Method and system for monitoring state of primary equipment in new energy power system

Abstract

The invention discloses a primary equipment state monitoring method and system in a new energy power system, and the method comprises the following steps: deploying a multi-mode sensor array on primary equipment, and synchronously collecting a voltage signal, a current signal, a temperature signal, a vibration signal and an environment parameter; inputting the collected data into an edge computing node for preprocessing to obtain a multi-modal signal sequence; variational mode decomposition is carried out to form a multi-dimensional feature vector; inputting a long-short-term memory neural network model containing an attention mechanism, performing training and reasoning by using an AdamW optimizer and a cosine annealing learning rate strategy, and outputting the health degree of equipment; determining the weight of each monitoring index based on an analytic hierarchy process, and dividing the equipment into a plurality of state grades; the fault probability is obtained through fuzzy Petri net reasoning, and a corresponding early warning mechanism is triggered according to a preset threshold value. According to the invention, through multi-dimensional data fusion and intelligent analysis, the accuracy and real-time performance of state monitoring are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of power system operation and maintenance technology, and specifically relates to a method and system for monitoring the status of primary equipment in a new energy power system. Background Technology

[0002] As a crucial component of modern power grids, the primary equipment (such as photovoltaic inverters, converters, and switchgear) of new energy power systems operates under complex and variable electrical and environmental conditions for extended periods. This makes them susceptible to performance degradation and malfunctions due to the combined effects of electrical stress, thermal stress, mechanical vibration, and environmental factors. Failure to monitor equipment operating status in a timely manner can not only impact the efficiency of new energy power generation and the safe and stable operation of the power grid but may also lead to widespread power outages. Therefore, conducting online monitoring and intelligent diagnostics of primary equipment is of paramount importance.

[0003] Existing methods for monitoring the condition of primary equipment mainly rely on single-type sensor data, such as collecting parameters like current, voltage, or temperature, and using thresholds to detect faults. While these methods are simple to implement, they have significant shortcomings: First, they lack sufficient monitoring dimensions and cannot reflect the true operating status of equipment under the coupling of multiple physical quantities; second, threshold-based early warning methods have high false alarm and false negative rates, making it difficult to meet the high-precision, low-latency condition monitoring requirements of new energy equipment; third, traditional manual inspection methods are labor-intensive, lack real-time performance, and cannot adapt to the intelligent operation and maintenance of large-scale new energy power plants.

[0004] To improve the accuracy and intelligence of monitoring, Chinese patent application CN120277594A proposes a multimodal-based method for monitoring the health status of power equipment. This method collects multimodal data such as oil chromatography, temperature, current, and vibration, and performs data cleaning and label correction using expert rules to ensure data reliability. In the modeling stage, an autoencoder is used to reduce feature dimensionality, and a graph model is constructed based on the equipment topology. A graph neural network is then used to achieve early anomaly detection. In the diagnostic stage, a fault mechanism knowledge graph is used for mechanism matching and consistency assessment to achieve interpretable diagnosis. Simultaneously, online incremental and transfer learning mechanisms are introduced to achieve adaptive model updates. This approach significantly improves upon traditional methods in terms of multi-source data fusion and interpretable diagnosis.

[0005] The above-mentioned approach still has certain limitations: First, the method mainly relies on autoencoders and graph neural networks for feature dimensionality reduction and anomaly detection, which does not adequately consider the temporal correlation and dynamic evolution of equipment multimodal signals, making it difficult to accurately predict the trend of equipment health changes; Second, the diagnosis relies on knowledge graph mechanism matching, and when the knowledge graph coverage is insufficient or the rules are incomplete, the real-time performance and accuracy of the diagnosis results are easily affected; Third, the method is still insufficient in terms of fault early warning mechanism and maintenance strategy generation, and lacks proactive operation and maintenance capabilities based on equipment remaining life prediction and intelligent scheduling optimization. Summary of the Invention

[0006] This invention provides a method and system for monitoring the status of primary equipment in a new energy power system, aiming to solve the problems of the real-time performance and accuracy being easily affected and the health trend being difficult to predict in existing technologies.

[0007] To address the aforementioned technical problems, this invention proposes a method for monitoring the status of primary equipment in a new energy power system, comprising the following steps: A multimodal sensor array is deployed on a primary device to simultaneously collect voltage signals, current signals, temperature signals, vibration signals, and environmental parameters; The collected data is input into the edge computing node for preprocessing to obtain a multimodal signal sequence; The multimodal signal sequence is subjected to variational mode decomposition to form a multidimensional feature vector; The feature vector is input into a long short-term memory neural network model that includes an attention mechanism. The model is trained and inferred using the AdamW optimizer and a cosine annealing learning rate strategy, and the health status of the device is output. The weights of each monitoring indicator are determined based on the analytic hierarchy process (AHP), and the membership degree of the health status is calculated using a cloud model. The equipment is then divided into several status levels. The state level, health status, and key features are input into a fuzzy Petri net to infer the failure probability, and the corresponding early warning mechanism is triggered according to a preset threshold.

[0008] Preferably, after triggering an early warning, the status monitoring method uses a genetic algorithm to optimize the scheduling of maintenance resources and generate a corresponding maintenance strategy.

[0009] Preferably, the condition monitoring method further includes estimating the remaining life of the primary equipment by combining the Weibull distribution with the Markov chain.

[0010] Preferably, the preprocessing includes wavelet denoising, moving average filtering, normalization, and outlier detection and interpolation compensation based on the 3σ criterion.

[0011] Preferably, the variational mode decomposition is based on adaptively adjusting the number of mode decompositions K and the penalty factor α according to signal characteristics.

[0012] Preferably, the long short-term memory neural network model includes: The input layer is used to receive the multidimensional feature vector extracted by the variational mode decomposition; At least one layer of long short-term memory unit is used to model the time series of the multidimensional feature vector, and its activation functions include the tanh function and the sigmoid function; The attention layer, located between the long short-term memory unit and the output layer, is used to assign attention weights to the hidden states at each time step and form a context vector by weighted summation to highlight features or time points that are more critical to health prediction. A fully connected layer is used to perform non-linear mapping on the context vector; The output layer is used to output a predicted device health value between 0 and 1.

[0013] Preferably, the attention weights are calculated as follows:

[0014] In the formula, For time steps Attention weights This indicates that the Long Short-Term Memory (LSTM) neural network model is in time step The hidden state, For learnable query vectors, This represents the similarity between the hidden state and the query vector. Represented by natural constant Exponentiation with base 0. This represents the total number of time steps, which is the length of the feature sequence input into the Long Short-Term Memory neural network model.

[0015] Preferably, the step of determining the weights of each monitoring indicator using the analytic hierarchy process includes: The monitoring indicators related to equipment status in the multidimensional feature vector are compared pairwise, and a judgment matrix is ​​constructed using the 1-9 scaling method; Calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix, and normalize the eigenvector to obtain the weight of each monitoring indicator; The consistency index is calculated based on the largest eigenvalue, and the consistency ratio is calculated in combination with the average random consistency index. When the consistency ratio is less than 0.1, the judgment matrix is ​​determined to meet the consistency requirements.

[0016] Preferably, the step of obtaining the fault probability through fuzzy Petri net inference includes: The device status level, health value, and key features are used as fuzzy values ​​in the input database. The fuzzy values ​​are in the range of 0 to 1 to represent the credibility of the corresponding status. Based on the preset arc weights, the fuzzy values ​​of the input library are weighted and summed to obtain the trigger strength of the transition; When the trigger strength is greater than or equal to the transition threshold, update the fuzzy value of the output library; Through cascading reasoning of multiple transitions, the final output is a fuzzy value representing the overall failure probability of the equipment.

[0017] In another aspect, the present invention also proposes a primary equipment condition monitoring system in a new energy power system, the system being used to implement the condition monitoring method as described in the first aspect of the present invention, comprising: A multimodal sensor array is deployed in key parts of primary equipment to synchronously acquire voltage signals, current signals, temperature signals, vibration signals, and environmental parameters. Edge computing nodes, connected to the multimodal sensor array, are used to perform wavelet denoising, moving average filtering, normalization processing, and outlier detection and interpolation compensation based on the 3σ criterion on the collected raw data to obtain a multimodal signal sequence. The feature extraction module is used to perform variational mode decomposition on the multimodal signal sequence, adjust the number of mode decompositions K and the penalty factor α based on adaptive parameters, and form a multidimensional feature vector; The health prediction module includes a long short-term memory neural network model incorporating an attention mechanism, used to receive the multidimensional feature vector and output the device health status; The status grading assessment module is used to determine the weight of each monitoring indicator based on the analytic hierarchy process and to calculate the membership degree of health status by combining the cloud model, thereby classifying the equipment into several status levels. The fault reasoning and early warning module includes a fuzzy Petri net reasoning unit, which takes the state level, health status and key features as input, calculates the fault probability of the device, and triggers the corresponding early warning mechanism when the threshold is exceeded. The maintenance decision module, connected to the fault reasoning and early warning module, is used to optimize the scheduling of maintenance resources based on a genetic algorithm after an early warning is triggered, and to generate corresponding maintenance strategies. The lifespan prediction and model self-learning module is used to estimate the remaining lifespan of equipment by combining Weibull distribution and Markov chain, and to update the monitoring model parameters through online incremental learning.

[0018] Compared with the prior art, the present invention has the following technical effects: 1. The condition monitoring method proposed in this invention can realize multimodal data fusion, intelligent modeling and reasoning, hierarchical early warning and intelligent maintenance decision-making. It effectively overcomes the problems of insufficient monitoring dimensions, high false alarm rate and reliance on human experience in operation and maintenance of traditional methods, and significantly improves the monitoring accuracy and operation and maintenance intelligence level of primary equipment in new energy power systems.

[0019] 2. The state monitoring method proposed in this invention performs wavelet denoising, moving average filtering, normalization processing, and outlier detection and compensation based on the 3σ criterion through edge computing nodes, which effectively improves the real-time performance and reliability of the signal and reduces the interference of noise and outliers on subsequent analysis.

[0020] 3. The condition monitoring method proposed in this invention extracts multidimensional features based on variational mode decomposition and uses a long short-term memory neural network with an attention mechanism to model time-dependent features, which can highlight key time points and key features, and improve the accuracy and stability of equipment health prediction.

[0021] 4. The status monitoring method proposed in this invention combines the analytic hierarchy process (AHP) and cloud model to calculate the membership degree and classify the status of health results, thereby achieving a scientific evaluation through the fusion of multiple indicators and overcoming the high false alarm rate of traditional threshold judgment methods.

[0022] 5. The condition monitoring method proposed in this invention uses fuzzy Petri nets to reason about the condition level, health status and key features, which can handle uncertainty and fuzziness, output more realistic fault probabilities, and construct a hierarchical early warning mechanism to improve the accuracy of fault diagnosis and early warning.

[0023] 6. The status monitoring method proposed in this invention, after the early warning is triggered, combines a genetic algorithm to optimize the scheduling of maintenance resources and generate a refined maintenance strategy, realizing intelligent closed-loop management from fault early warning to maintenance execution, thereby reducing downtime and operation and maintenance costs.

[0024] 7. The condition monitoring method proposed in this invention further combines Weibull distribution and Markov chain to estimate the remaining life of equipment, providing a basis for life prediction for operation and maintenance plans, and improving the reliability and life management level of the system.

[0025] 8. The condition monitoring method proposed in this invention deploys a multi-modal sensor array in key parts of primary equipment to simultaneously collect voltage, current, temperature, vibration and environmental parameters, thereby achieving comprehensive and coordinated monitoring of electrical parameters, mechanical characteristics and environmental factors, avoiding the problem of insufficient monitoring dimensions of traditional single sensors. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the state monitoring method described in this invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0028] Example 1 This embodiment describes a method for monitoring the status of primary equipment in a new energy power system, such as... Figure 1 As shown, it includes the following steps one through six: Step 1: Deploy a multimodal sensor array on the primary equipment to simultaneously collect voltage signals, current signals, temperature signals, vibration signals, and environmental parameters.

[0029] The multimodal sensor array in this embodiment includes a Rogowski coil current sensor, a fiber optic grating temperature sensor, and a MEMS (Micro-Electro-Mechanical Systems) vibration accelerometer.

[0030] Rogowski coil current sensor: used to collect current signals. Rogowski coils have the characteristics of wide bandwidth and fast response, and are suitable for monitoring complex current signals in new energy power systems.

[0031] Fiber Bragg grating temperature sensor: Responsible for collecting temperature parameters. Fiber Bragg grating sensors have advantages such as anti-electromagnetic interference, high measurement accuracy, and good stability, and can adapt to the complex electromagnetic environment in which power equipment is located.

[0032] MEMS vibration accelerometers: used to acquire vibration signals. MEMS (microelectromechanical systems) sensors are small in size, light in weight, and low in cost, making them easy to deploy in critical parts of equipment to monitor vibration characteristics.

[0033] Sensors are deployed in key parts of the equipment. The specific key parts need to be determined based on the structure and operating characteristics of different primary equipment. For example, for photovoltaic inverters, voltage and current sensors can be deployed at their input and output terminals; temperature sensors can be placed close to the surface of power devices (such as IGBT modules); vibration sensors can be installed on the inverter housing or the fixed structure of the core components; and environmental parameter sensors are deployed in appropriate locations around the equipment to accurately reflect the environmental conditions (such as temperature and humidity) of the equipment.

[0034] The following synchronization method is used between sensors: the edge computing node is used as the time reference, and the time of each sensor is synchronized through a unified clock signal (such as GPS clock, Beidou clock or local high-precision clock source); when each sensor collects data, it receives a synchronization trigger signal from the edge computing node to ensure that the data collection operation starts at the same time, thereby realizing the time alignment of multi-sensor data.

[0035] Multimodal signal synchronous acquisition: Under the coordinated control of edge computing nodes, the multimodal sensor array starts data acquisition simultaneously according to the set sampling frequency, ensuring the synchronization of voltage, current, temperature, vibration signals and environmental parameters in the time dimension, and avoiding data deviation caused by acquisition time difference.

[0036] Multimodal signal timestamps are used to add a timestamp to each set of data collected by each sensor, based on a unified time reference provided by the edge computing nodes. The timestamps are accurate to the millisecond or microsecond level, enabling accurate correlation of signal data from different modes during subsequent data preprocessing, feature extraction, and model analysis. This provides precise time dimension information for equipment status assessment and fault diagnosis.

[0037] Step 2: Input the collected data into the edge computing node for preprocessing to obtain a multimodal signal sequence.

[0038] The preprocessing includes wavelet denoising, moving average filtering, normalization, and outlier detection and interpolation compensation based on the 3σ criterion.

[0039] In this embodiment, the collected multimodal sensor data is input into an edge computing node for preprocessing to generate a multimodal signal sequence that can be used for subsequent feature extraction. The preprocessing steps include wavelet denoising, moving average filtering, normalization, and outlier detection and interpolation compensation based on the 3σ criterion.

[0040] Specifically, for vibration signals and high-frequency fluctuation signals of voltage and current, which are usually severely affected by noise, it is preferable to perform wavelet denoising, moving average filtering, and normalization in sequence, combined with outlier detection and interpolation compensation to improve signal quality. For low-frequency or slowly varying signals such as temperature signals and environmental parameters, moving average filtering and normalization are the main methods, while wavelet denoising can be selectively applied depending on the actual noise situation.

[0041] In the wavelet denoising step, Daubechies (e.g., db5) or Symlets (e.g., sym4) wavelet basis functions are preferred because they have good separation effects on abrupt signals and high-frequency noise, making them suitable for vibration signals and current signals containing impact components. The number of decomposition layers is typically 3 to 5, preferably determined based on the signal sampling frequency and noise characteristics: 4 to 5 layers are suitable for higher sampling frequencies, while 3 layers are suitable for lower signal frequencies. A soft thresholding method is used to threshold the detail coefficients during denoising, with the threshold estimated based on the Median Absolute Deviation (MAD) of the first-layer detail coefficients. Mirror expansion or zero-padding is preferred for edge regions to avoid distortion of the edge signal after filtering.

[0042] In the moving average filtering step, an appropriate window length is selected based on different signal characteristics. For example, for a 50Hz power frequency signal with a sampling frequency of 1kHz, the window length can be set to 5 to 20 sampling points. A window length of 10 corresponds to a filtering window of approximately 10ms, which balances smoothing effect and signal response speed. For slowly varying signals such as temperature and environmental signals, the filtering window can be set to 1 to 10 seconds. For edge data involved in the filtering process, mirror expansion or zero-padding methods are preferred.

[0043] In the outlier detection and interpolation compensation steps, a detection method based on the 3σ criterion is adopted, which calculates the local mean μ and standard deviation σ within a sliding window. When data points satisfy the 3σ criterion... When an outlier occurs, it is identified as an anomaly. If the outlier is isolated or its continuous length is less than a preset threshold (e.g., 5 to 20 sampling points), linear interpolation or cubic spline interpolation is used for compensation. If the outlier occurs continuously and exceeds the preset threshold, it is marked as a potential real fault, and no interpolation repair is performed. Instead, it is retained and reported for subsequent diagnostic analysis. For missing data at the beginning or end of the sequence, forward padding, backward padding, or mirror expansion can be used for compensation.

[0044] In the normalization process, the Min-Max normalization method is preferred, which linearly maps the data to the [0,1] interval, as shown in the formula:

[0045] In the formula, The normalized value. The original value, The minimum value of the processed signal. This represents the maximum value of the processed signal. For voltage, current, and temperature signals with a defined physical range, Min-Max normalization can effectively eliminate dimensional differences and avoid adverse effects of amplitude differences between different signals on subsequent model training. For signals without a fixed physical range or requiring standard normalization, the Z-score normalization method can also be used.

[0046] Through the above preprocessing steps, noise reduction, smoothing, anomaly correction and dimensional unification of multimodal signals can be achieved at edge computing nodes, resulting in stable and highly comparable multimodal signal sequences, laying the foundation for subsequent feature extraction based on variational mode decomposition.

[0047] Step 3: Perform Variational Mode Decomposition (VMD) on the multimodal signal sequence to form multidimensional feature vectors.

[0048] In this embodiment, the variational mode decomposition is based on adaptively adjusting the number of mode decompositions K and the penalty factor α according to signal characteristics.

[0049] Specifically, this step performs variational mode decomposition on the multimodal signal sequence obtained after preprocessing in step two to extract the intrinsic mode functions (IMFs) of the signal and form a multidimensional feature vector.

[0050] First, the multimodal signal sequence is input into the VMD model. Then, using the Lagrange multiplier method and alternating direction multiplier method (ADMM) iterative optimization framework, variational decomposition of the signal is performed under bandwidth constraints. Unlike existing methods that fix the number of mode decompositions K and the penalty factor α, this embodiment introduces a parameter adaptation mechanism: In the initial stage of iteration, the appropriate number of modes K is dynamically estimated by analyzing the spectral distribution and energy concentration of the signal. If the spectral distribution is sparse and the energy is concentrated in a few frequency bands, the value of K is decreased; if the signal spectrum is complex and the energy is dispersed, the value of K is increased. The penalty factor α is adjusted based on the convergence of the modal bandwidth constraint and the energy overlap between adjacent modes: when the mode separation is incomplete, α is increased appropriately; when the decomposition results are too scattered, α is decreased appropriately to obtain a convergent and stable decomposition result.

[0051] After completing the adaptive parameter adjustment, VMD outputs multiple narrowband intrinsic mode components, each of which can effectively characterize the signal features in different frequency bands. Subsequently, the time-domain features (mean, variance, kurtosis, skewness, etc.) and frequency-domain features (dominant frequency, bandwidth, energy spectral density, etc.) of each mode component are calculated, and the above features are combined to construct a multi-dimensional feature vector, which serves as the input to the deep learning model in the subsequent step four.

[0052] Through the above settings, this embodiment can achieve adaptive optimization of mode decomposition parameters under different operating conditions, improve the accuracy and robustness of feature extraction, and provide a more reliable data foundation for equipment health assessment.

[0053] In this embodiment, the feature vector extracted by VMD has 28 dimensions, specifically including kurtosis, variance, spectral entropy, mean, peak value, peak-to-peak value, root mean square (RMS), skewness, waveform factor, peak factor, impulse factor, kurtosis factor, margin factor, centroid frequency, mean square frequency, frequency variance, spectral peak value, spectral kurtosis, spectral entropy, and frequency band energy ratio (such as the proportion of energy in each frequency band to the total energy); as well as the above time-domain or frequency-domain features extracted based on each modal component after VMD decomposition (such as the mean and root mean square of each mode).

[0054] Step four: Input the feature vector into a long short-term memory neural network model that includes an attention mechanism, and use the AdamW optimizer and cosine annealing learning rate strategy for training and inference to output the health status of the device.

[0055] The long short-term memory neural network model includes an input layer, at least one long short-term memory unit, an attention layer, a fully connected layer, and an output layer.

[0056] The input layer is used to receive the multidimensional feature vector extracted by the variational mode decomposition. In this embodiment, it is 28-dimensional, that is, the number of nodes in the input layer is 28.

[0057] At least one layer of long short-term memory unit, with the number of nodes in each layer set to 64, 128 or 256, is used to model the time series of the multidimensional feature vector and capture the time dependency of the feature sequence. Its activation functions include the tanh function (for cell state update) and the sigmoid function (for control of forget gate, input gate and output gate).

[0058] The attention layer, located between the long short-term memory unit and the output layer, is used to assign attention weights to the hidden states at each time step and form a context vector by weighted summation to highlight features or time points that are more critical to health prediction.

[0059] A fully connected layer is used to perform non-linear mapping on the context vector. The context vector is non-linearly mapped through 1 to 2 fully connected layers. The number of nodes in the fully connected layer is gradually reduced, for example from 64 to 32. The ReLU function can be used as the activation function to enhance the expressive power of the model and alleviate the gradient vanishing problem.

[0060] The output layer consists of one node, which outputs a predicted device health value between 0 and 1, and its activation function is the sigmoid function.

[0061] The above attention weights are calculated as follows:

[0062] In the formula, For time steps Attention weights This indicates that the Long Short-Term Memory (LSTM) neural network model is in time step The hidden state, For learnable query vectors, This represents the similarity between the hidden state and the query vector. Represented by natural constant Exponentiation with base 0. This represents the total number of time steps, i.e., the length of the feature sequence input into the Long Short-Term Memory (LSTM) neural network model. Through this attention mechanism, the model can automatically focus on features or time segments that contribute more to health prediction.

[0063] During model training, the training data comes from the historical operating data of primary equipment in the new energy power system, including: historical multimodal sensor data (a 28-dimensional feature sequence after preprocessing in step two and feature extraction in step three); and corresponding operation and maintenance records for the equipment (such as fault occurrence time, maintenance records, and status assessment results). Equipment health labels are continuous values ​​between 0 and 1, marked by equipment operation and maintenance records and expert assessment results. The labeling criteria include: if the equipment is in normal operation (no fault records), the health label is close to 1; if the equipment shows signs of a fault (such as abnormal parameters but no shutdown), the label decreases with the degree of abnormality (e.g., 0.7-0.3); if the equipment fails (shutdown or maintenance), the label is 0; intermediate state label values ​​can be generated by interpolation using expert assessments or historical fault data.

[0064] The AdamW optimizer is used to reduce overfitting, and a cosine annealing learning rate strategy is combined to improve model convergence. The loss function is either mean squared error (MSE) or mean absolute error (MAE).

[0065] Step 5: Determine the weight of each monitoring indicator based on the analytic hierarchy process (AHP), calculate the membership degree of the health status using the cloud model, and classify the equipment into several status levels.

[0066] The steps for determining the weights of each monitoring indicator using the analytic hierarchy process include: The monitoring indicators related to equipment status in the multidimensional feature vector are compared pairwise, and a judgment matrix is ​​constructed using the 1-9 scaling method; Calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix, and normalize the eigenvector to obtain the weight of each monitoring indicator; The consistency index is calculated based on the largest eigenvalue, and the consistency ratio is calculated in combination with the average random consistency index. When the consistency ratio is less than 0.1, the judgment matrix is ​​determined to meet the consistency requirements.

[0067] Specifically, firstly, pairwise comparisons are performed on the monitoring indicators closely related to the equipment status from the multidimensional feature vectors extracted in step three, and a judgment matrix is ​​constructed using the 1-9 scaling method. For example, taking a photovoltaic inverter as an example, current fluctuation variance, peak temperature, vibration kurtosis, and ambient humidity are selected as core indicators to form a 4×4 judgment matrix. Subsequently, the maximum eigenvalue and corresponding eigenvector of this judgment matrix are calculated, and the eigenvector is normalized to obtain the weights of each monitoring indicator. In the above example, the obtained weights are: current fluctuation variance 0.25, peak temperature 0.52, vibration kurtosis 0.15, and ambient humidity 0.08.

[0068] To ensure the validity of the weighting results, a consistency check is performed on the judgment matrix. The calculation method is as follows: consistency index... , where n is the matrix order; look up the table to obtain the average random consistency index RI, when n=4, RI=0.90; calculate the consistency ratio CR=CI / RI, when CR<0.1, the judgment matrix is ​​considered to meet the consistency requirements, otherwise the judgment matrix needs to be readjusted by experts.

[0069] After obtaining the weights of each monitoring indicator, a weighted fusion method is used to calculate the overall health value of the equipment, and the health value is input into the cloud model. The cloud model uses a cloud generator to generate membership degrees for different state levels, and its cloud droplet distribution parameters include expectation Ex, entropy En, and hyperentropy He, which can effectively handle the fuzziness and uncertainty in equipment state classification.

[0070] Based on the calculation results of the cloud model, the membership values ​​of the equipment in four levels—"healthy," "sub-healthy," "prodromal," and "fault"—are obtained, and the final state level is determined by the principle of maximum membership. For example, when the equipment health level is 0.72, the cloud model calculates that its membership degree for the "sub-healthy" state is the highest, so the equipment is classified as "sub-healthy."

[0071] In practical implementation, device status levels can be classified according to the following criteria: Health: Health score ≥ 0.8, all monitored indicators are within the normal threshold (e.g., temperature ≤ 60℃, vibration kurtosis ≤ 3). Sub-health: 0.5≤health level<0.8, with some indicators slightly exceeding the standard (such as temperature 60-70℃, vibration kurtosis 3-5); Early signs of failure: 0.2 ≤ health level < 0.5, multiple indicators are significantly abnormal (such as temperature 70-85℃, vibration kurtosis 5-8); Fault: Health level <0.2, key indicators seriously exceed limits (such as temperature >85℃, vibration kurtosis >8).

[0072] Through the above method, this embodiment can obtain the weights of each monitoring indicator using the analytic hierarchy process and combine it with the cloud model to realize the classification of status levels. This not only improves the scientificity and rationality of equipment status assessment, but also provides an accurate classification basis for the subsequent fault probability reasoning and early warning triggering in step six.

[0073] Step six: Input the state level, health status and key features into a fuzzy Petri net, infer the failure probability, and trigger the corresponding early warning mechanism according to a preset threshold.

[0074] The steps for obtaining the fault probability through fuzzy Petri net inference include: The device status level, health value, and key features are used as fuzzy values ​​in the input database. The fuzzy values ​​are in the range of 0 to 1 to represent the credibility of the corresponding status. Based on the preset arc weights, the fuzzy values ​​of the input library are weighted and summed to obtain the trigger strength of the transition; When the trigger strength is greater than or equal to the transition threshold, update the fuzzy value of the output library; Through cascading reasoning of multiple transitions, the final output is a fuzzy value representing the overall failure probability of the equipment.

[0075] Specifically, this step inputs the status level, health status and key features obtained in step five into a fuzzy Petri net, infers the failure probability of the equipment, triggers the corresponding early warning mechanism according to the preset threshold, and generates a maintenance strategy.

[0076] The fuzzy Petri net includes places, transitions, and weighted arcs. Places represent states or conditions related to equipment failure, such as "temperature exceeds threshold," "abnormal vibration kurtosis," or "equipment in a sub-healthy state." Transitions represent causal relationships between states, such as "temperature exceeds threshold and vibration is abnormal → increased failure probability." Arcs connect places and transitions, and fuzzy weights are labeled on the arcs. The weights range from 0 to 1, representing the degree of influence of different conditions on the result.

[0077] In this embodiment, the input data for the fuzzy Petri net includes: the equipment status level and its membership degree output in step five, the equipment health value and its trend predicted in step four, the key feature parameters extracted in step three (such as vibration kurtosis anomalies and temperature exceeding threshold amplitude), and fault mode and feature association data from the historical fault database. The above input data is normalized and used as the fuzzy value for each location, where the fuzzy value is a decimal between 0 and 1, representing the confidence level of the corresponding status. For example, when the equipment temperature reaches 75℃, the fuzzy value for the location "temperature exceeding threshold" can be assigned 0.8.

[0078] During the inference process, the fuzzy Petri net calculates the transition trigger strength based on the fuzzy values ​​of the input places and the on-arc weights. When the transition trigger strength is greater than or equal to a set threshold, the fuzzy values ​​of the output places are updated, forming a new state result. Through cascaded inference of multi-level transitions, a fuzzy value representing the probability of equipment failure is finally obtained. For example, in the application of photovoltaic inverters, when the confidence level of the "temperature exceeds threshold" place is 0.8 and the confidence level of the "abnormal vibration" place is 0.7, the confidence level of the overall failure probability of the equipment obtained by inference is 0.85, that is, the confidence level of a failure probability greater than 80% is 85%.

[0079] Based on the obtained failure probability, a three-level early warning mechanism is triggered: When the equipment is in a "sub-healthy" state and the probability of failure is between 10% and 30%, a level one warning is triggered, indicating that there are potential hidden dangers in the equipment but that it does not affect operation at the moment. When the equipment is in a "predictable failure" state and the probability of failure is between 30% and 60%, a level 2 warning is triggered, indicating that the risk of equipment failure has increased significantly and an inspection should be arranged within 24-48 hours. When the equipment is in a state close to or reaches a "fault" state and the probability of failure is greater than 60%, a level 3 warning is triggered, indicating that the equipment may fail in the short term and needs to be shut down for maintenance immediately.

[0080] After an early warning is triggered, a corresponding maintenance strategy is generated based on the equipment location and the current maintenance resource status. Specifically, the maintenance strategy for a Level 1 warning is to increase monitoring frequency, clean heat dissipation components, or remotely adjust parameters; the maintenance strategy for a Level 2 warning is planned shutdown for inspection and replacement of vulnerable components; and the maintenance strategy for a Level 3 warning is emergency shutdown for repair, activation of backup equipment, and allocation of emergency maintenance resources.

[0081] Through the above method, this embodiment can achieve fault probability inference under uncertain conditions by using fuzzy Petri nets based on equipment status level assessment.

[0082] In other embodiments of the present invention, after triggering an early warning, the status monitoring method optimizes the scheduling of maintenance resources based on a genetic algorithm to generate a corresponding maintenance strategy. The maintenance strategy is generated using a resource scheduling optimization method based on a genetic algorithm, aiming to minimize downtime and reduce maintenance costs, and outputs a specific maintenance plan, including maintenance time, personnel arrangements, and spare parts allocation sequence.

[0083] When the equipment failure probability obtained by the fuzzy Petri net inference in step six exceeds a preset threshold and triggers an early warning, the system further executes maintenance resource scheduling optimization based on a genetic algorithm to generate a specific maintenance strategy. The process includes input data preparation, setting the scheduling objective function, genetic algorithm encoding and initial population generation, fitness function design, genetic operations, iteration and convergence, and maintenance strategy output.

[0084] The system calls the operations and maintenance database based on the warning level to obtain the following input information: Equipment status data: including the status level, health status, and key anomaly indicators output in step five; Maintenance resource data includes a list of available maintenance personnel on site (professional skills, shift schedule), current spare parts inventory (model, quantity, location), and availability of maintenance tools; Operation and maintenance constraints include equipment downtime limits, maintenance budget, and personnel attendance constraints.

[0085] The scheduling objective function aims to minimize equipment downtime and reduce total maintenance costs. The objective function is defined as follows:

[0086] In the formula, Indicates the duration of equipment downtime. This indicates the total cost of maintenance (including spare parts and labor costs). These are the weighting coefficients.

[0087] Individuals in the maintenance scheme are represented by integer encoding, where chromosomes contain personnel allocation genes, spare parts allocation genes, and maintenance time genes. The initial population is generated through heuristic rules, such as prioritizing the allocation of maintenance personnel with high professional skill matching and prioritizing the selection of spare parts with sufficient on-site inventory, to ensure the feasibility of the initial solution.

[0088] The fitness value of an individual is calculated based on the objective function mentioned above. At the same time, a constraint penalty mechanism is introduced to impose a penalty value on schemes that exceed the staffing time limit, have insufficient spare parts inventory, or have maintenance window time constraints.

[0089] During genetic manipulation, roulette wheel selection is used to select individuals with high fitness; partial mapped crossover (PMX) is used to recombine personnel and spare parts in the offspring; and individual maintenance personnel or spare parts allocation is randomly adjusted to enhance population diversity.

[0090] The genetic algorithm gradually optimizes the scheduling scheme during multiple generations of iteration. When the fitness function does not improve significantly within several consecutive generations or reaches the maximum number of iterations, the algorithm terminates and outputs the optimal maintenance strategy.

[0091] The optimization results are output in the form of a maintenance plan, including: Maintenance schedule (e.g., "Complete the shutdown maintenance within 24 hours"); Personnel allocation plan (e.g., "dispatch two electrical engineers and one mechanical maintenance technician"); Spare parts allocation sequence (e.g., "prioritize the allocation of inverter fans → then replace capacitor components").

[0092] The maintenance plan is simultaneously pushed to the operation and maintenance management system to guide on-site execution.

[0093] In other embodiments of the invention, the condition monitoring method further includes estimating the remaining lifetime of primary equipment by combining Weibull distribution and Markov chain.

[0094] Specifically, after triggering the fault probability reasoning and early warning mechanism described in step six, the system further combines Weibull distribution and Markov chain to estimate the remaining lifespan of the primary equipment, providing maintenance personnel with a more forward-looking decision-making basis. The process is as follows: Prepare data, including historical equipment operation data: long-term monitoring sequences of multi-modal signals such as voltage, current, temperature, and vibration; fault database: containing failure time, failure mode, and maintenance records of similar equipment; current health status: equipment health and trend information output from step four, and status levels classified in step five.

[0095] For Weibull distribution lifetime modeling, equipment failure time data is selected, and a two-parameter Weibull distribution is fitted using the Maximum Likelihood Estimation (MLE) method to obtain the shape parameters. and scale parameters The distribution function is used to obtain the time... The failure probability at time t is distributed as follows:

[0096] Then based on the current cumulative running time of the equipment The conditional reliability function is calculated to obtain the probability of the device's survival at future times.

[0097] In the formula, For equipment uptime (predicted time), i.e., the time at which we want to estimate the equipment's uptime. The probability that the time-tracking device is still functioning properly. Given the cumulative operating time of the current equipment (observation baseline), the significance of the conditional reliability function lies in: knowing that the equipment has been operating safely for... Continue operating into the future What is the probability of it? Let be the reliability function, representing the time since the device's arrival. The probability of maintaining normal operation. This represents the conditional reliability function, which means "when the equipment has been operating normally for..." Under the premise that it continues to operate normally until The probability of.

[0098] Next, Markov chain state transition modeling is performed, dividing the equipment health state into four discrete states: healthy (S1), sub-healthy (S2), fault precursor (S3), and fault (S4); based on historical operation and fault databases, the transition frequencies between different states are statistically analyzed, and a state transition probability matrix is ​​constructed.

[0099] Among them, S4 is the absorption state (i.e., equipment failure). In the above formula, four states are first defined: healthy state S1, sub-healthy state S2, fault precursor state S3, and fault state (absorption state, once entered, it will not exit) S4. This represents the probability of transitioning from state Si to state Sj, for example: This indicates the probability that a device will transition from a healthy state to a sub-healthy state. This indicates the probability that a device will progress from a sub-healthy state to a precursor to a malfunction. This indicates the probability that a device will directly progress from a fault precursor to a fault.

[0100] Based on Markov chain state transition modeling, the remaining lifetime is estimated. First, the Weibull distribution model is used to calculate the survival probability curves of the equipment at different future time points. Then, combined with the current health status of the equipment, the Markov chain is used to predict its future state transition path and the expected time to reach the fault state S4. Combining the results of both, the expected value and confidence interval of the remaining lifetime are obtained.

[0101] The system ultimately outputs an estimated remaining lifespan of the equipment (e.g., "estimated remaining lifespan is 850 hours, confidence level 90%), and links it with the early warning mechanism. If the remaining lifespan is higher than the threshold (e.g., >10,000 hours), then only preventative maintenance will be performed; If the remaining lifespan is in the middle range (3000–10000 hours), a planned maintenance task will be generated. If the remaining lifetime is below the threshold (<3000 hours), emergency maintenance and resource scheduling optimization will be triggered.

[0102] By using the above method, the lifetime statistics of Weibull distribution and the state transition characteristics of Markov chain are further introduced on the basis of fuzzy Petri net inference, so as to realize the dynamic estimation of the remaining lifetime of primary equipment, and provide long-term prediction and refined decision support for operation and maintenance management.

[0103] Example 2 This embodiment is a primary equipment condition monitoring system in a new energy power system. The system is used to implement the condition monitoring method described in Embodiment 1, including: A multimodal sensor array is deployed in key parts of primary equipment to synchronously acquire voltage signals, current signals, temperature signals, vibration signals, and environmental parameters. Edge computing nodes, connected to the multimodal sensor array, are used to perform wavelet denoising, moving average filtering, normalization processing, and outlier detection and interpolation compensation based on the 3σ criterion on the collected raw data to obtain a multimodal signal sequence. The feature extraction module is used to perform variational mode decomposition on the multimodal signal sequence, adjust the number of mode decompositions K and the penalty factor α based on adaptive parameters, and form a multidimensional feature vector; The health prediction module includes a long short-term memory neural network model incorporating an attention mechanism, used to receive the multidimensional feature vector and output the device health status; The status grading assessment module is used to determine the weight of each monitoring indicator based on the analytic hierarchy process and to calculate the membership degree of health status by combining the cloud model, thereby classifying the equipment into several status levels. The fault reasoning and early warning module includes a fuzzy Petri net reasoning unit, which takes the state level, health status and key features as input, calculates the fault probability of the device, and triggers the corresponding early warning mechanism when the threshold is exceeded. The maintenance decision module, connected to the fault reasoning and early warning module, is used to optimize the scheduling of maintenance resources based on a genetic algorithm after an early warning is triggered, and to generate corresponding maintenance strategies. The lifespan prediction and model self-learning module is used to estimate the remaining lifespan of equipment by combining Weibull distribution and Markov chain, and to update the monitoring model parameters through online incremental learning.

[0104] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for monitoring the condition of primary equipment in a new energy power system, characterized in that, Includes the following steps: A multimodal sensor array is deployed on a primary device to simultaneously collect voltage signals, current signals, temperature signals, vibration signals, and environmental parameters; The collected data is input into the edge computing node for preprocessing to obtain a multimodal signal sequence; The multimodal signal sequence is subjected to variational mode decomposition to form a multidimensional feature vector; The feature vector is input into a long short-term memory neural network model that includes an attention mechanism. The model is trained and inferred using the AdamW optimizer and a cosine annealing learning rate strategy, and the health status of the device is output. The weights of each monitoring indicator are determined based on the analytic hierarchy process (AHP), and the membership degree of the health status is calculated using a cloud model. The equipment is then divided into several status levels. The state level, health status, and key features are input into a fuzzy Petri net to infer the failure probability, and the corresponding early warning mechanism is triggered according to a preset threshold.

2. The method according to claim 1, characterized in that, After triggering an early warning, the condition monitoring method uses a genetic algorithm to optimize the scheduling of maintenance resources and generate corresponding maintenance strategies.

3. The method according to claim 1, characterized in that, The condition monitoring method also includes estimating the remaining life of primary equipment by combining Weibull distribution and Markov chain.

4. The method according to claim 1, characterized in that, The preprocessing includes wavelet denoising, moving average filtering, normalization, and outlier detection and interpolation compensation based on the 3σ criterion.

5. The method according to claim 1, characterized in that, The variational mode decomposition is based on adaptively adjusting the number of mode decompositions K and the penalty factor α according to signal characteristics.

6. The method according to claim 1, characterized in that, The long short-term memory neural network model includes: The input layer is used to receive the multidimensional feature vector extracted by the variational mode decomposition; At least one layer of long short-term memory unit is used to model the time series of the multidimensional feature vector, and its activation functions include the tanh function and the sigmoid function; The attention layer, located between the long short-term memory unit and the output layer, is used to assign attention weights to the hidden states at each time step and form a context vector by weighted summation to highlight features or time points that are more critical to health prediction. A fully connected layer is used to perform non-linear mapping on the context vector; The output layer is used to output a predicted device health value between 0 and 1.

7. The method according to claim 6, characterized in that, The method for calculating the attention weight is as follows: In the formula, For time step Attention weights This indicates that the Long Short-Term Memory (LSTM) neural network model is in time step The hidden state, For learnable query vectors, This represents the similarity between the hidden state and the query vector. Represented by natural constant Exponentiation with base 0. This represents the total number of time steps, which is the length of the feature sequence input into the Long Short-Term Memory neural network model.

8. The method according to claim 1, characterized in that, The steps for determining the weights of each monitoring indicator using the analytic hierarchy process include: The monitoring indicators related to equipment status in the multidimensional feature vector are compared pairwise, and a judgment matrix is ​​constructed using the 1-9 scaling method; Calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix, and normalize the eigenvector to obtain the weight of each monitoring indicator; The consistency index is calculated based on the largest eigenvalue, and the consistency ratio is calculated in combination with the average random consistency index. When the consistency ratio is less than 0.1, the judgment matrix is ​​determined to meet the consistency requirements.

9. The method according to claim 1, characterized in that, The steps for obtaining the fault probability through fuzzy Petri net inference include: The device status level, health value, and key features are used as fuzzy values ​​in the input database. The fuzzy values ​​are in the range of 0 to 1 to represent the credibility of the corresponding status. Based on the preset arc weights, the fuzzy values ​​of the input library are weighted and summed to obtain the trigger strength of the transition; When the trigger strength is greater than or equal to the transition threshold, update the fuzzy value of the output library; Through cascading reasoning of multiple transitions, the final output is a fuzzy value representing the overall failure probability of the equipment.

10. A primary equipment condition monitoring system in a new energy power system, characterized in that, The system is used to implement the condition monitoring method as described in any one of claims 1-9, including: A multimodal sensor array is deployed in key parts of primary equipment to synchronously acquire voltage signals, current signals, temperature signals, vibration signals, and environmental parameters. Edge computing nodes, connected to the multimodal sensor array, are used to perform wavelet denoising, moving average filtering, normalization processing, and outlier detection and interpolation compensation based on the 3σ criterion on the collected raw data to obtain a multimodal signal sequence. The feature extraction module is used to perform variational mode decomposition on the multimodal signal sequence, adjust the number of mode decompositions K and the penalty factor α based on adaptive parameters, and form a multidimensional feature vector; The health prediction module includes a long short-term memory neural network model incorporating an attention mechanism, used to receive the multidimensional feature vector and output the device health status; The status grading assessment module is used to determine the weight of each monitoring indicator based on the analytic hierarchy process and to calculate the membership degree of health status by combining the cloud model, thereby classifying the equipment into several status levels. The fault reasoning and early warning module includes a fuzzy Petri net reasoning unit, which takes the state level, health status and key features as input, calculates the fault probability of the device, and triggers the corresponding early warning mechanism when the threshold is exceeded. The maintenance decision module, connected to the fault reasoning and early warning module, is used to optimize the scheduling of maintenance resources based on a genetic algorithm after an early warning is triggered, and to generate corresponding maintenance strategies. The lifespan prediction and model self-learning module is used to estimate the remaining lifespan of equipment by combining Weibull distribution and Markov chain, and to update the monitoring model parameters through online incremental learning.

Images