Power equipment state monitoring method based on adaptive neural network

Abstract

The invention discloses a power equipment state monitoring method based on an adaptive neural network. The method comprises the following steps: S1, collecting multi-source time sequence data such as voltage, current, temperature, harmonic wave and vibration, and constructing an input data set; s2, performing missing completion, exception elimination, normalization and trend smoothing processing on the input data to generate standardized operation data; s3, constructing a neural slice network model containing a jump slice module, and extracting fusion state features; s4, inputting standardized data, and generating an initial state classification and risk scoring result; s5, jointly optimizing the spline control point, the path weight and the activation threshold by adopting a spectrum ant oscillation optimization algorithm; s6, re-evaluating the state of the equipment by using the optimization model, and generating classification, score and trend information; and S7, accessing an evaluation result to a power operation and maintenance system to realize display and early warning. According to the method, the fault early warning accuracy and the operation and maintenance response efficiency are remarkably improved while the state of the power equipment is accurately evaluated.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of electric power equipment, and in particular to a method for monitoring the state of electric power equipment based on an adaptive neural network. Background Art

[0002] In existing power systems, equipment condition monitoring, as a crucial foundation for ensuring the safe and stable operation of the power grid, has become a core component of intelligent O&M for transmission, distribution, and consumer terminals. Traditional condition monitoring methods primarily rely on manual inspections, scheduled maintenance, and alarm mechanisms based on fixed thresholds. These methods suffer from coarse detection granularity, significant response delays, and limited diagnostic accuracy, making them incapable of meeting the demands for refined identification of equipment operating status and predictive maintenance in complex operating scenarios. With the application of IoT technology, multi-source sensors are widely deployed in power equipment, enabling real-time collection of multi-dimensional operating data such as temperature, voltage, current, harmonics, and vibration, providing a data foundation for digital monitoring. However, relying solely on raw data collection and simple rule-based judgments remains insufficient to fully explore the underlying patterns in power equipment's operating behavior.

[0003] In recent years, deep learning technology has been gradually introduced into the field of power equipment status assessment, enabling the fitting and prediction of complex nonlinear relationships by constructing neural network models. However, existing methods mostly use static neural networks and lack the ability to dynamically adjust the model structure to different equipment and operating conditions. This results in insufficient model generalization performance and makes it difficult to adapt to the diversity and volatility of equipment status. Furthermore, most current models focus solely on the accuracy of prediction results, neglecting the joint modeling of indicators such as failure risk and trend evolution. This makes it difficult for prediction results to fully reflect the health status of equipment, limiting the model's guiding value in actual operation and maintenance scenarios.

[0004] When it comes to model optimization, existing strategies often rely on standard gradient descent methods such as SGD or Adam. These methods are prone to falling into local optima in models with complex structures, large parameter spaces, and significant non-convexity, and lack global search capabilities. Furthermore, the structural optimization process is often disconnected from the model training process, making it impossible to effectively jointly optimize network parameters such as spline function control points and path connection methods, hindering overall model performance.

[0005] Based on the above problems, the existing power equipment condition monitoring technology has technical bottlenecks such as poor structural adaptability, incomplete indicator coverage, and limited optimization strategies. There is an urgent need for an intelligent condition monitoring method that has structural adaptability, supports multi-dimensional state joint evaluation, and integrates efficient structural optimization algorithms to improve the accuracy, adaptability, and practicality of equipment monitoring.

[0006] Therefore, how to provide a power equipment status monitoring method based on an adaptive neural network is a problem that those skilled in the art urgently need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a method for monitoring the status of power equipment based on an adaptive neural network. The present invention fully integrates the neural slice network structure, the jump slice module and the spectrum ant oscillation optimization algorithm, and describes in detail the standardized processing of multi-source time series data, structural adaptive modeling and fault risk assessment process. It has the advantages of strong structural adjustability, high state recognition accuracy and excellent trend prediction ability.

[0008] According to an embodiment of the present invention, a method for monitoring the state of power equipment based on an adaptive neural network includes the following steps:

[0009] S1. Collect the timing signals of power equipment, perform time alignment and format unification, and construct the input data set;

[0010] S2. Perform missing value filling, outlier removal, normalization and trend smoothing on the input data set to obtain standardized data;

[0011] S3. Construct a neural slice network model including a skip-slice module. The skip-slice module is composed of several segments of spline function sub-networks. Combined with a path selection control unit, it realizes dynamic path activation based on input features. Different paths correspond to feature modeling accuracy at different scales. Multi-scale outputs are fused through a coupling mechanism to enhance state representation capabilities.

[0012] S4, inputting the standardized data into the neural slice network model to obtain the initial state evaluation results;

[0013] S5. Optimizing the structural parameters of the neural slice network model using a spectrum ant oscillation optimization algorithm, wherein the optimization includes jointly updating the spline function coefficients, jump path connection weights, and path activation thresholds based on a spectrum perturbation mechanism and an ant colony search mechanism;

[0014] S6. Reprocess the standardized data with the optimized model and output the operating status classification, fault risk score and trend prediction results to form comprehensive status assessment information;

[0015] S7. Transmit the comprehensive status assessment information to the power operation and maintenance system to realize status display, fault warning and health score output.

[0016] Optionally, the time alignment and format unification performed by S1 include: setting a unified time sampling interval for the voltage signal, current signal, temperature signal, harmonic signal and vibration signal respectively, reconstructing the time axis of data with different sampling frequencies by interpolation, and generating a data matrix with a unified time step; uniformly processing the data formats of various signals, converting the original signal into a fixed-length numerical sequence or vector form, encoding it in a unified data structure and storing it in the input data set.

[0017] Optionally, the S2 performs missing value filling, outlier removal, normalization and trend smoothing on the input data set, including: constructing each type of time series signal in the unified input data set as an equally spaced time series, and using the interpolation method to fill in the missing data at non-boundary positions; using the sliding window statistical method to detect outlier data that exceeds the normal fluctuation range and remove it; after outlier removal and missing value filling, normalizing each type of time series signal according to the maximum and minimum values ​​in the entire time series range to unify the feature value range; performing trend smoothing on the normalized time series signal based on the exponential sliding average method to suppress instantaneous fluctuations.

[0018] Optionally, the S3 specifically includes:

[0019] S31, construct the standardized running data into an input sequence X={x1,x2,…,x T}, where the input feature vector x at each moment t , as the input sequence of the neural slice network model;

[0020] S32, based on the input sequence X, construct a hierarchical representation through multiple spline function sub-networks, each spline function sub-network is used to represent the upper layer input Perform piecewise cubic spline transform and output the current layer representation Among them, l represents the number of network layers, and the control point set of the spline function subnetwork is recorded as Where m is the number of slices per layer;

[0021] S33, set a jump path set P between different depth layers = {p i,j}, where p i,j Represents the jump path from layer i to layer j, and defines the activation variable a for each jump path i,j ∈{0,1};

[0022] S34, construct a path selection control unit, based on the change rate Δx of the input sequence at time step t t =x t -x t-1 Calculate the path activation probability π i,j , based on the maximum probability strategy to activate the corresponding jump path, that is, when πi,j >ρ, set a i,j =1, where ρ is the preset threshold;

[0023] S35, perform several scale spline coupling fusions on the output representations of all activated paths and the current main path representation, and construct a fusion representation sequence H = {h1,h2,…,h T}, describing the time evolution characteristics of the power equipment status.

[0024] Optionally, the S4 specifically includes:

[0025] S41, construct the standardized running data into an input sequence X={x1,x2,…,x T}, where x t is the multi-source feature vector at time step t, which is input into the neural slice network model containing the skip-slice module;

[0026] S42: In the neural slice network model, the input sequence X is processed by multi-segment spline function sub-network and skip path selection in sequence to obtain a fusion state representation sequence H = {h1, h2, ..., h T}, where h t is the state feature representation vector at time step t;

[0027] S43, input the fusion state representation sequence H into the state evaluation output module, and generate the initial state evaluation result Y={(y1,r1),(y2,r2),…,(y T ,r T )}, where y t represents the running status classification label at time step t, r t Indicates the corresponding fault risk score value.

[0028] Optionally, the S5 specifically includes:

[0029] S51, based on the initial state evaluation result Y={(y1,r1),(y2,r2),…,(y T ,r T )} and the standard answer to construct the loss function L, where y t Represents the running status classification label, r t represents the fault risk score value, and the loss function L considers both classification error and regression error;

[0030] S52. Define the search parameter space of the spectral ant oscillation optimization algorithm, including the control point set C = {c1, c2, ..., c m}、Skip path connection weight set W={w i,j} and the path activation threshold set Θ = {θ i,j};

[0031] S53. Initialize multiple ant colony individuals, each individual represents a parameter vector v = [C, W, Θ], and introduce a parameter oscillation mechanism based on the spectral perturbation function to perturb the individuals to initialize the perturbation, thereby enhancing the globality of the parameter search;

[0032] S54, calculating the model loss value L under the current parameter combination for each generation of ant colony individuals, updating the pheromone matrix and path evaluation function according to the loss result, and performing weight enhancement on the high-quality paths;

[0033] S55. After the maximum number of iterations or the convergence threshold condition is met, the control point set C corresponding to the individual that minimizes the loss function is selected. * , connection weight set W * With the activation threshold set Θ * , used to update the structural parameter configuration of the neural slice network model.

[0034] Optionally, defining the search parameter space of the spectral ant oscillation optimization algorithm in S52 includes: constructing a parameter vector v = [C, W, Θ], wherein the spline function control point set C = {c1, c2, ..., c m} represents the node position and number of the piecewise cubic spline function in each layer of the neural slice network, and the jump path connection weight set W = {w i,j} represents the jump path weight between layer i and layer j, and the path activation threshold set Θ = {θ i,j} represents the threshold parameter that controls the activation state of the jump path from the i-th layer to the j-th layer; the parameter vector v is used as the search individual representation of the spectral ant oscillation optimization algorithm, and the search range is limited to the continuous feasible space within the preset parameter boundary, which serves as the basis for subsequent ant colony initialization and perturbation operations.

[0035] Optionally, the optimized neural slice network model in S6 will reprocess the standardized operating data, including using the optimal spline function control points, jump path connection weights and path activation thresholds to update the neural slice network model structure, using the standardized operating data as input, performing a complete forward calculation on the model, generating a new fusion state representation result, and the state assessment module outputs the final operating state classification, fault risk score and trend prediction results to form comprehensive state assessment information.

[0036] Optionally, the S7 inputs the comprehensive status assessment information into the power operation and maintenance system, including: uniformly formatting the operating status classification results, fault risk scores and trend prediction results into visual status data, and synchronizing them to the power operation and maintenance system through the operation and maintenance interface module, to realize equipment status chart display, abnormal operating condition warning information push and health management record update, and support real-time monitoring, decision reference and risk intervention operations of operation and maintenance personnel.

[0037] The beneficial effects of the present invention are:

[0038] (1) This paper constructs a neural slice network model that includes a skip-slice module, introduces a multi-segment spline function subnetwork and a path selection control unit, supports dynamic path activation and multi-scale feature fusion, and can adaptively select the optimal modeling path based on different operating states. This structure effectively improves the model's ability to fit nonlinear fluctuating data, enhances state recognition accuracy, and overcomes the poor adaptability of traditional static neural networks under complex operating conditions.

[0039] (2) During the model output phase, the present invention jointly generates operating status classification results, fault risk scores, and trend prediction values ​​to form comprehensive status assessment information. Compared to existing monitoring methods that only output a single result, the present invention achieves intelligent evaluation of the entire cycle and process of power equipment operation, enhances the system's ability to detect potential faults early and analyze trends, and effectively supports predictive maintenance decisions for power operations.

[0040] (3) This paper introduces a spectral ant oscillation optimization algorithm during the model optimization phase, performing a global search and iterative update of spline function control points, jump path connection weights, and activation thresholds, constructing a unified optimization space and strengthening the path structure selection mechanism. This optimization method possesses strong global optimization capabilities, overcoming the problem of conventional gradient optimization methods' insufficient support for complex structural models and significantly improving the model's robustness and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0042] Figure 1 This is an overall flow chart of a method for monitoring the state of power equipment based on an adaptive neural network proposed by the present invention;

[0043] Figure 2 This is a standardized operation data processing flow chart of a power equipment status monitoring method based on an adaptive neural network proposed by the present invention;

[0044] Figure 3This is a neural slice network structure diagram of the power equipment status monitoring method based on adaptive neural network proposed by the present invention. DETAILED DESCRIPTION

[0045] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0046] refer to Figure 1-3 , a method for monitoring the state of power equipment based on an adaptive neural network, comprising the following steps:

[0047] S1. Collect the timing signals of power equipment, perform time alignment and format unification, and construct the input data set;

[0048] S2. Perform missing value filling, outlier removal, normalization and trend smoothing on the input data set to obtain standardized data;

[0049] S3. Construct a neural slice network model including a skip-slice module. The skip-slice module is composed of several segments of spline function sub-networks. Combined with a path selection control unit, it realizes dynamic path activation based on input features. Different paths correspond to feature modeling accuracy at different scales. Multi-scale outputs are fused through a coupling mechanism to enhance state representation capabilities.

[0050] S4, inputting the standardized data into the neural slice network model to obtain the initial state evaluation results;

[0051] S5. Optimizing the structural parameters of the neural slice network model using a spectrum ant oscillation optimization algorithm, wherein the optimization includes jointly updating the spline function coefficients, jump path connection weights, and path activation thresholds based on a spectrum perturbation mechanism and an ant colony search mechanism;

[0052] S6. Reprocess the standardized data with the optimized model and output the operating status classification, fault risk score and trend prediction results to form comprehensive status assessment information;

[0053] S7. Transmit the comprehensive status assessment information to the power operation and maintenance system to realize status display, fault warning and health score output.

[0054] This paper constructs a neural slice network model that incorporates a skip-slice module and combines it with a spectral ant oscillation optimization algorithm to implement a comprehensive power equipment condition monitoring solution, from data acquisition and preprocessing to model building, structural optimization, and evaluation output. Compared to traditional static network models, this approach boasts structural adaptability, stronger optimization strategies, and a more comprehensive evaluation dimension. This improves the model's accuracy and adaptability under complex operating conditions, significantly enhancing the ability to detect fault trends early on.

[0055] In this embodiment, the time alignment and format unification of S1 include: setting a unified time sampling interval for the voltage signal, current signal, temperature signal, harmonic signal and vibration signal respectively, and reconstructing the time axis of data with different sampling frequencies using interpolation method to generate a data matrix with a unified time step; uniformly processing the data format of various signals, converting the original signal into a fixed-length numerical sequence or vector form, encoding it with a unified data structure and storing it in the input data set.

[0056] To address the issue of differences in time series data collection, the present invention introduces a unified time alignment mechanism and data format standardization solution to ensure that voltage, current, temperature and other signals from multiple sensors can be processed synchronously. This solves the modeling error problem caused by asynchronous and inconsistent formats of multi-source data in the existing technology, and provides a stable and high-quality input foundation for subsequent state recognition models.

[0057] In this embodiment, the missing value filling, outlier elimination, normalization and trend smoothing processing of the input data set in S2 include: constructing each type of time series signal in the unified input data set as an equally spaced time series, and using the interpolation method to fill the missing data at non-boundary positions; using the sliding window statistical method to detect outlier data that exceeds the normal fluctuation range and eliminate it; after outlier elimination and missing value filling, each type of time series signal is normalized according to the maximum and minimum values ​​in the entire time series range to unify the feature value range; based on the exponential sliding average method, the normalized time series signal is trend smoothed to suppress instantaneous fluctuations.

[0058] The present invention designs a systematic cleaning process in the data preprocessing stage, including missing value filling, outlier removal, normalization and trend smoothing operations, which improves data integrity and stability, and solves the problem of unstable performance of traditional algorithms when faced with missing data, severe fluctuations or inconsistent feature scales, laying the foundation for stable training and accurate inference of neural networks.

[0059] In this embodiment, S3 specifically includes:

[0060] S31, construct the standardized running data into an input sequence X={x1,x2,…,x T}, where the input feature vector x at each moment t , as the input sequence of the neural slice network model;

[0061] S32, based on the input sequence X, construct a hierarchical representation through multiple spline function sub-networks, each spline function sub-network is used to represent the upper layer input Perform piecewise cubic spline transform and output the current layer representation Among them, l represents the number of network layers, and the control point set of the spline function subnetwork is recorded as Where m is the number of slices per layer;

[0062] S33, set a jump path set P between different depth layers = {p i,j}, where p i,j Represents the jump path from layer i to layer j, and defines the activation variable a for each jump path i,j ∈{0,1};

[0063] S34, construct a path selection control unit, based on the change rate Δx of the input sequence at time step t t =x t -x t-1 Calculate the path activation probability π i,j , based on the maximum probability strategy to activate the corresponding jump path, that is, when π i,j >ρ, set a i,j =1, where ρ is the preset threshold;

[0064] S35, perform several scale spline coupling fusions on the output representations of all activated paths and the current main path representation, and construct a fusion representation sequence H = {h1,h2,…,h T}, describing the time evolution characteristics of the power equipment status.

[0065] The skip slicing module constructed in this paper introduces a multi-segment spline function subnetwork and a path selection control unit, enabling dynamic activation of modeling paths based on input features. This addresses the issues of fixed structure and sluggish response in existing neural networks. By coupling multi-scale path outputs, the model's ability to distinguish between different state modes and its structural flexibility are effectively enhanced, making it particularly suitable for applications where power equipment operating conditions are complex and changeable.

[0066] In this embodiment, the S4 specifically includes:

[0067] S41, construct the standardized running data into an input sequence X={x1,x2,…,x T}, where x t is the multi-source feature vector at time step t, which is input into the neural slice network model containing the skip-slice module;

[0068] S42: In the neural slice network model, the input sequence X is processed by multi-segment spline function sub-network and skip path selection in sequence to obtain a fusion state representation sequence H = {h1, h2, ..., h T}, where h t is the state feature representation vector at time step t;

[0069] S43, input the fusion state representation sequence H into the state evaluation output module, and generate the initial state evaluation result Y={(y1,r1),(y2,r2),…,(y T ,r T )}, where y t represents the running status classification label at time step t, r t Indicates the corresponding fault risk score value.

[0070] The present invention applies the aforementioned structure to the actual data processing process. Without changing the original data expression, it extracts the fusion state representation and generates the initial state assessment result. It has the dual output capabilities of state classification and risk scoring. Compared with the traditional single-label classification method, the output structure of the present invention is richer and the information value is higher, which helps to comprehensively judge the current and potential operating status of the equipment.

[0071] In this embodiment, the S5 specifically includes:

[0072] S51, based on the initial state evaluation result Y={(y1,r1),(y2,r2),…,(y T ,r T )} and the standard answer to construct the loss function L, where y t Represents the running status classification label, r t represents the fault risk score value, and the loss function L considers both classification error and regression error;

[0073] S52. Define the search parameter space of the spectral ant oscillation optimization algorithm, including the control point set C = {c1, c2, ..., c m}、Skip path connection weight set W={w i,j} and the path activation threshold set Θ = {θ i,j};

[0074] S53. Initialize multiple ant colony individuals, each individual represents a parameter vector v = [C, W, Θ], and introduce a parameter oscillation mechanism based on the spectral perturbation function to perturb the individuals to initialize the perturbation, thereby enhancing the globality of the parameter search;

[0075] S54, calculating the model loss value L under the current parameter combination for each generation of ant colony individuals, updating the pheromone matrix and path evaluation function according to the loss result, and performing weight enhancement on the high-quality paths;

[0076] S55. After the maximum number of iterations or the convergence threshold condition is met, the control point set C corresponding to the individual that minimizes the loss function is selected. * , connection weight set W * With the activation threshold set Θ* , used to update the structural parameter configuration of the neural slice network model.

[0077] This paper introduces the spectral ant oscillation optimization algorithm into the neural network structural parameter optimization process, constructing a joint parameter space consisting of spline control points, jump path weights, and activation thresholds. This overcomes the limitation of traditional deep learning models that only optimize weights, improving structural adaptability and model performance. This method possesses strong global search capabilities, avoids local optimality, and ensures comprehensive improvements in model stability and generalization.

[0078] In this embodiment, the search parameter space of the spectral ant oscillation optimization algorithm is defined in S52, which includes: constructing a parameter vector v = [C, W, Θ], wherein the spline function control point set C = {c1, c2, ..., c m} represents the node position and number of the piecewise cubic spline function in each layer of the neural slice network, and the jump path connection weight set W = {w i,j} represents the jump path weight between layer i and layer j, and the path activation threshold set Θ = {θ i,j} represents the threshold parameter that controls the activation state of the jump path from the i-th layer to the j-th layer; the parameter vector v is used as the search individual representation of the spectral ant oscillation optimization algorithm, and the search range is limited to the continuous feasible space within the preset parameter boundary, which serves as the basis for subsequent ant colony initialization and perturbation operations.

[0079] When constructing the parameter space of the spectral ant oscillation optimization algorithm, the present invention incorporates structural parameters such as spline function control points, path connection weights and activation thresholds into a unified optimization framework for the first time, clarifies the dependencies and optimization boundaries between structural levels, improves the integrity of the optimization operation and the refinement of the model tuning, and provides an algorithmic basis for dynamic structural adjustment, which is significantly different from traditional optimization strategies.

[0080] In this embodiment, the optimized neural slice network model in S6 is reprocessed with standardized operation data, including updating the neural slice network model structure using the optimal spline function control points, jump path connection weights and path activation thresholds, using the standardized operation data as input, performing a complete forward calculation on the model, generating a new fusion state representation result, and the state evaluation module outputs the final operation state classification, fault risk score and trend prediction results to form comprehensive state evaluation information.

[0081] After optimizing structural parameters, the present invention reprocesses the same input data based on the optimized neural slice network model, outputting a fused state representation and complete evaluation results, achieving linkage consistency between model structure and functional output. This mechanism significantly improves the reliability and stability of the final evaluation, resolving the disconnect between structural optimization and evaluation output in traditional algorithms.

[0082] In this embodiment, the S7 inputs the comprehensive status assessment information into the power operation and maintenance system, including: uniformly formatting the operating status classification results, fault risk scores and trend prediction results into visual status data, and synchronizing them to the power operation and maintenance system through the operation and maintenance interface module, to realize the display of equipment status charts, push of abnormal working condition warning information and update of health management records, and support real-time monitoring, decision reference and risk intervention operations of operation and maintenance personnel.

[0083] This invention formats the final comprehensive status assessment results and directly connects them to the power operation and maintenance system, enabling multi-dimensional visualization and alert push of operating status, risk scoring, and trend forecasts. This enhances the interpretability and application efficiency of the model results in practical systems. Compared to traditional models that require secondary processing, this invention offers greater system integration capabilities and practical application value.

[0084] Example 1:

[0085] To verify the feasibility and effectiveness of this invention, we applied it to a remote monitoring system for 110kV power transmission and transformation equipment at a comprehensive energy company. This system was deployed at multiple locations, covering key nodes such as transformers, circuit breakers, voltage transformers, current transformers, and distribution buses. Its goal was to achieve online intelligent identification of equipment operating status, early warning of risk trends, and visual analysis of status information.

[0086] In this scenario, the existing monitoring system primarily relied on a fixed threshold-triggered alarm mechanism, supplemented by periodic manual inspections. This resulted in high false alarm rates, delayed responses, and insufficient status detail. Sensor data fluctuations increased significantly under high-temperature, high-load, aging equipment, or harmonic interference conditions. Existing algorithm models were unable to effectively identify nonlinear fluctuation patterns, often leading to misjudgments of status or delayed warnings, posing a threat to system safety.

[0087] To address these issues, the project team employed the adaptive neural network-based power equipment condition monitoring method described in this paper. Sensor nodes were deployed to collect multi-source time-series signal data, including temperature, voltage, current, harmonic components, and vibration, and a unified input data set was constructed. During data processing, missing value interpolation, sliding window anomaly removal, and min-max normalization were used to clean and standardize the raw equipment signals. The average single-point data missing recovery rate reached 98.4%, significantly improving the stability of subsequent models.

[0088] During the modeling phase, a neural slice network model integrating a skip-slice module was constructed. This model activates optimal paths between spline subnetworks of varying depths based on the dynamic changes in input features, enabling hierarchical modeling of device operational characteristics. A spectral ant oscillation optimization algorithm was also introduced to globally optimize the model's structural parameters, including spline function control points, skip-path connection weights, and activation thresholds. Ultimately, the resulting structural combination achieved the optimal convergence performance within the current device environment.

[0089] During the status assessment phase, the model simultaneously outputs operating status classification labels, equipment failure risk scores, and trend prediction values. Test data verified that the model's accuracy in the early detection of circuit breaker anomalies increased from 81.6% of traditional models to 94.2%. In temperature rise overload trend prediction, the prediction deviation decreased from a mean square error of 0.36 to 0.12. The risk score correlation index increased to 0.91, demonstrating strong interpretability and foresight. Within the system operation and maintenance platform, these status assessment results are automatically connected to a visualization interface for use in business processes such as status curve display, risk map distribution, and automatic work order generation, significantly improving the efficiency of fault intervention response and the accuracy of inspection resource allocation.

[0090] To support the experimental results, the project team compiled typical data samples and model evaluation results, as shown in the following table.

[0091] Table 1: Statistics of equipment condition monitoring model test results

[0092] Test indicators Original model performance The method of the present invention shows Running status classification accuracy (%) 81.6 94.2 <![CDATA[Fault risk score correlation (R 2 )]]> 0.72 0.91 Trend forecast mean square error (MSE) 0.36 0.12 Abnormal identification lead time (average minutes) 8.5 21.7 Data loss recovery rate (%) 75.3 98.4 Calculation delay (seconds per device) 2.8 3.1

[0093] The aforementioned "Equipment Condition Monitoring Model Test Results Statistics" demonstrates the performance advantages of the adaptive neural network-based power equipment condition monitoring method proposed in this paper in typical practical applications. By comparing it with the existing model, the improved accuracy, predictive power, data processing capabilities, and actual operational efficiency are comprehensively demonstrated across six key performance indicators.

[0094] First, in terms of operating status classification accuracy, the original model achieved an accuracy of 81.6%, while the proposed method increased this to 94.2%, an improvement of over 12%. This significantly enhances the ability to identify the operating status of power equipment, demonstrating that its structural adaptive mechanism is more suitable for processing complex fluctuating signals. Second, in terms of fault risk score correlation, the original model's score output had a correlation of only 0.72 with the actual results, while the proposed method achieved a correlation of 0.91, demonstrating that the optimized model can more accurately quantify potential equipment risks and is more valuable for operational and maintenance decision-making.

[0095] In terms of trend prediction performance, the mean square error of the proposed method decreased from 0.36 to 0.12, demonstrating its greater accuracy in modeling equipment operating trends. It is particularly suitable for identifying gradual anomalies such as overload and temperature rise. Regarding the anomaly identification lead time metric, the original model could identify anomalies an average of 8.5 minutes in advance, while the proposed method could identify anomalies 21.7 minutes earlier, a nearly threefold improvement. This provides more time for intervention by operation and maintenance personnel, improving accident prevention efficiency.

[0096] In terms of data processing capabilities, the proposed method significantly outperforms the original system in terms of missing data recovery rate, increasing from 75.3% to 98.4%. This demonstrates that its preprocessing modules (such as interpolation and cleaning algorithms) are more comprehensive and can effectively address the issue of incomplete sensor data under actual working conditions. Finally, in terms of real-time computing power, the two are similar. The proposed method has an average processing latency of 3.1 seconds on a single device, slightly higher than the original model's 2.8 seconds. However, this difference is completely acceptable in exchange for higher performance evaluation results.

[0097] In summary, this table fully verifies the actual effect of the method of the present invention in power equipment status monitoring, especially in terms of status recognition accuracy, risk judgment ability, trend prediction ability and data completion stability, showing good engineering practical value and deployment feasibility.

[0098] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for monitoring the state of power equipment based on an adaptive neural network, characterized in that: The steps include: S1. Collect the timing signals of power equipment, perform time alignment and format unification, and construct the input data set; S2. Perform missing value filling, outlier removal, normalization and trend smoothing on the input data set to obtain standardized data; S3. Construct a neural slice network model including a skip-slice module. The skip-slice module is composed of several segments of spline function sub-networks. Combined with a path selection control unit, it realizes dynamic path activation based on input features. Different paths correspond to feature modeling accuracy at different scales. Multi-scale outputs are fused through a coupling mechanism to enhance state representation capabilities. S4, inputting the standardized data into the neural slice network model to obtain the initial state evaluation results; S5. Optimizing the structural parameters of the neural slice network model using a spectrum ant oscillation optimization algorithm, wherein the optimization includes jointly updating the spline function coefficients, jump path connection weights, and path activation thresholds based on a spectrum perturbation mechanism and an ant colony search mechanism; S6. Reprocess the standardized data with the optimized model and output the operating status classification, fault risk score and trend prediction results to form comprehensive status assessment information; S7. Transmit the comprehensive status assessment information to the power operation and maintenance system to realize status display, fault warning and health score output.

2. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: The time alignment and format unification performed by S1 include: setting a unified time sampling interval for voltage signals, current signals, temperature signals, harmonic signals and vibration signals respectively, reconstructing the time axis of data with different sampling frequencies using interpolation method, and generating a data matrix with a unified time step; uniformly processing the data formats of various signals, converting the original signals into fixed-length numerical sequences or vector forms, encoding them in a unified data structure and storing them in the input data set.

3. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: The S2 performs missing value filling, outlier removal, normalization and trend smoothing on the input data set, including: constructing each type of time series signal in the unified input data set into an equally spaced time series, and using the interpolation method to fill in the missing data at non-boundary positions; using the sliding window statistical method to detect outlier data that exceeds the normal fluctuation range and remove it; after outlier removal and missing value filling, normalize each type of time series signal according to the maximum and minimum values ​​in the entire time series range to unify the feature value range; and perform trend smoothing on the normalized time series signal based on the exponential sliding average method to suppress instantaneous fluctuations.

4. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: The S3 specifically includes: S31, construct the standardized running data into an input sequence X={x1,x2,…,x T }, where the input feature vector x at each moment t , as the input sequence of the neural slice network model; S32, based on the input sequence X, construct a hierarchical representation through multiple spline function sub-networks, each spline function sub-network is used to represent the upper layer input Perform piecewise cubic spline transform and output the current layer representation Among them, l represents the number of network layers, and the control point set of the spline function subnetwork is recorded as Where m is the number of slices per layer; S33, set a jump path set P between different depth layers = {p i,j }, where p i,j Represents the jump path from layer i to layer j, and defines the activation variable a for each jump path i,j ∈{0,1}; S34, construct a path selection control unit, based on the change rate Δx of the input sequence at time step t t =x t -x t-1 Calculate the path activation probability π i,j , based on the maximum probability strategy to activate the corresponding jump path, that is, when π i,j >ρ, set a i,j =1, where ρ is the preset threshold; S35, perform several scale spline coupling fusions on the output representations of all activated paths and the current main path representation, and construct a fusion representation sequence H = {h1,h2,…,h T }, describing the time evolution characteristics of the power equipment status.

5. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: The S4 specifically includes: S41, construct the standardized running data into an input sequence X={x1,x2,…,x T }, where x t is the multi-source feature vector at time step t, which is input into the neural slice network model containing the skip-slice module; S42: In the neural slice network model, the input sequence X is processed by multi-segment spline function sub-network and skip path selection in sequence to obtain a fusion state representation sequence H = {h1, h2, ..., h T }, where h t is the state feature representation vector at time step t; S43, input the fusion state representation sequence H into the state evaluation output module, and generate the initial state evaluation result Y={(y1,r1),(y2,r2),…,(y T ,r T )}, where y t represents the running status classification label at time step t, r t Indicates the corresponding fault risk score value.

6. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: The S5 specifically includes: S51, based on the initial state evaluation result Y={(y1,r1),(y2,r2),…,(y T ,r T )} and the standard answer to construct the loss function L, where y t Represents the running status classification label, r t represents the fault risk score value, and the loss function L considers both classification error and regression error; S52. Define the search parameter space of the spectral ant oscillation optimization algorithm, including the control point set C = {c1, c2, ..., c m }、Skip path connection weight set W={w i,j } and the path activation threshold set Θ = {θ i,j }; S53. Initialize multiple ant colony individuals, each individual represents a parameter vector v = [C, W, Θ], and introduce a parameter oscillation mechanism based on the spectral perturbation function to perturb the individuals to initialize the perturbation, thereby enhancing the globality of the parameter search; S54, calculating the model loss value L under the current parameter combination for each generation of ant colony individuals, updating the pheromone matrix and path evaluation function according to the loss result, and performing weight enhancement on the high-quality paths; S55. After the maximum number of iterations or the convergence threshold condition is met, the control point set C corresponding to the individual that minimizes the loss function is selected. * , connection weight set W * With the activation threshold set Θ * , used to update the structural parameter configuration of the neural slice network model.

7. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: The search parameter space of the spectral ant oscillation optimization algorithm defined in S52 includes: constructing a parameter vector v = [C, W, Θ], wherein the spline function control point set C = {c1, c2, ..., c m } represents the node position and number of the piecewise cubic spline function in each layer of the neural slice network, and the jump path connection weight set W = {w i,j } represents the jump path weight between layer i and layer j, and the path activation threshold set Θ = {θ i,j } represents the threshold parameter that controls the activation state of the jump path from the i-th layer to the j-th layer; the parameter vector v is used as the search individual representation of the spectral ant oscillation optimization algorithm, and the search range is limited to the continuous feasible space within the preset parameter boundary, which serves as the basis for subsequent ant colony initialization and perturbation operations.

8. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: In S6, the optimized neural slice network model is reprocessed with standardized operation data, including updating the neural slice network model structure using the optimal spline function control points, jump path connection weights and path activation thresholds, using the standardized operation data as input, performing a complete forward calculation on the model, generating a new fusion state representation result, and the state assessment module outputs the final operation state classification, fault risk score and trend prediction results to form comprehensive state assessment information.

9. The method for monitoring the state of power equipment based on an adaptive neural network according to claim 1, characterized in that: The S7 inputs the comprehensive status assessment information into the power operation and maintenance system, including: uniformly formatting the operating status classification results, fault risk scores and trend prediction results into visual status data, synchronizing them to the power operation and maintenance system through the operation and maintenance interface module, realizing the display of equipment status charts, the push of abnormal working condition warning information and the update of health management records, and supporting real-time monitoring, decision reference and risk intervention operations of operation and maintenance personnel.

Images