A locomotive traction transformer oil-paper insulation residual life model

Abstract

A residual lifetime model for the oil-paper insulation of locomotive traction transformers is proposed. The method first integrates multi-source operation and maintenance data; second, based on advanced maintenance data, it automatically classifies health states in an unsupervised manner; then, it performs two types of predictions in parallel: predicting individualized residual lifetime based on time-series data and a state degradation model, and predicting statistical residual lifetime based on a lifetime distribution fitted from censored fault data; finally, it dynamically weights and fuses the two types of results according to data quality, outputting the final predicted value. This invention solves the problems of missing labels, incomplete data, and poor model adaptability in actual operation and maintenance, improving the accuracy, robustness, and adaptability of predictions, and providing effective support for predictive maintenance.

Description

Technical Field

[0001] This invention relates to the field of railway locomotive operation and maintenance technology, and in particular to a condition assessment and remaining life prediction model for traction transformer oil-paper insulation based on multi-source operation and maintenance monitoring and maintenance data. Background Technology

[0002] Traction transformers are critical equipment in railway locomotives, and their condition directly affects the power supply safety and operational reliability of the locomotives. With the expansion of railway scale and the increase in the number of locomotives, railway locomotive maintenance has entered a "post-maintenance era" centered on predictive maintenance. Due to long-term high-load operation, condition assessment and remaining life prediction of this component have become major technical challenges.

[0003] Currently, traction transformer condition assessment and life prediction technologies still have several limitations. In condition assessment, existing methods are heavily reliant on specific operating conditions. For example, winding condition assessment requires overvoltage or impulse voltage data, while insulating oil aging assessment requires samples obtained through multi-stress aging test platforms. Furthermore, many methods heavily depend on design data and offline experimental data, lacking online adaptability, making limited use of massive amounts of operation and maintenance data, and exhibiting low levels of multi-source information fusion, making it difficult to achieve automatic and comprehensive condition judgment. In RUL prediction, existing physical-driven models generally rely on accurate condition input, and there are differences between test conditions and actual operating environments, making it difficult to generalize component-level predictions to the part-level. Data-driven methods are limited by the scarcity of fault samples, insufficient detailed health status labels, and limited model generalization ability.

[0004] Furthermore, the railway industry's current maintenance strategy, which combines regular preventative maintenance with post-fault repair, still faces challenges such as fixed inspection cycles, discontinuous key parameter collection, missing maintenance records, and incomplete monitoring feature coverage. These issues lead to difficulties in constructing degradation trajectories and low prediction accuracy. The involvement of multiple stakeholders in maintenance data and inconsistent levels of informatization exacerbate these problems. Therefore, given relatively weak data support and an imperfect detection system, there is an urgent need to develop transformer condition assessment and remaining life prediction technologies that can integrate multi-source information and adapt to sparse and incomplete data conditions. This will enhance fault early warning and health management capabilities, supporting the intelligent operation and maintenance transformation of the railway system. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a remaining life model of the oil-paper insulation of locomotive traction transformers. By using a multi-source data fusion method, discrete time series data of insulating oil, key detection point data of advanced maintenance, and fault data during vehicle operation are integrated to overcome the problem of insufficient information from a single data source.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is: a remaining life model of oil-paper insulation for locomotive traction transformers, comprising the following steps:

[0007] Step 1: Multi-source data integration

[0008] Integrating multi-source operation and maintenance data from locomotive operation and maintenance units, and classifying them as follows:

[0009] Multidimensional time-series data: including the physicochemical properties of insulating oil and gas detection data;

[0010] Advanced overhaul maintenance record data: Includes maintenance record data of transformers before and after maintenance during advanced overhaul;

[0011] Using fault information data: including fault data of different vehicle models after advanced repairs;

[0012] Step Two: Automatic Classification of Health Status

[0013] Based on advanced maintenance records, the health status is automatically classified through the following sub-steps:

[0014] (1) Calculate the status score of each test item, set the threshold according to the data quantiles before and after maintenance, and divide the items into four categories: healthy, good, normal, and faulty.

[0015] (2) Based on the principle of integrating data-driven and mechanism-driven approaches, the weights of each detection item are determined;

[0016] (3) Based on the item scores and weights, classify the overall health status of the traction transformer;

[0017] Step 3: Predicting Remaining Life Based on Health Status Deterioration

[0018] (1) Enhance the features of multidimensional time series data to generate new features such as gas standardized residuals, gas ratios, and physicochemical determination results;

[0019] (2) Using time series data as input and health status as output, construct and train a state prediction model;

[0020] (3) Use the trained model to supplement the missing state data of the repair process and construct a complete degradation trajectory;

[0021] (4) Calculate the state transition probability and remaining lifetime based on the continuous-time Markov model;

[0022] Step 4: Remaining useful life prediction based on the useful life analysis model

[0023] (1) Through data analysis, the fault information data is divided into two categories: fault data with interval censoring and right censoring data;

[0024] (2) The lifetime data were analyzed by combining probability plots and Anderson-Darling test to determine the lifetime distribution model of the right-censored data; the distribution parameters were estimated by using the likelihood function algorithm.

[0025] (3) Calculate the remaining life probability distribution under a given service time based on the conditional probability formula.

[0026] Step 5: Adaptive Weighted Fusion

[0027] (1) Evaluate the quality of the data used in the two prediction methods, “state degradation-based” and “lifetime analysis-based”, respectively, and calculate the quality score from three dimensions: completeness, accuracy and consistency;

[0028] (2) Dynamically allocate the weights of the two types of prediction results based on the data quality score;

[0029] (3) The two types of remaining lifetime prediction results are weighted and fused to output the final remaining lifetime prediction value.

[0030] This invention proposes an adaptive classification method for the health status of traction transformers under unlabeled scenarios, based on changes in statistical characteristics of data before and after advanced maintenance. Using machine learning as a framework, a health status prediction model for traction transformers is constructed, incorporating a time-series Markov model to characterize the persistence and transition patterns of health status, thereby achieving Reliability Limitation (RUL) prediction. Furthermore, for incomplete fault information data, censoring is used to correct incomplete fault information, fitting a lifetime distribution, and using conditional failure probability to characterize RUL. Finally, an adaptive weighted fusion mechanism is constructed to integrate RUL prediction results based on lifetime and status data, dynamically adjusting weights according to data quality to improve prediction robustness and adaptability.

[0031] Intentional effects of the present invention compared with the prior art

[0032] 1. Overcoming the limitations of data fragmentation and improving prediction robustness: Addressing the issues of traditional data-driven methods relying on a single data source, providing incomplete information, and being sensitive to data gaps, this invention integrates discrete time-series data of insulating oil, key inspection point data from advanced maintenance processes, and vehicle operation fault data to construct a multi-source information complementary system. It also possesses high data robustness, maintaining stable prediction even with partial data loss, effectively solving the prediction bias problem caused by insufficient information from a single data source. This allows RUL prediction to be based on a more comprehensive and realistic data foundation, significantly improving the environmental adaptability and anti-interference capability of the prediction results. 2. Solving the pain points of unlabeled scenarios and reducing reliance on manual intervention: Traditional methods are limited by insufficient detailed health status labels, often relying on manual annotation, resulting in low segmentation efficiency and high subjectivity. This invention, based on the changes in statistical characteristics of data before and after advanced maintenance, uses a three-step algorithm—"item score calculation - weight determination - comprehensive state classification"—to achieve adaptive classification of health status in unlabeled scenarios. It can automatically update the classification criteria based on maintenance records, avoiding the limitations of manual labeling and ensuring the matching degree between the state classification and actual operation and maintenance scenarios, providing an accurate state foundation for subsequent RUL prediction. 3. Optimized censored data processing to improve the reliability of lifespan analysis: Addressing the problem that traditional methods struggle to accurately fit lifespan distributions when faced with incomplete fault information (such as right-hand censored data), this invention uses censored data correction technology to process incomplete fault information, combines statistical methods to determine the censored lifespan distribution and parameters, and derives the RUL probability distribution based on conditional failure probability. This effectively solves the lifespan analysis bias problem caused by incomplete fault data, making the lifespan model-based RUL prediction more closely match actual fault patterns and significantly improving the reliability and credibility of lifespan analysis results. 4. Integration of dual-path prediction to improve accuracy and engineering practicality: Traditional methods often rely on a single prediction model, which is prone to poor prediction stability due to data quality fluctuations. The adaptive weighted fusion mechanism of this invention can be based on data quality ( , The weights of "health status degradation prediction" and "life analysis model prediction" are dynamically adjusted, and the optimal fusion of the two results is achieved through normalization calculation. The higher the data quality, the higher the proportion of the prediction method. This avoids the limitations of a single model and can adapt to different data quality scenarios. Ultimately, it greatly improves the accuracy, robustness and engineering practicality of RUL prediction for vehicle-mounted traction transformers, and provides more reliable technical support for locomotive operation and maintenance decisions. Attached Figure Description

[0033] Figure 1 This is the overall framework diagram.

[0034] Figure 2 This is a flowchart of the automatic state partitioning method.

[0035] Figure 3 This is a flowchart for calculating the weights of each item.

[0036] Figure 4 This is a diagram illustrating the RUL prediction approach based on state partitioning. Detailed Implementation

[0037] The present invention will now be further described with reference to the accompanying drawings.

[0038] Example 1

[0039] A method and model for condition assessment and remaining life prediction of traction transformer oil-paper insulation based on multi-source data fusion is proposed. This method is applicable to situations where the construction of degradation trajectories is difficult and the prediction accuracy is low due to factors such as scarce and low-quality fault samples and insufficient health status. The overall idea of ​​this method is as follows: Figure 1 As shown, the detailed steps are as follows.

[0040] Phase S1: Data Integration

[0041] This embodiment integrates multi-source operation and maintenance data provided by the operating unit and various maintenance units, and divides the raw data into three types: "multi-dimensional time-series data", "C6 repair record data" and "fault information data".

[0042] Among them, "Multidimensional Time Series Data" refers to "HXD1B Transformer Oil Physicochemical Properties and Gas Detection Data," which includes more than 3,038 oil analysis data (including physicochemical indicators and dissolved gas data) for 149 HXD1B model locomotives, used for the study of item change patterns, correlations, and life state transitions; "C6 Overhaul Record Data" refers to "C6 Overhaul Traction Transformer Overhaul Records," which includes overhaul record data of HXD1 series locomotive transformers before and after maintenance during the C6 overhaul process, used for condition assessment; and "Fault Information Data" refers to "Application of Fault Data Records," which includes fault data of HXD1B, C, D, and HXD3C models after C5 and C6 overhauls, used for life analysis.

[0043] Phase S2: Automatic State Partitioning Method

[0044] This embodiment, driven by C6 repair record data and combined with the proposed automatic health status classification method, achieves automatic classification of the health status of the on-board traction transformer, laying the foundation for subsequent training of the state prediction model and predicting the health status of other sections. The process of the automatic state classification method is as follows: Figure 2 As shown, the detailed implementation process is as follows.

[0045] Step 1: Calculate the scores for each item

[0046] This embodiment is based on multi-dimensional operation and maintenance data of the vehicle-mounted traction transformer (including information such as dissolved gases in the insulating oil, insulating oil temperature, and particulate matter). Several evaluation items are obtained through physicochemical and chromatographic analysis, such as appearance, flash point, acid value, dielectric loss factor, water content, breakdown voltage, and other physicochemical properties, as well as chromatographic analysis results for hydrogen, methane, ethylene, ethane, and acetylene. This embodiment combines data before and after the C6 overhaul and sets... It is the 0.25 quantile. The scoring method for the single test item status of insulating oil is shown in Table 1, with the 0.75 quantile.

[0047] Table 1 Classification Method of Transformer Maintenance Items

[0048] Whether the limits are exceeded is determined based on the maintenance standards for each point under national, industry, and enterprise standards.

[0049] Step 2: Determine the weights of each item

[0050] In this embodiment, the weight values ​​of different items are calculated based on the principle of integrating data-driven and mechanism-driven approaches.

[0051] 1) Based on the rate of change of the average index values ​​before and after the advanced overhaul. The ranking is performed such that the larger the rate of change of the mean, the greater the weight, indicating that advanced repair processes can significantly improve the average performance of this indicator.

[0052] 2) Based on the ratio of the standard deviations of the index values ​​before and after the advanced overhaul. When sorting, the larger the ratio of standard deviations, the greater the weight, indicating that advanced finishing processes can significantly reduce the dispersion of this index.

[0053] 3) Increase the weight of items that frequently exceed limits in advanced repair processes;

[0054] 4) Increase the weight of items related to failures that occur during emergency repairs after advanced maintenance.

[0055] The specific process is as follows: Figure 3 As shown, the percentage normalization calculation method is as follows: , This indicates the index value corresponding to the item; the manual evaluation standard is: if an item C6 frequently exceeds the limit after repair or if an item is related to a fault after advanced repair, the manual score for that item will be appropriately increased.

[0056] Step 3: Divide the overall state

[0057] In this embodiment, the health status of the components before the C6 repair is divided into four categories: healthy, good, average, and faulty. The specific classification rules are shown in Table 2.

[0058] Table 2 Transformer Health Status Classification Rules

[0059] Phase S3: RUL Prediction Based on State Partitioning

[0060] This embodiment is based on the state classification results of the on-board traction transformer C6 before overhaul in stage S2, according to... Figure 4 The RUL prediction approach shown involves sequentially training the state prediction model, supplementing the degradation process cross-section, and calculating the Markov state transition probability to achieve RUL prediction for the vehicle-mounted traction transformer. The core process includes the following four steps:

[0061] Step 1: Feature Processing

[0062] In this embodiment, during the feature processing stage, feature engineering methods from machine learning are used to add new features such as gas-standardized residuals, representative gas ratios, and physicochemical judgment results to the original features (single point detection data of insulating oil), thereby achieving feature enhancement.

[0063] Step 2: Training the State Prediction Model

[0064] This embodiment uses multi-dimensional time-series data of the equipment (including the physical and chemical indicators of insulating oil, dissolved gas data, etc.) as input and the health status (healthy, good, normal, faulty) divided into S2 stages as output to construct a state prediction model.

[0065] In this embodiment, a Multi-Layer Perceptron (MLP) is used as the basic model during the model training phase. It learns the mapping relationship between features and health status through training with historical data. The input features consist of both original and newly added features, while the output features refer to the state classification results of the on-board traction transformer before the C6 overhaul in stage S2.

[0066] Step 3: Completion of the cross-section of the degradation process

[0067] In the degradation process section of this embodiment, to address the problem of missing status data for basic repair procedures (such as C2 and C3 repairs), a well-trained MLP model is used to predict and complete the missing basic repair procedure statuses.

[0068] In this embodiment, for the detection data of each basic repair process (the processing method is consistent with step 1), the predicted state of each basic repair process is first obtained by using the state prediction model; then, according to the time sequence of the repair process, the predicted state of these basic repair processes is concatenated with the known state of advanced repair processes (such as C6 repair) to construct a complete "time-state" degradation trajectory.

[0069] Step 4: RUL Calculation Based on State Transition Model

[0070] This embodiment quantifies the degradation pattern and calculates the RUL based on the completed degradation trajectory through a state transition model. The state definition follows the four discrete states of the S2 stage (healthy, good, average, and faulty), and based on the assumption of irreversible degradation, it stipulates that state transitions are only allowed to proceed in a worse direction.

[0071] This embodiment uses a continuous-time Markov model to model state transitions, calculates the transition rate by statistically analyzing state dwell time, constructs a transition rate matrix, and then achieves RUL prediction for the on-board traction transformer. The specific steps are as follows:

[0072] Step 4.1: Calculate the state dwell time And estimate the transfer rate

[0073] This embodiment is based on the dwell time of the same type of component in various states: if the total number of a certain type of traction transformer is At the start of the statistics (All are in good health) up to the statistical cutoff time period, These represent the times when this type of device first undergoes a state transition after the start of the statistics. Therefore, it can be considered that... arrive The time interval between This refers to the time the device remains in an excellent state, which is also the time interval required to transition from state 1 to state 2; similarly, The dwell time in the good state is the time interval between the transition from state 2 to state 3. This represents the dwell time in the general state, i.e., the time interval between transitions from state 3 to state 4. Based on the state dwell time, the transition rate parameters can be estimated as follows:

[0074] Step 4.2: Construct the transfer rate matrix

[0075] For a continuous-time Markov chain consisting of various states, its transition rate is usually expressed as a matrix. It is expressed as follows

[0076] Step 4.3: Calculate the transition probability matrix at any time.

[0077] Given the transition rate matrix, a continuous-time Markov process can be constructed from the Kolmogorov backward equation.

[0078] The general solution to the differential equation is in matrix exponential form as follows:

[0079] Through matrix exponentiation Then the transition probability matrix at any time can be obtained. Based on the transfer rate matrix, the expected total time from the current state to the fault state is calculated, which is the RUL.

[0080] Phase S4: RUL Prediction Based on Lifetime Analysis Model

[0081] This embodiment uses the integrated "fault information data" to predict its RUL (Lifetime Limiting) based on a lifetime analysis model. Key steps include... Figure 1 As shown, the specific process is as follows.

[0082] Step 1: Data Analysis

[0083] This embodiment first divides the fault information data into two categories through data analysis: fault data with interval censoring and right-censored data. Due to cost and observation time limitations, researchers cannot observe the failure of all individuals. Therefore, the collected dataset includes some non-failed individuals. For these individuals, it is only known that their failure time is later than the research deadline (i.e., the truncation time), but the exact value cannot be obtained; this situation is called right censoring.

[0084] Step 2: Determine the deletion lifetime distribution and distribution model parameters

[0085] This embodiment analyzes lifetime data using a combination of probability plots and the Anderson-Darling (AD) test to determine the lifetime distribution model (such as Weibull, Gamma, and log-normal distributions) for right-censored data, and denotes its cumulative distribution function as follows. .

[0086] For right-censored data, taking a single individual as an example, its lifespan... exist The information within the interval is unclear. Regarding the observed values... In other words, there is a lower limit. Then the likelihood function of the observed values ​​is

[0087] in For component life The probability density function of this random variable; Represents random variables In parameter vector The cumulative distribution function under these conditions; The likelihood function can be understood as a function based on the result. ,parameter The corresponding probability. Having The joint likelihood function of the dataset of independent observations is

[0088] The maximum value of the likelihood function can be understood as finding the model parameters most likely to produce the outcome based on the outcome event, and then inferring the probability of unknown events based on the model parameters.

[0089] Step 3: Predict the probability of failure

[0090] This embodiment utilizes conditional probability formulas to derive the probability distribution of remaining lifespan (time until the next fault alarm) based on service time. If a component has been used for a certain period of time... Then the next part The probability of a failure occurring within a given time period is

[0091] Phase S5: Weighted Fusion

[0092] This embodiment sets up an adaptive weighted fusion mechanism based on data quality, which integrates the state-based RUL prediction method with the life analysis model-based RUL prediction method to achieve multi-source information fusion, thereby improving the accuracy, robustness and practicality of the remaining life prediction of vehicle-mounted traction transformers. The specific steps are as follows.

[0093] Step 1: Data quality assessment, calculation and

[0094] Step 1.1: Calculation (RUL prediction data quality based on health status degradation)

[0095] For "multidimensional time-series data + automatic state segmentation results", the evaluation is carried out from three dimensions: "completeness, accuracy, and consistency". The weights of each dimension are set according to the "influence priority" (completeness 0.35, accuracy 0.35, consistency 0.3), as shown in Table 3. The weighted summation formula is as follows.

[0096] The completeness score is calculated as follows: completeness score = number of non-empty data records / total number of records to be recorded; accuracy score = number of data records that match the actual test value / number of sampled data records; consistency score = number of logically consistent data records / total number of compared data records. The range of each score is as follows: .

[0097] Step 1.2: Calculation (RUL prediction data quality based on lifetime analysis model)

[0098] Regarding the "use of fault information data," the evaluation is also conducted from three dimensions: "completeness, accuracy, and consistency," as shown in Table 3. The weighted summation formula is as follows:

[0099] The completeness score is calculated as follows: completeness score = number of records containing key fault fields (fault type, occurrence time, etc.) / total number of fault records; accuracy score = number of fault records verified through maintenance / total number of fault records; consistency score = number of records with consistent fault classification / total number of fault records. The range for each score is as follows: .

[0100] Table 3. Weights and categorization criteria for each evaluation dimension.

[0101] Step 2: Calculate adaptive weights based on data quality

[0102] according to and The calculation results are used to calculate the adaptive weights of the two RUL prediction results using a normalization method, as shown in the following formula. (10) (11)

[0103] in, These are the weights of the RUL prediction method based on health status degradation, while These are the weights of the RUL prediction method based on the lifetime analysis model. Simultaneously, the weights sum to 1 to prevent any single path from dominating the prediction due to excessively low data quality, ensuring the rationality of the fusion.

[0104] Step 3: Weighted fusion of multi-source RUL prediction results

[0105] RUL prediction results based on health status degradation Compared with RUL prediction results based on lifetime analysis model The remaining lifetime prediction result is obtained by weighting and summing according to the above adaptive weights. The formula is as follows:

[0106] This embodiment, through the design of "dimensional adaptation to the scenario + weight matching influence", allows the fusion result to combine the advantages of both paths (state tracking for dynamic degradation and lifetime analysis for static patterns) and adapt to data quality fluctuations (such as automatic reduction of weight when data is missing in a certain path), effectively improving the accuracy and robustness of RUL prediction for vehicle-mounted traction transformers.

[0107] Example 2

[0108] The difference from Example 1 is:

[0109] 1. An alternative to automatic state classification: The preferred embodiment uses a "quantile threshold" (... The system achieves label-free status classification by "calculating sub-item scores + determining weights through a dual-drive approach of data and mechanism". The core logic can be implemented through other technical paths. The sub-item scoring method can adopt the fuzzy C-means (FCM) clustering algorithm, which automatically determines 4 clusters (corresponding to "healthy-good-average-faulty") through the Bayesian Information Criterion (BIC). The membership degree of the sample to each category is calculated to adapt to the non-normally distributed maintenance data, and the consistency between the status classification and expert judgment is still ≥88%. The weight calculation method can adopt the Analytic Hierarchy Process (AHP), which determines the item weights by constructing a judgment matrix of 5-8 railway operation and maintenance experts, and can be dynamically adjusted according to the working conditions of different train models to adapt to different operation and maintenance needs.

[0110] 2. Alternative solutions for RUL prediction based on health state degradation: The preferred embodiment calculates RUL using an “MLP state prediction model + continuous-time Markov model”. The core functionality can be achieved through other models. The state prediction model can be replaced with a Long Short-Term Memory (LSTM) network or a random forest. LSTM can automatically capture long dependencies in time-series data without manual feature enhancement, while random forest is more robust to outliers to adapt to scenarios with large fluctuations in detection data. The state transition model can be replaced with a semi-Markov model (SMM) that describes state dwell time with a Weibull distribution (rather than an exponential distribution), which also conforms to the rule that “state dwell time shortens with the degree of degradation”.

[0111] 3. Alternatives to RUL prediction based on lifetime analysis: The preferred embodiment calculates the distribution parameters through maximum likelihood estimation (MLE), while the parameter estimation method can be replaced by the expectation-maximization (EM) algorithm, which is suitable for scenarios with a high proportion of censored data (>60%). Parameter convergence is achieved through iterative optimization, avoiding parameter fluctuations caused by insufficient sample size in MLE. Alternatively, the Kaplan-Meier nonparametric estimation method can be used to replace parameter distribution fitting. It does not require assuming a lifetime distribution type and directly calculates the survival function based on sample data, which is suitable for scenarios where fault samples are extremely scarce (<10).

[0112] 4. Alternative to the weighted fusion mechanism: The preferred embodiment calculates data quality based on fixed weights of "integrity 0.35, accuracy 0.35, consistency 0.3". and Fusion is achieved through normalization, and the fusion logic can be further optimized—data quality dimensions can be added based on the characteristics of the data source, such as... A new "Timeliness" dimension has been added (weight 0.2, replacing part of the consistency weight), and [the following is unclear due to incomplete text]. A new "Effectiveness" dimension has been added (weight 0.2, replacing some consistency weights); the weight calculation method can use entropy weighting instead of fixed weights, dynamically allocating weights by calculating the information entropy of each dimension (the smaller the entropy value, the higher the information utility), such as... When the entropy value of "timeliness" is low, its weight is automatically increased from 0.2 to 0.3; the fusion algorithm can use DS evidence theory to replace weighted summation, regard the dual-path RUL prediction results as "evidence", and use synthesis rules to fuse the results to deal with the uncertainty of the prediction results.

Claims

1. A model for the remaining life of oil-paper insulation in a locomotive traction transformer, characterized in that, Includes the following steps: Step 1: Multi-source data integration Integrating multi-source operation and maintenance data from locomotive operation and maintenance units, and classifying them as follows: Multidimensional time-series data: including the physicochemical properties of insulating oil and gas detection data; Advanced overhaul maintenance record data: Includes maintenance record data of transformers before and after maintenance during advanced overhaul; Using fault information data: including fault data of different vehicle models after advanced repairs; Step Two: Automatic Classification of Health Status Based on advanced maintenance records, the health status is automatically classified through the following sub-steps: (1) Calculate the status score of each test item, set the threshold according to the data quantiles before and after maintenance, and divide the items into four categories: healthy, good, normal, and faulty. (2) Based on the principle of integrating data-driven and mechanism-driven approaches, the weights of each detection item are determined; (3) Based on the item scores and weights, classify the overall health status of the traction transformer; Step 3: Predicting Remaining Life Based on Health Status Deterioration (1) Enhance the features of multidimensional time series data to generate new features such as gas standardized residuals, gas ratios, and physicochemical determination results; (2) Using time series data as input and health status as output, construct and train a state prediction model; (3) Use the trained model to supplement the missing state data of the repair process and construct a complete degradation trajectory; (4) Calculate the state transition probability and remaining lifetime based on the continuous-time Markov model; Step 4: Remaining useful life prediction based on the useful life analysis model (1) Through data analysis, the fault information data is divided into two categories: fault data with interval censoring and right censoring data; (2) The lifetime data were analyzed by combining probability plots and Anderson-Darling test to determine the lifetime distribution model of the right-censored data; the distribution parameters were estimated by using the likelihood function algorithm. (3) Calculate the remaining lifetime probability distribution for a given service time based on the conditional probability formula; Step 5: Adaptive Weighted Fusion (1) Evaluate the quality of the data used in the two prediction methods, "based on state degradation" and "based on lifetime analysis", respectively, and calculate the quality score from three dimensions: completeness, accuracy and consistency; (2) Dynamically allocate the weights of the two types of prediction results based on the data quality score; (3) The two types of remaining lifetime prediction results are weighted and fused to output the final remaining lifetime prediction value.

2. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 1, characterized in that: In step two, (2) the specific steps for determining the weight of each detection item based on the principle of data-driven and mechanism-driven fusion are as follows: 1) Based on the rate of change of the average index values ​​before and after the advanced overhaul. The ranking is performed such that the larger the rate of change of the mean, the greater the weight, indicating that advanced repair processes can significantly improve the average performance of this indicator. 2) Based on the ratio of the standard deviations of the index values ​​before and after the advanced overhaul. When sorting, the larger the ratio of standard deviations, the greater the weight, indicating that advanced finishing processes can significantly reduce the dispersion of this index. 3) Increase the weight of items that frequently exceed limits in advanced repair processes; 4) Increase the weight of items related to failures that occur during emergency repairs after advanced maintenance.

3. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 1, characterized in that: In step three, the state prediction model training stage uses a multilayer perceptron as the basic model. It learns the mapping relationship between features and health status through training with historical data. The input features are composed of original features and newly added features, while the output features refer to the state classification results of the vehicle-mounted traction transformer C6 before maintenance in step two.

4. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 1, characterized in that: In step three, (4) the specific steps for calculating the state transition probability and remaining lifetime based on the continuous-time Markov model are as follows: 1) Calculate the state dwell time And estimate the transfer rate ; （1） 2) Construct the transfer rate matrix Q; For a continuous-time Markov chain consisting of various states, its transition rate is usually represented by matrix Q as follows: （2） 3) Calculate the transition probability matrix at any time. ; Given the transition rate matrix, a continuous-time Markov process can be constructed from the Kolmogorov backward equation. （3） The general solution to the differential equation is in matrix exponential form as follows: （4） Through matrix exponentiation Then the transition probability matrix at any time can be obtained. Based on the transfer rate matrix, the sum of the expected time from the current state to the fault state is calculated, which is the remaining lifetime.

5. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 1, characterized in that: In step four, For right-censored data, taking a single individual as an example, its lifespan... exist The information within the interval is unclear; for the observed values In other words, there is a lower limit. Then the likelihood function of the observed values ​​is （5） in For component life The probability density function of this random variable; Represents random variables In parameter vector The cumulative distribution function under these conditions; This represents the likelihood function, i.e., based on the result. ,parameter The corresponding probability; have The joint likelihood function of the dataset of independent observations is （6）。 6. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 5, characterized in that: Using the conditional probability formula, derive the probability distribution of remaining lifespan based on service time; if a component has been used for a certain period of time... Then the next part The probability of a failure occurring within a given time period is （7）。 7. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 1, characterized in that: Step five, the specific steps for data quality assessment, are as follows: 1) Calculation of remaining life expectancy prediction data quality based on health status decline ; For "multidimensional time-series data + automatic state segmentation results", the evaluation is based on three dimensions: "completeness, accuracy, and consistency". The weights of each dimension are set according to the "priority of impact": completeness 0.35, accuracy 0.35, consistency 0.3; the weighted summation formula is as follows: （8） The completeness score is calculated as follows: completeness score = number of non-empty data records / total number of records to be recorded; accuracy score = number of data records that match the actual test value / number of sampled data records; consistency score = number of logically consistent data records / total number of compared data records. The range of each score is as follows: ; 2) Calculation of Remaining Life Prediction Data Quality Based on Life Analysis Model ; Regarding the "use of fault information data", it is also evaluated from three dimensions: "completeness, accuracy, and consistency". The weighted summation formula is as follows: （9） The completeness score is calculated as follows: completeness score = number of records containing key fault fields (fault type, occurrence time, etc.) / total number of fault records; accuracy score = number of fault records verified through maintenance / total number of fault records; consistency score = number of records with consistent fault classification / total number of fault records. The range for each score is as follows: .

8. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 7, characterized in that: The specific steps for dynamically assigning weights to the two types of prediction results based on the data quality score are as follows: according to and The calculation results are used to calculate the adaptive weights of the two remaining lifetime prediction results using a normalization method, as shown in the following formula. （10） （11） in, It is the weight of the remaining lifespan prediction method based on the decline in health status, while It is the weight of the remaining lifetime prediction method based on the lifetime analysis model; at the same time, the weight sum is 1 to avoid a certain path from dominating the prediction due to low data quality, and to ensure the rationality of the fusion.

9. The remaining life model of the oil-paper insulation of a locomotive traction transformer according to claim 8, characterized in that: The specific steps for weighted fusion of the two types of remaining lifetime prediction results are as follows: Life expectancy prediction results based on health status decline Comparison with remaining lifetime prediction results based on lifetime analysis model The remaining lifetime prediction result is obtained by weighting and summing according to the above adaptive weights. The formula is as follows: （12）。

Images