Intelligent monitoring method and system for distribution cable branch box

Abstract

The invention discloses a distribution cable branch box state monitoring method and system. According to the method, multi-dimensional information such as temperature, partial discharge and gas concentration of a branch box is acquired by deploying a multi-source sensor array, and dynamic feature vectors are extracted from the multi-dimensional information; the method is characterized in that a state transition model based on a hidden Markov model is constructed, and equipment operation states are divided into five hidden states including normal, concerned, abnormal, early warning and faults. And analyzing the real-time feature sequence by using an observation probability model and a Viterbi decoding algorithm, identifying the most possible current implicit state of the equipment, and predicting the future state evolution probability of the equipment. And a dual early warning mechanism of an instantaneous threshold and a state probability threshold is combined, so that comprehensive coverage from early warning to sudden faults is realized. According to the invention, the problems of high false alarm rate and lack of early degradation identification capability of traditional single-threshold monitoring are effectively solved, and accurate identification and predictive maintenance of the state of the power distribution branch box are realized.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, specifically to an intelligent monitoring method and system for distribution cable branch boxes. Background Technology

[0002] As a critical node in the power distribution network for connecting and distributing electrical energy, the operational reliability of distribution cable branch boxes directly affects power supply quality and safety. Traditional monitoring methods mainly rely on regular manual inspections and simple temperature threshold alarms, which have the following obvious drawbacks: Single monitoring dimension: Most systems only monitor the joint temperature, which cannot fully reflect the multi-dimensional status of internal insulation, sealing, microenvironment, etc., resulting in a high risk of missed reports.

[0003] The alarm logic is simple: it uses a fixed threshold for comparison, which is easily affected by ambient temperature, instantaneous load surges, etc., leading to frequent false alarms and causing "alarm fatigue" for maintenance personnel.

[0004] Lack of condition assessment: It can only provide binary judgments such as "normal" or "overheating", and cannot identify intermediate states such as early deterioration and sub-health of equipment, thus missing the best opportunity for preventive maintenance.

[0005] Lack of predictive ability: Unable to predict future state evolution based on current data trends, resulting in passive and delayed operation and maintenance decisions.

[0006] In existing technologies, although there have been attempts to use multiple sensors, most of them remain at the level of data accumulation or simple logical "AND-OR" judgments, failing to build an intelligent identification model that can simulate the gradual degradation process of equipment from the perspective of system dynamic evolution. Summary of the Invention

[0007] This invention aims to address the problems of incomplete monitoring, high false alarm rate, lack of early warning and status prediction capabilities in existing technologies, and provides an intelligent monitoring method that can accurately identify the continuous state spectrum of branch boxes from healthy to faulty, and realize early warning and predictive maintenance.

[0008] To achieve the above objectives, this invention provides an intelligent monitoring method for distribution cable branch boxes based on multi-source features and a state transition model. The core of this method lies in treating the branch box as a dynamic system with internal implicit states, fusing multi-source sensor information, and using a probabilistic graphical model to decode and predict the state evolution process. The method mainly includes the following steps: A method for intelligent monitoring of power distribution cable branch boxes, characterized by comprising the following steps: S1: A heterogeneous sensor array is deployed at key electrical connection points and within the cavity space inside the branch box for synchronous or quasi-synchronous data acquisition. The sensor array includes at least a distributed temperature sensor, an ultra-high frequency partial discharge sensor, a gas microenvironment sensor, and a door status sensor. The distributed temperature sensor is used to acquire the temperature values ​​and their spatial distribution at cable joints and busbar connections. The ultra-high frequency partial discharge sensor is used to sense and quantify the electromagnetic wave signals generated by partial discharge of insulating materials. The gas microenvironment sensor is used to monitor at least the oxygen concentration, characteristic sulfide concentration, and relative humidity. The door status sensor is used to monitor the opening and closing status and sealing integrity of the box door. S2. Preprocess the raw time-series data collected in step S1. The preprocessing includes filtering, denoising, and data alignment. From the preprocessed data, extract a feature vector containing multi-dimensional physical meaning. This feature vector is composed of at least a temperature feature subset, a partial discharge feature subset, and a gas and comprehensive feature subset. The temperature feature subset includes: the highest temperature value in the current monitoring period, the maximum temperature gradient value between key measuring points, and the slope of the temperature change trend calculated based on the highest temperature sequence in the past preset time window. The partial discharge feature subset includes: the average discharge intensity, the number of discharge pulses per unit time, and the correlation distribution characteristics between the discharge pulse and the power frequency voltage phase. The gas and comprehensive feature subset includes: the instantaneous change rate of characteristic sulfide concentration, the slow decrease rate of oxygen concentration, and the correlation coefficient between humidity and temperature time-series data. S3. Predefine the hidden state set M = {M1: normal state, M2: state of concern, S3: abnormal state, warning state, M5: fault state} of the branch box; establish a state transition model, the model parameters of which include: a state transition probability matrix characterizing the transition law between states, where the matrix elements represent the probability of transitioning from state Si to state Sj; and an observation probability distribution characterizing the likelihood of observing feature vectors in each state; during the online monitoring stage, input the real-time feature vector sequence obtained in step S2 into the state transition model, use a dynamic decoding algorithm to calculate the most likely hidden state sequence at the current moment, and determine the final state of the sequence as the current identification state of the branch box; at the same time, calculate the probability distribution of each hidden state at the next moment; S4. Implement a parallel dual-path decision-making mechanism; the first path is a rapid alarm path based on instantaneous physical quantities: when the raw data collected by any sensor or its simple derivative exceeds the preset absolute safety threshold, the highest level of emergency alarm is immediately triggered, and this triggering is independent of the state recognition result; the second path is an intelligent early warning path based on state recognition: according to the probability distribution P of the current identified state and the implicit state output in step S3, a graded early warning is performed: if the current identified state is "attention state" or "abnormal state", a planned inspection work order is generated; if the current identified state is "early warning state", a maintenance work order requiring emergency intervention is generated; if the predicted probability of being in a "fault state" at the next moment exceeds the preset risk probability threshold, a predictive alarm is triggered, indicating a high risk of a fault occurring in the short term.

[0009] The present invention further includes, in step S2, a multidimensional dynamic feature extraction step that also includes an advanced feature generation step: using signal processing and time series analysis methods, further mining deep correlation features from the original sensor data; for temperature data, calculating the standard deviation of the temperature at all measuring points to characterize the uniformity of the temperature field, and using a sliding window to calculate the approximate entropy or sample entropy of the temperature time series to quantify the complexity of temperature fluctuations; for partial discharge data, in addition to intensity and frequency, extracting the equivalent duration and equivalent bandwidth of the discharge signal, constructing a joint time-frequency domain feature, and calculating the coefficient of variation of the interval between adjacent discharge pulses to characterize the randomness or periodicity of the discharge; for gas concentration data, calculating the mutual information value between oxygen concentration and characteristic sulfide concentration to indicate whether the changes in the two types of gases originate from the same potential fault process; all the extracted basic features and advanced features together constitute a higher-dimensional, more information-rich enhanced feature vector, which is used to replace or supplement the basic feature vector input to the state transition model.

[0010] The present invention further specifies that the state transition model in step S3 is an improved Hidden Markov Model or its extended model; the observation probability distribution is modeled using a Gaussian mixture model; the initialization of the state transition probability matrix is ​​based on the physical model of equipment failure and expert experience, and its default setting follows the following principles: the probability of directly jumping from "normal state" to "fault state" is minimized, while the transition path from "normal state" through "attention state", "abnormal state", and "warning state" has a higher initial probability, and the state reversal probability is set to a non-zero value to accommodate the situation of fault elimination or environmental interference fading; the training and parameter optimization of the model adopt a supervised or semi-supervised learning algorithm based on historical monitoring data and corresponding equipment maintenance records, using a forward-backward algorithm to estimate model parameters, or using a Bayesian method to incorporate prior information into the transition matrix; during online decoding, the Viterbi algorithm, which can find the globally optimal path, is used to determine the most likely hidden state sequence, thereby obtaining a more stable current state identification result that is more consistent with the long-term degradation logic of the equipment.

[0011] The present invention further incorporates an environmental context adaptive correction mechanism into the state recognition process in step S3. Specifically, an environmental context word vector is added or associated to the feature vector, which includes at least environmental temperature, environmental humidity, current load current of the branch box, and day / night time information. When constructing the observation probability distribution model or post-processing the observation probability, environmental conditions are introduced as condition variables. By establishing an adaptive adjustment model for the feature benchmark value under different environmental context conditions, the state recognition model can distinguish between feature fluctuations caused by external environment or load changes and feature degradation caused by internal faults.

[0012] The present invention further integrates the intelligent early warning path in the dual threshold early warning and maintenance decision-making process described in step S4 with state health quantification and remaining effective life prediction functions. Specifically, after each state identification, not only are discrete state labels output, but a continuous state health index is also calculated. This index is the probability weighted sum of the current state at each benign state, or the information entropy calculated based on the state probability distribution. The state health index ranges from 0 to 1, where 1 represents absolute health and 0 represents complete failure. Simultaneously, based on the current state transition probability matrix and the probability distribution of the current state, the system utilizes... Markov chain Monte Carlo simulation or incorporating Markov chain theory simulates the future evolution path of equipment states, statistically analyzes the average time or time distribution required for the first arrival at a "fault state," and thus estimates the predictive remaining useful life (RUL). Early warning decisions are based on the probability distribution of the current identified state and the implied state, and comprehensively consider the rate of decline of the state health index and the estimated value of the predictive remaining useful life. When the rate of decline of the state health index exceeds the threshold or the RUL is lower than the preset safe operating cycle, an early warning is triggered even if the current identified state has not reached the "early warning" level, enabling more forward-looking maintenance planning.

[0013] The present invention also provides an intelligent monitoring system for distribution cable branch boxes. The system adopts an edge-cloud collaborative architecture, including a field monitoring unit, an edge computing gateway, a cloud platform analysis service center, and a user terminal. The field monitoring unit is fixedly installed inside the power distribution cable branch box and includes the heterogeneous sensor array, sensor signal conditioning circuit, analog-to-digital conversion module, and first wireless communication module; the sensor array is responsible for raw data acquisition, and the acquired data is transmitted through the first wireless communication module after conditioning and conversion. The edge computing gateway is deployed near the branch box or in the power distribution station, and has a built-in edge processor and a second wireless communication module. It communicates with one or more field monitoring units through the second wireless communication module to receive raw data. The edge processor is configured to perform the multi-dimensional dynamic feature extraction task described in step S2, compress the massive amount of raw data into a low-data-volume feature vector, and execute the instantaneous threshold fast alarm logic of the first path in step S4 to achieve local fast response. The cloud platform analysis service center includes a cloud server cluster and a database. The cloud server receives feature vector data and preprocessing results uploaded from various edge computing gateways and is configured to centrally execute the state recognition based on the state transition model described in step S3, as well as the intelligent early warning path decision of the second path in step S4. The database is used to store all historical feature data, state recognition records, model parameters, and device files. The cloud platform is also responsible for the centralized training, optimization, and downward deployment of the state transition model. The user terminal is a computer or mobile device running client software. It is connected to the cloud platform analysis service center via a network to receive alarm information, view real-time device status, health trends, predictive RUL reports, and execute the dispatch and closed-loop management of maintenance work orders.

[0014] The present invention further includes a modular, self-organizing network design for the sensors in the field monitoring unit; specifically, it includes: a main control module and multiple flexibly arranged sensor sub-modules; the main control module integrates power management, a core processing chip, and the first wireless communication module; each sensor sub-module is dedicated to the measurement of a physical quantity, including at least a temperature sensor sub-module, a partial discharge sensor sub-module, and a gas sensor sub-module; the sensor sub-modules communicate with the main control module via wired or short-range wireless means; the temperature sensor sub-module uses contact or non-contact temperature probes, arranged in an array at all cable joints; the partial discharge sensor sub-module uses an ultra-high frequency antenna sensor with a built-in pre-amplification circuit, installed at the optimal detection position inside the enclosure; the gas sensor sub-module uses a microelectromechanical system (MEMS) gas sensor array to monitor the concentration of multiple gases; all sensor sub-modules are synchronized by the main control module to achieve data acquisition synchronization; the main control module has local data caching capabilities, temporarily storing data during network interruptions and resuming subsequent transmissions to ensure data continuity.

[0015] The present invention further provides that the edge computing gateway has local lightweight state recognition and autonomous decision-making capabilities. Specifically, in addition to running feature extraction algorithms, the edge processor also pre-loads a lightweight version of the state recognition engine. This lightweight engine is a simplified version of the complete state transition model of the cloud platform, which may use a smaller subset of features, a simplified set of states, or a more computationally efficient classification algorithm. When the edge computing gateway is connected to the cloud platform network normally, it uploads the feature vectors and receives more accurate state recognition results from the cloud as feedback for online fine-tuning of the local lightweight model. When the network connection is interrupted or the cloud service is unavailable, the edge computing gateway automatically switches to autonomous mode, uses the local lightweight state recognition engine to analyze the feature vectors in real time, directly outputs a preliminary state judgment, and executes corresponding local early warning actions to ensure that it still has basic state monitoring and fault alarm capabilities under network isolation conditions, thereby improving system reliability.

[0016] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent monitoring method for power distribution cable branch boxes. Specifically, the program is configured to control a computing device to perform the following functions: receiving raw monitoring data streams from a heterogeneous sensor array; calling a data preprocessing and feature extraction module to generate multidimensional dynamic feature vectors according to a method; calling or accessing a pre-trained state transition model, the model parameters of which are stored in the medium or an associated database; executing a state decoding algorithm to calculate the probability distribution of the current implicit state and future state of the computing device; implementing dual threshold early warning logic to generate graded alarms and maintenance instructions based on instantaneous over-limit and state identification results; and optionally executing model parameter updates, health index calculations, and remaining life prediction algorithms. The storage medium includes, but is not limited to, server hard drives, solid-state drives, flash memory of edge computing devices, or cloud storage space, and the program code stored therein can be loaded and executed in a distributed computing environment from edge computing nodes to cloud servers.

[0017] The beneficial effects of this invention are as follows: It achieves continuous spectrum identification of equipment status: By dividing and probabilistically identifying five implicit states, it can accurately capture the intermediate degradation stages of equipment from health to failure, providing the possibility for early intervention. It improves the accuracy and reliability of alarms: By integrating multi-dimensional features and using Hidden Pattern Models (HMM) instead of single-point thresholds for judgment, it effectively suppresses false alarms caused by environmental interference or instantaneous fluctuations. It possesses predictive maintenance capabilities: By probabilistically predicting future states through a state transition probability matrix and calculating a health index, it realizes the transformation from "reactive maintenance" and "periodic maintenance" to "predictive maintenance." The model has self-learning and adaptive capabilities: Through online or periodic parameter updates, the model can evolve with the actual aging process of the equipment, maintaining long-term accuracy. It provides an intuitive quantitative assessment tool: The Health Index (HI) provides a unified and quantitative decision-making basis for equipment asset management and full lifecycle cost analysis. Attached image description: Figure 1 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0018] The embodiments of this application will be described in detail below, providing a clear and complete description of the technical solutions within this application. Obviously, the described embodiments are merely a portion of the embodiments of this application, and not all of them. The components of this application described and shown herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0019] This invention provides an intelligent monitoring method for power distribution cable branch boxes, comprising the following steps: S1. Multi-source heterogeneous data acquisition and synchronization: A heterogeneous sensor network consisting of distributed temperature sensors, ultra-high frequency (UHF) partial discharge sensors, gas microenvironment sensors, and door status sensors is deployed at key locations inside the branch box. All sensors are synchronized via a unified clock source or synchronization protocol to ensure data timing consistency.

[0020] S2. Multidimensional dynamic feature extraction and fusion: After preprocessing the raw data, time-domain, frequency-domain, and time-frequency-domain features with clear physical meaning are extracted to form a high-dimensional feature vector. Feature extraction not only includes basic statistics, but also focuses on the correlation and trends between features, such as temperature gradient, discharge signal phase distribution, and gas concentration change rate.

[0021] S3. Implicit state identification based on state transition model: The operating status of the branch box is discretized into five implicit states: Normal S1, Attention S2, Abnormal S3, Warning S4, and Fault S5. The set of implicit states is represented as S = {S1, S2, S3, S4, S5}.

[0022] Establish a state transition model: Use a Hidden Markov Model (HMM) or its extended model as the state recognition framework; the model consists of triples. Parameterized definition: State transition probability matrix ,in .

[0023] Observation probability distribution B: Gaussian mixture model (GMM) is used for the given state. Observed eigenvectors Model the probability: ; in, It is the number of GMM components. It is the weight of the m-th Gaussian component. and These are the mean vector and covariance matrix of the component, respectively. This represents the probability density function of a multivariate Gaussian distribution.

[0024] In the initial state distribution ,in .

[0025] Then, using historical data, the Baum-Welch algorithm (forward-backward algorithm) was employed to train the model parameters λ.

[0026] Where the forward probability The recursive calculation is as follows: ; ; Where the backward probability The recursive calculation is as follows: ; ; use and Calculate the state at time t given the observation sequence. probability and time t from state Transferred to probability The details are as follows:

[0027]

[0028] use and Iteratively update the GMM parameters in model parameters A and B.

[0029] Online state decoding (recognition) for real-time observed feature sequences The Viterbi algorithm is used to find the optimal state sequence. .

[0030] definition The recursive calculation is as follows:

[0031]

[0032] Simultaneously record the state index that maximizes the above expression. ; Ultimately, the optimal final state And through backtracking The entire optimal state sequence is obtained.

[0033] Future state prediction: Probability distribution of states based on the current time T Given the state transition matrix A, predict the state probability distribution after k steps. : .

[0034] Dual threshold early warning and health measurement: Instantaneous physical quantity hard threshold alarm: Set absolute safety upper limits for raw temperature, partial discharge intensity, etc. Once exceeded, the highest level alarm is triggered immediately to ensure rapid response to sudden serious faults.

[0035] Soft threshold warning based on state probability: If the current identified state is... Generate maintenance work orders at the corresponding level for S3 or S4. (This is based on the predicted failure state probability.) Exceeding the preset risk probability threshold (e.g., 0.4) triggers a predictive alarm.

[0036] The health index is calculated as follows: ,in, Is a state Weighting coefficients, for example =1.0, =0.7, =0.4, =0.1; It is the current state. The probability (which can be derived from) get).

[0037] The system periodically collects new operational data and corresponding maintenance records, and updates the HMM parameters using incremental learning or sliding window retraining. The structure of GMM enables the model to adapt to equipment aging and changes in the operating environment.

[0038] If certain terms are used in the specification and claims to refer to specific components, those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The term "comprising" as used throughout the specification and claims is an open-ended term and should be interpreted as "comprising but not limited to." "Approximately" means that within an acceptable margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error.

[0039] It should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a product or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the product or system that includes that element.

[0040] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A power distribution cable branch box intelligent monitoring method, characterized in that, The method comprises the following steps: S1: In the key electrical connection points and cavity space inside the branch box, a heterogeneous sensor array is arranged for synchronous or quasi-synchronous data acquisition; the sensor array at least comprises a distributed temperature sensor, an ultrahigh frequency partial discharge sensor, a gas microenvironment sensor and a box door state sensor; the distributed temperature sensor is used to obtain the temperature value and its spatial distribution of the cable joint and busbar connection; the ultrahigh frequency partial discharge sensor is used to sense and quantify the electromagnetic wave signal generated by the partial discharge of the insulating material; the gas microenvironment sensor is used to monitor at least the oxygen concentration, characteristic sulfide concentration and environmental relative humidity; the box door state sensor is used to monitor the opening and closing state and sealing integrity of the box body door; S2. The raw time series data collected in step S1 is preprocessed, and the preprocessing comprises filtering, denoising and data alignment; from the preprocessed data, a feature vector containing multiple dimensions of physical meaning is extracted, which is composed of at least a temperature feature subset, a partial discharge feature subset and a gas and comprehensive feature subset; the temperature feature subset comprises the maximum temperature value in the current monitoring period, the maximum temperature gradient value between key measuring points, and the temperature change trend slope calculated based on the highest temperature sequence in the past preset time window; the partial discharge feature subset comprises the average discharge intensity, the number of discharge pulses per unit time, and the correlation distribution characteristics of the discharge pulse and the power frequency voltage phase; the gas and comprehensive feature subset comprises the instantaneous change rate of the characteristic sulfide concentration, the slow decline rate of the oxygen concentration, and the correlation coefficient between the humidity and temperature time series data; S3. The implicit state set of the branch box is predefined as M={M1: normal state, M2: attention state, S3: abnormal state, prewarning state, M5: fault state}; a state transition model is established, and the model parameters comprise: a state transition probability matrix representing the state transition law, wherein the matrix elements represent the probability of transition from state Si to state Sj; and an observation probability distribution representing the likelihood of observing the feature vector in each state; in the online monitoring stage, the real-time feature vector sequence obtained in step S2 is input into the state transition model, the most probable implicit state sequence at the current time is calculated by using a dynamic decoding algorithm, and the last state of the sequence is determined as the current recognition state of the branch box; at the same time, the probability distribution of each implicit state at the next time is calculated. S4. Implement parallel dual-path decision mechanism; the first path is a fast alarm path based on instantaneous physical quantities: when the raw data collected by any sensor or its simple derivative exceeds the preset absolute safety threshold, the highest level of emergency alarm is triggered immediately, which is independent of the state recognition result; the second path is an intelligent early warning path based on state recognition: according to the current recognition state and the probability distribution P of the implied state output in step S3, hierarchical early warning is performed: if the current recognition state is "attention state" or "abnormal state", a planned inspection work order is generated; if the current recognition state is "warning state", a maintenance work order requiring emergency intervention is generated; if the probability of the next moment predicted to be in the "fault state" exceeds the preset risk probability threshold, a predictive alarm is triggered, prompting a high risk of failure in the short term.

2. The intelligent monitoring method for power distribution cable branch box according to claim 1, characterized in that, The multi-dimensional dynamic feature extraction in step S2 further includes a high-level feature generation step: using signal processing and time series analysis methods, further mining deep associated features from the original sensor data; for temperature data, the standard deviation of the temperature of all measuring points is calculated to represent the uniformity of the temperature field, and the approximate entropy or sample entropy of the temperature time series is calculated using a sliding window to quantify the complexity feature of temperature fluctuation; for partial discharge data, in addition to intensity and frequency, the equivalent time length and equivalent bandwidth of the discharge signal are extracted, joint time-frequency domain features are constructed, and the coefficient of variation of the interval time between adjacent discharge pulses is calculated to represent the randomness or periodicity mode of the discharge; for gas concentration data, the mutual information value between oxygen concentration and characteristic sulfide concentration is calculated to indicate whether the changes of the two types of gas are caused by the same potential failure process; all the extracted basic features and high-level features together constitute an enhanced feature vector of higher dimension and richer information, which is used to replace or supplement the basic feature vector input into the state transition model.

3. The intelligent monitoring method for power distribution cable branch box according to claim 1, characterized in that, The state transition model in step S3 is specifically an improved hidden Markov model or an extended model thereof; the observation probability distribution is modeled by a Gaussian mixture model; the initialization of the state transition probability matrix is based on the device failure physical model and expert experience knowledge, and the default setting follows the following principles: the probability of jumping directly from "normal state" to "fault state" is a minimum value, while the transition path from "normal state" through "attention state", "abnormal state", "warning state" gradually deteriorates, and has a higher initial probability, and the state reversal probability is set to a non-zero value to accommodate the situation of failure elimination or environmental disturbance subsidence; The training and parameter optimization of the model use supervised or semi-supervised learning algorithms based on historical monitoring data and corresponding device maintenance records, use the forward-backward algorithm to estimate the model parameters, or use the Bayesian method to incorporate prior information into the transition matrix; in online decoding, the Viterbi algorithm capable of finding the global optimal path is used to determine the most likely hidden state sequence, so as to obtain a more stable and more consistent with the long-term degradation logic of the current state recognition result of the device.

4. The intelligent monitoring method for power distribution cable branch box according to claim 1, characterized in that, The state recognition process in step S3 further introduces an environmental context adaptive correction mechanism; specifically, in the feature vector, an environmental context vector is additionally added or associated, which at least contains environmental temperature, environmental humidity, current load current of the branch box, and day and night period information; when constructing the observation probability distribution model or post-processing the observation probability, the environmental conditions are introduced as conditional variables; by establishing an adaptive adjustment model of feature reference values under different environmental context conditions, the state recognition model can distinguish between feature fluctuations caused by external environment or load changes and feature degradation caused by internal faults.

5. The intelligent monitoring method of the power distribution cable branch box according to claim 1, characterized in that, The intelligent early warning path in the double threshold early warning and maintenance decision in step S4 further integrates state health quantification and residual useful life prediction functions; specifically, after each state recognition, not only a discrete state label is output, but also a continuous state health index is calculated, which is the probability weighted sum of being in each benign state at the current time or the information entropy calculated based on the state probability distribution; the state health index ranges from 0 to 1, with 1 representing absolute health and 0 representing complete failure; at the same time, based on the current state transition probability matrix and the probability distribution of the current state, the future evolution path of the device state is simulated using Markov chain Monte Carlo simulation or absorbing Markov chain theory, the average time or time distribution required for the first arrival of the "fault state" is calculated, and the predictive residual useful life RUL is estimated; the early warning decision is based on the current recognized state and the probability distribution of the implicit state, and comprehensively considers the decline rate of the state health index and the estimated value of the predictive residual useful life; when the decline rate of the state health index exceeds the threshold or the RUL is lower than the preset safe operation period, early warning is triggered even if the current recognized state does not reach "warning", realizing more forward-looking maintenance planning.

6. A power distribution cable branch box intelligent monitoring system, characterized in that, The system adopts an edge-cloud collaborative architecture, including a field monitoring unit, an edge computing gateway, a cloud platform analysis service center, and a user terminal; The field monitoring unit is fixedly installed inside the power distribution cable branch box and includes the heterogeneous sensor array, the sensor signal conditioning circuit, the analog-to-digital conversion module, and the first wireless communication module of claim 1; the sensor array is responsible for raw data acquisition, and the acquired data is sent through the first wireless communication module after conditioning and conversion; The edge computing gateway is deployed near the branch box or in the power distribution station and is internally provided with an edge processor and a second wireless communication module; it communicates with one or more field monitoring units through the second wireless communication module to receive raw data; the edge processor is configured to perform the multi-dimensional dynamic feature extraction task in step S2 of claim 1, compresses massive raw data into a low-data feature vector, and performs the instantaneous threshold rapid alarm logic of the first path in step S4 to realize local rapid response; The cloud platform analysis service center includes a cloud server cluster and a database; The cloud server receives the feature vector data and preprocessing results uploaded from each edge computing gateway, is configured to centrally execute the state recognition based on the state transition model described in step S3 of claim 1, and the intelligent early warning path decision of the second path in step S4; the database is used to store all historical feature data, state recognition records, model parameters and device archives; the cloud platform is also responsible for centralized training, optimization and downward deployment of the state transition model; The user terminal, which is a computer or mobile device running client software, is connected to the cloud platform analysis service center through the network, and is used to receive alarm information, view real-time device status, health trend, predictive RUL report and perform dispatch and closed-loop management of operation and maintenance work orders.

7. The intelligent monitoring system of distribution cable branch box according to claim 6, characterized in that, The sensors in the field monitoring unit adopt a modular and self-organizing network design, specifically including a master control module and multiple flexible arrangement sensor sub-modules; the master control module integrates power management, a core processing chip and the first wireless communication module; each sensor sub-module is dedicated to measuring a physical quantity, including at least a temperature sensor sub-module, a partial discharge sensor sub-module and a gas sensor sub-module; the sensor sub-modules communicate with the master control module through wired or short-range wireless means; the temperature sensor sub-module adopts a contact or non-contact temperature measurement probe and is arranged in an array at all cable joints; the partial discharge sensor sub-module adopts a UHF antenna sensor with a built-in pre-amplification circuit and is installed at the best detection position in the internal space of the box; the gas sensor sub-module adopts a micro-electro-mechanical system gas sensor array to monitor multiple gas concentrations; all sensor sub-modules are uniformly time-synchronized by the master control module to achieve synchronization of data acquisition; the master control module has local data caching capability to temporarily store data when the network is interrupted and subsequently transmit the data after recovery, thereby ensuring data continuity. 8.The intelligent monitoring system of a power distribution cable branch box according to claim 6, characterized in that, The edge computing gateway further has local lightweight state recognition and autonomous decision-making capability; specifically, in addition to running the feature extraction algorithm, the edge processor also preloads a lightweight version of the state recognition engine; when the edge computing gateway and the cloud platform network connection is normal, the feature vector is uploaded and the more accurate state recognition result from the cloud is received as feedback for online fine-tuning of the local lightweight model; when the network connection is interrupted or the cloud service is unavailable, the edge computing gateway automatically switches to autonomous mode, uses the local lightweight state recognition engine to analyze the feature vector in real time, directly outputs the preliminary state judgment and performs the corresponding local early warning action, ensuring that the system still has basic state monitoring and fault alarm capability in the case of network isolation, and improving the system reliability.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, implements the steps of the power distribution cable branch box intelligent monitoring method of any one of claims 1 to 5; specifically, the program is configured to control the computing device to perform the following functions: receiving raw monitoring data streams from a heterogeneous sensor array; The data preprocessing and feature extraction module is called to generate the multi-dimensional dynamic feature vector according to the method of claim 2; a pre-trained state transition model is called or accessed, and the model parameters are stored in a medium or an associated database; a state decoding algorithm is executed to calculate the current hidden state and the future state probability distribution of the equipment; a double threshold early warning logic is implemented to generate a hierarchical alarm and maintenance instruction according to the instantaneous overrun and state recognition result; and optionally, model parameter updating, health index calculation and residual life prediction algorithms are executed; and the storage medium includes but is not limited to a server hard disk, a solid state disk, a flash memory of an edge computing device, or a cloud storage space, and the program code stored therein can be loaded and executed in a distributed computing environment from an edge computing node to a cloud server.

Images