Integrated distribution network low-voltage cable branch box state monitoring system

Abstract

The application provides a state monitoring system for integrated network low-voltage cable branch box, relates to the technical field of power equipment monitoring, and comprises a temperature sensor, an image acquisition unit and an edge computing unit. The system collects temperature data at a cable joint of an outgoing line of the branch box and image data inside the branch box in real time, and marks a collection time stamp; calculates a temperature change rate and outputs a first type of abnormal signal and a confidence degree thereof; performs multi-modal recognition on the image and outputs a second type of abnormal signal and a confidence degree thereof; pairs the two types of abnormal signals closest in time, calculates a comprehensive risk measure through Bayesian fusion, filters a high evidence group in a sliding window and performs smooth statistics; and finally, adopts the Laplace criterion to perform double outlier detection on the number of the high evidence group and the average comprehensive risk measure, and judges that there is a risk when both indexes are outliers. The application improves the accuracy and reliability of the state monitoring of the branch box through the cross-modal time sequence correlation and double outlier judgment mechanism.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, specifically to an integrated distribution network low-voltage cable branch box status monitoring system. Background Technology

[0002] Currently, the status monitoring of low-voltage cable branch boxes in distribution networks mainly relies on threshold judgments of single physical quantities, such as monitoring cable joint temperatures through temperature sensors or identifying anomalies inside the box through image acquisition. However, single-mode monitoring methods have inherent drawbacks: temperature changes are lag-dependent, often only showing a significant temperature rise after the fault has developed to a certain extent; while image recognition, although capable of detecting instantaneous anomalies such as arcs and smoke, is easily affected by factors such as lighting and obstruction, leading to false alarms. Existing technologies, such as the patent with publication number CN109375061A, "An Integrated Status Monitoring System for Low-Voltage Cable Branch Boxes in Distribution Networks," disclose an integrated status monitoring system for low-voltage cable branch boxes in distribution networks. This system collects electrical parameters and environmental data through current transformers, temperature sensors, and temperature and humidity sensors, but it still operates on a multi-sensor independent acquisition and judgment model, lacking a mechanism for temporal correlation analysis and fusion judgment between different modal data.

[0003] When early faults such as poor contact or insulation aging occur inside the branch box, temperature anomalies and image anomalies often show a correlation in time and trend—for example, localized heating may be accompanied by slight smoke, but a single mode may be missed due to insufficient confidence. Existing monitoring systems cannot utilize this cross-modal temporal consistency to improve the reliability of judgments, resulting in high false alarm and false negative rates, making it difficult to meet the accuracy requirements of distribution network condition monitoring.

[0004] Therefore, how to perform cross-modal temporal correlation between temperature data and image data, and improve the accuracy and robustness of anomaly identification through multi-level fusion and statistical decision mechanisms, is a technical problem that urgently needs to be solved in this field. The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated low-voltage cable branch box status monitoring system for power distribution networks, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The integrated distribution network low-voltage cable branch box status monitoring system includes: The data acquisition module is used to acquire temperature data at the outgoing cable joints of each branch and image data inside the branch box in real time, and to mark the acquisition timestamp. The confidence assessment module is used to calculate the temperature change rate of each branch. If the temperature change rate of any branch exceeds the first threshold, the module outputs the first type of abnormal signal of that branch and its confidence level, and records the collection timestamp. The multimodal recognition module is used to perform multimodal recognition on image data, determine in real time whether there is a preset risk event, and if so, generate a second type of abnormal signal and its confidence level, and record the collection timestamp. The signal processing module is used to select another type of abnormal signal with the smallest time difference from the occurrence of the abnormal signal when an abnormal signal occurs, and pair them to form an abnormal signal group. The posterior risk probability between the confidence levels of the two types of abnormal signals in the abnormal signal group is used as the comprehensive risk measure of the abnormal signal group. The cross-modal time difference is calculated based on the acquisition timestamps of the two abnormal signals, and the cross-modal temporal consistency is calculated based on the temporal change trend of the confidence levels of the two types of abnormal signals. The risk measurement module is used to set a sliding window. Within the sliding window corresponding to each monitoring time, the number of abnormal signal groups is counted, and abnormal signal groups with cross-modal time difference less than the time threshold and cross-modal temporal consistency greater than the time threshold are selected as high evidence groups. The number of high evidence groups is counted, and their comprehensive risk measurement is calculated as the average comprehensive risk measurement. The risk assessment module is used to obtain the number of high evidence groups and the average comprehensive risk measure within the nearest several sliding windows at each monitoring time, forming a quantity sequence and a measure sequence. If the number of high evidence groups in the current sliding window is zero, the module outputs that there is no risk at that time. Otherwise, the module performs outlier detection on the quantity sequence and the measure sequence. If both the number of high evidence groups and the average comprehensive risk measure in the current sliding window are detected as outliers, the module determines that there is a risk at that time.

[0007] Furthermore, when identifying image data, the preset risk events include: arc light, smoke, and fire.

[0008] Furthermore, the specific steps for performing multimodal recognition on image data are as follows: Pre-store the reference image features corresponding to each of the arc light, smoke, and firelight; Real-time acquisition of image data inside the branch box, and extraction of abnormal pixel areas in the image data; The features of the abnormal pixel region are compared with the features of each reference image to calculate the degree of overlap. The type of risk event is determined based on the reference image with the highest degree of overlap, and this degree of overlap is used as the discrimination score for the corresponding risk event.

[0009] Furthermore, the specific steps for calculating the confidence levels of the first and second types of anomalous signals are as follows: Calculate the ratio of the temperature change rate to the first threshold. The confidence level of the first type of abnormal signal is 1 minus the reciprocal of the ratio. Obtain the discrimination scores for each preset risk event during the image recognition process; Based on the preset risk level of each risk event, a weight is assigned to each risk event. The higher the risk level, the greater the weight assigned. The weighted sum of the scores for each risk event and their corresponding weights is used to obtain the weighted fusion score. If multiple risk events have the same weighted score during the weighted summation process, the weighted score corresponding to the highest priority among them will be selected as the weighted fusion score according to the preset risk event priority. Multiple score intervals are preset, each interval corresponds to a confidence value, and the confidence value corresponding to the interval into which the weighted fused score falls is used as the second confidence value.

[0010] Furthermore, when an abnormal signal occurs, another type of abnormal signal with the smallest time difference from the occurrence of the abnormal signal is selected for pairing. The specific steps are as follows: Based on the current timestamp of the abnormal signal acquisition, a preset time window is extended forward and backward, and all unpaired abnormal signals of the other type are searched within the time window. If at least one other type of abnormal signal is found, calculate the absolute value of the difference between the current abnormal signal and the timestamp of each found abnormal signal, and select the one with the smallest absolute value as the pairing object. If multiple abnormal signals have the same absolute value of timestamp difference with the current abnormal signal and all of them are the smallest, then the one with the higher confidence level is selected as the pairing object. After a successful pairing, the current abnormal signal and the paired object are marked as paired, forming a group of abnormal signals, and the pairing timestamp of the group is recorded. If no other type of abnormal signal is found within the preset time window, the current abnormal signal will not be paired and will wait for a new abnormal signal to appear. If the pairing is still unsuccessful after the preset timeout period, it will be marked as an independent abnormal signal and will not participate in the subsequent sliding window statistics.

[0011] Furthermore, based on the anomalous signal group paired with each anomalous signal, the relevant comprehensive risk metric, cross-modal time difference, and cross-modal temporal consistency are calculated. The specific steps are as follows: The comprehensive risk measurement is obtained through the following methods: The confidence levels of the first and second types of abnormal signals in the abnormal signal set are used as observational evidence; Based on the preset prior probability and the preset conditional probability, the posterior probability is calculated using Bayes' theorem. The posterior probability is normalized and mapped to the interval between 0 and 1, serving as a comprehensive risk measure. The cross-modal time difference is obtained through the following method: Obtain the acquisition timestamps of the first type of abnormal signal and the second type of abnormal signal in the abnormal signal group, and calculate the absolute value of the difference between the two acquisition timestamps as the cross-modal time difference; Cross-modal timing consistency is achieved through the following methods: Obtain the confidence level of the first type in the current abnormal signal group, and the confidence level of the first type in the N adjacent abnormal signal groups in the current abnormal signal group, to form the first confidence level sequence; Obtain the confidence level of the second type in the current abnormal signal group, and the confidence level of the second type in the N adjacent abnormal signal groups before this abnormal signal group, to form the second confidence level sequence; Calculate the correlation coefficient between the first confidence sequence and the second confidence sequence as cross-modal time series consistency.

[0012] Furthermore, the comprehensive risk measure for the number of high evidence groups is calculated using the following steps: Let the arithmetic mean of the comprehensive risk measure of the high-evidence group within the sliding window corresponding to the current monitoring time be taken as the current observation value; Let the average comprehensive risk measure at the previous monitoring time be the historical value; A constant between 0 and 1 is preset as the smoothing coefficient. The average comprehensive risk measure at the current monitoring time is calculated as: smoothing coefficient × current observation value + (1 - smoothing coefficient) × historical value.

[0013] Furthermore, the number of high-evidence groups and the average comprehensive risk measure within several sliding windows are used to construct a quantity sequence and a measure sequence, and outlier detection is performed on the quantity sequence and measure sequence. The specific steps are as follows: Take the current monitoring window and the N previous sliding windows, arrange the number of high evidence groups in each sliding window in chronological order to form a quantity sequence, calculate the mean and standard deviation of the quantity sequence, and if the number of high evidence groups in the current sliding window exceeds the sum of the mean and K times the standard deviation, it is determined to be a quantity outlier. Take the same sliding window at the current monitoring time and the N sliding windows before it, arrange the average comprehensive risk measures in all sliding windows in chronological order to form a measurement sequence, calculate the mean and standard deviation of the average comprehensive risk measures of the measurement sequence, and if the average comprehensive risk measure of the current sliding window exceeds the sum of the mean and K times the standard deviation, it is determined to be a measurement outlier. When both quantitative and metric outliers are true, the current sliding window is deemed to be at risk. Where N and K are both preset positive numbers.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention is based on; The present invention also applies. Attached Figure Description

[0015] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a schematic diagram of the overall method flow of the present invention; Figure 3 This is a graph showing the relationship between the rate of temperature change and the first type of confidence level in this invention. Figure 4 This is a graph showing the relationship between the weighted fusion score and the second type of confidence level in this invention. Figure 5 This is a graph showing the change in the comprehensive risk measurement of this invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0017] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0018] Example: Please see Figure 1-5 The present invention provides a technical solution: The integrated low-voltage cable branch box status monitoring system for power distribution networks includes the following steps: The data acquisition module is used to acquire temperature data at the outgoing cable joints of each branch and image data inside the branch box in real time, and to mark the acquisition timestamp.

[0019] In this embodiment, the integrated distribution network low-voltage cable branch box status monitoring system includes: multiple temperature sensors, at least one image acquisition unit, and an embedded edge computing unit. The temperature sensors are surface-mount thermistor temperature sensors, specifically NTC thermistors, which are fixed to a flexible printed circuit board using a surface mount process. The temperature sensors are tightly attached to the metal surface of each branch cable connector using thermally conductive silicone grease and secured with high-temperature resistant cable ties or clamps to ensure accurate temperature measurement. The image acquisition unit uses a wide-angle infrared night vision camera, model OV5640 module, with a field of view of 120 degrees. This camera is installed in the center of the top cover inside the branch box, and its field of view covers all branch cable connector areas and the interior space of the branch box. The embedded edge computing unit uses an STM32F407 microcontroller, which integrates a real-time clock module. This microcontroller is electrically connected to each temperature sensor via an I²C bus and to the image acquisition unit via a DCMI interface.

[0020] All operations of the system, including data acquisition, timestamp marking, temperature change rate calculation, anomaly detection, confidence calculation, image multimodal recognition, anomaly signal pairing, sliding window statistics, and Raida criterion risk judgment, are completed within the embedded edge computing unit.

[0021] The monitoring system continuously collects temperature and image data at a fixed sampling period. Specifically, the temperature data sampling period is set to 2 seconds, meaning that every 2 seconds, the edge computing unit sequentially reads the current resistance value of each temperature sensor via the I²C bus and converts it into a Celsius temperature value based on the resistance-temperature characteristic curve of the thermistor. The image data sampling period is set to 2 frames per minute, meaning that every 30 seconds, the edge computing unit triggers the image acquisition unit to capture an image of the inside of the branch box via the DCMI interface. Temperature acquisition and image acquisition are asynchronous in time.

[0022] Suppose the branch box has a total of M branch outgoing cables, M 1. For each branch cable, after the temperature is collected at each sampling time, the temperature value is bound and stored with the branch number.

[0023] In this embodiment, the image data sampling period is set to 30 seconds. Every 30 seconds, the edge computing unit triggers the image acquisition unit to capture an image of the inside of the branch box via the DCMI interface. The image format is JPEG, and the resolution is 800×600 pixels.

[0024] Each time temperature and image data are collected, the edge computing unit reads the current time from the internal real-time clock module as the timestamp of the data collection. Temperature data is stored in a data set bound to its timestamp, and image data is stored in a data set bound to its timestamp.

[0025] The confidence assessment module is used to calculate the temperature change rate of each branch. If the temperature change rate of any branch exceeds the first threshold, the module outputs the first type of abnormal signal and its confidence level for that branch, and records the collection timestamp.

[0026] For any branch Let its current sampling timestamp be... The collected temperature is The temperature last collected on this branch is recorded as follows: Its bound collection timestamp is The rate of temperature change of this branch at the current moment Defined as: That is, the absolute value of temperature change per unit time, expressed in degrees Celsius per second. This refers to the time interval between two temperature measurements. In this embodiment, the interval is equal to the temperature sampling period of 2 seconds.

[0027] For each newly acquired temperature data point, the edge computing unit calculates the rate of temperature change for that branch at the current acquisition timestamp using the formula described above. When previous temperature data is unavailable, no calculation is performed; instead, the current temperature data is cached for the next acquisition.

[0028] A first threshold is set to determine whether temperature changes are abnormal. In this embodiment, the first threshold is set to 0.5℃ / s. The basis for this threshold value is that under normal operating conditions, the temperature change rate of a low-voltage cable joint is usually less than 0.1℃ / s; when poor contact, overload, or short circuit precursors occur, the temperature change rate will increase significantly. Using 0.5℃ / s as the threshold can effectively filter out normal fluctuations and retain abnormal signs. The first threshold is pre-set in the storage unit of the edge computing unit.

[0029] The edge computing unit calculates the temperature change rate of each branch at the current acquisition time stamp and compares the temperature change rate with a first threshold. If the temperature change rate is greater than the first threshold, it determines that the branch has a temperature anomaly at the current acquisition time stamp and outputs a first-type anomaly signal for that branch. The output first-type anomaly signal contains at least the following three elements: Anomaly type identifier: indicates that the anomaly originates from an excessive rate of temperature change; The corresponding branch number: that is, the branch that meets the conditions; Data collection timestamp: The data collection timestamp that is associated with the current temperature data.

[0030] When judging and outputting the first type of abnormal signal, the confidence level of the first type of abnormal signal is calculated as follows: First, calculate the ratio of the current temperature change rate of this branch to the first threshold. ,in The first threshold is used as the output condition for an abnormal signal, since the rate of temperature change must exceed the first threshold. Next, calculate the reciprocal of the ratio. Finally, the confidence level of the first type of anomalous signal is defined as: Confidence level The value range is greater than 0 and less than 1. When the rate of temperature change just exceeds the threshold, the confidence level is close to 0; when the rate of temperature change is much greater than the threshold, the confidence level is close to 1. The higher the confidence level, the more reliable the abnormal signal.

[0031] Table 1 shows the confidence mapping relationship between the two types of abnormal signals. Based on Table 1 and... Figure 5 It can be seen that the confidence level of the first type of anomalous signal increases monotonically with the increase of the temperature change rate. When the temperature change rate just exceeds the first threshold of 0.5 degrees Celsius per second, the confidence level is close to 0; when the temperature change rate rises to 1.0 degrees Celsius per second, the confidence level reaches 0.5; when the temperature change rate further rises to 5.0 degrees Celsius per second, the confidence level reaches 0.9. Thereafter, even if the temperature change rate continues to increase, the growth rate of the confidence level gradually slows down, and the curve tends to flatten. This indicates that when the present invention converts the physical quantity of temperature change rate into confidence level, it achieves both refined quantification of the anomaly intensity and reflects a reasonable design of diminishing marginal returns—when the anomaly is already very obvious, the confidence level approaches saturation, and there is no need to infinitely distinguish the severity. According to Table 1 and Figure 5 It can be seen that there is a step-like mapping relationship between the confidence level and the weighted fusion score of the second type of abnormal signal. When the weighted fusion score is between 0 and 0.3, the confidence level is 0.2; when the score is between 0.3 and 0.5, the confidence level is 0.5; when the score is between 0.5 and 0.7, the confidence level is 0.7; and when the score exceeds 0.7, the confidence level is capped at 0.9. The advantage of this step-like mapping is that when the weighted fusion score fluctuates within the same range, the confidence level remains stable, avoiding frequent jumps in confidence level due to small score fluctuations; only when the score crosses the boundary of the range does the confidence level jump, representing a qualitative improvement in the anomaly level. At the same time, the design of capping the confidence level at 0.9 reflects the objective fact that image recognition technology has a reasonable recognition upper limit.

[0032] Table 1: Confidence Mapping Relationship between Two Types of Abnormal Signals When the first type of abnormal signal is output, the edge computing unit combines the following information into an abnormal record and stores it in the cache: Abnormal signal record: {collection timestamp, abnormal type, branch number, confidence level}. If the temperature change rate of all branches does not exceed the first threshold, no abnormal signal is output, the system returns to the previous step and continues to collect the next round of data.

[0033] The multimodal recognition module is used to perform multimodal recognition on image data, determine in real time whether there are preset risk events, and if so, generate a second type of abnormal signal and its confidence level, and record the collection timestamp.

[0034] In this embodiment, the preset risk events include: arcing, smoke, and fire. These three events are all abnormal conditions that may occur inside the low-voltage cable branch box. Arcing is usually caused by poor electrical contact or short circuit, while smoke and fire represent more serious overheating or combustion conditions.

[0035] Before the system runs, the edge computing unit's storage unit pre-stores reference image features corresponding to each of the arc light, smoke, and firelight. Specifically: The reference image features of the arc light include visual feature vectors such as high-brightness areas, white or bluish-white light spots, and irregular radial edges; Reference image features of smoke: including visual feature vectors such as semi-transparent gray-white areas, blurred edges, and changes in diffusion over time; Reference image features of the firelight include visual feature vectors such as orange-red areas, edge jerking, and accompanying brightness and color fluctuations.

[0036] The features of the above reference images are all stored in the form of feature vectors for subsequent comparison.

[0037] The edge computing unit acquires an image of the inside of the branch box every 30 seconds in JPEG format with a resolution of 800×600 pixels.

[0038] After each image data acquisition, the edge computing unit immediately preprocesses the image, including: converting the color image to a grayscale image or retaining the RGB three channels, performing median filtering to remove noise, and enhancing image contrast through histogram equalization to facilitate the subsequent extraction of abnormal pixel regions.

[0039] The edge computing unit extracts abnormal pixel regions from the preprocessed image. The extraction method is as follows: The current image and the background image (a reference image pre-acquired under normal operating conditions inside the branch box) are differentially analyzed pixel-by-pixel, calculating the difference in grayscale or color value for each pixel. Pixels with differences exceeding a preset difference threshold (30 in this embodiment, grayscale range 0-255) are marked as abnormal pixels. Connectivity analysis is performed on all adjacent abnormal pixels, and interconnected abnormal pixels constitute one or more abnormal pixel regions. Each abnormal pixel region serves as a candidate risk event region for subsequent identification.

[0040] For each extracted anomalous pixel region, the edge computing unit extracts features of that region, including: color histogram, mean brightness, texture features, and edge gradient features. The features of this anomalous pixel region are then compared with pre-stored reference image features of arc light, smoke, and firelight to calculate the overlap degree. The overlap degree is calculated by performing a cosine similarity calculation between the feature vector of the anomalous pixel region and the feature vector of the reference image; the resulting similarity value is the overlap degree, ranging from 0 to 1, with a higher value indicating greater similarity.

[0041] The risk event type with the highest overlap is taken as the determination result of the abnormal pixel region, and the highest overlap is used as the discrimination score of the corresponding risk event. If the highest overlap value is lower than the preset effective threshold, which is set to 0.5 in this embodiment, the abnormal pixel region is considered not to belong to any preset risk event and is ignored.

[0042] If there are multiple abnormal pixel regions in the same image, the above identification is performed separately for each region, and the region with the highest discrimination score is taken as the final result of this image recognition.

[0043] If, after the above identification, the highest overlap of a certain abnormal pixel region is greater than or equal to a preset effective threshold, and is thus determined to be one of the following categories: arc light, smoke, or fire, then a second type of abnormal signal is generated. The generated second type of abnormal signal must contain at least the following three elements: Anomaly type identifier: indicates that the anomaly originates from image multimodal recognition; Risk event types: specifically "arc light", "smoke", or "fire". Acquisition timestamp: This is the acquisition timestamp associated with the image data, and it must be consistent with the image data timestamp.

[0044] For the acquired second type of anomalous signals, their confidence level is calculated using a weighted fusion method. The specific steps are as follows: For the results of this identification, the edge computing unit records the discrimination scores for each of the three risk events: arc light, smoke, and fire. The discrimination score for the event determined to have occurred is the overlap degree calculated in step 5; the discrimination score for the other events determined not to have occurred is 0.

[0045] After obtaining the discrimination scores for each preset risk event, weights are assigned according to the risk level. The higher the risk level, the greater the weight assigned. In this embodiment, the preset risk level and weight correspondence is as follows: The weighting is based on the following: flames represent combustion that has already occurred and are the most dangerous; smoke is a precursor or accompanying phenomenon of combustion and is the second most dangerous; arcs are usually a sign of electrical faults and, although they require attention, are less urgent than the former two.

[0046] The sum of the scores for each risk event, multiplied by their corresponding weights, is calculated using the following formula: in, This represents the weighted fusion score, a value greater than or equal to 0 and less than or equal to the sum of the weights, used for subsequent mapping to the confidence level of the second type of anomalous signal; This is the subscript index used to iterate through the three preset risk events. Represents arc light, Represents smoke, Represents firelight; Indicates subscript The corresponding risk events are weighted as follows: in this embodiment, the weight of arc light is 0.3, the weight of smoke is 0.5, and the weight of fire is 0.7.

[0047] If multiple risk events have the same weighted score during the weighted summation process, the weighted score corresponding to the highest priority among them is selected as the weighted fusion score according to the preset risk event priority. In this embodiment, the preset risk event priorities from high to low are: fire > smoke > arc. The basis for this setting is: fire represents open flame combustion, which poses the most direct and serious threat to the safety of the branch box and requires priority response; smoke is usually a precursor or accompanying phenomenon of combustion, and its danger level is second; arc often manifests as an electrical fault but may not immediately cause a fire, and its priority is relatively the lowest. This priority order conforms to the common judgment standard for the severity of anomalies in the field of power safety. In other embodiments or application scenarios, the priorities can be reordered according to the environment of the branch box (such as whether it is located in a densely populated area or whether flammable materials are stored) or the importance of power supply.

[0048] Multiple score intervals are preset, each interval corresponding to a confidence value. The confidence value corresponding to the interval into which the weighted fusion score calculated in the above steps falls is used as the confidence value of the second type of abnormal signal.

[0049] The setting is based on the following: the theoretical maximum value of the weighted fusion score is the sum of the weights, i.e., 0.3 + 0.5 + 0.7 = 1.5. However, in practical applications, since the discrimination scores for each risk event are between 0 and 1, and usually only one event has a non-zero score (generally only one type of anomaly appears in the image), the actual common range of the weighted fusion score is between 0 and 0.7. Setting the thresholds at 0.3, 0.5, and 0.7 as dividing points allows for the division of confidence into four levels: low (0.2), medium (0.5), relatively high (0.7), and high (0.9). The intervals are divided using a left-closed and right-open or double-closed method to ensure that each weighted fusion score uniquely corresponds to a confidence value, without ambiguity. In other embodiments, the number of score intervals, the values ​​of the dividing points, and the corresponding confidence values ​​can be adjusted according to the actual application requirements for sensitivity and false alarm rate.

[0050] When a second type of anomaly signal is generated, the edge computing unit combines the following information into an anomaly record and stores it in the cache: Anomaly Record: {Collection Timestamp, Anomaly Type, Risk Event Type, Confidence Level}. If no preset risk event is identified in the current image, i.e., the highest overlap of all abnormal pixel regions is less than the preset effective threshold of 0.5, no anomaly signal is output, and the system returns to step 1 to continue collecting the next round of image data. The preset effective threshold of 0.5 means that only when the overlap reaches 50% or more is the identification considered valid. This threshold avoids misjudging noise or irrelevant interference as risk events, reducing the false alarm rate. This effective threshold can be adjusted according to factors such as image quality, ambient lighting conditions, and camera performance. By setting an effective threshold, low-confidence identification results are filtered out, ensuring that only anomalies with high confidence trigger the second type of anomaly signal, thereby reducing invalid calculations and false triggers in subsequent steps.

[0051] The signal processing module is used to select another type of abnormal signal with the smallest time difference from the occurrence of the abnormal signal when an abnormal signal occurs, and pair them to form an abnormal signal group. The posterior risk probability between the confidence levels of the two types of abnormal signals in the abnormal signal group is used as the comprehensive risk measure of the abnormal signal group. The cross-modal time difference is calculated based on the acquisition timestamps of the two abnormal signals, and the cross-modal temporal consistency is calculated based on the temporal change trend of the confidence levels of the two types of abnormal signals.

[0052] When any abnormal signal is detected, that abnormal signal is used as the current abnormal signal to initiate the pairing process. If the current abnormal signal has already been marked as "paired", the process is skipped and will not be repeated.

[0053] Based on the current timestamp of the abnormal signal acquisition, a preset time window is extended both forward and backward. Within this time window, all unpaired abnormal signals of the other type are searched. In this embodiment, the preset time window length is set to 30 seconds, based on a 2-second sampling period for the temperature sensor and a 30-second sampling period for the image. Temperature anomalies typically appear within seconds to tens of seconds after a fault occurs, while image anomalies may appear at the moment of the fault or shortly thereafter. The 30-second time window can cover the time difference between most temperature and image anomalies, while avoiding incorrectly pairing two temporally unrelated anomalies. The length of the preset time window can be adjusted according to factors such as the physical environment of the branch box, the sensor sampling frequency, and the fault propagation speed. By setting a limited time window, it is ensured that temporally related cross-modal anomalies can be successfully paired, while avoiding incorrect association of anomalies that are too far apart in time and have no causal relationship, thus improving the accuracy of the pairing results and the reliability of subsequent risk measurement.

[0054] Within a preset time window, search for all unpaired abnormal signals of the other type. If a match is found, calculate the absolute value of the difference between the current abnormal signal and the timestamp of each of the searched abnormal signals. Select the abnormal signal with the smallest absolute difference between the timestamps as the pairing object. The basis for this is that the smaller the absolute value of the time difference, the closer the two abnormal signals are in time, and the greater the probability that they originate from the same fault event. Prioritizing the pairing object with the smallest time difference can maximize the reliability of cross-modal correlation.

[0055] If multiple abnormal signals have the same absolute value of timestamp difference with the current abnormal signal and all of them are the smallest, then the one with the higher confidence level is selected as the pairing object.

[0056] If no other type of abnormal signal is found to be matched within the preset time window, the current abnormal signal will not be matched and will be placed in a waiting queue to await subsequent new abnormal signals. In this embodiment, the preset timeout is set to 2 minutes. If the current abnormal signal enters the waiting queue and still fails to match within 2 minutes, it will be marked as an "independent abnormal signal" and will no longer participate in subsequent sliding window statistics.

[0057] After successful pairing, the current abnormal signal and its paired object are marked as "paired," forming an abnormal signal group, and the pairing timestamp of this group is recorded. The pairing timestamp is taken as the collection timestamp of the later of the two abnormal signals, which serves as the time reference for this abnormal signal group.

[0058] For each successfully paired abnormal signal group, its comprehensive risk metric, cross-modal time difference, and cross-modal timing consistency are calculated. The specific steps are as follows: The calculation method for comprehensive risk measurement is as follows: The anomalous signal group contains two types of anomalous signals: a first type of anomalous signal (from the rate of temperature change) and its confidence level, and a second type of anomalous signal (from image recognition) and its confidence level. These two types of confidence levels are used as observational evidence. In this embodiment, the following parameters are preset: Prior probability P(H): Represents the prior probability that a branch box will fail, i.e., there is a real risk, in the absence of any observational evidence. In this embodiment, P(H) = 0.1, meaning that under the default condition, the probability of the branch box being in normal operation is 90%, and the probability of failure is 10%. This value is based on historical fault statistics of low-voltage branch boxes in the distribution network. Conditional probability P(E|H): represents the probability of observing evidence with the current confidence level under the condition that a fault actually occurs. In this embodiment, P(E|H) is calculated by adding the confidence value of the first type of anomalous signal to the confidence value of the second type of anomalous signal and then dividing by 2. This means that if a fault actually exists, the higher the confidence level of the two modes, the greater the probability of observing the evidence. Conditional probability P(E|¬H): This represents the probability of observing evidence with the current confidence level under the condition that no fault has occurred, i.e., the abnormal signal is a false alarm. In this embodiment, P(E|¬H) is calculated by subtracting the conditional probability P(E|H) from 1. This means that if no fault exists, the lower the confidence level of the two modalities, the greater the probability of observing the evidence. The above probability model reflects the basic logic that high confidence levels are more likely to be genuine faults, while low confidence levels are more likely to be false alarms. Using the arithmetic mean of confidence levels as a simplified model of conditional probability facilitates rapid calculation by edge computing units and is suitable for the computing capabilities of embedded devices. In other embodiments, the prior probability P(H) can be adjusted based on the actual historical failure rate of the branch box; the conditional probabilities P(E|H) and P(E|¬H) can employ more complex models, such as a probability density function based on a Gaussian distribution, or a weighted average. Specific probability parameters and functional forms are not essential features of this invention, as long as the monotonic relationship of "the higher the confidence level, the larger P(E|H) and the smaller P(E|¬H)" is satisfied.

[0059] The posterior probability P(H|E) is calculated using Bayes' theorem, which represents the probability that a fault actually exists given the observed evidence of current confidence level: P(H|E) = , where P(¬H)=1-P(H)=0.9.

[0060] The P(H|E) calculated by the Bayesian formula above is already in the range of 0 to 1, so no additional normalization is needed. This posterior probability value can be directly used as a comprehensive risk measure.

[0061] Table 2 shows a comparison of Bayesian fusion results at different intensities. As can be seen from Table 2, when both confidence levels are low (Scenario 1), the overall risk metric is only 0.023, and the system will not misclassify it as risky. When both confidence levels increase simultaneously (Scenarios 3 to 6), the overall risk metric gradually rises from 0.135 to 0.268. It is worth noting that even if one confidence level reaches saturation (e.g., image confidence caps at 0.9), the overall risk metric can still continue to increase through the improvement of the other confidence level (Scenarios 5 to 6), demonstrating the complementary advantages of cross-modal fusion.

[0062] Table 2: Comparison of Bayesian fusion results at different intensities The method for calculating the cross-modal time difference is as follows: The acquisition timestamps of the first type of abnormal signal and the second type of abnormal signal in the abnormal signal group are obtained. The absolute value of the difference between the two acquisition timestamps is calculated as the cross-modal time difference. In this embodiment, the unit of the cross-modal time difference is seconds, accurate to milliseconds.

[0063] The calculation method for cross-modal timing consistency is as follows: The confidence scores of the first type of abnormal signal in the current abnormal signal group and the confidence scores of the first type of abnormal signal in the N adjacent abnormal signal groups before this abnormal signal group are obtained to form the first confidence score sequence.

[0064] The confidence scores of the second type of abnormal signals in the current abnormal signal group and the confidence scores of the second type of abnormal signals in the N adjacent abnormal signal groups before this abnormal signal group are obtained to form the second confidence score sequence.

[0065] In this embodiment, N is set to 3, meaning the three successfully paired abnormal signal groups preceding the current abnormal signal group are taken. If the number of previous abnormal signal groups is less than N, the actual number is used, but the sequence length must be at least 2 to calculate the correlation coefficient. If there are fewer than two, the cross-modal temporal consistency is set to the default value of 0.5. The rationale for this setting is that N=3 is a balance point sufficient to observe the trend of confidence changes without over-reliance on historical data due to excessively long sequences. Taking the sequence preceding the current group (rather than including subsequent groups) ensures that the indicator can be calculated in real time at the current moment, without relying on future data.

[0066] The risk measurement module is used to set a sliding window. Within the sliding window corresponding to each monitoring time, the number of abnormal signal groups is counted, and abnormal signal groups with cross-modal time difference less than the time threshold and cross-modal temporal consistency greater than the time threshold are selected as high evidence groups. The number of high evidence groups is counted, and their comprehensive risk measurement is calculated as the average comprehensive risk measurement.

[0067] In this embodiment, the sliding window duration is set to 5 minutes, meaning the current moment is the end of the sliding window, and the window starts 5 minutes prior. This setting is based on the fact that a 5-minute window can capture a sufficient number of abnormal signal groups for statistical analysis (abnormalities occur less frequently under normal conditions; under fault conditions, abnormal signal groups may appear densely), while avoiding excessive smoothing of historical information due to an overly long window, which could lead to a delay in responding to current risks. The sliding window duration can be adjusted based on factors such as the importance of the line where the branch box is located, the speed of fault development, and system response requirements. Through the sliding window mechanism, the system can output a statistical result based on recent data at each monitoring moment, avoiding random fluctuations caused by a single abnormal signal group while maintaining good real-time response capabilities.

[0068] In this embodiment, each monitoring moment is synchronized with the image data acquisition moment, meaning that subsequent statistical and decision-making steps are performed every 30 seconds. Synchronizing the monitoring moments with the image sampling period ensures the timeliness of risk assessment while avoiding meaningless frequent calculations. Image anomalies are typically more sudden than temperature anomalies, making the image sampling period a reasonable decision-making period. The interval between monitoring moments can be adjusted according to actual needs.

[0069] At each monitoring time, the edge computing unit performs the following statistical operations: Within the current sliding window interval, the number of all successfully paired anomalous signal groups is counted. From all anomalous signal groups, those that simultaneously meet the following two conditions are selected as high-evidence groups: Condition 1: Cross-modal time difference is less than a time threshold; Condition 2: Cross-modal temporal consistency is greater than a time threshold. In this embodiment, the time threshold is set to 5 seconds, meaning only anomalous signal groups with a cross-modal time difference of less than 5 seconds meet Condition 1. The time threshold is set to 0.6, meaning only anomalous signal groups with a cross-modal temporal consistency greater than 0.6 meet Condition 2. The settings are based on the following criteria: a time threshold of 5 seconds: the temperature sampling period is 2 seconds, and the image sampling period is 30 seconds. If the time difference between temperature anomalies and image anomalies is less than 5 seconds, it indicates that the two are highly synchronized in time and are very likely to originate from the same fault event. If the time difference is greater than 5 seconds (for example, the temperature anomaly occurs first, and the image anomaly appears 20 seconds later), the two may only be indirectly related or coincidental. Using 5 seconds as the threshold can filter out the pairings with the closest time correlation. A time series threshold of 0.6: a Pearson correlation coefficient of 0.6 or higher indicates a moderate to strong positive correlation. This threshold can filter out anomalous signal groups with good consistency in the trend of confidence changes, and exclude pairings with divergent trends or random fluctuations. Through dual-condition screening, high-quality anomalous signal groups with close time correlation and high trend consistency are retained, while interference from accidental or low-quality pairings is eliminated, making subsequent risk measurement more reliable.

[0070] Obtain the comprehensive risk measure of all high-evidence groups, calculate their arithmetic mean, and use it as the observation value for the current sliding window. The calculation formula is as follows: in, Indicates the number of high-evidence groups. is the subscript index, indicating the first... A high-evidence group Indicates the first A comprehensive risk measure for high-evidence groups. These are the observations for the current sliding window.

[0071] The average comprehensive risk measure at the current monitoring time is calculated according to the following formula: in, This is the average comprehensive risk measure at the current monitoring time, with a value ranging from 0 to 1. This represents the smoothing coefficient, which is set to 0.3 in this embodiment, and ranges from 0 to 1. This represents the observations in the current sliding window, i.e., the arithmetic mean of the composite risk measure for the high-evidence group. This represents the average comprehensive risk measure at the previous monitoring time. For the first monitoring time after system startup, since there is no historical value from the previous monitoring time, it will be... Initialize to 0.

[0072] Smoothing coefficient This means that the current observation has a 30% weight in the final result, while historical values ​​have a 70% weight. This setting makes the average composite risk measure respond more smoothly to changes in the current observation, preventing drastic jumps in the risk measure due to high-risk observations within a single window, thus avoiding frequent false alarms. At the same time, continuous high-risk observations will gradually increase the average composite risk measure, ensuring that the system can respond to persistent risks.

[0073] The choice of value is a trade-off. The larger the value, the more sensitive the system is to the current observation, the faster the response, but the greater the fluctuation. The smaller the value, the smoother the system and the higher its stability, but the slower its response to new changes. 0.3 is a balanced value, suitable for most distribution network monitoring scenarios.

[0074] At each monitoring moment, the edge computing unit stores the following information in a cache and records it in a sliding window: [monitoring moment, total number of anomalous signal groups within the window, number of high-evidence groups, observation value, average comprehensive risk measure]. If the total number of anomalous signal groups within the current sliding window is zero, then the number of high-evidence groups = 0, and the number of observations = 0. In this case, the average composite risk measure is calculated using the smoothing formula. This means that in the absence of abnormal signal groups, the average comprehensive risk metric will gradually decay towards 0, with the decay rate determined by the smoothing coefficient. control.

[0075] Table 3 compares the filtering effects under different smoothing coefficients. Figure 5 As can be seen, during the entire monitoring process, the observed value (blue solid line in the figure) fluctuated drastically as the fault developed, especially during the fault escalation phase (around the 10th monitoring time), when the observed value rapidly climbed to above 0.8, and then quickly fell back during the fault descent phase. In contrast, the average comprehensive risk measure after exponential smoothing (red solid line in the figure) was much smoother: during the fault escalation phase, the solid line steadily rose following the dashed line but lagged behind; during the fault descent phase, the solid line also declined smoothly, without jumps caused by drastic fluctuations in a single observed value. This demonstrates the technical effectiveness of the exponential smoothing method used in this invention, which retains the ability to follow the fault development trend while effectively suppressing accidental fluctuations caused by single anomalies, avoiding frequent false alarms due to drastic changes in observed values, and making the risk judgment results more stable and reliable.

[0076] Table 3: Comparison of filtering effects under different smoothing coefficients The risk assessment module is used to obtain the number of high evidence groups and the average comprehensive risk measure within the nearest several sliding windows at each monitoring time, forming a quantity sequence and a measure sequence. If the number of high evidence groups in the current sliding window is zero, the module outputs that there is no risk at that time. Otherwise, the module performs outlier detection on the quantity sequence and the measure sequence. If both the number of high evidence groups and the average comprehensive risk measure in the current sliding window are detected as outliers, the module determines that there is a risk at that time.

[0077] At each monitoring time, the sliding window corresponding to the current monitoring time and the N sliding windows before it are taken. The number of high evidence groups in each sliding window is arranged in chronological order to form a quantity sequence; the average comprehensive risk measure in each sliding window is arranged in chronological order to form a measure sequence.

[0078] In this embodiment, N is set to 9. That is, the current sliding window and the nine sliding windows preceding it, totaling 10 sliding windows, constitute a sample set. Let the sliding windows be numbered 1, 2, 3, ..., 10 in order of monitoring time from earliest to latest, where number 10 is the current sliding window. Then: Quantity sequence: denoted as ,in The number of high evidence groups within the first sliding window. The number of high evidence groups within the current sliding window; Metric sequence: denoted as ,in The average composite risk measure within the earliest first sliding window. This is the average overall risk measure within the current sliding window.

[0079] Setting the sliding window size to N=9 means using the current window and the nine preceding historical windows (a total of 10 samples) as the statistical baseline. Ten samples is one of the minimum sample sizes applicable to the Raida criterion, ensuring statistical significance of the mean and standard deviation while preventing sluggish response to changes in the current state due to an excessively large sample size. The sliding window duration is 5 minutes, and the 10 windows correspond to 50 minutes of historical data, covering a sufficiently long timeframe to establish a baseline for normal operation. By using a sample set with a fixed window size, the system can dynamically adapt to the current branch box's operating environment, automatically establishing a statistical baseline for normal operation without requiring manual setting of fixed thresholds, demonstrating good adaptability.

[0080] If the number of high-evidence groups in the current sliding window is 0, then the system directly outputs that there is no risk at this monitoring moment, and no further outlier detection is performed. Otherwise, outlier detection is performed on the quantity sequence and the metric sequence to determine whether the number of high-evidence groups in the current sliding window is an outlier value. Specifically: Calculate the mean and standard deviation of a quantity series: Let the quantity series be... Include In this embodiment, there are 10 samples. There are N+1 samples in total, and the quantity sequence is as follows. The formula for calculating the mean is: ,in Represents the arithmetic mean of a sequence of quantities. This indicates a total sample size of 10. The index is the subscript index, indicating the first position in the sequence. One sample, From 1 to , Indicates the first The number of high evidence groups within a sliding window.

[0081] Standard deviation of quantity series The calculation formula is: ,in, The sample standard deviation represents the number of sequences. Indicates the first The deviation of each sample from the mean This represents the square of the deviation; Determining outliers: In this embodiment, the preset multiplier is 3. If the number of high evidence groups in the current sliding window is... The following conditions must be met: Then determine This represents the outlier value. The default multiplier of 3 is the standard value for the Raida criterion. Under the assumption of a normal distribution, the value falls within the mean. The probability within the specified range is approximately 99.73%, and beyond that range... The probability of this is approximately 0.27%, which is a low-probability event and is therefore considered an outlier. Using the Raida criterion for outlier detection eliminates the need for a pre-set fixed threshold for the number of high-evidence groups. Instead, it dynamically adjusts the judgment criteria based on the statistical characteristics of historical data, adapting to the normal operating baseline of different branch bins and improving the robustness of the decision.

[0082] Used for quantity sequences The same method is used to perform outlier detection on the metric sequence to determine the current... Is it a measure of outliers?

[0083] A risk is considered to exist at the current monitoring time when both of the following conditions are met: Condition 1: The number of high evidence groups in the current sliding window It was identified as a quantity outlier; Condition 2: The average comprehensive risk measure of the current sliding window It was determined to be a measure of outliers.

[0084] Otherwise, it is determined that there is no risk at the current monitoring time.

[0085] This embodiment uses a "simultaneous occurrence of two outliers" decision logic to avoid false alarms from a single indicator. The following situations may occur: There are many high-evidence groups, but the overall risk measure of each group is very low (e.g., a large number of low-confidence pairings). In this case, the number is out of the ordinary but the measure is not, so it is not judged as risk. If a group has an extremely high overall risk measure, but other groups have very low risk measures, resulting in a high average but no outliers, then the measure is outlier but the number of outliers is not considered risk.

[0086] Risk is only output when both dimensions point to an anomaly, effectively reducing the false alarm rate.

[0087] The dual confirmation mechanism significantly improves the reliability of risk judgment, avoids false alarms caused by the accidental fluctuation of a single indicator, and retains the ability to detect when two indicators rise simultaneously when a real fault occurs.

[0088] If the system has not accumulated N+1 sliding windows after startup, the mean and standard deviation are calculated based on the actual number of existing windows. If the actual number of windows is less than 3, outlier detection is not performed, and the system is directly output as "no risk".

[0089] Table 4 shows the risk judgment process deduction table. As shown in Table 4, in this embodiment, the mean of the quantity sequence plus three standard deviations is used as the upper limit for judging quantity outliers, and the mean of the measurement sequence plus three standard deviations is used as the upper limit for judging measurement outliers. During the fault development stage (14:45-14:50), both the number of high-evidence groups and the average comprehensive risk measurement exceed their respective upper limits, satisfying the double outlier condition, and the system accurately outputs "risk exists." During the initial fault stage (14:40) and the fault recovery stage (14:55), only a single indicator exceeds the upper limit, and the system will not issue a false alarm. This demonstrates the technical advantages of the dual confirmation mechanism of this invention in improving the accuracy of risk identification and reducing the false alarm rate.

[0090] Table 4: Risk Judgment Process Deduction Table The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0091] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0092] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0093] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An integrated low-voltage cable branch box status monitoring system for power distribution networks, characterized in that, include: The data acquisition module is used to acquire temperature data at the outgoing cable joints of each branch and image data inside the branch box in real time, and to mark the acquisition timestamp. The confidence assessment module is used to calculate the temperature change rate of each branch. If the temperature change rate of any branch exceeds the first threshold, the module outputs the first type of abnormal signal of that branch and its confidence level, and records the collection timestamp. The multimodal recognition module is used to perform multimodal recognition on image data, determine in real time whether there is a preset risk event, and if so, generate a second type of abnormal signal and its confidence level, and record the collection timestamp. The signal processing module is used to select another type of abnormal signal with the smallest time difference from the occurrence of the abnormal signal when an abnormal signal occurs, and pair them to form an abnormal signal group. The posterior risk probability between the confidence levels of the two types of abnormal signals in the abnormal signal group is used as the comprehensive risk measure of the abnormal signal group. The cross-modal time difference is calculated based on the acquisition timestamps of the two abnormal signals, and the cross-modal temporal consistency is calculated based on the temporal change trend of the confidence levels of the two types of abnormal signals. The risk measurement module is used to set a sliding window. Within the sliding window corresponding to each monitoring time, the number of abnormal signal groups is counted, and abnormal signal groups with cross-modal time difference less than the time threshold and cross-modal temporal consistency greater than the time threshold are selected as high evidence groups. The number of high evidence groups is counted, and their comprehensive risk measurement is calculated as the average comprehensive risk measurement. The risk assessment module is used to obtain the number of high evidence groups and the average comprehensive risk measure within the nearest few sliding windows at each monitoring time, forming a quantity sequence and a measure sequence; if the number of high evidence groups within the current sliding window is zero, then the module outputs that there is no risk at that time. Otherwise, outlier detection is performed on the quantity sequence and the measure sequence. If both the number of high evidence groups and the average comprehensive risk measure in the current sliding window are detected as outliers, then it is determined that there is a risk at that moment.

2. The integrated distribution network low-voltage cable branch box status monitoring system according to claim 1, characterized in that: When identifying image data, the preset risk events include: arc light, smoke, and fire.

3. The integrated distribution network low-voltage cable branch box status monitoring system according to claim 2, characterized in that: The specific steps for performing multimodal recognition on image data are as follows: Pre-store the reference image features corresponding to each of the arc light, smoke, and firelight; Real-time acquisition of image data inside the branch box, and extraction of abnormal pixel areas in the image data; The features of the abnormal pixel region are compared with the features of each reference image to calculate the degree of overlap; The type of risk event is determined based on the reference image with the highest degree of overlap, and this degree of overlap is used as the discrimination score for the corresponding risk event.

4. The integrated distribution network low-voltage cable branch box status monitoring system according to claim 3, characterized in that: The specific steps for calculating the confidence levels of Type I and Type II anomalous signals are as follows: Calculate the ratio of the temperature change rate to the first threshold. The confidence level of the first type of abnormal signal is 1 minus the reciprocal of the ratio. Obtain the discrimination scores for each preset risk event during the image recognition process; Based on the preset risk level of each risk event, a weight is assigned to each risk event. The higher the risk level, the greater the weight assigned. The weighted sum of the scores for each risk event and their corresponding weights is used to obtain the weighted fusion score. If multiple risk events have the same weighted score during the weighted summation process, the weighted score corresponding to the highest priority among them will be selected as the weighted fusion score according to the preset risk event priority. Multiple score intervals are preset, each interval corresponds to a confidence value, and the confidence value corresponding to the interval into which the weighted fused score falls is used as the second confidence value.

5. The integrated distribution network low-voltage cable branch box status monitoring system according to claim 1, characterized in that: When an abnormal signal occurs, select another type of abnormal signal with the smallest time difference from the occurrence of the abnormal signal for pairing. The specific steps are as follows: Based on the current timestamp of the abnormal signal acquisition, a preset time window is extended forward and backward, and all unpaired abnormal signals of the other type are searched within the time window. If at least one other type of abnormal signal is found, calculate the absolute value of the difference between the current abnormal signal and the timestamp of each found abnormal signal, and select the one with the smallest absolute value as the pairing object. If multiple abnormal signals have the same absolute value of timestamp difference with the current abnormal signal and all of them are the smallest, then the one with the higher confidence level is selected as the pairing object. After a successful pairing, the current abnormal signal and the paired object are marked as paired, forming a group of abnormal signals, and the pairing timestamp of the group is recorded. If no other type of abnormal signal is found within the preset time window, the current abnormal signal will not be paired and will wait for a new abnormal signal to appear. If the pairing is still unsuccessful after the preset timeout period, it will be marked as an independent abnormal signal and will not participate in the subsequent sliding window statistics.

6. The integrated distribution network low-voltage cable branch box status monitoring system according to claim 5, characterized in that: Based on the anomalous signal group paired with each anomalous signal, the relevant comprehensive risk metric, cross-modal time difference, and cross-modal time series consistency are calculated. The specific steps are as follows: The comprehensive risk measurement is obtained through the following methods: The confidence levels of the first and second types of abnormal signals in the abnormal signal set are used as observational evidence; Based on the preset prior probability and the preset conditional probability, the posterior probability is calculated using Bayes' theorem. The posterior probability is normalized and mapped to the interval between 0 and 1, serving as a comprehensive risk measure. The cross-modal time difference is obtained through the following method: Obtain the acquisition timestamps of the first type of abnormal signal and the second type of abnormal signal in the abnormal signal group, and calculate the absolute value of the difference between the two acquisition timestamps as the cross-modal time difference; Cross-modal timing consistency is achieved through the following methods: Obtain the confidence level of the first type in the current abnormal signal group, and the confidence level of the first type in the N adjacent abnormal signal groups in the current abnormal signal group, to form the first confidence level sequence; Obtain the confidence level of the second type in the current abnormal signal group, and the confidence level of the second type in the N adjacent abnormal signal groups before this abnormal signal group, to form the second confidence level sequence; Calculate the correlation coefficient between the first confidence sequence and the second confidence sequence as cross-modal time series consistency.

7. The integrated distribution network low-voltage cable branch box status monitoring system according to claim 1, characterized in that: The specific steps for calculating the overall risk measure of a high number of evidence groups are as follows: Let the arithmetic mean of the comprehensive risk measure of the high-evidence group within the sliding window corresponding to the current monitoring time be taken as the current observation value; Let the average comprehensive risk measure at the previous monitoring time be the historical value; A constant between 0 and 1 is preset as the smoothing coefficient. The average comprehensive risk measure at the current monitoring time is calculated as: smoothing coefficient × current observation value + (1 - smoothing coefficient) × historical value.

8. The integrated distribution network low-voltage cable branch box status monitoring system according to claim 7, characterized in that: The number of high-evidence groups and the average comprehensive risk measure within several sliding windows are used to construct a quantity sequence and a measure sequence. Outlier detection is then performed on the quantity sequence and the measure sequence. The specific steps are as follows: Take the current monitoring window and the N previous sliding windows, arrange the number of high evidence groups in each sliding window in chronological order to form a quantity sequence, calculate the mean and standard deviation of the quantity sequence, and if the number of high evidence groups in the current sliding window exceeds the sum of the mean and K times the standard deviation, it is determined to be a quantity outlier. Take the same sliding window at the current monitoring time and the N sliding windows before it, arrange the average comprehensive risk measures in all sliding windows in chronological order to form a measurement sequence, calculate the mean and standard deviation of the average comprehensive risk measures of the measurement sequence, and if the average comprehensive risk measure of the current sliding window exceeds the sum of the mean and K times the standard deviation, it is determined to be a measurement outlier. When both quantitative and metric outliers are true, the current sliding window is deemed to be at risk. Where N and K are both preset positive numbers.

Images