Intelligent diagnosis and adaptive maintenance method and system for substation monitoring data

Abstract

The application discloses a kind of intelligent diagnosis and adaptive maintenance methods and systems for substation monitoring data, it is related to the field of substation diagnosis and maintenance, comprising: constructing multimodal monitoring database;Extract the time-frequency domain stable feature of vibration waveform, establish equipment health fingerprint benchmark;Based on multi-core learning support vector machine, output the fault type probability distribution of this operation and the contribution weight of each modal feature;Based on dynamic Bayesian network, output the fault probability evolution trend in future operation window;Dynamically generate adaptive maintenance strategy;Verify maintenance effect, according to the health degree change before and after maintenance, correct dynamic Bayesian network parameter;Trigger new fault mode mining, update the fault category library of multi-core learning support vector machine.The application has the advantages that: adaptive evaluation, degradation trend prediction and closed-loop maintenance strategy optimization of equipment health status are realized, and the accuracy of substation key equipment fault diagnosis and the intelligent level of maintenance are improved.

Description

Technical Field

[0001] This invention relates to the field of substation diagnosis and maintenance, specifically to an intelligent diagnosis and adaptive maintenance method and system based on substation monitoring data. Background Technology

[0002] With the deepening of smart grid construction, unmanned operation of substations has become the mainstream mode of modern power system operation. Circuit breakers, disconnectors, and on-load tap changers, as key execution units in substations, directly determine power supply reliability and operation and maintenance costs based on their mechanical condition. Existing condition monitoring methods mostly rely on single electrical quantities (such as opening and closing coil current) or simple threshold alarms, making it difficult to capture early mechanical degradation characteristics (such as buffer spring fatigue, mechanism jamming, and contact wear). This often leads to faults being discovered only after sudden failures, causing unplanned power outages or even equipment damage. Wideband vibration waveforms and ultrasonic acoustic waveforms generated during equipment operation contain rich mechanical condition information, reflecting subtle changes such as component wear, loosening, and poor lubrication. However, single vibration or acoustic waveform analysis is easily affected by environmental noise and fluctuations in operating conditions, resulting in a high false alarm rate. Multimodal fusion of vibration, acoustic waveforms, electrical quantities, and condition quantities can create complementary advantages and effectively suppress the randomness and uncertainty of single-source information.

[0003] Methods based on multimodal sensing (vibration, acoustic signature, electrical, and state variables) and machine learning enable refined analysis of equipment operation processes and health status assessments, representing a crucial direction for improving the intelligence level of substation equipment operation and maintenance. Furthermore, integrating temporal probabilistic prediction and dynamic Bayesian networks allows for early prediction of fault evolution trends, providing technical support for the shift from "periodic maintenance" to "predictive maintenance." Simultaneously, through maintenance effect feedback and automatic discovery of unknown fault modes, the diagnostic model can continuously evolve, adapting to new degradation mechanisms during long-term operation.

[0004] Traditional methods often rely on manually preset rule bases or single threshold alarms, lacking early identification capabilities for complex mechanical faults (such as spring fatigue, contact wear, and mechanism jamming). Furthermore, rule updates depend on expert experience, making it difficult to adapt to performance drift during long-term equipment operation. Existing maintenance strategies mostly employ periodic inspections or reactive maintenance, failing to dynamically optimize maintenance timing and content based on the equipment's current health status, degradation trends, and specific fault modes, leading to both over-maintenance and neglect. Regarding diagnostic model evolution, there is a lack of automatic discovery mechanisms for unknown fault modes; newly emerging fault types often require extensive manual analysis and labeling before being incorporated into the diagnostic system, resulting in long response cycles. In terms of decision feedback, there is a lack of adaptive closed-loop mechanisms that utilize maintenance effects to correct predictive models, preventing continuous improvement in model accuracy with the accumulation of operational data. Summary of the Invention

[0005] To address the aforementioned technical issues, this paper provides an intelligent diagnostic and adaptive maintenance method and system for substation monitoring data. This technical solution solves the problems mentioned in the background technology, such as reliance on manually preset rule bases or single threshold alarms, lack of early identification capability for complex mechanical faults, reliance on periodic or post-event maintenance, lack of automatic mining of unknown fault modes, and lack of adaptive closed-loop correction of prediction models using maintenance effects.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An intelligent diagnostic and adaptive maintenance method for substation monitoring data includes:

[0008] Collect three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical quantities and status quantities during the operation of substation equipment. Preprocess the raw data to construct a multimodal monitoring database.

[0009] Based on continuous operation data confirmed to be in a healthy state after initial commissioning or maintenance, the time-frequency domain stability characteristics of the vibration waveform are extracted to establish a health fingerprint benchmark for the equipment, and the benchmark is adaptively updated using the exponential weighted moving average method.

[0010] For each equipment operation, multimodal feature vectors are extracted from multimodal monitoring data. Based on multi-core learning support vector machine, the probability distribution of fault type and the contribution weight of each modality feature for this operation are output.

[0011] The probability distribution sequence of failure types from multiple consecutive operations is used as the observation sequence, and the trend term of the key modal feature weight in the contribution weight sequence is used as the driving parameter of the state transition model. Both are input into the dynamic Bayesian network to output the evolution trend of failure probability within the future operation window.

[0012] Based on the evolution trend of failure probability and the mutual verification of diagnostic results of multiple devices within the same interval, an adaptive maintenance strategy is dynamically generated.

[0013] After performing maintenance operations, waveform data for the next equipment operation is collected again. The hidden state estimation is updated by multi-core learning support vector machine and dynamic Bayesian network to verify the maintenance effect. The parameters of dynamic Bayesian network are then corrected based on the changes in health status before and after maintenance.

[0014] When the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold, new fault mode mining is triggered. The vibration waveform is converted into a symbol sequence by symbol aggregation approximation, and frequent patterns are extracted by sequence pattern mining algorithm to update the fault category library of the multi-core learning support vector machine.

[0015] Preferably, the preprocessing of the raw data, including the acquisition of three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical and status quantities during the operation of substation equipment, specifically includes:

[0016] Install a triaxial MEMS accelerometer and an ultrasonic microphone in the mechanism box of a circuit breaker, disconnector, or on-load tap changer. The acquisition and recording of waveform data from before to after the operation are triggered by the equipment operation command.

[0017] Simultaneously acquire the current waveform of the opening and closing coils, SF6 gas density / pressure, cumulative number of operations, and the amplitude of the disconnection current;

[0018] All collected data are interpolated and downsampled to align with a unified timestamp, and each physical quantity is normalized to the [0,1] interval to eliminate the influence of dimensions.

[0019] After alignment and normalization, outlier detection based on the 3σ criterion is performed to remove impact interference spikes in the vibration waveform that exceed the mean ± 3 times the standard deviation, and the removed points are replaced with the mean of the preceding and following neighborhoods.

[0020] The wavelet soft thresholding denoising method is adopted, and the Sym8 wavelet basis is selected to perform a 4-level decomposition of the vibration signal to obtain the low-frequency approximation coefficients and the high-frequency detail coefficients of each level.

[0021] The low-frequency approximation coefficients are retained, and a threshold is set for the high-frequency detail coefficients. The high-frequency detail coefficients that exceed the threshold are retained, and the denoised vibration waveform is reconstructed.

[0022] The ambient background noise is suppressed by spectral subtraction on the acoustic waveform, that is, the noise power spectrum when there is no operation is estimated, and the time-domain acoustic waveform is reconstructed after subtracting the noise spectrum from the power spectrum of the operation waveform.

[0023] A multimodal monitoring database is constructed, and the pre-processed vibration waveform, acoustic waveform, and aligned electrical and state quantities are stored together in the multimodal monitoring database.

[0024] Preferably, the step of extracting the time-frequency domain stability features of the vibration waveform, establishing a device health fingerprint benchmark, and adaptively updating the benchmark using an exponentially weighted moving average method specifically includes:

[0025] After the equipment is put into operation for the first time or its health status is confirmed during maintenance, the vibration waveforms of the previous several normal operations are continuously collected.

[0026] For each waveform, the stable features of the time domain peak, frequency domain centroid, and waveform entropy are extracted, and the median of each feature is taken as the baseline vector of the device health fingerprint.

[0027] For each subsequent operation, the Euclidean distance between the extracted feature vector and the reference vector is calculated as the deviation.

[0028] The baseline vector is updated using an exponentially weighted moving average, and the update is performed only when the deviation of the current operation is less than a distance threshold.

[0029] Preferably, the step of outputting the fault type probability distribution and the contribution weights of each modality feature based on multi-core learning support vector machine specifically includes:

[0030] For each device operation, time-domain features are extracted from the denoised vibration waveform. These time-domain features include: peak value, peak-to-peak value, number of pulses, waveform entropy, and energy decay rate.

[0031] The same vibration waveform is subjected to a fast Fourier transform to extract frequency domain features, which include: frequency centroid, dominant frequency shift, and harmonic energy ratio.

[0032] The energy proportion, zero-crossing rate, and Mel frequency cepstral coefficients of the frequency band above 20kHz are extracted from the denoised audio waveform.

[0033] Extract electrical characteristics such as peak current of the opening and closing coils, operating time, and harmonic distortion rate from electrical quantities;

[0034] Extract state characteristics from the cumulative number of operations and the SF6 pressure change rate from the state variables;

[0035] Based on time-domain features, frequency-domain features, voiceprint features, electrical features, state features, and current operation deviation, a multimodal feature vector is formed by feature concatenation.

[0036] A fault category library is pre-built, including but not limited to the following categories: normal, buffer spring fatigue, contact wear, mechanism jamming, connecting rod loosening, base bolt loosening, and changes in buffer characteristics caused by SF6 leakage;

[0037] For feature groups of different modalities, RBF kernel functions are constructed independently, and the kernel width parameter is determined by cross-validation;

[0038] The SimpleMKL algorithm is used to learn the linear combination weights of each kernel, and the kernel weights are the contribution weights of the corresponding features in this diagnosis.

[0039] The combined kernel is input into the standard C-SVM classifier, which outputs the decision value for each fault category. Then, the decision value is transformed into a probability distribution by Platt scaling.

[0040] Output the probability distribution of the fault type and the contribution weight of each modal feature for this operation.

[0041] Preferably, the step of using the probability distribution sequence of failure types from multiple consecutive operations as the observation sequence, and using the trend term of the key modal feature weights in the contribution weight sequence as the driving parameter of the state transition model, and inputting them together into the dynamic Bayesian network to output the evolution trend of failure probability within the future operation window specifically includes:

[0042] Construct a dynamic Bayesian network and set the hidden state variables to include continuous health variables and discrete remaining life interval variables;

[0043] The probability distribution sequence of fault types obtained from multiple consecutive operations is used as the observation variable and input into the network.

[0044] The key modal feature weight trend term in the contribution weight sequence obtained from multiple consecutive operations is used as the driving parameter of the state transition model.

[0045] The state transition model adopts a linear drift form with load-driven and degradation mode correction;

[0046] The observation model uses logistic regression to map hidden health status to a failure probability distribution;

[0047] Particle filtering is used for inference. At each step, the next health level is predicted based on the state transition model with correction. The likelihood is calculated as the particle weight based on the current failure probability distribution. After normalization and resampling, the weighted average of the health levels of all particles is taken as the health level estimate.

[0048] When making forward predictions, only the iterative state transition model is used to obtain the future health distribution, which is then converted into failure probabilities by the observation model, and the evolution trend of failure probabilities within the future operating window is output.

[0049] Preferably, the step of dynamically generating an adaptive maintenance strategy based on the failure probability evolution trend and the mutual verification of diagnostic results from multiple devices within the same interval specifically includes:

[0050] Based on the evolution trend of failure probability and the mutual verification of diagnostic results of multiple devices within the same interval, an adaptive maintenance strategy is dynamically generated, including: multi-device collaborative diagnosis and adaptive maintenance strategy generation.

[0051] Multi-device collaborative diagnosis verifies and confirms the confidence level of abnormal alarms of a single device by examining the causal and temporal relationships between related devices. When multiple devices exhibit consistent degradation characteristics, it is determined to be a common-cause fault, triggering joint maintenance for the entire interval.

[0052] The adaptive maintenance strategy generates predictive maintenance work orders when the predicted failure probability exceeds a probability threshold, and determines the dominant degradation type based on feature weight trends to determine the maintenance content.

[0053] Adjust operating parameters for specific fault types and generate joint maintenance plans for common-cause faults.

[0054] Preferably, the verification of maintenance effectiveness and the correction of dynamic Bayesian network parameters based on changes in health before and after maintenance specifically include:

[0055] After the maintenance operation is completed, the vibration waveform, acoustic waveform, electrical quantity, and status quantity of the equipment are collected again for the next normal operation to obtain a new health estimate.

[0056] Based on data statistics, a high health threshold, a high confidence threshold, and a minimum effective improvement amount are set;

[0057] If the health estimate is higher than the high health threshold and the normal class probability output by the multi-core learning support vector machine is higher than the high confidence threshold, then the maintenance is deemed effective and the frequency of subsequent maintenance is reduced.

[0058] If the estimated improvement in health after maintenance is less than the minimum effective improvement, the maintenance is deemed ineffective and will be upgraded from routine maintenance to in-depth maintenance, triggering a comprehensive inspection.

[0059] The deviation between the actual degradation rate and the model prediction is calculated, and the transition model parameters of the dynamic Bayesian network are corrected through Bayesian updates.

[0060] Preferably, triggering new fault mode mining when the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than a set threshold specifically includes:

[0061] Based on data statistics, low confidence threshold, minimum support threshold, and similarity threshold are set;

[0062] When the maximum probability of the output of a multi-core learning support vector machine for multiple consecutive operations is lower than the low confidence threshold, the new fault mode discovery process is automatically triggered. The specific steps are as follows:

[0063] Extract the vibration waveforms from these operations, divide the waveforms into several segments along the time axis, calculate the mean value of the waveform amplitude within each time window, and map the mean value onto a set alphabet to obtain a string of fixed length.

[0064] A sequence pattern mining algorithm is used to mine common subsequence patterns that occur at a frequency no less than the minimum support from multiple strings;

[0065] The discovered patterns are compared one by one with the regular expressions in the existing fault mode library to calculate the string similarity. If the similarity is lower than the similarity threshold, the pattern is pushed to the operation and maintenance experts.

[0066] After experts confirmed the new fault type through on-site inspection, they converted the symbol sequence pattern into a regular expression and added it as a new fault category to the fault category library of the multi-kernel learning support vector machine. At the same time, they generated a set of typical feature vectors for this category for kernel function calculation.

[0067] Furthermore, this solution proposes an intelligent diagnostic and adaptive maintenance system for substation monitoring data, used to implement the aforementioned intelligent diagnostic and adaptive maintenance method for substation monitoring data, including:

[0068] The data acquisition module is used to acquire three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical quantities and status quantities during the operation of substation equipment, preprocess the raw data, and construct a multimodal monitoring database.

[0069] The probability prediction module is used to extract the time-frequency domain stability features of vibration waveforms based on multiple consecutive operation data confirmed to be in a healthy state after initial commissioning or maintenance, establish a device health fingerprint benchmark, and adaptively update the benchmark using an exponentially weighted moving average method. For each device operation, multimodal feature vectors are extracted from multimodal monitoring data, and based on a multi-kernel learning support vector machine, the probability distribution of fault types and the contribution weights of each modal feature for this operation are output. The sequence of fault type probability distributions from multiple consecutive operations is used as the observation sequence, and the trend term of the key modal feature weights in the contribution weight sequence is used as the driving parameter of the state transition model. Both are input into a dynamic Bayesian network to output the evolution trend of fault probability within the future operation window.

[0070] The maintenance and update module is used to dynamically generate adaptive maintenance strategies based on the evolution trend of fault probability and the mutual verification of diagnostic results of multiple devices within the same interval. After the maintenance operation is performed, waveform data of the next device operation is collected again, and the hidden state estimation is updated through a multi-core learning support vector machine and a dynamic Bayesian network to verify the maintenance effect. The parameters of the dynamic Bayesian network are also corrected based on the changes in health before and after maintenance. When the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold, a new fault mode mining is triggered. The vibration waveform is converted into a symbol sequence using symbol aggregation approximation, and frequent patterns are extracted through a sequence pattern mining algorithm to update the fault category library of the multi-core learning support vector machine.

[0071] Preferably, the probability prediction module includes:

[0072] The stability feature unit is used to extract the time-frequency domain stability features of the vibration waveform based on multiple consecutive operation data confirmed to be in a healthy state after initial commissioning or maintenance, establish a health fingerprint benchmark for the equipment, and adaptively update the benchmark using an exponentially weighted moving average method.

[0073] The fault prediction unit is used to extract multimodal feature vectors from multimodal monitoring data for each equipment operation, and output the probability distribution of fault type and the contribution weight of each modality feature based on multi-core learning support vector machine.

[0074] The probability evolution unit is used to take the probability distribution sequence of failure types from multiple consecutive operations as the observation sequence and the key modal feature weight trend term in the contribution weight sequence as the driving parameter of the state transition model, and input them together into the dynamic Bayesian network to output the failure probability evolution trend within the future operation window.

[0075] The maintenance and update module includes:

[0076] The maintenance strategy unit is used to dynamically generate an adaptive maintenance strategy based on the failure probability evolution trend and the mutual verification of the diagnostic results of multiple devices within the same interval.

[0077] The verification and optimization unit is used to re-collect waveform data of the next equipment operation after performing maintenance operations, update the hidden state estimation through multi-core learning support vector machine and dynamic Bayesian network, verify the maintenance effect, and correct the dynamic Bayesian network parameters according to the changes in health before and after maintenance.

[0078] The fault category update unit is used to trigger new fault mode mining when the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold. The unit uses symbol aggregation approximation to convert the vibration waveform into a symbol sequence, extracts frequent patterns through a sequence pattern mining algorithm, and updates the fault category library of the multi-core learning support vector machine.

[0079] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0080] This invention constructs a multimodal monitoring database by synchronously collecting vibration, acoustic, electrical, and state variables. It then utilizes a multi-core learning support vector machine to adaptively fuse features from various modes, enabling earlier detection of early mechanical degradation characteristics such as buffer spring fatigue, contact wear, and mechanism jamming, significantly reducing false alarm and missed alarm rates. By using the fault probability distribution sequence from multiple consecutive operations as observed variables and the contribution weight trend term as the driving parameter for the state transition model, a dynamic Bayesian network with degradation mode correction is constructed. Through particle filtering and inference of health evolution, the fault probability evolution trend within future operating windows can be output in advance, achieving a shift from "periodic maintenance" to "predictive maintenance." The contribution weight sequence is not only used for diagnostic interpretability but also serves as a dynamic correction parameter for the state transition model, enabling... The predictive model can respond to changes in degradation patterns in real time, improving the accuracy and robustness of trend prediction. It utilizes the fixed operating sequence of circuit breakers and disconnectors within the same bay for multi-device collaborative diagnosis and causal verification, effectively reducing false alarms caused by single-point interference and identifying bay-level common-cause faults that trigger joint maintenance of the entire bay. Through maintenance effect feedback and Bayesian updates, it corrects model parameters in real time. Simultaneously, it automatically discovers unknown fault modes and updates the fault database using low-confidence-driven symbolic aggregation approximation and sequence pattern mining, forming a continuously evolving closed loop. Finally, based on the evolution trend of fault probability and the monotonicity of feature weights, it determines the dominant degradation type, automatically generates differentiated maintenance strategies and adjusts operating parameters, achieving an upgrade from "one-size-fits-all" periodic maintenance to state-driven, refined, and precise maintenance. Attached Figure Description

[0081] Figure 1 This is a flowchart of the intelligent diagnosis and adaptive maintenance method for substation monitoring data according to the present invention;

[0082] Figure 2 The flowchart below shows the process of establishing a device health fingerprint benchmark and using an exponentially weighted moving average method to adaptively update the benchmark for extracting the time-frequency domain stability characteristics of the vibration waveform in this invention.

[0083] Figure 3 The flowchart for the multi-core learning support vector machine of this invention outputs the probability distribution of fault types and the contribution weights of each modal feature in this operation.

[0084] Figure 4 This is a flowchart illustrating the evolution trend of fault probability within the future operation window of the present invention. Detailed Implementation

[0085] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0086] Implementation Example 1

[0087] Reference Figure 1 As shown, an intelligent diagnostic and adaptive maintenance method for substation monitoring data includes:

[0088] Collect three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical quantities and status quantities during the operation of substation equipment. Preprocess the raw data to construct a multimodal monitoring database.

[0089] Based on continuous operation data confirmed to be in a healthy state after initial commissioning or maintenance, the time-frequency domain stability characteristics of the vibration waveform are extracted to establish a health fingerprint benchmark for the equipment, and the benchmark is adaptively updated using the exponential weighted moving average method.

[0090] For each equipment operation, multimodal feature vectors are extracted from multimodal monitoring data. Based on multi-core learning support vector machine, the probability distribution of fault type and the contribution weight of each modality feature for this operation are output.

[0091] The probability distribution sequence of failure types from multiple consecutive operations is used as the observation sequence, and the trend term of the key modal feature weight in the contribution weight sequence is used as the driving parameter of the state transition model. Both are input into the dynamic Bayesian network to output the evolution trend of failure probability within the future operation window.

[0092] Based on the evolution trend of failure probability and the mutual verification of diagnostic results of multiple devices within the same interval, an adaptive maintenance strategy is dynamically generated.

[0093] After performing maintenance operations, waveform data for the next equipment operation is collected again. The hidden state estimation is updated by multi-core learning support vector machine and dynamic Bayesian network to verify the maintenance effect. The parameters of dynamic Bayesian network are then corrected based on the changes in health status before and after maintenance.

[0094] When the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold, new fault mode mining is triggered. The vibration waveform is converted into a symbol sequence by symbol aggregation approximation, and frequent patterns are extracted by sequence pattern mining algorithm to update the fault category library of the multi-core learning support vector machine.

[0095] The preprocessing of the raw data collected during the operation of substation equipment, including three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical and status quantities, specifically includes:

[0096] Install a triaxial MEMS accelerometer and an ultrasonic microphone in the mechanism box of a circuit breaker, disconnector, or on-load tap changer. The acquisition and recording of waveform data from before to after the operation are triggered by the equipment operation command.

[0097] Simultaneously acquire the current waveform of the opening and closing coils, SF6 gas density / pressure, cumulative number of operations, and the amplitude of the disconnection current;

[0098] All collected data are interpolated and downsampled to align with a unified timestamp, and each physical quantity is normalized to the [0,1] interval to eliminate the influence of dimensions.

[0099] After alignment and normalization, outlier detection based on the 3σ criterion is performed to remove impact interference spikes in the vibration waveform that exceed the mean ± 3 times the standard deviation, and the removed points are replaced with the mean of the preceding and following neighborhoods.

[0100] The wavelet soft thresholding denoising method is adopted, and the Sym8 wavelet basis is selected to perform a 4-level decomposition of the vibration signal to obtain the low-frequency approximation coefficients and the high-frequency detail coefficients of each level.

[0101] The low-frequency approximation coefficients are retained, and a threshold is set for the high-frequency detail coefficients. The high-frequency detail coefficients that exceed the threshold are retained, and the denoised vibration waveform is reconstructed.

[0102] The ambient background noise is suppressed by spectral subtraction on the acoustic waveform, that is, the noise power spectrum when there is no operation is estimated, and the time-domain acoustic waveform is reconstructed after subtracting the noise spectrum from the power spectrum of the operation waveform.

[0103] A multimodal monitoring database is constructed, and the pre-processed vibration waveform, acoustic waveform, and aligned electrical and state quantities are stored together in the multimodal monitoring database.

[0104] In the data acquisition phase, a triaxial MEMS accelerometer and an ultrasonic microphone are deployed within the mechanism box of the circuit breaker, disconnector, or on-load tap changer. The triaxial MEMS accelerometer directly outputs the waveform of the vibration acceleration of the equipment surface over time (i.e., vibration waveform), with the X, Y, and Z channels reflecting mechanical impact and friction information in different directions, respectively. The ultrasonic microphone outputs a high-frequency waveform of sound pressure changing over time (i.e., ultrasonic acoustic waveform), used to capture high-frequency ultrasonic signals generated by partial discharge or abnormal friction. Both sensors are synchronously triggered by equipment operation commands, recording from before the operation to after the operation, thus obtaining a complete transient process waveform. Simultaneously, the current waveform of the opening and closing coils is synchronously acquired to obtain the action sequence and status of the electromagnetic operating mechanism; SF6 gas density / pressure is acquired to monitor whether the arc-extinguishing medium is normal; the cumulative number of operations and the amplitude of the interrupted current are recorded to assess the electrical life and wear degree of the equipment. The vibration waveform and acoustic waveform are converted from analog to digital to become discrete time series, which, together with the synchronously acquired electrical quantities and status quantities, constitute the basic data of the multimodal database. After alignment with a unified timestamp, a holographic record of the operation process can be established.

[0105] Reference Figure 2 As shown, the extraction of the time-frequency domain stability features of the vibration waveform, the establishment of a device health fingerprint benchmark, and the adaptive updating of the benchmark using an exponentially weighted moving average method specifically include:

[0106] After the equipment is put into operation for the first time or its health status is confirmed during maintenance, the vibration waveforms of the previous several normal operations are continuously collected.

[0107] For each waveform, the stable features of the time domain peak, frequency domain centroid, and waveform entropy are extracted, and the median of each feature is taken as the baseline vector of the device health fingerprint.

[0108] For each subsequent operation, the Euclidean distance between the extracted feature vector and the reference vector is calculated as the deviation.

[0109] The baseline vector is updated using an exponentially weighted moving average, and the update is performed only when the deviation of the current operation is less than a distance threshold.

[0110] In the feature extraction and health fingerprint benchmark establishment stage, after the equipment is put into operation for the first time or confirmed to be healthy after maintenance, the vibration waveforms of the previous normal operation are continuously collected, and the stable features of the time domain peak value, frequency domain centroid and waveform entropy are extracted. The median of each feature is taken as the health fingerprint benchmark vector. After each subsequent operation, the feature vector is extracted, and the Euclidean distance between it and the benchmark vector is calculated as the deviation. The benchmark is updated adaptively using the exponential weighted moving average method, and the update is only performed when the current deviation is lower than the distance threshold. This not only adapts to the benign aging of the equipment, but also avoids the benchmark being contaminated by abnormal data, providing an accurate comparison basis for fault diagnosis.

[0111] Reference Figure 3 As shown, the step of using a multi-kernel learning support vector machine to output the probability distribution of fault types and the contribution weights of each modality feature for this operation specifically includes:

[0112] For each device operation, time-domain features are extracted from the denoised vibration waveform. These time-domain features include: peak value, peak-to-peak value, number of pulses, waveform entropy, and energy decay rate.

[0113] The same vibration waveform is subjected to a fast Fourier transform to extract frequency domain features, which include: frequency centroid, dominant frequency shift, and harmonic energy ratio.

[0114] The energy proportion, zero-crossing rate, and Mel frequency cepstral coefficients of the frequency band above 20kHz are extracted from the denoised audio waveform.

[0115] Extract electrical characteristics such as peak current of the opening and closing coils, operating time, and harmonic distortion rate from electrical quantities;

[0116] Extract state characteristics from the cumulative number of operations and the SF6 pressure change rate from the state variables;

[0117] Based on time-domain features, frequency-domain features, voiceprint features, electrical features, state features, and current operation deviation, a multimodal feature vector is formed by feature concatenation.

[0118] A fault category library is pre-built, including but not limited to the following categories: normal, buffer spring fatigue, contact wear, mechanism jamming, connecting rod loosening, base bolt loosening, and changes in buffer characteristics caused by SF6 leakage;

[0119] For feature groups of different modalities, RBF kernel functions are constructed independently, and the kernel width parameter is determined by cross-validation;

[0120] The SimpleMKL algorithm is used to learn the linear combination weights of each kernel, and the kernel weights are the contribution weights of the corresponding features in this diagnosis.

[0121] The combined kernel is input into the standard C-SVM classifier, which outputs the decision value for each fault category. Then, the decision value is transformed into a probability distribution by Platt scaling.

[0122] Output the probability distribution of the fault type and the contribution weight of each modal feature for this operation.

[0123] The number of pulses is defined as the number of impact events in the vibration waveform that exceed the mean ± 3 standard deviations. The energy decay rate is the time constant required for the waveform envelope trailing edge to decay from the peak value to 1 / e of it. The dominant frequency offset refers to the difference between the current vibration frequency centroid and the healthy reference dominant frequency. The harmonic energy ratio is the ratio of the second harmonic to the fundamental frequency energy. The energy proportion of the frequency band above 20kHz is the energy proportion of the ultrasonic acoustic waveform components above 20kHz. The action time is the time interval from the start of the opening and closing coil current to the change of the contact. The SF6 pressure change rate is the ratio of the pressure difference before and after the operation to the operation duration.

[0124] Reference Figure 4 As shown, the specific trends in the evolution of fault probability within the output future operation window include:

[0125] Construct a dynamic Bayesian network and set the hidden state variables to include continuous health variables and discrete remaining life interval variables;

[0126] The probability distribution sequence of fault types obtained from multiple consecutive operations is used as the observation variable and input into the network.

[0127] The key modal feature weight trend term in the contribution weight sequence obtained from multiple consecutive operations is used as the driving parameter of the state transition model.

[0128] The state transition model adopts a linear drift form with load-driven and degradation mode correction;

[0129] The observation model uses logistic regression to map hidden health status to a failure probability distribution;

[0130] Particle filtering is used for inference. At each step, the next health level is predicted based on the state transition model with correction. The likelihood is calculated as the particle weight based on the current failure probability distribution. After normalization and resampling, the weighted average of the health levels of all particles is taken as the health level estimate.

[0131] When making forward predictions, only the iterative state transition model is used to obtain the future health distribution, which is then converted into failure probabilities by the observation model, and the evolution trend of failure probabilities within the future operating window is output.

[0132] This paper utilizes a dynamic Bayesian network to predict future equipment failure trends. By using equipment health and remaining lifespan as hidden states and the probability distribution of fault types diagnosed in each operation as the observation input, the health degradation rate is dynamically adjusted using the modal contribution weights of each sensor. This allows the state transition model to adaptively reflect the current degradation mode. The observation model uses logistic regression to map health to failure probability and compares it with actual observations. Particle filtering is used to simulate health evolution: the health at the next moment is predicted based on the state transition model, and then particle weights are calculated based on the observation likelihood. After resampling, the weighted average is taken as the health estimate. When predicting forward, only the state transition model is iterated without introducing new observations, thus outputting the evolution trend of failure probability within the future operation window. This provides a basis for predictive maintenance, thereby integrating diagnostic results and weight trends to achieve adaptive health inference and early prediction of failure probability.

[0133] The process of dynamically generating an adaptive maintenance strategy based on the evolution trend of failure probability and the mutual verification of diagnostic results from multiple devices within the same interval specifically includes:

[0134] Based on the evolution trend of failure probability and the mutual verification of diagnostic results of multiple devices within the same interval, an adaptive maintenance strategy is dynamically generated, including: multi-device collaborative diagnosis and adaptive maintenance strategy generation.

[0135] Multi-device collaborative diagnosis verifies and confirms the confidence level of abnormal alarms of a single device by examining the causal and temporal relationships between related devices. When multiple devices exhibit consistent degradation characteristics, it is determined to be a common-cause fault, triggering joint maintenance for the entire interval.

[0136] The adaptive maintenance strategy generates predictive maintenance work orders when the predicted failure probability exceeds a probability threshold, and determines the dominant degradation type based on feature weight trends to determine the maintenance content.

[0137] Adjust operating parameters for specific fault types and generate joint maintenance plans for common-cause faults.

[0138] The multi-device collaborative diagnosis includes:

[0139] When there is a causal relationship between the operation time sequences of circuit breakers and disconnectors within the same bay (for example, the electrical isolation operation of the disconnector is necessarily accompanied by the circuit breaker's opening operation, and the two have a fixed sequence in time), if only a single device is diagnosed as abnormal while the associated devices are diagnosed as normal, the following collaborative verification is performed:

[0140] When a single device diagnoses an anomaly, the initial confidence level of the anomaly is taken as the maximum fault probability output by the multi-core learning support vector machine.

[0141] If the associated device diagnosis is normal, the abnormal confidence level is corrected to the product of the initial confidence level and (1 − co-attenuation coefficient) to reduce the risk of false alarms;

[0142] The vibration waveform, acoustic waveform, and electrical quantities of the device are re-acquired during the next operation. Multimodal features are extracted again and input into a multi-kernel learning support vector machine to obtain a new fault probability distribution.

[0143] If the second diagnostic result is still abnormal and the confidence level is higher than the first confidence threshold, the alarm will be maintained; otherwise, it will be marked as transient interference and no maintenance work order will be generated.

[0144] If multiple devices simultaneously exhibit an increase in vibration characteristic weights and the diagnosed fault types are consistent, it is determined to be a common cause fault at the interval level. In this case, the whole interval joint maintenance strategy is triggered, and a maintenance work order is generated that includes the maintenance plans for all related devices in the same interval. It is also recommended to handle the faults centrally within a unified power outage window.

[0145] Adaptive maintenance strategy generation includes:

[0146] When the probability of failure in the Mth operation predicted by the dynamic Bayesian network exceeds the probability threshold, a predictive maintenance work order is generated, and the original plan of maintenance according to a fixed cycle is adjusted to maintenance on demand.

[0147] When the average weight of the vibration characteristic group (including time domain peak, frequency domain centroid, harmonic energy ratio, etc.) in the contribution weight sequence shows a monotonically increasing trend (confirmed by the Mann-Kendall trend test), it is determined that the equipment degradation is dominated by mechanical wear, and the maintenance content is set to mechanical repair of the operating mechanism (such as lubrication, clearance adjustment, spring replacement).

[0148] When buffer spring fatigue is diagnosed (fault type probability exceeds probability threshold), the following mitigation measures are automatically implemented:

[0149] For on-load tap changers, reduce the subsequent tap adjustment speed;

[0150] For circuit breakers and disconnectors, reduce the frequency of high-frequency operation of the equipment or adjust the operation sequence to reduce impact load;

[0151] For common-cause faults at the interval level, a maintenance plan is generated to simultaneously inspect all related equipment within the same interval, and the maintenance work order is marked "Common-cause fault - joint inspection of the entire interval is recommended" to eliminate the risk that other equipment may still have hidden dangers after the repair of a single equipment.

[0152] The verification of maintenance effectiveness and the correction of dynamic Bayesian network parameters based on changes in health before and after maintenance specifically include:

[0153] After the maintenance operation is completed, the vibration waveform, acoustic waveform, electrical quantity, and status quantity of the equipment are collected again for the next normal operation to obtain a new health estimate.

[0154] Based on data statistics, a high health threshold, a high confidence threshold, and a minimum effective improvement amount are set;

[0155] If the health estimate is higher than the high health threshold and the normal class probability output by the multi-core learning support vector machine is higher than the high confidence threshold, then the maintenance is deemed effective and the frequency of subsequent maintenance is reduced.

[0156] If the estimated improvement in health after maintenance is less than the minimum effective improvement, the maintenance is deemed ineffective and will be upgraded from routine maintenance to in-depth maintenance, triggering a comprehensive inspection.

[0157] The deviation between the actual degradation rate and the model prediction is calculated, and the transition model parameters of the dynamic Bayesian network are corrected through Bayesian updates.

[0158] The actual degradation rate reflects how quickly the health of the equipment declines in the period before maintenance. The specific calculation method is as follows: take the estimated health values ​​of the equipment during the most recent normal operation before the maintenance operation, calculate the total change in these health values ​​over time or operation frequency, and then divide it by the corresponding operation frequency or time interval to obtain the health value decrease per unit operation frequency or unit time, which is the actual degradation rate. The model-predicted degradation rate for comparison is taken from the load driving term in the dynamic Bayesian network state transition model, that is, the expected value of health value decrease at each operation step predicted based on the current operation cut-off current amplitude and basic degradation parameters. The difference between the two is used to evaluate whether the model prediction is accurate.

[0159] When the change in health status before and after maintenance is less than the minimum effective improvement, it indicates that the model may have underestimated the actual degradation rate of the equipment or overestimated the maintenance effect. In this case, it is necessary to correct the key parameters in the state transition model. The Bayesian update method is adopted, the basic idea of ​​which is to use the observation data before and after maintenance to adjust the probability distribution of the parameters, so that the model is closer to the actual situation. The specific implementation steps are as follows:

[0160] Initial probability distributions are set for the load sensitivity coefficient and the basic degradation parameter, respectively, and it is assumed that they are independent of each other. These initial distributions are given based on historical statistical data of similar equipment or design manuals.

[0161] The sequence of health changes from the most recent operations before maintenance is taken as the observation sample, and the posterior probability distribution of the parameters under the given observation sample is calculated using Bayes' theorem.

[0162] Since the normal distribution and the normal likelihood function form a conjugate pair, the posterior distribution is still a normal distribution. Its mean and variance can be updated by a simple recursive formula: the new mean is equal to the sum of the weighted average of the old mean and the weighted average of the observed values. The weights are determined by the prior variance and the variance of the observation noise. The posterior variance decreases accordingly, indicating that the uncertainty of parameter estimation is reduced.

[0163] After each maintenance, if the observed actual degradation rate deviates from the model prediction by more than the set deviation, such an update is performed to continuously correct the parameters;

[0164] For noise variance that reflects random fluctuations, a similar method can be used to set an inverse gamma prior, and the prediction error of each operation can be used for recursive updates.

[0165] Through the Bayesian update method, the state transition model of dynamic Bayesian networks can gradually evolve with the accumulation of maintenance practices, improving the prediction accuracy of future degradation trends, thereby guiding more reasonable maintenance timing and repair strategies.

[0166] The step of triggering new fault mode mining when the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than a set threshold specifically includes:

[0167] Based on data statistics, low confidence threshold, minimum support threshold, and similarity threshold are set;

[0168] When the maximum probability of the output of a multi-core learning support vector machine for multiple consecutive operations is lower than the low confidence threshold, the new fault mode discovery process is automatically triggered. The specific steps are as follows:

[0169] Extract the vibration waveforms from these operations, divide the waveforms into several segments along the time axis, calculate the mean value of the waveform amplitude within each time window, and map the mean value onto a set alphabet to obtain a string of fixed length.

[0170] A sequence pattern mining algorithm is used to mine common subsequence patterns that occur at a frequency no less than the minimum support from multiple strings;

[0171] The discovered patterns are compared one by one with the regular expressions in the existing fault mode library to calculate the string similarity. If the similarity is lower than the similarity threshold, the pattern is pushed to the operation and maintenance experts.

[0172] After experts confirmed the new fault type through on-site inspection, they converted the symbol sequence pattern into a regular expression and added it as a new fault category to the fault category library of the multi-kernel learning support vector machine. At the same time, they generated a set of typical feature vectors for this category for kernel function calculation.

[0173] For a string set corresponding to multiple consecutive low-confidence operations, the PrefixSpan algorithm is used to mine common subsequence patterns: First, scan all strings and find all frequent letters with a length of 1 whose occurrence frequency is not less than the minimum support. Then, using each frequent letter as a prefix, recursively generate a projection database and mine frequent subsequences with a length of +1 in it. Repeat this process until no new frequent subsequences can be generated. Finally, output the set of the longest patterns that meet the conditions.

[0174] Implementation Example 2

[0175] Based on Example 1, the objective function of the SimpleMKL algorithm is: ,in, The constraints are , , , The inner layer of this objective function is about The standard C-SVM dual maximization problem has an outer layer concerning kernel weights. The minimization problem is solved by SimpleMKL through alternating optimization: fixed Solving the standard C-SVM yields the optimal solution. Then fix Update along gradient direction Iterate until convergence, and finally obtain That is, the contribution weight of each modal feature;

[0176] The parameters in the objective function of the SimpleMKL algorithm are defined as follows: This represents the total number of historical device operation events (each operation is considered a sample). Let i be the multimodal feature vector extracted in the i-th operation. The label for the actual fault category of the i-th operation. Let be the Lagrange multipliers to be solved, with non-zero values ​​corresponding to support vectors, C be the penalty parameter controlling the tradeoff between classification margin and misclassification, and D be the number of kernels. The kernel weights for the k-th modal feature are... Let be the RBF kernel function of dimension k.

[0177] Implementation Example 3

[0178] Building upon Example 1, a dynamic Bayesian network is constructed, with hidden state variables including continuous health variables. and discrete remaining life interval variables The probability distribution sequence of fault types obtained from multiple consecutive operations. As observed variable Input a dynamic Bayesian network, where G is the number of fault categories, and input the contribution weight sequence obtained from multiple consecutive operations. The key modal feature weight trend term is used as a dynamic correction parameter for the state transition model;

[0179] The state transition model adopts a load-driven linear drift form: ,in, As the amount of health deterioration, Let t be the normalized current amplitude of the cutoff operation. For load sensitivity coefficient, Based on the degradation parameter, The sum of the contribution weights of the i-th mode in the t-th operation, i.e. ,in Let i be the set of feature indices contained in the i-th mode. This is the linear trend term of the modality weights for the most recent operations (e.g., using the linear regression slope of the most recent 5 operations). Let be the acceleration factor for the degradation mode of the i-th mode. For random noise, model parameters , , , Discrete hidden states are learned from historical data using the expectation-maximization algorithm. The value is determined by the current health level range: A value >0.8 indicates a high threshold (significant number of reliable operations remaining); 0.4≤ The middle interval is when the value is ≤0.8. A value less than 0.4 indicates a low range (close to failure).

[0180] The observation model uses logistic regression to represent hidden health status. Mapped to a fault probability distribution:

[0181]

[0182] In the formula, For the index of the specific fault category whose probability is currently being calculated, The summation index is used to iterate through all fault categories in the denominator, excluding the reference class. and The values ​​are the same but traversed independently, where, and The first The gain coefficients and bias terms for the fault class are obtained offline from historical fault cases and normal operation data using the expectation-maximization algorithm. For the reference class... =G (usually the normal class), whose probability is determined by Give;

[0183] It should be noted that in this embodiment, the contribution weight sequence is used to dynamically correct the degradation rate of the state transition model. Its contribution weight reflects the importance of different sensor modes in the current diagnosis. When the equipment has mechanical wear-related faults (such as mechanism jamming or spring fatigue), the feature weights of the vibration modes will show an upward trend. On the other hand, electrical faults (such as SF6 leakage) are more reflected in changes in electrical feature weights. By extracting the trend term of the vibration mode weights, it is possible to sense in real time whether the current degradation mode is accelerating and feed it back into the health evolution equation, so that the state transition model has the ability to adapt to the degradation mode and improves the accuracy of future fault probability trend prediction.

[0184] Therefore, the likelihood function of the particle filter is given solely by the fault probability distribution:

[0185]

[0186] in, Given by the logistic regression observation model;

[0187] Inference is performed using particle filtering: The number of particles is set to 500. During initialization, the initial health values ​​of all particles are randomly generated within the range of 0.9 to 1.0 (representing the health status of new equipment or equipment just after maintenance). For each time step t, the health status of each particle is predicted according to the state transition model to obtain the predicted health status for the next time step. Based on current observations (i.e., the failure probability distribution and contribution weight sequence), calculate the importance weight of each particle. After normalizing the weights of all particles, a resampling method is used to eliminate particles with low weights and replicate particles with high weights; the final health state estimate is the weighted average of the health states of all particles. ;

[0188] When making forward predictions, no new observation data is introduced; only the state transition model is repeatedly applied, and the weights are not updated using the observation model. The trend term is taken from the most recent known value or remains unchanged, to obtain the health distribution in the Mth step in the future. Then, based on the observation model, the predicted failure probability P( Take the average or probability distribution of particles and output the evolution trend of the failure probability within the future operation window.

[0189] Furthermore, based on the same inventive concept as the aforementioned intelligent diagnosis and adaptive maintenance method for substation monitoring data, this solution proposes an intelligent diagnosis and adaptive maintenance system for substation monitoring data, comprising:

[0190] The data acquisition module is used to acquire three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical quantities and status quantities during the operation of substation equipment, preprocess the raw data, and construct a multimodal monitoring database.

[0191] The probability prediction module is used to extract the time-frequency domain stability features of vibration waveforms based on multiple consecutive operation data confirmed to be in a healthy state after initial commissioning or maintenance, establish a device health fingerprint benchmark, and adaptively update the benchmark using an exponentially weighted moving average method. For each device operation, multimodal feature vectors are extracted from multimodal monitoring data, and based on a multi-kernel learning support vector machine, the probability distribution of fault types and the contribution weights of each modal feature for this operation are output. The sequence of fault type probability distributions from multiple consecutive operations is used as the observation sequence, and the trend term of the key modal feature weights in the contribution weight sequence is used as the driving parameter of the state transition model. Both are input into a dynamic Bayesian network to output the evolution trend of fault probability within the future operation window.

[0192] The maintenance and update module is used to dynamically generate adaptive maintenance strategies based on the evolution trend of fault probability and the mutual verification of diagnostic results of multiple devices within the same interval. After the maintenance operation is performed, waveform data of the next device operation is collected again, and the hidden state estimation is updated through a multi-core learning support vector machine and a dynamic Bayesian network to verify the maintenance effect. The parameters of the dynamic Bayesian network are also corrected based on the changes in health before and after maintenance. When the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold, new fault mode mining is triggered. The vibration waveform is converted into a symbol sequence using symbol aggregation approximation, and frequent patterns are extracted through a sequence pattern mining algorithm to update the fault category library of the multi-core learning support vector machine.

[0193] The probability prediction module includes:

[0194] The stability feature unit is used to extract the time-frequency domain stability features of the vibration waveform based on multiple consecutive operation data confirmed to be in a healthy state after initial commissioning or maintenance, establish a health fingerprint benchmark for the equipment, and adaptively update the benchmark using an exponentially weighted moving average method.

[0195] The fault prediction unit is used to extract multimodal feature vectors from multimodal monitoring data for each equipment operation, and output the probability distribution of fault type and the contribution weight of each modality feature based on multi-core learning support vector machine.

[0196] The probability evolution unit is used to take the probability distribution sequence of failure types from multiple consecutive operations as the observation sequence and the key modal feature weight trend term in the contribution weight sequence as the driving parameter of the state transition model, and input them together into the dynamic Bayesian network to output the failure probability evolution trend within the future operation window.

[0197] The maintenance and update module includes:

[0198] The maintenance strategy unit is used to dynamically generate an adaptive maintenance strategy based on the failure probability evolution trend and the mutual verification of the diagnostic results of multiple devices within the same interval.

[0199] The verification and optimization unit is used to re-collect waveform data of the next equipment operation after performing maintenance operations, update the hidden state estimation through multi-core learning support vector machine and dynamic Bayesian network, verify the maintenance effect, and correct the dynamic Bayesian network parameters according to the changes in health before and after maintenance.

[0200] The fault category update unit is used to trigger new fault mode mining when the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold. The unit uses symbol aggregation approximation to convert the vibration waveform into a symbol sequence, extracts frequent patterns through a sequence pattern mining algorithm, and updates the fault category library of the multi-core learning support vector machine.

[0201] In summary, the advantages of this invention are: by deeply integrating multimodal monitoring data with dynamic Bayesian network prediction, it achieves adaptive assessment of equipment health status, prediction of degradation trends, and optimization of closed-loop maintenance strategies, significantly improving the accuracy of fault diagnosis and the level of intelligent maintenance of key equipment in substations.

[0202] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for intelligent diagnosis and adaptive maintenance based on substation monitoring data, characterized in that, include: Collect three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical quantities and status quantities during the operation of substation equipment. Preprocess the raw data to construct a multimodal monitoring database. Based on continuous operation data confirmed to be in a healthy state after initial commissioning or maintenance, the time-frequency domain stability characteristics of the vibration waveform are extracted to establish a health fingerprint benchmark for the equipment, and the benchmark is adaptively updated using the exponential weighted moving average method. For each equipment operation, multimodal feature vectors are extracted from multimodal monitoring data. Based on multi-core learning support vector machine, the probability distribution of fault type and the contribution weight of each modality feature for this operation are output. The probability distribution sequence of failure types from multiple consecutive operations is used as the observation sequence, and the trend term of the key modal feature weight in the contribution weight sequence is used as the driving parameter of the state transition model. Both are input into the dynamic Bayesian network to output the evolution trend of failure probability within the future operation window. Based on the evolution trend of failure probability and the mutual verification of diagnostic results of multiple devices within the same interval, an adaptive maintenance strategy is dynamically generated. After performing maintenance operations, waveform data for the next equipment operation is collected again. The hidden state estimation is updated by multi-core learning support vector machine and dynamic Bayesian network to verify the maintenance effect. The parameters of dynamic Bayesian network are then corrected based on the changes in health status before and after maintenance. When the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold, new fault mode mining is triggered. The vibration waveform is converted into a symbol sequence by symbol aggregation approximation, and frequent patterns are extracted by sequence pattern mining algorithm to update the fault category library of the multi-core learning support vector machine.

2. The intelligent diagnosis and adaptive maintenance method for substation monitoring data according to claim 1, characterized in that, The preprocessing of the raw data collected during the operation of substation equipment, including three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical and status quantities, specifically includes: Install a triaxial MEMS accelerometer and an ultrasonic microphone in the mechanism box of a circuit breaker, disconnector, or on-load tap changer. The acquisition is triggered by the equipment operation command to record waveform data from before to after the operation. Simultaneously acquire the current waveform of the opening and closing coils, SF6 gas density / pressure, cumulative number of operations, and the amplitude of the disconnection current; All collected data are interpolated and downsampled to align with a unified timestamp, and each physical quantity is normalized to the [0,1] interval to eliminate the influence of dimensions. After alignment and normalization, outlier detection based on the 3σ criterion is performed to remove impact interference spikes in the vibration waveform that exceed the mean ± 3 times the standard deviation, and the removed points are replaced with the mean of the preceding and following neighborhoods. The wavelet soft thresholding denoising method is adopted, and the Sym8 wavelet basis is selected to perform a 4-level decomposition of the vibration signal to obtain the low-frequency approximation coefficients and the high-frequency detail coefficients of each level. The low-frequency approximation coefficients are retained, and a threshold is set for the high-frequency detail coefficients. The high-frequency detail coefficients that exceed the threshold are retained, and the denoised vibration waveform is reconstructed. The ambient background noise is suppressed by spectral subtraction on the acoustic waveform, that is, the noise power spectrum when there is no operation is estimated, and the time-domain acoustic waveform is reconstructed after subtracting the noise spectrum from the power spectrum of the operation waveform. A multimodal monitoring database is constructed, and the pre-processed vibration waveform, acoustic waveform, and aligned electrical and state quantities are stored together in the multimodal monitoring database.

3. The intelligent diagnosis and adaptive maintenance method for substation monitoring data according to claim 2, characterized in that, The extraction of the time-frequency domain stability characteristics of the vibration waveform, the establishment of a device health fingerprint benchmark, and the adaptive updating of the benchmark using an exponentially weighted moving average method specifically include: After the equipment is put into operation for the first time or its health status is confirmed during maintenance, the vibration waveforms of the previous several normal operations are continuously collected. For each waveform, the stable features of the time domain peak, frequency domain centroid, and waveform entropy are extracted, and the median of each feature is taken as the baseline vector of the device health fingerprint. For each subsequent operation, the Euclidean distance between the extracted feature vector and the reference vector is calculated as the deviation. The baseline vector is updated using an exponentially weighted moving average, and the update is performed only when the deviation of the current operation is less than a distance threshold.

4. The intelligent diagnosis and adaptive maintenance method for substation monitoring data according to claim 3, characterized in that, The multi-kernel learning support vector machine-based output of the fault type probability distribution and the contribution weights of each modal feature for this operation specifically includes: For each device operation, time-domain features are extracted from the denoised vibration waveform. These time-domain features include: peak value, peak-to-peak value, number of pulses, waveform entropy, and energy decay rate. The same vibration waveform is subjected to a fast Fourier transform to extract frequency domain features, which include: frequency centroid, dominant frequency shift, and harmonic energy ratio. The energy proportion, zero-crossing rate, and Mel frequency cepstral coefficients of the frequency band above 20kHz are extracted from the denoised audio waveform. Extract electrical characteristics such as peak current of the opening and closing coils, operating time, and harmonic distortion rate from electrical quantities; Extract state characteristics from the cumulative number of operations and the SF6 pressure change rate from the state variables; Based on time-domain features, frequency-domain features, voiceprint features, electrical features, state features, and current operation deviation, a multimodal feature vector is formed by feature concatenation. A fault category library is pre-built, including but not limited to the following categories: normal, buffer spring fatigue, contact wear, mechanism jamming, connecting rod loosening, base bolt loosening, and changes in buffer characteristics caused by SF6 leakage; For feature groups of different modalities, RBF kernel functions are constructed independently, and the kernel width parameter is determined by cross-validation; The SimpleMKL algorithm is used to learn the linear combination weights of each kernel, and the kernel weights are the contribution weights of the corresponding features in this diagnosis. The combined kernel is input into the standard C-SVM classifier, which outputs the decision value for each fault category. Then, the decision value is transformed into a probability distribution by Platt scaling. Output the probability distribution of the fault type and the contribution weight of each modal feature for this operation.

5. The intelligent diagnosis and adaptive maintenance method for substation monitoring data according to claim 4, characterized in that, The step of using the probability distribution sequence of failure types from multiple consecutive operations as the observation sequence, and the key modal feature weight trend term from the contribution weight sequence as the driving parameter of the state transition model, and inputting them together into a dynamic Bayesian network to output the failure probability evolution trend within the future operation window specifically includes: Construct a dynamic Bayesian network and set the hidden state variables to include continuous health variables and discrete remaining life interval variables; The probability distribution sequence of fault types obtained from multiple consecutive operations is used as the observation variable and input into the network. The key modal feature weight trend term in the contribution weight sequence obtained from multiple consecutive operations is used as the driving parameter of the state transition model. The state transition model adopts a linear drift form with load-driven and degradation mode correction; The observation model uses logistic regression to map hidden health status to a failure probability distribution; Particle filtering is used for inference. At each step, the next health level is predicted based on the state transition model with correction. The likelihood is calculated as the particle weight based on the current failure probability distribution. After normalization and resampling, the weighted average of the health levels of all particles is taken as the health level estimate. When making forward predictions, only the iterative state transition model is used to obtain the future health distribution, which is then converted into failure probabilities by the observation model, and the evolution trend of failure probabilities within the future operating window is output.

6. The intelligent diagnosis and adaptive maintenance method for substation monitoring data according to claim 5, characterized in that, The process of dynamically generating an adaptive maintenance strategy based on the evolution trend of failure probability and the mutual verification of diagnostic results from multiple devices within the same interval specifically includes: Based on the evolution trend of failure probability and the mutual verification of diagnostic results of multiple devices within the same interval, an adaptive maintenance strategy is dynamically generated, including: multi-device collaborative diagnosis and adaptive maintenance strategy generation. Multi-device collaborative diagnosis verifies and confirms the confidence level of abnormal alarms of a single device by examining the causal and temporal relationships between related devices. When multiple devices exhibit consistent degradation characteristics, it is determined to be a common-cause fault, triggering joint maintenance for the entire interval. The adaptive maintenance strategy generates predictive maintenance work orders when the predicted failure probability exceeds a probability threshold, and determines the dominant degradation type based on feature weight trends to determine the maintenance content. Adjust operating parameters for specific fault types and generate joint maintenance plans for common-cause faults.

7. The intelligent diagnosis and adaptive maintenance method for substation monitoring data according to claim 6, characterized in that, The verification of maintenance effectiveness and the correction of dynamic Bayesian network parameters based on changes in health before and after maintenance specifically include: After the maintenance operation is completed, the vibration waveform, acoustic waveform, electrical quantity, and status quantity of the equipment are collected again for the next normal operation to obtain a new health estimate. Based on data statistics, a high health threshold, a high confidence threshold, and a minimum effective improvement amount are set; If the health estimate is higher than the high health threshold and the normal class probability output by the multi-core learning support vector machine is higher than the high confidence threshold, then the maintenance is deemed effective and the frequency of subsequent maintenance is reduced. If the estimated improvement in health after maintenance is less than the minimum effective improvement, the maintenance is deemed ineffective and will be upgraded from routine maintenance to in-depth maintenance, triggering a comprehensive inspection. The deviation between the actual degradation rate and the model prediction is calculated, and the transition model parameters of the dynamic Bayesian network are corrected through Bayesian updates.

8. The intelligent diagnosis and adaptive maintenance method for substation monitoring data according to claim 7, characterized in that, The step of triggering new fault mode mining when the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than a set threshold specifically includes: Based on data statistics, low confidence threshold, minimum support threshold, and similarity threshold are set; When the maximum probability of the output of a multi-core learning support vector machine for multiple consecutive operations is lower than the low confidence threshold, the new fault mode discovery process is automatically triggered. The specific steps are as follows: Extract the vibration waveforms from these operations, divide the waveforms into several segments along the time axis, calculate the mean value of the waveform amplitude within each time window, and map the mean value onto a set alphabet to obtain a string of fixed length. A sequence pattern mining algorithm is used to mine common subsequence patterns that occur at a frequency no less than the minimum support from multiple strings; The discovered patterns are compared one by one with the regular expressions in the existing fault mode library to calculate the string similarity. If the similarity is lower than the similarity threshold, the pattern is pushed to the operation and maintenance experts. After experts confirmed the new fault type through on-site inspection, they converted the symbol sequence pattern into a regular expression and added it as a new fault category to the fault category library of the multi-kernel learning support vector machine. At the same time, they generated a set of typical feature vectors for this category for kernel function calculation.

9. An intelligent diagnostic and adaptive maintenance system for substation monitoring data, characterized in that, The method for intelligent diagnosis and adaptive maintenance of substation monitoring data as described in any one of claims 1-8 includes: The data acquisition module is used to acquire three-dimensional broadband vibration waveforms, ultrasonic acoustic waveforms, and conventional electrical quantities and status quantities during the operation of substation equipment, preprocess the raw data, and construct a multimodal monitoring database. The probability prediction module is used to extract the time-frequency domain stability features of vibration waveforms based on multiple consecutive operation data confirmed to be in a healthy state after initial commissioning or maintenance, establish a device health fingerprint benchmark, and adaptively update the benchmark using an exponentially weighted moving average method. For each device operation, multimodal feature vectors are extracted from multimodal monitoring data, and based on a multi-kernel learning support vector machine, the probability distribution of fault types and the contribution weights of each modal feature for this operation are output. The sequence of fault type probability distributions from multiple consecutive operations is used as the observation sequence, and the trend term of the key modal feature weights in the contribution weight sequence is used as the driving parameter of the state transition model. Both are input into a dynamic Bayesian network to output the evolution trend of fault probability within the future operation window. The maintenance and update module is used to dynamically generate adaptive maintenance strategies based on the evolution trend of fault probability and the mutual verification of diagnostic results of multiple devices within the same interval. After the maintenance operation is performed, waveform data of the next device operation is collected again, and the hidden state estimation is updated through a multi-core learning support vector machine and a dynamic Bayesian network to verify the maintenance effect. The parameters of the dynamic Bayesian network are also corrected based on the changes in health before and after maintenance. When the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold, a new fault mode mining is triggered. The vibration waveform is converted into a symbol sequence using symbol aggregation approximation, and frequent patterns are extracted through a sequence pattern mining algorithm to update the fault category library of the multi-core learning support vector machine.

10. The intelligent diagnostic and adaptive maintenance system for substation monitoring data according to claim 9, characterized in that, The probability prediction module includes: The stability feature unit is used to extract the time-frequency domain stability features of the vibration waveform based on multiple consecutive operation data confirmed to be in a healthy state after initial commissioning or maintenance, establish a health fingerprint benchmark for the equipment, and adaptively update the benchmark using an exponentially weighted moving average method. The fault prediction unit is used to extract multimodal feature vectors from multimodal monitoring data for each equipment operation, and output the probability distribution of fault type and the contribution weight of each modality feature based on multi-core learning support vector machine. The probability evolution unit is used to take the probability distribution sequence of failure types from multiple consecutive operations as the observation sequence and the key modal feature weight trend term in the contribution weight sequence as the driving parameter of the state transition model, and input them together into the dynamic Bayesian network to output the failure probability evolution trend within the future operation window. The maintenance and update module includes: The maintenance strategy unit is used to dynamically generate an adaptive maintenance strategy based on the failure probability evolution trend and the mutual verification of the diagnostic results of multiple devices within the same interval. The verification and optimization unit is used to re-collect waveform data of the next equipment operation after performing maintenance operations, update the hidden state estimation through multi-core learning support vector machine and dynamic Bayesian network, verify the maintenance effect, and correct the dynamic Bayesian network parameters according to the changes in health before and after maintenance. The fault category update unit is used to trigger new fault mode mining when the output confidence of the multi-core learning support vector machine for multiple consecutive operations is lower than the threshold. The unit uses symbol aggregation approximation to convert the vibration waveform into a symbol sequence, extracts frequent patterns through a sequence pattern mining algorithm, and updates the fault category library of the multi-core learning support vector machine.

Images