Method for predicting life of power capacitor based on time series learning

Abstract

The present application relates to the field of artificial intelligence and power equipment state prediction technology, specifically to a power capacitor life prediction method based on time series learning. Multivariate state time series data formed by continuous operation of the power capacitor is obtained, time series state representation of local operation segments is extracted, degradation stage variables are determined according to state evolution direction, historical change trajectory and cumulative operation context, and local operation segments with different degradation speeds are mapped to a unified degradation coordinate; an implicit health state sequence is constructed based on the aligned time series state representation, irreversible degradation direction constraints are used to distinguish continuous degradation and transient observation recovery, and the remaining life is determined in combination with the stage advancing rate and the life termination state interval. The present application reduces the life mapping deviation caused by the misplacement of natural operation time and actual degradation degree, suppresses the reverse jump of health state caused by short-time working condition changes, makes the remaining life correspond to the degradation stage and the degradation speed, and improves the stability of life prediction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and power equipment condition prediction technology, specifically a method for predicting the lifespan of power capacitors based on time-series learning. Background Technology

[0002] Power capacitor lifetime prediction typically relies on long-term operational time-series data, including voltage, current, temperature, switching status, and harmonics, as the basis for analysis. The collected data is organized into a continuous sequence according to a uniform sampling period, and local variation features are extracted using a sliding time window. Existing implementations often establish sample order based on natural operating time, considering data from different capacitors operating near the same duration as comparable. They utilize cyclic time-series models, attention-based time-series models, or multi-level sequence prediction models to learn the mapping relationship between historical states and remaining lifetime. The training phase usually uses the actual lifetime of failed samples or manually defined health stages as supervisory information, determining model parameters by minimizing the deviation between predicted and actual lifetimes. The prediction phase inputs the most recent operating sequence of the capacitor under test into the trained model, directly outputting the remaining lifetime or lifetime range based on changes in time-series characteristics. This approach has a clear implementation path and can use continuous operational data to reflect the overall trend of capacitor state changes over time.

[0003] In the specific data alignment process, existing lifetime models typically continue to use sampling time, cumulative operating hours, or fixed-length windows as the reference for sequence location. To mitigate the impact of inconsistent operating cycles between different devices, some schemes interpolate, normalize, truncate, or resample the time series to ensure that each training sample has the same input length; other schemes add conditional variables such as load level, temperature, and switching frequency, allowing the model to learn the relationship between operating condition differences and lifetime changes. For parameter declines or short-term recoveries during operation, the model generally still directly writes the observed changes into the hidden state, allowing the time series network to learn its statistical meaning from a large number of samples. Since the aging rate of different power capacitors may continuously change with long-term load, thermal stress, and switching history, even if the natural time is close, their actual degradation degree may not be consistent; conversely, capacitors with similar degradation degrees may be distributed at completely different operating time locations.

[0004] Lifetime models based on naturally time-organized samples cannot stably establish a consistent correspondence between sequence positions and actual degradation stages. When slow-aging and fast-aging samples are simultaneously included in the training set, the same runtime may correspond to different health levels, and the same health level may be scattered across multiple time points. Therefore, the lifetime mapping learned by the model is mixed with differences in runtime and actual degradation. When operating conditions change, observations such as voltage, current, temperature, or harmonics may show a phased decline. If this decline is directly interpreted as a recovery in health, the implicit state will exhibit a reverse jump, causing the remaining lifetime to show increases or fluctuations that do not conform to the continuous aging process with short-term changes in operating conditions. Existing processing lacks a joint mechanism that can map time-series data with different degradation rates to a unified degradation scale while simultaneously constraining the continuous evolution of health status along the actual aging direction. This makes lifetime prediction susceptible to interference from time scale misalignment and transient parameter recovery. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the lifetime of power capacitors based on time-series learning, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The power capacitor lifetime prediction method based on time series learning includes: acquiring multivariate state time series data formed by the continuous operation of power capacitors, and generating state sequences according to time correlation relationships; Based on the state sequence, extract the temporal state representation of the local running segment; Based on the state evolution direction, historical state change trajectory and cumulative operational context between adjacent local operational segments, the degradation stage variables corresponding to each local operational segment are determined; Based on the aforementioned degradation stage variables, local runtime segments at different degradation rates are aligned according to their degradation stages. Based on the aligned temporal state representation, a sequence of implicit healthy states that progresses along the degradation stage is constructed, and an irreversible degradation direction constraint is imposed on the sequence of implicit healthy states. The remaining lifetime of the power capacitor is determined based on the implicit health state sequence, the progression of the degradation stage, and their temporal changes.

[0007] Preferably, the determination of the degradation stage variables includes: forming a stage discrimination sequence for each local running segment, consisting of the current temporal state representation, the state representation of the previous local running segment, and the cumulative running context; Calculate the state displacement along the historical degradation direction between adjacent stage discrimination sequences, and decompose the state displacement into a propulsion component consistent with the existing degradation trajectory and a disturbance component deviating from the existing degradation trajectory; The candidate degradation stage of the current local running segment is updated based on the propagation component, and the candidate degradation stage is continuously corrected with the historically determined degradation stage as the order constraint to obtain the degradation stage variable that advances unidirectionally with the running sequence.

[0008] Preferably, the degradation stage alignment includes: constructing non-uniform stage coordinates using degradation stage variables corresponding to each power capacitor, and mapping local running segments on the natural time axis to the non-uniform stage coordinates; For local running segments that are in adjacent degradation stage intervals but have inconsistent sampling times, a stage correspondence is established based on the evolutionary continuity between the temporal state representations; According to the stage correspondence, the state representation of the local running segment is resampled in the stage domain to form an aligned state sequence with a unified degenerate stage position, and the stage advancement rate formed when each local running segment is mapped from natural time to the degenerate stage is retained.

[0009] Preferably, the irreversible degradation direction constraint includes: generating healthy state candidate values ​​based on the temporal state representations corresponding to adjacent degradation stages in the aligned state sequence, and determining the main degradation direction using the healthy state candidate values ​​of multiple consecutive degradation stages; The changes in health status along the main degradation direction are written into the implicit health status sequence, and the changes in status that are opposite to the main degradation direction and do not continuously cross multiple subsequent degradation stages are retained in the observed change components. When the reverse state change is maintained in the continuous degradation phase and changes the evolution trajectory of the subsequent healthy state, the main degradation direction of the corresponding phase is re-determined and the implicit healthy state sequence is updated.

[0010] Preferably, the decomposition of the state displacement includes: constructing a local degradation trajectory using a temporal state representation formed by multiple historical local running segments, and determining the degradation tangential representation of the current stage using the local degradation trajectory; The state change of the current local running segment relative to the previous local running segment is projected onto the degenerate tangential representation and its remaining state subspace to form the propulsion component and the disturbance component, respectively. The degenerate tangential representation is updated based on the degree of consistency of the direction of the propulsion components in the continuous local operation segments, and the state changes that frequently reverse direction and do not form a continuous displacement are restricted to the perturbation components, so that the stage discrimination sequence progresses segment by segment according to the self-updating local degradation trajectory.

[0011] Preferably, the stage domain resampling includes: for a state sequence of multiple local running segments within the same degradation stage interval, determining the stage position weight based on the stage advancement rate and the degree of continuity of state evolution of each local running segment; The stage position weights are used to map local running segments with different natural time intervals to a unified stage node; For segments where there is a lack of direct observation between adjacent stage nodes, an intermediate stage representation constrained by continuity is generated based on the temporal state representation of the preceding and following stage nodes, the stage advancement rate, and the degradation direction. The state representation corresponding to the original observation is then combined with the intermediate stage representation to form a stage domain temporal sequence.

[0012] Preferably, the determination of the observed change components includes: for the degradation stage in the implicit health state sequence where a reverse change occurs, extracting the state representations of multiple stages before and after the degradation stage, and constructing a baseline before the reverse change and a baseline after the reverse change, respectively; The state after the reverse change is compared to indicate whether it reverts to the baseline before the reverse change, and it is simultaneously determined whether the main degradation direction in the subsequent stage remains the original direction; State changes that satisfy the baseline regression relationship and do not change the subsequent main degradation direction are included in the observed change component, and state changes that continuously deviate from the baseline before the reverse change and form a new degradation trajectory are included in the implicit health status update.

[0013] Preferably, the update of the implicit health state sequence includes: establishing a stage conditional association between the degradation stage variable and the health state candidate value, and recording the state difference with the preceding health state and the corresponding stage span for each degradation stage; A stage consistency record is formed based on the state difference direction of multiple consecutive stages, and the state difference that satisfies the main degradation direction is written into the health state recursion process according to the stage span. For a degradation stage with the observed change components, the preceding implicit healthy state is kept as the recursive benchmark, and only the continuous state changes that subsequently re-enter the main degradation direction are incorporated into the recursive result, forming an implicit healthy state sequence consistent with the unidirectional progression relationship of the degradation stage.

[0014] Preferably, the determination of the stage advancement rate includes: forming a basic advancement rate based on the natural time span and degradation stage span corresponding to adjacent stage nodes in the stage domain time sequence, and using the continuity of the formed advancement component as a rate reliability constraint; When the base propulsion rate of a continuous stage node changes abruptly, it is compared whether the corresponding propulsion component synchronously forms a continuous change. If no continuous change is formed, the base propulsion rate is continuously corrected along the preceding degenerate tangent. If the propulsion component changes synchronously with the basic propulsion rate, a new local propulsion segment is established, and the subsequent stage propulsion rate sequence is re-formed within the local propulsion segment.

[0015] Preferably, the determination of the remaining lifetime includes: constructing a lifetime time series representation based on the implicit health state corresponding to the current degradation stage, the health state change sequence formed by multiple consecutive stages, and the determined stage advancement rate sequence; The degradation stage is pushed forward along the current local advancement segment, and the implicit healthy state is maintained to evolve along the defined main degradation direction during the pushing process; When the recursive health state reaches the lifetime end state interval formed by the corresponding degradation termination stage in the historical failure sequence, the remaining stage distance between the current degradation stage and the lifetime end state interval is determined, and the remaining stage distance is converted into the remaining lifetime of the power capacitor according to the stage advancement rate sequence.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By remapping local operating segments on the natural timeline to a unified degradation coordinate system using degradation stage variables, different power capacitors are no longer directly compared based on operating time, but rather their degradation positions are determined by the direction of state evolution, historical change trajectories, and cumulative operating context. Local state displacements are decomposed into a propulsion component along the existing degradation trajectory and a disturbance component deviating from the trajectory. The propulsion component continuously updates the stage position, while the disturbance component is restricted from stage discrimination, allowing rapid aging and slow aging sequences to form a correspondence at similar degradation levels. After stage alignment, the implicit health state continues to be constrained by the irreversible degradation direction. Reverse changes that do not continuously cross subsequent stages are retained as observed changes and are not directly written into the health state recursion process. This reduces interference caused by natural time misalignment and short-term state recovery entering the lifetime mapping simultaneously, ensuring a continuous correspondence between remaining lifetime changes and the actual degradation stage and stage propagation speed, and reducing the probability of reverse jumps and large fluctuations in lifetime results during operating condition switching.

[0017] 2. Stage-domain resampling and continuous advance rate correction further improve sequence consistency under different sampling rhythms and degradation rates. Local operating segments are reorganized in non-uniform degradation coordinates according to stage position weights. For adjacent stage segments lacking direct observation, continuous intermediate stage representations are formed using the state representations of preceding and following stages, stage advance rates, and degradation directions, avoiding fixed-time interpolation that brings natural time differences back into the degradation sequence. During health status updates, the baseline relationship before and after reverse changes is used to determine whether the state decline truly changes the subsequent degradation trajectory, and changes that do not change the main degradation direction are left in the observation layer. When the advance component and the base advance rate form a continuous change synchronously, a new local advance segment is established. The remaining lifetime is jointly determined by the current health status, the continuous stage change sequence, and the advance rate, enabling lifetime calculation to adjust with the actual degradation rhythm, reducing prediction discrepancies caused by differences in sampling intervals, operating durations, and local recovery processes between different devices, and maintaining the temporal stability of lifetime output during continuous operation. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall process of predicting the lifetime of power capacitors based on time-series learning. Figure 2 A recurrence relation for variables in the degradation stage; Figure 3 Flowcharts for self-alignment during the degradation phase and irreversible healthy states; Figure 4 Flowchart for stage advance rate correction and remaining lifetime determination; Figure 5 Contribution map for segmented remaining lifetime. Detailed Implementation

[0019] In one embodiment, please refer to Figure 1A time-series learning-based method for predicting the lifetime of power capacitors includes: acquiring multivariate state time-series data generated by continuous operation of power capacitors and generating state sequences according to temporal relationships; extracting time-series state representations of local operating segments based on the state sequences; determining degradation stage variables corresponding to each local operating segment based on the state evolution direction, historical state change trajectory, and cumulative operating context between adjacent local operating segments; aligning degradation stages of local operating segments at different degradation rates based on the degradation stage variables; constructing a hidden healthy state sequence advancing along the degradation stages based on the aligned time-series state representations and imposing irreversible degradation direction constraints on the hidden healthy state sequence; and determining the remaining lifetime of the power capacitor based on the hidden healthy state sequence, the advancing degradation stages, and their time-series relationship. In practical implementation, multivariate state time series data are associated according to the operating time sequence of the same power capacitor. Each sampling time corresponds to a state vector. Before entering the time series learning process, the data of each dimension in the state vector are marked with missing positions, retained with abnormal mutations, and scaled according to their own historical distribution. The missing positions are not directly replaced by fixed constants, but missing markers that can participate in time series coding are generated synchronously. Abnormal mutations are corrected only when it can be determined that they are acquisition errors. The original change relationship is maintained for mutations that may reflect changes in operating state. This avoids the elimination of time series changes that can characterize the degradation process in advance during the preprocessing stage. The processed state vectors are arranged into a continuous state sequence according to the time sequence, and the natural time information corresponding to each state vector is retained for use in the subsequent stage of advance rate conversion.

[0020] In this embodiment, local operating segments are not extracted according to the principle that different power capacitors must have the same operating time. Instead, they are constructed in the continuous state sequence of a single power capacitor according to the information continuity between adjacent states. Adjacent local operating segments are allowed to have partial time overlap so that a state change process can participate in the encoding of the end state of the previous local segment and the beginning state of the next local segment simultaneously. Each local operating segment is input into a temporal coding network in chronological order. The temporal coding network performs correlation calculations between the current state vector and the preceding implicit representation, and compresses the state level, change direction, change duration, and dependencies within the local segment into a temporal state representation with a fixed dimension. The temporal state representation is not directly used as the regression input for the remaining lifetime label, but as the basis for determining the actual degradation stage. For two local operating segments with similar natural operating times but different state evolution directions, their temporal state representations can maintain different trajectory positions. For local operating segments with large differences in natural operating times but whose state change patterns have entered similar degradation intervals, similar stage positions are allowed to be formed in the subsequent degradation stage mapping process. This decouples the fixed binding relationship between natural time position and degradation degree at the data organization level.

[0021] Furthermore, the historical state change trajectory is composed of multiple consecutive time-series state representations preceding the current local operating segment in the order of their generation, and forms a local degradation path through the differential relationship between adjacent time-series state representations. The cumulative operating context is composed of the natural operating time preceding the current local operating segment, the historical stage advancement record, and the long-term state representation maintained across segments. The stage discrimination process jointly compares the current time-series state representation with the historical state change trajectory. Instead of directly determining the degradation progress based on the increase or decrease of a single state value, it identifies the projection direction and duration of the current state change in the historical degradation path. When the current state change is consistent with the degradation direction formed by multiple previous local segments, the corresponding change is used as the basis for candidate stage advancement. When the current state change deviates significantly from the historical degradation path and then returns to the vicinity of the original trajectory, the corresponding change is retained as an operating disturbance without causing the degradation stage to move in the opposite direction. This forms a degradation stage sequence that can nonlinearly expand and contract with the aging rate of different power capacitors. The degradation stage sequence only describes the relative position of advancement from the initial operating state to the end-of-life state. Natural time is only used as the basis for calculating the speed of stage advancement and does not directly determine the stage position.

[0022] In this embodiment, the degradation stage variables are recursively calculated according to the positive projection of the current state change along the existing local degradation trajectory. One feasible calculation form is as follows: ; in, This represents the degradation stage variable corresponding to the t-th local execution segment. This variable represents the degradation stage of the previous local execution segment. This represents the state shift vector between the current time-series state representation and the previous time-series state representation. This represents the localized degenerate tangential representation determined based on the displacements of multiple previous states in the current segment. This represents the inner product of the current state displacement along the existing degradation direction. The square norm of the locally degenerate tangential representation is denoted by . This represents a positive stable term used to avoid a denominator of zero. This means that when the projection result is less than 0, it is taken as 0; when the projection result is greater than or equal to 0, the projection result is retained. In a specific calculation case, if the current state displacement is consistent with the direction of the local degradation tangential representation, the inner product is positive, and the degradation stage variable is increased by the corresponding advance based on the previous stage. If the current state displacement is mainly caused by the recovery of short-term conditions and the direction is opposite to the local degradation tangential representation, the inner product is negative. After the maximum value calculation, the stage increment is 0, so that the natural time continues to increase while the degradation stage does not produce reverse movement. If the current state displacement is approximately perpendicular to the local degradation direction, the stage increment remains small and is combined with the trajectory update result in subsequent continuous segments to determine whether it constitutes a new continuous degradation direction.

[0023] Please refer to Figure 2 , Figure 2 The horizontal axis represents the stage position of the degradation stage variable in the complete running sequence, with different segment intervals represented by the start segment, the beginning segment, the middle-early segment, the middle segment, the middle-late segment, the end segment, and the final segment; the vertical axis represents the degradation stage variable. The value of is given. The black curve represents the stage evolution result of the power capacitor in the local time series state during the recursive process. The dashed line and dotted line represent the stage stability reference line and the key review reference line, respectively, which are used to help determine whether the current segment has entered the stable degradation range and whether it needs to be included in the subsequent lifetime review.

[0024] from Figure 2 visible, As the running segments progress, the overall trend is not downward, but a more obvious step-like increase occurs from the early to the middle stages, indicating that the state displacement... Tangential to historical local degradation in these fragments Maintain strong consistency, so that the variables in the current stage are relative to Effective progress is achieved; if the curve remains horizontal in a local interval, it indicates that the observed changes in that interval have not formed a continuous degradation direction, and the relevant changes are more suitable to be classified as disturbance components or short-term operating condition fluctuations, rather than directly rewriting the degradation stage position.

[0025] Once the curve crosses the stage stability reference line, the degradation stage variables enter a more defined middle to later region, indicating that the forward projection advancements of multiple consecutive local operating segments have accumulated to a level that can stably characterize the degree of degradation. When the curve approaches or exceeds the key verification reference line, subsequent remaining lifetime prediction should focus on the changes caused by… Determining the current stage position, stage progress rate, and the continuity of the implied health status helps avoid misjudging short-term state decline as life recovery.

[0026] In a preferred embodiment, please refer to Figure 3The determination of degradation stage variables includes: forming a stage discrimination sequence for each local operation segment, consisting of the current temporal state representation, the state representation of the previous local operation segment, and the cumulative operation context; calculating the state displacement along the historical degradation direction between adjacent stage discrimination sequences, and decomposing the state displacement into a propulsion component consistent with the existing degradation trajectory and a disturbance component deviating from the existing degradation trajectory; updating the candidate degradation stage of the current local operation segment based on the propulsion component, and performing continuous correction on the candidate degradation stages using historically determined degradation stages as sequence constraints, to obtain degradation stage variables that advance unidirectionally with the operation sequence. Specifically, the stage discrimination... The current temporal state representation in the sequence describes the current local state, the state representation of the preceding local running segment provides a short-term change reference, and the cumulative running context provides a long-term reference across multiple local segments. These three types of information are associated in the same representation space, enabling the current change to be compared with the recent change direction and the long-term evolution path simultaneously. The propulsion component is obtained by projecting the state displacement onto the local degradation tangential direction, and the perturbation component is obtained by the remaining vector after subtracting the propulsion component from the original state displacement. The candidate degradation stage only absorbs the positive persistence part of the propulsion component, avoiding a reverse state change from directly overwriting the previously accumulated degradation progress.

[0027] In a preferred embodiment, continuous correction is constrained by the sequential relationship of the historical degradation stage sequence. Instead of simply truncating all negative increments and directly accumulating them, it establishes a local continuity relationship between candidate stages formed by multiple adjacent local operating segments. When a candidate stage suddenly moves forward significantly but its subsequent state returns to the vicinity of the previous degradation trajectory, the current mutation amount is retained in the perturbation component. The candidate stage is re-determined according to the continuous advancement relationship of the preceding and following stages. When the advancement magnitude of a candidate stage changes and remains in the same direction in multiple subsequent local operating segments, the stage advancement slope is allowed to change, while maintaining the stage position no less than the previously determined stage position. This can represent the process of the power capacitor moving from slow degradation to a faster degradation range. The historically determined stages serve as irreversible sequential anchors in subsequent calculations. Each update only determines how much the current stage increases relative to the previous anchor without rewriting the already formed stage order. During training, the sequential relationship from the initial state to the failure state of the complete lifetime sequence can be used to constrain the stage representation, so that the stage coordinates formed by different samples have the same starting and ending directions, while not requiring the same natural running length to correspond to the same stage value.

[0028] Further, in a preferred embodiment, degradation stage alignment includes: constructing non-uniform stage coordinates using degradation stage variables corresponding to each power capacitor, and mapping local operating segments on the natural time axis to the non-uniform stage coordinates; for local operating segments that are in adjacent degradation stage intervals but have inconsistent sampling times, establishing a stage correspondence relationship based on the evolutionary continuity between time-series state representations; resampling the state representations of local operating segments in the stage domain according to the stage correspondence relationship to form an aligned state sequence with a unified degradation stage position, and retaining the stage advancement rate formed when each local operating segment is mapped from natural time to the degradation stage. The non-uniform stage coordinates use degradation stage variables as the horizontal position rather than sampling time as the horizontal position. Therefore, operating segments that are longer in natural time but have slower state advancement are compressed in the stage coordinates, while operating segments that are shorter in natural time but have continuous degradation advancement retain a relatively large stage span in the stage coordinates. Local operating segments of different power capacitors can participate in the correspondence calculation as long as they are in similar degradation stage intervals, without needing to have the same sampling time, the same cumulative running time, or the same original sequence length.

[0029] When establishing the correspondence between stages, the temporal state representation within the target stage interval and the preceding and following state representations within adjacent stage intervals are input into the continuity discrimination process. The continuity discrimination simultaneously compares the distance in the representation space, the direction of state displacement, and the local stage advancement trend. If two local running segments from different natural time positions are close in stage variables and their preceding and following state evolution directions can form a continuous trajectory, then they are mapped to the same or adjacent unified stage nodes. If the stage variables are close but the state evolution directions are obviously opposite, then their respective stage positions are maintained and the correspondence between them is reduced to avoid incorrectly mixing samples with different state trajectories based solely on the proximity of stage values. After the correspondence is established, the local state representation of each power capacitor is rearranged according to the unified stage node. The same unified stage node allows observations from original segments at different natural time positions, enabling the training process to compare the state patterns of different power capacitors at the same degradation position. The original natural time information is not deleted but forms a stage advancement rate record together with the mapping span between unified stage nodes. Subsequent lifetime conversion can still recover the relationship from stage distance to natural time distance.

[0030] In a preferred embodiment, the irreversible degradation direction constraint includes: generating health state candidate values ​​based on the temporal state representations corresponding to adjacent degradation stages in the aligned state sequence, and determining the main degradation direction using the health state candidate values ​​of multiple consecutive degradation stages; writing health state changes along the main degradation direction into the implicit health state sequence, and retaining state changes opposite to the main degradation direction and not continuously crossing multiple subsequent degradation stages in the observed change component; when the reverse state change is maintained in the consecutive degradation stages and changes the subsequent health state evolution trajectory, the main degradation direction of the corresponding stage is re-determined and the implicit health state sequence is updated. The health state candidate values ​​are obtained by continuous mapping from the aligned temporal state representations. The continuous mapping maintains the order relationship between adjacent stages, so that the candidate values ​​represent the health position of the current state relative to the historical degradation path. The main degradation direction is not fixed in advance as an increase or decrease of a certain data dimension, but is determined based on the stable evolution direction of the health state candidate values ​​in multiple consecutive degradation stages of the same power capacitor. Therefore, when different state quantities have inconsistent response directions to aging, a consistent degradation direction can be formed in a unified implicit representation space.

[0031] Specifically, when a candidate value for a health state in a certain stage shifts in the opposite direction to the previous stage, the previously formed implicit health state is not immediately modified. Instead, the difference between the current candidate value and the implicit health state is retained as the observed change component, and candidate values ​​for subsequent stages are read. If the subsequent candidate values ​​gradually return to the vicinity of the trajectory corresponding to the original main degradation direction, the previous reverse shift is maintained at the observation layer and gradually exits the health state recursive input as the local state recovery process proceeds. If the subsequent candidate values ​​do not return to the original trajectory, and the new state after the reverse shift continues to form a stable evolution direction, the previous reverse change is regarded as the starting point for the change in the degradation trajectory structure. The main degradation direction is re-determined based on the new continuous segment. At the same time, the existing implicit health state near the changed position is used as the connection anchor point, so that the updated implicit health state sequence maintains the continuity of the stage order without generating unfounded lifetime regression. Through this processing, the parameter recovery caused by short-term changes in operating conditions and the situation where the long-term state trajectory has changed in reality can be distinguished in the same time-series representation space without setting a fixed recovery threshold for each observation.

[0032] Further, in a preferred embodiment, the decomposition of state displacement includes: constructing a local degradation trajectory using temporal state representations formed by multiple historical local running segments, and determining the degradation tangential representation of the current stage from the local degradation trajectory; projecting the state change of the current local running segment relative to the previous local running segment onto the degradation tangential representation and its remaining state subspace to form a propulsion component and a disturbance component, respectively; updating the degradation tangential representation based on the degree of directional consistency of the propulsion component in consecutive local running segments, and restricting state changes with frequent directional reversals and without forming a continuous displacement to the disturbance component, so that the stage discrimination sequence progresses segment by segment according to the self-updated local degradation trajectory. The local degradation trajectory is formed by connecting a set of consecutive temporal state representations before the current segment in chronological order, and the current degradation tangential representation is obtained by directional normalization and historical weighting of adjacent state displacements. Among them, state displacements that are closer to the current segment and have the same direction as the longer historical trajectory receive a higher trajectory contribution, while state displacements that appear briefly and then immediately reverse are reduced in contribution after continuity judgment. This allows the degradation tangential representation to slowly turn with the long-term degradation path without violently rotating with a single local abnormal segment.

[0033] During the state subspace decomposition process, the current state change is first orthogonally projected along the degenerate tangential representation. The part along the tangential direction constitutes the advancing component, and the remaining part forms the perturbation component. If the advancing components of multiple consecutive local running segments have the same sign and the direction difference remains within the range that the current trajectory can connect, then these consecutive advancing components are used to re-estimate the degenerate tangential representation of the next stage, so that the local degenerate trajectory can smoothly transition from the previous degenerate direction to the new continuous direction. If the advancing component frequently changes sign between adjacent segments, and the cumulative state displacement returns to the vicinity of the original local trajectory after passing through multiple segments, then the corresponding change does not enter the degenerate tangential representation update, but is continuously retained in the perturbation subspace. The stage discrimination process thus obtains two mutually constrained sources of information: the advancing component determines whether the stage moves forward, and the perturbation component records the degree of deviation of the current observation from the degenerate trajectory. In the subsequent health state judgment, the perturbation component can be used to confirm whether a candidate health state reverse movement occurs simultaneously with the stage advancement, avoiding the conversion of pure observation fluctuations into new degenerate directions.

[0034] In a preferred embodiment, stage domain resampling includes: for a state sequence of multiple local running segments within the same degradation stage interval, determining stage position weights based on the stage advancement rate and the continuity of state evolution of each local running segment; mapping local running segments with different natural time intervals to a unified stage node using the stage position weights; for segments lacking direct observation between adjacent stage nodes, generating an intermediate stage representation constrained by continuity based on the temporal state representation of the preceding and following stage nodes, the stage advancement rate, and the degradation direction, and combining the state representation corresponding to the original observation with the intermediate stage representation to form a stage domain temporal sequence. The stage position weights are jointly determined by the distance between the degradation stage position of the local running segment and the target unified stage node, and the consistency relationship between the state direction of the segment before and after and the degradation direction of the target segment. When two original segments are close to the same target stage node, segments with more continuous stage advancement and whose preceding and following states can be connected along the main degradation direction receive higher position weights. When a segment is close in stage position but belongs to a short-term disturbance segment, its corresponding disturbance component increases and decreases its contribution to the state representation of the unified stage node.

[0035] For unified stage nodes with direct original observations, local running segments falling within the neighborhood of that node are combined according to stage position weights. The combination process preserves the source marker and natural time position of each segment so that the actual time span corresponding to the stage change can be traced when calculating the stage advancement rate. For segments between two adjacent unified stage nodes without any original observations, the state is not directly generated by linear interpolation over natural time. Instead, it starts from the state representation of the previous stage node and gradually moves along the local degradation direction jointly determined by the previous and next stage nodes. The state change amplitude of the intermediate position is adjusted according to the advancement rate of the previous and next stages so that the intermediate stage representation is on a trajectory in the representation space that can be continuously reached from the previous node to the next node. When the degradation direction corresponding to the next node has changed continuously, the intermediate stage representation adopts a gradually turning trajectory instead of forcibly maintaining the previous direction. This stage domain completion method forms a continuous input sequence while preserving the distinguishing markers between the original observation node and the inferred intermediate node, so that the subsequent health state recursion can use different confidence constraints for direct observation and continuous completion.

[0036] Further, in a preferred embodiment, the determination of the observed change component includes: for the degradation stage in the implicit health state sequence where a reverse change occurs, extracting the state representations of multiple stages before and after the degradation stage, and constructing a baseline before the reverse change and a baseline after the reverse change, respectively; comparing whether the state representation after the reverse change regresses to the baseline before the reverse change, and simultaneously determining whether the main degradation direction of the subsequent stage maintains the original direction; including state changes that satisfy the baseline regression relationship and do not change the subsequent main degradation direction into the observed change component, and including state changes that continuously deviate from the baseline before the reverse change and form a new degradation trajectory into the implicit health state update. The baseline before the reverse change is formed by weighting the state representations of multiple consecutive unified stage nodes before the reverse change according to the stage distance. The closer the node is to the reverse change position, the higher its contribution to the current baseline. At the same time, the degradation direction formed in the earlier stage is retained as the baseline direction constraint. The baseline after the reverse change is updated step by step using the consecutive stage nodes obtained after the reverse change. Therefore, as subsequent observations increase, it is possible to continuously determine whether the current state regresses to the old trajectory or forms a new continuous trajectory.

[0037] When determining the baseline regression relationship, each state representation after the reverse change is compared with the baseline before the reverse change and the current trajectory after the reverse change. If the distance between the continuous state representation and the old baseline trajectory gradually decreases, and its stage advancement component remains positive along the main degradation direction before the reverse change, then the state difference corresponding to the reverse change only enters the observed change component, and the implicit healthy state continues to use the previous stable stage as the recursive benchmark. If the distance between the continuous state representation and the old baseline trajectory continues to increase, and a stable same-direction displacement is formed between the new states, then the new trajectory is regarded as a continuous change that needs to enter the healthy state update, and the degradation stage is kept at the trajectory connection point without regression. This avoids directly taking a state recovery with a large amplitude as a lifetime recovery, and also avoids permanently excluding state changes that have indeed formed long-term trajectory changes from the healthy state. The observed change component retains its duration and recovery path in the subsequent time-series coding, so that when the same short-term change reappears, it can be compared with the existing operational disturbance mode.

[0038] In a preferred embodiment, updating the implicit health state sequence includes: establishing a stage-conditional association between degradation stage variables and candidate health state values; recording the state difference between each degradation stage and the preceding health state, as well as the corresponding stage span; forming a stage consistency record based on the direction of the state differences of multiple consecutive stages, and writing the state differences that satisfy the main degradation direction into the health state recursion process according to the stage span; for degradation stages with observed change components, maintaining the preceding implicit health state as the recursion benchmark, and only incorporating the continuous state changes that subsequently re-enter the main degradation direction into the recursion result, forming an implicit health state sequence consistent with the unidirectional progression relationship of degradation stages; the stage-conditional association uses the degradation stage position as the conditional input, so that the same magnitude of state change can be mapped to different health state changes in the early, middle, and near-terminal stages of life, and training this association through the sequential order of each stage in the complete life sample, avoiding processing the state differences of all stages according to a uniform scale.

[0039] Specifically, each unified stage node generates a candidate health state value and a candidate state difference relative to the previous stable health state. The stage span is determined by the difference in degradation stage variables between the current node and the previous stable node. When the direction of the candidate state difference is consistent with the main degradation direction formed by multiple consecutive historical nodes, the candidate state difference is normalized according to the stage span and written into the implicit health state for recursion. When the candidate state difference is opposite to the main degradation direction, the current implicit health state remains unchanged and the candidate state difference is stored in the observation change record. After subsequent nodes continue to advance along the main degradation direction, new degradation segments are connected starting from the previous stable implicit health state. If subsequent nodes form a new continuous degradation direction, the main degradation direction is updated after the trajectory change is confirmed by the aforementioned baseline judgment and recursion continues from the connection anchor point. The entire health state sequence thus maintains the same order attribute as the degradation stage variables. During training, the health states near the end of the lifespan are constrained to enter the unified termination interval through the complete lifespan sequence, and the current health state is constrained not to be mapped to the termination position that has already exceeded its actual running stage through the running sequence that has not yet expired.

[0040] Furthermore, in a preferred embodiment, the stage advancement rate is jointly determined by the degradation stage change and the corresponding natural time span. The stage advancement rate describes the degree to which the degradation stage moves forward within a unit of natural operating time. The same power capacitor can form a continuously changing and relatively slow advancement rate in the long-term stable operating range. After the continuous change in operating conditions causes the actual degradation path to accelerate, a new local advancement segment is allowed to form. The stage advancement rate itself does not participate in the determination of whether the degradation stage can be reversed. Instead, it is responsible for retaining the natural time scale information after the degradation stage has been determined by the advancement component. Therefore, the stage domain alignment will not lose the differences in the aging rate of different devices. During the model training process, the same unified degradation stage node can aggregate the state patterns of different power capacitors, and the natural time experienced by each sample from the previous stage node to the current node is still stored in the corresponding advancement rate. This allows the subsequent remaining lifetime calculation to determine the current degree of degradation based on the unified stage position and to determine the natural time required to complete the remaining stage distance based on the current device's own advancement rate.

[0041] In a preferred embodiment, the determination of the stage advancement rate includes: forming a basic advancement rate based on the natural time span and degradation stage span corresponding to adjacent stage nodes in the stage domain time sequence, and using the aforementioned advancement component continuity as a rate reliability constraint; when the basic advancement rate of a continuous stage node changes abruptly, comparing whether the corresponding advancement component synchronously forms a continuous change; if no continuous change is formed, the basic advancement rate is continuously corrected along the preceding degradation tangential representation; if the advancement component changes synchronously with the basic advancement rate, a new local advancement segment is established, and the subsequent stage advancement rate sequence is re-formed within the local advancement segment. The basic advancement rate is obtained by dividing the stage span between two stage nodes with direct observation sources by the corresponding natural time span. For nodes formed by completing intermediate stage representations, their advancement rate is jointly estimated by the basic advancement rate of the preceding and following directly observed nodes and the continuity of the local degradation direction, without using the completion node alone to generate a new rate abrupt change.

[0042] When the base propulsion rate changes significantly, the propulsion component sequence and the perturbation component sequence of the same segment are read for consistency judgment. If the propulsion rate suddenly increases but the state displacement mainly falls in the perturbation subspace, or if the propulsion component increases only in a single segment and then recovers quickly, the original local propulsion segment is maintained and the current rate is continuously connected according to the previous tangential direction to avoid short-term observed changes being converted into long-term lifetime consumption acceleration. If the propulsion component maintains a large positive displacement in multiple consecutive segments, and the local degradation tangential indicates the formation of a stable new direction, a new local propulsion segment is established at the starting point of the change. The new segment re-estimates the propulsion rate using its own natural time span and stage span, while the rate record of the old segment remains unchanged, forming a segmented continuous propulsion rate sequence. Subsequent predictions can select the corresponding rate evolution state based on the current local propulsion segment and combine the rate changes of the most recent stages to determine whether the current segment should continue or gradually revert to the historical stable rate in the next few stages.

[0043] In a preferred embodiment, please refer to Figure 4 The determination of remaining lifetime includes: constructing a lifetime time series representation based on the implicit health state corresponding to the current degradation stage, the health state change sequence formed by multiple consecutive stages, and the stage advancement rate sequence determined in the aforementioned manner; recursively pushing the degradation stage along the current local advancement segment, while maintaining the implicit health state evolving along the aforementioned main degradation direction during the recursion process; when the recursed health state reaches the lifetime termination state interval formed by the corresponding degradation termination stage in the historical failure time series, determining the remaining stage distance between the current degradation stage and the lifetime termination state interval, and converting the remaining stage distance into the remaining lifetime of the power capacitor based on the stage advancement rate sequence. The lifetime termination state interval is jointly determined by the implicit health states formed by multiple consecutive stages before failure of the power capacitors whose complete failure process has been recorded in the training data. Instead of using a fixed threshold of a single original observation as the termination criterion, a termination state region that can accommodate normal individual differences is formed within a unified stage coordinate and implicit health representation space.

[0044] In this embodiment, the period from the current stage to the end of the lifespan is divided into several adjacent stage segments, and the stage distance is converted into a natural time distance using the advancement rate corresponding to each stage segment. One feasible calculation method is as follows: ; in, This represents the remaining lifetime at the current moment, k represents the index of the current stage segment, and K represents the index of the termination stage corresponding to the lifetime termination state segment. This represents the span of the degradation stages in the j-stage segment. This represents the stage advancement rate corresponding to the j-th stage segment. This means that the natural time required for each segment between the current stage and the final stage is accumulated sequentially. In a specific calculation scenario, if the current local advancement segment remains stable, then the time required for subsequent adjacent segments... Based on the current rate sequence, continuous extrapolation can be performed. If changes in health status and the advancing component jointly indicate entry into a new segment of continuous degradation, then a re-estimated rate can be used starting from the new segment. When the advancement rate of a certain segment is small and the stage span is fixed, the natural time distance corresponding to that segment is relatively long. When the advancement rate continues to increase, the natural time distance corresponding to the same stage span is shortened. Therefore, the remaining lifetime is determined by the remaining degradation distance and the current degradation rate, rather than by regression of the current health state alone.

[0045] Please refer to Figure 5 , Figure 5 The horizontal axis represents the remaining stage segment from the current degradation position to the end-of-life state, divided into the starting segment, preceding segment, subsequent segment, and final segment; the vertical axis represents the natural operating time contribution calculated for each segment. The height of each bar corresponds to the stage segment's contribution to... The contribution of the dashed line is the short-term maintenance reference line, used to observe which degradation stage segments mainly dominate the remaining life results.

[0046] from Figure 5 It can be seen that the contribution of remaining lifetime is not uniformly distributed across different segments, with the highest value observed in the final segment. This indicates that the segment span is relatively large in the intervals approaching the termination state. With propulsion rate These factors collectively determine the significant time contribution. If a segment has a lower column height, it indicates that the distance of the degradation stage to be traversed in that segment is smaller, or that the advancement rate of that segment is faster, thus impacting the total remaining lifetime. The impact is relatively limited.

[0047] Figure 5 Each bar is not an isolated empirical value, but is formed by the cumulative contribution of multiple segments from the current stage index k to the final stage index K. Reflects the distance to be overcome in the degradation stage. It reflects the stage-by-stage progress speed within a unit of natural operating time; both together determine the converted time for a single segment, and are ultimately summarized as follows: .

[0048] During lifetime regression, the current implicit health state serves as the starting point. A sequence of health state changes formed by multiple consecutive stages is used to determine the current main degradation direction and range of change. The stage advancement rate sequence is used to determine the natural time corresponding to each stage advance. Each time the regression state enters a new unified stage node, a candidate value for the next health state is generated according to the current main degradation direction. The evolution trajectory of the health state adjacent to the current stage position in the historical complete lifetime sequence is used to limit the continuous range of the candidate value. When the candidate state reaches the lifetime termination state interval, the stage regression stops. If the current state is still far from the termination interval, the subsequent stages are generated based on the local advancement segment. If the current device has recently formed a new continuous degradation direction and a new advancement rate segment, the future regression will preferentially continue along the new segment. As the regression distance increases, the constraints of the historical similar stage trajectory on the subsequent state are gradually increased to avoid infinitely extending sudden changes based solely on the most recent state segment. The remaining lifetime obtained also retains the time contribution of each stage segment. The degradation stage position, the local advancement segment, and the stage distance from the termination state interval corresponding to the current result can be further output as intermediate state records for lifetime calculation.

[0049] In a complete training process, training samples are organized according to the temporal continuity of a single power capacitor, without randomly shuffling the segment order within the same lifetime sequence. The complete lifetime sequence is used to establish the stage sequence from the initial operating state to the end-of-life state. Operating sequences that have not reached the end-of-life state are used to constrain the observed stage positions from exceeding the actual supported end-of-life range. The training process first enables the time-series coding network to generate continuous state representations based on local operating segments, and then forms degradation stage variables based on state displacements and historical trajectories. Simultaneously, the conditional association between healthy state candidate values ​​and degradation stages is trained on a unified stage coordinate system. Stage sequence constraints require that the degradation stage variables of the later local operating segment are not less than those of the previous local operating segment. Trajectory continuity constraints require that the direction of the propulsion component in the same continuously degraded segment remain connectable. Observation change constraints require that reverse changes identified as short-term recovery cannot directly change the implicit healthy state. End-of-life constraints require that the implicit healthy state at the end of the complete failure sequence enter the common termination state interval. All constraints work together on the same time-series representation process, forming a continuous data dependency relationship between degradation stages, implicit healthy states, and remaining lifetime.

[0050] Preferably, to avoid the natural time length differences of different training sequences from dominating the stage learning again, samples are selected according to the coverage of the degradation stage when forming training batches. This ensures that a training batch simultaneously includes local running segments located in the early, middle, and near-end-of-life regions, while maintaining that each segment can be traced back to its previous historical state. The scaling transformation of the temporal state representation only uses the statistics already observed in the training data, and the same transformation rules are fixedly applied to subsequent prediction data. The local degradation tangential representation is formed by several initial continuous state displacements at the beginning of a new lifetime sequence. In the initial interval where the stable tangential direction is not yet determined, only the state trajectory is accumulated without large-scale stage advancement. After the continuous displacements form a consistent direction, the normal stage is recursively advanced. This avoids the small amount of random fluctuations in the early stage of operation from directly determining the subsequent degradation coordinate direction. As the historical segments increase, the early tangential representation is gradually updated by the subsequent continuous degradation trajectory, but the already formed degradation stage order is not written back and modified.

[0051] During the prediction process, the continuous state time series data of the power capacitor to be predicted up to the current moment is converted into local operating segments according to the same data organization rules as the training phase, and time series state representations are generated sequentially. The prediction process updates the historical state change trajectory, cumulative operating context, degradation tangential representation, and degradation stage variables one by one in chronological order, without waiting for the formation of a complete lifetime sequence. When a new local operating segment arrives, the advancing component and the disturbance component are determined according to the relationship between the current state displacement and the existing degradation tangential representation. The updated degradation stage is then mapped to a unified stage coordinate to determine the corresponding health state candidate value. If the candidate value advances along the main degradation direction, the implicit health state and stage advancement rate are updated simultaneously. If the candidate value changes in the opposite direction, the observation change discrimination is entered and the previous stable implicit health state is maintained. As subsequent local operating segments continue to be input, the baseline regression relationship and the new trajectory are continuously updated. After confirmation, it is decided whether the reverse change enters the new health state trajectory. In this way, each lifetime update in the online prediction can be traced back to the current stage position, state advancement direction, and natural time span.

[0052] In a further implementation, when different power capacitors have different initial state baselines, the absolute value of the original state is not directly used as the starting point of the degradation stage. Instead, the initial stable segment after each capacitor is put into continuous operation is used to form an individual initial state representation. Then, the continuous state displacement relative to the individual initial state representation is used to participate in the degradation trajectory calculation. The starting region in the unified stage coordinate is jointly defined by the initial state representations of multiple lifetime sequences and their initial continuous change direction. Therefore, it is allowed that the absolute values ​​of the original states of different individuals are different. At the same time, it is ensured that the variables of the subsequent stages describe the cumulative degradation progress relative to their own initial states. For data that is recorded in the middle and lacks a complete initial stage, the initial mapping interval can be determined according to the relationship between its current time-series state representation and the state trajectory of each unified stage node in the training set. This mapping interval is used as the starting point for the recursion of the subsequent stages. After that, it is still continuously updated according to the same calculation logic of the advancement component, disturbance component, irreversible health state and stage advancement rate. The lack of initial history is not directly interpreted as a lower or higher lifetime state.

[0053] In another preferred embodiment, when a long data gap occurs in the running sequence, the local running segments before and after the gap are not directly used as consecutive adjacent segments to calculate the state displacement. Instead, the natural time information of the gap's duration is retained, and the last stable state representation before the gap and the earliest stable state representation after the gap are calculated separately. If the state change between the two can form a consistent connection along the degradation tangential direction before the gap and the degradation direction of the consecutive segments after the gap, the gap segment is treated as a stage interval that has not been directly observed. The stage span is estimated based on the stage advancement state at both ends of the gap and the degradation direction formed by subsequent observations. Then, the intermediate stage representation is generated through the continuity of the stage domain. If the state after the gap and the degradation trajectory before the gap do not have a continuous connection relationship, the state difference at both ends is not written into the degradation stage for the time being. Instead, multiple local running segments after the gap form a new stable trajectory, and the stage update method is determined based on the connection position between the new and old trajectories. This avoids the single large-span state difference caused by data loss being directly interpreted as continuous lifetime consumption.

[0054] Specifically, the temporal coding, stage discrimination, health status mapping, and lifetime recursion in the aforementioned embodiments adopt a shared data index relationship. Each unified stage node stores the original local running segment index, natural time position, degradation stage variables, advancement components, disturbance components, implicit health status, observed change components, and stage advancement rate. The index relationship allows any remaining lifetime result to be traced back to the state segment that participated in the current lifetime recursion, while ensuring that different calculation processes do not lose their temporal correspondence due to reordering. When a new local running segment enters the state sequence, it is only necessary to continue updating from the most recent stable stage node, without having to recalculate the entire historical sequence. Historically determined degradation stages, implicit health statuses, and local advancement segments are retained as continuous states. New data only updates the current segment and subsequent prediction parts that have not yet been determined. This forms a recursive processing procedure suitable for continuous lifetime prediction, and maintains the same set of data relationships for the definition of degradation stages, the discrimination of observed changes, and the lifetime termination state intervals in the training and prediction stages.

[0055] In this embodiment, before outputting the remaining lifetime, the model checks whether the current state has sufficient continuous history. The basis for this check is not setting a fixed hardware threshold related to the device structure, but rather judging whether a continuous state trajectory that can determine the local degradation tangential representation has been formed before the current degradation stage, whether the current health state has a traceable stable recursive benchmark, and whether the stage advancement rate has at least one connectable local advancement segment. When historical information is insufficient, the current local running segment still participates in the temporal state representation and trajectory accumulation, but the remaining lifetime uses the historical stage advancement distribution corresponding to the current unified stage position as a temporary conversion basis. As the number of continuous running segments increases, the advancement rate formed by the current power capacitor itself gradually replaces the historical conversion basis. When a stable local degradation trajectory has been formed, the lifetime result is completely recursively calculated and time-converted according to the current implicit health state, the health state change sequence, and the stage advancement rate, so that the entire temporal learning process maintains a continuous computational closed loop from data entry, stage formation, state constraints to lifetime output.

Claims

1. A method for predicting the lifetime of power capacitors based on time-series learning, characterized in that, include: Acquire multivariable state time-series data generated by the continuous operation of power capacitors, and generate state sequences according to time correlation relationships; Based on the state sequence, extract the temporal state representation of the local running segment; Based on the state evolution direction, historical state change trajectory and cumulative operational context between adjacent local operational segments, the degradation stage variables corresponding to each local operational segment are determined; Based on the aforementioned degradation stage variables, local runtime segments at different degradation rates are aligned according to their degradation stages. Based on the aligned temporal state representation, a sequence of implicit healthy states that progresses along the degradation stage is constructed, and an irreversible degradation direction constraint is imposed on the sequence of implicit healthy states. The remaining lifetime of the power capacitor is determined based on the implicit health state sequence, the progression of the degradation stage, and their temporal changes.

2. The method for predicting the lifetime of power capacitors based on time-series learning according to claim 1, characterized in that, The determination of the degradation stage variables includes: forming a stage discrimination sequence for each local running segment, consisting of the current temporal state representation, the state representation of the previous local running segment, and the cumulative running context; Calculate the state displacement along the historical degradation direction between adjacent stage discrimination sequences, and decompose the state displacement into a propulsion component consistent with the existing degradation trajectory and a disturbance component deviating from the existing degradation trajectory; The candidate degradation stage of the current local running segment is updated based on the propagation component, and the candidate degradation stage is continuously corrected with the historically determined degradation stage as the order constraint to obtain the degradation stage variable that advances unidirectionally with the running sequence.

3. The method for predicting the lifetime of power capacitors based on time-series learning according to claim 1, characterized in that, The degradation stage alignment includes: constructing non-uniform stage coordinates using degradation stage variables corresponding to each power capacitor, and mapping local running segments on the natural time axis to the non-uniform stage coordinates; For local running segments that are in adjacent degradation stage intervals but have inconsistent sampling times, a stage correspondence is established based on the evolutionary continuity between the temporal state representations; According to the stage correspondence, the state representation of the local running segment is resampled in the stage domain to form an aligned state sequence with a unified degenerate stage position, and the stage advancement rate formed when each local running segment is mapped from natural time to the degenerate stage is retained.

4. The power capacitor lifetime prediction method based on time-series learning according to claim 3, characterized in that, The irreversible degradation direction constraint includes: generating healthy state candidate values ​​based on the temporal state representations corresponding to adjacent degradation stages in the aligned state sequence, and determining the main degradation direction using the healthy state candidate values ​​of multiple consecutive degradation stages; The changes in health status along the main degradation direction are written into the implicit health status sequence, and the changes in status that are opposite to the main degradation direction and do not continuously cross multiple subsequent degradation stages are retained in the observed change components. When the reverse state change is maintained in the continuous degradation phase and changes the evolution trajectory of the subsequent healthy state, the main degradation direction of the corresponding phase is re-determined and the implicit healthy state sequence is updated.

5. The method for predicting the lifetime of power capacitors based on time-series learning according to claim 2, characterized in that, The decomposition of the state displacement includes: constructing a local degradation trajectory using a temporal state representation formed by multiple historical local running segments, and determining the degradation tangential representation of the current stage using the local degradation trajectory; The state change of the current local running segment relative to the previous local running segment is projected onto the degenerate tangential representation and its remaining state subspace to form the propulsion component and the disturbance component, respectively. The degenerate tangential representation is updated based on the degree of consistency of the direction of the propulsion components in the continuous local operation segments, and the state changes that frequently reverse direction and do not form a continuous displacement are restricted to the perturbation components, so that the stage discrimination sequence progresses segment by segment according to the self-updating local degradation trajectory.

6. The method for predicting the lifetime of power capacitors based on time-series learning according to claim 3, characterized in that, The stage domain resampling includes: for the state sequence of multiple local running segments within the same degradation stage interval, determining the stage position weight based on the stage advancement rate and the degree of continuity of state evolution of each local running segment; The stage position weights are used to map local running segments with different natural time intervals to a unified stage node; For segments where there is a lack of direct observation between adjacent stage nodes, an intermediate stage representation constrained by continuity is generated based on the temporal state representation of the preceding and following stage nodes, the stage advancement rate, and the degradation direction. The state representation corresponding to the original observation is then combined with the intermediate stage representation to form a stage domain temporal sequence.

7. The power capacitor lifetime prediction method based on time-series learning according to claim 4, characterized in that, The determination of the observed change components includes: for the degradation stage in the implicit health state sequence where a reverse change occurs, extracting the state representations of multiple stages before and after the degradation stage, and constructing the baseline before the reverse change and the baseline after the reverse change, respectively; The state after the reverse change is compared to indicate whether it reverts to the baseline before the reverse change, and it is simultaneously determined whether the main degradation direction in the subsequent stage remains the original direction; State changes that satisfy the baseline regression relationship and do not change the subsequent main degradation direction are included in the observed change component, and state changes that continuously deviate from the baseline before the reverse change and form a new degradation trajectory are included in the implicit health status update.

8. The power capacitor lifetime prediction method based on time-series learning according to claim 7, characterized in that, The update of the implicit health state sequence includes: establishing a stage condition association between the degradation stage variable and the health state candidate value, and recording the state difference with the previous health state and the corresponding stage span for each degradation stage; A stage consistency record is formed based on the state difference direction of multiple consecutive stages, and the state difference that satisfies the main degradation direction is written into the health state recursion process according to the stage span. For a degradation stage with the observed change components, the preceding implicit healthy state is kept as the recursive benchmark, and only the continuous state changes that subsequently re-enter the main degradation direction are incorporated into the recursive result, forming an implicit healthy state sequence consistent with the unidirectional progression relationship of the degradation stage.

9. The method for predicting the lifetime of power capacitors based on time-series learning according to claim 6, characterized in that, The determination of the stage advancement rate includes: forming a basic advancement rate based on the natural time span and degradation stage span corresponding to adjacent stage nodes in the stage domain time sequence, and using the continuity of the formed advancement component as a rate reliability constraint; When the base propulsion rate of a continuous stage node changes abruptly, it is compared whether the corresponding propulsion component synchronously forms a continuous change. If no continuous change is formed, the base propulsion rate is continuously corrected along the preceding degenerate tangent. If the propulsion component changes synchronously with the basic propulsion rate, a new local propulsion segment is established, and the subsequent stage propulsion rate sequence is re-formed within the local propulsion segment.

10. The method for predicting the lifetime of power capacitors based on time-series learning according to claim 1, characterized in that, The determination of the remaining lifetime includes: constructing a lifetime time series representation based on the implicit health state corresponding to the current degradation stage, the health state change sequence formed by multiple consecutive stages, and the determined stage advancement rate sequence; The degradation stage is pushed forward along the current local advancement segment, and the implicit healthy state is maintained to evolve along the defined main degradation direction during the pushing process; When the recursive health state reaches the lifetime end state interval formed by the corresponding degradation termination stage in the historical failure sequence, the remaining stage distance between the current degradation stage and the lifetime end state interval is determined, and the remaining stage distance is converted into the remaining lifetime of the power capacitor according to the stage advancement rate sequence.

Images