Fault Early Warning System for Power Transmission and Transformation Equipment Based on Online Monitoring

Abstract

The invention discloses a fault early warning system for power transmission and transformation equipment based on online monitoring, including the following modules: a multi-parameter sensor module for collecting data in real time; a wireless sensor network module for transmitting the data to a data processing center for processing; a data fusion module that uses a multi-modal data fusion algorithm based on a self-evolving nested network to generate a comprehensive health score of the equipment; a trend prediction module based on the non-linear time series analysis technology of a generative adversarial adaptive residual network, and introducing an adaptive residual module for trend prediction; an early warning threshold adjustment module that uses an adaptive evolution algorithm of meta-reinforcement learning to dynamically adjust the early warning threshold; a maintenance suggestion generation module for generating maintenance suggestions; and a power grid dispatching linkage module for automatically linking with the power grid dispatching system to detect and adjust the equipment load. The invention adopts a multi-parameter fusion and generative adversarial network method, and introduces an adaptive residual module to achieve accurate fault early warning of power transmission and transformation equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring and early warning, and particularly to a fault early warning system for power transmission and transformation equipment based on on-line monitoring. Background Art

[0002] With the continuous expansion of the scale of the power system, as the core components of the power grid operation, the safety and stability of power transmission and transformation equipment become particularly important. However, due to the long-term operation of power transmission and transformation equipment under high voltage, high load and complex environments, the occurrence of faults is unpredictable. Most traditional maintenance methods for power transmission and transformation equipment rely on regular inspections and manual patrols, which not only involve huge workloads, but also may fail to detect equipment hidden dangers in time, thus leading to serious power accidents. Therefore, how to monitor and early warn the faults of power transmission and transformation equipment through intelligent means has become an important issue in the power industry.

[0003] Existing fault monitoring technologies for power transmission and transformation equipment mainly rely on the monitoring of single or a small number of parameters, such as physical quantities like temperature, current, voltage, etc. The abnormal detection methods based on these parameters usually use preset thresholds, and when a certain parameter exceeds the threshold, the system will issue an alarm. This single-parameter monitoring method has multiple problems. First of all, equipment faults are often the result of the combined action of multiple factors, and the change of a single parameter may not be sufficient to comprehensively reflect the true state of the equipment; secondly, the fixed early warning threshold cannot adapt to the changes in the equipment operation environment, such as seasonal climate changes or load fluctuations, which is likely to lead to false alarms or missed alarms; furthermore, due to the complexity of equipment operation and the huge amount of data, relying on a single parameter cannot effectively capture the non-linear change trend of equipment faults and cannot achieve accurate fault early warning.

[0004] In order to improve the accuracy of fault early warning, in recent years, methods of multi-parameter fusion analysis have emerged, attempting to fuse the data collected by multiple sensors and generate a more accurate equipment health status through data fusion algorithms. However, most existing data fusion technologies are based on simple weighted averaging or logical reasoning. Although these methods have improved the detection accuracy to a certain extent, due to their inability to dynamically and adaptively adjust the fusion weights, there are still large limitations for data under complex working conditions. In addition, most existing fusion methods are based on static models and cannot be optimized in real time according to the changes in equipment status. This makes the early warning ability of the system prone to decline when the equipment operation environment and load conditions change.

[0005] Another challenge is that most existing fault prediction models rely on linear or simple machine learning algorithms, such as support vector machines (SVM), decision trees, etc. These methods have certain effects when dealing with simple fault patterns, but they show obvious deficiencies when facing more complex fault patterns. Especially, the operating states and fault trends of power transmission and transformation equipment usually exhibit non-linear, time-varying, and multi-dimensional characteristics, and traditional linear models cannot capture the dynamic change trends of equipment health states. In addition, these methods cannot adaptively optimize and update the prediction model, making it difficult to meet the personalized needs of different equipment.

[0006] In addition, there are also significant limitations in the existing technologies for generating maintenance suggestions. Traditional methods usually rely on historical data and experience for simple inferences. When equipment shows anomalies, the system can only give fixed suggestions, such as arranging maintenance or replacing components. However, such experience-based suggestions often lack pertinence and are difficult to provide precise maintenance plans according to the specific fault patterns and operating states of the equipment. This not only reduces the efficiency of fault handling but may even lead to an increase in maintenance costs.

[0007] Regarding the linkage problem of the power grid dispatching system, existing technologies are usually limited to manual operations. When the early warning level of equipment is relatively high, operators need to manually adjust the load distribution of power grid equipment, start redundant equipment, etc. This way that relies on human judgment and operation is not only inefficient but also prone to delays, unable to respond immediately at the moment of fault occurrence, thus affecting the stability of the power system.

[0008] In summary, there are many deficiencies in the existing technologies for fault monitoring, early warning, and maintenance of power transmission and transformation equipment: First, the single-parameter monitoring method cannot comprehensively reflect the operating state of the equipment; second, the existing data fusion methods cannot dynamically adjust weights and are difficult to cope with complex working condition changes; third, traditional fault prediction models are difficult to capture the non-linear changes in equipment health states and cannot achieve precise trend prediction; finally, the maintenance suggestions lack pertinence, and the linkage response efficiency of the power grid dispatching system is low. These technical defects lead to the instability of the fault early warning system in practical applications and increase the risk of equipment operation.

[0009] Therefore, how to provide a fault early warning system for power transmission and transformation equipment based on online monitoring is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0010] One purpose of the present invention is to propose a power transmission and transformation equipment fault warning system based on online monitoring. The present invention adopts a multi-parameter fusion technology combined with a self-evolving nested network and a method of generating adversarial adaptive residual networks. By real-time collection of the operating status data of the power transmission and transformation equipment, and multi-modal data fusion and fault trend prediction, and further combined with the adaptive evolutionary algorithm of meta-reinforcement learning, the dynamic adjustment of the fault warning threshold is realized. The system can not only accurately predict the equipment failure trend, but also generate intelligent maintenance suggestions according to different warning levels, and link with the power grid dispatching system, improving the real-time and accuracy of fault warning and the efficiency of system emergency response, and has the advantages of high safety, precise maintenance and strong linkage.

[0011] The power transmission and transformation equipment fault early warning system based on online monitoring according to an embodiment of the present invention includes the following modules:

[0012] Multi-parameter sensor module, used to collect real-time operating status data of power transmission and transformation equipment;

[0013] A wireless sensor network module is used to transmit the collected operating status data to a data processing center for processing;

[0014] The data fusion module uses a multimodal data fusion algorithm based on a self-evolving nested network and combines it with an adaptive weight distribution model to fuse the data of each sensor modality and generate a comprehensive health score for the device;

[0015] The trend prediction module is based on the nonlinear time series analysis technology of the generative adversarial adaptive residual network, and introduces an adaptive residual module to predict the trend of the comprehensive health score and historical operation data of the equipment;

[0016] The warning threshold adjustment module uses an adaptive evolutionary algorithm of meta-reinforcement learning to dynamically adjust the warning threshold based on the historical data and real-time operation status of the equipment;

[0017] The maintenance suggestion generation module is used to generate maintenance suggestions based on the fault risk index and warning level when the system detects an abnormality;

[0018] The power grid dispatch linkage module is used to automatically link the power grid dispatch system when the equipment warning level reaches orange or red, detect and adjust the equipment load, and distribute the load to the backup equipment or start the redundant equipment when necessary.

[0019] According to an embodiment of the present invention, a method for early warning of power transmission and transformation equipment failure based on online monitoring comprises the following steps:

[0020] S1. Collecting the operating status data of the power transmission and transformation equipment in real time through multi-parameter sensors, wherein the operating status data includes temperature, vibration, current, voltage and partial discharge information;

[0021] S2. Transmit the collected operation status data to the data processing center through the wireless sensor network for processing;

[0022] S3. Adopt a multi-modal data fusion algorithm based on a self-evolving nested network, and combine an adaptive weight allocation model for data fusion to generate a comprehensive health score of the device;

[0023] S4. Based on the non-linear time series analysis technology of the generative adversarial adaptive residual network, introduce an adaptive residual module in the generator to predict the trend of the comprehensive health score and historical operation data of the device, monitor the change of the device health index, and identify the fault trend;

[0024] S5. According to the historical data and real-time operation status of the device, adopt an adaptive evolution algorithm based on meta-reinforcement learning to dynamically adjust the warning threshold;

[0025] S6. When the system detects an anomaly, generate maintenance suggestions according to the warning level, and the warning level includes yellow warning, orange warning and red warning;

[0026] S7. When the warning level of the device reaches orange or red, the system is linked with the power grid dispatching system to automatically detect and adjust the device load, and when necessary, allocate the load to the standby device or start the redundant device.

[0027] Optionally, the S3 specifically includes:

[0028] S31. For the operation status data of the power transmission and transformation equipment collected by the multi-parameter sensor, for each sensor modality D i Perform preprocessing respectively, where i = 1, 2,..., n, and n represents the number of sensor types. The preprocessing includes data denoising, outlier removal, normalization and time window sliding operation to obtain a standardized data set D′ i ;

[0029] S32. Establish a self-evolving nested network structure, and construct an initial nested network N for each sensor modality D′ i The weights of the nested network are initialized to a random normal distribution, and the nested network model N i is defined as: i For:

[0030] F i = N(D′ i ; θ i , L i );

[0031] Among them, θ i represents the parameter set of the nested network N i , and L irepresents the number of layers of the nested network, and initially for each modal data D′ i processes are carried out using a network structure with the same number of layers, and the network output is F i ;

[0032] S33. Use an adaptive weight assignment model to weight the contribution degrees of the outputs of each modality, and the weight coefficient w i is dynamically updated through a multi-layer adaptive adjustment formula, and the weight adjustment mechanism is based on the multi-modal non-linear dependence relationship of sensor data;

[0033] Calculate the initial contribution degree of modality i at time t

[0034]

[0035] where, represents the output of modality i at time t, represents the output of modality j at time t, n represents the total number of modalities, and L total represents the total loss function of multi-modal fusion:

[0036]

[0037] where, represents the weight coefficient of modality i at time t, and y true represents the true device health state value, λ represents the regularization coefficient, and ∥θ i ∥ 2 represents the L2 norm of the network parameters;

[0038] By adding a time decay term, calculate the weight update value of the modality:

[0039]

[0040] where, represents the weight update value of modality i at time t, and ρ represents the time decay coefficient;

[0041] Add a non-linear self-adjustment function to perform a non-linear transformation on the weights:

[0042]

[0043] where, exp represents the exponential function, κ represents the control factor, and μ represents the dynamic offset;

[0044] S34. Perform self-evolution processing on the nested network N i The evolution process is based on the combination of a genetic algorithm and an adaptive neural network structure search algorithm to achieve dynamic optimization of the network structure, including increasing or decreasing the number of network layers, changing the type of activation function, and adjusting the number of neurons;

[0045] S35. Calculate the fitness function of the network structure after each round of evolution. The fitness function Fitness(N i ) is defined as:

[0046]

[0047] where γ represents the weight balance coefficient and L i represents the network depth;

[0048] S36. Dynamically adjust the nested network structure during the evolution process and select the optimal network structure based on the fitness results The optimal fusion result obtained by each round of iterative calculation

[0049] S37. Repeat steps S33 - S36 until the preset convergence condition is reached, complete the optimal fusion of the self-evolving nested network, and finally generate the comprehensive health score of the device.

[0050] Optionally, the specific steps of S4 include:

[0051] S41. Perform multi-dimensional non-linear time series modeling on the comprehensive health score data H t and historical operation data D t . Define the historical comprehensive health score sequence and operation data sequence as:

[0052] H t ={H t ,H t-1 ,…,H t-k}, D t ={D t ,D t-1 ,…,D t-k};

[0053] where k represents the time window length, and the generator G generates the health score at the future moment. The generated output is:

[0054]

[0055] where represents the generated predicted score, and θ G represents the initial parameter of the generator;

[0056] S42. Calculate the error between the output of the generator and the actual health score by introducing an adaptive residual module and perform feedback correction;

[0057] S43. After the correction by the adaptive residual module, the output of the generator is updated to:

[0058]

[0059] Among them, represents the updated generator output, and α t represents the dynamic adjustment coefficient, and R t represents the residual defined at the current time t, and R t-i represents the residual at time t - i, exp represents the exponential function, k represents the total number of steps, and W (i) represents the weight matrix at the i-th step;

[0060] S44. The generator and the discriminator are adversarially trained through an adaptive residual module. The discriminator D is used to distinguish the true health score and the predicted value generated by the generator:

[0061]

[0062] Among them, m represents the number of samples processed;

[0063] S45. Through multiple adjustments and corrections of the adaptive residual module, the final output prediction of the generator is obtained, and the fault trend is identified.

[0064] Optionally, the correction of the adaptive residual module in S42 specifically includes:

[0065] S421. Define the residual R t at the current time as:

[0066] R t = H t - G(H t-1 , D t-1 ; θ G );

[0067] Among them, G(H t-1 , D t-1 ; θ G ) represents the predicted output of the generator at time t - 1;

[0068] S422. When the absolute value of the residual |R t | exceeds the threshold δ1, the generator will adjust the network depth according to the size of the residual:

[0069]

[0070] Among them, L new represents the adjusted network depth, L represents the current network depth, β represents the adjustment factor, represents the gradient of the generator loss function L G with respect to the parameter θ G , represents the gradient of the residual;

[0071] S423. When the residual δ2 < |R t | ≤ δ1, the adaptive residual module optimizes the network width by adjusting the number of neurons N in each layer l :

[0072]

[0073] Wherein, represents the number of neurons in the adjusted l-th layer, N l represents the current number of neurons in the l-th layer, γ1 represents the learning rate, represents the weight matrix at the i-th step of the l-th layer, ∈ represents the smoothing term, and k represents the total number of steps;

[0074] S424. When the residual δ2 < |R t | ≤ δ1, the generator introduces cross-layer connections to enhance the network's expressive ability:

[0075]

[0076] Wherein, y l represents the output of the l-th layer, y l+k represents the output of the (l + k)-th layer, and λ1 represents the adjustment coefficient of the cross-layer connection.

[0077] This cross-layer connection directly transmits low-level information to the high level, enhancing the ability to identify complex non-linear fault trends.

[0078] Optionally, the specific content of S5 includes:

[0079] S51. Construct the state space of the device's historical operation data D t and the comprehensive health score H t , and define the state space S t as:

[0080] S t = {D t , H t , …, D t-k , H t-k};

[0081] Wherein, t represents the current time point, and k represents the time window length of the historical data;

[0082] S52. According to the state space S t , define the action space A t of the device, and the action space includes the operation of dynamically adjusting the warning threshold:

[0083]

[0084] Wherein, Denote the $i$-th action, representing different warning threshold adjustment strategies;

[0085] S53. Train an adaptive evolutionary strategy using the meta-reinforcement learning algorithm, specifically including the design of the reward function and policy update. The reward function $Q$ t is used to measure the effect of each action:

[0086]

[0087] where, denotes the true health score $H$ t and the loss between the predicted health score . $\lambda_2$ represents the regularization coefficient, and $\|\theta$ t \|$ 2 denotes the L2 regularization term of the parameters;

[0088] S54. Based on the meta-learning mechanism of reinforcement learning, define the policy update rule as: According to the current state $S$ t and the action $a$ t , update the policy function $\pi(a$ t |$S$ t ), and select the optimal action $a$ t at each time step according to the current state:

[0089]

[0090] where, $\tau$ represents the learning rate, $Q$ t denotes the actual reward, denotes the estimated reward, denotes the gradient of the parameter $\pi$;

[0091] S55. The model adaptively adjusts the warning threshold according to the historical operation data and the current health state of the device. The formula for threshold adjustment is:

[0092]

[0093] where, $\theta$ t+1 denotes the warning threshold at the next moment $t + 1$, $\theta$ t denotes the warning threshold at the current $t$ moment, $\xi$ represents the adjustment step size, and $\zeta$ represents the discount factor

[0094] S56. When the operating state of the device changes significantly, the model updates the warning model through the accumulated reward information.

[0095] Optionally, the generation of maintenance suggestions according to the warning level in S6 specifically includes:

[0096] Yellow warning: Prompt the operation and maintenance personnel to pay attention to the device during the next routine inspection and check the key components of the device;

[0097] Orange warning: The system generates more detailed fault speculation and maintenance suggestions, including analyzing historical fault patterns, suggesting replacing or repairing specific components, and arranging maintenance in advance;

[0098] Red warning: The system recommends immediate shutdown or emergency maintenance, analyzes possible fault patterns in detail, such as electrical short circuits, component wear, or insulation damage, and prompts the operation and maintenance personnel to take emergency response measures.

[0099] The beneficial effects of the present invention are as follows:

[0100] First of all, the system realizes real-time monitoring of power transmission and transformation equipment through a multi-parameter sensor module, and collects various operation state data including temperature, vibration, current, voltage, and partial discharge. This multi-parameter monitoring method makes up for the deficiencies of traditional single-parameter monitoring methods, can more comprehensively reflect the operation state of the equipment, especially in complex working conditions with multi-dimensional operation of the equipment, and can significantly improve the accuracy and comprehensiveness of monitoring.

[0101] Secondly, the system adopts a multi-modal data fusion algorithm based on a self-evolving nested network, and combines an adaptive weight assignment model to perform fusion analysis on various sensor data to generate a comprehensive health score. This dynamic weight assignment mechanism enables the data fusion process to adapt to changes in the equipment operation environment, adjusts the contribution degree of each modal data in real time, and avoids false alarms and missed alarms caused by fixed weights in the prior art. In addition, the nested network can continuously optimize its own structure according to the equipment operation state, further improving the accuracy and robustness of data fusion.

[0102] In terms of fault trend prediction, the system introduces a generative adversarial adaptive residual network, and corrects the output and adjusts the architecture of the generator through an adaptive residual module, making the fault prediction of the equipment more accurate. Compared with traditional linear models that cannot handle the complex non-linear changes of equipment, the generative adversarial network of the present invention can capture the dynamic change trend of the equipment health state, continuously optimize the generator output through the feedback mechanism of the residual error, greatly improves the accuracy and timeliness of fault prediction, and significantly reduces the risk of faults.

[0103] In addition, the system adopts an adaptive evolution algorithm based on meta-reinforcement learning, which can dynamically adjust the warning threshold according to the historical operation data and real-time state of the equipment. This technology avoids the problem of inflexible warnings caused by fixed thresholds in the prior art. Through the reinforcement learning mechanism, the system can adaptively optimize the warning strategy according to different equipment and working conditions, improve the accuracy and adaptability of fault warnings, thereby effectively reducing the false alarm rate and missed alarm rate, and enhancing the warning reliability of the system.

[0104] In terms of fault maintenance, the present invention generates corresponding maintenance suggestions based on the fault risk index and warning level. The system can automatically generate yellow, orange, and red warnings, corresponding to different levels of fault handling strategies. Different from the prior art that relies on experience to generate fixed suggestions, the present invention provides a more detailed and accurate maintenance strategy based on multi-dimensional data analysis. The operation and maintenance personnel can perform targeted maintenance according to specific suggestions, thereby reducing the fault handling time and improving the maintenance efficiency.

[0105] Finally, the system realizes the linkage with the power grid dispatching system. When the warning level of the equipment reaches orange or red, the system can automatically detect and adjust the load of the equipment. When necessary, the load is distributed to standby equipment or redundant equipment is started to ensure the stability of the power grid operation. This intelligent linkage mechanism greatly improves the emergency response ability of the system, can effectively control risks in the initial stage of the fault, prevent the fault from further deteriorating, and ensures the stability and safety of the entire power system. Brief Description of the Drawings

[0106] The 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, but do not constitute a limitation to the present invention. In the drawings:

[0107] Figure 1 is a schematic structural diagram of a power transmission and transformation equipment fault warning system based on online monitoring proposed by the present invention;

[0108] Figure 2 is the overall flowchart of a power transmission and transformation equipment fault warning method based on online monitoring proposed by the present invention. Detailed Embodiments

[0109] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0110] Refer to Figure 1 , a power transmission and transformation equipment fault warning system based on online monitoring includes the following modules:

[0111] A multi-parameter sensor module for real-time collection of the operation status data of power transmission and transformation equipment;

[0112] A wireless sensor network module for transmitting the collected operation status data to the data processing center for processing;

[0113] A data fusion module that uses a multi-modal data fusion algorithm based on a self-evolving nested network and combines an adaptive weight allocation model to fuse the modal data of each sensor to generate a comprehensive health score of the equipment;

[0114] A trend prediction module, based on the non - linear time - series analysis technology of the generative adversarial adaptive residual network, and introducing an adaptive residual module to predict the trend of the comprehensive health score and historical operation data of the device;

[0115] An early warning threshold adjustment module, based on the historical data and real - time operation status of the device, adopts an adaptive evolutionary algorithm of meta - reinforcement learning to dynamically adjust the early warning threshold;

[0116] A maintenance recommendation generation module, used to generate maintenance recommendations according to the fault risk index and early warning level when the system detects an anomaly;

[0117] A power grid dispatching linkage module, used to automatically link the power grid dispatching system when the early warning level of the device reaches orange or red, detect and adjust the device load, and allocate the load to standby devices or start redundant devices when necessary.

[0118] Reference Figure 2 , a fault early warning method for power transmission and transformation equipment based on online monitoring, includes the following steps:

[0119] S1. Real - time collect the operation status data of the power transmission and transformation equipment through a multi - parameter sensor, and the operation status data includes temperature, vibration, current, voltage and partial discharge information;

[0120] S2. Transmit the collected operation status data to the data processing center through a wireless sensor network for processing;

[0121] S3. Adopt a multi - modal data fusion algorithm based on a self - evolving nested network, and combine an adaptive weight allocation model for data fusion to generate a comprehensive health score of the device;

[0122] S4. Based on the non - linear time - series analysis technology of the generative adversarial adaptive residual network, introduce an adaptive residual module into the generator to predict the trend of the comprehensive health score and historical operation data of the device, monitor the change of the device health index, and identify the fault trend;

[0123] S5. According to the historical data and real - time operation status of the device, adopt an adaptive evolutionary algorithm based on meta - reinforcement learning to dynamically adjust the early warning threshold;

[0124] S6. When the system detects an anomaly, generate maintenance recommendations according to the early warning level, and the early warning level includes yellow warning, orange warning and red warning;

[0125] S7. When the early warning level of the device reaches orange or red, the system is linked with the power grid dispatching system, automatically detects and adjusts the device load, and allocates the load to standby devices or starts redundant devices when necessary.

[0126] In this embodiment, the specific content of S3 includes:

[0127] S31. For each sensor modality D, the operation status data of the power transmission and transformation equipment collected by the multi-parameter sensor i is preprocessed separately, where i = 1, 2, …, n, and n represents the number of sensor types. The preprocessing includes data denoising, outlier removal, normalization, and time window sliding operation to obtain a standardized data set D′ i ;

[0128] S32. Establish a self-evolving nested network structure. For each sensor modality D′ i Construct an initial nested network N i , the weights of the nested network are initialized with a random normal distribution, and the nested network model N i is defined as:

[0129] F i = N(D′ i ; θ i , L i );

[0130] where θ i represents the parameter set of the nested network N i , L i represents the number of layers of the nested network, and initially, the same number of layers of network structure is used to process each modal data D′ i , and the network output is F i ;

[0131] S33. Use an adaptive weight allocation model to weight the contribution degrees of the outputs of each modality. The weight coefficient w i is dynamically updated through a multi-layer adaptive adjustment formula, and the weight adjustment mechanism is based on the multi-modal non-linear dependence relationship of the sensor data;

[0132] Calculate the initial contribution degree of modality i at time t

[0133]

[0134] where represents the output of modality i at time t, represents the output of modality j at time t, n represents the total number of modalities, and L total represents the total loss function of multi-modal fusion:

[0135]

[0136] where represents the weight coefficient of modality i at time t, y true represents the true equipment health status value, λ represents the regularization coefficient, ∥θi ∥ 2 Denote the L2 norm of network parameters;

[0137] By adding a time decay term, calculate the weight update value of the modality:

[0138]

[0139] wherein, denotes the weight update value of modality i at time t, and ρ denotes the time decay coefficient;

[0140] Add a non - linear self - adjustment function to perform non - linear transformation on the weights:

[0141]

[0142] wherein, exp denotes the exponential function, κ denotes the control factor, and μ denotes the dynamic offset;

[0143] S34. Perform self - evolution processing on the nested network N i The evolution process is based on the combination of genetic algorithm and adaptive neural network structure search algorithm to achieve dynamic optimization of the network structure, including increasing or decreasing the number of network layers, changing the type of activation function, and adjusting the number of neurons;

[0144] S35. Calculate the fitness function of the network structure after each round of evolution. The fitness function Fitness(N i ) is defined as:

[0145]

[0146] wherein, γ denotes the weight balance coefficient, and L i denotes the network depth;

[0147] S36. Dynamically adjust the nested network structure during the evolution process, and select the optimal network structure based on the fitness results The optimal fusion result obtained by each round of iterative calculation

[0148] S37. Repeat steps S33 - S36 until the preset convergence condition is reached, complete the optimal fusion of the self - evolving nested network, and finally generate the comprehensive health score of the device.

[0149] In this embodiment, the specific content of S4 includes:

[0150] S41. Perform multi - dimensional non - linear time - series modeling on the comprehensive health score data H t of the device and the historical operation data D t Define the historical comprehensive health score sequence and the operation data sequence as:

[0151] H t = {H t , H t-1 , …, H t-k}, D t = {D t , D t-1 , …, D t-k};

[0152] Among them, k represents the time window length, and the generator G generates the health score at the future moment. The generated output is:

[0153]

[0154] Among them, represents the generated predicted score, and θ G represents the initial parameter of the generator;

[0155] S42. Calculate the error between the output of the generator and the actual health score by introducing an adaptive residual module and perform feedback correction;

[0156] S43. After being corrected by the adaptive residual module, the output of the generator is updated to:

[0157]

[0158] Among them, represents the updated output of the generator, α t represents the dynamic adjustment coefficient, R t represents the residual defined at the current moment t, R t-i represents the residual at the moment t - i, exp represents the exponential function, k represents the total number of steps, and W (i) represents the weight matrix of the i-th step;

[0159] S44. The generator and the discriminator perform adversarial training through the adaptive residual module. The discriminator D is used to distinguish the real health score and the predicted value generated by the generator:

[0160]

[0161] Among them, m represents the number of samples processed;

[0162] S45. Through multiple adjustments and corrections of the adaptive residual module, obtain the final output prediction of the generator and identify the fault trend.

[0163] In this embodiment, the correction of the adaptive residual module in S42 specifically includes:

[0164] S421. Define the residual R t at the current moment as:

[0165] R t =H t -G(H t-1 ,D t-1 ;θ G );

[0166] where G(H t-1 ,D t-1 ;θ G ) represents the predicted output of the generator at time t - 1;

[0167] S422. When the residual |R t | exceeds the threshold δ1, the generator adjusts the network depth according to the size of the residual:

[0168]

[0169] where L new represents the adjusted network depth, L represents the current network depth, β represents the adjustment factor, represents the generator loss function L G with respect to the parameter θ G gradient, represents the gradient of the residual;

[0170] S423. When the residual δ2 < |R t | ≤ δ1, the adaptive residual module optimizes the network width by adjusting the number of neurons N l per layer:

[0171]

[0172] where, represents the adjusted number of neurons in the l-th layer, N l represents the current number of neurons in the l-th layer, γ1 represents the learning rate, represents the weight matrix at the i-th step of the l-th layer, ∈ represents the smoothing term, and k represents the total number of steps;

[0173] S424. When the residual δ2 < |R t | ≤ δ1, the generator introduces cross-layer connections to enhance the network's expressive ability:

[0174]

[0175] where y l represents the output of the l-th layer, y l+k represents the output of the l + k-th layer, and λ1 represents the adjustment coefficient of the cross-layer connection.

[0176] This cross-layer connection directly transmits low-level information to the high level, enhancing the ability to identify complex non-linear fault trends.

[0177] In this embodiment, step S5 specifically includes:

[0178] S51. Construct the state space of the historical operation data D t and the comprehensive health score H t of the device, and define the state space S t as:

[0179] S t ={D t , H t ,…, D t-k , H t-k};

[0180] where t represents the current time point, and k represents the time window length of the historical data;

[0181] S52. According to the state space S t , define the action space A t of the device. The action space includes operations for dynamically adjusting the warning threshold:

[0182]

[0183] where represents the i-th action, representing different warning threshold adjustment strategies;

[0184] S53. Use the meta-reinforcement learning algorithm to train the adaptive evolution strategy, which specifically includes the design of the reward function and the update of the strategy. The reward function Q t is used to measure the effect of each action:

[0185]

[0186] where represents the loss between the true health score H t and the predicted health score . λ2 represents the regularization coefficient, and ∥θ t ∥ 2 represents the L2 regular term of the parameters;

[0187] S54. Based on the meta-learning mechanism of reinforcement learning, define the strategy update rule as: According to the current state S t and the action a t , update the policy function π(a t |S t ), and select the optimal action a t at each time step according to the current state:

[0188]

[0189] Among them, τ represents the learning rate, Q t represents the actual reward, represents the estimated reward, represents the gradient of the parameter π;

[0190] S55. The model adaptively adjusts the warning threshold according to the historical operation data and the current health status of the device. The formula for threshold adjustment is:

[0191]

[0192] Among them, θ t+1 represents the warning threshold at the next moment t + 1, θ t represents the warning threshold at the current moment t, ξ represents the adjustment step size, and ζ represents the discount factor

[0193] S56. When the operation state of the device changes significantly, the model updates the warning model through the accumulated reward information.

[0194] In this embodiment, the generation of maintenance suggestions according to the warning level in S6 specifically includes:

[0195] Yellow warning: Prompt the operation and maintenance personnel to focus on the device during the next routine inspection and check the key components of the device;

[0196] Orange warning: The system generates more detailed fault speculation and maintenance suggestions, including analyzing historical fault modes, suggesting replacing or repairing specific components, and arranging maintenance in advance;

[0197] Red warning: The system recommends immediate shutdown or emergency repair, analyzes possible fault modes in detail, such as electrical short - circuit, component wear or insulation damage, and prompts the operation and maintenance personnel to take emergency response measures.

[0198] Example 1:

[0199] To verify the feasibility of the present invention in implementation, the present invention is applied to a certain high - voltage transmission and transformation substation in the North China Power Grid. Transmission and transformation equipment has been operating in an environment of high voltage and high load for a long time. The occurrence of equipment failures poses a great threat to the stability and security of the power grid. In the past, since most of the equipment in this substation adopted a traditional single - parameter monitoring system, only physical parameters such as temperature, current or voltage of the equipment were monitored. Once these parameters exceeded the preset fixed threshold, the system would issue an alarm. However, this method has a high false alarm rate, especially in the case of temperature changes and load fluctuations, the false alarm frequency is relatively high, resulting in a large number of ineffective equipment maintenance works. At the same time, since the traditional method fails to fully consider the multi - dimensional operation state of the equipment, many potential equipment failures cannot be detected in time, resulting in a high frequency of sudden equipment failures in the past few years, bringing great hidden dangers to the power grid security.

[0200] To solve these problems, North China Power Grid introduced a fault warning system for power transmission and transformation equipment based on on-line monitoring. This system uses a multi-parameter sensor module, which can collect various operation status data of power transmission and transformation equipment in real time, such as temperature, vibration, current, voltage, partial discharge, etc. These data are transmitted to the data processing center through a wireless sensor network. The system is based on a multi-modal data fusion algorithm of a self-evolving nested network to comprehensively process various sensor data, thereby generating a comprehensive health score of the equipment.

[0201] In a certain application, the temperature data of a main transformer continuously remained higher than the normal range during the peak summer load period. However, due to the traditional system using a fixed warning threshold, it failed to determine whether this anomaly was a real fault hidden danger. The new system not only collected temperature data but also combined multi-dimensional data such as vibration, partial discharge, and current, and performed data fusion through a self-evolving nested network to generate a comprehensive health score of the equipment. The results showed that the comprehensive health score only decreased slightly and did not reach the warning standard of the system. The system dynamically adjusted the weight of the temperature parameter according to the fused health score, avoiding false alarms caused by single data anomalies. Finally, through the trend prediction module, the equipment was determined to be operating normally and did not require shutdown for maintenance, greatly saving maintenance costs.

[0202] In another actual application, for the switchgear in a substation, the vibration and current fluctuated abnormally in the multi-dimensional monitoring data. The traditional system only monitored the current parameter and failed to give an early warning in the initial stage of the anomaly, resulting in an accidental fault tripping event of the switchgear during use, which had a greater impact on the operation of the power grid. After the new system was introduced, in the case where the fault did not occur, the system used the trend prediction technology of a generative adversarial adaptive residual network to identify the fault trend of the equipment in advance. The adaptive residual module deeply analyzed the comprehensive health score and historical operation data of the equipment, and the generated trend prediction data showed that the decline rate of the health index of the equipment increased significantly. The system issued an orange warning, suggesting that the operation and maintenance personnel conduct further inspections on the equipment.

[0203] According to the maintenance suggestions generated by the system, the operation and maintenance personnel quickly inspected the switchgear and found that due to long-term high-load operation, the mechanical components were worn, posing a potential tripping risk. Based on the maintenance suggestions provided by the system, the operation and maintenance personnel replaced the relevant components in advance, successfully avoiding a possible equipment fault tripping event. This example proves that the prediction accuracy of the new system has been greatly improved, which can identify equipment fault risks in advance and provide sufficient time windows for power grid dispatching and maintenance.

[0204] In addition, in another 500 kV substation located in a certain place in Hebei Province, the new system was effectively linked with the power grid dispatching system. During a high-load operation, the partial discharge data of the main transformer equipment showed abnormal fluctuations. Through the trend analysis of the adaptive residual network, the system predicted that the health score of the equipment would continue to decline, and immediately issued a red warning. The system automatically linked the power grid dispatching center, first reduced the load of the equipment, and then automatically started the standby redundant equipment, ensuring the stable operation of the power grid. Finally, the equipment was further detected and repaired during the reduced-load operation, without affecting the operation of the power grid.

[0205] According to the data provided by the substation, the system conducted a total of 256 equipment status evaluations during the one-year period from June 2019 to June 2020. Among them, 78 equipment anomalies were detected, and the system generated 52 yellow warnings, 19 orange warnings, and 7 red warnings respectively. In orange and red warning events, the system linked the power grid dispatching system and initiated 5 redundant equipment startups and 13 load transfer operations. All these operations effectively avoided sudden failures of power grid equipment. Compared with the previous traditional warning system, the false alarm rate of this system was reduced by approximately 60%, and no power supply interruption events caused by sudden equipment failures occurred.

[0206] From these actual application scenarios, it can be seen that the power transmission and transformation equipment fault warning system based on online monitoring of the present invention can significantly improve the accuracy and timeliness of fault warning when dealing with complex power equipment operating environments, reduce false alarms and missed alarms, improve the efficiency of equipment maintenance, and reduce the risks and costs of power system operation. Significant beneficial effects have been achieved in aspects such as equipment fault prediction, warning threshold adjustment, maintenance recommendation generation, and power grid dispatching linkage, greatly improving the safety and stability of the power system.

[0207] Table 1 Comparative analysis table of the status of power transmission and transformation equipment based on traditional and intelligent fault warning systems

[0208]

[0209]

[0210] Through the comparative analysis in Table 1, significant differences were shown between the traditional power transmission and transformation equipment fault warning system and the newly introduced intelligent monitoring system in terms of equipment status monitoring and fault warning. First of all, in terms of the number of equipment status evaluations, the new system evaluated the equipment status 256 times in one year, an increase of 28% compared with 200 times of the traditional system. This improvement is due to the fact that the new system can conduct more accurate and automated monitoring, reducing the need for human intervention and significantly enhancing the frequency and efficiency of equipment status monitoring.

[0211] In terms of the number of device anomalies detected, the traditional system only detected 60 device anomalies, while the new system detected 78, representing a 30% increase. This increase reflects that the new system, through multimodal data fusion and trend prediction technologies, monitors the operating status of devices more comprehensively and precisely, and can identify more potential failure risks.

[0212] Regarding the warning levels, the new system had 52 yellow warning times, while the traditional system had 30, indicating that the new system can detect mild anomalies earlier, which helps to take preventive measures in advance and avoid minor problems evolving into serious faults. At the same time, the number of orange and red warnings also increased significantly, reaching 19 and 7 respectively, which shows that the new system can accurately identify moderate and severe anomalies of the device and issue more targeted alarms, thus improving the response speed and maintenance efficiency of the operation and maintenance personnel.

[0213] In terms of the startup of redundant devices and load transfer operations, the new system is significantly superior to the traditional system. The number of redundant device startups increased from 2 to 5, and the load transfer operations increased from 5 to 13. This reflects the intelligent linkage between the new system and the power grid dispatching system, which can quickly take effective dispatching measures when severe warnings occur in the device, reduce the impact of device failures on the power grid operation, and ensure the stability of the power grid.

[0214] Finally, it is worth noting the significant decrease in the false alarm rate. The false alarm rate of the new system decreased by approximately 60% compared to the traditional system. This result shows that the multi-parameter fusion and adaptive warning threshold adjustment mechanism of the new system can effectively avoid false alarms and reduce ineffective maintenance work. In addition, within one year, there were no power supply interruption events caused by device failures in the new system, while the traditional system had 2 power supply interruptions during the same period. This result fully demonstrates the superiority of the new system in preventing major faults.

[0215] Generally speaking, based on the comparative analysis of this table, it can be concluded that the new system is not only significantly superior to the traditional system in the comprehensiveness and accuracy of device status monitoring, but also effectively reduces the device failure risk through intelligent warning and dispatching linkage, improving the overall operation stability and safety of power transmission and transformation equipment. This shows that the new system has strong application prospects and technical advantages in the field of power transmission and transformation equipment warning and maintenance.

[0216] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An on-line monitoring-based power transmission and transformation equipment fault warning system, characterized in that It includes the following modules: A multi-parameter sensor module for collecting the operation status data of power transmission and transformation equipment in real time; A wireless sensor network module for transmitting the collected operation status data to a data processing center for processing; A data fusion module that uses a multi-modal data fusion algorithm based on a self-evolving nested network and combines an adaptive weight allocation model to fuse the modal data of each sensor, generating a comprehensive health score for the equipment. Specifically, it includes the following steps: S31. For the operation status data of power transmission and transformation equipment collected by multi-parameter sensors, preprocess each sensor modality separately, where , represents the number of sensor types, and the preprocessing includes data denoising, outlier removal, normalization, and time window sliding operation to obtain a standardized data set ; S32. Establish a self-evolving nested network structure for each sensor modality Construct an initial nested network , initialize the weights of the nested network to a random normal distribution, and define the nested network model as follows: ； Among them, represents the parameter set of the nested network , represents the number of layers of the nested network, and initially, for each modality data the same number of layers of network structure is used for processing, and the network output is ; S33. Weight the contribution degrees of the outputs of each modality using an adaptive weight allocation model, and the weight coefficients are dynamically updated through a multi-layer adaptive adjustment formula, and the weight adjustment mechanism is based on the multi-modal non-linear dependence relationship of the sensor data; Calculation mode At Initial contribution degree at a moment : ； Among them, represents the output at time modality , represents the output at time modality , represents the total number of modalities, represents the total loss function of multimodal fusion: ； Among them, represents the modality at the weight coefficient at the moment, represents the true device health status value, represents the regularization coefficient, represents the L2 norm of the network parameters; By adding a time decay term, calculate the weight update value of the modality: ； Among them, represents the modality at the weight update value at the moment, represents the time decay coefficient; Add a non-linear self-regulation function to perform a non-linear transformation on the weights: ； Among them, represents an exponential function, represents a control factor, represents a dynamic offset; S34. Perform self-evolution processing on the nested network The evolution process is based on the combination of genetic algorithm and adaptive neural network structure search algorithm to achieve dynamic optimization of the network structure, including increasing or decreasing the number of network layers, changing the type of activation function, and adjusting the number of neurons; S35. Calculate the fitness function of the network structure after each round of evolution. The fitness function is defined as: ； Among them, represents the weight balance coefficient, represents the network depth; S36. Dynamically adjust the nested network structure during the evolution process, and select the optimal network structure based on the fitness results , the optimal fusion result obtained by each round of iteration ; S37. Repeat steps S33 - S36 until a preset convergence condition is reached, complete the optimal fusion of the self-evolving nested network, and finally generate a comprehensive health score for the equipment; A trend prediction module that uses a non-linear time series analysis technique based on a generative adversarial adaptive residual network and introduces an adaptive residual module to predict the trend of the comprehensive health score and historical operation data of the equipment; An early warning threshold adjustment module that uses an adaptive evolution algorithm of meta-reinforcement learning to dynamically adjust the early warning threshold based on the historical data and real-time operation status of the equipment; A maintenance recommendation generation module for generating maintenance recommendations according to the failure risk index and early warning level when the system detects an anomaly; A power grid dispatching linkage module for automatically linking the power grid dispatching system when the equipment early warning level reaches orange or red, detecting and adjusting the equipment load, and if necessary, allocating the load to standby equipment or starting redundant equipment.

2. The on-line monitoring-based power transmission and transformation equipment fault warning system according to claim 1 performs an on-line monitoring-based power transmission and transformation equipment fault warning method, characterized in that, It includes the following steps: S1. Collect the operation status data of power transmission and transformation equipment in real time through a multi-parameter sensor. The operation status data includes temperature, vibration, current, voltage, and partial discharge information; S2. Transmit the collected operation status data to a data processing center for processing through a wireless sensor network; S3. Use a multi-modal data fusion algorithm based on a self-evolving nested network and combine an adaptive weight allocation model for data fusion to generate a comprehensive health score for the equipment; S4. Based on a non-linear time series analysis technique of a generative adversarial adaptive residual network, introduce an adaptive residual module in the generator to predict the trend of the comprehensive health score and historical operation data of the equipment, monitor the change of the equipment health index, and identify the fault trend; S5. Dynamically adjust the early warning threshold according to the historical data and real-time operation status of the equipment by using an adaptive evolution algorithm based on meta-reinforcement learning; S6. When the system detects an anomaly, generate maintenance recommendations according to the early warning level. The early warning level includes yellow warning, orange warning, and red warning; S7. When the equipment early warning level reaches orange or red, the system is linked with the power grid dispatching system to automatically detect and adjust the equipment load, and if necessary, allocate the load to standby equipment or start redundant equipment.

3. The on-line monitoring-based power transmission and transformation equipment fault early warning system according to claim 2, characterized in that, The specific content of S4 includes: S41. The comprehensive health score data of the device and historical operation data are used to perform multi-dimensional non-linear time series modeling. The historical comprehensive health score sequence and the operation data sequence are defined as follows: ； Among them, represents the time window length, and the generator generates the health score at future times, and the generated output is: ； Among them, represents the generated predicted score, represents the initial parameters of the generator; S42. Calculate the error between the output of the generator and the actual health score by introducing an adaptive residual module and perform feedback correction; S43. After the correction by the adaptive residual module, the output of the generator is updated to: ； Among them, represents the updated generator output, represents the dynamic adjustment coefficient, represents defining the current moment's residual, represents moment's residual, represents the exponential function, represents the time window length, represents the weight matrix at the step; S44. The generator and discriminator are adversarially trained through an adaptive residual module, and the discriminator is used to distinguish between the true health score and the predicted value generated by the generator: ； Among them, represents the number of samples processed; S45. Through multiple adjustments and corrections of the adaptive residual module, the final output prediction of the generator is obtained, and the fault trend is identified.

4. The on-line monitoring-based power transmission and transformation equipment fault warning system according to claim 3, characterized in that The correction of the adaptive residual module in S42 specifically includes: S421. Define the residual at the current moment as follows: ； Among them, represents the predicted output of the generator at time ; S422. When the residual exceeds the threshold , the generator will adjust the network depth according to the size of the residual: ； Among them, represents the adjusted network depth, represents the current network depth, represents the adjustment factor, represents the generator loss function with respect to the parameter gradient, represents the gradient of the residual; S423. When the residual is present, the adaptive residual module optimizes the network width by adjusting the number of neurons in each layer as follows: ； Among them, represents the number of neurons in the layer after adjustment, represents the current number of neurons in the layer, represents the learning rate, represents the layer, step weight matrix, represents the smoothing term, represents the time window length; S424. When the residual is present, the generator introduces cross-layer connections to enhance the network's expressive power: ； Among them, represents the output of the layer, represents the output of the layer, represents the adjustment coefficient of the cross-layer connection.

5. The fault warning system for power transmission and transformation equipment based on online monitoring according to claim 2, wherein, S5 specifically includes: S51. Construct the historical operation data of the device and the comprehensive health score of the state space, and define the state space as follows: ； Among them, represents the current time point, represents the time window length; S52. Define the action space of the device according to the state space , where the action space includes an operation of dynamically adjusting the warning threshold: ​ ； Among them, represents the th action, representing different warning threshold adjustment strategies; S53. Training an adaptive evolutionary strategy using a meta-reinforcement learning algorithm, specifically including the design of a reward function and policy update. The reward function is used to measure the effect of each action: ； Among them, represents the true health score and the predicted health score the loss between them, represents the regularization coefficient, represents the L2 regularization term of the parameter; S54. The meta - learning mechanism based on reinforcement learning defines the policy update rule as follows: According to the current state and action , update the policy function , and select the optimal action according to the current state at each time step : ； Among them, represents the learning rate, represents the actual reward, represents the estimated reward, represents the parameter gradient; S55. The model adaptively adjusts the warning threshold according to the historical operation data and the current health status of the device. The formula for threshold adjustment is: ； Among them, represents the early warning threshold for the next moment , represents the early warning threshold for the current moment represents the adjustment step size represents the discount factor; S56. When the operation status of the device changes significantly, the model updates the warning model through the accumulated reward information.

6. The fault warning system for power transmission and transformation equipment based on online monitoring according to claim 2, characterized in that The generation of maintenance suggestions according to the warning level in S6 specifically includes: Yellow warning: Prompt the operation and maintenance personnel to pay attention to the device during the next routine inspection and check the key components of the device; Orange warning: The system generates more detailed fault speculation and maintenance suggestions, including analyzing historical fault patterns, suggesting replacing or repairing specific components, and arranging maintenance in advance; Red warning: The system recommends immediate shutdown or emergency maintenance, analyzes possible fault patterns in detail, including electrical short circuits, component wear or insulation damage, and prompts the operation and maintenance personnel to take emergency response measures.

Images