A method for evaluating life of ox horn capacitor based on attention mechanism

Abstract

This invention discloses a method for assessing the lifespan of horn capacitors based on an attention mechanism, relating to the field of capacitor lifespan assessment technology. The method includes the following steps: collecting and standardizing horn capacitor lifespan assessment data to generate a horn capacitor lifespan assessment dataset; constructing a multi-domain state sequence of horn capacitors according to their number, acquisition time, and operating conditions; generating a horn capacitor operating condition calibration deviation sequence based on rated parameters, temperature derating relationships, and ripple load relationships; identifying degradation initiation points, degradation duration segments, and degradation inflection points based on deviation changes; generating an attention-based degradation characterization of the horn capacitors using an operating condition-gated DA-RNN model; constructing a degradation monotonic constraint lifespan evolution trajectory to generate horn capacitor lifespan prediction results; determining the failure risk level and matching maintenance trigger conditions to generate horn capacitor lifespan assessment results. This invention can improve the accuracy of degradation identification, lifespan prediction, and maintenance decision-making.

Description

Technical Field

[0001] This invention relates to the field of capacitor life assessment technology, and in particular to a method for assessing the life of horn capacitors based on an attention mechanism. Background Technology

[0002] In recent years, with the continuous development of power electronic equipment, high-frequency switching power supplies, industrial frequency converters, and energy storage systems, horn capacitors, as filtering, energy storage, and pulsating current absorption devices, have a direct impact on equipment stability due to their service life. Existing technologies typically collect data on capacitance, equivalent series resistance, leakage current, loss tangent, ripple current, and temperature. They then identify the capacitor's performance degradation state through rated threshold comparisons, trend analysis, accelerated aging tests, or machine learning models, and predict the remaining service life based on capacitance retention rate, equivalent series resistance growth rate, or historical failure records. Some solutions also incorporate load rate, ambient temperature, and operating time to refine the life prediction results, providing recommendations for maintenance or replacement.

[0003] However, existing methods for assessing the lifespan of horn capacitors still suffer from the problem of difficulty in separating the influence of operating conditions from the actual degradation state. Different temperatures, ripple loads, charge / discharge states, and switching frequencies cause fluctuations in electrical and thermal parameters. Existing solutions often directly input the collected parameters into the prediction model, lacking the processing to construct a unified reference state based on rated parameters, temperature derating relationships, and ripple load relationships. This easily leads to misjudging changes in operating conditions as performance degradation. Furthermore, existing solutions typically use fixed time windows or overall characteristics for lifespan prediction, failing to adequately utilize the differences in contributions from different degradation parameters and time windows. They also fail to characterize the degradation evolution process by combining the degradation initiation point, degradation duration, and degradation inflection point, resulting in a lack of monotonic constraints on the lifespan trajectory and inaccurate correlation between failure risk levels and maintenance triggering conditions. Summary of the Invention

[0004] One objective of this invention is to propose a life assessment method for horn capacitors based on an attention mechanism. This invention fully utilizes electrothermal coupling calibration and condition-gated DA-RNN attention modeling methods to achieve horn capacitor degradation state identification, remaining service life prediction, and failure risk classification. It has the advantages of accurate degradation feature extraction, reliable life prediction, and clear maintenance decisions.

[0005] A method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to an embodiment of the present invention includes the following steps: Collect horn capacitor life assessment data and perform standardization processing to generate horn capacitor life assessment dataset; The life assessment dataset of the horn capacitor is windowed and associated according to the horn capacitor number, acquisition time and operating conditions. Electrical state, thermal state and operating condition are combined to generate a multi-domain state sequence of the horn capacitor. Based on the rated parameters, temperature derating relationship and ripple load relationship, an electrothermal coupling operating condition reference state is constructed. The deviation between the multi-domain state sequence of the horn capacitor and the electrothermal coupling operating condition reference state is calculated to generate the horn capacitor operating condition calibration deviation sequence. Based on the deviation direction, deviation amplitude, and parameter co-change within a continuous time window, the degradation initiation point, degradation duration, and degradation inflection point are identified, generating a degradation event sequence for the horn capacitor. The calibration deviation sequence and degradation event sequence of the horn capacitor under operating conditions are input into the operating condition gated DA-RNN model to determine the operating condition gate weight, degradation parameter weight and degradation time window weight, and generate the attention degradation characterization of the horn capacitor. Based on the attention degradation characterization of the horn capacitor, a degradation monotonic constrained lifetime evolution trajectory is constructed to determine the health status, aging stage, remaining service life and lifetime decline rate, and generate horn capacitor lifetime prediction results. Based on the predicted lifespan of the horn capacitor and the degradation inflection point, the failure risk level is determined and maintenance triggering conditions are matched to generate the lifespan assessment results of the horn capacitor.

[0006] Optionally, the life assessment data of the horn capacitor includes basic parameters, electrical status data, thermal status data, operating condition data, environmental status data, and historical life data of the horn capacitor. The standardization process includes timestamp alignment, numbering unification, data format conversion, measurement unit unification, sensor deviation correction, outlier removal, missing value completion, sampling frequency unification, operating condition coding, and numerical normalization.

[0007] Optionally, the generation of the horn capacitor multi-domain state sequence includes: The dataset for assessing the lifespan of horn capacitors was grouped according to their serial numbers, and the data in each group was sorted in ascending order by the time of collection. Based on the time window length and window sliding step, continuous time windows are divided, and data with the same horn capacitor number and acquisition time within the same continuous time window are time-aligned and aggregated. The operating conditions corresponding to each continuous time window are determined based on terminal voltage, ripple current, load rate, charging and discharging status, switching frequency, ambient temperature, and ambient humidity. Adjacent continuous time windows are associated according to the order of change of operating conditions. The actual capacitance, equivalent series resistance, leakage current, loss tangent, terminal voltage and ripple current are extracted from each continuous time window and combined to form the electrical state. Extract the shell temperature, estimated internal temperature, and temperature rise rate, and combine them to form the thermal state; Extract load rate, charging and discharging status, switching frequency, cumulative running time, ambient temperature and ambient humidity, and combine them to form the operating condition status; The electrical state, thermal state, and operating state are bound together according to the time sequence of a continuous time window to generate a multi-domain state sequence of the horn capacitor.

[0008] Optionally, the generation of the bullhorn capacitor operating condition calibration deviation sequence includes: Match the rated capacitance, rated voltage, rated ripple current, rated operating temperature, initial equivalent series resistance, initial leakage current, and initial loss tangent value to the horn capacitor number; The temperature offset is determined based on the ambient temperature, shell temperature and internal estimated temperature of each continuous time window, and the rated capacitance, initial equivalent series resistance, leakage current and loss tangent are corrected for temperature based on the temperature derating relationship. The ripple load factor is determined based on the ratio of ripple current to rated ripple current, load factor and switching frequency. The load is then adjusted for terminal voltage, ripple current, temperature rise rate and internal estimated temperature in combination with the ripple load relationship. The temperature correction results and load correction results are combined according to the corresponding continuous time window to generate electrical state reference values ​​and thermal state reference values, and to construct the electrothermal coupling operating condition reference state. The electrical, thermal, and operating states of each continuous time window in the multi-domain state sequence of the horn capacitor are compared with the corresponding reference values ​​by performing difference and ratio calculations to determine the deviation direction, deviation magnitude, and deviation changes between adjacent continuous time windows for each state parameter. The horn capacitor operating condition calibration deviation sequence is generated by arranging the parameters according to the end time of the window.

[0009] Optionally, the generation of the horn capacitor degradation event sequence includes: The calibration deviation sequence of the horn capacitor is arranged according to the end time of the window, and the deviation direction, deviation amplitude and deviation change rate of the key degradation parameters in each continuous time window are extracted. By aligning key degradation parameters with consistent degradation direction over time, and by counting the number of consecutive time windows with consistent deviation direction and the number of key degradation parameters with synchronously changing deviation magnitude, the state of parameter co-change is determined. The continuous time window in which the parameter coordinated change state first continuously reaches the degradation judgment condition is determined as the degradation start point. Starting from the degradation start point, adjacent continuous time windows in which the deviation direction is consistent, the deviation amplitude is continuously accumulated, and the parameter coordinated change state is not interrupted are connected to form the degradation duration segment. The rate of change of deviation is continuously compared within the duration of degradation, and the continuous time window in which the rate of change of deviation shows a continuous increase, a continuous decrease, or a change in the direction of change is determined as the degradation inflection point. By binding the degradation start point, degradation duration, and degradation inflection point with the corresponding key degradation parameters, deviation states, and durations, a degenerative event sequence of the horn capacitor is generated.

[0010] Optionally, the generation of the bullhorn capacitor attention degradation characterization includes: The calibration deviation sequence and degradation event sequence of the horn capacitor under operating conditions are input into the operating condition-gated DA-RNN model. The operating condition-gated DA-RNN model includes a sequence construction layer, an operating condition gating layer, a parameter attention encoding layer, an event fusion layer, and a time window attention output layer. In the sequence construction layer, the deviation direction, deviation magnitude, and deviation change rate of each state parameter are time-aligned with the degradation start point, degradation duration, and degradation inflection point according to the end time of the window, generating a degradation input sequence and a degradation event marker sequence. In the working condition gating layer, the gating value is calculated based on the working condition status and corresponding parameter deviation of each continuous time window. The working condition gating weight is determined by range constraints, and the degenerate input sequence is weighted item by item. In the parameter attention encoding layer, the parameter attention score is calculated based on the deviation between the previous encoded state and the weighted parameter. The degenerate parameter weight is determined after normalization. The weighted aggregation result of each state parameter is input into the recurrent encoding unit to generate a degenerate hidden state sequence. In the event fusion layer, the degenerate event tag sequence is mapped to degenerate event features and fused with the degenerate hidden states of the same continuous time window to generate an event-enhanced state sequence. In the time window attention output layer, the time attention score is calculated based on the initial decoding state and the event enhancement state sequence. After normalization, the weight of the degradation time window is determined. The event enhancement states of each consecutive time window are weighted and aggregated to generate the horn capacitive attention degradation representation.

[0011] Optionally, the generation of the horn capacitor lifetime prediction result includes: A fully connected mapping is performed on the attention degradation characterization of the horn capacitor to generate the current health value and the lifetime decay increment corresponding to each predicted time step. The lifetime decay increment is nonnegated, and the lifetime decay increment is subtracted from the current health value sequentially according to the predicted time step. The health value of the next predicted time step is restricted to not be higher than the health value of the previous predicted time step, thus generating a degenerate monotonic constrained lifetime evolution trajectory. The current health value is matched with the health status classification threshold to determine the health status, and the aging stage is determined based on the degradation start point, degradation inflection point and the position of the current health value in the degradation monotonic constrained lifetime evolution trajectory. The predicted time step when the health value first reaches the failure threshold is retrieved along the degradation monotonic constraint lifetime evolution trajectory, and the remaining lifetime is determined based on the time interval between the current time and the corresponding predicted time. Calculate the ratio of the decrease in health value between adjacent prediction time steps to the time interval, and weight and aggregate the ratios to determine the rate of lifespan decline; By combining the health status, aging stage, remaining service life, life decline rate, and degradation monotonic constraint life evolution trajectory, the life prediction results of the horn capacitor are generated.

[0012] Optionally, the generation of the horn capacitor lifetime assessment results includes: Extract the current lifetime status and its changing trend from the lifetime prediction results of the horn capacitor, and locate the degradation inflection point closest to the current assessment time. The failure risk assessment value is calculated based on the change in the overall degradation rate before and after the degradation inflection point, the duration after the inflection point, and the remaining service life. The failure risk level is determined according to the range to which the failure risk assessment value belongs, and the failure risk level is adjusted upward when the rate of life decline continues to increase after the degradation inflection point. Match maintenance trigger conditions based on failure risk level; By binding the failure risk level, maintenance triggering conditions, and corresponding judgment criteria with the horn capacitor number and assessment time, the horn capacitor life assessment result is generated.

[0013] The beneficial effects of this invention are: This invention uses a windowed association method to correlate multi-source lifetime assessment data according to the capacitor number, acquisition time, and operating conditions. This unifies the electrical, thermal, and operational states into a multi-domain state sequence for the capacitor, enabling analysis of parameters under different acquisition frequencies, operating stages, and conditions within a continuous time window. Based on rated parameters, temperature derating relationships, and ripple load relationships, an electrothermal coupling reference state is constructed. By calculating the deviation between the actual state and the corresponding reference state, the interference of ambient temperature, ripple load, and operating condition changes on parameter fluctuations is reduced. This allows the capacitor calibration deviation sequence to more accurately reflect capacitance decay, equivalent series resistance growth, and thermal deterioration.

[0014] This invention identifies degradation initiation points, degradation durations, and degradation inflection points based on deviation direction, deviation amplitude, and coordinated parameter changes. It organizes scattered parameter anomalies into a sequence of horn-shaped capacitor degradation events with time ranges and evolutionary relationships, enhancing the ability to identify early degradation, continuous degradation, and degradation rate changes. By employing a condition-gated DA-RNN model to hierarchically weight the influence of operating conditions, state parameter contributions, and time window contributions, and integrating degradation event characteristics, it can highlight state parameters highly correlated with lifetime decay and key degradation time periods, reducing the impact of short-term operating condition fluctuations and local anomalies on lifetime prediction results.

[0015] This invention constructs a monotonic constraint lifetime evolution trajectory based on the attention degradation characterization of horn capacitors, ensuring that the predicted health value remains non-incremental over operating time. Based on this, the health status, aging stage, remaining service life, and lifetime decline rate are determined, improving the consistency between lifetime prediction results and the actual aging patterns of horn capacitors. By combining degradation inflection points and lifetime change trends to determine the failure risk level and matching maintenance trigger conditions corresponding to inspection cycles, re-inspections, replacement preparations, and shutdown procedures, the lifetime prediction results can be transformed into clear maintenance decisions, providing a basis for horn capacitor condition monitoring, maintenance scheduling, and failure risk control. Attached Figure Description

[0016] 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 an overall flowchart of a horn capacitor lifetime assessment method based on an attention mechanism proposed in this invention; Figure 2 This is a schematic diagram illustrating the generation of attention degradation characterization of a horn capacitor, based on an attention mechanism, as proposed in this invention for evaluating the lifetime of horn capacitors. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-2 A method for evaluating the lifetime of a horn capacitor based on an attention mechanism includes the following steps: Collect horn capacitor life assessment data and perform standardization processing to generate horn capacitor life assessment dataset; The life assessment dataset of the horn capacitor is windowed and associated according to the horn capacitor number, acquisition time and operating conditions. Electrical state, thermal state and operating condition are combined to generate a multi-domain state sequence of the horn capacitor. Based on the rated parameters, temperature derating relationship and ripple load relationship, an electrothermal coupling operating condition reference state is constructed. The deviation between the multi-domain state sequence of the horn capacitor and the electrothermal coupling operating condition reference state is calculated to generate the horn capacitor operating condition calibration deviation sequence. Based on the deviation direction, deviation amplitude, and parameter co-change within a continuous time window, the degradation initiation point, degradation duration, and degradation inflection point are identified, generating a degradation event sequence for the horn capacitor. The calibration deviation sequence and degradation event sequence of the horn capacitor under operating conditions are input into the operating condition gated DA-RNN model to determine the operating condition gate weight, degradation parameter weight and degradation time window weight, and generate the attention degradation characterization of the horn capacitor. Based on the attention degradation characterization of the horn capacitor, a degradation monotonic constrained lifetime evolution trajectory is constructed to determine the health status, aging stage, remaining service life and lifetime decline rate, and generate horn capacitor lifetime prediction results. Based on the predicted lifespan of the horn capacitor and the degradation inflection point, the failure risk level is determined and maintenance triggering conditions are matched to generate the lifespan assessment results of the horn capacitor.

[0019] In this embodiment, the life assessment data of the horn capacitor includes basic parameters of the horn capacitor, electrical status data, thermal status data, operating condition data, environmental status data, and historical life data. The standardization process includes timestamp alignment, numbering unification, data format conversion, measurement unit unification, sensor deviation correction, outlier removal, missing value completion, sampling frequency unification, operating condition coding, and numerical normalization.

[0020] In this embodiment, the generation of the multi-domain state sequence of the horn capacitor includes: The dataset for assessing the lifespan of horn capacitors was grouped according to their serial numbers, and the data in each group was sorted in ascending order by the time of collection. Continuous time windows are divided based on the time window length and window sliding step size. Data with the same capacitive capacitor number and acquisition time falling within the same continuous time window are time-aligned and aggregated. Specifically, for each capacitive capacitor number, the corresponding data is arranged in ascending order of acquisition time. The earliest acquisition time is used as the start time of the first continuous time window. The continuous time window length is set to 30 minutes, and the window sliding step size is set to 10 minutes, generating continuous time windows with window numbers, window start times, and window end times. Data acquired between the start and end times of the corresponding window are grouped into the corresponding continuous time window, with 1 minute as the uniform time interval. Time alignment nodes are generated within a continuous time window. Various electrical status data, thermal status data, and operating condition data are matched to the time alignment node with the smallest time difference. Data with a time difference of no more than 30 seconds are bound to the corresponding time alignment node. When multiple continuous data exist in the same time alignment node, the average value is calculated. When multiple status data exist, the status value with the longest duration is retained. For the time-aligned data, the window average, maximum value, minimum value, standard deviation, first and last difference, and change slope are calculated according to the data category. The duration of the status and the number of status transitions are also counted. The calculation results are bound according to the horn capacitor number and the window number to generate the window status record corresponding to the continuous time window. The operating conditions corresponding to each continuous time window are determined based on terminal voltage, ripple current, load rate, state of charge / discharge, switching frequency, ambient temperature, and ambient humidity. Adjacent continuous time windows are associated according to the order of change in operating conditions. Specifically, for each continuous time window, the ratio of terminal voltage to rated voltage and the ratio of ripple current to rated ripple current are calculated, and the load rate, state of charge / discharge, switching frequency, ambient temperature, and ambient humidity are read. The terminal voltage ratio is divided into four intervals: less than 0.85, 0.85 to 0.95, 0.95 to 1.05, and greater than 1.05; the ripple current ratio is divided into four intervals: less than 0.50, 0.50 to 0.80, 0.80 to 1.00, and greater than 1.00; and the load rate is divided into four intervals: less than 30%, 30% to 60%, and 60% to 85%. The system identifies four intervals: 1) the maximum and 85% of the operating conditions. Based on the duration of the charging, discharging, and holding states within a continuous time window, the state with the longest duration is determined as the charging / discharging state. Corresponding state codes are generated according to the frequency range of the switching frequency, the temperature range of the ambient temperature, and the humidity range of the ambient humidity. The terminal voltage range, ripple current range, load rate range, charging / discharging state, switching frequency state code, ambient temperature state code, and ambient humidity state code are combined sequentially to determine the operating conditions corresponding to each continuous time window. The continuous time windows are arranged in ascending order according to their end time. The operating conditions of adjacent continuous time windows are compared. If the operating conditions are consistent, a continuous operating condition association is established. If the operating conditions are inconsistent, a operating condition transition association is established based on the operating conditions before and after the change. The actual capacitance, equivalent series resistance, leakage current, loss tangent, terminal voltage and ripple current are extracted from each continuous time window and combined to form the electrical state. Extract the shell temperature, estimated internal temperature, and temperature rise rate, and combine them to form the thermal state; Extract load rate, charging and discharging status, switching frequency, cumulative running time, ambient temperature and ambient humidity, and combine them to form the operating condition status; The electrical state, thermal state, and operating state are bound together according to the time sequence of a continuous time window to generate a multi-domain state sequence of the horn capacitor.

[0021] In this embodiment, the generation of the horn capacitor operating condition calibration deviation sequence includes: Match the rated capacitance, rated voltage, rated ripple current, rated operating temperature, initial equivalent series resistance, initial leakage current, and initial loss tangent value to the horn capacitor number; The temperature offset is determined based on the ambient temperature, shell temperature and internal estimated temperature of each continuous time window, and the rated capacitance, initial equivalent series resistance, leakage current and loss tangent are corrected for temperature based on the temperature derating relationship. The determination and correction of temperature offset are as follows: For each continuous time window, the average values ​​of ambient temperature, shell temperature, and internal estimated temperature are calculated. The internal estimated temperature is used as the primary temperature, the shell temperature as the check temperature, and the ambient temperature as the boundary temperature. These are then weighted and combined with weights of 0.5, 0.3, and 0.2 to obtain the comprehensive operating temperature corresponding to the continuous time window. The difference between the comprehensive operating temperature and the rated operating temperature is calculated to determine the temperature offset. A temperature derating relationship is then matched according to the temperature offset interval. For every 10°C increase in temperature offset, the rated capacitance reference value is reduced by 0.8%, and the initial equivalent series capacitance is adjusted accordingly. The reference value for resistance is increased by 6%, the reference value for leakage current is increased by 12%, and the reference value for loss tangent is increased by 4%. When the temperature offset is between adjacent temperature ranges, linear interpolation is used to determine the corresponding correction ratio. When the temperature offset is less than 0℃, the corresponding correction ratio is limited by reverse change. The reference value for rated capacitance is not higher than the rated capacitance, and the reference value for initial equivalent series resistance is not lower than the initial equivalent series resistance. The corrected rated capacitance, initial equivalent series resistance, leakage current, and loss tangent are bound according to the horn capacitor number and window number to form the temperature correction result corresponding to the continuous time window. The ripple load factor is determined based on the ratio of ripple current to rated ripple current, load factor and switching frequency. The load is then adjusted for terminal voltage, ripple current, temperature rise rate and internal estimated temperature in combination with the ripple load relationship. The determination of the ripple load factor and load correction are as follows: For each continuous time window, calculate the ratio of the average ripple current to the rated ripple current, convert the load rate to a load ratio between 0 and 1, limit the switching frequency to between 0 and 20 kHz, and divide the limited switching frequency by 20 kHz to obtain the switching frequency ratio; combine the ripple current ratio, load ratio, and switching frequency ratio with weights of 0.5, 0.3, and 0.2 respectively to obtain the ripple load factor, which is then limited to between 0 and 1.5; match the ripple load relationship according to the range to which the ripple load factor belongs, determining a low ripple load state when the ripple load factor is less than 0.5, and a medium ripple load state when it is between 0.5 and 0.8. The load condition is defined as follows: a ripple load of 0.8 to 1.0 indicates a high ripple load condition, and a ripple load of more than 1.0 indicates an over-ripple load condition. Using the rated voltage as the terminal voltage reference value, for every 0.1 increase in the ripple load factor, the terminal voltage reference value is adjusted down by 0.15% of the rated voltage, the ripple current reference value is adjusted up by 10% of the rated ripple current, the temperature rise rate reference value is increased by 0.08℃ / min, and the internal estimated temperature reference value is increased by 0.6℃. When the ripple load factor is between adjacent intervals, linear interpolation is used to determine the corresponding correction amount. The corrected terminal voltage, ripple current, temperature rise rate, and internal estimated temperature are bound according to the capacitor number and window number to form the load correction result corresponding to the continuous time window. The temperature correction results and load correction results are combined according to the corresponding continuous time window to generate electrical state reference values ​​and thermal state reference values, and to construct the electrothermal coupling operating condition reference state. The generation of electrical condition reference values ​​and thermal condition reference values ​​is as follows: The temperature-corrected rated capacitance, initial equivalent series resistance, leakage current, and loss tangent are used as the capacitance reference value, equivalent series resistance reference value, leakage current reference value, and loss tangent reference value within the corresponding continuous time window, respectively. The load-corrected terminal voltage and ripple current are used as the terminal voltage reference value and ripple current reference value, respectively, and combined according to the parameter order of the electrical condition. The load-corrected internal estimated temperature is used as the internal estimated temperature reference value, and 70% of the difference between the ambient temperature and the internal estimated temperature reference value is added to the ambient temperature. The ambient temperature is used to obtain the shell temperature reference value. The temperature rise rate after load correction is used as the temperature rise rate reference value. The parameters are combined in the order of shell temperature, internal estimated temperature and temperature rise rate. The median value of the interval to which the load rate belongs is used as the load rate reference value. The code corresponding to the longest charging and discharging state within the continuous time window is used as the charging and discharging state reference value. The median values ​​of the intervals to which the switching frequency, ambient temperature and ambient humidity belong are used as the switching frequency reference value, ambient temperature reference value and ambient humidity reference value, respectively. The cumulative running time corresponding to the end time of the continuous time window is used as the cumulative running time reference value. The electrical, thermal, and operating states of each continuous time window in the multi-domain state sequence of the horn capacitor are compared with their corresponding reference values ​​using difference and ratio calculations to determine the deviation direction, deviation magnitude, and deviation changes between adjacent continuous time windows. These are then arranged according to the end time of each window to generate a horn capacitor operating condition calibration deviation sequence. Specifically, based on the horn capacitor number and window number, the electrical, thermal, and operating states of each continuous time window in the multi-domain state sequence are aligned with their corresponding reference values ​​in the electrothermal coupling operating condition reference state. For continuous state parameters, the actual value is subtracted from the corresponding reference value to obtain the parameter difference. The parameter difference is divided by the absolute value of the corresponding reference value to obtain the parameter deviation ratio. The absolute value of the parameter deviation ratio is used as the deviation magnitude. A deviation ratio greater than 0.5% is marked as a positive deviation, less than -0.5% is marked as a negative deviation, and between -0.5% and 0.5% is marked as no significant deviation. For the charging and discharging states and their corresponding state codes, the actual state code is compared with the operating condition reference value... The state code performs difference calculation. When the difference is zero, it is marked as a consistent state. When the difference is not zero, the state transition direction and state difference are recorded. The parameter deviation ratios of the same state parameter are arranged in ascending order according to the end time of the window. The parameter deviation ratio of the current continuous time window is subtracted from the parameter deviation ratio of the previous continuous time window to obtain the deviation change between adjacent continuous time windows. The deviation change is divided by the window sliding step size to obtain the deviation change rate. The parameter difference, parameter deviation ratio, deviation direction, deviation amplitude, deviation change, and deviation change rate corresponding to each continuous time window are combined according to the parameter order of electrical state, thermal state, and operating condition state, and bound to the horn capacitor number, window number, and window end time to generate the horn capacitor operating condition calibration deviation sequence.

[0022] In this embodiment, the generation of the horn capacitor degradation event sequence includes: The calibration deviation sequence of the horn capacitor is arranged according to the end time of the window, and the deviation direction, deviation amplitude and deviation change rate of the key degradation parameters in each continuous time window are extracted. By aligning key degradation parameters with consistent degradation direction over time, and by counting the number of consecutive time windows with consistent deviation direction and the number of key degradation parameters with synchronously changing deviation magnitude, the state of parameter co-change is determined. The determination of the parameter co-change state is as follows: Based on the horn capacitor number and window number, the actual capacitance, equivalent series resistance, leakage current, loss tangent, internal estimated temperature, and temperature rise rate are correspondingly bound within the same continuous time window; a negative deviation in the actual capacitance is determined as a degradation direction, and a positive deviation in the other parameters is determined as a degradation direction. The number of continuous time windows in which each parameter maintains its degradation direction is counted. If the number of continuous time windows is not less than 3, the corresponding parameter is marked as a continuously degrading parameter; the deviation amplitude of the current continuous time window is compared with the deviation amplitude of the previous continuous time window. If the deviation amplitude increases by at least 10%, the corresponding parameter is marked as... When the deviation amplitude of the synchronization enhancement parameter changes by a percentage between -10% and 10%, the corresponding parameter is marked as a synchronization persistence parameter. The number of persistently degrading parameters, synchronization enhancement parameters, and synchronization persistence parameters within the same continuous time window are counted. A state with fewer than two persistently degrading parameters is defined as no coordination; a state with at least two persistently degrading parameters and a total of at least two synchronization enhancement and persistence parameters is defined as a coordination persistence state; and a state with at least three persistently degrading parameters and at least two synchronization enhancement parameters is defined as a coordination enhancement state. The corresponding state is then bound to the window number, the number of persistently degrading parameters, the number of synchronization enhancement parameters, and the number of consecutively maintained windows to form a parameter coordination change state. The continuous time window in which the parameter coordinated change state first continuously reaches the degradation judgment condition is determined as the degradation start point. Starting from the degradation start point, adjacent continuous time windows in which the deviation direction is consistent, the deviation amplitude is continuously accumulated, and the parameter coordinated change state is not interrupted are connected to form the degradation duration segment. The generation of the degradation duration segment is as follows: The parameters' cooperative change states are sorted in ascending order according to the end time of the window. Degradation criteria are defined as the continuous occurrence of a cooperative sustained state or a cooperative enhanced state, the number of continuously degraded parameters being no less than two, and the number of consecutive holding time windows being no less than three. Starting from the beginning of the calibration deviation sequence under the horn capacitor operating condition, the first consecutive time window where all three consecutive time windows meet the degradation criteria is determined as the degradation start point, and the corresponding window number and end time are recorded. From the degradation start point, the continuously degraded parameters are compared window by window, and at least two continuously degraded parameters within adjacent consecutive time windows are maintained in degradation. If the direction and corresponding deviation amplitude are not less than 90% of the deviation amplitude of the previous continuous time window, and the parameter cooperative change state remains in a cooperative continuous state or a cooperative enhanced state, it is determined as a degenerate continuous state; connect adjacent continuous time windows that are continuously in the degenerate continuous state in chronological order, and accumulate the deviation amplitude of each continuously degenerate parameter and the number of connected windows; if two consecutive time windows are in a non-cooperative state, or the number of continuously degenerate parameters decreases to less than one, or the deviation amplitude decreases by more than 20% of the previous continuous time window, terminate the current connection, and determine the time range from the degenerate start point to the termination of the previous continuous time window as the degenerate duration segment; The rate of change of deviation is continuously compared within the duration of degradation, and the continuous time window in which the rate of change of deviation shows a continuous increase, a continuous decrease, or a change in the direction of change is determined as the degradation inflection point. By binding the degradation start point, degradation duration, and degradation inflection point with the corresponding key degradation parameters, deviation states, and durations, a degenerative event sequence of the horn capacitor is generated.

[0023] In this embodiment, the generation of the attention degradation characterization of the bullhorn capacitor includes: The calibration deviation sequence and degradation event sequence of the horn capacitor under operating conditions are input into the operating condition-gated DA-RNN model. The operating condition-gated DA-RNN model includes a sequence construction layer, an operating condition gating layer, a parameter attention encoding layer, an event fusion layer, and a time window attention output layer. The improvements to the condition-gated DA-RNN model are specifically reflected in the following aspects: A condition-gated layer is added before the input attention encoder of the standard DA-RNN model. The condition state of each consecutive time window is concatenated with the corresponding parameter deviation. The condition-gated weights are calculated using a fully connected mapping and a sigmoid function, and then multiplied term-by-term by the horn capacitor condition calibration deviation sequence to form the condition-calibrated input features. Event fusion processing is added between the encoder and the temporal attention decoder. The degradation start point, degradation duration, and degradation inflection point are encoded as corresponding degradation event features, and concatenated and linearly mapped with the encoded states of the same consecutive time window to form the event-enhanced state. The temporal attention mechanism of the standard DA-RNN model is adjusted to a time-window attention mechanism constrained by degradation events. The attention score of each consecutive time window is calculated based on the initial decoded state and the event-enhanced state. Degradation time-window weights are generated using a Softmax function, and the event-enhanced states are weighted and aggregated to generate the horn capacitor attention degradation representation. The training data for the condition-gated DA-RNN model includes accelerated aging test data, isothermal ripple load test data, charge-discharge cycle test data, equipment field operation data, and failure detection records of horn capacitors with different rated capacities and rated voltages. Continuous data from the initial operation to the failure time of each horn capacitor are processed using 30-minute and 10-minute sliding window steps, forming a training sample of 48 consecutive time windows. Based on capacitance retention rate, equivalent series resistance growth rate, actual failure time, and degradation event identification results, the training samples are labeled with health status, aging stage, remaining service life, and degradation. The model identifies the starting point, duration of degradation, and inflection point of degradation. It divides the data into training, validation, and test sets based on the capacitor's serial number, with proportions of 70%, 20%, and 10%, respectively. Samples from the same capacitor are assigned to only one dataset. The mean and standard deviation of the training set are used to normalize each dataset. During model training, the capacitor's operating condition calibration deviation sequence and degradation event sequence are simultaneously input into the model. The operating condition gating weights, degradation parameter weights, and degradation time window weights are calculated sequentially. Joint training is performed through the lifetime prediction branch, health status recognition branch, and degradation event recognition branch. Backpropagation is used to update the model parameters. The training batch size was set to 64, the encoding and decoding hidden dimensions were both set to 128, the degradation event feature dimension was set to 32, the dropout rate was set to 0.2, the AdamW optimizer was used, the initial learning rate was set to 0.001, the weight decay coefficient was set to 0.00001, the gradient clipping threshold was set to 5, and the maximum number of training epochs was set to 200. If the validation set loss did not decrease for 8 consecutive epochs, the learning rate was adjusted to 0.5 of the current learning rate, and training was terminated if it did not decrease for 20 consecutive epochs, while retaining the model parameters with the lowest total validation set loss. The loss function consisted of remaining lifespan prediction loss, health status prediction loss, aging stage classification loss, degradation event recognition loss, lifespan change monotonic constraint loss, and attention sparsity constraint loss, with corresponding weights set to 0.40, 0.20, 0.15, 0.10, 0.10, and 0.05, respectively. In the sequence construction layer, the deviation direction, deviation magnitude, and deviation change rate of each state parameter are time-aligned with the degradation start point, degradation duration, and degradation inflection point according to the end time of the window, generating a degradation input sequence and a degradation event marker sequence. The sequence construction layer is specifically processed as follows: The calibration deviation sequence and degradation event sequence of the horn capacitors are grouped according to their horn capacitor numbers, and arranged in ascending order according to the end time of the window, with 48 consecutive time windows forming a sequence construction interval; the deviation direction, deviation amplitude, and deviation change rate of each state parameter are extracted according to the parameter order in the horn capacitor calibration deviation sequence; positive deviation, no significant deviation, and negative deviation are encoded as 1, 0, and -1 respectively; the deviation amplitude is divided by 0.05, and the calculation result is limited to between 0 and 3; the deviation change rate is divided by the upper limit of the change rate of the corresponding parameter during the historical normal operation phase, and the calculation result is limited to -1. Between 3 and 3; arrange the direction code, amplitude code, and rate code of each state parameter within the same continuous time window in sequence to form a window deviation feature vector; combine the window deviation feature vectors according to the end time of the window to generate a degradation input sequence; set three-dimensional event markers according to the window numbers corresponding to the degradation start point, degradation duration, and degradation inflection point. When the current continuous time window is the degradation start point, set the start marker to 1; when it is within the degradation duration, set the duration marker to 1; when it is the degradation inflection point, set the inflection marker to 1; and set the other markers to 0; arrange the three-dimensional event markers of each continuous time window according to the end time of the window to generate a degradation event marker sequence; In the working condition gating layer, the gating value is calculated based on the working condition status and corresponding parameter deviation of each continuous time window. The working condition gating weight is determined by range constraints, and the degenerate input sequence is weighted item by item. The specific processing of the operating condition gating layer is as follows: Load rate, switching frequency, cumulative runtime, ambient temperature, and ambient humidity are normalized. Specifically, the load rate and ambient humidity values ​​are scaled to between 0 and 1. The switching frequency is proportionally calculated to 20kHz, and the cumulative runtime is proportionally calculated to 5000h, with the corresponding calculation results limited to between 0 and 1. The ambient temperature is mapped to a temperature range of -20℃ to 85℃, with -20℃ corresponding to 0 and 85℃ corresponding to 1, and mapping results outside the temperature range are limited to between 0 and 1. The charging state is encoded as 1, 0, 0; the holding state is encoded as 0, 1, 0; and the discharging state is encoded as 0, 0, 1, combining to form the operating condition gating layer. The working condition feature vector is concatenated with the window deviation feature vector and input into a fully connected unit containing 64 neurons for linear mapping and ReLU activation to generate a working condition deviation fusion feature. The working condition deviation fusion feature is input into a gating mapping unit, and a gating value corresponding to each state parameter is generated through the Sigmoid activation function. The gating value is restricted to between 0.05 and 1, and the restricted gating value is directly determined as the working condition gating weight of the corresponding state parameter. Each working condition gating weight is multiplied by the direction code, amplitude code, and rate code of the corresponding state parameter in the degenerate input sequence. The three types of codes corresponding to the same state parameter use the same working condition gating weight to generate a working condition-gated degenerate input sequence. In the parameter attention encoding layer, the parameter attention score is calculated based on the deviation between the previous encoded state and the weighted parameter. The degenerate parameter weight is determined after normalization. The weighted aggregation result of each state parameter is input into the recurrent encoding unit to generate a degenerate hidden state sequence. The parameter attention encoding layer processes the degenerate input sequence after condition gating sequentially according to the end time of the window. It combines the orientation, amplitude, and rate codes corresponding to the same state parameter within each consecutive time window into a parameter bias vector, generating parameter mapping features through a linear mapping unit containing 32 neurons. The encoded hidden state and encoded memory state corresponding to the previous consecutive time window are each set as 128-dimensional vectors, while the encoded hidden state and encoded memory state of the first consecutive time window are initialized to zero vectors. For each state parameter within the current consecutive time window, the parameter mapping features are concatenated with the previous encoded hidden state and input into an attention mapping unit containing 64 neurons. Hyperbolic tangent activation and single-neuron linear mapping are then used to obtain the corresponding parameters for each state parameter. The parameter attention scores are calculated; Softmax normalization is performed on all parameter attention scores within the same continuous time window to make the sum of the parameter attention scores equal to 1, and the degenerate parameter weights corresponding to each state parameter are determined; the degenerate parameter weights are multiplied by the corresponding parameter deviation vectors respectively, and the weighted parameter deviation vectors are summed to generate the parameter weighted aggregated features of the current continuous time window; the parameter weighted aggregated features, the previous encoded hidden state, and the previous encoded memory state are input into a long short-term memory unit with a hidden dimension of 128 to update the encoded hidden state and encoded memory state of the current continuous time window; 48 continuous time windows are processed in the same way, and the encoded hidden states generated by each continuous time window are arranged according to the end time of the window to generate a degenerate hidden state sequence; In the event fusion layer, the degenerate event tag sequence is mapped to degenerate event features and fused with the degenerate hidden states of the same continuous time window to generate an event-enhanced state sequence. The event fusion layer processes the following steps: The degenerate event label sequence is mapped window-by-window to the degenerate hidden state sequence according to window numbering. The start, duration, and transition labels for each consecutive time window are extracted and combined to form a three-dimensional event label vector. This three-dimensional event label vector is then input into an event mapping unit containing 32 neurons, where linear mapping, ReLU activation, and random inactivation with a dropout rate of 0.2 are performed sequentially to generate 32-dimensional degenerate event features. Finally, these features are concatenated with 128-dimensional degenerate hidden states from the same consecutive time window to form a 160-dimensional fusion input. The event fusion features are generated by performing dimensionality compression and hyperbolic tangent activation on a fusion mapping unit containing 128 neurons. These features are then input into a gated mapping unit, where a 128-dimensional event fusion weight is generated using a sigmoid activation function. The event fusion weights are multiplied term-by-term by the event fusion features, and then the event fusion weights are multiplied term-by-term by term-by-term by the degenerate hidden states. The two products are then summed to generate the event enhancement state for the current consecutive time window. Finally, the event enhancement states for each consecutive time window are arranged according to their end times to generate an event enhancement state sequence. In the time window attention output layer, the time attention score is calculated based on the initial decoding state and the event enhancement state sequence. After normalization, the weight of the degradation time window is determined. The event enhancement states of each consecutive time window are weighted and aggregated to generate the horn capacitive attention degradation representation.

[0024] The processing of the time window attention output layer is as follows: The event enhancement state corresponding to the last consecutive time window in the event enhancement state sequence is input into two linear mapping units containing 128 neurons each. After hyperbolic tangent activation, the initial decoded hidden state and the initial decoded memory state are generated. For each consecutive time window in the event enhancement state sequence, the corresponding 128-dimensional event enhancement state is concatenated with the initial decoded hidden state and the initial decoded memory state, and then input into a time attention mapping unit containing 64 neurons for linear mapping and hyperbolic tangent activation. Finally, a single-neuron mapping unit outputs the time attention score for the corresponding consecutive time window. The temporal attention scores of 48 consecutive time windows are subjected to Softmax normalization so that the sum of all temporal attention scores is 1, and the weight of the degenerate time window corresponding to each consecutive time window is determined. Each degenerate time window weight is multiplied by the corresponding event enhancement state, and the 48 weighted event enhancement states are summed to generate a 128-dimensional time window aggregate feature. The time window aggregate feature is concatenated with the event enhancement state corresponding to the last consecutive time window, and the input is linearly mapped, ReLU activated, and randomly deactivated with a dropout rate of 0.2 by an output mapping unit containing 128 neurons to generate a horn capacitive attention degenerate representation.

[0025] In this embodiment, the generation of the horn capacitor lifetime prediction result includes: A fully connected mapping is performed on the attention degradation characterization of the horn capacitor to generate the current health value and the lifetime decay increment corresponding to each predicted time step. The generation of lifetime decay increments is as follows: A 128-dimensional horn-capacitor attention degradation representation is input into both the health value mapping path and the lifetime decay mapping path. In the health value mapping path, the horn-capacitor attention degradation representation is input into a fully connected layer containing 64 neurons for linear mapping and ReLU activation, then into a fully connected layer containing 1 neuron. The output value is limited to between 0 and 1 using the Sigmoid activation function to generate the current health value. A health value closer to 1 indicates a lifespan state closer to the initial state. In the lifetime decay mapping path, the horn-capacitor attention degradation representation is input into a fully connected layer containing 128 neurons for linear mapping, ReLU activation, and random inactivation with a dropout rate of 0.2, then into a fully connected layer containing 50 neurons. This outputs decay values ​​for the next 50 predicted time steps, each corresponding to 100 hours. The Softplus activation function is used to convert each decay value to a non-negative number, and the decay value for a single predicted time step is limited to between 0 and 0.05. These values ​​are then arranged in the order of the predicted time steps to generate the lifetime decay increment for each predicted time step. The lifetime decay increment is non-negatively processed. The lifetime decay increment is subtracted sequentially from the current health value according to the predicted time step, limiting the health value of the next predicted time step to no higher than the health value of the previous predicted time step, thus generating a degenerate monotonic constrained lifetime evolution trajectory. Specifically, the lifetime decay increments corresponding to each predicted time step are arranged in the order of prediction time to form a non-negative lifetime decay increment sequence. The current health value is taken as the health value of the 0th predicted time step. Starting from the 1st predicted time step, the lifetime decay increment corresponding to the current predicted time step is subtracted from the health value of the previous predicted time step to obtain the candidate health value of the current predicted time step. The lifetime decay increment corresponding to the current predicted time step is subtracted from the health value of the previous predicted time step, and the lower limit of the calculation result is limited to 0 to obtain the health value of the current predicted time step. The health values ​​of the next 50 predicted time steps are calculated sequentially according to a 100-hour time interval corresponding to each predicted time step. The current health value and the health values ​​of the next 50 predicted time steps are arranged according to their corresponding times to form a degenerate monotonic constrained lifetime evolution trajectory where the health value changes non-incrementally with running time. The current health value is matched with the health status classification threshold to determine the health status. The aging stage is determined based on the degradation initiation point, degradation inflection point, and the position of the current health value in the degradation monotonic constrained lifetime evolution trajectory. Specifically, the current health value is compared with the health status classification threshold: a current health value greater than or equal to 0.80 is determined to be in a healthy state; a current health value greater than or equal to 0.60 and less than 0.80 is determined to be in a slightly deteriorated state; a current health value greater than or equal to 0.30 and less than 0.60 is determined to be in a moderately deteriorated state; and a current health value less than 0.30 is determined to be in a severely deteriorated state. Using a health value of 1 as the lifetime evolution start position and a failure threshold of 0.20 as the lifetime evolution end position, the current health value relative to the lifetime evolution start position is calculated. The cumulative decline rate determines the current health value's position in the degradation monotonic constraint lifetime evolution trajectory. Using a health value of 1 as the starting point of lifetime evolution and a failure threshold of 0.20 as the ending point, the difference between 1 and the current health value is divided by 0.80, and the result is limited to between 0 and 1 to obtain the lifetime progression position. An accelerated aging stage is defined as when the lifetime progression position is not less than 75%, or when the rate of deviation change continuously increases within three consecutive time windows after the most recent degradation inflection point. A stable aging stage is defined as when the accelerated aging stage has not been reached but a degradation inflection point has been identified. An early aging stage is defined as when no degradation inflection point has been identified but a degradation starting point has been identified, or when the lifetime progression position is not less than 25%. All other cases are defined as non-aging stages. The predicted time step when the health value first reaches the failure threshold is retrieved along the degradation monotonic constraint lifetime evolution trajectory, and the remaining lifetime is determined based on the time interval between the current time and the corresponding predicted time. The determination of remaining useful life is as follows: using 0.20 as the failure threshold, the health values ​​in the degradation monotonic constraint lifetime evolution trajectory are compared sequentially according to the predicted time steps. The prediction time step corresponding to the first health value less than or equal to 0.20 is determined as the failure prediction time step. When the current health value is less than or equal to 0.20, the remaining useful life is determined as 0 hours. When the health value of the previous prediction time step is higher than 0.20, linear interpolation is performed within a 100-hour prediction interval based on the health value of the previous prediction time step, the health value of the failure prediction time step, and the failure threshold to determine the failure prediction time when the health value reaches the failure threshold. The difference between the failure prediction time and the current time is calculated to obtain the remaining useful life. When the health values ​​of the next 50 prediction time steps are all higher than 0.20, the decrease in health value of the last 5 prediction time steps is extracted, the corresponding average decrease is calculated, and the degradation monotonic constraint lifetime evolution trajectory is extended according to the average decrease until the extended health value first reaches 0.20. The remaining useful life is determined based on the time interval between the extended time and the current time. Calculate the ratio of the decrease in health value between adjacent prediction time steps to the time interval, and weight and aggregate the ratios to determine the rate of lifespan decline; The determination of the lifespan decline rate is as follows: Following the predicted time step sequence, the health value of the previous predicted time step is subtracted from the health value of the current predicted time step to obtain the health value decline between adjacent predicted time steps. This health value decline is then divided by 100 hours to obtain the health value decline ratio per unit time for the corresponding predicted time interval. The next 50 predicted time steps are divided into five consecutive segments. The weights of the ratios corresponding to the 1st to 10th predicted time steps are set to 0.05, and the weights of the ratios corresponding to the 11th to 20th predicted time steps are set to... The weights of the ratios corresponding to the 21st to 30th prediction time steps are set to 0.0125, the weights of the ratios corresponding to the 31st to 40th prediction time steps are set to 0.0075, and the weights of the ratios corresponding to the 41st to 50th prediction time steps are set to 0.005. The ratio of the decline in health value per unit time is multiplied by the corresponding weight, and the product results are summed to determine the rate of lifespan decline. If there are fewer than 50 future prediction time steps, the weights corresponding to the existing prediction time steps are normalized before weighted aggregation is completed. By combining the health status, aging stage, remaining service life, life decline rate, and degradation monotonic constraint life evolution trajectory, the life prediction results of the horn capacitor are generated.

[0026] In this embodiment, the generation of the horn capacitor lifetime assessment results includes: Extract the current lifetime status and its changing trend from the lifetime prediction results of the horn capacitor, and locate the degradation inflection point closest to the current assessment time. The failure risk assessment value is calculated based on the change in the overall degradation rate before and after the degradation inflection point, the duration after the inflection point, and the remaining service life. The calculation of the failure risk assessment value is as follows: Locate the degradation inflection point closest to the current assessment time, extract the deviation change rate of key degradation parameters within three consecutive time windows before and after the degradation inflection point, multiply each deviation change rate by the corresponding degradation parameter weight and sum them to obtain the comprehensive degradation rate for each consecutive time window, calculate the average comprehensive degradation rate before the degradation inflection point and the average comprehensive degradation rate after the degradation inflection point; subtract the average lifetime decline rate before the inflection point from the average lifetime decline rate after the inflection point, and then divide by the absolute value of the average lifetime decline rate before the inflection point to obtain the rate growth ratio. Rates less than 0 are considered as growth rates. The rate growth ratio is corrected to 0, and rate growth ratios greater than 1 are corrected to 1; the time length between the end of the window corresponding to the degradation inflection point and the current assessment time is calculated, the time length is divided by 24h, and the calculation result is restricted to between 0 and 1 to obtain the inflection duration ratio; 2000h is subtracted from the remaining useful life and then divided by 2000h, and the calculation result is restricted to between 0 and 1 to obtain the remaining useful life reduction ratio; the rate growth ratio, inflection duration ratio, and remaining useful life reduction ratio are weighted and summed according to weights of 0.4, 0.2, and 0.4 respectively to generate a failure risk judgment value ranging from 0 to 1; The failure risk level is determined according to the range to which the failure risk assessment value belongs. The failure risk level is upgraded when the rate of lifespan decline continues to increase after the degradation inflection point. Specifically, the failure risk assessment value is matched with the risk level range: a failure risk assessment value less than 0.25 is classified as normal; between 0.25 and 0.50 as a concern; between 0.50 and 0.75 as a warning; and greater than or equal to 0.75 as an emergency. The lifespan decline rate after the most recent degradation inflection point is arranged according to the end time of the window. The lifespan decline rates of adjacent consecutive time windows are compared sequentially. If the lifespan decline rate in three consecutive time windows is higher than the previous consecutive time window, and the growth rate of the lifespan decline rate in the third time window relative to the first time window is not less than 20%, the lifespan decline rate is determined to be continuously increasing. When the lifespan decline rate continues to increase, the normal, concern, or warning level is upgraded by one level, while the emergency level remains unchanged. If the continuously increasing condition is not met, the failure risk level determined based on the failure risk assessment value is maintained. Maintenance trigger conditions are matched based on the failure risk level. Specifically, maintenance response rules are established according to the failure risk level. When the failure risk level is normal, a routine inspection cycle of 30 days is maintained, and a life assessment is re-executed at the next inspection time. When the failure risk level is of concern, the inspection cycle is shortened to 7 days, and primary capacitance, equivalent series resistance, and casing temperature are added for inspection. When the failure risk level is warning, a maintenance task is generated to complete re-inspection and replacement preparation within 72 hours. Replacement is triggered when the re-inspection result is still at the warning level or upgraded to the emergency level. When the failure risk level is emergency, a shutdown inspection is immediately triggered, the corresponding capacitor is taken out of operation, and replacement is performed. The inspection cycle, execution time limit, and maintenance actions corresponding to each level are combined to form maintenance trigger conditions, which are then bound to the failure risk level. By binding the failure risk level, maintenance triggering conditions, and corresponding judgment criteria with the horn capacitor number and assessment time, the horn capacitor life assessment result is generated.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a frequency converter rectifier power supply cabinet in an equipment manufacturing plant. This power supply cabinet is equipped with a DC bus filter unit, which consists of 24 parallel-connected horn capacitors with a rated voltage of 450V and a rated capacitance of 4700μF. The rated output power of the power supply cabinet is 320kW, and the switching frequency varies between 6kHz and 14kHz. During equipment operation, the ambient temperature ranges from 18℃ to 46℃, and the load rate fluctuates between 40% and 95%. When the load increases rapidly, the ripple current, casing temperature, internal estimated temperature, equivalent series resistance, and leakage current all increase simultaneously. The existing maintenance method mainly sets fixed alarm thresholds based on the capacitance reduction ratio and the equivalent series resistance increase ratio. This method cannot effectively distinguish between transient parameter fluctuations caused by high temperature and high ripple load and the degradation of the horn capacitors themselves, easily leading to repeated inspections and affecting the timely identification of continuous degradation.

[0028] In this scenario, the implementer deploys voltage and ripple current sensors on the DC bus side, sets up a leakage current detection circuit at the horn capacitor connection terminal, and deploys a temperature sensor in the middle of the casing. Data is collected on actual capacitance, equivalent series resistance, leakage current, loss tangent, terminal voltage, ripple current, casing temperature, estimated internal temperature, temperature rise rate, load factor, charge / discharge status, switching frequency, cumulative runtime, ambient temperature, and ambient humidity. After standardization, the collected data is correlated and organized according to 30-minute and 10-minute sliding window steps, forming a multi-domain state sequence for the horn capacitor. Based on rated parameters, temperature derating relationships, and ripple load relationships, reference states for electrothermal coupling conditions corresponding to each continuous time window are generated. Deviations between the actual electrical and thermal states and the corresponding reference values ​​are calculated to identify degradation initiation points, degradation durations, and degradation inflection points. The calibration deviation sequence and degradation event sequence of the horn capacitor under operating conditions are input into the operating condition-gated DA-RNN model to generate the attention degradation characterization of the horn capacitor. Based on this, the degradation monotonic constrained lifetime evolution trajectory is constructed to determine the health status, aging stage, remaining service life, lifetime degradation rate, failure risk level, and maintenance triggering conditions.

[0029] During a single operation where the load rate increased from 52% to 90%, the casing temperature of one of the capacitors rose from 34.6℃ to 45.2℃, and the estimated internal temperature increased from 39.8℃ to 54.1℃. The equivalent series resistance increased by 8.7% compared to the low-load state, and the leakage current increased by 11.3%. With a fixed alarm threshold, these parameter changes would be directly flagged as abnormal. This invention generates a corresponding electrothermal coupling reference state based on the current ambient temperature, ripple current, load rate, and switching frequency. After calibration, the equivalent series resistance deviation was 1.6%, and the leakage current deviation was 2.1%. Within subsequent consecutive time windows, no more than two key degradation parameters consistently showed a consistent degradation trend, and the coordinated parameter changes did not meet the degradation judgment criteria. After the load decreased, the casing temperature, equivalent series resistance, and leakage current returned to the corresponding reference range. The system maintained a regular detection cycle, avoiding misinterpreting short-term fluctuations caused by changes in operating conditions as lifespan degradation.

[0030] Another capacitor exhibited continuous degradation under similar load conditions. Its actual capacitance decreased continuously relative to the electrical state reference value over six consecutive time windows, with the deviation increasing from 3.2% to 7.9%. The deviation of the equivalent series resistance increased from 4.5% to 13.6%, and the leakage current and temperature rise rate also showed a consistent degradation trend. The system determined the degradation initiation point after the parameters continuously met the degradation criteria, and connected the subsequent continuous time windows maintaining the coordinated state as the degradation duration. When the rate of change of the deviations of the equivalent series resistance and the internally estimated temperature continuously increased, the system determined the degradation inflection point. The condition-gated DA-RNN model allocated higher degradation parameter weights to the equivalent series resistance, actual capacitance, and internally estimated temperature, and increased the degradation time window weights of the continuous time windows near the degradation inflection point, so that the degradation characterization focused on reflecting the continuous decay state rather than the instantaneous fluctuations during high load phases.

[0031] Based on the attention degradation characterization of the horn capacitor, the system generated a current health value of 0.43, classifying its health state as moderate decay. Combining the degradation initiation point, degradation inflection point, and lifetime progression position, it was determined to be in a stable aging stage. The health value in the degradation monotonic constraint lifetime evolution trajectory continuously decreased according to the predicted time step. The system determined the remaining lifetime to be 1380 hours, with a lifetime degradation rate of 0.016 per 100 hours, and assigned a failure risk level as a warning level, matching the maintenance trigger condition of completing re-inspection and replacement preparation within 72 hours. During the maintenance personnel's review, it was found that the actual capacitance of the horn capacitor had decreased to 82.6% of the rated capacitance, the equivalent series resistance reached 1.64 times the initial equivalent series resistance, and a slight bulge appeared in the top sealing area. Replacement was subsequently completed during equipment downtime maintenance.

[0032] In the continuous operation record of the same power cabinet, the fixed threshold method generated 11 abnormality alerts, of which 7 corresponding parameters dropped back after the load was reduced. After adopting the present invention, the corresponding short-term operating condition fluctuations did not lead to horn capacitor degradation events. Only for horn capacitors with continuous parameter coordinated changes, degradation duration, or degradation inflection points, the detection cycle was shortened, replacement preparation was initiated, or shutdown was carried out. This shows that the present invention can reduce the interference of changes in ambient temperature, ripple load, and switching frequency on life assessment, identify the actual degradation process of horn capacitors, and convert health status, remaining service life, and failure risk level into clear maintenance trigger conditions.

[0033] Table 1. Comparison of Comprehensive Performance of Horn Capacitor Life Assessment Methods under Multiple Operating Conditions and Disturbances

[0034] The comparative data in Table 1 is based on the same set of horn capacitor operating data. The traditional fixed threshold-LSTM method directly identifies anomalies based on the relationship between capacitance, equivalent series resistance, leakage current, and temperature parameters and a fixed threshold, and then uses a standard LSTM model to predict the remaining service life. The method of this invention adds operating condition correlation, electrothermal coupling condition calibration, degradation event identification, operating condition gating, and attention weight calculation to the same data. Both methods use the same data partitioning method, continuous time window, and failure threshold to reduce the impact of differences in training samples and evaluation conditions on the comparison results.

[0035] In terms of operating condition interference identification, the false alarm rate of traditional methods for operating condition fluctuations is 17.8%, while that of the method of this invention is 11.9%. This result indicates that when load rate, ambient temperature, ripple current, and switching frequency change, the method of this invention can reduce the occurrence of misclassifying normal operating condition fluctuations as capacitor degradation. This is because this invention does not directly compare the acquired parameters with a fixed threshold. Instead, it generates a reference state for the electrothermal coupling operating condition corresponding to the current continuous time window based on rated parameters, temperature derating relationships, and ripple load relationships, and then calculates the deviation of the actual electrical and thermal states relative to the corresponding reference state. The synchronous rise in parameters caused by high temperature or high ripple load can be corrected during the operating condition calibration process; only parameter changes that persist and exceed the corresponding operating condition reference range will be included in the subsequent degradation analysis.

[0036] In terms of lifetime prediction, the traditional method has an average absolute error of 194 hours and an average relative error of 13.8% for remaining lifetime; the corresponding values ​​for the method of this invention are 166 hours and 11.9%, respectively. The improvement in prediction accuracy mainly comes from the differentiated processing of different information contribution levels by the condition-gated DA-RNN model. The condition-gated weights are used to reduce invalid parameter changes caused by the current operating condition, the degradation parameter weights are used to highlight parameters with a high correlation to lifetime decay among the actual capacitance, equivalent series resistance, and internally estimated temperature, and the degradation time window weights are used to increase the contribution of the window near the degradation start point and degradation inflection point, so that the model reduces its dependence on normal operation phase and short-term disturbance data.

[0037] The traditional method has a lifetime trajectory monotonicity violation rate of 7.1%, indicating a local rebound phenomenon in the future health value output by ordinary time-series prediction models, which is inconsistent with the irreversible aging process of horn capacitors. The method of this invention has a lifetime trajectory monotonicity violation rate of 0 because it restricts the lifetime decay increment at each prediction time step to a non-negative value and subtracts the lifetime decay increment from the current health value sequentially according to the prediction time step, ensuring that the health value at the next prediction time step is no higher than that at the previous prediction time step. This constraint is not achieved by deleting abnormal prediction points later, but by directly limiting the direction of health value change during the lifetime evolution trajectory construction process, ensuring that the remaining useful life and the rate of lifetime decay have a consistent calculation basis.

[0038] Regarding failure risk and maintenance decisions, the method of this invention achieves an accuracy rate of 89.1% in determining the failure risk level and 87.6% in matching maintenance trigger conditions, which are higher than the 83.9% and 81.7% of traditional methods, respectively. Traditional methods mainly trigger re-inspection or replacement based on single parameter exceeding limits, making it difficult to distinguish between temporary parameter exceeding limits and continuous deterioration. The method of this invention comprehensively utilizes remaining service life, overall degradation rate, degradation inflection point, and the persistent state after the inflection point to determine the failure risk level, and matches maintenance trigger conditions corresponding to routine inspection, shortened inspection cycle, re-inspection, and replacement preparation according to different levels. Therefore, there is a more stable correspondence between failure risk level and actual maintenance needs.

[0039] The average time for a single lifetime assessment using the method of this invention is 0.67s, which is higher than the 0.48s of the traditional method. This is mainly due to the addition of electrothermal coupling reference state calculation, degradation event correlation, and multi-layer attention weight calculation. This increase does not have a significant impact on practical applications because the lifetime change of the horn capacitor is a slow evolution process. Data is sampled on a minute-by-minute basis and evaluated in continuous time windows. The 0.67s single calculation time is sufficient to meet the needs of online condition monitoring and maintenance decision-making.

[0040] The above are merely preferred embodiments 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 method for evaluating the lifetime of a horn capacitor based on an attention mechanism, characterized in that, Includes the following steps: Collect horn capacitor life assessment data and perform standardization processing to generate horn capacitor life assessment dataset; The life assessment dataset of the horn capacitor is windowed and associated according to the horn capacitor number, acquisition time and operating conditions. Electrical state, thermal state and operating condition are combined to generate a multi-domain state sequence of the horn capacitor. Based on the rated parameters, temperature derating relationship and ripple load relationship, an electrothermal coupling operating condition reference state is constructed. The deviation between the multi-domain state sequence of the horn capacitor and the electrothermal coupling operating condition reference state is calculated to generate the horn capacitor operating condition calibration deviation sequence. Based on the deviation direction, deviation amplitude, and parameter co-change within a continuous time window, the degradation initiation point, degradation duration, and degradation inflection point are identified, generating a degradation event sequence for the horn capacitor. The calibration deviation sequence and degradation event sequence of the horn capacitor under operating conditions are input into the operating condition gated DA-RNN model to determine the operating condition gate weight, degradation parameter weight and degradation time window weight, and generate the attention degradation characterization of the horn capacitor. Based on the attention degradation characterization of the horn capacitor, a degradation monotonic constrained lifetime evolution trajectory is constructed to determine the health status, aging stage, remaining service life and lifetime decline rate, and generate horn capacitor lifetime prediction results. Based on the predicted lifespan of the horn capacitor and the degradation inflection point, the failure risk level is determined and maintenance triggering conditions are matched to generate the lifespan assessment results of the horn capacitor.

2. The method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to claim 1, characterized in that, The life assessment data for the horn capacitor includes basic parameters, electrical status data, thermal status data, operating condition data, environmental status data, and historical life data. The standardization process includes timestamp alignment, numbering unification, data format conversion, measurement unit unification, sensor deviation correction, outlier removal, missing value completion, sampling frequency unification, operating condition coding, and numerical normalization.

3. The method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to claim 1, characterized in that, The generation of the horn capacitor multi-domain state sequence includes: The dataset for assessing the lifespan of horn capacitors was grouped according to their serial numbers, and the data in each group was sorted in ascending order by the time of collection. Based on the time window length and window sliding step, continuous time windows are divided, and data with the same horn capacitor number and acquisition time within the same continuous time window are time-aligned and aggregated. The operating conditions corresponding to each continuous time window are determined based on terminal voltage, ripple current, load rate, charging and discharging status, switching frequency, ambient temperature, and ambient humidity. Adjacent continuous time windows are associated according to the order of change of operating conditions. The actual capacitance, equivalent series resistance, leakage current, loss tangent, terminal voltage and ripple current are extracted from each continuous time window and combined to form the electrical state. Extract the shell temperature, estimated internal temperature, and temperature rise rate, and combine them to form the thermal state; Extract load rate, charging and discharging status, switching frequency, cumulative running time, ambient temperature and ambient humidity, and combine them to form the operating condition status; The electrical state, thermal state, and operating state are bound together according to the time sequence of a continuous time window to generate a multi-domain state sequence of the horn capacitor.

4. The method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to claim 1, characterized in that, The generation of the bullhorn capacitor operating condition calibration deviation sequence includes: Match the rated capacitance, rated voltage, rated ripple current, rated operating temperature, initial equivalent series resistance, initial leakage current, and initial loss tangent value to the horn capacitor number; The temperature offset is determined based on the ambient temperature, shell temperature and internal estimated temperature of each continuous time window, and the rated capacitance, initial equivalent series resistance, leakage current and loss tangent are corrected for temperature based on the temperature derating relationship. The ripple load factor is determined based on the ratio of ripple current to rated ripple current, load factor and switching frequency. The load is then adjusted for terminal voltage, ripple current, temperature rise rate and internal estimated temperature in combination with the ripple load relationship. The temperature correction results and load correction results are combined according to the corresponding continuous time window to generate electrical state reference values ​​and thermal state reference values, and to construct the electrothermal coupling operating condition reference state. The electrical, thermal, and operating states of each continuous time window in the multi-domain state sequence of the horn capacitor are compared with the corresponding reference values ​​by performing difference and ratio calculations to determine the deviation direction, deviation magnitude, and deviation changes between adjacent continuous time windows for each state parameter. The horn capacitor operating condition calibration deviation sequence is generated by arranging the parameters according to the end time of the window.

5. The method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to claim 1, characterized in that, The generation of the bullhorn capacitor degradation event sequence includes: The calibration deviation sequence of the horn capacitor is arranged according to the end time of the window, and the deviation direction, deviation amplitude and deviation change rate of the key degradation parameters in each continuous time window are extracted. By aligning key degradation parameters with consistent degradation direction over time, and by counting the number of consecutive time windows with consistent deviation direction and the number of key degradation parameters with synchronously changing deviation magnitude, the state of parameter co-change is determined. The continuous time window in which the parameter coordinated change state first continuously reaches the degradation judgment condition is determined as the degradation start point. Starting from the degradation start point, adjacent continuous time windows in which the deviation direction is consistent, the deviation amplitude is continuously accumulated, and the parameter coordinated change state is not interrupted are connected to form the degradation duration segment. The rate of change of deviation is continuously compared within the duration of degradation, and the continuous time window in which the rate of change of deviation shows a continuous increase, a continuous decrease, or a change in the direction of change is determined as the degradation inflection point. By binding the degradation start point, degradation duration, and degradation inflection point with the corresponding key degradation parameters, deviation states, and durations, a degenerative event sequence of the horn capacitor is generated.

6. The method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to claim 1, characterized in that, The generation of the bullhorn capacitor attention degradation characterization includes: The calibration deviation sequence and degradation event sequence of the horn capacitor under operating conditions are input into the operating condition-gated DA-RNN model. The operating condition-gated DA-RNN model includes a sequence construction layer, an operating condition gating layer, a parameter attention encoding layer, an event fusion layer, and a time window attention output layer. In the sequence construction layer, the deviation direction, deviation magnitude, and deviation change rate of each state parameter are time-aligned with the degradation start point, degradation duration, and degradation inflection point according to the end time of the window, generating a degradation input sequence and a degradation event marker sequence. In the working condition gating layer, the gating value is calculated based on the working condition status and corresponding parameter deviation of each continuous time window. The working condition gating weight is determined by range constraints, and the degenerate input sequence is weighted item by item. In the parameter attention encoding layer, the parameter attention score is calculated based on the deviation between the previous encoded state and the weighted parameter. The degenerate parameter weight is determined after normalization. The weighted aggregation result of each state parameter is input into the recurrent encoding unit to generate a degenerate hidden state sequence. In the event fusion layer, the degenerate event tag sequence is mapped to degenerate event features and fused with the degenerate hidden states of the same continuous time window to generate an event-enhanced state sequence. In the time window attention output layer, the time attention score is calculated based on the initial decoding state and the event enhancement state sequence. After normalization, the weight of the degradation time window is determined. The event enhancement states of each consecutive time window are weighted and aggregated to generate the horn capacitive attention degradation representation.

7. The method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to claim 1, characterized in that, The generation of the bullhorn capacitor lifetime prediction results includes: A fully connected mapping is performed on the attention degradation characterization of the horn capacitor to generate the current health value and the lifetime decay increment corresponding to each predicted time step. The lifetime decay increment is nonnegated, and the lifetime decay increment is subtracted from the current health value sequentially according to the predicted time step. The health value of the next predicted time step is restricted to not be higher than the health value of the previous predicted time step, thus generating a degenerate monotonic constrained lifetime evolution trajectory. The current health value is matched with the health status classification threshold to determine the health status, and the aging stage is determined based on the degradation start point, degradation inflection point and the position of the current health value in the degradation monotonic constrained lifetime evolution trajectory. The predicted time step when the health value first reaches the failure threshold is retrieved along the degradation monotonic constraint lifetime evolution trajectory, and the remaining lifetime is determined based on the time interval between the current time and the corresponding predicted time. Calculate the ratio of the decrease in health value between adjacent prediction time steps to the time interval, and weight and aggregate the ratios to determine the rate of lifespan decline; By combining the health status, aging stage, remaining service life, life decline rate, and degradation monotonic constraint life evolution trajectory, the life prediction results of the horn capacitor are generated.

8. The method for evaluating the lifetime of a horn capacitor based on an attention mechanism according to claim 1, characterized in that, The generation of the horn capacitor life assessment results includes: Extract the current lifetime status and its changing trend from the lifetime prediction results of the horn capacitor, and locate the degradation inflection point closest to the current assessment time. The failure risk assessment value is calculated based on the change in the overall degradation rate before and after the degradation inflection point, the duration after the inflection point, and the remaining service life. The failure risk level is determined according to the range to which the failure risk assessment value belongs, and the failure risk level is adjusted upward when the rate of life decline continues to increase after the degradation inflection point. Match maintenance trigger conditions based on failure risk level; By binding the failure risk level, maintenance triggering conditions, and corresponding judgment criteria with the horn capacitor number and assessment time, the horn capacitor life assessment result is generated.

Images