A transformer voiceprint monitoring method based on fault prediction and health management

Abstract

This invention discloses a transformer acoustic signature monitoring method based on fault prediction and health management, comprising the following steps: acquiring multi-channel acoustic signature signals and operating condition parameters, and performing data preprocessing to form an acoustic signature feature sequence and an operating condition parameter sequence; reconstructing the operating condition parameter sequence into a time-delay input block; constructing an acoustic signature feature output block, and simultaneously performing N4SID numerical subspace state space identification to obtain the operating condition-acoustic signature state space representation relationship; determining the health observation subspace, and generating a dynamic acoustic signature baseline based on the acoustic signature baseline components; performing fault-sensitive acoustic signature direction decomposition to obtain fault-sensitive degradation components; identifying the current observation subspace and constructing a transformer health index; correcting the transformer health index, updating the operating condition-acoustic signature state space representation relationship, and generating transformer acoustic signature monitoring results. This invention improves the fault prediction capability and health status assessment reliability of transformers under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a transformer acoustic signature monitoring method based on fault prediction and health management. Background Technology

[0002] Transformers are crucial equipment in power systems, responsible for voltage transformation and energy transmission. Their operational status directly impacts power grid security and equipment maintenance costs. During long-term operation, transformers are affected by factors such as load fluctuations, temperature changes, winding stress, core vibration, cooling system activation and deactivation, and external environmental noise. This can lead to potential problems such as core loosening, winding deformation, partial discharge, cooling system malfunctions, and loose fasteners. These problems often exhibit changes in acoustic signature characteristics in their early stages. Therefore, online monitoring of transformers using acoustic signature signals offers advantages such as non-contact operation, convenient deployment, and timely response.

[0003] Existing transformer acoustic signature monitoring methods typically collect the operating sounds of the transformer using microphones, acoustic sensors, or vibration sensors, and then filter, analyze the spectrum, extract features, and identify anomalies in the collected acoustic signature signals. Some methods employ fixed thresholds, template matching, traditional machine learning classifiers, or deep learning models to judge acoustic signature features and thus identify whether there are abnormalities in the transformer. However, the operating sound of a transformer is not static; its acoustic signature characteristics change with load rate, current harmonics, oil temperature, winding temperature, the start / stop status of the cooling device, and the intensity of ambient noise. If only a fixed acoustic signature template or a uniform threshold is used, it is easy to misjudge changes in normal operating conditions as faults or anomalies, and it is also easy to miss early degradation of acoustic signatures under complex noise conditions.

[0004] Therefore, how to provide a transformer acoustic signature monitoring method based on fault prediction and health management is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a transformer acoustic signature monitoring method based on fault prediction and health management. This invention combines acoustic signature features, operating conditions, and N4SID numerical state space identification to construct a dynamic acoustic signature baseline and extract fault-sensitive degradation components. This reduces the interference of operating condition fluctuations on acoustic signature judgment, improves the accuracy of early transformer anomaly identification, fault trend prediction, and health index assessment, and provides a reliable basis for predictive maintenance and intelligent operation and maintenance.

[0006] A transformer acoustic signature monitoring method based on fault prediction and health management according to an embodiment of the present invention includes the following steps: Multi-channel acoustic signature signals and operating parameters are collected during transformer operation, and data preprocessing is performed to form acoustic signature feature sequences and operating parameter sequences. Based on the time delay difference in the response of the operating condition parameter sequence to the voiceprint feature sequence, the operating condition parameter sequence is reconstructed into a time delay input block; Construct the voiceprint feature output block and combine it with the time-delay input block to construct the phase-delay block Hankel matrix of the operating condition modulation. At the same time, perform N4SID numerical subspace state space identification to obtain the operating condition-voiceprint state space representation relationship. The health observation subspace is determined based on the relationship between operating conditions and voiceprint state space. The real-time voiceprint features are decomposed into voiceprint baseline components that can be explained by the current operating conditions and unexplainable health degradation components. A dynamic voiceprint baseline is generated based on the voiceprint baseline components. Based on the dynamic voiceprint baseline, the health degradation component is decomposed into fault-sensitive voiceprint direction to obtain the fault-sensitive degradation component. The current observation subspace is identified based on the working condition-acoustic print state space representation relationship, the structural drift of the current observation subspace is calculated, and a transformer health index is constructed. The transformer health index is corrected based on the fault-sensitive degradation component and structural drift. The operating condition-acoustic signature state space representation relationship is updated based on the corrected transformer health index, and the transformer acoustic signature monitoring results are generated.

[0007] Optionally, obtaining the voiceprint feature sequence and the operating condition parameter sequence specifically includes: Multi-channel acoustic fingerprint signals and operating condition parameters during transformer operation are collected according to the same sampling time reference. The multi-channel acoustic fingerprint signals are organized according to the sampling time and sampling point order, and the operating condition parameters are organized according to the parameter category and parameter value order to obtain the original acoustic fingerprint data and original operating condition data. The original voiceprint data is preprocessed to obtain preprocessed voiceprint data, and the operating condition parameter data corresponding to the preprocessed voiceprint data is extracted from the original operating condition data. The voiceprint data is continuously segmented according to the set sliding window length and sliding step size to obtain voiceprint data segments. Voiceprint feature extraction is performed on each voiceprint data segment to obtain a voiceprint feature sequence. Then, the operating condition parameter data of each voiceprint data segment is aggregated to obtain the corresponding operating condition parameter sequence.

[0008] Optionally, obtaining the time-delay input block specifically includes: Read the voiceprint feature sequence and the operating condition parameter sequence, calculate the voiceprint change amplitude between adjacent time windows in the voiceprint feature sequence, and determine the time window that meets the preset change condition as the voiceprint change anchor point; For each type of operating condition parameter in the operating condition parameter sequence, based on the voiceprint change anchor point, the operating condition parameter segments before the current voiceprint change anchor point, at the current corresponding time, and after the current voiceprint change anchor point are extracted respectively. Based on the start position, duration position, and fall position of parameter change in the operating condition parameter segment, the influence phase record of the operating condition parameter relative to the voiceprint change anchor point is generated. Based on the influence phase records formed by the same type of operating parameters under different acoustic text change anchor points, the stable influence phase is determined, and the operating parameters are divided into leading response operating parameters, synchronous response operating parameters, lagging response operating parameters, or maintenance response operating parameters according to the stable influence phase. According to the stable influence phase, configure phase attribute identifiers for various operating condition parameters, and perform phase alignment reconstruction on the sequence of various operating condition parameters based on the phase attribute identifiers. The lead response operating condition quantity is entered into the input position before the acoustic text change anchor point, the synchronous response operating condition quantity is entered into the input position corresponding to the acoustic text change anchor point, the hysteresis response operating condition quantity is entered into the input position after the acoustic text change anchor point, and the maintenance response operating condition quantity is entered into the continuous input position covering the acoustic text change anchor point, thus obtaining the phase-separated operating condition sub-blocks carrying phase attributes. According to the time sequence of the voiceprint change anchor points, the phase-separated sub-blocks carrying phase attributes are spliced ​​together, and the phase order of the lead response, synchronization response, hysteresis response and maintenance response corresponding to each phase-separated sub-block is maintained to obtain the time-delay input block.

[0009] Optionally, obtaining the working condition-voiceprint state space representation relationship specifically includes: Based on the phase attributes carried by the time-delay input block, the operating condition parameter segments associated with the same acoustic text change anchor point are compressed into phase-separated operating condition input blocks, and past acoustic text output blocks and future acoustic text output blocks are constructed from the acoustic text feature sequence. Based on the phase position of each operating condition parameter segment in the phase-separated operating condition input block relative to the acoustic text change anchor point, the phase contribution intensity of the phase-separated operating condition input block to the past acoustic text output block is calculated, and the phase-separated operating condition input block is rearranged and weighted according to the phase contribution intensity to obtain the phase-constrained operating condition input block. The phase-constrained operating condition input block, past voiceprint output block and future voiceprint output block are arranged in layers according to the block Hankel arrangement rule to form the phase-delay block Hankel matrix of the operating condition modulation. The N4SID numerical subspace state space identification is performed on the phase-delay block Hankel matrix of the modulated operating condition. The phase-constrained operating condition input block and the past acoustic print output block are combined to form a phase-constrained historical data space. The future acoustic print output block is orthogonally projected onto the phase-constrained historical data space to obtain the future acoustic print projection data constrained by the operating condition phase. Singular value decomposition is performed on the future acoustic signature projection data constrained by the working condition phase, and the number of acoustic signature change anchor points is counted within a set identification window. The ratio between the number of acoustic signature change anchor points and the set identification window length is determined as the acoustic signature change anchor point density. The working condition-acoustic signature hidden state order is determined by the singular value energy contribution, phase contribution intensity and acoustic signature change anchor point density. Extended observation information and working condition-acoustic signature hidden state sequence are extracted based on the working condition-acoustic signature hidden state order. Based on the degree to which the phase-constrained working condition input block interprets the working condition-soundprint hidden state sequence, the working condition-soundprint hidden state sequence is divided into the baseline hidden state sequence and the degenerate hidden state sequence. The baseline recursive relationship of the acoustic print driven by the baseline hidden state sequence is determined, the acoustic print degradation recursive relationship is determined by the degradation hidden state sequence, the working condition input action relationship is determined by the phase-constrained working condition input block, and the acoustic print observation relationship is determined by the acoustic print feature output block. The acoustic print baseline recursive relationship, the acoustic print degradation recursive relationship, the working condition input action relationship and the acoustic print observation relationship are combined to form the working condition-acoustic print state space representation relationship.

[0010] Optionally, obtaining the dynamic voiceprint baseline specifically includes: Extract the working condition input relationship, the voiceprint baseline recursive relationship, the voiceprint degradation recursive relationship, and the voiceprint observation relationship from the working condition-voiceprint state space representation relationship; Substitute the sequence of operating parameters corresponding to the historical healthy operating status into the operating condition input action relationship and the voiceprint baseline recursion relationship to obtain the baseline hidden state sequence corresponding to the historical healthy operating status. Substitute the baseline hidden state sequence into the voiceprint observation relationship to obtain the corresponding healthy voiceprint observation sequence, and determine the healthy observation subspace from the healthy voiceprint observation sequence. Substitute the current operating condition parameters into the operating condition input action relationship and the voiceprint baseline recursion relationship to obtain the candidate baseline hidden state corresponding to the current operating condition. Substitute the candidate baseline hidden state into the health observation subspace for interpretation qualification determination, retain the hidden state part that satisfies the health observation subspace constraints, and obtain the health constraint baseline hidden state. Substituting the hidden state of the health constraint baseline into the voiceprint observation relationship, we obtain the voiceprint baseline component that can be explained by the current operating conditions. The part of the real-time voiceprint feature that can be explained by the voiceprint baseline component is marked as the operating condition-explainable voiceprint part, and the part of the real-time voiceprint feature that cannot be explained by the voiceprint baseline component is marked as the operating condition-unexplainable voiceprint part. Substitute the unexplainable voiceprint component under operating conditions into the voiceprint degradation recursive relation to obtain the health degradation component. Then, store the health degradation component in isolation from the voiceprint baseline component so that the health degradation component does not participate in the generation of the dynamic voiceprint baseline. Arrange the voiceprint baseline components in the order of the time window to generate the dynamic voiceprint baseline under the current operating conditions.

[0011] Optionally, obtaining the fault-sensitive degradation component specifically includes: Read the dynamic voiceprint baseline and health degradation component, match the dynamic voiceprint baseline and health degradation component within the same time window according to the time window order, and align each voiceprint baseline feature in the dynamic voiceprint baseline with each degradation feature in the health degradation component according to the voiceprint feature dimension order to obtain baseline degradation alignment data. Based on the baseline degradation aligned data, the health degradation component within each time window is subjected to baseline normalization processing to obtain baseline normalized degradation data. Read historical fault voiceprint samples, perform data preprocessing and voiceprint feature extraction on the historical fault voiceprint samples, and divide the historical fault voiceprint samples into different fault sample groups according to the corresponding confirmed fault category. The degradation features corresponding to each historical fault voiceprint sample in the fault sample group are aggregated according to the voiceprint feature dimension, and the direction of the aggregated degradation features is determined as the fault-sensitive voiceprint direction corresponding to the current confirmed fault category. The fault-sensitive voiceprint directions are arranged in the order of the confirmed fault categories to form a fault-sensitive voiceprint direction set. The baseline normalized degradation data is matched with each fault-sensitive acoustic fingerprint direction in the fault-sensitive acoustic fingerprint direction set to obtain the matching results of the health degradation component in different fault-sensitive acoustic fingerprint directions within each time window. Based on the matching results, the baseline normalized degradation data is assigned to the corresponding fault-sensitive acoustic fingerprint direction to obtain fault-sensitive degradation components corresponding to different fault-sensitive acoustic fingerprint directions, and the fault-sensitive degradation components are arranged in order of time window.

[0012] Optionally, the construction of the transformer health index specifically includes: Read the working condition-voiceprint state space representation relationship, health observation subspace, dynamic voiceprint baseline and fault-sensitive degradation component sequence, and extract the real-time voiceprint features, dynamic voiceprint baseline and fault-sensitive degradation components in the current identification window according to the time window order to form the voiceprint analysis data of the current window; Based on the working condition-voiceprint state space representation relationship, the real-time voiceprint features and corresponding working condition parameters in the current identification window are reconstructed to obtain the current window working condition-voiceprint hidden state sequence, and the current observation subspace is generated from the current window working condition-voiceprint hidden state sequence. Align the current observation subspace with the healthy observation subspace, calculate the principal angle of the current observation subspace relative to the healthy observation subspace, and use the weighted result of all principal angles as the structural drift. The baseline deviation is calculated based on the dynamic voiceprint baseline and real-time voiceprint features. The baseline deviation is the degree of deviation of the real-time voiceprint features from the dynamic voiceprint baseline within the same time window. The fault-sensitive degradation intensity is calculated based on the fault-sensitive degradation component sequence. A transformer health index is constructed based on structural drift, baseline deviation, and fault-sensitive degradation intensity.

[0013] Optionally, the generation of the transformer acoustic signature monitoring results specifically includes: The transformer health index, structural drift, and fault-sensitive degradation components are aligned one by one according to the same time window to obtain the data to be corrected. The increase of fault-sensitive degradation components, the increase of structural drift, and the decrease of transformer health index in the data to be corrected are checked sequentially along the time window. The health index deduction intensity is increased within the time window when all three conditions are met simultaneously, and the health index deduction intensity caused by structural drift is reduced within the time window when the three conditions are not met simultaneously, so as to obtain the corrected transformer health index. The corrected transformer health index is arranged in order of time window, the numerical changes between adjacent time windows are compared, and the health trend judgment result is formed based on the number of consecutively decreasing windows and the magnitude of the decrease. The working condition-soundprint state space representation relationship is updated based on the corrected transformer health index. When the corrected transformer health index meets the baseline update condition, the soundprint baseline change in the corresponding time window is written into the soundprint baseline recursive relationship. When the corrected transformer health index does not meet the baseline update condition, the fault-sensitive degradation component in the corresponding time window is written into the soundprint degradation recursive relationship. Based on the updated working condition-soundprint state space representation relationship, the corrected transformer health index, and the health trend judgment results, transformer soundprint monitoring results are generated.

[0014] The beneficial effects of this invention are: This invention reconstructs the operating condition parameter sequence into a time-delay input block based on the time delay difference in the response of the operating condition parameter sequence to the acoustic signature feature sequence. This is further constructed into a phase-delayed time-delay block Hankel matrix modulated by the operating condition, and then N4SID numerical subspace state-space identification is performed to obtain the operating condition-acoustic signature state-space representation relationship. This allows for a more accurate description of the impact of different operating conditions on transformer acoustic signature changes, reducing false alarms and missed alarms caused by changes in operating conditions, and improving the adaptability of the dynamic acoustic signature baseline.

[0015] This invention is based on the working condition-soundprint state space representation relationship. It decomposes real-time soundprint features into soundprint baseline components that can be explained by the current operating condition and unexplainable health degradation components. It generates a dynamic soundprint baseline based on the soundprint baseline components, which can distinguish between normal operating condition changes and suspected fault degradation changes. This avoids early fault soundprints being mistakenly absorbed into the normal baseline and improves the ability to identify early faults such as loose core, abnormal winding, partial discharge, and abnormal cooling device.

[0016] This invention decomposes the health degradation component into fault-sensitive acoustic direction based on the dynamic acoustic baseline to obtain the fault-sensitive degradation component. It also constructs and corrects the transformer health index by combining the structural drift of the current observation subspace. This not only determines whether the current acoustic is abnormal, but also reflects the degree of change in the transformer operating condition-acoustic coupling relationship, thereby realizing a quantitative assessment of the equipment health status and a prediction of the fault development trend.

[0017] This invention updates the working condition-acoustic fingerprint state space representation relationship based on the corrected transformer health index and generates transformer acoustic fingerprint monitoring results. This enables the monitoring process to form a closed-loop processing flow from acoustic fingerprint acquisition, dynamic baseline generation, degradation component extraction, health index assessment to monitoring result output, thereby improving the transformer fault prediction capability, health status assessment reliability and predictive operation and maintenance decision-making level. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a transformer acoustic signature monitoring method based on fault prediction and health management proposed in this invention; Figure 2 This is a schematic diagram of N4SID numerical state space identification driven by the acoustic signature change anchor point in a transformer acoustic signature monitoring method based on fault prediction and health management proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A transformer acoustic signature monitoring method based on fault prediction and health management includes the following steps: Multi-channel acoustic signature signals and operating parameters are collected during transformer operation, and data preprocessing is performed to form acoustic signature feature sequences and operating parameter sequences. Based on the time delay difference in the response of the operating condition parameter sequence to the voiceprint feature sequence, the operating condition parameter sequence is reconstructed into a time delay input block; Construct the voiceprint feature output block and combine it with the time-delay input block to construct the phase-delay block Hankel matrix of the operating condition modulation. At the same time, perform N4SID numerical subspace state space identification to obtain the operating condition-voiceprint state space representation relationship. The health observation subspace is determined based on the relationship between operating conditions and voiceprint state space. The real-time voiceprint features are decomposed into voiceprint baseline components that can be explained by the current operating conditions and unexplainable health degradation components. A dynamic voiceprint baseline is generated based on the voiceprint baseline components. Based on the dynamic voiceprint baseline, the health degradation component is decomposed into fault-sensitive voiceprint direction to obtain the fault-sensitive degradation component. The current observation subspace is identified based on the working condition-acoustic print state space representation relationship, the structural drift of the current observation subspace is calculated, and a transformer health index is constructed. The transformer health index is corrected based on the fault-sensitive degradation component and structural drift. The operating condition-acoustic signature state space representation relationship is updated based on the corrected transformer health index, and the transformer acoustic signature monitoring results are generated.

[0021] In this embodiment, obtaining the voiceprint feature sequence and the operating condition parameter sequence specifically includes: Multi-channel acoustic fingerprint signals and operating condition parameters during transformer operation are collected according to the same sampling time reference. The multi-channel acoustic fingerprint signals are organized according to the sampling time and sampling point order, and the operating condition parameters are organized according to parameter category and parameter value order to obtain the original acoustic fingerprint data and original operating condition data. The multi-channel acoustic fingerprint signals include sound signals collected by acoustic fingerprint acquisition channels arranged on the transformer body side, bushing side, radiator side and cooling device side. The operating condition parameters include load rate, current, voltage, current harmonics, oil temperature, winding temperature, cooling device start / stop status and ambient noise intensity. The original voiceprint data is preprocessed, including channel synchronization, abnormal sampling point removal, amplitude normalization, filtering and noise reduction, and effective voiceprint segment extraction, to obtain preprocessed voiceprint data. The operating condition parameter data corresponding to the preprocessed voiceprint data is extracted from the original operating condition data. The voiceprint data is continuously segmented according to the set sliding window length and sliding step size to obtain voiceprint data segments. Voiceprint feature extraction is performed on each voiceprint data segment to obtain a voiceprint feature sequence. Then, the operating condition parameter data of each voiceprint data segment is aggregated to obtain the corresponding operating condition parameter sequence.

[0022] In this embodiment, obtaining the time-delay input block specifically includes: Read the voiceprint feature sequence and the operating condition parameter sequence, calculate the voiceprint change amplitude between adjacent time windows in the voiceprint feature sequence, and determine the time window that meets the preset change condition as the voiceprint change anchor point; For each type of operating condition parameter in the operating condition parameter sequence, based on the voiceprint change anchor point, the operating condition parameter segments before the current voiceprint change anchor point, at the current corresponding time, and after the current voiceprint change anchor point are extracted respectively. Based on the start position, duration position, and fall position of parameter change in the operating condition parameter segment, the influence phase record of the operating condition parameter relative to the voiceprint change anchor point is generated. Based on the influence phase records formed by the same type of operating parameters under different acoustic text change anchor points, the stable influence phase is determined, and the operating parameters are divided into leading response operating parameters, synchronous response operating parameters, lagging response operating parameters, or maintenance response operating parameters according to the stable influence phase. According to the stable influence phase, configure phase attribute identifiers for various operating condition parameters, and perform phase alignment reconstruction on the sequence of various operating condition parameters based on the phase attribute identifiers. The lead response operating condition quantity is entered into the input position before the acoustic text change anchor point, the synchronous response operating condition quantity is entered into the input position corresponding to the acoustic text change anchor point, the hysteresis response operating condition quantity is entered into the input position after the acoustic text change anchor point, and the maintenance response operating condition quantity is entered into the continuous input position covering the acoustic text change anchor point, thus obtaining the phase-separated operating condition sub-blocks carrying phase attributes. According to the time sequence of the voiceprint change anchor points, the phase-separated sub-blocks carrying phase attributes are spliced ​​together, and the phase order of the lead response, synchronization response, hysteresis response and maintenance response corresponding to each phase-separated sub-block is maintained to obtain the time-delay input block.

[0023] This invention analyzes the starting, duration, and fall positions of various operating parameters based on the anchor point of voiceprint changes, determines the stable influence phase of the operating parameters relative to voiceprint changes, and constructs phase-separated operating condition sub-blocks and time-delay input blocks carrying phase attributes. This allows the leading, synchronizing, lagging, and sustaining effects of different operating factors on voiceprint changes to be accurately expressed, thereby improving the construction accuracy of the phase-separated time-delay block Hankel matrix of operating condition modulation and enhancing the adaptability of operating condition-voiceprint state space identification to complex operating conditions.

[0024] In this embodiment, obtaining the working condition-voiceprint state space representation relationship specifically includes: Based on the phase attributes carried by the time-delay input block, the operating condition parameter segments associated with the same acoustic text change anchor point are compressed into phase-separated operating condition input blocks, and past acoustic text output blocks and future acoustic text output blocks are constructed from the acoustic text feature sequence. Based on the phase position of each operating condition parameter segment in the phase-separated operating condition input block relative to the acoustic text change anchor point, the phase contribution intensity of the phase-separated operating condition input block to the past acoustic text output block is calculated, and the phase-separated operating condition input block is rearranged and weighted according to the phase contribution intensity to obtain the phase-constrained operating condition input block. The phase-constrained operating condition input block, past voiceprint output block and future voiceprint output block are arranged in layers according to the block Hankel arrangement rule to form the phase-delay block Hankel matrix of the operating condition modulation. The N4SID numerical subspace state space identification is performed on the phase-delay block Hankel matrix of the modulated operating condition. The phase-constrained operating condition input block and the past acoustic print output block are combined to form a phase-constrained historical data space. The future acoustic print output block is orthogonally projected onto the phase-constrained historical data space to obtain the future acoustic print projection data constrained by the operating condition phase. Singular value decomposition is performed on the future acoustic signature projection data constrained by the working condition phase, and the number of acoustic signature change anchor points is counted within a set identification window. The ratio between the number of acoustic signature change anchor points and the set identification window length is determined as the acoustic signature change anchor point density. The working condition-acoustic signature hidden state order is determined by the singular value energy contribution, phase contribution intensity and acoustic signature change anchor point density. Extended observation information and working condition-acoustic signature hidden state sequence are extracted based on the working condition-acoustic signature hidden state order. Based on the degree of interpretation of the condition-soundprint hidden state sequence by the phase-constrained condition input block, the condition-soundprint hidden state sequence is divided into a baseline hidden state sequence and a degenerate hidden state sequence. The baseline hidden state sequence is used to characterize the normal soundprint changes driven by the operating condition, while the degenerate hidden state sequence is used to characterize the unexplained soundprint degradation changes of the operating condition. The baseline recursive relationship of the acoustic print driven by the baseline hidden state sequence is determined, the acoustic print degradation recursive relationship is determined by the degradation hidden state sequence, the working condition input action relationship is determined by the phase-constrained working condition input block, and the acoustic print observation relationship is determined by the acoustic print feature output block. The acoustic print baseline recursive relationship, the acoustic print degradation recursive relationship, the working condition input action relationship and the acoustic print observation relationship are combined to form the working condition-acoustic print state space representation relationship.

[0025] This invention introduces the phase attributes of operating condition parameters relative to the anchor point of acoustic text change into the state space identification process of the N4SID numerical subspace, constructs a phase-delay block Hankel matrix of operating condition modulation and a phase-constrained historical data space, so that the sequential relationship of the operating condition on acoustic text change participates in the state space identification, and distinguishes the operating condition-acoustic text latent state sequence into baseline latent state sequence and degenerate latent state sequence, thereby improving the ability of the operating condition-acoustic text state space representation relationship to separate normal acoustic text change and degenerate acoustic text change, and enhancing the accuracy of dynamic acoustic text baseline generation and fault prediction.

[0026] In this embodiment, obtaining the dynamic voiceprint baseline specifically includes: Extract the working condition input relationship, the voiceprint baseline recursive relationship, the voiceprint degradation recursive relationship, and the voiceprint observation relationship from the working condition-voiceprint state space representation relationship; Substituting the operating condition parameter sequence corresponding to the historical healthy operating status into the operating condition input action relationship and the acoustic fingerprint baseline recursion relationship, the baseline hidden state sequence corresponding to the historical healthy operating status is obtained. Substituting the baseline hidden state sequence into the acoustic fingerprint observation relationship, the corresponding healthy acoustic fingerprint observation sequence is obtained. The healthy observation subspace is determined by the healthy acoustic fingerprint observation sequence. The historical healthy operating status refers to the operating status of the transformer during the historical period when no confirmed fault occurred, no abnormal alarm was triggered, the operating parameters were within the allowable operating range, and the acoustic fingerprint characteristics did not show continuous abnormal deviation. Substitute the current operating condition parameters into the operating condition input action relationship and the voiceprint baseline recursion relationship to obtain the candidate baseline hidden state corresponding to the current operating condition. Substitute the candidate baseline hidden state into the health observation subspace for interpretation qualification determination, retain the hidden state part that satisfies the health observation subspace constraints, and obtain the health constraint baseline hidden state. Substituting the hidden state of the health constraint baseline into the voiceprint observation relationship, we obtain the voiceprint baseline component that can be explained by the current operating conditions. The part of the real-time voiceprint feature that can be explained by the voiceprint baseline component is marked as the operating condition-explainable voiceprint part, and the part of the real-time voiceprint feature that cannot be explained by the voiceprint baseline component is marked as the operating condition-unexplainable voiceprint part. Substitute the unexplainable voiceprint component under operating conditions into the voiceprint degradation recursive relation to obtain the health degradation component. Then, store the health degradation component in isolation from the voiceprint baseline component so that the health degradation component does not participate in the generation of the dynamic voiceprint baseline. Arrange the voiceprint baseline components in the order of the time window to generate the dynamic voiceprint baseline under the current operating conditions.

[0027] This invention constrains the hidden state of candidate baselines corresponding to the current operating condition through a health observation subspace, and divides real-time voiceprint features into an interpretable voiceprint part and an uninterpretable voiceprint part under the operating condition. This ensures that the dynamic voiceprint baseline is generated only from the interpretable normal voiceprint changes under the current operating condition, while isolating the uninterpretable voiceprint degradation changes as a health degradation component. This prevents early fault voiceprints from being absorbed into the normal baseline, thereby improving the reliability of the dynamic voiceprint baseline and the sensitivity of fault degradation identification under complex operating conditions.

[0028] In this embodiment, obtaining the fault-sensitive degradation component specifically includes: Read the dynamic voiceprint baseline and health degradation component, match the dynamic voiceprint baseline and health degradation component within the same time window according to the time window order, and align each voiceprint baseline feature in the dynamic voiceprint baseline with each degradation feature in the health degradation component according to the voiceprint feature dimension order to obtain baseline degradation alignment data. Based on the baseline degradation alignment data, the health degradation component within each time window is subjected to baseline normalization processing. The baseline normalization processing normalizes and converts the value of each degradation feature according to the amplitude of the voiceprint baseline feature under the same voiceprint feature dimension to obtain baseline normalized degradation data. Read historical fault acoustic fingerprint samples, perform data preprocessing and acoustic fingerprint feature extraction on the historical fault acoustic fingerprint samples, and divide the historical fault acoustic fingerprint samples into different fault sample groups according to the corresponding confirmed fault categories. The historical fault acoustic fingerprint samples refer to the acoustic fingerprint data and related operating condition data corresponding to the transformer before the confirmed fault occurred, during the fault occurred, and after the fault was confirmed. The acoustic fingerprint data has a clear fault category label. The degradation features corresponding to each historical fault voiceprint sample in the fault sample group are aggregated according to the voiceprint feature dimension, and the direction of the aggregated degradation features is determined as the fault-sensitive voiceprint direction corresponding to the current confirmed fault category. The fault-sensitive voiceprint directions are arranged in the order of the confirmed fault categories to form a fault-sensitive voiceprint direction set. The baseline normalized degradation data is matched with each fault-sensitive acoustic fingerprint direction in the fault-sensitive acoustic fingerprint direction set to obtain the matching results of the health degradation component in different fault-sensitive acoustic fingerprint directions within each time window. Based on the matching results, the baseline normalized degradation data is assigned to the corresponding fault-sensitive acoustic fingerprint direction to obtain fault-sensitive degradation components corresponding to different fault-sensitive acoustic fingerprint directions, and the fault-sensitive degradation components are arranged in order of time window.

[0029] This invention performs dimension-wise alignment and baseline normalization between the health degradation component and the dynamic acoustic print baseline, and uses historical fault acoustic print samples with confirmed fault category markers to form a fault-sensitive acoustic print direction set. This enables the health degradation component to be matched and decomposed according to different fault acoustic print directions, thereby transforming ordinary acoustic print deviations into fault-sensitive degradation components with fault directionality, improving the accuracy of early fault type identification, degradation trend tracking, and acoustic print monitoring result determination for transformers.

[0030] In this embodiment, the construction of the transformer health index specifically includes: Read the working condition-voiceprint state space representation relationship, health observation subspace, dynamic voiceprint baseline and fault-sensitive degradation component sequence, and extract the real-time voiceprint features, dynamic voiceprint baseline and fault-sensitive degradation components in the current identification window according to the time window order to form the voiceprint analysis data of the current window; Based on the working condition-voiceprint state space representation relationship, the real-time voiceprint features and corresponding working condition parameters in the current identification window are reconstructed to obtain the current window working condition-voiceprint hidden state sequence, and the current observation subspace is generated from the current window working condition-voiceprint hidden state sequence. Align the current observation subspace with the healthy observation subspace, calculate the principal angle of the current observation subspace relative to the healthy observation subspace, and use the weighted result of all principal angles as the structural drift. The principal angle is used to characterize the degree of deviation between the coupling relationship between the operating condition and the acoustic signature under the current operating state and the coupling relationship between the operating condition and the acoustic signature under the healthy operating state. The baseline deviation is calculated based on the dynamic voiceprint baseline and real-time voiceprint features. The baseline deviation is the degree of deviation of the real-time voiceprint features from the dynamic voiceprint baseline within the same time window. The fault-sensitive degradation intensity is calculated based on the fault-sensitive degradation component sequence. The fault-sensitive degradation intensity is the cumulative intensity of the fault-sensitive degradation component in the time dimension within the current identification window. A transformer health index is constructed based on structural drift, baseline deviation, and fault-sensitive degradation intensity. The transformer health index is the value obtained by subtracting the structural drift deduction, baseline deviation deduction, and fault-sensitive degradation deduction from the full health score. The structural drift deduction is determined by the structural drift amount, the baseline deviation deduction is determined by the baseline deviation amount, and the fault-sensitive degradation deduction is determined by the fault-sensitive degradation intensity.

[0031] This invention constructs a transformer health index by combining the structural drift between the current observation subspace and the healthy observation subspace, the baseline deviation of real-time acoustic signature features relative to the dynamic acoustic signature baseline, and the cumulative intensity of fault-sensitive degradation components in the time dimension. This allows the health status assessment to no longer rely on a single acoustic signature anomaly, but to simultaneously reflect the shift in the operating condition-acoustic signature coupling relationship, the degree of acoustic signature baseline deviation, and the characteristics of continuous fault degradation. This improves the accuracy, stability, and characterization ability of the quantitative assessment of transformer health status under complex operating conditions.

[0032] In this embodiment, the generation of transformer acoustic signature monitoring results specifically includes: The transformer health index, structural drift, and fault-sensitive degradation components are aligned one by one according to the same time window to obtain the data to be corrected. The increase of fault-sensitive degradation components, the increase of structural drift, and the decrease of transformer health index in the data to be corrected are checked sequentially along the time window. The health index deduction intensity is increased within the time window when all three conditions are met simultaneously, and the health index deduction intensity caused by structural drift is reduced within the time window when the three conditions are not met simultaneously, so as to obtain the corrected transformer health index. The corrected transformer health index is arranged in order of time window, the numerical changes between adjacent time windows are compared, and the health trend judgment result is formed based on the number of consecutively decreasing windows and the magnitude of the decrease. The working condition-acoustic signature state space representation relationship is updated based on the corrected transformer health index. When the corrected transformer health index meets the baseline update conditions, the acoustic signature baseline change within the corresponding time window is written into the acoustic signature baseline recursive relationship. When the corrected transformer health index does not meet the baseline update conditions, the fault-sensitive degradation component within the corresponding time window is written into the acoustic signature degradation recursive relationship. The baseline update conditions refer to the corrected transformer health index being within the allowable update range of the health index, the structural drift being within the allowable range of the structural drift, and the fault-sensitive degradation component being within the allowable range of the degradation component. Based on the updated working condition-soundprint state spatial representation relationship, the corrected transformer health index, and the health trend judgment results, transformer soundprint monitoring results are generated. The transformer soundprint monitoring results include soundprint abnormal state, fault sensitive direction, fault risk level, transformer health index, health trend judgment results, and health management recommendations.

[0033] Example 1: To verify the feasibility of the present invention in practice, it was applied to a long-term operating oil-immersed power transformer acoustic signature online monitoring scenario. The transformer is equipped with a cooling fan, oil temperature monitoring device, current acquisition device, voltage acquisition device, current harmonic acquisition device, environmental noise acquisition device, and multi-channel acoustic signature acquisition device. The acoustic signature acquisition channels are respectively arranged on the transformer body side, bushing side, radiator side, and cooling device side to collect acoustic signature signals generated by the transformer at different operating locations. The operating parameters include load rate, current, voltage, current harmonics, oil temperature, winding temperature, cooling device start / stop status, and environmental noise intensity.

[0034] In this scenario, the transformer's operating load fluctuates continuously, the cooling system starts and stops periodically, and environmental noise changes due to the operation of surrounding equipment. Existing fixed-threshold acoustic signature monitoring methods are prone to misinterpreting normal acoustic signature changes caused by increased load, fan startup, and enhanced environmental noise as faults in such scenarios. Furthermore, weak degraded acoustic signatures generated by early faults such as loose core, loose fasteners, and partial discharge are easily missed because their amplitude has not yet significantly exceeded the fixed threshold.

[0035] In the application of this invention, the system first preprocesses the multi-channel acoustic signature signals to form an acoustic signature feature sequence, and simultaneously organizes the operating condition parameters into an operating condition parameter sequence. Subsequently, the system determines the acoustic signature change anchor point based on the acoustic signature change amplitude in the acoustic signature feature sequence, and uses the acoustic signature change anchor point as a benchmark to analyze the initial change position, continuous change position, and fallback change position of various operating condition parameters relative to the acoustic signature change anchor point, thereby determining the stable influence phase of the operating condition parameters on the acoustic signature change. For load rate and current changes, the system identifies their leading influence on low-frequency acoustic signature energy; for the start-up and shutdown state of the cooling device, the system identifies its synchronous influence on the fan frequency band acoustic signature; for oil temperature and winding temperature, the system identifies its sustaining influence on some structural vibration acoustic signatures; and for environmental noise intensity, the system identifies its synchronous interference influence on broadband noise components.

[0036] Based on the aforementioned phase attributes, this invention reconstructs the operating condition parameter sequence into a time-delay input block, and jointly constructs the phase-delay block Hankel matrix of the operating condition modulation with the voiceprint feature output block. During the N4SID numerical subspace state space identification process, this invention combines the phase-constrained operating condition input block and the past voiceprint output block into a phase-constrained historical data space, and projects the future voiceprint output block onto this phase-constrained historical data space, thereby obtaining the operating condition-voiceprint state space representation relationship. This representation relationship can distinguish between voiceprint changes explainable by the current operating condition and health degradation changes not explainable by the current operating condition, ensuring that the dynamic voiceprint baseline is generated only by the voiceprint baseline component, without absorbing health degradation components into the normal baseline.

[0037] During monitoring, when the transformer load rate increases and causes an increase in low-frequency acoustic signature energy, conventional fixed threshold methods easily identify this change as abnormal. However, this invention can determine that the acoustic signature change belongs to the acoustic signature baseline component that can be explained under the current operating condition by using the operating condition-acoustic signature state space representation relationship, and incorporate it into the dynamic acoustic signature baseline. When the transformer exhibits a continuously increasing low-frequency periodic degradation component under relatively stable load conditions, this invention can identify this part as a health degradation component that cannot be explained under the current operating condition, and match it to the fault-sensitive acoustic signature direction related to core loosening or fastener loosening based on fault-sensitive acoustic signature direction decomposition. Subsequently, the system constructs a transformer health index by combining the dynamic acoustic signature baseline, fault-sensitive degradation component, and real-time acoustic signature features with the structural drift between the current observation subspace and the health observation subspace, and generates transformer acoustic signature monitoring results based on the corrected health index.

[0038] To verify the beneficial effects of the present invention, this embodiment compares the present invention with a conventional fixed-threshold voiceprint monitoring method. The conventional method only judges anomalies based on voiceprint energy, dominant frequency amplitude, and frequency band energy threshold. The present invention, however, constructs a time-delay input block carrying phase attributes according to the correspondence between voiceprint feature sequences and operating condition parameter sequences, and further constructs a phase-delayed time-delay block Hankel matrix modulated by the operating conditions to perform N4SID numerical subspace state-space identification. Specific comparison data is shown in Table 1. Table 1 Comparison of Transformer Acoustic Monitoring Performance under Complex Operating Conditions

[0039] As shown in Table 1, under the same data basis of 21,600 effective voiceprint data segments, the number of valid abnormal segments verified by the present invention increased from 286 segments to 351 segments, the number of false alarm segments decreased from 446 segments to 61 segments, and the anomaly identification accuracy increased from 39.1% to 85.2%. This indicates that the present invention can effectively reduce misjudgments caused by operating condition fluctuations through dynamic voiceprint baseline and fault-sensitive degradation components. At the same time, the number of early anomaly missed segments decreased from 98 segments to 23 segments, the average number of early warning windows increased from 11 windows to 38 windows, the fault-sensitive direction matching accuracy increased from 63.5% to 88.7%, and the consistency rate between the health index and the inspection and re-verification results increased from 68.2% to 90.5%. This shows that the present invention can improve the early fault prediction capability, fault direction determination capability, and health status assessment reliability.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A transformer acoustic signature monitoring method based on fault prediction and health management, characterized in that, Includes the following steps: Multi-channel acoustic signature signals and operating parameters are collected during transformer operation, and data preprocessing is performed to form acoustic signature feature sequences and operating parameter sequences. Based on the time delay difference in the response of the operating condition parameter sequence to the voiceprint feature sequence, the operating condition parameter sequence is reconstructed into a time delay input block; Construct the voiceprint feature output block and combine it with the time-delay input block to construct the phase-delay block Hankel matrix of the operating condition modulation. At the same time, perform N4SID numerical subspace state space identification to obtain the operating condition-voiceprint state space representation relationship. The health observation subspace is determined based on the relationship between operating conditions and voiceprint state space. The real-time voiceprint features are decomposed into voiceprint baseline components that can be explained by the current operating conditions and unexplainable health degradation components. A dynamic voiceprint baseline is generated based on the voiceprint baseline components. Based on the dynamic voiceprint baseline, the health degradation component is decomposed into fault-sensitive voiceprint direction to obtain the fault-sensitive degradation component. The current observation subspace is identified based on the working condition-acoustic print state space representation relationship, the structural drift of the current observation subspace is calculated, and a transformer health index is constructed. The transformer health index is corrected based on the fault-sensitive degradation component and structural drift. The operating condition-acoustic signature state space representation relationship is updated based on the corrected transformer health index, and the transformer acoustic signature monitoring results are generated.

2. The transformer acoustic signature monitoring method based on fault prediction and health management according to claim 1, characterized in that, The specific steps for obtaining the voiceprint feature sequence and the operating condition parameter sequence include: Multi-channel acoustic fingerprint signals and operating condition parameters during transformer operation are collected according to the same sampling time reference. The multi-channel acoustic fingerprint signals are organized according to the sampling time and sampling point order, and the operating condition parameters are organized according to the parameter category and parameter value order to obtain the original acoustic fingerprint data and original operating condition data. The original voiceprint data is preprocessed to obtain preprocessed voiceprint data, and the operating condition parameter data corresponding to the preprocessed voiceprint data is extracted from the original operating condition data. The voiceprint data is continuously segmented according to the set sliding window length and sliding step size to obtain voiceprint data segments. Voiceprint feature extraction is performed on each voiceprint data segment to obtain a voiceprint feature sequence. Then, the operating condition parameter data of each voiceprint data segment is aggregated to obtain the corresponding operating condition parameter sequence.

3. The transformer acoustic signature monitoring method based on fault prediction and health management according to claim 1, characterized in that, The specific steps to obtain the time-delay input block include: Read the voiceprint feature sequence and the operating condition parameter sequence, calculate the voiceprint change amplitude between adjacent time windows in the voiceprint feature sequence, and determine the time window that meets the preset change condition as the voiceprint change anchor point; For each type of operating condition parameter in the operating condition parameter sequence, based on the voiceprint change anchor point, the operating condition parameter segments before the current voiceprint change anchor point, at the current corresponding time, and after the current voiceprint change anchor point are extracted respectively. Based on the start position, duration position, and fall position of parameter change in the operating condition parameter segment, the influence phase record of the operating condition parameter relative to the voiceprint change anchor point is generated. Based on the influence phase records formed by the same type of operating parameters under different acoustic text change anchor points, the stable influence phase is determined, and the operating parameters are divided into leading response operating parameters, synchronous response operating parameters, lagging response operating parameters, or maintenance response operating parameters according to the stable influence phase. According to the stable influence phase, configure phase attribute identifiers for various operating condition parameters, and perform phase alignment reconstruction on the sequence of various operating condition parameters based on the phase attribute identifiers. The lead response operating condition quantity is entered into the input position before the acoustic text change anchor point, the synchronous response operating condition quantity is entered into the input position corresponding to the acoustic text change anchor point, the hysteresis response operating condition quantity is entered into the input position after the acoustic text change anchor point, and the maintenance response operating condition quantity is entered into the continuous input position covering the acoustic text change anchor point, thus obtaining the phase-separated operating condition sub-blocks carrying phase attributes. According to the time sequence of the voiceprint change anchor points, the phase-separated sub-blocks carrying phase attributes are spliced ​​together, and the phase order of the lead response, synchronization response, hysteresis response and maintenance response corresponding to each phase-separated sub-block is maintained to obtain the time-delay input block.

4. The transformer acoustic signature monitoring method based on fault prediction and health management according to claim 1, characterized in that, The specific methods for obtaining the working condition-voiceprint state space representation relationship include: Based on the phase attributes carried by the time-delay input block, the operating condition parameter segments associated with the same acoustic text change anchor point are compressed into phase-separated operating condition input blocks, and past acoustic text output blocks and future acoustic text output blocks are constructed from the acoustic text feature sequence. Based on the phase position of each operating condition parameter segment in the phase-separated operating condition input block relative to the acoustic text change anchor point, the phase contribution intensity of the phase-separated operating condition input block to the past acoustic text output block is calculated, and the phase-separated operating condition input block is rearranged and weighted according to the phase contribution intensity to obtain the phase-constrained operating condition input block. The phase-constrained operating condition input block, past voiceprint output block and future voiceprint output block are arranged in layers according to the block Hankel arrangement rule to form the phase-delay block Hankel matrix of the operating condition modulation. The N4SID numerical subspace state space identification is performed on the phase-delay block Hankel matrix of the modulated operating condition. The phase-constrained operating condition input block and the past acoustic print output block are combined to form a phase-constrained historical data space. The future acoustic print output block is orthogonally projected onto the phase-constrained historical data space to obtain the future acoustic print projection data constrained by the operating condition phase. Singular value decomposition is performed on the future acoustic signature projection data constrained by the working condition phase, and the number of acoustic signature change anchor points is counted within a set identification window. The ratio between the number of acoustic signature change anchor points and the set identification window length is determined as the acoustic signature change anchor point density. The working condition-acoustic signature hidden state order is determined by the singular value energy contribution, phase contribution intensity and acoustic signature change anchor point density. Extended observation information and working condition-acoustic signature hidden state sequence are extracted based on the working condition-acoustic signature hidden state order. Based on the degree to which the phase-constrained working condition input block interprets the working condition-soundprint hidden state sequence, the working condition-soundprint hidden state sequence is divided into the baseline hidden state sequence and the degenerate hidden state sequence. The baseline recursive relationship of the acoustic print driven by the baseline hidden state sequence is determined, the acoustic print degradation recursive relationship is determined by the degradation hidden state sequence, the working condition input action relationship is determined by the phase-constrained working condition input block, and the acoustic print observation relationship is determined by the acoustic print feature output block. The acoustic print baseline recursive relationship, the acoustic print degradation recursive relationship, the working condition input action relationship and the acoustic print observation relationship are combined to form the working condition-acoustic print state space representation relationship.

5. The transformer acoustic signature monitoring method based on fault prediction and health management according to claim 1, characterized in that, The dynamic voiceprint baseline is obtained specifically through: Extract the working condition input relationship, the voiceprint baseline recursive relationship, the voiceprint degradation recursive relationship, and the voiceprint observation relationship from the working condition-voiceprint state space representation relationship; Substitute the sequence of operating parameters corresponding to the historical healthy operating status into the operating condition input action relationship and the voiceprint baseline recursion relationship to obtain the baseline hidden state sequence corresponding to the historical healthy operating status. Substitute the baseline hidden state sequence into the voiceprint observation relationship to obtain the corresponding healthy voiceprint observation sequence, and determine the healthy observation subspace from the healthy voiceprint observation sequence. Substitute the current operating condition parameters into the operating condition input action relationship and the voiceprint baseline recursion relationship to obtain the candidate baseline hidden state corresponding to the current operating condition. Substitute the candidate baseline hidden state into the health observation subspace for interpretation qualification determination, retain the hidden state part that satisfies the health observation subspace constraints, and obtain the health constraint baseline hidden state. Substituting the hidden state of the health constraint baseline into the voiceprint observation relationship, we obtain the voiceprint baseline component that can be explained by the current operating conditions. The part of the real-time voiceprint feature that can be explained by the voiceprint baseline component is marked as the operating condition-explainable voiceprint part, and the part of the real-time voiceprint feature that cannot be explained by the voiceprint baseline component is marked as the operating condition-unexplainable voiceprint part. Substitute the unexplainable voiceprint component under operating conditions into the voiceprint degradation recursive relation to obtain the health degradation component. Then, store the health degradation component in isolation from the voiceprint baseline component so that the health degradation component does not participate in the generation of the dynamic voiceprint baseline. Arrange the voiceprint baseline components in the order of the time window to generate the dynamic voiceprint baseline under the current operating conditions.

6. The transformer acoustic signature monitoring method based on fault prediction and health management according to claim 1, characterized in that, The specific steps to obtain the fault-sensitive degradation component include: Read the dynamic voiceprint baseline and health degradation component, match the dynamic voiceprint baseline and health degradation component within the same time window according to the time window order, and align each voiceprint baseline feature in the dynamic voiceprint baseline with each degradation feature in the health degradation component according to the voiceprint feature dimension order to obtain baseline degradation alignment data. Based on the baseline degradation aligned data, the health degradation component within each time window is subjected to baseline normalization processing to obtain baseline normalized degradation data. Read historical fault voiceprint samples, perform data preprocessing and voiceprint feature extraction on the historical fault voiceprint samples, and divide the historical fault voiceprint samples into different fault sample groups according to the corresponding confirmed fault category. The degradation features corresponding to each historical fault voiceprint sample in the fault sample group are aggregated according to the voiceprint feature dimension, and the direction of the aggregated degradation features is determined as the fault-sensitive voiceprint direction corresponding to the current confirmed fault category. The fault-sensitive voiceprint directions are arranged in the order of the confirmed fault categories to form a fault-sensitive voiceprint direction set. The baseline normalized degradation data is matched with each fault-sensitive acoustic fingerprint direction in the fault-sensitive acoustic fingerprint direction set to obtain the matching results of the health degradation component in different fault-sensitive acoustic fingerprint directions within each time window. Based on the matching results, the baseline normalized degradation data is assigned to the corresponding fault-sensitive acoustic fingerprint direction to obtain fault-sensitive degradation components corresponding to different fault-sensitive acoustic fingerprint directions, and the fault-sensitive degradation components are arranged in order of time window.

7. The transformer acoustic signature monitoring method based on fault prediction and health management according to claim 1, characterized in that, The construction of the transformer health index specifically includes: Read the working condition-voiceprint state space representation relationship, health observation subspace, dynamic voiceprint baseline and fault-sensitive degradation component sequence, and extract the real-time voiceprint features, dynamic voiceprint baseline and fault-sensitive degradation components in the current identification window according to the time window order to form the voiceprint analysis data of the current window; Based on the working condition-voiceprint state space representation relationship, the real-time voiceprint features and corresponding working condition parameters in the current identification window are reconstructed to obtain the current window working condition-voiceprint hidden state sequence, and the current observation subspace is generated from the current window working condition-voiceprint hidden state sequence. Align the current observation subspace with the healthy observation subspace, calculate the principal angle of the current observation subspace relative to the healthy observation subspace, and use the weighted result of all principal angles as the structural drift. The baseline deviation is calculated based on the dynamic voiceprint baseline and real-time voiceprint features. The baseline deviation is the degree of deviation of the real-time voiceprint features from the dynamic voiceprint baseline within the same time window. The fault-sensitive degradation intensity is calculated based on the fault-sensitive degradation component sequence. A transformer health index is constructed based on structural drift, baseline deviation, and fault-sensitive degradation intensity.

8. The transformer acoustic signature monitoring method based on fault prediction and health management according to claim 1, characterized in that, The generation of the transformer acoustic signature monitoring results specifically includes: The transformer health index, structural drift, and fault-sensitive degradation components are aligned one by one according to the same time window to obtain the data to be corrected. The increase of fault-sensitive degradation components, the increase of structural drift, and the decrease of transformer health index in the data to be corrected are checked sequentially along the time window. The health index deduction intensity is increased within the time window when all three conditions are met simultaneously, and the health index deduction intensity caused by structural drift is reduced within the time window when the three conditions are not met simultaneously, so as to obtain the corrected transformer health index. The corrected transformer health index is arranged in order of time window, the numerical changes between adjacent time windows are compared, and the health trend judgment result is formed based on the number of consecutively decreasing windows and the magnitude of the decrease. The working condition-soundprint state space representation relationship is updated based on the corrected transformer health index. When the corrected transformer health index meets the baseline update condition, the soundprint baseline change in the corresponding time window is written into the soundprint baseline recursive relationship. When the corrected transformer health index does not meet the baseline update condition, the fault-sensitive degradation component in the corresponding time window is written into the soundprint degradation recursive relationship. Based on the updated working condition-soundprint state space representation relationship, the corrected transformer health index, and the health trend judgment results, transformer soundprint monitoring results are generated.

Images