy metering box operation state model, the prediction comprises:

[0029] Pretreat and extract features from real-time monitoring data to generate a feature vector;

[0030] The base learning layer outputs intermediate results in parallel, and the meta-learning layer comprehensively calculates the final fault probability;

[0031] Use sliding window recursion to predict the parameter change in the next hour, and trigger early warning when the predicted value exceeds the preset threshold.

[0032] Preferably, the collected image information is preprocessed;

[0033] An abnormal feature recognition algorithm is used to construct a multi-class target detection model, and the model parameters are optimized by calling the k-fold cross-validation method;

[0034] Feature extraction is performed on the preprocessed image, and the trained model is used to detect abnormal areas in real time, and output the abnormal type and confidence.

[0035] Preferably, the multi-class target detection model uses an improved DeepLabv3+ semantic segmentation network architecture, which includes:

[0036] The backbone network adopts a ResNet50 network as a feature extraction backbone, including a deep feature extraction layer, a middle feature extraction layer and a low feature extraction layer, which respectively extract low-level detail features, middle-level texture features and deep-level semantic features;

[0037] The attention mechanism: an attention module (SimAM) is added after the deep, middle and low feature extraction layers to calculate the energy function after feature normalization.

[0038] Multi-scale feature fusion: the deep features are processed by an atrous spatial convolution pyramid (ASPP) and then 4 times up-sampled, the middle features are processed by a 1x1 convolution and then 2 times up-sampled, and the low features are refined by a 3x3 depth separable convolution after being spliced with the low features, and the segmentation result is output.

[0039] According to the technical solution, compared with the prior art, the electric energy metering box with intelligent monitoring and fault warning function is provided, which can realize deep analysis and prediction of component aging degree, line contact condition and humidity related hidden danger by real-time monitoring of temperature, current and humidity and other parameters of key parts in the box, using advanced data processing technology and machine learning algorithm, and timely issuing fault warning, so as to improve the reliability of electric energy metering box operation, reduce the failure rate, ensure the stable operation of power system, meet the demand of modern power system efficient, stable and safe operation. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0041] Figure 1 The structure diagram of the intelligent monitoring and fault warning device provided by the present application is shown. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0043] The embodiment of the present application discloses an electric energy metering box with intelligent monitoring and fault warning function, comprising:

[0044] The box body and the intelligent monitoring and fault early warning device are integrated on the box body.

[0045] As shown in Figure 1 The intelligent monitoring and fault early warning device includes an intelligent monitoring module, a data processing module, a data transmission module, and a fault early warning module.

[0046] The intelligent monitoring module is used to acquire monitoring data of the electric energy metering box in real time.

[0047] The data processing module is used to integrate machine learning algorithms to establish an electric energy metering box operation state model and a multi-class target detection model, and to perform deep analysis and prediction on the monitoring data based on the electric energy metering box operation state model and the multi-class target detection model to obtain processed data.

[0048] The data transmission module is used to upload the monitoring data and the processed data to a remote monitoring center in real time.

[0049] The fault early warning module is used to set an early warning threshold value according to a data processing result, to trigger an early warning mechanism when the monitoring data exceeds the threshold value, to remind an operation and maintenance personnel, and to send early warning information to the remote monitoring center. The early warning information includes a fault type, a location of occurrence, and a severity, etc.

[0050] The data processing and transmission module analyzes and processes data collected by the sensor, and establishes an electric energy metering box operation state model through a machine learning algorithm. Based on the electric energy metering box operation state model, deep analysis and prediction of the aging degree of components, the contact condition of lines, and humidity-related hidden dangers are realized. At the same time, an adaptive data transmission technology is used to upload the processed data to a remote monitoring center in real time to realize remote monitoring.

[0051] The fault early warning module sets an early warning threshold value according to a data processing result, and the system immediately triggers an early warning mechanism when the monitoring data exceeds the threshold value, to remind an operation and maintenance personnel, and to send early warning information to the remote monitoring center, including a fault type, a location of occurrence, and a severity, etc., so as to take timely measures for processing.

[0052] Specifically, the intelligent monitoring module includes a temperature sensor, a current sensor, a humidity sensor, and an image sensor. The temperature sensor, the current sensor, and the humidity sensor respectively collect temperature, current, and humidity data of the electric energy metering box.

[0053] The image sensor collects image information of key areas of a metering box door body, a lock, and a shell.

[0054] Specifically, the temperature sensor is used to monitor the temperature of the key parts in the box in real time to prevent overheating from causing failure; the current sensor is used to monitor the current in real time to determine whether the line is overloaded or short-circuited; and the humidity sensor is used to monitor the humidity in the box to avoid corrosion of electrical components caused by high humidity.

[0055] The image sensor is used to monitor the shell of the metering box in real time, the high-definition camera is used to capture the appearance image of the box body, and the image recognition algorithm is used to analyze whether the shell has abnormal opening, damage, drilling and other traces. When the integrity of the shell is detected to be damaged, the system immediately determines that there is a potential electricity stealing risk, synchronously triggers the early warning mechanism, uploads the abnormal image and position information to the remote monitoring center, and facilitates the operation and maintenance personnel to quickly check and effectively curb the electricity stealing behavior.

[0056] Specifically, the temperature sensor is a PT1000 platinum resistance sensor installed on the surface of the metering chip, terminal, circuit breaker and other components prone to heat, with a sampling frequency of 1 Hz, a measurement range of -50℃ to 150℃, and an accuracy of ±0.5℃.

[0057] The current sensor is a closed-loop Hall current sensor connected in series in the line between the incoming line and the outgoing line, supporting a range of 0-100A, a sampling frequency of 5kHz, and an accuracy of ±0.2%.

[0058] The humidity sensor is a SHT30 digital temperature and humidity integrated sensor installed in the middle of the metering box, with a sampling frequency of 0.5Hz, a measurement range of 0-100%RH, and an accuracy of ±2%RH.

[0059] The image sensor is a 200 million pixel high-definition camera with the lens facing the metering box door and lock position, supporting infrared night vision function, frame rate of 25fps, and image resolution of 1920x1080.

[0060] Specifically, the machine learning algorithm uses a Stacking integrated learning framework, including a base learning layer, a meta-learning layer and a parameter optimization module.

[0061] The base learning layer includes random forest (RF), gradient boosting decision tree (GBDT) and light gradient boosting machine (LightGBM), which respectively extract features from temperature, current and humidity time series data, and output element aging index, line health degree and humidity risk value.

[0062] The meta-learning layer uses a linear regression model to fuse the output results of the base learning layer and the original monitoring data, establishes a multi-dimensional evaluation model, and the calculation formula is:

[0063] Y = β0 + β1X temp + β2X curr + β3X hum + ξ.

[0064] wherein X temp is a temperature feature vector, X curr is a current feature vector, X hum is a humidity feature vector, β0-β3 are regression coefficients, and ξ is a random error term.

[0065] The parameter optimization module optimizes the model parameters using the k-fold cross-validation method to avoid overfitting. The model training period is 24 hours / time, and the prediction weights are updated in real time.

[0066] Specifically, training the electric energy metering box operation state model includes:

[0067] Training phase:

[0068] Data set construction: integrate historical monitoring data and fault records (including component aging, line fault, moisture, etc. labels) to form a labeled sample set;

[0069] Base learning layer training:

[0070] Random forest RF: input temperature peak value and temperature rise rate, output component aging index (0-1) through 50 decision tree integration, the higher the index, the more serious the aging;

[0071] Gradient boosting decision tree GBDT: input current fluctuation amplitude and current-time integral feature, output line contact risk score (0-100) through 100 rounds of iterative training, score ≥80 is determined as high risk;

[0072] Lightweight gradient boosting machine LightGBM: input humidity cumulative change and moisture risk index, use histogram algorithm to accelerate training, output humidity risk level (1-5 levels);

[0073] Meta-learning layer fusion splices the output results of the base learning layer and the original features, trains a linear regression model through k-fold cross-validation, and the objective function is to minimize the mean square error (MSE) of the predicted value and the true label.

[0074] Specifically, based on the trained electric energy metering box operation state model for prediction, including:

[0075] Prediction phase:

[0076] Real-time feature input: pre-process and feature extract real-time monitoring data to generate feature vectors;

[0077] Multi-dimensional prediction: the base learning layer outputs intermediate results in parallel, and the meta-learning layer comprehensively calculates the final failure probability (P);

[0078] Trend prediction: Based on the data of the past 24 hours, the sliding window recursive method is used to predict the parameter changes in the next hour. When the predicted value exceeds the preset threshold, an early warning is triggered.

[0079] The monitoring data is preprocessed and feature extracted, including:

[0080] The preprocessing unit includes:

[0081] Data normalization: The preprocessed data is mapped to the [0, 1] interval using the min-max normalization method, and the formula is:

[0082]

[0083] where s is the original data, s max and s min are the maximum and minimum values of the features, respectively, to eliminate the influence of dimensional differences on model training;

[0084] The time series feature extraction unit extracts the following key features for temperature, current, and humidity time series data:

[0085] Time domain features: mean, variance, peak, and valley within a sliding window (window size 10 min);

[0086] Trend features: temperature rise rate (℃ / min), current fluctuation amplitude (%), and humidity cumulative change amount (%RH).

[0087] Specifically, the current-time integral feature is constructed and represented as:

[0088]

[0089] where Q represents the cumulative effect of line load, i(t) is the real-time current collected by the current sensor, and [t0, t1] is a 10-minute sliding window. The trapezoidal method is used for numerical integration;

[0090] The temperature-current coupling feature is extracted, and the Pearson correlation coefficient of temperature rise rate and current fluctuation amplitude is calculated to quantify the trend of their coordinated changes, which is used for early identification of line contact failure;

[0091] Based on the humidity duration feature, the cumulative duration (T hum ) of humidity exceeding the threshold (≥75%RH) is calculated, and the damp risk index (I hum = T hum × temperature mean) is constructed combining the temperature feature.

[0092] Specifically, the acquired image information is preprocessed; the preprocessing includes removing noise through Gaussian filtering, enhancing the contrast between light and dark areas through histogram equalization, and automatically turning on infrared fill light for nighttime scenes to ensure that the image clarity meets the recognition requirements.

[0093] A multi-class target detection model is constructed using an anomaly feature recognition algorithm, and the model parameters are optimized by calling the k-fold cross-validation method.

[0094] Feature extraction is performed on the preprocessed image, and abnormal regions are detected in real time using a trained model, outputting the abnormality type and confidence level.

[0095] The multi-class object detection model uses an improved DeepLabv3+ semantic segmentation network. The training set contains 5000+ labeled images (covering four types of electricity theft-related anomalies, including "abnormal door opening", "lock damage", "box drilling", and "external wire connection"). The model parameters are optimized by calling k-fold cross-validation.

[0096] Real-time identification: Features are extracted from the preprocessed image, and abnormal regions are detected in real time using a trained model. The abnormality type and confidence score are output (≥0.85 is considered a valid abnormality).

[0097] Exception detection logic:

[0098] Constructing a multi-condition decision mechanism by combining time series analysis:

[0099] The first labeling is triggered when an abnormal feature is identified in a single instance with a confidence level ≥ 0.9.

[0100] The duration of anomalies is statistically analyzed using a sliding time window (5 seconds). If the same anomaly is detected in 3 consecutive frames (0.04 seconds apart) and the duration is ≥10 seconds, it is determined as a "valid risk of electricity theft".

[0101] Early warning response mechanism: Local early warning: Triggers a 110dB buzzer + red flashing warning light, which continues until maintenance personnel reset it on-site. At the same time, the internal backup power supply of the metering box (with a battery life of ≥4 hours) is activated to ensure monitoring continuity.

[0102] Remote uploads include: abnormal images (including timestamps and GPS coordinates); abnormal type (e.g., "lock damage"), occurrence time and duration; device number and its corresponding substation information.

[0103] The early warning information will be pushed to the remote monitoring center simultaneously.

[0104] Specifically, the multi-class object detection model adopts an improved DeepLabv3+ semantic segmentation network architecture, including:

[0105] Main network: ResNet50 network is used as the feature extraction backbone, including deep feature extraction, middle feature extraction layer and low feature extraction layer, containing 14 layers of structure (conv1_x to conv4_x), among which the conv1_x layer is a combination of 7x7 convolution and maximum pooling, and the conv2_x to conv4_x layers are residual block structures, which extract low-level detail features (lock edge), middle-level texture features (box surface integrity) and deep-level semantic features (abnormal area outline) respectively;

[0106] Attention mechanism: After the deep, middle and low feature extraction layers, attention modules (SimAM) are added respectively to calculate the energy function of the normalized features;

[0107]

[0108] wherein, represents the calculated neuron importance quantitative value, which is used to measure the importance of neurons in the feature extraction process, and the larger the value, the more important the corresponding neuron in the current feature extraction. t represents the feature value of the neuron being calculated. is the mean value estimate of the feature value, which reflects the average level of the group of feature values. is the variance estimate of the feature value, which is used to describe the dispersion degree of the feature value, and the larger the variance, the more dispersed the feature value distribution. λ is a hyperparameter, usually a constant, which is used to adjust the calculation result of the formula, and can be set and adjusted according to experimental results, etc.

[0109] By calculating the energy function of the normalized features, the importance of the neurons is quantified, the feature weight of the abnormal area is enhanced, and the background noise is suppressed;

[0110] Multi-scale feature fusion: After the deep features are processed by the atrous spatial convolution pyramid (ASPP), they are upsampled by 4 times, the middle features are upsampled by 2 times after 1x1 convolution, and the low features are upsampled by 4 times after being refined by 3x3 depth separable convolution after being spliced. The final segmentation result is output. Realize the fine-grained and accurate identification of the key areas of the metering box door, lock and shell;

[0111] Model training: Adam optimizer (initial learning rate 0.007) and cross-entropy loss function are used, based on the expanded training set (containing rotation, scaling and other enhanced samples) for iterative training, and the hyperparameters are adjusted through k-fold cross-validation to ensure the robustness of the model in different light and shielding scenes.

[0112] The various embodiments described in this specification are implemented in a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the apparatus disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0113] The above description of disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An energy metering box with intelligent monitoring and fault early warning functions, characterized in that, include: The enclosure and the intelligent monitoring and fault early warning device are integrated on the enclosure; The intelligent monitoring and fault early warning device includes an intelligent monitoring module, a data processing module, a data transmission module, and a fault early warning module; The intelligent monitoring module is used to acquire monitoring data from the power metering box in real time; The data processing module is used to integrate machine learning algorithms to establish an operating status model of the power metering box and a multi-category target detection model. Based on the operating status model of the power metering box and the multi-category target detection model, the monitoring data is analyzed and predicted in depth to obtain the processed data. The data transmission module is used to upload the monitoring data and the processed data to the remote monitoring center in real time. The fault early warning module is used to set early warning thresholds based on data processing results. When the monitored data exceeds the threshold, the early warning mechanism is triggered to remind maintenance personnel and send early warning information to the remote monitoring center.

2. The power metering box with intelligent monitoring and fault early warning functions according to claim 1, characterized in that, The intelligent monitoring module includes a temperature sensor, a current sensor, a humidity sensor, and an image sensor. The temperature sensor, current sensor, and humidity sensor respectively collect temperature, current, and humidity data of the power metering box. The image sensor acquires image information of key areas of the metering box door, lock, and outer shell.

3. The power metering box with intelligent monitoring and fault early warning functions according to claim 2, characterized in that, The machine learning algorithm adopts the Stacking ensemble learning framework, which includes a base learning layer, a meta-learning layer, and a parameter optimization module; The base learning layer includes random forest, gradient boosting decision tree and lightweight gradient boosting machine, which extract features from time series data of temperature, current and humidity, and output component aging index, line health and humidity risk value. The meta-learning layer uses a linear regression model to fuse the output of the base learning layer with the original monitoring data, establishing a multi-dimensional evaluation model. The calculation formula is as follows: Y=β0+β1X temp +β2X curr +β3X hum +ξ; Among them, X temp X is the temperature feature vector. curr Let X be the current eigenvector. hum Let β0-β3 be the humidity feature vector, β0-β3 be the regression coefficients, and ξ be the random error term; The parameter optimization module uses k-fold cross-validation to optimize the model parameters.

4. The power metering box with intelligent monitoring and fault early warning functions according to claim 3, characterized in that, Training the operating status model of the electricity metering box includes: Integrate historical monitoring data and fault records to form a tagged sample set; Using peak temperature and rate of temperature rise as inputs, the component aging index is output through decision tree integration. Using current fluctuation amplitude and current-time integral characteristics as inputs, the line contact risk score is iteratively trained and output; Using the cumulative change in humidity and the moisture risk index as inputs, a histogram algorithm is used for training to output the humidity risk level. Meta-learning layer fusion concatenates the output of the base learning layer with the original features, and trains a linear regression model through k-fold cross-validation. The objective function is to minimize the mean squared error between the predicted value and the true label.

5. An energy metering box with intelligent monitoring and fault early warning functions according to claim 4, characterized in that, Predictions are made based on the trained operational status model of the electricity metering box, including: Preprocess and extract features from real-time monitoring data to generate feature vectors; The base learning layer outputs intermediate results in parallel, and the meta-learning layer calculates the final failure probability. The sliding window recursive method is used to predict future parameter changes, and an early warning is triggered when the predicted value exceeds a preset threshold.

6. An energy metering box with intelligent monitoring and fault early warning functions according to claim 2, characterized in that, Preprocess the acquired image information; A multi-class target detection model is constructed using an anomaly feature recognition algorithm, and the model parameters are optimized by calling the k-fold cross-validation method. Feature extraction is performed on the preprocessed image, and abnormal regions are detected in real time using a trained model, outputting the abnormality type and confidence level.

7. An energy metering box with intelligent monitoring and fault early warning functions according to claim 6, characterized in that, The multi-class object detection model employs an improved DeepLabv3+ semantic segmentation network architecture, including: Backbone network: The ResNet50 network is used as the feature extraction backbone, including deep feature extraction, mid-level feature extraction layer and low-level feature extraction layer, which extract low-level detail features, mid-level texture features and deep semantic features respectively; Attention mechanism: Attention modules are added after the deep, mid-level and low-level feature extraction layers respectively, and the energy function after feature normalization is calculated. Multi-scale feature fusion: Deep features are upsampled by 4 times after being processed by dilated spatial convolution pooling pyramid, and mid-level features are upsampled by 2 times after being processed by 1×1 convolution. After being concatenated with low-level features, they are refined by 3×3 depth-separable convolution to output the segmentation result.

Images