Dual time scale mamba state space method for medium voltage cable joint aging assessment

Abstract

The application discloses a kind of double time scale Mamba state space's medium voltage cable joint aging evaluation method, comprising: obtaining the multi-source monitoring data of medium voltage cable joint, equipment attribute, life cycle identification and maintenance record;Calculate joint temperature rise residual and extract short-time event characteristics, obtain fine, coarse time scale sequence and cumulative stress characteristics;Fast, slow scale Mamba branch obtains fast event representation and preliminary slow state respectively, and fast event representation is written back to slow state by reliability and degradation correlation degree gate;Project fast event representation and slow state to public dimension and fuse, obtain continuous aging score;According to aging score, determine grade, according to maintenance record, manage life cycle or update state.The application distinguishes short-time reversible event and long-term irreversible degradation, maintains the causality of online processing, and realizes the reasonable management of life cycle and historical cumulative stress.

Description

Technical Field

[0001] This invention belongs to the field of cable accessory condition monitoring and intelligent assessment technology, specifically involving a method for assessing the aging of medium-voltage cable joints in a dual-timescale Mamba state space. Background Technology

[0002] Medium-voltage cable joints are used to complete conductor connections, insulation restoration, electric field control, continuous metallic shielding, and sealing protection. The conductor connection interface, insulation interface, and sealing interface of the joint are easily affected by the combined effects of installation quality, contact pressure, load changes, thermal cycling, moisture intrusion, mechanical stress, and local electric field distortion. Joint aging is usually manifested as a gradual increase in contact resistance, changes in temperature rise characteristics, enhanced partial discharge activity, or a decline in the performance of the insulation material.

[0003] Existing methods for assessing the condition of medium-voltage cable joints include periodic offline testing, material surface or dielectric parameter detection, fixed threshold monitoring, multi-index comprehensive evaluation, and machine learning methods based on temperature or partial discharge data. Offline methods can obtain relatively direct insulation condition information, but they are difficult to continuously reflect the operation process; fixed thresholds are prone to confusing short-term load disturbances with continuous degradation; and general machine learning methods usually rely on fixed-length windows, making it difficult to simultaneously retain short-term anomalies and long-term trends.

[0004] Monitoring data for medium-voltage cable joints have different physical time constants. Partial discharge pulse clusters, abrupt changes in temperature rise slope, abnormal sheath current, and vibration shocks can occur on the order of seconds to minutes; thermal cycling accumulation, contact interface relaxation, slow moisture intrusion, and insulation aging typically manifest over hours, days, or longer periods. Directly concatenating all high-frequency data into a long sequence would not only increase communication and inference overhead but could also dilute a small number of critical events with a large amount of stable data.

[0005] Existing multi-scale or bi-branch time-series models typically employ direct splicing of local and global branches after separate encoding. This general structure fails to address four key issues in medium-voltage cable joint scenarios: First, short-term events may originate from noise, load transients, or reversible operational disturbances, and cannot be unconditionally solidified as long-term aging simply because the event actually occurred; second, during online processing, the current coarse window has not yet ended, and its complete statistics cannot be used in advance to modulate the data preceding that window; third, the risk of short-term events can decrease as operating conditions recover, and cannot be confused with the long-term aging scores, which have weak monotonicity, in the same output; fourth, after maintenance or replacement, simply resetting the final score without distinguishing between the new life cycle, historical accumulated stress, and network internal states can lead to confusion in physical meaning or the retention of historical information.

[0006] Mamba state-space networks can handle long-term sequences through input-selective state update mechanisms, but directly replacing the original time-series model with a general Mamba network cannot automatically generate the fast and slow state interactions required for the aging mechanism of medium-voltage cable joints. If the fast and slow branches are only spliced ​​at the output, short-term anomalies will not change the subsequent long-term memory; if the information that has not yet been completed in the current window is used to modulate the already occurred fine window, it will lead to future information leakage.

[0007] To address the technical problems in the existing technologies mentioned above, such as the difficulty in simultaneously considering short-term anomalies and long-term degradation, the lack of dedicated state interaction for medium-voltage cable joints in general dual-branch structures, the potential introduction of future information in online processing, and the residual historical state after maintenance, a medium-voltage cable joint aging assessment method is needed that can distinguish between short-term reversible events and long-term irreversible degradation, maintain a strict causal relationship in time, and accurately manage the life cycle and historical strain according to the maintenance type. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a dual-time-scale Mamba state space-based method for evaluating the aging of medium-voltage cable joints, thereby resolving the issues present in the prior art.

[0009] To achieve the above objectives, this invention provides a method for aging assessment of medium-voltage cable joints in a dual-timescale Mamba state space, comprising: Acquire multi-source monitoring data, equipment attributes, lifecycle identifiers, and maintenance records for medium-voltage cable joints; Based on the multi-source monitoring data, the joint temperature rise residual is calculated using the health load-environment baseline model, and short-term event features are extracted based on the actual available monitoring modes in the multi-source monitoring data; a fine time-scale sequence is obtained based on the joint temperature rise residual and the short-term event features, a coarse time-scale sequence is obtained based on the fine time-scale sequence, and cumulative stress features are generated. Based on the fine-scale time sequence and the slow-scale hidden state of the previous coarse-scale time window, fast event representations and short-term event risks are obtained through the fast-scale Mamba branch. Based on the coarse timescale sequence and the slow-scale hidden state of the previous coarse time window, the initial slow-scale hidden state is obtained through the slow-scale Mamba branch. The fast event representation is written back to the initial slow-scale hidden state through event reliability and degradation correlation gating to obtain the slow-scale hidden state of the current coarse time window. The fast event representation and the slow-scale hidden state of the current coarse time window are projected onto a common dimension, and the projected fast and slow representations, the cumulative stress characteristics, and the device attributes are fused to obtain a continuous aging score. The aging level is determined based on the continuous aging score, and lifecycle management or status update is performed based on the lifecycle identifier and the maintenance record.

[0010] Optionally, the health load-environment baseline model is established from early health data of joints of the same life cycle, data of healthy joints of the same type, or a thermal circuit model; the joint temperature rise residual is the difference between the measured temperature rise of the medium-voltage cable joint and the expected health temperature rise output by the baseline model under the same load, ambient temperature, and available environmental variables, wherein the measured temperature rise of the medium-voltage cable joint is the difference between the measured temperature of the joint and the ambient temperature.

[0011] Optionally, based on the first time span, the joint temperature rise residual and the short-term event characteristics are used to form the fine time-scale sequence. The fine time-scale sequence includes at least the resampled monitoring quantity, the rate of change, the abnormal event characteristics corresponding to the actual available monitoring modes, the data quality marker, and the missing measurement mask. Among them, the event-type data does not generate fictitious events through interpolation, and the short-term missingness of continuous variables is filled by the adjacent valid data. Based on a second time span greater than the first time span, multiple time-continuous fine-time-scale sequences are statistically aggregated to form the coarse-time-scale sequence. The coarse-time-scale sequence includes at least six of the following: mean, standard deviation, quantile, trend slope, duration of exceeding limits, event occurrence rate, load rate, data integrity rate, and cumulative stress increment. A coarse time window is composed of multiple time-continuous fine-time windows.

[0012] Optionally, the cumulative stress characteristics include modal cumulative values ​​formed over time by at least two of the following: temperature stress rate per unit time, partial discharge stress rate, humidity stress rate, and load cycle stress rate; the modal cumulative values ​​are updated in the following manner: ; in, Indicates a coarse time window index; Indicates stress mode index, ; and They represent the first The and the first At the end of the first coarse time window The cumulative value of each stress mode; This indicates that the mode is in the first... Does the coarse time window contain data that can be used for stress calculation? This represents the corresponding data quality coefficient; Indicates the first Duration of a coarse time window; Indicates the first The unit time, dimensionless and non-negative stress rate of a certain mode.

[0013] Optionally, both the fast-scale Mamba branch and the slow-scale Mamba branch include an input mapping layer, a state space layer with an input-selective state update mechanism, a gated residual layer, and a normalization layer; the hidden state of the fast-scale Mamba branch is only recursively processed within the current coarse time window and is reinitialized based on the previous slow-scale hidden state at the beginning of the next coarse time window; before the current coarse time window has ended, the fast-scale Mamba branch is not modulated using the complete statistical features that have not yet been formed in the coarse time window.

[0014] Optionally, in the fast-scale Mamba branch, gating convergence is performed based on the event intensity, duration, data quality, and modal availability of each fine time window within the current coarse time window to obtain the fast event representation; the gating convergence includes relative event weights and an overall event gate, the relative event weights are used to compare the importance of each fine time window within the same coarse time window, and the overall event gate is determined based on the event intensity, duration, data quality, and modal availability corresponding to the actual available monitoring modal.

[0015] Optionally, the event reliability and degradation correlation are respectively formed into an event reliability gate and a degradation correlation gate, and combined element by element into a write-back gate; the event reliability gate is determined based on sensor health status, data integrity rate, noise level, and modal availability; the degradation correlation gate is determined based on at least two of the following: the persistence of the event in multiple fine time windows, the reproducibility in adjacent coarse time windows, the residual degree after load recovery, and the mutual corroboration relationship between different monitoring modalities.

[0016] Optionally, the fast event representation and the slow-scale hidden state of the current coarse time window are projected onto a common dimension to obtain fast representation and slow representation. The fast representation, the slow representation, the difference features and interaction features between the fast representation and the slow representation, the cumulative stress features, and the device attributes are fused together to obtain a continuous aging score between 0 and 1. The aging level is divided according to multiple ordered thresholds determined by independent verification data.

[0017] Optionally, when the maintenance record does not contain valid maintenance records, trend correction is only performed on abnormal declines in the continuous aging score exceeding the allowable range, and the trend correction does not apply to short-term event risks; when the maintenance record contains records of complete replacement or re-creation, a new lifecycle identifier is generated, equipment attributes are updated, the pre-maintenance scroll window and event cache are cleared, and the health baseline, accumulated stress, and slow-scale initial state are reconstructed; when the maintenance record contains records of partial repairs, the pre-maintenance historical accumulated stress is retained, the short-term event cache is cleared, and a post-maintenance candidate slow state is generated based on the maintenance type, re-inspection results, and post-maintenance observation period data to update the slow-scale hidden state.

[0018] Optionally, it further includes generating an aging trend curve based on the time series of the continuous aging scores, and combining the aging level, the rate of change of the continuous aging scores, the assessment uncertainty, and the short-term event risk to output graded maintenance prompts or supplementary inspection prompts. Compared with the prior art, the present invention has the following advantages and technical effects: First, the difference between the measured temperature rise and the expected temperature rise under the same load and environment is used to reduce the interference of changes in normal operating conditions. Second, fast and slow sequences are constructed based on different physical time constants, and the current fast branch is modulated by the previous slow state to maintain online causal relationships. Third, event reliability gates and degradation correlation gates are used to avoid unconditionally writing noise or reversible transients into the long-term state. Fourth, short-term event risks and long-term aging scores are output separately, and trend constraints only apply to the long-term aging score. Fifth, each stress mode is accumulated separately, and missing measurements will not cause historical stress to regress. Sixth, overall replacement establishes a new life cycle, local repair retains historical stress, and slow states are updated with re-inspection and observation period data, and maintenance processing has clear physical meaning. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the overall process of the medium-voltage cable joint aging assessment method according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a method for evaluating the aging of medium-voltage cable joints using a dual-timescale Mamba state space, including: S1 acquires multi-source monitoring data, equipment attributes, lifecycle identifiers, test or maintenance results, and maintenance records for medium-voltage cable joints.

[0023] Specifically, the joint temperature, ambient temperature, and load current of the target medium-voltage cable joint are obtained, and at least one of the following is obtained based on the field sensing conditions: partial discharge signal, sheath current, ambient humidity, leakage current, and vibration signal. At the same time, the equipment attributes, life cycle identifier, and maintenance records of the target medium-voltage cable joint are obtained.

[0024] Specifically, the equipment attributes include at least two of the following: connector type, voltage level, insulation material, commissioning time, laying environment, manufacturing batch, and construction batch; the life cycle identifier is used to distinguish different connectors used successively at the same installation location, and a new life cycle identifier is generated after the connector is completely replaced or remanufactured; the maintenance record includes at least the connector location identifier, life cycle identifier, maintenance time, maintenance type, work order number, and re-inspection result.

[0025] S2 performs time alignment, anomaly identification, and missing measurement marking on the data. It obtains the temperature rise residual based on the health load-environment baseline model, extracts short-term event features according to the actual available monitoring modes, and forms fine time-scale sequences, coarse time-scale sequences, and cumulative stress features maintained by mode.

[0026] Specifically, the acquired data is time-aligned, anomaly data is identified, and missing data is marked. The joint temperature rise residual is calculated using the load-environment baseline model under healthy conditions. Based on the actual available monitoring modes, the temperature rise change rate and load mutation characteristics, as well as at least one of partial discharge events, sheath current anomalies, humidity mutations, leakage current anomalies, and vibration and shock characteristics are extracted from the short-term change data. A fine time-scale sequence is formed according to the first time span, and a coarse time-scale sequence is formed according to the second time span, which is greater than the first time span.

[0027] Furthermore, the load-environment baseline model under the healthy state is established from early health data of joints of the same life cycle, data of healthy joints of the same type, or thermal circuit models; the joint temperature rise residual is the difference between the measured temperature rise of the joint and the expected temperature rise of the health given by the baseline model under the same load, ambient temperature, and available environmental variables, wherein the measured temperature rise of the joint is the difference between the measured temperature of the joint and the ambient temperature; the range where the load is close to zero is not amplified by directly dividing the temperature difference by the load rate.

[0028] Furthermore, the fine-scale time series includes at least resampled monitoring quantities, rate of change, abnormal event characteristics corresponding to the actual available modes, data quality markers, and missing measurement masks; short-time interpolation can be performed on continuous variables in temperature, load, and environmental quantities, and fictitious events are not generated through interpolation for event-type data in partial discharge and vibration shock; the coarse-scale time series includes at least six of the following: mean, standard deviation, quantiles, trend slope, duration of exceeding limits, event occurrence rate, load rate, data integrity rate, and cumulative stress increment; a coarse time window is composed of multiple time-continuous fine time windows.

[0029] Furthermore, the accumulated stress characteristics include at least two of the following: temperature stress rate per unit time, partial discharge stress rate, humidity stress rate, and load cycle stress rate, which are modal cumulative values ​​formed over time. The partial discharge quantity integrated by the window is first converted into activity per unit time before being accumulated. When a certain sensing mode is missing, the stress rate of that mode is not set to zero and its historical cumulative value is reduced accordingly. Instead, the previous effective cumulative value of that mode is maintained, the missing measurement mask is recorded, and the evaluation uncertainty is increased. When the operating conditions exceed the calibration range, the fusion weight of the corresponding physical feature is reduced and a supplementary calibration is prompted.

[0030] S3, within the current coarse time window, initialize and modulate the fast-scale Mamba branch using the slow-scale hidden state already formed in the previous coarse time window, and encode each fine time window in sequence to form fast event representation and short-term event risk.

[0031] Specifically, within the k-th coarse time window, the fine time-scale sequence is input into the fast-scale Mamba branch, and the window state of the fast-scale Mamba branch is initialized and conditionally modulated using the slow-scale hidden state already formed in the (k-1)-th coarse time window. Gated aggregation is performed based on the event intensity, persistence, data quality, and modal availability of each fine time window within the current coarse time window to obtain the fast event representation and short-term event risk of the k-th coarse time window.

[0032] S4. After the current coarse time window ends, a preliminary slow state is formed using the slow-scale Mamba branch. The reliability and degradation correlation of the event are calculated respectively. Only fast events that are reliable and have continuity, reproducibility, residual after load recovery, or cross-modal verification are written into the slow state, so that short-term anomalies that have real degradation significance affect the subsequent coarse time window.

[0033] Specifically, after the kth coarse time window ends, the coarse-scale features formed by the current coarse time window are used as the current time step of the slow-scale sequence and input into the slow-scale Mamba branch to obtain the initial slow-scale hidden state. Event reliability gate and degradation correlation gate are generated according to event reliability and degradation correlation, respectively. The two are combined into a write-back gate. Only reliable fast events that are persistent, reproducible, have residual after load recovery, or are cross-modal verified are written into the initial slow-scale hidden state through bounded mapping to form the slow-scale hidden state of the kth coarse time window.

[0034] Both the fast-scale Mamba branch and the slow-scale Mamba branch include an input mapping layer, a state space layer with an input-selective state update mechanism, a gated residual layer, and a normalization layer. The fast-scale Mamba branch and the slow-scale Mamba branch are allowed to use different hidden dimensions. The initial slow-scale hidden state is generated from healthy baseline features and device attributes, or, when historical data is lacking, a zero vector or a trained initial vector is used. The hidden state of the fast-scale Mamba branch is only recursively applied within the current coarse time window and is reinitialized based on the previous slow-scale hidden state at the beginning of the next coarse time window. Before the current coarse time window has ended, the complete statistical features not yet formed in the coarse time window are not used to modulate the fast-scale Mamba branch.

[0035] The gating aggregation includes relative event weights and an overall event gate. The relative event weights are used to compare the importance of each fine time window within the same coarse time window. The overall event gate is determined based on the event intensity, duration, data quality, and modality availability corresponding to the actual available monitoring modalities. The event reliability gate is used to reduce the amount of data written for noise, sensor anomalies, and incomplete data events. The degradation correlation gate is determined based on at least two of the following: the persistence of the event within multiple fine time windows, the reproducibility within adjacent coarse time windows, the residual degree after load recovery, and the mutual corroboration relationship between different monitoring modalities. Only when both event reliability and degradation correlation meet the requirements is the amount of data written by fast events to the hidden state at the slow scale increased.

[0036] S5 projects the fast and slow characteristics onto a common dimension, integrates them with the cumulative stress and equipment properties maintained by mode, outputs a continuous aging score, and further determines the assessment uncertainty; and classifies the aging level based on the threshold determined by the validated data.

[0037] Specifically, the fast event representation and the slow-scale hidden state of the kth coarse time window are projected onto the same common dimension, and the projected fast representation, slow representation, their difference features, interaction features, cumulative stress features maintained according to the monitoring mode, and equipment attributes are fused to obtain a continuous aging score between 0 and 1, and the evaluation uncertainty is further determined; the aging level of the medium-voltage cable joint is obtained based on the grading threshold determined by the verified data.

[0038] S6 performs tolerance-based trend correction on long-term aging scores and does not impose monotonic constraints on short-term event risks; after overall replacement or remanufacturing, a new life cycle is established and reinitialized; after local repair, historical accumulated stress is retained and the slow state is updated based on re-inspection and observation period data.

[0039] Specifically, when no valid maintenance records exist, trend correction is only applied to abnormal declines in the continuous aging score exceeding the allowable range. This trend correction does not apply to short-term event risks. When the joint is completely replaced or remanufactured, a new lifecycle identifier is generated, equipment attributes are updated, the pre-maintenance scroll window and event cache are cleared, and the health baseline, accumulated stress, and slow-scale initial state are reconstructed. During local repairs, the pre-maintenance historical accumulated stress is retained, the short-term event cache is cleared, and post-maintenance candidate slow states are generated based on the maintenance type, re-inspection results, and post-maintenance observation period data to update the slow-scale hidden state.

[0040] S7 generates an aging trend curve based on the time series of continuous aging scores, and outputs graded maintenance prompts or supplementary testing prompts by combining the aging level, the rate of change of continuous aging scores, assessment uncertainty and short-term event risk.

[0041] The above steps S1 to S7 provide the overall process framework for this embodiment. The following, in conjunction with... Figure 1 This document further explains the key technical details of each step involved in this embodiment, including data sources and aging annotation, healthy baseline correction and dual-timescale sequence construction, causal dual-timescale Mamba state encoding, common dimension fusion and aging output, model training and validation, online evaluation and post-maintenance state processing.

[0042] (a) Data sources and aging labeling; At the target joint or similar joint, the joint temperature, ambient temperature, and load current are collected. Depending on the field sensing conditions, one or more of the following are also collected: partial discharge, sheath current, ambient humidity, leakage current, or vibration. The joint temperature and ambient temperature are used to calculate the measured temperature rise, while the load current and environmental variables are used to establish the expected temperature rise under healthy operating conditions. Unconfigured monitoring modes participate in subsequent processing through mode availability markers and missing measurement masks; their data are not assumed to exist.

[0043] For the original waveform of high-frequency partial discharge, pulse detection and noise suppression can be performed at the sensing terminal, and the pulse number, equivalent amplitude, phase distribution, repetition rate, pulse cluster duration, and noise level can be uploaded. The so-called partial discharge intensity or energy characteristics are equivalent indicators obtained from the sensing signal and are not directly equivalent to the actual discharge energy inside the connector.

[0044] Equipment attributes must record at least two of the following: connector type, voltage rating, insulation material, commissioning time, installation environment, and manufacturing or construction batch. In addition to installation location identification, a lifecycle identifier should also be provided to distinguish different connectors installed or remanufactured at the same location. Maintenance records should include connector location identification, lifecycle identifier, maintenance time, work order number, maintenance type, and re-inspection results, avoiding the use of other connectors, old lifecycles, or invalid work orders for status updates.

[0045] Aging indicators can be formed by one or more of the following: accelerated thermal aging, thermal cycling, electrical aging, partial discharge, dielectric loss, insulation resistance, withstand voltage test, disassembly inspection, and on-site maintenance conclusions. Continuous aging scores use a normalized scale of 0 to 1, where 0 represents the baseline health state and 1 represents reaching a predefined critical degradation state.

[0046] Different experimental indicators are first oriented and normalized based on the health baseline and critical boundaries, and then continuous labels are formed based on the reliability of the indicators. Samples with only grade conclusions are used for grade or interval supervision, and the median of a certain grade is not directly regarded as an exact continuous label; when different experimental results conflict with each other, a simple averaging is not performed directly, but the label reliability is determined based on the experimental effectiveness, the time proximity to the monitoring window, and the retest results; when it is necessary to form a weak continuous label, the training weight of the sample is reduced.

[0047] (ii) Healthy baseline correction and dual-timescale series; Timestamps from different sensors are standardized, and resampling is performed according to the target sampling intervals allowed on-site. For continuous variables such as temperature, load current, ambient temperature, and ambient humidity, short-term missing data can be filled in using nearby valid data or local statistics. For event-type data such as partial discharge pulses, vibration shocks, and sudden changes in leakage current, fictitious events are not generated through interpolation; only the missing measurement mask, modal availability markers, and data integrity rate are retained. Continuous missing data exceeding a set duration is not continuously filled using a fixed constant. Abnormal data should be judged in conjunction with physical upper and lower limits, sensor health status, and neighborhood statistics to avoid deleting real partial discharge pulses as ordinary outliers.

[0048] To distinguish between temperature rise caused by normal load and environmental conditions and additional temperature rise caused by deterioration of connection condition, this embodiment establishes a healthy load-environment baseline model. The model can be obtained from early health data of joints within the same lifecycle, data of healthy joints of the same type, a piecewise regression model, or a thermal circuit model. The measured temperature rise and temperature rise residual of the joint are defined as follows: ; ; In the formula, Indicates the sampling time; and Representing time respectively The measured temperature of the connector and the ambient temperature; This indicates the measured temperature rise at the joint; This represents the expected temperature rise in health output from the health load-environment baseline model; Indicates the load current; This represents a vector of environmental variables consisting of ambient humidity, wind speed, laying conditions, or other available environmental quantities. Indicates the parameters of the health baseline model; This represents the difference between the measured temperature rise and the expected healthy temperature rise, i.e., the residual temperature rise at the joint, which has the same dimensions as temperature. In low-load ranges, the temperature difference is not amplified by dividing by a near-zero load rate to avoid numerical distortion.

[0049] Fine-scale time windows are used to retain the rate of temperature rise, load transitions, and actual field-available partial discharge pulse clusters, sheath current abrupt changes, humidity abrupt changes, leakage current anomalies, or vibration shocks, while simultaneously retaining data quality, missing measurement masks, and modal availability markers. Coarse-scale time windows consist of multiple consecutive fine windows and are used to statistically analyze the mean and quantiles of temperature rise residuals, trend slope, duration of exceedances, event occurrence rate, load cycles, data integrity rate, and stress increments for each modality.

[0050] The first time span is determined based on the duration of partial discharge activity, sensor sampling rate, and transient thermal response; the second time span is determined based on load cycling and thermal aging evolution cycles. Both can be adjusted according to the connector type, but the second time span is longer than the first time span. The resulting dual time scale does not arbitrarily copy the same data into two windows, but rather corresponds to short-term event processes and long-term degradation processes, respectively.

[0051] Within each coarse time window, state variables such as temperature, partial discharge, humidity, and load cycling are converted into dimensionless, non-negative stress rates per unit time based on the health baseline and calibration upper limit. To avoid misinterpreting a missing sensing mode as a stress reduction, each type of stress retains its cumulative value, and the increment within this window is controlled by a modality availability flag and a quality coefficient. The update method is as follows: ; ; In the formula, Indicates a coarse time window index; Indicates stress mode index, ; and They represent the first The and the first At the end of the first coarse time window The cumulative value of each stress mode; This indicates that the mode is in the first... Does the coarse time window contain data that can be used for stress calculation? This represents the corresponding data quality coefficient; Indicates the first Duration of a coarse time window; Indicates the first The unit time, dimensionless and non-negative stress rate of a certain mode; Represents the cumulative stress vector consisting of the cumulative values ​​of all modes; superscript This represents the transpose of a vector. hour, Therefore, missing measurements will not decrease the historical cumulative value; at the same time, the uncertainty of the assessment is increased by using a missing measurement mask. If the partial discharge index has been integrated by window, it should be converted into activity per unit time before being entered. .

[0052] (iii) Causal dual-timescale Mamba state coding; Both the fast-scale and slow-scale Mamba branches employ a state-space layer with an input-selective state update mechanism. The current input is used to adjust the amount of state retention and the amount of new information written, enabling the network to process continuous sequences with low computational complexity. The specific discretization implementation of the Mamba layer can use a publicly available stable implementation. The core of this invention does not lie in redefining the general Mamba recursive formula, but in constructing a causal interaction between fast and slow states oriented towards the aging mechanism of medium-voltage cable joints.

[0053] The fast-scale Mamba branch processes the fine windows within the current coarse window sequentially, focusing on representing the occurrence, duration, and recovery of events; the slow-scale Mamba branch processes long-term statistical trends in the order of the coarse windows. The two branches can use different hidden dimensions. The initial slow-scale hidden state is generated from healthy baseline features and device attributes; when historical data is lacking, a zero vector or a trained initial vector can be used. The fast branch of the k-th coarse window initializes and conditionally modulates only the slow-scale hidden state already completed in the (k-1)-th coarse window, and its hidden state within the window is not persistently stored across coarse windows.

[0054] The fast state of each fine window first obtains relative event weights and overall event gates based on explicit event intensity, persistence, modal availability, and data quality, forming a fast event representation. Then, event reliability and degradation correlation are calculated separately: event reliability reflects whether the event is supported by valid sensor data; degradation correlation reflects whether the event persists in multiple fine windows, reproduces in adjacent coarse windows, remains after load recovery, or is mutually corroborated by different modes such as temperature rise residuals, partial discharge, sheath current, or leakage current. Real but transient and reversible operational disturbances can create short-term event risks, but their reliable occurrence does not automatically write them into the long-term slow state.

[0055] After the current coarse window ends, the slow-scale Mamba branch forms an initial slow state based on the current coarse-scale features and the previous slow state. The event reliability gate and the degradation correlation gate together constitute a write-back gate, which writes reliable fast events with degradation correlation into the initial slow state through bounded mapping, and limits the state amplitude through normalization. This process can be summarized as follows: ; ; ; In the formula, Indicates a coarse time window index, superscript and These represent the fast scale and the slow scale, respectively. Indicates the first A fine-scale sequence arranged by time within a coarse time window; This represents the current coarse-scale feature formed after the end of the coarse time window; This represents the slow-scale hidden state of the previous coarse time window; and These represent the fast-scale Mamba encoding map and the slow-scale Mamba encoding map, respectively. Represents fast events; This represents the initial slow-scale hidden state before a fast event is written back; This represents an auxiliary characteristic of event reliability formed by sensor health status, data integrity rate, noise level, and modal availability; This represents the auxiliary feature of degradation correlation formed by event persistence, reproducibility of adjacent coarse windows, persistence after load recovery, and cross-modal corroboration. , and For trainable mapping matrix, and It is the bias vector; and These represent the event reliability gate and the degradation correlation gate, respectively. This represents the joint write-back gate obtained by combining the two elements one by one; Represents the Sigmoid function; This indicates vector concatenation; Represents element-wise product; Used to limit the write-back increment value; This represents a normalized mapping; This represents the slow-scale hidden state after a bounded fast event write-back is completed. The joint write-back gate only increases if the events are reliable and have a degenerate correlation.

[0056] (iv) Integration of public dimensions and aging output; Since fast and slow branches can have different hidden dimensions, two projection layers are first used to map the fast event representation and slow state to the same common dimension, and then the absolute difference and element-wise interaction are calculated. The difference term is used to represent the inconsistency between the short-term event activity level and the long-term state, and the interaction term is used to represent the degradation mode that is jointly enhanced by the two. The stress increment in the coarse-scale sequence describes the new exposure in the current window, while the modal cumulative stress used in the fusion stage describes the historical exposure up to the current moment; the two have different physical meanings.

[0057] The integration of common dimensions, continuous aging scores, and output of assessment uncertainty can be summarized as follows: ; ; ; ; In the formula, and These represent the slow-scale state projection matrix and the fast-scale characterization projection matrix, respectively. and This represents the slow-scale representation and the fast-scale representation after projection onto the same common dimension; It represents a cross-scale feature composed of fast and slow representations, absolute differences per element, and interactions per element; This indicates a common dimension fusion mapping; This represents the cumulative stress vector maintained according to the monitored modes. This indicates that it is a bounded mapping; This indicates the embedded equipment attributes formed by factors such as connector type, voltage level, insulation material, commissioning time, and installation environment; Indicates fusion characteristics; , These represent the weight vectors of the aging score output head and the uncertainty output head, respectively. , Indicates the corresponding bias; Indicates the continuous aging score; Indicates the uncertainty of the assessment; This indicates the data uncertainty caused by missing measurements, noise, and operating conditions deviating from the calibration range. This is used to ensure that the output uncertainty is positive. Coverage calibration is also required on an independent validation set.

[0058] Aging levels are categorized based on multiple ordered thresholds determined from independently validated data, such as normal, mild, moderate, severe, and critical. Thresholds can be individually calibrated for different voltage levels or connector types. Using a continuous score plus threshold approach avoids forcibly training class probabilities when sufficient level samples are lacking.

[0059] The assessment uncertainty is determined by combining the uncertainty output header with the data uncertainty, and coverage calibration is performed on an independent validation set. The data uncertainty is determined based on missing sensing modes, data completeness, sensor noise, and the degree to which operating conditions deviate from the training or healthy baseline calibration range; model ensemble or conformal prediction can be used as further calibration methods. Fixed-fold ranges of data that have not been independently validated and calibrated are not directly referred to as statistical confidence intervals.

[0060] (v) Model training and validation; Prioritize dividing the training, validation, and test sets according to the connector lifecycle identifier, ensuring that data from connectors within the same lifecycle only enters one dataset. When the sample size is limited and the data is divided by time period, a strict chronological order is adopted, and time isolation intervals that do not participate in training and evaluation are set between the training, validation, and test segments; it is prohibited to randomly assign adjacent windows of the same connector to different datasets to avoid leakage of device attributes and state continuity.

[0061] During training, continuous aging annotation error is the primary supervisory factor; samples with graded conclusions are used to supplement the constraint on the intervals in which the scores lie; in continuous intervals where no valid maintenance records exist, only abnormal decreases in long-term aging scores exceeding the allowable tolerance are penalized, without imposing monotonic constraints on short-term event risks. The allowable tolerance is determined based on adjacent evaluation intervals, sensor noise, and data integrity, thus allowing for reasonable small variations caused by measurement corrections and condition recovery.

[0062] The accumulated stress of each mode is used only as a priori for trend consistency. When a key sensing mode is missing, the previous valid accumulated value for that mode is maintained; the unknown interval is not treated as zero stress, nor is the historical accumulated value reduced accordingly, while the evaluation uncertainty is increased. When the load or environmental conditions exceed the baseline calibration range, the fusion weight of the corresponding physical prior is reduced to avoid unreliable stress estimation dominating the network output. The training sequence is split at the maintenance time, and trend constraints do not cross the valid maintenance time.

[0063] Pre-training can be performed using accelerated aging and thermal cycling data from the laboratory, followed by fine-tuning using data from confirmed defects, maintenance, and disassembly inspections in the field. When the sensor bandwidth or operating condition distribution in the laboratory and in the field is inconsistent, fine-tuning can be performed through device attribute embedding, grouping normalization, domain adaptation, or minor field annotations, without assuming that the two are completely identically distributed.

[0064] The validation phase should at least check continuous score error, classification accuracy, false positive rate, false negative rate, trend stability, uncertainty coverage, and marginal inference latency. Model weights, feature versions, health baseline parameters, classification thresholds, and uncertainty calibration parameters should be managed as a single version package.

[0065] (vi) Online assessment and post-maintenance status handling; During online operation, the fast-scale sequence and event cache within the window are updated after the fine-scale time window is completed. After the coarse-scale time window meets the termination condition, coarse-scale features, fast event representations, and new slow-scale hidden states are formed, and a formal continuous aging score and assessment uncertainty are generated. When a higher frequency of safety alerts is required on-site, short-term event risks can be output by the fast branch; these risks can decrease as operating conditions recover and are not used as formal long-term aging scores for incomplete coarse-scale windows.

[0066] In the absence of valid maintenance records, trend correction applies only to long-term aging scores, ensuring they do not fall below the previous long-term aging score minus the allowable decline amount. The allowable decline amount is not a fixed constant but adjusts with the assessment time interval, sensor noise, and data integrity rate. This trend constraint is not imposed on short-term event risks to prevent them from being permanently retained after recovery; for events proven by subsequent data to have no degradation relevance, their impact on subsequent slow states is allowed to gradually diminish through normal state updates.

[0067] Maintenance records must match the joint location identifier, lifecycle identifier, maintenance time, work order number, and re-inspection results. When a joint is completely replaced or remanufactured, a new lifecycle identifier is generated for the new joint, the commissioning time, manufacturing batch, and construction batch are updated, and the scrolling window and event cache of the old lifecycle are deleted. The cumulative stress of each mode in the new lifecycle is recalculated from the initial value, and the health baseline and initial slow-scale hidden state are reconstructed using re-inspection data and health observation period data. Old lifecycle data is retained for historical traceability but is no longer involved in the online status recursion of new joints.

[0068] Sealing repairs, tightening, or localized treatments typically cannot eliminate existing material stress exposure, thus failing to reduce the historical accumulated stress of each mode prior to maintenance. Instead, the pre-maintenance short-term event cache is cleared, the maintenance status and necessary health baseline parameters are updated, and candidate slow states are generated post-maintenance based on the maintenance type, re-inspection results, and post-maintenance observation data. These candidate slow states are updated in conjunction with the pre-maintenance slow states while preserving residual aging information, without unconditionally scaling any hidden vectors.

[0069] An observation period is set after maintenance. During the observation period, the candidate slow status after maintenance is updated gradually based on the degree of improvement after re-inspection, the recovery of temperature rise residual, the results of partial discharge or leakage current re-measurement, and the data integrity rate. If the amount of data during the observation period is insufficient, the re-inspection information is incomplete, or different test results conflict, the assessment uncertainty is increased and supplementary tests are prioritized to avoid judging that the joint has been fully restored based solely on short-term normal data.

[0070] The early warning rules consider both long-term aging risks and short-term event risks. Long-term aging risks are assessed by combining continuous aging scores, aging levels, score growth rates, number of consecutive exceedances, and the upper limit of assessment uncertainty. Short-term event risks are used to alert for partial discharge, sudden temperature rise, abnormal sheath current, or abnormal leakage current that require immediate verification. When the long-term risk is high and the uncertainty is low, a graded maintenance alert is generated; when the risk is high but the uncertainty is too large, an alert is first issued to conduct partial discharge retesting, infrared thermography, or offline insulation testing.

[0071] This embodiment is used to illustrate the technical solution of the present invention and is not intended to limit the scope of protection. The monitoring quantity type, window span, number of Mamba layers, state dimension, hierarchical threshold, and deployment location can be adjusted according to the connector structure, voltage level, and on-site sensing conditions.

[0072] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for evaluating the aging of medium-voltage cable joints using a dual-timescale Mamba state space, characterized in that, Includes the following steps: Acquire multi-source monitoring data, equipment attributes, lifecycle identifiers, and maintenance records for medium-voltage cable joints; Based on the multi-source monitoring data, the joint temperature rise residual is calculated using the health load-environment baseline model, and short-term event features are extracted based on the actual available monitoring modes in the multi-source monitoring data. A fine-scale time-scale sequence is obtained based on the joint temperature rise residual and the short-time event characteristics. A coarse-scale time-scale sequence is then obtained based on the fine-scale time-scale sequence, and cumulative stress characteristics are generated. Based on the fine-scale time sequence and the slow-scale hidden state of the previous coarse-scale time window, fast event representations and short-term event risks are obtained through the fast-scale Mamba branch. Based on the coarse timescale sequence and the slow-scale hidden state of the previous coarse time window, the initial slow-scale hidden state is obtained through the slow-scale Mamba branch. The fast event representation is written back to the initial slow-scale hidden state through event reliability and degradation correlation gating to obtain the slow-scale hidden state of the current coarse time window. The fast event representation and the slow-scale hidden state of the current coarse time window are projected onto a common dimension, and the projected fast and slow representations, the cumulative stress characteristics, and the device attributes are fused to obtain a continuous aging score. The aging level is determined based on the continuous aging score, and lifecycle management or status update is performed based on the lifecycle identifier and the maintenance record.

2. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, The health load-environment baseline model is established from early health data of joints of the same life cycle, data of healthy joints of the same type, or thermal circuit models; the joint temperature rise residual is the difference between the measured temperature rise of the medium-voltage cable joint and the expected health temperature rise output by the baseline model under the same load, ambient temperature, and available environmental variables, wherein the measured temperature rise of the medium-voltage cable joint is the difference between the measured temperature of the joint and the ambient temperature.

3. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, Based on the first time span, the joint temperature rise residual and the short-term event characteristics are used to form the fine time-scale sequence. The fine time-scale sequence includes at least the resampled monitoring quantity, the rate of change, the abnormal event characteristics corresponding to the actual available monitoring modes, the data quality marker, and the missing measurement mask. Among them, the event-type data does not generate fictitious events through interpolation, and the short-term missing continuous variables are filled by the adjacent valid data. Based on a second time span greater than the first time span, multiple time-continuous fine-time-scale sequences are statistically aggregated to form the coarse-time-scale sequence. The coarse-time-scale sequence includes at least six of the following: mean, standard deviation, quantile, trend slope, duration of exceeding limits, event occurrence rate, load rate, data integrity rate, and cumulative stress increment. A coarse time window is composed of multiple time-continuous fine-time windows.

4. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, The cumulative stress characteristics include modal cumulative values ​​formed over time by at least two of the following: temperature stress rate per unit time, partial discharge stress rate, humidity stress rate, and load cycle stress rate; the modal cumulative values ​​are updated in the following manner: ； in, Indicates a coarse time window index; Indicates stress mode index, ; and They represent the first The and the first At the end of the first coarse time window The cumulative value of each stress mode; This indicates that the mode is in the first... Does the coarse time window contain data that can be used for stress calculation? This represents the corresponding data quality coefficient; Indicates the first Duration of a coarse time window; Indicates the first The unit time, dimensionless and non-negative stress rate of a certain mode.

5. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, Both the fast-scale Mamba branch and the slow-scale Mamba branch include an input mapping layer, a state space layer with an input-selective state update mechanism, a gated residual layer, and a normalization layer. The hidden state of the fast-scale Mamba branch is only recursively processed within the current coarse time window and is reinitialized based on the previous slow-scale hidden state at the beginning of the next coarse time window. Before the current coarse time window has ended, the fast-scale Mamba branch is not modulated using the complete statistical features that have not yet been formed in the coarse time window.

6. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, In the fast-scale Mamba branch, gating convergence is performed based on the event intensity, duration, data quality, and modal availability of each fine time window within the current coarse time window to obtain the fast event representation. The gating convergence includes relative event weights and an overall event gate. The relative event weights are used to compare the importance of each fine time window within the same coarse time window, and the overall event gate is determined based on the event intensity, duration, data quality, and modal availability corresponding to the actual available monitoring modal.

7. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, The event reliability and degradation correlation are respectively formed into an event reliability gate and a degradation correlation gate, and are combined element by element into a write-back gate; the event reliability gate is determined based on sensor health status, data integrity rate, noise level and modal availability; the degradation correlation gate is determined based on at least two of the following: the persistence of the event in multiple fine time windows, the reproducibility in adjacent coarse time windows, the residual degree after load recovery, and the mutual corroboration relationship between different monitoring modes.

8. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, The fast event representation and the slow-scale hidden state of the current coarse time window are projected onto a common dimension to obtain fast representation and slow representation. The fast representation, the slow representation, the difference features and interaction features between the fast representation and the slow representation, the cumulative stress features, and the device attributes are fused together to obtain a continuous aging score between 0 and 1. The aging level is divided according to multiple ordered thresholds determined by independent verification data.

9. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, When the maintenance record does not contain valid maintenance records, trend correction is only applied to abnormal declines in the continuous aging score that exceed the allowable range. This trend correction does not apply to short-term event risks. When the maintenance record contains records of complete replacement or re-creation, a new lifecycle identifier is generated, equipment attributes are updated, the pre-maintenance scroll window and event cache are cleared, and the health baseline, accumulated stress, and slow-scale initial state are reconstructed. When the maintenance record contains records of partial repairs, the pre-maintenance historical accumulated stress is retained, the short-term event cache is cleared, and post-maintenance candidate slow states are generated based on the maintenance type, re-inspection results, and post-maintenance observation period data to update the slow-scale hidden state.

10. The method for evaluating the aging of medium-voltage cable joints in a dual-time-scale Mamba state space according to claim 1, characterized in that, It also includes generating an aging trend curve based on the time series of the continuous aging score, determining the rate of change and assessment uncertainty of the continuous aging score, and outputting graded maintenance prompts or supplementary inspection prompts by combining the aging level, the rate of change, the assessment uncertainty and the short-term event risk.

Images