Device residual life dynamic prediction method and device, equipment and medium

Abstract

This invention discloses a method, apparatus, device, and medium for dynamically predicting the remaining life of equipment. The method includes: constructing a historical feature matrix based on historical multi-dimensional feature data and historical remaining life values, determining initial feature weights and historical fluctuation confidence intervals; constructing a real-time feature matrix and a real-time weight set during equipment operation; monitoring feature change amplitudes and comparing them with historical fluctuation confidence intervals to dynamically adjust feature weights; adaptively processing real-time multi-dimensional feature data by combining feature change amplitudes and the real-time weight set; generating model input data and inputting it into a dynamic prediction model; and outputting the remaining life prediction result of the equipment. This invention, by introducing multi-dimensional features and dynamically adjusting feature weights, enables the model to adaptively update the input data structure according to changes in operating conditions, thereby more accurately depicting the equipment degradation process and improving the accuracy and stability of remaining life prediction.

Description

Technical Field

[0001] This invention relates to the field of machine learning and state prediction technology, and in particular to a method, apparatus, device and medium for dynamic prediction of remaining equipment life. Background Technology

[0002] Equipment remaining life prediction is a crucial technical means to ensure safe equipment operation and optimize maintenance strategies. This is especially true for equipment with significant degradation characteristics, such as energy storage batteries, where the prediction results directly impact the reliability of operational decisions. Existing life prediction methods typically model and extrapolate equipment degradation trends based on historical and real-time data during operation to obtain the predicted remaining life.

[0003] In existing technologies, some lifespan prediction methods primarily focus on modeling and analyzing a single or limited number of key indicators of the equipment. For example, using the change in energy storage battery capacity as the core characteristic, battery lifespan is predicted by analyzing the rate of capacity decay over time and its changing patterns. While these methods can reflect the overall degradation trend of the equipment to some extent, their modeling foundation heavily relies on a single performance indicator and fails to fully incorporate multi-dimensional operational characteristics such as temperature, current, voltage, state of charge, and internal resistance changes. This makes it difficult to comprehensively characterize the complex degradation behavior of the equipment under different operating conditions. When the equipment's operating environment or load conditions change, lifespan models built based on a single feature are prone to insufficient information, leading to predictions that deviate from the actual degradation state.

[0004] Furthermore, existing technologies generally have shortcomings in characterizing and dynamically adjusting feature importance. Some schemes, after determining feature parameters through statistical analysis or factor decomposition during the modeling phase, often assume relatively stable feature contribution relationships during subsequent prediction, adapting to changes in operating conditions only by validating real-time data or selecting different analysis orders. This approach lacks the ability to characterize the dynamic evolution of feature weights with changing operating conditions, making it difficult to continuously reflect the real-time impact of different features on lifespan degradation during equipment operation. When equipment experiences load fluctuations, sudden changes in ambient temperature, or switching of operating modes, the relative importance between features may change significantly, making it difficult for models with fixed or weak adjustment mechanisms to maintain prediction accuracy. Summary of the Invention

[0005] The main objective of this invention is to provide a method, apparatus, device, and storage medium for dynamic prediction of remaining equipment lifespan. This invention aims to solve the technical problem that existing technologies, due to the lack of real-time mapping and dynamic weight adjustment mechanisms for the contribution relationships of multi-dimensional operating characteristics, are unable to continuously and accurately reflect the true lifespan degradation state of equipment under conditions of changing operating conditions and non-stationary degradation processes.

[0006] To achieve the above objectives, the present invention provides a method for dynamically predicting the remaining life of equipment, comprising: Obtain historical multi-dimensional feature data of the device and the corresponding historical remaining lifespan value, and construct a historical feature matrix based on the historical multi-dimensional feature data; Based on the historical feature matrix and the historical remaining lifetime value, the initial weight value of each feature is determined, and the historical fluctuation confidence interval of each feature is calculated based on the historical feature matrix. Real-time multi-dimensional feature data during device operation is acquired, and a real-time feature matrix is ​​constructed based on the real-time multi-dimensional feature data; Using the initial weight values, initialize the real-time weight set for each feature; Monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval; Based on the comparison results, the weight values ​​of each feature in the real-time weight set are dynamically adjusted to obtain the real-time weight set at the current moment. Based on the variation magnitude of each feature and the real-time weight set, the real-time multi-dimensional feature data is adaptively processed to generate model input data. The model input data is input into the dynamic prediction model to perform prediction calculations and output the remaining life prediction results of the device.

[0007] Furthermore, to achieve the above objectives, the present invention provides a device for dynamic prediction of remaining equipment life, comprising: The historical feature construction module is used to acquire the device's historical multi-dimensional feature data and corresponding historical remaining lifetime value, and to construct a historical feature matrix based on the historical multi-dimensional feature data. The feature weight initialization module is used to determine the initial weight value of each feature based on the historical feature matrix and the historical remaining lifetime value, and to calculate the historical fluctuation confidence interval of each feature based on the historical feature matrix. The real-time feature construction module is used to acquire real-time multi-dimensional feature data during the operation of the device and construct a real-time feature matrix based on the real-time multi-dimensional feature data. The weight set initialization module is used to initialize the real-time weight set of each feature using the initial weight values; The feature change monitoring module is used to monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval. The weight dynamic update module is used to dynamically adjust the weight values ​​of each feature in the real-time weight set based on the comparison results, so as to obtain the real-time weight set at the current moment. The model input adaptive generation module is used to adaptively process the real-time multi-dimensional feature data based on the change magnitude of each feature and the real-time weight set to generate model input data. The remaining lifetime prediction module is used to input the model input data into the dynamic prediction model to perform prediction calculations and output the remaining lifetime prediction results of the device.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a device remaining lifetime dynamic prediction program stored in the memory and executable on the processor, wherein when the device remaining lifetime dynamic prediction program is executed by the processor, it implements the steps of the device remaining lifetime dynamic prediction method as described above.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a device remaining lifetime dynamic prediction program, wherein when the device remaining lifetime dynamic prediction program is executed by a processor, it implements the steps of the device remaining lifetime dynamic prediction method as described above.

[0010] Beneficial Effects: This invention discloses a method, apparatus, device, and medium for dynamic prediction of remaining equipment lifespan, comprising: constructing a historical feature matrix based on historical multi-dimensional feature data and historical remaining lifespan values, determining initial feature weights and historical fluctuation confidence intervals; constructing a real-time feature matrix and a real-time weight set during equipment operation; monitoring feature change amplitudes and comparing them with historical fluctuation confidence intervals to dynamically adjust feature weights; adaptively processing real-time multi-dimensional feature data by combining feature change amplitudes and the real-time weight set; generating model input data and inputting it into a dynamic prediction model; and outputting the remaining lifespan prediction result of the equipment. This invention, by introducing multi-dimensional features and dynamically adjusting feature weights, enables the model to adaptively update the input data structure according to changes in operating conditions, thereby more accurately depicting the equipment degradation process and improving the accuracy and stability of remaining lifespan prediction. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for the dynamic prediction method of equipment remaining life in one embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the dynamic prediction method for remaining equipment lifespan according to the present invention. Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the device for dynamic prediction of remaining lifespan of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] The dynamic prediction method for remaining equipment life provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can construct a historical feature matrix based on historical multi-dimensional feature data and historical remaining lifetime values, and determine the initial feature weights and historical fluctuation confidence intervals. During device operation, it constructs a real-time feature matrix and a real-time weight set, monitors feature changes, and compares them with historical fluctuation confidence intervals to dynamically adjust feature weights. Combining feature change magnitudes and the real-time weight set, it adaptively processes real-time multi-dimensional feature data, generates model input data, inputs it into a dynamic prediction model, and outputs the device's remaining lifetime prediction result. This invention, by introducing multi-dimensional features and dynamically adjusting feature weights, enables the model to adaptively update the input data structure according to changes in operating conditions, thereby more accurately depicting the device degradation process and improving the accuracy and stability of remaining lifetime prediction. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this invention.

[0014] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the dynamic prediction method for remaining equipment life provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0015] like Figure 2 As shown, the dynamic prediction method for remaining equipment life proposed in this invention includes the following steps: S10, acquire the historical multi-dimensional feature data of the device and the corresponding historical remaining lifespan value, and construct a historical feature matrix based on the historical multi-dimensional feature data; In this embodiment, the historical multi-dimensional feature data and corresponding historical remaining lifetime values ​​of the device are obtained. The historical multi-dimensional feature data comes from the device's operation records, containing multiple observable variables with timestamps or sample numbers. The corresponding historical remaining lifetime value is a lifetime label matching each historical record. A one-to-one correspondence is established through timestamp alignment or sample number alignment to avoid feature-label mismatch. Preprocessing of the historical multi-dimensional feature data covers outlier handling, missing value handling, duplicate handling, and time alignment. Outlier handling identifies and removes or replaces out-of-bounds points based on physical or statistical boundaries. Missing value handling performs interpolation to complete or fill missing segments. Duplicate handling removes duplicates based on timestamps and retains valid records. Time alignment maps variables with different sampling frequencies to a unified time raster to form a multi-dimensional vector at the same moment. The historical feature matrix is ​​constructed based on the preprocessed data. The historical feature matrix is ​​expressed in a two-dimensional structure, with rows corresponding to time points or samples, columns corresponding to feature variables, and cells filled with preprocessed feature values. The column order and column names are fixed, allowing the matrix to be directly read and accessed.

[0016] Historical multi-dimensional feature data can be queried and batch-loaded from the running database by time range, or it can be parsed from log files and merged by timestamp; outlier handling can be done by upper and lower limit pruning or deviation judgment based on sliding window; missing data handling can be done by linear interpolation, forward padding, or segmented completion; duplication handling can be done by deduplication based on the consistency of timestamp and variable set; time alignment can be done by fixed interval resampling or by nearest neighbor matching of other variables based on the timestamp of the main variable; historical feature matrix can be organized by single time point or by aggregation by fixed window and then organized by window as rows.

[0017] This embodiment establishes a correspondence between historical multi-dimensional feature data and historical remaining lifetime values, and constructs a historical feature matrix after processing for anomalies, missing data, duplicates, and time alignment. This ensures that historical data maintains consistency in temporal semantics and structural expression, reduces interference from noise points, missing segments, and asynchronous sampling on matrix construction, and improves the usability and comparability of historical data.

[0018] S20, based on the historical feature matrix and the historical remaining lifetime value, determine the initial weight value of each feature, and calculate the historical fluctuation confidence interval of each feature based on the historical feature matrix; In this embodiment, the initial weight values ​​for each feature are determined based on the historical feature matrix and historical remaining lifetime values. The historical feature matrix provides the value distribution of each feature across historical samples, and the historical remaining lifetime values ​​provide the lifetime label for each sample. The two form a supervised association through the corresponding sample row index. The initial weight values ​​are used to characterize the contribution of each feature to the lifetime label. The initial weight values ​​can be determined by constructing a regression map and extracting the feature contribution. The regression map takes the historical feature matrix as input and the historical remaining lifetime values ​​as output. After training, a contribution score is calculated for each feature. The contribution score can be derived from the cumulative split gain within the model, the sensitivity to error increment, or the loss change caused by permutation perturbation. The contribution score is then normalized to obtain the initial weight value, thereby making the weights of different features comparable. In calculating the historical fluctuation confidence interval for each feature based on the historical feature matrix, the historical feature matrix is ​​traversed column by column. Each column corresponds to a historical time series or historical sample series of a feature. First, the central statistic and discrete statistic of the column are calculated. The central statistic can be the mean or median, and the discrete statistic can be the standard deviation or quantile. Then, the central statistic and discrete statistic are combined to generate upper and lower boundaries. The upper boundary and the lower boundary together constitute the historical fluctuation confidence interval. The historical fluctuation confidence interval expresses the allowable fluctuation range of the feature in the historical period in the form of a numerical range and is bound to the feature name for storage for subsequent retrieval.

[0019] This embodiment establishes a feature contribution metric and generates initial weight values ​​by using a historical feature matrix and historical remaining lifetime values, making the importance of each feature quantifiable. At the same time, it generates a historical fluctuation confidence interval for each feature by using the historical feature matrix, making the feature fluctuation range have a searchable historical benchmark, thereby establishing a unified measurement basis for subsequent weight-based and fluctuation-based processing.

[0020] S30, acquire real-time multi-dimensional feature data during the operation of the device, and construct a real-time feature matrix based on the real-time multi-dimensional feature data; In this embodiment, real-time multi-dimensional feature data acquired during device operation is represented by data originating from the current operating state of the device rather than historical archives. Multi-dimensional data simultaneously covers multiple physical or operational attributes. Data acquisition can be achieved through continuous reception via sensor networks, controller interfaces, or communication buses, with the acquisition frequency determined by the device's operating rhythm or event triggering conditions. Preprocessing of the real-time multi-dimensional feature data aims to eliminate noise, missing data, and scale differences introduced during real-time acquisition, enabling joint modeling of data from different sources and with different scales. The preprocessing process includes outlier identification and removal, numerical completion of missing segments, and scale unification of feature values. Outlier identification can be based on instantaneous jump amplitude or deviation from neighboring statistics. Missing data completion can be achieved through interpolation of adjacent time points or short-term prediction. Scale unification can be achieved through linear normalization or standard score transformation. After obtaining the real-time preprocessed feature data, the real-time feature matrix is ​​constructed by combining the features of each dimension corresponding to the same time point into a row, and arranging the continuous time points in chronological order to form a two-dimensional structure. This allows the real-time feature matrix to reflect the evolution of time in the row dimension and the distribution of features in the column dimension. This matrix structure provides a data carrier for subsequent joint processing based on time and features.

[0021] This embodiment preprocesses real-time multi-dimensional feature data and constructs a real-time feature matrix with a unified structure, enabling real-time operational information to form a continuous expression under conditions of noise suppression, scale consistency, and time alignment. This provides a stable and computable data foundation for subsequent analysis and prediction based on real-time status.

[0022] S40, Initialize the real-time weight set of each feature using the initial weight value; In this embodiment, the real-time weight set of each feature is initialized using initial weight values. These initial weight values ​​are derived from the correlation analysis results between the historical feature matrix and the historical remaining lifetime values, representing the relative influence of different features in the early stages of lifetime evolution. The core of the initialization is to transform the statically obtained weight information into a weight carrier that can be updated with the running state. The real-time weight set represents a data structure continuously maintained during operation, with each item corresponding one-to-one with a specific feature, used to store the weight value of that feature at the current moment. The initialization process is not a simple copy; instead, the initial weight values ​​are mapped and written into the real-time weight set according to the feature identifier, and a unified data access interface is reserved for subsequent dynamic adjustments, giving the real-time weight set readable, writable, and updatable attributes, thus becoming the starting state for runtime re-evolution.

[0023] Real-time weight sets can be implemented as vectors, mapping tables, or indexed arrays. Different features are associated with initial weight values ​​through names, numbers, or position indices. During initialization, all initial weight values ​​can be loaded directly, or they can be loaded on demand when the corresponding feature data is first detected. To address differences in operating environments, boundary constraints or numerical smoothing can be applied to the weight values ​​during the initialization phase to prevent extreme weights from causing instability in subsequent calculations.

[0024] This embodiment establishes a real-time weight set consistent with historical analysis results in the early stage of operation, so that each feature has a clear and traceable initial degree of influence when entering the dynamic operation stage, thereby ensuring the continuity and controllability of the subsequent weight update process.

[0025] S50, monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval; In this embodiment, the real-time changes of each feature in the real-time feature matrix are monitored, with the core being the continuous perception of the dynamic fluctuations of feature values ​​over time during operation. The real-time feature matrix organizes multi-dimensional feature data in chronological order, with each column corresponding to a specific feature and each row corresponding to a time point. The monitoring process involves reading newly entered data rows in the matrix and associating them with the values ​​of the same feature at previous time points, thereby forming a continuous time series.

[0026] The calculation of the magnitude of change is based on a preset time window, which limits the range of data traced back and is used to characterize the intensity of characteristic fluctuations within a short time scale. The characteristic sequences within the window can be used to calculate differences, extreme value spans, or statistical dispersion, thereby obtaining quantitative results that reflect the level of characteristic fluctuations in the current operating phase.

[0027] The historical fluctuation confidence interval is derived from the statistical analysis results in the historical feature matrix. Its upper and lower boundaries characterize the reasonable fluctuation range of the feature under historical operating conditions. By comparing the real-time calculated change amplitude with the corresponding historical fluctuation interval, it can be determined whether the current fluctuation is within the range allowed by historical experience, thus providing a status basis for subsequent processing.

[0028] This embodiment enables the system to distinguish between normal fluctuations and abnormal offset states by continuously quantifying characteristic fluctuations during operation and comparing them with historical intervals, thus providing a stable and clear basis for subsequent dynamic processing.

[0029] S60, Based on the comparison results, dynamically adjust the weight values ​​of each feature in the real-time weight set to obtain the real-time weight set at the current moment; In this embodiment, when the calculated change in a certain feature within a preset time window exceeds the historical fluctuation confidence interval, it means that the behavior of that feature under current operating conditions has deviated from historical statistical patterns, and its impact on lifetime degradation may have changed. At this point, it is necessary to reassess the contribution of that feature under the current operating conditions. The change in contribution is used to characterize the relative trend of the strength of the feature's influence on the remaining lifetime prediction result. Its source can be the change in the correlation between the feature's numerical change and the lifetime degradation trend, or the increase or decrease in influence derived based on the historical model structure.

[0030] The real-time weight set stores the weight values ​​of each feature at the current moment, reflecting the relative importance of different features in the prediction calculation. When a change in contribution is calculated, this change is combined with the current weight value of the corresponding feature, and a new weight value is formed through additive or proportional updates, thus adjusting the weights according to changes in operating conditions. For features whose changes do not exceed the historical fluctuation confidence interval, their operating status remains within the historical statistical range, and the original weights can continue to reflect the stable influence of the feature; therefore, the weights remain unchanged. After the above processing, the weights of all features are uniformly organized to form the real-time weight set corresponding to the current moment.

[0031] This embodiment dynamically adjusts the weights when feature fluctuations exceed the historical range, enabling the weight allocation to evolve with changes in operating status. This continuously reflects the actual impact of each feature under different operating conditions, while avoiding ineffective adjustments to stable features, thus improving the timeliness and rationality of the weight set.

[0032] S70, based on the change magnitude of each feature and the real-time weight set, adaptive processing is performed on the real-time multi-dimensional feature data to generate model input data; In this embodiment, the magnitude of change of each feature reflects the degree of fluctuation of the feature relative to historical behavior within the current time scale, and the real-time weight set characterizes the relative importance of different features to lifetime prediction under the current operating state. Combining the magnitude of change with the real-time weight set allows for targeted adjustment of the feature data before it enters the prediction model. Adaptive processing is applied to the real-time multi-dimensional feature data or the corresponding real-time preprocessed feature data itself. By introducing weight adjustment and structural adjustment, the feature data is made to better meet the prediction requirements under the current operating conditions at both the numerical and combinatorial levels.

[0033] During processing, the magnitude of change is used to determine the overall stability of the operating state and the activity level of individual features, while the real-time weight set is used to determine the proportion of influence of each feature in the model input. Weighting, filtering, or enhancement operations are performed on real-time multi-dimensional feature data or corresponding real-time preprocessed feature data to make features with drastic changes and high weights occupy a more prominent position in the model input, while features with smaller changes or lower weights have a correspondingly weaker impact on the model input. The model input data generated in this way maintains the integrity of the original information while better reflecting the current operating state.

[0034] This embodiment adaptively processes the feature change magnitude with real-time weights, enabling the model input to dynamically adjust with changes in operating status. This helps to highlight more influential feature information under the current operating conditions, suppress redundant or weakly correlated features, and thus improve the relevance and effectiveness of the input data for lifetime prediction.

[0035] S80, input the model input data into the dynamic prediction model to perform prediction calculations, and output the remaining life prediction result of the device.

[0036] In this embodiment, the model input data is a structured dataset formed after preprocessing, which has been adapted to the current operating state in terms of numerical scale, feature contribution relationship, and combination form. The dynamic prediction model is used to characterize the mapping relationship between feature evolution over time and lifetime changes. It contains a computational structure capable of handling temporal dependencies and a mapping unit for outputting continuous prediction results. Inputting the model input data into the dynamic prediction model means sending the input data into the model's computational path according to the data interface and time dimension organization method defined by the model, enabling the model to perform prediction calculations based on historically learned degradation patterns and the current input state.

[0037] During the prediction calculation process, the model performs multi-layered calculations and parameter mapping on the input data, progressively extracting implicit representations reflecting degradation trends, and converting these representations into numerical outputs of remaining lifetime. The output equipment remaining lifetime prediction results are given in the form of continuous numerical values ​​or time lengths, used to characterize the sustainable operating time of the equipment in its current state.

[0038] In implementation, the input data can be organized into a fixed-dimensional vector or time series format according to the input specifications of the dynamic prediction model, and then passed to the model calculation module. The model can use a loop structure, a gating structure, or other computational units capable of handling time correlations to complete the prediction calculation. After the calculation is completed, the corresponding predicted value is read from the model output and output or stored as the remaining lifetime prediction result of the equipment for subsequent decision-making.

[0039] This embodiment feeds the adaptively processed input data into the dynamic prediction model, enabling the prediction calculation to be completed based on the combined effect of the current operating state and historical degradation patterns. This improves the responsiveness of the remaining life prediction results to real-time operating condition changes and enhances the overall prediction stability.

[0040] In one embodiment, step S10 above includes: S101, extract historical multi-dimensional feature data and corresponding historical remaining lifespan values ​​from the equipment's historical operation database; S102, the historical multi-dimensional feature data is subjected to anomaly detection using the standard deviation multiple statistical criterion, and numerical points that deviate from the mean by more than a preset multiple of the standard deviation are identified and removed to obtain preliminary cleaned data; S103, perform interpolation to complete or remove duplicates from the missing segments or duplicate records in the preliminary cleaned data to obtain the cleaned historical feature data. S104, The cleaned historical feature data is transformed into dimensionless form using the standard fraction standardization method to eliminate dimensional differences and obtain the preprocessed data; S105, the preprocessed data is arranged and aligned according to the timestamp order of the preprocessed data, and the arranged and aligned data is organized into a two-dimensional matrix in which rows correspond to time points and columns correspond to feature variables, so as to construct a historical feature matrix.

[0041] In this embodiment, the historical multi-dimensional feature data comes from the equipment's historical operation database. This database can be formed by an equipment-side recording system, a host computer acquisition system, or an operation and maintenance platform. The data content covers multiple sensor channels or multiple operational indicator dimensions. The historical remaining lifetime value comes from the lifetime annotation results or lifetime inference results under the same time reference as the historical multi-dimensional feature data. The two are linked through timestamps, batch identifiers, cycle counts, or operational stage identifiers, so that each feature record can be associated with a historical remaining lifetime value. The extraction process uses field mapping and filtering rules to limit the extraction scope. Field mapping is used to determine the column name, dimension, and sampling granularity of each feature variable in the database. Filtering rules are used to limit conditions such as equipment number, operating range, operating condition type, and data completeness, thereby obtaining structured historical multi-dimensional feature data and a matching set of historical remaining lifetime values, providing source data input for subsequent processing.

[0042] Anomaly detection is based on the standard deviation multiple statistical criterion. The key to this criterion is using the mean and standard deviation to characterize the central position and dispersion of feature variables within their historical distribution, and using a preset multiple of the standard deviation to form the deviation judgment boundary. The calculation process establishes a statistical window or a full statistical set for each feature variable. The statistical window can be segmented by time, by operating stage, or by working condition. The mean and standard deviation are calculated within the corresponding segments to reduce threshold drift caused by cross-working-condition mixing. For each numerical point, the deviation from the mean is calculated, and it is determined whether the deviation exceeds a preset multiple of the standard deviation. Numerical points exceeding the boundary are identified as outliers and removed. Removal actions include setting the numerical point to null, deleting it from the sequence, or recording it as an invalid marker. The removal results form preliminary cleaned data. This processing isolates outliers caused by obvious jump noise, sensor transient distortion, and communication jitter from the training data, preventing outliers from being amplified and propagated in subsequent alignment and standardization stages.

[0043] The processing of missing segments and duplicate records revolves around the initial data cleaning. Missing segments are identified by consecutive timestamp gaps, consecutive invalid markers, or insufficient sampling points. Duplicate records are identified by duplicate timestamps, primary keys, or feature vectors. Interpolation completion is performed on missing segments. Interpolation strategies can be selected based on the characteristics of the feature variables, such as linear interpolation, spline interpolation, forward imputation, or neighborhood-based statistical interpolation. The selection criteria include the rate of change of the feature variables, the existence of physical boundaries, and whether they are periodic. The interpolation results need to maintain the monotonicity of the time series and control the interpolation span to avoid filling the real abrupt change interval with interpolations that span too long. Deduplication is performed on duplicate records. Deduplication rules can be selected to retain the earliest record, retain the latest record, select the best record based on data quality score, or aggregate duplicate records. The aggregation method can be to take the mean, median, or weighted mean to reduce the bias introduced by duplicate reporting. After interpolation completion and deduplication, the cleaned historical feature data is formed. The data has enhanced continuity and reduced redundancy on the time axis, providing a controllable input for subsequent dimensionless and matrix organization.

[0044] The dimensionless transformation is performed using standard score standardization. The basic form of standard score standardization is to center and scale the feature variables using their mean and standard deviation, mapping feature variables of different dimensions to a comparable numerical scale. This reduces the unbalanced impact of variables with different dimensions, such as temperature, current, and voltage, on the gradient and loss contributions during model training. In implementation, the mean and standard deviation are calculated for each feature variable dimension on the cleaned historical feature data. The mean and standard deviation can be calculated on the full historical data or on subsets of different operating conditions to maintain scale consistency. The calculated parameters need to be saved along with the data for subsequent reuse and consistency verification. For each sampling point, a subtraction of the mean and division by the standard deviation are performed. For feature variables with a standard deviation close to zero, protection rules need to be configured, such as setting a minimum standard deviation lower limit, skipping that dimension, or replacing it with a fixed scaling factor to avoid numerical instability. After standard score standardization, the preprocessed data is obtained. This data is scale-uniform and has more stable statistical properties, providing normalized input for constructing the historical feature matrix.

[0045] The arrangement and alignment process revolves around the preprocessed data, ordered by timestamp. This timestamp order defines the sequential relationship between time-series samples and supports subsequent time-related modeling. The arrangement sorts each record in ascending order by timestamp and aligns them when there are inconsistencies in sampling frequencies from multiple sources. Alignment methods can include resampling, time window aggregation, or nearest neighbor matching, mapping feature variables of different frequencies to a unified time grid. The aligned data must maintain a complete set of feature vectors for each time point. Missing dimensions can be reverted to the aforementioned interpolation completion rules or filled with missing markers and recorded during the alignment phase. After arrangement and alignment, the data is organized into a two-dimensional matrix. Rows correspond to time points to represent the time-series sample sequence, and columns correspond to feature variables to represent the multi-dimensional feature space. A mapping relationship is established between the matrix's row and column indices and the feature name table, facilitating subsequent weight calculations, fluctuation interval statistics, and column-level feature processing during model training. The generation of the historical feature matrix transforms discrete historical records into a structured tensor entry point, preserving both the temporal and feature structures, thus establishing a unified data foundation for subsequent matrix-based statistical calculations and supervised training.

[0046] This embodiment extracts and binds historical multi-dimensional feature data and historical remaining lifetime values ​​from a historical operational database to form a supervised learning data correspondence. Then, outliers are removed by standard deviation multiple statistical criteria, and missing segments and duplicate records are interpolated or deduplicated to improve the continuity and effectiveness of the data on the time axis. Subsequently, standard score standardization is used to achieve cross-dimensional unification, and the historical feature matrix is ​​constructed after being arranged and aligned in timestamp order. This allows subsequent training and statistical calculations to be performed on a unified scale, a unified time base, and a unified matrix structure, thereby reducing the interference of abnormal noise and scale differences on modeling parameters and improving the usability of historical data.

[0047] In one embodiment, step S20 above includes: S201, Construct a random forest regression model, using the historical feature matrix as input data and the historical remaining lifetime value as output label; S202, the random forest regression model is trained under supervision using the historical feature matrix and the historical remaining lifetime value, and the average gain value generated by each feature during the model training process is calculated. S203, normalize the average gain value to convert it into contribution scores for each feature; S204, assign the contribution score to the corresponding feature as the initial weight value; S205, For each feature in the historical feature matrix, calculate the mean parameter and standard deviation parameter of the feature on the historical time series; S206, Based on the mean parameter and the standard deviation parameter, set the upper limit boundary and lower limit boundary of the allowable fluctuation; S207, the numerical range formed by the upper limit boundary and the lower limit boundary is defined as the historical fluctuation confidence interval of the feature.

[0048] In this embodiment, a random forest regression model is used to establish a nonlinear mapping relationship between the historical feature matrix and historical remaining lifetime values. The modeling object uses the column-corresponding feature variables of the historical feature matrix as input dimensions and the historical remaining lifetime values ​​as supervision signals. During the construction phase, the number of trees, the maximum depth of a single tree, the minimum number of samples for node splitting, the feature subsampling ratio, and the sample bootstrapping ratio are determined to achieve a controllable balance between fitting ability and generalization ability. Input data is fed into the model in the form of row vectors of the historical feature matrix, with each row corresponding to a time point or a sample unit, and each column corresponding to a feature variable. The output labels are aligned one-to-one with the input row vectors, enabling the training process to calculate errors at the sample granularity and drive parameter updates.

[0049] Supervised training is performed using a joint training set of the historical feature matrix and historical remaining lifetime values. During training, each tree grows on a subset of samples obtained through bootsampling, and at node splits, splitting features that can reduce error are selected from the candidate feature subset. The average gain value is derived by summarizing the gain contribution of each feature across all trees and splitting nodes. The gain can be characterized by the decrease in loss or variance before and after splitting. During summarization, the contributions of the same feature across different trees and different levels of nodes are accumulated, and the accumulated results are normalized and scaled according to the number of trees or the number of splits to obtain an average gain value that can be compared across features. This statistical process requires clarifying the correspondence between feature identifiers and column indices to avoid confusion in feature contribution attribution under resampling and feature subsampling conditions, while retaining the source of the cumulative contribution of each feature for subsequent consistency verification.

[0050] Normalization of average gain values ​​is used to transform gain results of different orders of magnitude into contribution scores on a uniform scale. Normalization can be performed using a sum-of-the-parts normalization method, dividing the average gain value of each feature by the sum of the average gain values ​​of all features, thus placing the contribution scores in an interpretable scale space; or a maximum-value normalization method, dividing the average gain value of each feature by the maximum average gain value, thus placing the contribution scores in an interval scale space. During normalization, dimensions with average gain values ​​of zero or close to zero need to be handled to avoid numerical instability caused by excessively small denominators. After normalization, a consistency check is performed on the contribution scores, such as checking whether the scores satisfy non-negativity and whether they meet the expected normalization constraints.

[0051] When the contribution score is assigned as the initial weight value, an index table or key-value mapping between features and weights needs to be established to ensure that the initial weight value is strictly aligned with the columns in the historical feature matrix. The assignment process includes not only writing the numerical value but also naming and binding the weights and recording their versions. Naming and binding are used to read the corresponding weights based on the same-named features in subsequent real-time calculations, while version recording is used to trace the source batch of the initial weight values ​​after model iterations or data updates. To prevent a feature from having an excessively large or small weight due to sampling anomalies, truncation or smoothing rules can be introduced before and after the assignment. For example, upper and lower limits can be set for the contribution score, or neighborhood average smoothing can be used to ensure that the initial weight value has a controllable numerical range.

[0052] The calculation of historical volatility confidence intervals is based on a historical feature matrix. For each feature, a one-dimensional sequence is generated on the historical time series, and then the mean and standard deviation parameters are calculated. The mean parameter characterizes the centrality of the feature within the historical sample set, while the standard deviation parameter characterizes the dispersion of the feature within the historical sample set. The calculation requires specifying the time series range and sample granularity, such as calculating based on the full historical time series, or segmenting by operating phase or operating condition type. Segmented calculation can reduce mean drift and standard deviation inflation caused by cross-state mixing. For features exhibiting trend drift, the mean and standard deviation parameters can be calculated within a sliding time window, allowing the confidence interval to be updated as historical phases change, while still maintaining the interval's closeness to the actual volatility range.

[0053] Upper and lower bounds define the allowable range of fluctuations. These bounds can be derived from a combination of mean and standard deviation parameters; for example, a two-sided range can be formed by adding or subtracting a preset multiple from the mean parameter. The preset multiple controls the interval width; a larger multiple allows for a wider fluctuation range, while a smaller multiple increases sensitivity to abnormal fluctuations. The multiple selection can be related to the noise level in the training data, sensor accuracy, and sampling period. After setting the boundaries, their usability needs to be verified. This includes checking if the upper bound is greater than the lower bound, if the boundaries fall within the physically feasible range of the features, and if there are abrupt changes between different segments to avoid false triggers or missed triggers in subsequent comparison stages.

[0054] When defining the numerical range formed by the upper and lower bounds as the historical fluctuation confidence interval, it is necessary to bind and store the interval of each feature with its feature identifier, forming a searchable set of confidence interval parameters. The stored content should at least include the feature identifier, upper bound, lower bound, mean parameter, standard deviation parameter, and corresponding data statistical range marker. The data statistical range marker indicates whether the interval is calculated from the full historical data or segmented historical data. This binding relationship is directly invoked in subsequent real-time comparisons, enabling the fluctuation determination of each feature to be compared based on its own historical scale, rather than a coarse-grained judgment based on a globally uniform threshold.

[0055] For example, the formula for calculating feature importance (average gain):

[0056] in, Representation of features The importance score (i.e., "average gain value"). T represents the total number of decision trees in the random forest regression model. Representation of features The information gain (or reduction in Gini impurity) resulting from splitting a node in the t-th decision tree.

[0057] This embodiment obtains the average gain value through supervised training of a random forest regression model and normalizes it into a contribution score. The contribution score is then assigned as the initial weight value, forming the weight initialization basis aligned with the columns of the historical feature matrix. At the same time, the mean and standard deviation parameters are calculated for each feature of the historical feature matrix, and upper and lower bounds are set. This forms a historical fluctuation confidence interval independently bound by the feature dimension, enabling subsequent processing to have a searchable and comparable benchmark in both the feature importance and the scale of historical fluctuations. This reduces the subjectivity of feature contribution estimation and improves the consistency and traceability of fluctuation judgment.

[0058] In one embodiment, step S30 above includes: S301 collects real-time, multi-dimensional feature data reflecting the operating status of the equipment through a sensor network or data interface deployed on the equipment at fixed time intervals or event-triggered methods. S302, perform real-time anomaly detection and removal on the collected real-time multi-dimensional feature data, and interpolate and complete the missing segments in the real-time multi-dimensional feature data to obtain real-time cleaned data. S303, The real-time cleaning data is transformed into dimensionless form using the standard fraction standardization method to eliminate dimensional differences and obtain real-time preprocessed feature data; S304, Maintain a sliding time window, and append the real-time preprocessed feature data to the sliding time window in the order of the timestamps of the real-time preprocessed feature data; S305, extract all data from the sliding time window, and organize all data into a two-dimensional matrix with rows corresponding to time points and columns corresponding to feature variables in chronological order to construct a real-time feature matrix.

[0059] In this embodiment, a sensor network or data interface is used to form a real-time multi-dimensional feature data acquisition channel. The sensor network covers observation points that characterize the operating state, such as temperature, voltage, current, vibration, pressure, rotational speed, and internal resistance. The data interface covers the controller bus, communication link, or host computer acquisition link. Both establish unified constraints on data format, sampling frequency, and timestamp source for subsequent alignment. Fixed time intervals are used to achieve periodic acquisition. A reading action is triggered by setting the sampling period, and the acquisition results are written to a buffer. Event-triggered methods are used to respond to state changes or alarm conditions. Reading actions are triggered by detecting event signals such as threshold exceeding limits, mode switching, load changes, and charge / discharge switching, and event identifiers are attached. To avoid timing misalignment caused by multi-source acquisition, a timestamp is bound to each piece of real-time multi-dimensional feature data during the acquisition phase. Feature values ​​from different channels are aggregated and marked within the same timestamp, forming a traceable acquisition batch.

[0060] Real-time anomaly detection and removal are used to eliminate significantly distorted observations before data enters subsequent standardization. Anomaly sources include sensor glitches, communication jitter, range overflow, and recovery spikes after short-term link interruptions. Anomaly detection can be performed jointly based on statistical rules, physical constraints, and change constraints. Statistical rules use sliding statistics to identify outliers deviating from the distribution; physical constraints use the feasible range of features to identify impossible values; and change constraints use the rate of change of adjacent sampling points to identify abrupt changes. Removal is performed on a feature-dimensional basis, marking outliers as gaps and recording the reasons for removal, enabling subsequent interpolation completion to distinguish between true missing data and gaps caused by anomaly removal. Interpolation completion of missing segments is used to restore the continuity of the time series. Missing segments consist of consecutive missing timestamps or consecutive gap markers. Completion strategies can be selected based on the missing length and feature type, including linear interpolation, spline interpolation, forward imputation, or estimation completion based on adjacent window statistics. The completion process writes a completion flag to each completion point, so that the real-time cleaned data contains both numerical values ​​and quality labels, which facilitates the subsequent stages of downgrading or isolating low-confidence completion segments.

[0061] The standard score standardization method maps real-time cleaned data with different dimensions to a uniform scale. Standardization is performed separately for each feature dimension, constructing real-time standardization parameters for each feature and performing a dimensionless transformation. Standardization parameters can be historical statistical parameters or real-time window statistical parameters. When using historical statistical parameters, the mean and standard deviation are kept consistent with historical data to ensure scale stability. When using real-time window statistical parameters, the mean and standard deviation are derived from the current window to ensure scale adaptation. The output of the dimensionless transformation forms real-time preprocessed feature data. Each feature value in the output is bound to a corresponding feature identifier, and a timestamp and quality label are retained, ensuring both numerical consistency and data quality traceability in subsequent matrix construction. To reduce numerical amplification caused by excessively small standard deviations, a lower limit can be set on the standard deviation parameter or a small stabilizing term can be added to ensure the numerical stability of the standardization results.

[0062] A sliding time window is used to maintain a time-series cache of real-time preprocessed feature data. The window is defined by its length and the advancement strategy. The window length can be represented by the time span or the number of samples, and the advancement strategy can be represented by a fixed step size or event-driven advancement. Append operations are performed in timestamp order. Before appending, a consistency check is performed between the newly arrived data and the timestamp at the end of the window. If out-of-order data is found, rearrangement or buffering is performed. If duplicate timestamps are found, aggregation or overwriting strategies are performed to maintain the uniqueness of the feature vector at a single time point. Within the window, data is organized using timestamps as the primary key and feature identifiers as column indices, ensuring that each append operation locates the corresponding row and writes to the corresponding column. Missing features are reserved as placeholders and inherit quality labels. Window maintenance also handles the eviction of data exceeding the window length, prioritizing the earliest timestamp for deletion, ensuring that the window always retains data from the most recent consecutive time periods.

[0063] The entire data is extracted from the sliding time window to form the complete real-time feature matrix. The extraction process first determines the set of timestamps within the window and sorts them chronologically, then generates row indices based on the sorting results. For the two-dimensional matrix, each row corresponding to a time point must contain a feature vector at the same timestamp, and each column corresponding to a feature variable must be fixedly mapped to the same feature identifier. During matrix construction, real-time preprocessed feature data is written according to row and column indices. Missing values ​​are filled according to preset rules or missing markers are retained. Filling rules can use zero-padding, mean-padding, or the most recent valid value, and are linked to the quality label. After matrix construction, a structure validation is performed. Validation includes whether the column set matches the target feature set, whether the rows monotonically increase chronologically, whether the matrix dimensions match the window length, and whether the standardized scale of each column meets the expected range. After successful validation, the real-time feature matrix is ​​output for subsequent monitoring and comparison.

[0064] For example, the representation of a multidimensional feature vector at a given time point:

[0065] Where t represents the time point or sampling time index; M represents the number of features; Indicates the first These features at a given time point The possible values ​​of ; Indicates a point in time The multidimensional feature vectors are used as a row of a two-dimensional matrix to participate in the construction of the real-time feature matrix.

[0066] This embodiment collects real-time multi-dimensional feature data through a sensor network or data interface at fixed time intervals or event-triggered modes. After collection, real-time anomaly detection and removal, as well as interpolation and completion of missing segments, are performed to form real-time cleaned data. Then, a dimensionless transformation is performed using a standard score standardization method to obtain real-time preprocessed feature data. Combined with a sliding time window, data is appended and extracted in time stamp order and organized into a two-dimensional matrix with rows corresponding to time points and columns corresponding to feature variables. This ensures that the real-time feature matrix has a consistent data foundation in terms of time alignment, scale uniformity, and data quality traceability, reducing the interference of multi-source acquisition noise and dimensional differences on subsequent fluctuation judgment and weight update links.

[0067] In one embodiment, step S50 above includes: S501, Traverse each feature in the real-time feature matrix and extract the real-time data sequence of the feature within a preset time window before the current time. S502, calculate the statistical change of the real-time data sequence within the preset time window, and define the statistical change as the change amplitude of the feature; S503, retrieve the upper limit value and lower limit value of the historical fluctuation confidence interval corresponding to the feature from the stored confidence interval parameters; S504, determine whether the change amplitude is greater than the upper limit of the historical fluctuation confidence interval or less than the lower limit of the historical fluctuation confidence interval; S505, determine whether the change range of the feature exceeds the historical fluctuation confidence interval based on the judgment result.

[0068] In this embodiment, the real-time feature matrix carries a structured expression of the evolution of multi-dimensional features over time during device operation. Row dimensions are bound to time points, and column dimensions are bound to features. Monitoring the real-time changes of each feature in the real-time feature matrix is ​​first implemented through a scheduling mechanism for traversing feature columns. The traversal order can be sorted by feature identifier or executed in parallel by feature grouping to match computing resource configuration. After locating a feature each time, extracting the real-time data sequence within a preset time window before the current moment requires simultaneously defining the time window boundaries and data indexing rules. The time window can be expressed using time span or the number of sampling points. Boundary calculation uses the current time stamp as the right endpoint, obtains the left endpoint by backtracking, and then truncates continuous segments based on the timestamp index or circular buffer index. During the truncation process, missing values ​​and completion flags are handled consistently. Missing values ​​can be filled according to rules during the truncation stage, or missing flags can be retained and a robust statistical strategy can be adopted during the statistical calculation stage to avoid statistical distortion caused by local missing values.

[0069] The statistical change of a real-time data sequence within a preset time window is used to map the intensity of fluctuations to a comparable numerical scale. The statistical change must be calculated using the same caliber as the historical fluctuation confidence interval to ensure effective numerical comparisons at the same scale. The statistical change can be expressed as the difference between the maximum and minimum values ​​within the window, the difference between the last and initial values, or the difference between the mean and the last value. The choice of caliber depends on how the historical fluctuation confidence interval is constructed and the type of anomaly to be captured. If the historical fluctuation confidence interval reflects the normal range of deviations of the feature around the mean, the statistical change can be selected as the deviation of the last value relative to the window mean, retaining the sign to distinguish between upward and downward deviations. If the historical fluctuation confidence interval reflects the normal range of fluctuation amplitude within the window, the statistical change can be selected as the range or standard deviation to characterize the degree of dispersion within the window. Once the statistical change is determined, it needs to be defined as the change magnitude of the feature and written into a traceable data structure. The data structure should at least include the feature identifier, the current time stamp, the preset time window parameter, and the change magnitude value, so that subsequent weight updates and input adaptation can directly reference the same change magnitude result, avoiding duplicate calculations and caliber drift.

[0070] The confidence interval parameter storage ensures the reusability and consistent retrieval of historical fluctuation confidence intervals. Retrieval actions use the feature identifier as the primary key and the version number or effective time as the secondary key to read the upper and lower limits of the historical fluctuation confidence interval from the parameter storage. The parameter storage can be a key-value database, a configuration center, or a memory-mapped table. During retrieval, parameter integrity must be verified, at least ensuring that both the upper and lower limits exist, the upper limit is not less than the lower limit, and the parameter's standardized definition matches the current feature. To reduce latency caused by frequent access, a cache of features up to and including upper and lower limits can be maintained in memory, with a consistent update strategy to ensure the cache is refreshed synchronously when parameters are updated, avoiding misjudgments caused by using expired upper and lower limits.

[0071] The comparison of the magnitude of change with the upper and lower limits requires bilateral numerical judgment. The comparison logic uses whether the magnitude of change is greater than the upper limit of the historical fluctuation confidence interval or less than the lower limit of the historical fluctuation confidence interval as the judgment condition. To avoid frequent switching caused by boundary jitter, a tolerance band or hysteresis band can be introduced in the comparison stage. The tolerance band is used to treat small out-of-bounds deviations near the boundary as acceptable fluctuations, and the hysteresis band is used to set regression conditions when the out-of-bounds deviation recovers, thereby suppressing repeated triggering near the boundary. The comparison stage also needs to address numerical stability issues, including truncation of extreme values ​​after standardization, weight reduction judgment of outlier completion segments, and correction of statistical scale changes caused by changes in window length. The judgment result is output as a feature-level out-of-bounds marker, which is bound to the feature identifier and the current time to form a structured record. The structured record can include out-of-bounds direction information to distinguish between cases greater than the upper limit and cases less than the lower limit, so that subsequent contribution change calculations can use the direction information to select a more suitable regression input segment or update strategy. Finally, based on the out-of-bounds marker, it is determined whether the change range of the feature exceeds the historical fluctuation confidence interval. This determination is written into the state table as a Boolean value or an enumerated state, and can be summarized to form the out-of-bounds feature set at the current moment, providing deterministic input for the subsequent real-time weight set update.

[0072] For example, the characteristic fluctuation anomaly triggering condition (3σ criterion) is:

[0073] in, Representation of features The real-time value at the current time t. Representation of features The statistical mean within the current monitoring time window. Representation of features The statistical standard deviation within the current monitoring time window. The logical expression of the entire formula is that if the real-time value deviates from the window mean by more than 3 times the standard deviation, it is judged as "out of confidence interval".

[0074] Formula for calculating the difference in characteristic variation amplitude:

[0075] in, This indicates the magnitude or amount of change at time point t; This represents the characteristic value at time point t; Representing time point t The characteristic value of 1; || represents absolute value operation. This formula uses the absolute value of the difference between adjacent time points as the realization of the change amplitude, and can be used to support the calculation of the change amplitude within a preset time window.

[0076] This embodiment extracts real-time data sequences within a preset time window from the real-time feature matrix feature by feature and calculates the change amplitude with a unified caliber. Then, it retrieves the upper and lower limits of the historical fluctuation confidence interval from the confidence interval parameters and performs a bilateral comparison. This enables the formation of traceable out-of-bounds markers and out-of-bounds direction information for the fluctuation state of each feature at the current moment. Thus, without changing the structure of the real-time feature matrix, continuous temporal fluctuations are compressed into structured judgment results that can be used for triggering conditions. This reduces the impact of single-point noise, missing completion, and scale differences on out-of-bounds judgment and provides a stable and consistent data basis for subsequent weight updates and input adaptation based on out-of-bounds results.

[0077] In one embodiment, step S60 above includes: S601, for features whose change magnitude is greater than the upper limit of the historical fluctuation confidence interval or less than the lower limit of the historical fluctuation confidence interval, the real-time correlation coefficient between the real-time data sequence of the feature and the equipment life degradation trend is calculated using an online support vector machine regression model, and the real-time correlation coefficient is used as the contribution change. S602, the contribution change is multiplied by the preset learning rate parameter to obtain the weighted adjustment amount; S603, read the current weight value of the feature from the real-time weight set, add the current weight value to the weighted adjustment amount to obtain the updated weight value, and replace the corresponding value in the real-time weight set with the updated weight value to update the weight value of the feature; S604, for features whose change amplitude is not greater than the upper limit of the historical fluctuation confidence interval and not less than the lower limit of the historical fluctuation confidence interval, the original weight value of the feature in the real-time weight set is retained so as to keep the weight value of the feature in the real-time weight set unchanged. S605, after completing the update or maintenance operation on all features, outputs the final determined real-time weight set as the real-time weight set at the current moment.

[0078] In this embodiment, two sets are established at the feature granularity: one set is the out-of-bounds feature set, judged by the change amplitude being greater than the upper limit of the historical fluctuation confidence interval or less than the lower limit of the historical fluctuation confidence interval; the other set is the non-out-of-bounds feature set, judged by the change amplitude being neither greater than the upper limit of the historical fluctuation confidence interval nor less than the lower limit of the historical fluctuation confidence interval. This set division needs to be consistent with the time window for calculating the change amplitude to avoid drift caused by using window A for out-of-bounds judgment while using window B for contribution change calculation. Each feature in the out-of-bounds feature set corresponds to a real-time data sequence, which originates from a column vector segment aligned with the time window in the real-time feature matrix. During extraction, timestamps are read synchronously and arranged chronologically. If necessary, missing points are filled using an interpolation completion strategy consistent with the preceding preprocessing, or missing markers are retained and a masking mechanism is used on the regression input side to avoid introducing spurious fluctuations during completion.

[0079] The calculation of contribution changes is based on an online support vector machine (SVM) regression model. The input to this model is jointly constructed from real-time data sequences and equipment lifespan degradation trends. The equipment lifespan degradation trend can be represented by a trend sequence obtained through local regression extrapolation of historical remaining lifespan values, a smoothed curve of the remaining lifespan prediction results within the most recent time window, or a trend sequence constructed using degradation surrogate variables strongly correlated with lifespan. The key is that the trend sequence and the real-time data sequence are strictly aligned in timestamps. To ensure the contribution of features with different dimensions to the regression process is controllable, the input to the online SVM regression model can undergo standardization consistent with historical stages. The standardization parameters can reuse the mean and standard deviation parameters obtained from historical statistics, or use sliding update mean and standard deviation parameters to adapt to distribution drift. The supervision signal for the online SVM regression model comes from the equipment lifespan degradation trend. The training method reflects its online nature, allowing for incremental fitting triggered at each time window, as well as incremental updates on new samples upon event triggering. During incremental updates, the support vector set is preserved, and the maximum number of support vectors is limited to prevent unbounded growth in memory and computational costs over time.

[0080] The real-time correlation coefficient, as a measure of contribution change, requires a clear definition of its calculation object and symbolic meaning. The real-time correlation coefficient can be derived from the marginal sensitivity of the real-time data sequence by the online support vector machine regression model. For example, it can be obtained by aggregating the local gradients of the regression output relative to the input within a window to obtain the sensitivity coefficient. Alternatively, the goodness-of-fit index between the regression predicted value and the trend sequence can be mapped to a correlation measure, or the correlation structure between the input features and the regression residuals can be converted into a correlation coefficient. To make contribution changes comparable across different features, the real-time correlation coefficient can be further normalized. Normalization can be based on scaling the maximum absolute value within the window, or it can be mapped based on the quantile scale of the contribution scores in historical periods. Normalization also preserves the sign to express whether the feature has a positive or negative association with the degradation trend. Once the contribution change is determined, it needs to be written into a data structure bound to the feature identifier, serving as the sole input for subsequent learning rate weighting and weight updates to avoid inconsistent results caused by repeated calculations within the same window.

[0081] The learning rate parameter is introduced to control the step size of weight updates. The weighted adjustment obtained by multiplication is equivalent to converting the contribution change into an incremental term for weight updates. The learning rate parameter can be a global constant or a grouping parameter bound to features, and it can also decay over time to suppress long-term cumulative drift. To avoid drastic weight jumps caused by abnormal windows, the weighted adjustment can be superimposed with a limiting strategy, performing upper and lower bound truncation on the adjustment, and combining out-of-bounds direction information to perform asymmetric limiting, so that the update magnitude triggered by upward and downward offsets conforms to engineering constraints. Subsequently, the current weight value is read from the real-time weight set. This reading needs to ensure concurrency consistency. In a multi-threaded or multi-process environment, read-write locks or version number mechanisms can be used on the real-time weight set to ensure that the current weight value used for calculation at the same time is consistent with the index position of the final write-back. After adding the current weight value to the weighted adjustment amount to obtain the updated weight value, boundary constraint processing can be performed on the updated weight value. For example, a minimum and maximum weight value can be set to avoid negative or excessively large weights that could lead to imbalance in subsequent weighted inputs. Soft constraints can also be introduced near the boundaries to maintain differentiability. Writing the updated weight value back to the real-time weight set requires using a feature identifier-to-index mapping table to locate the write-back position. The mapping table can be built and cached during the initialization phase. The write-back adopts an atomic write or batch write strategy. Batch write is suitable for vectorized computation to reduce write amplification and lock contention.

[0082] The processing of the non-boundary feature set follows the principle of maintenance, without modifying the corresponding values ​​in the real-time weight set. However, the non-boundary state still needs to be recorded for subsequent auditing and debugging. The state record can share the same state table as the boundary feature set. To reduce frequent switching near the boundary, a hysteresis condition can be introduced between the boundary judgment and the maintenance judgment. This requires features that have just crossed the boundary to meet stricter regression conditions before returning to the maintenance state, thereby reducing repeated updates of weights within adjacent windows. Finally, after completing the update or maintenance operation for all features, the real-time weight set is output as the real-time weight set at the current moment. The output format can be a vector, a key-value table, or a structure with a version number. The version number is formed by combining the current time-to-date timestamp and the window number, ensuring that downstream applications can identify a consistent weight snapshot.

[0083] For example, the real-time weight incremental update formula:

[0084] This represents the weight value of the i-th feature at time point t, which is an element belonging to the real-time weight set; This represents the weight value of the i-th feature at the previous time point; α represents the learning rate parameter, used to control the magnitude of weight adjustment; This represents the change in the contribution of the i-th feature at time point t. The formula uses the contribution change scaled by the learning rate as an adjustment, and adds it to the current weights to obtain the updated weights.

[0085] Formula for calculating the correlation of changes in contribution:

[0086] in, This represents the change in the contribution of the i-th feature at time point t; () represents the correlation calculation function, and the output can be the correlation coefficient or a comparable correlation measure. This represents the value of the i-th feature at time point t, which in engineering implementation typically corresponds to a real-time data sequence within a time window; This represents the remaining lifetime or lifetime degradation characteristic at time point t. In engineering implementation, it can be composed of historical remaining lifetime values, online estimated lifetime degradation trend values, etc. This formula is used to map the real-time correlation between the feature and the lifetime degradation characteristic into a change in contribution.

[0087] This embodiment binds the determination of out-of-bounds changes in the magnitude of changes to the calculation of the contribution changes of the online support vector machine regression model within the same time window, and introduces a learning rate parameter to form a controllable weighting adjustment. Then, it performs incremental updates or maintenance operations on the real-time weight set in a read-write consistent manner. This enables timely correction of feature importance when the operating conditions change abruptly, while avoiding unnecessary weight disturbances when the operating conditions are stable. This allows the real-time weight set to maintain a continuous and controllable evolution trajectory in the time dimension, reducing the impact of drastic weight oscillations on the subsequent model input weighting and prediction stability, and improving the adaptability to degradation trend drift caused by load fluctuations and environmental changes.

[0088] In one embodiment, step S70 above includes: S701, calculate the statistical mean of the change range of each feature, and compare the statistical mean with the preset feature change benchmark threshold. S702, if the statistical mean is greater than the feature change benchmark threshold, the time step of the dynamic prediction model is shortened; if the statistical mean is less than the feature change benchmark threshold, the time step of the dynamic prediction model is increased. S703, Set a time smoothing window, traverse the real-time weight set, identify features whose weight values ​​are continuously lower than the preset minimum effective weight threshold within the time smoothing window, and remove the data corresponding to the features from the real-time preprocessed feature data corresponding to the real-time multi-dimensional feature data. S704, identify the feature in the real-time weight set whose weight value is continuously higher than the preset enhanced weight threshold within the time smoothing window, and enhance the corresponding value of the feature in the real-time preprocessed feature data; S705, using the weight values ​​in the real-time weight set, the real-time preprocessed feature data after feature removal and numerical enhancement is weighted and combined to generate model input data.

[0089] In this embodiment, driven by the magnitude of change and the real-time weight set, adaptive reconstruction of real-time preprocessed feature data is performed to output model input data. The magnitude of change comes from the statistical change within a preset time window. First, the statistical mean of the magnitude of change of all features is calculated. The calculation of the statistical mean needs to be consistent with the window size of the magnitude of change. It is aggregated by feature dimension to obtain a single scalar, which is used to characterize the overall fluctuation level within the current time window. The feature change benchmark threshold is used as a reference. It can be derived from the quantile statistics of the distribution of magnitude of change in historical stages, or it can be an engineering threshold set based on the safety margin of equipment operation. After comparing the statistical mean with the feature change benchmark threshold, the time step of the dynamic prediction model is adjusted. When the statistical mean is greater than the feature change benchmark threshold, the time step is shortened; when the statistical mean is less than the feature change benchmark threshold, the time step is increased. The adjustment of the time step affects the temporal unfolding granularity and inference update frequency of the dynamic prediction model. It does not directly change the numerical ontology of the real-time preprocessed feature data, but it will change the sampling density and alignment rhythm of the subsequent model input data in the time dimension. To avoid frequent switching of the time step between adjacent windows, a hysteresis strategy or minimum adjustment interval can be superimposed on the comparison stage to ensure that the time step maintains controllable continuity within a short period.

[0090] A time-smoothing window is used to stabilize the judgment logic of the real-time weight set, avoiding erroneous removal or enhancement caused by single-point weight fluctuations. The time-smoothing window can be implemented using a sliding queue or a circular buffer, caching the weight value sequence of each feature across multiple consecutive windows, and forming a steady-state judgment based on consistently lower or higher weight values. The determination of consistently lower than the minimum effective weight threshold can be implemented using a counter, triggering removal only when the weight value is less than the minimum effective weight threshold at every moment within the time-smoothing window; alternatively, it can be implemented by comparing the window mean or maximum value with the minimum effective weight threshold, triggering removal only when the maximum value is still lower than the minimum effective weight threshold, thus excluding occasional spikes from the removal triggering conditions. The removal action targets the data corresponding to the feature in the real-time preprocessed feature data. During execution, it relies on the mapping relationship between feature identifiers and column indices, deleting or invalidating the corresponding column from the data structure and synchronously updating the feature mask, so that subsequent weighting and combination will no longer include this feature.

[0091] The determination of values ​​consistently exceeding the enhancement weight threshold also relies on a time smoothing window. The enhancement weight threshold and the minimum effective weight threshold form a stratified interval. Enhancement triggers and removal triggers should be mutually exclusive to avoid the same feature simultaneously satisfying both conditions within the same window. Enhancement actions are applied to the numerical values ​​corresponding to the feature in the real-time preprocessed feature data. Enhancement methods can employ numerical enhancement mechanisms such as amplitude amplification, contrast stretching, nonlinear mapping, or confidence recalibration. Amplification can be achieved using enhancement coefficients, which can be calculated from the proportion of the weight value exceeding the enhancement weight threshold, or generated by a joint function of the change amplitude and weight value, ensuring that the enhancement strength reflects both feature importance and fluctuation intensity. Nonlinear mapping can use piecewise functions to improve the resolution of key intervals while maintaining numerical monotonicity, avoiding excessive expansion of outliers caused by simple linear amplification. The enhanced values ​​need to be compatible with the standardized caliber of the real-time preprocessed feature data. If necessary, consistent truncation and normalization should be performed after enhancement to prevent dimensional drift introduced by enhancement.

[0092] After feature removal and numerical augmentation, the process moves to weighting and combination. Weighting uses weight values ​​from the real-time weight set, performing element-wise operations on the weight values ​​and corresponding feature data to form weighted feature data. Element-wise operations can employ multiplicative weight scaling or a combination of additive bias and multiplicative scaling, allowing the weights to express both relative importance and calibration offset. Combination generates a structured representation of the model input data. The combination method is consistent with the input interface of the dynamic prediction model, and can be organized as a two-dimensional matrix, a three-dimensional tensor, or a time-indexed batch sequence. Before combination, the feature order table and index mapping need to be reconstructed to ensure that the arrangement of the removed and augmented feature sets is definite and reproducible. The time alignment rhythm after time step adjustment is incorporated into the time index generation rules to ensure that the model input data is consistent with the time step of the dynamic prediction model in the time dimension. To support the throughput and stability of online processing, weighting and combination can use batch vectorized computation, and a version identifier and timestamp are appended to the output to form a traceable snapshot of the model input data for direct reading in subsequent prediction calculations.

[0093] For example, one of the time step adaptive scaling formulas: when

[0094] in, Indicates the adjusted time step; Indicates the default time step; Indicates the magnitude or amount of change at the current moment; This represents the baseline threshold for feature changes. This formula gives the time step scaling relationship when the change exceeds the baseline threshold, which is used to quantify the action of "shortening or increasing the time step" into a computable expression.

[0095] Time step adaptive scaling formula two:

[0096] in, Indicates the adjusted time step; Indicates the default time step; Indicates the magnitude or amount of change at the current moment; This represents the baseline threshold for feature changes. This formula provides the time step scaling relationship when the change magnitude does not exceed the baseline threshold, used to quantify the action of "increasing the time step" into a computable expression.

[0097] Weighted filtering and weighted input set construction:

[0098] in, The weighted set of input features at time point t represents a form of expression of the model's input data. This represents the weight value of the i-th feature at time point t; This represents the value of the i-th feature at time point t; Represents the minimum effective weight threshold; condition This means that only features with weights greater than or equal to the threshold are retained, achieving feature elimination and retention. This formula simultaneously reflects feature selection based on the threshold and numerical weighting based on the weights, aligning with the feature elimination and weighted combination used to generate the input components.

[0099] Single-feature weighted mapping:

[0100] in, This represents the i-th weighted eigenvalue at time point t; This represents the weight value of the i-th feature at time point t; This represents the value of the i-th feature at time point t. This formula is a single-feature layer expansion of the previous set formula, used to explain how the value of each feature is mapped to the weights during weighted combination.

[0101] This embodiment uses the comparison result of the statistical mean of the change amplitude and the feature change benchmark threshold to adjust the time step of the dynamic prediction model. It uses a time smoothing window to stabilize the low-weight elimination and high-weight enhancement of the real-time weight set. Then, it performs weighting and combination based on the real-time weight set on the eliminated and enhanced real-time preprocessed feature data. This can increase the time granularity to capture degenerate changes when the overall fluctuation increases and decrease the time granularity to reduce redundant calculations when the overall fluctuation decreases. At the same time, it suppresses the noise interference caused by low-contribution features and strengthens the expression intensity of high-contribution features. This enables the model input data to achieve synchronous adaptation in both time resolution and feature expression dimensions, thereby improving the consistency and usability of the input data in representing changes in working conditions.

[0102] In one embodiment, a device for dynamically predicting the remaining life of equipment is provided, which corresponds one-to-one with the method for dynamically predicting the remaining life of equipment in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the device for dynamic prediction of remaining lifespan of the present invention. The modules include: historical feature construction module 10, feature weight initialization module 20, real-time feature construction module 30, weight set initialization module 40, feature change monitoring module 50, weight dynamic update module 60, model input adaptive generation module 70, and remaining lifespan prediction module 80. Detailed descriptions of each functional module are as follows: The historical feature construction module 10 is used to acquire the historical multi-dimensional feature data of the device and the corresponding historical remaining life value, and to construct a historical feature matrix based on the historical multi-dimensional feature data. The feature weight initialization module 20 is used to determine the initial weight value of each feature based on the historical feature matrix and the historical remaining lifetime value, and to calculate the historical fluctuation confidence interval of each feature based on the historical feature matrix. The real-time feature construction module 30 is used to acquire real-time multi-dimensional feature data during the operation of the device and construct a real-time feature matrix based on the real-time multi-dimensional feature data. The weight set initialization module 40 is used to initialize the real-time weight set of each feature using the initial weight value; The feature change monitoring module 50 is used to monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval. The weight dynamic update module 60 is used to dynamically adjust the weight values ​​of each feature in the real-time weight set based on the comparison results, so as to obtain the real-time weight set at the current moment. The model input adaptive generation module 70 is used to adaptively process the real-time multi-dimensional feature data based on the change magnitude of each feature and the real-time weight set to generate model input data. The remaining life prediction module 80 is used to input the model input data into the dynamic prediction model to perform prediction calculations and output the remaining life prediction results of the device.

[0103] Specific limitations regarding the dynamic prediction device for remaining equipment life can be found in the aforementioned limitations on the dynamic prediction method for remaining equipment life, and will not be repeated here. Each module in the aforementioned dynamic prediction device for remaining equipment life can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0104] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a dynamic prediction method for the remaining life of a device on the server side.

[0105] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the client-side functions or steps of a dynamic prediction method for the remaining lifespan of a device.

[0106] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain historical multi-dimensional feature data of the device and the corresponding historical remaining lifespan value, and construct a historical feature matrix based on the historical multi-dimensional feature data; Based on the historical feature matrix and the historical remaining lifetime value, the initial weight value of each feature is determined, and the historical fluctuation confidence interval of each feature is calculated based on the historical feature matrix. Real-time multi-dimensional feature data during device operation is acquired, and a real-time feature matrix is ​​constructed based on the real-time multi-dimensional feature data; Using the initial weight values, initialize the real-time weight set for each feature; Monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval; Based on the comparison results, the weight values ​​of each feature in the real-time weight set are dynamically adjusted to obtain the real-time weight set at the current moment. Based on the variation magnitude of each feature and the real-time weight set, the real-time multi-dimensional feature data is adaptively processed to generate model input data. The model input data is input into the dynamic prediction model to perform prediction calculations and output the remaining life prediction results of the device.

[0107] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Obtain historical multi-dimensional feature data of the device and the corresponding historical remaining lifespan value, and construct a historical feature matrix based on the historical multi-dimensional feature data; Based on the historical feature matrix and the historical remaining lifetime value, the initial weight value of each feature is determined, and the historical fluctuation confidence interval of each feature is calculated based on the historical feature matrix. Real-time multi-dimensional feature data during device operation is acquired, and a real-time feature matrix is ​​constructed based on the real-time multi-dimensional feature data; Using the initial weight values, initialize the real-time weight set for each feature; Monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval; Based on the comparison results, the weight values ​​of each feature in the real-time weight set are dynamically adjusted to obtain the real-time weight set at the current moment. Based on the variation magnitude of each feature and the real-time weight set, the real-time multi-dimensional feature data is adaptively processed to generate model input data. The model input data is input into the dynamic prediction model to perform prediction calculations and output the remaining life prediction results of the device.

[0108] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

Claims

1. A method for dynamically predicting the remaining life of equipment, characterized in that, Includes the following steps: Obtain historical multi-dimensional feature data of the device and the corresponding historical remaining lifespan value, and construct a historical feature matrix based on the historical multi-dimensional feature data; Based on the historical feature matrix and the historical remaining lifetime value, the initial weight value of each feature is determined, and the historical fluctuation confidence interval of each feature is calculated based on the historical feature matrix. Real-time multi-dimensional feature data during device operation is acquired, and a real-time feature matrix is ​​constructed based on the real-time multi-dimensional feature data; Using the initial weight values, initialize the real-time weight set for each feature; Monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval; Based on the comparison results, the weight values ​​of each feature in the real-time weight set are dynamically adjusted to obtain the real-time weight set at the current moment. Based on the variation magnitude of each feature and the real-time weight set, the real-time multi-dimensional feature data is adaptively processed to generate model input data. The model input data is input into the dynamic prediction model to perform prediction calculations and output the remaining life prediction results of the device.

2. The method for dynamically predicting the remaining life of equipment as described in claim 1, characterized in that, Acquire historical multi-dimensional feature data of the device and corresponding historical remaining lifetime values, and construct a historical feature matrix based on the historical multi-dimensional feature data, including: Extract historical multi-dimensional feature data and corresponding historical remaining lifespan values ​​from the equipment's historical operation database; The historical multidimensional feature data is anomaly detected by using the standard deviation multiple statistical criterion, and numerical points that deviate from the mean by more than a preset multiple of the standard deviation are identified and removed to obtain preliminary cleaned data; The missing segments or duplicate records in the preliminary cleaned data are interpolated to complete or deduplicated to obtain the cleaned historical feature data. The cleaned historical feature data is transformed into dimensionless form using the standard score standardization method to eliminate dimensional differences and obtain preprocessed data. The preprocessed data is arranged and aligned according to the timestamp order of the preprocessed data, and the arranged and aligned data is organized into a two-dimensional matrix in which rows correspond to time points and columns correspond to feature variables, so as to construct a historical feature matrix.

3. The method for dynamically predicting the remaining life of equipment as described in claim 1, characterized in that, Based on the historical feature matrix and the historical remaining lifetime values, initial weight values ​​for each feature are determined, and historical fluctuation confidence intervals for each feature are calculated based on the historical feature matrix, including: A random forest regression model is constructed, with the historical feature matrix as input data and the historical remaining lifetime value as output label; The random forest regression model is trained under supervision using the historical feature matrix and the historical remaining lifetime values, and the average gain value generated by each feature during the model training process is calculated. The average gain value is normalized and converted into a contribution score for each feature. The contribution score is assigned to the corresponding feature as the initial weight value; For each feature in the historical feature matrix, calculate the mean parameter and standard deviation parameter of the feature over the historical time series; Based on the mean parameter and the standard deviation parameter, the upper and lower limits of the allowable fluctuation are set; The numerical range formed by the upper limit boundary and the lower limit boundary is defined as the historical fluctuation confidence interval of the feature.

4. The method for dynamically predicting the remaining life of equipment as described in claim 1, characterized in that, Real-time acquisition of multi-dimensional feature data during device operation, and construction of a real-time feature matrix based on the real-time multi-dimensional feature data, including: By using sensor networks or data interfaces deployed on the equipment, real-time multi-dimensional feature data reflecting the operating status of the equipment can be collected at fixed time intervals or in an event-triggered manner. Real-time anomaly detection and removal are performed on the collected real-time multi-dimensional feature data, and missing segments in the real-time multi-dimensional feature data are interpolated and filled to obtain real-time cleaned data. The real-time cleaning data is transformed into dimensionless form using the standard score standardization method to eliminate dimensional differences and obtain real-time preprocessed feature data. Maintain a sliding time window and append the real-time preprocessed feature data to the sliding time window in the order of the timestamps of the real-time preprocessed feature data. All data is extracted from the sliding time window and organized into a two-dimensional matrix with rows corresponding to time points and columns corresponding to feature variables in chronological order to construct a real-time feature matrix.

5. The method for dynamically predicting the remaining life of equipment as described in claim 1, characterized in that, Monitoring the real-time changes of each feature in the real-time feature matrix, calculating the magnitude of change of each feature within a preset time window, and comparing the magnitude of change with the corresponding historical fluctuation confidence interval, including: Iterate through each feature in the real-time feature matrix and extract the real-time data sequence of the feature within a preset time window before the current time. Calculate the statistical change of the real-time data sequence within the preset time window, and define the statistical change as the change amplitude of the feature; Retrieve the upper limit and lower limit of the historical fluctuation confidence interval corresponding to the feature from the stored confidence interval parameters; Determine whether the magnitude of the change is greater than the upper limit of the historical fluctuation confidence interval, or less than the lower limit of the historical fluctuation confidence interval; Based on the judgment result, determine whether the change range of the feature exceeds the historical fluctuation confidence interval.

6. The method for dynamically predicting the remaining life of equipment as described in claim 1, characterized in that, Based on the comparison results, the weight values ​​of each feature in the real-time weight set are dynamically adjusted to obtain the real-time weight set at the current moment, including: For features whose change magnitude is greater than the upper limit of the historical fluctuation confidence interval or less than the lower limit of the historical fluctuation confidence interval, the real-time correlation coefficient between the real-time data sequence of the feature and the equipment life degradation trend is calculated using an online support vector machine regression model, and the real-time correlation coefficient is used as the contribution change. The change in contribution is multiplied by a preset learning rate parameter to obtain a weighted adjustment amount; The current weight value of the feature is read from the real-time weight set, the current weight value is added to the weighted adjustment amount to obtain the updated weight value, and the updated weight value is used to replace the corresponding value in the real-time weight set to update the weight value of the feature. For features whose change amplitude is not greater than the upper limit of the historical fluctuation confidence interval and not less than the lower limit of the historical fluctuation confidence interval, the original weight value of the feature in the real-time weight set is retained to keep the weight value of the feature in the real-time weight set unchanged. After updating or maintaining all features, the final set of real-time weights is output as the set of real-time weights at the current moment.

7. The method for dynamically predicting the remaining life of equipment as described in claim 1, characterized in that, Based on the variation magnitude of each feature and the real-time weight set, the real-time multi-dimensional feature data is adaptively processed to generate model input data, including: Calculate the statistical mean of the change range of each feature, and compare the statistical mean with a preset feature change benchmark threshold. If the statistical mean is greater than the feature change benchmark threshold, the time step of the dynamic prediction model is shortened; if the statistical mean is less than the feature change benchmark threshold, the time step of the dynamic prediction model is increased. Set a time smoothing window, traverse the real-time weight set, identify features whose weight values ​​are consistently lower than the preset minimum effective weight threshold within the time smoothing window, and remove the data corresponding to the features from the real-time preprocessed feature data corresponding to the real-time multi-dimensional feature data. The feature in the real-time weight set whose weight value is continuously higher than a preset enhancement weight threshold within the time smoothing window is identified, and the corresponding value of the feature in the real-time preprocessed feature data is enhanced. Using the weight values ​​in the real-time weight set, the real-time preprocessed feature data after feature removal and numerical enhancement is weighted and combined to generate model input data.

8. A device for dynamically predicting the remaining life of equipment, characterized in that, The device for dynamically predicting the remaining lifespan of the equipment includes: The historical feature construction module is used to acquire the device's historical multi-dimensional feature data and corresponding historical remaining lifetime value, and to construct a historical feature matrix based on the historical multi-dimensional feature data. The feature weight initialization module is used to determine the initial weight value of each feature based on the historical feature matrix and the historical remaining lifetime value, and to calculate the historical fluctuation confidence interval of each feature based on the historical feature matrix. The real-time feature construction module is used to acquire real-time multi-dimensional feature data during the operation of the device and construct a real-time feature matrix based on the real-time multi-dimensional feature data. The weight set initialization module is used to initialize the real-time weight set of each feature using the initial weight values; The feature change monitoring module is used to monitor the real-time changes of each feature in the real-time feature matrix, calculate the change amplitude of each feature within a preset time window, and compare the change amplitude with the corresponding historical fluctuation confidence interval. The weight dynamic update module is used to dynamically adjust the weight values ​​of each feature in the real-time weight set based on the comparison results, so as to obtain the real-time weight set at the current moment. The model input adaptive generation module is used to adaptively process the real-time multi-dimensional feature data based on the change magnitude of each feature and the real-time weight set to generate model input data. The remaining lifetime prediction module is used to input the model input data into the dynamic prediction model to perform prediction calculations and output the remaining lifetime prediction results of the device.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a device remaining lifetime dynamic prediction program stored in the memory and executable on the processor. When the device remaining lifetime dynamic prediction program is executed by the processor, it implements the steps of the device remaining lifetime dynamic prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a dynamic prediction program for the remaining life of the device, which, when executed by a processor, implements the steps of the dynamic prediction method for the remaining life of the device as described in any one of claims 1-7.

Images