Intelligent power distribution cabinet state dynamic monitoring method and system

Abstract

The invention relates to the technical field of power distribution cabinet detection, and discloses an intelligent power distribution cabinet state dynamic monitoring method and system, and the method comprises the following steps: collecting multi-source data generated in the operation process of a power distribution cabinet; preprocessing the multi-source data to generate a standardized data set; a random field model is constructed and used for describing the spatial relevance between units in the power distribution cabinet and the relation between the state of each unit and multi-source data, and model parameters are estimated through the maximum logarithmic posterior probability. An intelligent power distribution simulation model and a multi-source data acquisition technology are adopted, multi-dimensional parameters are acquired in real time through a high-precision sensor, monitoring efficiency and decision accuracy are remarkably improved, data analysis integrity and model sensitivity are ensured through dynamic time modeling and capturing of space and time relevance of the power distribution cabinet, and the power distribution cabinet can be monitored more accurately. And the state distribution visualization and maintenance priority strategy generation module optimizes the maintenance plan, reduces the operation and maintenance cost, and guarantees the stability of the power distribution system.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution cabinet detection, and specifically to a method and system for dynamically monitoring the status of intelligent distribution cabinets. Background Art

[0002] With the acceleration of the modern industrialization process, the safety and reliability of the power system have become the key to ensuring the normal development of social and economic activities. However, traditional power monitoring methods, especially the monitoring and management of the distribution system, face problems such as long cycle and poor timeliness, and it is difficult to meet the requirements of high-efficiency operation; Existing distribution cabinet monitoring methods mostly rely on regular manual inspections or the monitoring of a small number of single-point parameters. This method has obvious defects - long cycle and slow response. Once an abnormality occurs in the equipment, it is very likely that key information cannot be captured in time, resulting in the failure to give an early warning of equipment faults. In severe cases, it may even cause power outages. Especially in industrial scenarios with high load operation, traditional monitoring methods cannot meet the real-time monitoring requirements of the dynamic operating state; Currently, many distribution cabinet detection systems only perform simple filtering and analysis on the collected data, lacking systematic multi-dimensional data processing capabilities. In complex industrial environments, problems such as noise interference and sensor errors will lead to a decline in data quality. Static analysis models are even more unable to effectively capture the dynamic change characteristics of the data, which makes the existing technology lack depth in analyzing the status of distribution cabinets and is difficult to meet the requirements of high-precision status assessment in complex scenarios; Traditional distribution cabinet status prediction technologies mostly rely on rule models or simple machine learning algorithms, usually only able to output status classification results, and lacking quantification of the prediction credibility. The direct problem brought by this method is that when the prediction result is not accurate enough, it is difficult for maintenance personnel to judge its credibility and it is also difficult to formulate a scientific operation and maintenance plan based on this. These methods usually cannot dynamically adjust the decision-making priority according to the uncertainty in the prediction, and this deficiency limits the applicability of the existing system in key scenarios; The current maintenance plan mainly relies on manual experience and lacks systematic scientific basis. Maintenance personnel usually maintain the equipment according to the established inspection cycle, and it is difficult to accurately locate potential risk points of faults or prioritize the processing of key units. This maintenance method not only increases unnecessary operation and maintenance costs, but also easily misses the best maintenance opportunity for high-risk equipment, further threatening the operation reliability of the distribution system. The existing technology is significantly lacking in the intelligence and efficiency of generating maintenance strategies and urgently needs improvement; Therefore, a method and system for dynamically monitoring the status of intelligent distribution cabinets are proposed to solve the above problems. Summary of the Invention

[0003] In view of the deficiencies of the prior art, the present invention provides a method and system for dynamically monitoring the status of intelligent power distribution cabinets. Through an intelligent power distribution simulation model, multi-source data acquisition and analysis technology, and dynamic modeling, combined with preprocessing, and the joint modeling of conditional random fields and Bayesian deep learning, the complex spatial associations and temporal dependencies of the status are captured. The status distribution visualization and maintenance priority strategy module efficiently optimizes the operation and maintenance process, solving problems such as monitoring lag, large noise interference, insufficient prediction transparency, and waste of maintenance resources compared with the prior art, and comprehensively ensuring the stability and reliability of the power distribution system.

[0004] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for dynamically monitoring the status of an intelligent power distribution cabinet, including the following steps: Collect multi-source data generated during the operation of the power distribution cabinet; Preprocess the multi-source data to generate a standardized data set; Construct a random field model to describe the spatial correlation between units in the power distribution cabinet and the relationship between the states of each unit and the multi-source data, and estimate the model parameters by maximizing the log posterior probability; Based on the dynamic time Markov random field extension model, capture the temporal dependence of the power distribution cabinet unit status, and obtain a dynamic distribution model describing the power distribution cabinet status; Use Bayesian deep learning to predict the status of the power distribution cabinet units, and output the status probability distribution of each unit and the corresponding uncertainty quantification value; Generate a visual distribution map of the power distribution cabinet status based on the prediction results, and generate a maintenance priority strategy in combination with the distribution map.

[0005] Preferably, the multi-source data includes electrical parameters, environmental parameters, and physical state parameters, where: The electrical parameters include current, voltage, and power; The environmental parameters include temperature and humidity; The physical state parameters include vibration and noise.

[0006] Preferably, the preprocessing includes the following steps: Perform wavelet transform denoising on the multi-source data to remove high-frequency noise generated during the acquisition process; Normalize the multi-source data with different dimensions, map all data to the same scale, and facilitate feature fusion for subsequent modeling; Serialize the multi-source data in chronological order for constructing a data structure with temporal dependence.

[0007] Preferably, the construction of the random field model includes the following steps: Use a conditional random field model to describe the spatial correlation between the states of units in a distribution cabinet. The model is built based on the conditional probability distribution of state variables and observed data; Define the interaction weights between states to describe the interaction strength between the units in the distribution cabinet; Define the potential function of a single-point state, and describe the dependence between the unit state and the observed data based on the linear relationship between the unit characteristics and the model parameters.

[0008] Preferably, the model parameters of the logarithmic posterior probability are calculated by the following formula: Where, represents the logarithmic posterior probability of the random field model; θ is the parameter of the random field model, including the interaction weights between states and the weights of the potential function of a single-point state; C represents the neighborhood set of the units in the distribution cabinet; ψ C (S C ,X C ) is the potential function; logZ(X) is the logarithm of the normalization factor.

[0009] Preferably, the dynamic time Markov random field model is implemented through the following steps: Define the time series state of each unit in the distribution cabinet, associate the state at the current moment with the state at the previous time point, and establish a time series model; Use the state transition function to describe the time dependence of the unit state in the distribution cabinet, parameterize the state transition function, and optimize the time correlation weights; Calculate the state probability distribution at all times according to the state transition function, and thus generate a dynamic distribution describing the comprehensive state of the units in the distribution cabinet.

[0010] Preferably, the steps of Bayesian deep learning for predicting the state of the units in the distribution cabinet include the following: Assume that the weights of the neural network are random variables, and estimate the weights by maximizing the posterior distribution of the data; Obtain the probability distribution of state prediction and the uncertainty quantification value through Monte Carlo sampling; Combine the uncertainty quantification value with the state probability distribution to generate a highly reliable prediction result.

[0011] Preferably, the prediction of the uncertainty quantification value includes the following steps: Sample the posterior distribution of the weights of the Bayesian neural network to generate multiple groups of weight samples; Use the sampled weight samples to calculate the prediction probability distribution of the unit state in the distribution cabinet respectively; Generate the uncertainty quantification value by calculating the variance of the prediction distribution, which is used to measure the reliability of the prediction result; When the uncertainty quantification value is higher than the preset threshold, mark the corresponding unit as a high-risk state and output a risk warning.

[0012] Preferably, the uncertainty quantification value is calculated by the following formula: where Uncertainty represents the uncertainty quantification value, which is used to describe the credibility of the power distribution cabinet state prediction result; N is the number of Monte Carlo sampling times; ω i represents the network weight sampled from the posterior distribution for the i-th time; P(S|X,ω i ) represents the prediction probability of the Bayesian neural network for the state S under the i-th sampling; S is the state of the power distribution cabinet unit; X is the input multi-source observation data; is the mean of all sampling prediction probabilities, representing the expected value of the prediction result.

[0013] Preferably, the present invention also provides an intelligent power distribution cabinet state dynamic monitoring system, including: A data acquisition module for acquiring electrical parameters, environmental parameters, and physical state parameters of the power distribution cabinet; A data preprocessing module for denoising, normalizing, and time-series processing of the acquired data; A random field modeling module for constructing a spatial association model and a dynamic time Markov random field model of the power distribution cabinet unit state; a deep learning prediction module for predicting the power distribution cabinet state based on Bayesian deep learning and quantifying the prediction uncertainty; a visualization and maintenance strategy module for generating a visualization distribution map of the power distribution cabinet state and a maintenance priority strategy.

[0014] The present invention provides an intelligent power distribution cabinet state dynamic monitoring method and system. It has the following beneficial effects: 1. The present invention adopts an intelligent power distribution simulation model and multi-source data acquisition and analysis technology, and uses high-precision sensors to obtain multi-dimensional operation parameters of the power distribution cabinet in real time, and combines dynamic modeling to accurately reflect the state. This solution significantly improves the efficiency of monitoring and the accuracy of data analysis. Compared with the traditional methods based on regular manual inspections and single-parameter monitoring in the prior art, the present invention solves the defects of monitoring lag and insufficient state coverage, and fundamentally meets the growing requirements of modern power systems for real-time and high efficiency.

[0015] 2. Through multi-level preprocessing methods such as wavelet denoising, normalization, and time serialization, the present invention ensures the integrity and consistency of data input. Meanwhile, by combining conditional random field and dynamic time Markov random field modeling techniques, the complex spatial and temporal correlations between power distribution cabinet units are captured, realizing the dynamic description of the state. Compared with the simple single-point monitoring and static analysis models in the prior art, the present invention effectively overcomes the problems of large data noise interference and insufficient sensitivity of the model to dynamic changes, providing a solid data and algorithm basis for state analysis and prediction.

[0016] 3. The present invention introduces Bayesian deep learning technology. By making high-precision predictions on the state of the power distribution cabinet and combining uncertainty quantification analysis, the reliability evaluation of the prediction results is realized. This technical means can clarify the credibility of the prediction results and dynamically adjust the operation and maintenance strategies according to the level of uncertainty. Compared with the prior art that relies on traditional deep learning models and lacks the reliability evaluation of prediction results, the present invention significantly improves the transparency of prediction and the accuracy of decision-making, solving the deficiency in the combination of credible prediction and practical application in the prior art.

[0017] 4. Through the state distribution visualization and maintenance priority strategy generation module, the present invention presents the complex state analysis results in an intuitive graphical manner, provides scientific maintenance suggestions, and clarifies the key units and time sequence for maintenance. Different from the traditional method of formulating maintenance plans relying on manual experience in the prior art, the present invention effectively solves the problems of low maintenance decision-making efficiency and poor resource utilization rate. By optimizing the operation and maintenance process, it not only reduces the economic losses caused by power outages due to failures but also reduces the maintenance cost, further ensuring the stability and reliability of the power distribution system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flowchart of the method of the present invention; Figure 2 is a schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] Please refer to the attached Figure 1 , the embodiment of the present invention provides a method for dynamically monitoring the state of an intelligent power distribution cabinet, including the following steps: S1. Collect multi-source data generated during the operation of the power distribution cabinet; S2. Preprocess the multi-source data to generate a standardized data set; As the first step of the present invention, the implementation of step S1 provides a necessary data basis for subsequent steps. Its main purpose is to collect multi-dimensional data during the operation of the power distribution cabinet through multi-source sensors. Step S2, on the other hand, systematically processes the collected data to meet the input requirements of subsequent random field modeling and Bayesian deep learning prediction. The quality and standardization degree of the data are directly related to the performance of the subsequent model. To ensure the integrity and effectiveness of the data, it is carried out from two aspects: data collection and preprocessing.

[0021] In step S1, the sources of data collection include electrical parameters, environmental parameters, and physical state parameters, covering the key information affecting the unit state during the operation of the power distribution cabinet. For example, the current and voltage in electrical parameters can reflect the circuit load state, the temperature and humidity in environmental parameters describe the impact of the external environment on electrical equipment, and physical state parameters such as vibration and noise can indicate whether there are mechanical abnormalities or external impacts on the equipment. These data are obtained through multi-type sensors distributed inside the power distribution cabinet to ensure coverage of each key unit.

[0022] In this embodiment, the collection process of multi-source data is based on the deployment of a distributed sensor network. Each sensor node separately collects electrical parameters (such as current, voltage, and power), environmental parameters (such as temperature and humidity), and physical state parameters (such as vibration and noise). The sensor nodes can transmit the collected data to the central processing unit through wired or wireless communication methods (such as LoRa, Wi-Fi, or ZigBee protocols). Generally, the transmission frequency can be dynamically adjusted according to the importance of the power distribution cabinet operation. For example, for industrial scenarios with frequent load fluctuations, a higher sampling frequency can be set, while for relatively stable environments, the transmission frequency can be appropriately reduced to save communication resources.

[0023] In step S2, as an option, after collecting the data, the preliminary processing module will perform denoising on the data. Specifically, in this embodiment, the denoising process uses the wavelet transform method to decompose the collected original signal into different frequency bands, retain the low-frequency components containing the main information, and remove the high-frequency components containing noise. This can effectively filter out high-frequency interference signals while maintaining the original characteristics of the data.

[0024] To further ensure the consistency and usability of the data, this embodiment normalizes multi-source data with different dimensions. Generally, the mathematical form of normalization is as follows: Where X ′ is the normalized data value, X is the original data; μ is the mean of this data; σ is the standard deviation of this data.

[0025] Through this normalization process, data with different dimensions are mapped to a unified scale range, thereby eliminating the influence of unit differences on subsequent analyses.

[0026] In some embodiments, the data acquisition time points may not be consistent, especially when there is asynchronous acquisition in multi-source sensors. Therefore, it is necessary to perform time alignment and serialization processing on the data. In this embodiment, interpolation algorithms are used to supplement missing data, and the data is sorted according to timestamps to generate a continuous time series. Specifically, linear interpolation methods are used to process incomplete temperature and humidity data, thereby ensuring that the data stream has consistent time dependence.

[0027] S3. Construct a random field model to describe the spatial correlation between units in the power distribution cabinet and the relationship between the states of each unit and multi-source data, and estimate the model parameters by maximizing the log posterior probability; In the power distribution cabinet state detection method of the present invention, step S3 aims to describe the spatial correlation between units in the power distribution cabinet by constructing a random field model and establish a mapping relationship between unit states and multi-source data. The construction of the random field model is one of the core steps of the whole method, and its result directly determines the accuracy and reliability of subsequent state prediction and maintenance strategy generation. Step S3 receives the output of the data preprocessing module (standardized data set) and provides a basis for subsequent dynamic modeling and deep learning prediction. Generally, the design of the random field model needs to comprehensively consider the physical structure of each unit in the power distribution cabinet, the characteristic dimensions of the operation data, and the complexity of their mutual correlations.

[0028] The random field model in this step is based on the theory of conditional random field (CRF) and is further extended to the dynamic time Markov random field model (DTMRF). In some embodiments, in order to capture the spatio-temporal correlation between power distribution cabinet units, the potential function of the random field model is designed as a joint expression of state variables and observation data. The technical content of this step will be described in detail below in combination with specific embodiments.

[0029] The random field model is constructed through the mathematical framework of conditional random fields. The random field model defines a probability distribution used to describe the dependence relationship between the states S = {s1, s2,..., s n} of each unit in the power distribution cabinet and the multi-source observation data X = {x1, x2,..., x n}. The model adopts the form of conditional probability distribution, and the specific expression is: Among them, P(S|X) is the conditional probability distribution of the conditional random field model. S represents the state vector of all units in the power distribution cabinet. Each unit state indicates that the unit is in the "normal" or "abnormal" state, which are represented by +1 and -1 respectively; X represents the multi-source observation data vector of all units in the power distribution cabinet, including electrical parameters, environmental parameters, and physical state parameters; C represents the neighborhood set of units, which is used to define the spatial association between units; ψ C (S C ,X C ) is the potential function, which represents the joint relationship between the state variable S C and the observation data X C ; Z(X) is the normalization factor, which is used to ensure that the sum of the probability distribution is 1.

[0030] The normalization factor Z(X) is specifically defined as: Among them, S is all possible combinations of unit states.

[0031] Specifically, the potential function ψ C (S C ,X C ) contains two main parts: the state interaction weight part and the single-point potential state function part. Its expression is: u i (x i )=θ T φ(x i ) Among them, w ij is the state interaction weight, which is used to describe the association strength between unit i and unit j in the power distribution cabinet; s i and s j represent the state variables of unit i and unit j; u i (x i ) represents the single-point potential state function, which is used to describe the dependence relationship between the state of unit i and its observation data x i ; θ is the model parameter; φ(x i ) is the feature mapping function of the observation data.

[0032] As a possible implementation, the neighborhood set C can be defined based on the actual layout of the physical units in the power distribution cabinet. For example, the neighborhood relationship can be determined by the physical distance between adjacent power distribution cabinet modules, or constructed through the circuit connection relationship. Generally, using a local neighborhood model can reduce the computational complexity while maintaining sufficient state expression ability.

[0033] In another possible implementation, the interaction weight w ijIt can be initialized with historical data. For example, the weight values can be calculated based on the correlation of unit states in historical operation data. The higher the weight value, the stronger the mutual dependence between the two unit states.

[0034] In this embodiment, to achieve the optimal fitting of the state-data relationship, the model parameters θ and w ij are learned by maximizing the log posterior probability. The objective function of the log posterior probability is: where, represents the log posterior probability of the random field model; θ is the parameter of the random field model, including the interaction weights between states and the weights of the single-point state potential function; C represents the neighborhood set of the units in the power distribution cabinet; ψ C (S C , X C ) is the potential function; logZ(X) is the logarithm of the normalization factor.

[0035] To efficiently solve the parameter optimization problem, in some embodiments, the natural gradient optimization strategy combining the gradient descent method and the Fisher information matrix is adopted. By introducing natural gradient optimization, the nonlinear problems in the parameter space can be effectively avoided, and the convergence speed can be improved.

[0036] S4. Based on the dynamic time Markov random field extended model, capture the time dependence of the power distribution cabinet unit state to obtain a dynamic distribution model describing the power distribution cabinet state; Step S4 plays a connecting role in the overall process. The dynamic time Markov random field model (DTMRF) of the present invention is extended on the basis of the foregoing random field modeling to capture the time dependence of the power distribution cabinet unit state. In the foregoing steps, the spatial correlation relationship between the power distribution cabinet units has been constructed through the random field model. In the actual operation of the power distribution cabinet, the state of the unit not only has a spatial correlation with other units, but also has significant time evolution characteristics. For example, the state of increasing temperature may occur continuously in multiple time segments, while the abnormal fluctuation of current may only appear in a short period. Therefore, the dynamic time dependence is crucial for the description and prediction of the unit state.

[0037] In this step, by introducing the idea of time series modeling, the current state of the power distribution cabinet unit is associated with the state at the previous time point to construct a dynamic time Markov random field model. The following is a detailed description of the specific implementation manner of step S4 in this embodiment.

[0038] In this embodiment, it is assumed that the states of n units in the power distribution cabinet at a certain moment t are represented by the state variable S t ={s 1,t , s 2,t ,..., sn,t}, where s i , t ∈ {-1, +1}. The state +1 indicates that the unit is in a normal state, while the state -1 indicates that the unit is in an abnormal state.

[0039] Observation data X t = {x 1,t , x 2,t ,..., x n,t} represents the multi-source observation information of the unit at time t, including but not limited to electrical parameters (such as current, voltage, power), environmental parameters (such as temperature, humidity), and physical state parameters (such as vibration, noise). These observation data have a conditional dependence relationship with the state variables.

[0040] In the dynamic time Markov random field model, the joint probability distribution of the state S t is defined as: where S t represents the set of states of all units at time t; X t represents the set of observation data of all units at time t; C represents the neighborhood set of the power distribution cabinet units in the random field; is the neighborhood potential function, used to describe the conditional dependence relationship between the states in the neighborhood C and the observation data ; φ(s i,t , s i,t-1 ) is the state transition potential function, describing the state dependence of the power distribution cabinet unit i at times t and t - 1; Z(X t ) is the normalization factor.

[0041] Specifically, the neighborhood potential function is defined as: u i (x i,t ) = θ T φ(x i,t ) where w ij represents the interaction weight between unit i and unit j, used to describe the spatial correlation intensity; represents the set of states of all units in the neighborhood C at time t; is the set of observation data of all units in the neighborhood C at time t; u i (x i,t ) represents the potential function of unit i; φ(x i,t ) is the high-dimensional mapping of the input feature x i,t , and θ is the model parameter.

[0042] The state transition potential function φ(s i,t , s i,t-1 ) is defined as: φ(s i,t , s i,t-1 ) = α i s i,t s i,t-1 where s i,t is the state of unit i at time t; s i,t-1 is the state of unit i at the previous time point t - 1; α i is the time correlation weight, which is used to quantify the dependence of the state of unit i over time.

[0043] In a possible implementation, in this embodiment, the model parameters θ, w, α are estimated by maximizing the log posterior probability of the dynamic time Markov random field. The objective function of the log posterior probability is defined as: where θ is the model parameter of the neighborhood potential function; w is the set of interaction weight parameters between neighborhood states; α is the set of time correlation weight parameters; T is the total duration of the time series.

[0044] In some embodiments, to improve the computational efficiency, the calculation of the normalization factor Z(X t ) can use approximate inference algorithms, such as variational inference or Markov chain Monte Carlo (MCMC) methods.

[0045] S5. Use Bayesian deep learning to predict the states of the power distribution cabinet units, and output the state probability distribution and the corresponding uncertainty quantification values of each unit; Step S5 is an important part for predicting the states of the power distribution cabinet units according to the constructed dynamic time Markov random field model and the analysis results of the foregoing multi-source data. By using the Bayesian deep learning method, the operating states of the units can be predicted with high confidence, and at the same time, the uncertainty quantification values of the predictions are output, so as to generate a maintenance priority strategy subsequently. This step is closely combined with the dynamic time modeling in step S4, and further provides a direct input basis for visualization and maintenance strategies.

[0046] In some possible implementation manners, the main function of step S5 is to achieve the transformation from complex data input to high-reliability prediction output. Especially in the case where the data distribution is uneven or contains noise, the Bayesian deep learning method can effectively quantify the uncertainty of the prediction results and improve the overall robustness of the system.

[0047] In general, when using a Bayesian neural network for state prediction, it is first necessary to assume that the weights of the network are random variables, so that the prediction has a probabilistic interpretation. Specifically, the Bayesian neural network estimates the weights by maximizing the posterior probability, and the posterior distribution can be expressed in the following form: P(ω|D)∝P(D|ω)P(ω) where ω represents the weight parameters of the neural network; D represents the observed dataset used for training; P(D|ω) is the likelihood function, which describes the probability of the data D appearing given the weight ω; P(ω) is the prior distribution of the weight parameters, which is used to introduce regularization constraints.

[0048] As an option, the posterior distribution is usually difficult to analyze. Therefore, in some implementations, the posterior distribution can be approximated by variational inference or Monte Carlo methods. Specifically, variational inference can be used to approximate the posterior distribution P(ω|D) with a simple distribution q(ω) to minimize the Kullback-Leibler (KL) divergence between the two.

[0049] In some specific implementations, to efficiently implement state prediction, the posterior distribution of the neural network weights can be sampled, and the sampled weight parameters are used to generate the state prediction probability distribution. The specific steps are as follows: First, use the output state of the dynamic time Markov random field model as the initial input of the network; Then, by sampling from q(ω) multiple times, N groups of weight samples ω1, ω2, …, ω N ; Subsequently, use each group of weight samples to calculate the predicted probability distribution P(S|X, ω i ).

[0050] Calculate the expected value of the prediction and the uncertainty quantification value using the sampled predicted probability distribution. The specific formulas are as follows: where represents the expected value of the predicted probability distribution, which is used to describe the state prediction of the unit; Uncertainty represents the uncertainty quantification value of the prediction, which reflects the credibility of the system in the state prediction of a specific unit; N is the number of Monte Carlo samplings. The more samplings, the higher the prediction accuracy; P(S|X, ω i ) is the predicted probability of the Bayesian neural network for the state S in the i-th sampling; X represents the input multi-source observed data.

[0051] In a possible implementation, when the uncertainty quantification value exceeds a certain preset threshold, the system will automatically mark the corresponding unit as "high-risk state" and output a risk warning message. Specifically, the threshold can be dynamically adjusted according to historical data or empirical values to meet the risk assessment requirements under different working conditions.

[0052] In some scenarios, due to the possible existence of outliers or uneven distributions in the input data, Bayesian deep learning can effectively address these issues. For example, when the unit load in the power distribution cabinet fluctuates violently, traditional neural networks may output misleading results due to overfitting, while the Bayesian framework can capture the uncertainty of the input through weight distribution and make more reasonable judgments on abnormal patterns.

[0053] In one implementation, time series analysis techniques can also be combined to further process the output of the dynamic time Markov random field. For example, by introducing time decay weights, the model pays more attention to the current state and appropriately weakens the influence of historical data, thereby improving the real-time performance and accuracy of the prediction.

[0054] Generally, after the above prediction steps, the output results include the following two parts: The state probability distribution of the unit. For example, the probability that a certain unit is in a normal state is 85%, and the probability of being in an abnormal state is 15%.

[0055] The uncertainty quantification value. For example, for the prediction result of a certain unit, its uncertainty is 0.03, indicating a relatively high confidence in the prediction of the unit state.

[0056] These output results will directly serve as the input for subsequent state visualization and maintenance strategy generation, providing a basis for actual maintenance decisions. By combining the state probability distribution and the uncertainty quantification value, "high-risk areas", "low-risk areas", and "normal areas" can be distinguished to achieve hierarchical management of the state of the power distribution cabinet units.

[0057] S6. Generate a visualization distribution map of the power distribution cabinet state based on the prediction results, and generate a maintenance priority strategy in combination with the distribution map.

[0058] In step S6, after the prediction result of the power distribution cabinet state is generated, in order to more intuitively reflect the operating state of the internal units of the power distribution cabinet and the abnormal risk areas, a state distribution map is generated through visualization technology. At the same time, combined with the distribution map, a targeted maintenance priority strategy is formulated to guide subsequent maintenance operations. This step is closely related to the foregoing steps. In particular, the outputs of the random field model and the Bayesian deep learning module are directly used as the visualization input to provide data support for the maintenance priority strategy. Generally, the prediction results include the state probability distribution and the uncertainty quantification value of each unit. Through these results, the abnormal distribution patterns and risk areas can be further analyzed, and finally an optimized maintenance plan can be formulated. In some embodiments, the state probability distribution and the uncertainty quantification value of different units are mapped to a two-dimensional plane, and a state distribution map is generated through a specific dimensionality reduction algorithm (such as t-SNE or UMAP). These algorithms compress the state feature points into a low-dimensional space by calculating the similarity between points in the high-dimensional space for intuitive display. Specifically, the t-SNE algorithm minimizes the difference between the high-dimensional and low-dimensional similarities through the following objective function: where argmin Y represents that the goal is to minimize the function value, Y is the data representation in the low-dimensional space, p ij represents the similarity between units i and j in the high-dimensional space, and q ij represents the corresponding similarity in the low-dimensional space. In this way, the aggregation areas of high-risk units can be clearly identified.

[0059] As an option, the color and size of the distribution map can be dynamically adjusted according to the state probability or the uncertainty quantification value of the unit. For example, in some embodiments, the color gradually changes from green (normal state) to red (high-risk state), and the depth of the color is proportional to the uncertainty value. At the same time, the size of the unit icon can represent the risk level of the state anomaly, which can help the operation and maintenance personnel quickly identify the areas that need to be inspected key points.

[0060] In some possible implementation manners, in order to enhance the practical value of the distribution map, the unit states in the distribution map can be further clustered. For example, a clustering method based on the Gaussian mixture model (GMM) is used to classify the states to obtain clusters of high-risk units. The probability density function of this clustering method is: where P(x) represents the probability density of the data point x, K represents the number of Gaussian mixture distributions, π k represents the weight of the k-th Gaussian distribution, represents the probability density function of the K-th Gaussian distribution, x represents the data point, μ kis the mean vector of the K-th Gaussian distribution, representing the central position of the cluster, and ∑ k is the covariance matrix of the K-th Gaussian distribution, representing the distribution shape and range of the cluster.

[0061] In the process of generating the maintenance priority strategy by combining the visualization distribution map, historical data and actual maintenance experience are also referred to. In a possible embodiment, a comprehensive risk score is calculated based on the uncertainty value of the unit and the historical anomaly rate. Generally, the comprehensive risk score is calculated by the following formula: Risk Score=α·Uncertainty+β·History Risk where α and β are weighting coefficients, Uncertainty is the current uncertainty value, and History Risk is the failure rate of the corresponding unit in the historical maintenance record.

[0062] Specifically, the maintenance priority strategy is divided into the following categories according to the comprehensive risk score: High priority: Units with a comprehensive risk score exceeding the preset threshold should be repaired immediately; Medium priority: Units with a comprehensive risk score within a medium range are recommended for regular monitoring; Low priority: Units with a relatively low comprehensive risk score only require routine maintenance.

[0063] In another implementation, the visualization distribution map is combined with the actual physical layout of the power distribution cabinet, and the status information is projected onto the 3D model of the power distribution cabinet. The physical position and its status distribution of the unit can be intuitively presented through the 3D model. This method is particularly suitable for complex power distribution cabinet systems and can greatly improve the maintenance efficiency.

[0064] Generally, the maintenance priority strategy generated by combining the distribution map can be automatically output as a maintenance report. The report content includes: the risk score of the unit, priority classification, specific recommended operations, and estimated maintenance time. As an extended implementation, the maintenance report can also combine intelligent scheduling algorithms to provide the optimal maintenance path planning for the operation and maintenance personnel to reduce the maintenance cost and time.

[0065] Please refer to the appendix Figure 2 , the present invention also provides an intelligent power distribution cabinet status dynamic monitoring system, including: A data acquisition module for collecting electrical parameters, environmental parameters, and physical state parameters of the power distribution cabinet; The data acquisition module is the basic component of the system and is responsible for collecting key parameter information during the operation of the power distribution cabinet, including electrical parameters, environmental parameters, and physical state parameters. This module obtains multi-source data through a distributed sensor network and provides accurate and real-time data input for the subsequent modules; For key monitoring points, the data acquisition module can adopt a multi-sensor redundant acquisition mechanism to improve the reliability of data. For example, current data can be collected simultaneously by two independent sensors and consistency verification is performed.

[0066] The data preprocessing module is used to perform denoising, normalization, and time serialization processing on the collected data; The data preprocessing module cleans, standardizes, and time serializes the collected raw data to ensure that the data quality meets the requirements of subsequent random field modeling and deep learning; The data preprocessing module performs denoising processing on the collected signal through wavelet transform. Wavelet transform decomposes the signal into different frequency bands, retains the low-frequency effective signal, and removes high-frequency noise. It is particularly suitable for denoising vibration and noise signals and can significantly improve the accuracy of data analysis.

[0067] The module has a built-in anomaly detection function and can mark possible data anomaly points in the preprocessing stage. For example, data points that deviate significantly from the mean are detected based on statistical methods, and the time and location of the abnormal data are recorded.

[0068] The random field modeling module is used to construct a spatial association model and a dynamic time Markov random field model for the state of the power distribution cabinet unit; The random field modeling module describes the spatial correlation of the power distribution cabinet unit state through conditional random field (CRF) and captures time dependence through the extended model of dynamic time Markov random field (DTMRF); The module uses the random field method to establish the spatial association relationship between each unit by using the collected multi-source data. For example, adjacent electronic components on the circuit board may have state linkages due to overload or short circuit, and the module quantifies the association strength through probability modeling.

[0069] The deep learning prediction module is used to predict the state of the power distribution cabinet based on Bayesian deep learning and quantify the prediction uncertainty; The deep learning prediction module is based on a Bayesian neural network, combines the results of random field modeling, predicts the operating state of the power distribution cabinet unit, and outputs the uncertainty quantification value of the prediction result; The module combines the foregoing modeling results to predict the state of the power distribution cabinet unit (such as normal, abnormal, faulty, etc.). The state prediction is based on multi-dimensional input data, including electrical, environmental, and physical state characteristics; To improve the credibility of the prediction, the module performs uncertainty quantification analysis on each prediction result. For example, when the predicted state of a certain unit is close to the fault threshold, its uncertainty is quantified to evaluate the reliability of the result. The module supports online learning and allows new data to be continuously input during the operation of the system to dynamically adjust the model weights and prediction results.

[0070] Visualization and maintenance strategy module, which is used to generate a visualization distribution map of the power distribution cabinet status and a maintenance priority strategy.

[0071] The visualization and maintenance strategy module presents the system's status analysis results in a graphical manner and generates a maintenance priority strategy for the power distribution cabinet unit based on the prediction results; The module projects high-dimensional status data into a two-dimensional space through data dimensionality reduction technology to generate an intuitive status distribution map. In the map, units in different statuses are marked with colors or symbols. For example, green indicates normal and red indicates abnormal; According to the predicted uncertainty quantification value and the status distribution result, the module assigns maintenance priorities to each unit: High priority: The unit status is predicted to be high risk and immediate maintenance is required. Medium priority: The unit status is predicted to be medium risk and periodic inspection is required. Low priority: The unit status is normal and only routine maintenance is required; The system has a built-in real-time alarm function. When the status prediction result of a certain unit shows high risk or the uncertainty quantification value exceeds the set threshold, an alarm is triggered and relevant information is pushed to the operation and maintenance personnel.

[0072] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically monitoring the state of an intelligent power distribution cabinet, characterized in that: The following steps are involved: Collect multi-source data generated during the operation of the power distribution cabinet; Preprocessing the multi-source data to generate a standardized data set; A random field model is constructed to describe the spatial correlation between the units in the distribution cabinet and the relationship between the state of each unit and multi-source data. The model parameters are estimated by maximizing the logarithmic posterior probability. Based on the extended model of dynamic time Markov random field, the time dependence of the state of the distribution cabinet unit is captured, and the dynamic distribution model describing the state of the distribution cabinet is obtained; Use Bayesian deep learning to predict the state of the distribution cabinet unit, and output the state probability distribution of each unit and the corresponding uncertainty quantification value; A visual distribution map of the power distribution cabinet status is generated based on the prediction results, and a maintenance priority strategy is generated based on the distribution map.

2. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The multi-source data includes electrical parameters, environmental parameters and physical state parameters, wherein: Electrical parameters include current, voltage, and power; Environmental parameters include temperature and humidity; Physical state parameters include vibration and noise.

3. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The pre-processing comprises the following steps: Perform wavelet transform denoising on multi-source data to remove high-frequency noise generated during the acquisition process; Normalize multi-source data of different dimensions and map all data to the same scale to facilitate feature fusion in subsequent modeling; Multi-source data is serialized in chronological order to construct a time-dependent data structure.

4. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The construction of the random field model comprises the following steps: A conditional random field model is used to describe the spatial correlation between the states of the units in the distribution cabinet, wherein the model is modeled based on the conditional probability distribution of the state variables and the observed data; Define the interaction weights between states to describe the interaction strength between the distribution cabinet units; The potential function of a single-point state is defined to describe the dependence of the cell state on the observed data based on the linear relationship between the cell characteristics and the model parameters.

5. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The logarithmic posterior probability estimation model parameters are calculated by the following formula: in, represents the logarithmic posterior probability of the random field model; θ is the parameter of the random field model, including the weight of the interaction between states and the weight of the single-point state potential function; C represents the neighborhood set of the unit in the distribution cabinet; ψ C (S C ,X C ) is the potential function; logZ(X) is the logarithm of the normalization factor.

6. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The dynamic time Markov random field model is implemented by the following steps: Define the time series state of each distribution cabinet unit, associate the current state with the state at the previous time point, and establish a time series model; The state transfer function is used to describe the time dependency of the state of the distribution cabinet unit, and the state transfer function is parameterized to optimize the time correlation weight; The state probability distribution at all times is calculated according to the state transfer function, thereby generating a dynamic distribution describing the comprehensive state of the distribution cabinet unit.

7. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The Bayesian deep learning step of predicting the state of the distribution cabinet unit includes the following: Assuming that the weights of the neural network are random variables, the weights are estimated by maximizing the posterior distribution of the data; The probability distribution and uncertainty quantification value of state prediction are obtained through Monte Carlo sampling; The uncertainty quantification value is combined with the state probability distribution to generate high-confidence prediction results.

8. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The prediction of the uncertainty quantification value comprises the following steps: Sampling the posterior distribution of the Bayesian neural network weights to generate multiple sets of weight samples; Use the weighted samples obtained by sampling to calculate the predicted probability distribution of the distribution cabinet unit status respectively; By calculating the variance of the prediction distribution, a quantitative uncertainty value is generated to measure the credibility of the prediction results; When the uncertainty quantification value is higher than the preset threshold, the corresponding unit is marked as a high-risk state and a risk warning is output.

9. The method for dynamic monitoring of the state of an intelligent power distribution cabinet according to claim 1, characterized in that: The uncertainty quantification value is calculated by the following formula: Among them, Uncertainty represents the uncertainty quantification value, which is used to describe the credibility of the distribution cabinet state prediction result; N is the number of Monte Carlo sampling; ω i represents the network weight sampled from the posterior distribution for the i-th time; P(S|X,ω i ) represents the prediction probability of the Bayesian neural network for state S under the i-th sampling; S is the state of the distribution cabinet unit; X is the input multi-source observation data; P(S|X) is the mean of all sampling prediction probabilities, which represents the expected value of the prediction result.

10. A system for dynamically monitoring the state of an intelligent power distribution cabinet, applied to a method for dynamically monitoring the state of an intelligent power distribution cabinet according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, used to collect electrical parameters, environmental parameters and physical state parameters of the power distribution cabinet; Data preprocessing module, used to perform denoising, normalization and time series processing on the collected data; Random field modeling module, used to construct the spatial correlation model of the distribution cabinet unit state and the dynamic time Markov random field model; Deep learning prediction module, used to predict the state of the distribution cabinet based on Bayesian deep learning and quantify the prediction uncertainty; The visualization and maintenance strategy module is used to generate a visualization distribution map of the distribution cabinet status and a maintenance priority strategy.

Images