Training method of high-voltage switch cabinet defect prediction model and defect prediction method

Abstract

The application discloses a high-voltage switch cabinet defect prediction model training method and a defect prediction method. The training method comprises the following steps: acquiring observation data of various sensors installed on a high-voltage switch cabinet, and constructing a data set based on the observation data; using a hidden semi-Markov model to model a high-voltage switch cabinet state transition process and a state maintenance process to obtain a defect prediction model; based on the data set, learning a state sequence of the defect prediction model online and unsupervised, and optimizing parameters of the defect prediction model according to a state sequence fitting effect to obtain a trained switch cabinet defect prediction model. The defect prediction model trained by the application can efficiently learn a switch cabinet state transition process, and can be learned online and unsupervised, without relying on a large amount of labeled data, thereby saving cost.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment defect diagnosis technology, specifically to a training method and a defect prediction method for a high-voltage switchgear defect prediction model. Background Technology

[0002] The safe and stable operation of switchgear is directly related to the equipment safety and power supply reliability of the power grid. Switchgear failures can cause fires, main transformer tripping, and other phenomena, making it difficult to meet the equipment reliability requirements of the rapid development of the power grid. To improve the stability of the power grid system, it is necessary to have a comprehensive understanding of the health status of switchgear and to detect and eliminate potential faults as early as possible.

[0003] High-voltage switchgear is the most widely used and numerous type of switchgear in power systems, playing a role in control, protection, and operation during power generation, transmission, distribution, energy conversion, and consumption. High-voltage switchgear is mainly composed of insulating and conductive materials. Common causes of power system failures include component aging and wear, improper design and installation of parts, moisture, contamination, and malfunctions in mechanisms and secondary equipment. Common high-voltage switchgear failures include malfunctions such as failure to operate or accidental operation, opening and closing failures, insulation failures, current-carrying failures, and external force failures.

[0004] Traditional methods for defect prediction in high-voltage switchgear rely primarily on manual labor or intelligent auxiliary equipment such as robots. However, these methods suffer from limitations such as limited monitoring data, significant background interference variations, and uncertain time intervals, making it difficult to meet the reliability demands of rapidly evolving power grids. With the development of power sensor technology, various sensors are designed to detect subtle changes in the operating status of switchgear components. Furthermore, supported by big data analytics, the collection, storage, and analysis of diverse and heterogeneous sensor data, such as temperature data, partial discharge data, and mechanical fault data, have become possible. Defect prediction then uses this sensor data to perform real-time assessments of the operating status of each component, promptly predicting potential sensor defects that could lead to switchgear failures based on data distribution changes and the occurrence of outliers. This is crucial for preventing significant economic losses due to power outages caused by defect deterioration. For example, patent application CN115900816A proposes a high-voltage switchgear defect prediction method based on multi-source heterogeneous data. This method utilizes multi-source heterogeneous data acquired from temperature sensors, partial discharge sensors, and mechanical fault characteristic sensors, combined with a neural network model, to predict high-voltage switchgear defects.

[0005] However, the prediction of defects in high-voltage switchgear still faces the following difficulties: (1) Difficulty in modeling the state transition process: During the operation of high-voltage switchgear, it often remains in the same state for a long time. Traditional machine learning models are difficult to model the state transition process of high-voltage switchgear. (2) High cost of expert manual annotation: In order to ensure the accuracy of high-voltage switchgear defect prediction, a large amount of sensor data needs to be collected for model training. Traditional machine learning models are supervised training methods, which require manual annotation of defect states, which is too costly. Supervised training methods are not suitable for actual use needs.

[0006] Therefore, how to model the correlation between the state transition process of high-voltage switchgear and the change process of sensor observation data, accurately perceive data defects, and carry out efficient and accurate perception and reliable assessment analysis of the defect status of intelligent high-voltage switchgear is an urgent problem to be solved. Summary of the Invention

[0007] The technical problem to be solved by this invention is how to accurately perceive data defects and improve the reliability of switchgear defect prediction.

[0008] The present invention solves the above-mentioned technical problems through the following technical means:

[0009] This invention proposes a training method for a defect prediction model of high-voltage switchgear, the method comprising:

[0010] Acquire observation data from various sensors installed on the high-voltage switchgear, and construct a dataset based on the observation data;

[0011] A hidden semi-Markov model is used to model the state transition process and state holding process of the high-voltage switchgear to obtain a defect prediction model.

[0012] Based on the dataset, the state sequence of the defect prediction model is learned online in an unsupervised manner, and the parameters of the defect prediction model are optimized according to the fitting effect of the state sequence to obtain a trained switch cabinet defect prediction model.

[0013] Furthermore, the various sensors installed on the switchgear include temperature sensors, partial discharge sensors, mechanical characteristic sensors, gas sensors, and opening and closing coil current monitoring devices.

[0014] The sensor defect types include insulation defects, structural defects, abnormal heating defects, secondary equipment abnormal defects, and composite defects containing at least two defect categories.

[0015] Furthermore, the temperature sensor is arranged on the inner wall of the handcart, the moving contact, and the contact arm;

[0016] The partial discharge sensors include ultrasonic partial discharge sensors, TEV transient ground wave partial discharge sensors, and UHF ultra-high frequency partial discharge sensors.

[0017] The opening and closing coil current monitoring device and the mechanical characteristic sensor are respectively installed on the secondary circuit of the instrument room and the circuit breaker operating shaft.

[0018] Furthermore, the process of modeling the state transition and state holding processes of the high-voltage switchgear using a hidden semi-Markov model to obtain a defect prediction model includes:

[0019] For time t, let the state of the high-voltage switchgear be s. t ∈S, state s t Expected duration is The sensor's observation data is e t S represents the set of states of the high-voltage switchgear. It is a set of positive integers;

[0020] The state transition process of the high-voltage switchgear is represented as follows:

[0021]

[0022] In the formula, δ is the indicator function, δ(s) t+1 ,s t )=1{s t+1 =s t};f θ (s t+1 |s t ) indicates when d t When = 0, the defect prediction model transitions to state s at time t+1. t+1 The probability distribution is given by a state transition probability matrix θ. f describe;

[0023] The process of maintaining the state of a high-voltage switchgear is represented as follows:

[0024]

[0025] In the formula, δ is the indicator function, δ(d t+1 ,d t -1)=1{d t+1 =d t -1};d t+1 |s t+1 This indicates that the defect prediction model is in state s t+1 The expected duration follows a Poisson distribution h. θ h θ The parameter θ h Indicates the expected duration of each state;

[0026] Observational data e t The state s at the current moment t Decision: et ~g θ (e t |s t ,e t-k:t-1 ), g θ The parameter is θ g Gaussian mixture model; s t Indicates the current status of the high-voltage switchgear; e t-k:t-1 This represents the observation data from the first k time steps; e t |s t ,e t-k:t-1 Indicates a given s t and e t-k:t-1 The current observation data under the given conditions; ~ indicates that it follows a probability distribution.

[0027] Further, the step of learning the state sequence of the defect prediction model online in unsupervised learning based on the dataset, and optimizing the parameters of the defect prediction model according to the fitting effect of the state sequence to obtain a trained switchgear defect prediction model includes:

[0028] Based on the dataset, the probability of the state sequence and model parameters of the high-voltage switchgear is estimated using the particle filter algorithm.

[0029] The model parameters are resampled based on the probabilities of the model parameters to obtain new model parameters;

[0030] The new parameters of the model were adjusted using the Markov chain Monte Carlo method based on the Metropolis Hastings kernel to obtain a well-trained switchgear defect prediction model.

[0031] Furthermore, the step of estimating the state sequence and model parameters of the high-voltage switchgear using a particle filter algorithm based on the dataset includes:

[0032] 1) Randomly initialize N sets of parameters θ for the defect prediction model. 1 ,θ 2 ,…,θ N θ i =(θ f ,θ h ,θ g ,θ π ), θ π The initial state distribution is π θ θ g For the parameters of the Gaussian mixture model, θ h h is a Poisson distribution θ The parameter, θ f This is the state transition probability matrix;

[0033] 2) Using normalized particle weights To probabilistically resample M particles, for each resampled particle, sample the state at time t+1 from the proposed distribution to obtain the state of the new particle;

[0034] 3) Calculate the non-normalized weights of each new particle, and normalize the non-normalized weights of each new particle to obtain the normalized particle weights of each particle.

[0035] 4) Update the log probabilities of the model parameters based on the initial log probabilities of the model parameters;

[0036] Repeat steps 2) to 4) until one particle is resampled. Use the state of this particle as the estimated result of the state sequence of the high-voltage switchgear, and calculate the probability of the model parameters based on the logarithmic probability of the model parameters at this time.

[0037] Furthermore, the resampling of model parameters based on the probabilities of the model parameters to obtain new parameters of the model includes:

[0038] Based on the probabilities of the model parameters, calculate the normalized probabilities of the model parameters;

[0039] The model parameters and state sequences are resampled based on the normalized probabilities of the model parameters to obtain new model parameters.

[0040] Furthermore, the method of adjusting the new parameters of the model using the Markov chain Monte Carlo method based on the Metropolis Hastings kernel to obtain the trained switchgear defect prediction model includes:

[0041] For each state transition of the Markov chain of the defect prediction model, the Metropolis Hastings kernel randomly samples candidate parameters from the new parameters of the model according to the proposal distribution;

[0042] The dataset is iterated through in a loop, and the particle filter algorithm is used to estimate the state sequence of the high-voltage switchgear and the probability of new parameters of the model.

[0043] The candidate parameters are received with the minimum probability among the probabilities of the new parameters of the model, and an estimated state sequence for each candidate parameter is sampled.

[0044] When traversing each observation data segment in the dataset, the expected value of the model parameters at this time is calculated as the parameters of the trained switchgear defect prediction model.

[0045] Furthermore, after obtaining the trained switchgear defect prediction model, the method further includes:

[0046] The defect prediction model is initialized using the expected values ​​of the model parameters;

[0047] Based on the test dataset, the defect prediction model is estimated online using the particle filter algorithm, and the expected particle state is used as the estimated state.

[0048] Furthermore, this invention proposes a high-voltage switchgear defect prediction method based on a hidden semi-Markov model, used to predict high-voltage switchgear defects using a well-trained switchgear defect prediction model as described above, comprising:

[0049] Real-time acquisition of observation data from various sensors installed on the high-voltage switchgear at the current moment;

[0050] A particle filter algorithm is used to randomly sample a number of particles, and a proposed distribution is used to sample the state of each particle.

[0051] The normalized particle weights of each particle are calculated based on the state of each particle and the observation data at the current moment.

[0052] Resampling is performed based on the normalized particle weights of each particle, and the expected particle state is calculated as the current state of the high-voltage switchgear.

[0053] The advantages of this invention are:

[0054] (1) This invention uses a hidden semi-Markov model to model a high-voltage switchgear defect prediction model, uses a random variable of expected duration to simulate the state holding process, and uses a random variable of state to simulate the state transition process. The expected duration of the state is introduced into the state variable. Compared with the traditional hidden Markov model, it is more robust to states with longer durations, and thus can efficiently learn the state transition process of the high-voltage switchgear. The model can be learned online without supervision, without the need to label the sensor data in advance, saving costs.

[0055] (2) When representing sensor data, the observation data of the sensor is collected multiple times as feature vectors to consider the data change process, thereby improving the accuracy of model training.

[0056] (3) By using SMC 2 The algorithm learns the state sequence of the defect prediction model and accurately fits the model parameters online, ensuring that the accuracy of the model parameters is only affected by Monte Carlo error. SMC is one such component. 2 The Quadratic Sequential Monte Carlo algorithm is an improvement on Sequential Monte Carlo (SMC).

[0057] (4) Since the Hidden Semi-Markov Model introduces the expected duration of the state into the state variable, the computational complexity increases dramatically. By applying the particle filter algorithm in the early stage of the training data sequence to estimate the state sequence of the high-voltage switchgear and the probability of the model parameters, the time complexity can be reduced and the inference efficiency of the Hidden Semi-Markov Model can be greatly improved.

[0058] (5) Based on the trained defect prediction model, the particle filtering algorithm can be applied to achieve efficient and accurate online defect status prediction based on real-time sensor data.

[0059] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0060] Figure 1 This is a flowchart of a training method for a high-voltage switchgear defect prediction model proposed in an embodiment of the present invention;

[0061] Figure 2 This is a structural diagram of the defect prediction model modeled using a hidden semi-Markov model in an embodiment of the present invention;

[0062] Figure 3 This is a block diagram illustrating the training principle of the defect prediction model in this embodiment of the invention.

[0063] Figure 4 This is a flowchart of a high-voltage switchgear defect prediction method based on a hidden semi-Markov model proposed in an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] like Figure 1 As shown, the first embodiment of the present invention proposes a training method for a defect prediction model of a high-voltage switchgear, the method comprising:

[0066] S10. Obtain observation data from various sensors installed on the high-voltage switchgear, and construct a dataset based on the observation data;

[0067] S20. Use a hidden semi-Markov model to model the state transition process and state holding process of the high-voltage switchgear to obtain a defect prediction model.

[0068] S30. Based on the dataset, the state sequence of the defect prediction model is learned online in an unsupervised manner, and the parameters of the defect prediction model are optimized according to the fitting effect of the state sequence to obtain a trained switch cabinet defect prediction model.

[0069] This embodiment uses a hidden semi-Markov model to model the expected duration of a state. Compared with the traditional hidden Markov model, it is more robust to states with longer durations, thus enabling efficient learning of the state transition process of a high-voltage switchgear. The expected duration random variable is used to simulate the state holding process, and the state random variable is used to simulate the state transition process. This separation of the expected duration of the state and the state random variable improves the model's flexibility. Furthermore, the model can be learned online without supervision, without the need for pre-labeling of sensor data or reliance on dense labeled data, greatly reducing training costs and facilitating practical applications.

[0070] In one embodiment, in step S10, the various sensors installed on the switchgear include temperature sensors, partial discharge sensors, mechanical characteristic sensors, gas sensors, and opening and closing coil current monitoring devices.

[0071] The sensor defect types include insulation defects, structural defects, abnormal heating defects, secondary equipment abnormal defects, and composite defects containing at least two defect categories.

[0072] Specifically, this embodiment pre-constructs a sensor database comprising multiple sub-databases, then collects observation data from the sensors, performs calibration tests on the collected data, and stores the observation data returned by the sensors into multiple sub-databases according to different sensor types. By analyzing the data based on different sensor types, problems such as missing data and inaccurate data sample collection are resolved, and it is convenient to update the data in each sub-database.

[0073] Furthermore, the specific defect types of the sensor are as follows:

[0074] (1) Insulation defects: The air clearance inside the cabinet does not meet the requirements, the cabinet is damp or condensed, the wall bushing and contact box are not equipped with shielding wires, the conductor tip causes the local electric field to be too large, and the cable terminal joint is poorly made, etc.

[0075] (2) Defects in the mechanism: jamming of the mechanism, loose or deformed parts, burnt-out opening and closing coils, etc.

[0076] (3) Abnormal heating defects: poor contact of disconnecting switch blade, poor contact of circuit breaker contacts, insufficient depth of handcart push-in, and loose bolts in the current guiding part.

[0077] (4) Abnormal defects of secondary equipment: poor contact of auxiliary switch, loose micro switch and failure to switch normally, operating voltage of opening and closing coil does not meet the requirements, rectifier module breakdown, etc.

[0078] It should be understood that a composite defect is a composite defect that contains at least two different types of defects, including but not limited to insulation defects, structural defects, abnormal heating defects, and abnormal secondary equipment defects.

[0079] In one embodiment, the temperature sensor is arranged on the inner wall of the handcart, the moving contact, and the contact arm;

[0080] The partial discharge sensors include ultrasonic partial discharge sensors, TEV transient ground wave partial discharge sensors, and UHF ultra-high frequency partial discharge sensors.

[0081] The opening and closing coil current monitoring device and the mechanical characteristic sensor are respectively installed on the secondary circuit of the instrument room and the circuit breaker operating shaft.

[0082] Specifically, sensor defects are typically a process; therefore, this embodiment uses sensor data collected over the past ten times as feature vectors to account for the data change process when designing the sensor data representation. The sensor structure data representation at each time point is as follows:

[0083] (1) Temperature sensor: A high-sensitivity temperature sensor is fixed at three points (T1, T2, T3) on the inner wall of the handcart. In addition, it is also fixed with screws to the moving contact (T4) and the contact arm (T5) to indirectly measure the temperature rise of the contact.

[0084] (2) Partial discharge sensor: The partial discharge signal in the switch cabinet is measured by ultrasonic partial discharge sensor (S1), TEV transient ground wave partial discharge sensor (S2) and UHF ultra-high frequency partial discharge sensor (S3).

[0085] (3) Mechanical characteristic sensor: The opening and closing coil current monitoring device and the switch mechanical characteristic sensor are installed on the secondary circuit and the circuit breaker operating shaft in the instrument room. The opening and closing coil current (A) is collected from the opening and closing coil current monitoring device, and the opening and closing speed (v) and opening and closing stroke (s) are collected from the switch mechanical characteristic sensor.

[0086] (4) Gas sensor: When the insulation components inside the high voltage switchgear are damaged due to partial discharge and abnormal heating, the insulation material will deteriorate and decompose due to high temperature. A gas sensor unit is used to detect the gas content (V) in real time.

[0087] The observation data collected using the aforementioned sensors includes: real-time monitoring of the switching devices in the intelligent high-voltage switchgear by mechanical characteristic sensors; capturing effective waveforms of opening and closing from the opening and closing current detection device to calculate the opening and closing time; and collecting opening and closing speeds and strokes from the switch mechanical characteristic sensors. Finally, the sensor data is recorded and stored in the corresponding mechanical characteristic sensor database, including fields such as measured switch characteristic time, calculated opening and closing speed, measured opening / closing stroke, and initial opening / closing speed.

[0088] In one embodiment, it is defined as follows: Figure 2 The hidden semi-Markov model modeling shown includes defining the number of high-voltage switchgear states and the state transition matrix, defining the probability mass function for the duration of each state, and defining the sensor data observation probability distribution using the current high-voltage switchgear state and sensor data from previous times as input. The defined high-voltage switchgear states include normal states and various defect states. That is, step S20 above includes the following sub-steps:

[0089] S21. For time t, let the state of the high-voltage switchgear be s. t ∈S, state s t Expected duration is The sensor's observation data is e t S represents the set of states of the high-voltage switchgear. It is a set of positive integers;

[0090] The indicator function δ is defined as follows:

[0091] δ(x,y)=1{x=y}

[0092] S22. The state transition process of the high-voltage switchgear is represented as follows:

[0093]

[0094] In the formula, δ is the indicator function, δ(s) t+1 ,s t )=1{s t+1 =s t};f θ (s t+1 |s t ) indicates when d t When = 0, the defect prediction model transitions to state s at time t+1. t+1 The probability distribution is given by a state transition probability matrix. describe, Represents the set of real numbers;

[0095] S23. The process of maintaining the state of the high-voltage switchgear is represented as follows:

[0096]

[0097] In the formula, δ is the indicator function, δ(d t+1 ,d t -1)=1{d t+1 =d t -1};d t+1 |s t+1 This indicates that the defect prediction model is in state s t+1 The expected duration follows a Poisson distribution h. θ h θ parameters Indicates the expected duration of each state;

[0098] S24, Observational Data e t The state s at the current moment t Decision: e t ~g θ (e t |s t ,e t-k:t-1 ), g θ The parameter is θ g Gaussian mixture model; s t Indicates the current status of the high-voltage switchgear; e t-k:t-1 This represents the observation data from the first k time steps; e t |s t ,e t-k:t-1 Indicates a given s t and e t-k:t-1 The current observation data under the given conditions; ~ indicates that it follows a probability distribution.

[0099] The initial state distribution of the constructed defect prediction model is denoted as π. θ All parameters are denoted as θ = (θ f ,θ h ,θ g ,θ π ), θ π The initial state distribution is π θ .

[0100] Then, N hidden semi-Markov models are randomly initialized, as follows:

[0101] (1) Randomly initialize N sets of parameters θ 1 ,θ 2 ,…,θ N Constructing a hidden semi-Markov model: For each hidden semi-Markov model, M particles are randomly initialized, and the initial state distribution is denoted as... The initial duration is denoted as

[0102] (2) Select an appropriate proposal distribution q(s) t+1 |s t ,e t+1 ), calculate the nonnormalized weights of each particle:

[0103]

[0104] Normalize the particle weights:

[0105]

[0106] In the formula, The nonnormalized weight of the m-th particle at the initial time; For a given Under the given conditions, the probability that the observed data at the initial time is e1; The initial state is The probability of; The proposed probability distribution; The normalized weight of the m-th particle at the initial time; The nonnormalized weight of the i-th particle at the initial time; s t The state at time t; s t+1 The state at time t+1; e t+1 The data is the observation data at time t+1;

[0107] (3) Calculate the initial logarithmic probability of the parameters:

[0108]

[0109] In the formula, Let be the initial logarithmic probability of the model parameters, and log be the logarithmic function.

[0110] In one embodiment, such as Figure 3 As shown, step S30: Based on the dataset, the state sequence of the defect prediction model is learned online in an unsupervised manner, and the parameters of the defect prediction model are optimized according to the fitting effect of the state sequence to obtain the trained switch cabinet defect prediction model, including the following steps:

[0111] S31. Based on the dataset, the probability of the state sequence and model parameters of the high-voltage switchgear is estimated using the particle filter algorithm;

[0112] S32. Resample the model parameters based on the probabilities of the model parameters to obtain new model parameters;

[0113] S33. Use the Markov chain Monte Carlo method based on Metropolis Hastings kernel to adjust the new parameters of the model and obtain a trained switch cabinet defect prediction model.

[0114] It should be noted that this embodiment uses SMC 2 The algorithm learns the state sequence of the hidden semi-Markov model without relying on dense labeled data, and accurately fits the model parameters online, ensuring that the accuracy of the model parameters is only affected by Monte Carlo error.

[0115] Because Hidden Semi-Markov Models (HSMs) introduce the expected duration of a state into the state variable, their computational complexity increases dramatically. By applying a particle filter algorithm early in the training data sequence to estimate the state sequence of the high-voltage switchgear and the probabilities of the model parameters, the time complexity can be reduced, significantly improving the inference efficiency of the HSM. For example, the computational complexity of a traditional Markov model is... Where K represents the number of states and T represents the length of the sensor data sequence. However, because the hidden semi-Markov model introduces the expected duration of each state into the state variable, the computational complexity increases dramatically. Here, D represents the longest duration of the state. In practical applications, D is much larger than K, and may even be positive infinity. However, using a particle filter algorithm can reduce the time complexity to... Where N represents the number of particles, which is usually taken as T / 2.

[0116] In one embodiment, step S31: Based on the dataset, the probability of estimating the state sequence and model parameters of the high-voltage switchgear using a particle filter algorithm, specifically includes the following sub-steps:

[0117] 1) Randomly initialize N sets of parameters θ for the defect prediction model. 1 ,θ 2 ,…,θ N θ i =(θ f ,θ h ,θ g ,θ π ), θ π The initial state distribution is π θ θ g For the parameters of the Gaussian mixture model, θ h h is a Poisson distribution θ The parameter, θ f This is the state transition probability matrix;

[0118] 2) Using normalized particle weights To probabilistically resample M particles, for each resampled particle, sample the state at time t+1 from the proposed distribution to obtain the state of the new particle;

[0119] It should be noted that, Let be the normalized weight of the m-th particle. The weights of all particles are represented by an array.

[0120] Specifically, using normalized particle weights To probabilistically resample M particles, their states are denoted as:

[0121]

[0122] For each resampled particle, sample the state at time t+1 from the proposed distribution:

[0123]

[0124] The new particle state is then represented as:

[0125]

[0126] In the formula, This is the state sequence of the m-th particle at time t. This is the state sequence of the m-th particle at time t+1. The state of the m-th particle at time t+1; e 1:t+1 This is the sequence of observation data from time t+1.

[0127] 3) Calculate the non-normalized weights of each new particle, and normalize the non-normalized weights of each new particle to obtain the normalized particle weights of each particle.

[0128] Specifically, the nonnormalized weights of each new particle are calculated as follows:

[0129]

[0130] Normalize the non-normalized weights of each particle:

[0131]

[0132] In the formula, The nonnormalized weight of the m-th particle at time t+1; For a given Under the given conditions, the observed data at time t+1 is e t+1 The probability of; for arrive The state transition probability; Let be the probability distribution of the proposal.

[0133] 4) Update the log probabilities of the model parameters based on the initial log probabilities of the model parameters;

[0134] Specifically, the logarithmic probabilities of the updated model parameters are:

[0135]

[0136] In the formula, The logarithmic probability of the updated model parameters; The initial logarithmic probability of the parameter; The weight is the non-normalized weight of the i-th particle at time t+1.

[0137] 5) Repeat steps 2) to 4) until one particle is resampled. Use the state of the particle as the estimated result of the state sequence of the high-voltage switchgear, and calculate the probability of the model parameters based on the logarithmic probability of the model parameters at this time.

[0138] In this embodiment, when executing T0 times repeatedly, the normalized particle weights are used. One particle is sampled for probability, and its state is used as the estimate of the predicted state of the defect prediction model by the particle filtering algorithm. Simultaneously, parameter probabilities are calculated based on the logarithmic likelihood of the parameters to be used in subsequent steps to evaluate parameter quality. The parameter probability formula is expressed as:

[0139]

[0140] In the formula, Indicates the probability of the parameter; exp represents the logarithmic probability of the parameter; exp represents an exponential function with base e.

[0141] It should be noted that this embodiment resamples the particle states and estimates the new particle weights using a particle filtering algorithm, thereby converging the particle state distribution to the true distribution and realizing state estimation of the hidden semi-Markov model based on observation data. Furthermore, this step updates the logarithmic probabilities of the parameters, facilitating the training and optimization of the hidden semi-Markov model.

[0142] In one embodiment, step S32: resampling the model parameters based on the probabilities of the model parameters to obtain new parameters of the model includes the following steps:

[0143] S321. Calculate the normalized probability of the model parameters based on the probability of the model parameters;

[0144] S322. Resample the model parameters and state sequences according to the normalized probabilities of the model parameters to obtain new model parameters.

[0145] Specifically, the normalized probability of the calculated parameters in this embodiment is:

[0146]

[0147] With normalized parameter probability Let N sets of parameters be resampled with probability, denoted as:

[0148] θ n n = 1, 2, ..., N

[0149] In the formula, Let be the normalized weight of the nth particle at time t; For θ n The parameter probability; For θ i The parameter probability.

[0150] This embodiment resamples the parameters and state sequences based on the normalized probabilities of the parameters, and then uses Metropolis Hastings sampling to obtain new parameters. The training dataset is iterated through repeatedly. Specifically, step S33 involves adjusting the model's new parameters using a Markov chain Monte Carlo method based on the Metropolis Hastings kernel to obtain a trained switchgear defect prediction model, including the following steps:

[0151] S331. For each state transition of the Markov chain of the defect prediction model, the Metropolis Hastings kernel randomly samples candidate parameters from the new parameters of the model according to the proposal distribution.

[0152] Specifically, a Markov chain Monte Carlo method based on the Metropolis Hastings kernel is used to sample new parameters. For each state transition of the Markov chain, the Metropolis Hastings kernel first samples the proposed distribution q′(θ). ＊ |θ) Randomly sample candidate parameters θ ＊ .

[0153] S332. Iterate through the dataset and use the particle filter algorithm to estimate the state sequence of the high-voltage switchgear and the probability of new parameters of the model.

[0154] S333. The candidate parameters are received with the minimum probability among the probabilities of the new parameters of the model, and the estimated state sequence of each candidate parameter is sampled.

[0155] Specifically, the particle filter algorithm described above is used to estimate... With probability min(1,a(θ) ＊ ,θ)) accept θ ＊ Otherwise, keep the original parameter θ unchanged, where:

[0156]

[0157] In the formula, For θ ＊ The parameter probability; p θ (e 1:t) represents the parametric probability of θ; q′(·|·) represents the proposal distribution.

[0158] S334. After traversing each segment of observation data in the dataset, calculate the expected value of the model parameters at this time as the parameters of the trained switchgear defect prediction model.

[0159] Specifically, the particle filtering algorithm described above is used to sample each new parameter θ obtained from the Markov chain Monte Carlo method. ＊ Estimate the state sequence:

[0160]

[0161] In the formula, Let be the state sequence of the nth particle at time t.

[0162] Repeat steps S332 to S334 above until the observed sequence e is reached. 1:T The entire process is traversed, and the expected values ​​of the model parameters at this point are calculated as the parameters of the trained switchgear defect prediction model.

[0163] Furthermore, the constructed dataset is divided into training and testing datasets according to a set ratio. For each observation sequence e in the training dataset... 1:T Using the methods described in steps S332 to S334, several sets of parameters are obtained, and all parameters are denoted as:

[0164] θ n n = 1, 2, ..., N.

[0165] The corresponding probability is denoted as:

[0166]

[0167] Calculate the normalized probability of the parameters:

[0168]

[0169] Calculate the expected value of the parameters based on the normalized probability:

[0170]

[0171] In one embodiment, after obtaining the trained switchgear defect prediction model, the method further includes the following steps:

[0172] The defect prediction model is initialized using the expected values ​​of the model parameters;

[0173] Based on the test dataset, the defect prediction model is estimated online using the particle filter algorithm, and the expected particle state is used as the estimated state.

[0174] It should be noted that after training the model, it is necessary to test the model. The Hidden Semi-Markov Model is initialized using the expected parameters, and the state of the Hidden Semi-Markov Model is estimated online using the particle filter algorithm. The expected particle state is then used as the estimated state.

[0175]

[0176] In the formula, s is the normalized probability of the nth particle; m Let be the state distribution of the nth particle.

[0177] This embodiment uses a hidden semi-Markov model to model the state maintenance and transition processes of high-voltage switchgear, enabling efficient and accurate model parameter learning without expert annotation. Simultaneously, the particle filtering algorithm used in the model prediction process ensures the real-time performance and accuracy of the prediction results. Compared to traditional high-voltage switchgear defect prediction methods, the accuracy is improved by 5.3%.

[0178] In addition, such as Figure 4 As shown, the second embodiment of the present invention proposes a high-voltage switchgear defect prediction method based on a hidden semi-Markov model, which is used to predict high-voltage switchgear defects using a switchgear defect prediction model trained as described in the first embodiment above, and includes the following steps:

[0179] S1. Real-time acquisition of observation data from various sensors installed on the high-voltage switchgear at the current moment;

[0180] S2. Randomly sample several particles using a particle filtering algorithm, and use a proposed distribution to sample the state of each particle.

[0181] S3. Calculate the normalized particle weight of each particle based on the state of each particle and the observation data at the current moment.

[0182] S4. Resample according to the normalized particle weights of each particle, and calculate the expected particle state as the current state of the high-voltage switchgear.

[0183] This embodiment is based on a trained defect prediction model and applies a particle filtering algorithm. It can achieve efficient and accurate online defect state prediction based on real-time sensor data. By using the particle filtering algorithm to estimate the state of the hidden semi-Markov model, the inference latency of the hidden semi-Markov model is greatly reduced, and the real-time performance of the system is improved.

[0184] It should be noted that other embodiments or implementation methods of the defect prediction model used in the high-voltage switchgear defect prediction method based on the hidden semi-Markov model described in this invention can refer to the above-described method embodiment one, and will not be repeated here.

[0185] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0186] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0187] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A training method for a defect prediction model of a high-voltage switchgear, characterized in that, The method includes: Acquire observation data from various sensors installed on the high-voltage switchgear, and construct a dataset based on the observation data; A hidden semi-Markov model is used to model the state transition process and state holding process of the high-voltage switchgear to obtain a defect prediction model. Based on the dataset, the state sequence of the defect prediction model is learned online in an unsupervised manner, and the parameters of the defect prediction model are optimized according to the fitting effect of the state sequence to obtain a trained switchgear defect prediction model. The process includes: estimating the state sequence of the high-voltage switchgear and the probability of model parameters using a particle filter algorithm based on the dataset; resampling the model parameters based on the probability of the model parameters to obtain new model parameters; and adjusting the new model parameters using a Markov chain Monte Carlo method based on the Metropolis Hastings kernel to obtain a trained switchgear defect prediction model. Based on the dataset, the probability of estimating the state sequence and model parameters of the high-voltage switchgear using a particle filter algorithm includes: 1) random initialization. The parameters of the defect prediction model described in the group , , Initial state distribution , The parameters are for the Gaussian mixture model. Poisson distribution The parameters, 1) State transition probability matrix; 2) Normalized particle weights For probability resampling For each resampled particle, sample from the proposed distribution. 1) Determine the state of the new particle at each time step; 2) Calculate the non-normalized weights of each new particle and normalize the non-normalized weights of each new particle to obtain the normalized particle weights of each particle; 3) Update the logarithmic probability of the model parameters based on the initial logarithmic probability of the model parameters; Repeat steps 2) to 4) until one particle is resampled, and use the state of that particle as the estimated result of the state sequence of the high-voltage switchgear, and calculate the probability of the model parameters based on the logarithmic probability of the model parameters at this time; The step of adjusting the new parameters of the model using the Markov chain Monte Carlo method based on the Metropolis Hastings kernel to obtain a trained switchgear defect prediction model includes: for each state transition of the Markov chain of the defect prediction model, the Metropolis Hastings kernel randomly samples candidate parameters from the new parameters of the model according to the proposal distribution; iterates through the dataset and uses a particle filter algorithm to estimate the state sequence of the high-voltage switchgear and the probability of the new parameters of the model; receives the candidate parameters with the minimum probability among the probabilities of the new parameters of the model, and samples the estimated state sequence of each candidate parameter; after traversing each segment of observation data in the dataset, calculates the expected value of the model parameters at this time as the parameters of the trained switchgear defect prediction model.

2. The training method for the high-voltage switchgear defect prediction model as described in claim 1, characterized in that, The various sensors installed on the switchgear include temperature sensors, partial discharge sensors, mechanical characteristic sensors, gas sensors, and opening and closing coil current monitoring devices. The sensor defect types include insulation defects, structural defects, abnormal heating defects, secondary equipment abnormal defects, and composite defects containing at least two defect categories.

3. The training method for the high-voltage switchgear defect prediction model as described in claim 2, characterized in that, The temperature sensor is arranged on the inner wall of the handcart, the moving contact, and the contact arm; The partial discharge sensors include ultrasonic partial discharge sensors, TEV transient ground wave partial discharge sensors, and UHF ultra-high frequency partial discharge sensors. The opening and closing coil current monitoring device and the mechanical characteristic sensor are respectively installed on the secondary circuit of the instrument room and the circuit breaker operating shaft.

4. The training method for the high-voltage switchgear defect prediction model as described in claim 1, characterized in that, The method of using a hidden semi-Markov model to model the state transition and state holding processes of the high-voltage switchgear yields a defect prediction model, including: For time Assume the state of the high-voltage switchgear is as follows: ,state Expected duration is The sensor's observation data is , This is a set of states for high-voltage switchgear. It is a set of positive integers; The state transition process of the high-voltage switchgear is represented as follows: In the formula, For indicator functions, ; Indicates when At that time, the defect prediction model in Time to transition to state The probability distribution is given by a state transition probability matrix. describe; The process of maintaining the state of a high-voltage switchgear is represented as follows: In the formula, For indicator functions, ; This indicates that the defect prediction model is in state The expected duration follows a Poisson distribution. , parameters Indicates the expected duration of each state; Observational data From the current state Decide: , The parameter is Gaussian mixture model; This indicates the current status of the high-voltage switchgear. Indicates the preceding Observational data at each moment; Indicates a given and Current observation data under the given conditions; It follows a probability distribution.

5. The training method for the high-voltage switchgear defect prediction model as described in claim 1, characterized in that, The resampling of model parameters based on probabilities to obtain new model parameters includes: Based on the probabilities of the model parameters, calculate the normalized probabilities of the model parameters; The model parameters and state sequences are resampled based on the normalized probabilities of the model parameters to obtain new model parameters.

6. The training method for the high-voltage switchgear defect prediction model as described in claim 1, characterized in that, After obtaining the trained switchgear defect prediction model, the method further includes: The defect prediction model is initialized using the expected values ​​of the model parameters; Based on the test dataset, the defect prediction model is estimated online using the particle filter algorithm, and the expected particle state is used as the estimated state.

7. A method for predicting defects in high-voltage switchgear based on a hidden semi-Markov model, characterized in that, For predicting defects in high-voltage switchgear using a switchgear defect prediction model trained according to any one of claims 1 to 6, the method includes: Real-time acquisition of observation data from various sensors installed on the high-voltage switchgear at the current moment; A particle filter algorithm is used to randomly sample a number of particles, and a proposed distribution is used to sample the state of each particle. The normalized particle weights of each particle are calculated based on the state of each particle and the observation data at the current moment. Resampling is performed based on the normalized particle weights of each particle, and the expected particle state is calculated as the current state of the high-voltage switchgear.

Images