Equipment residual life prediction method and system based on AI ensemble learning

Abstract

The invention discloses an equipment residual life prediction method and system based on AI ensemble learning, and relates to the technical field of predictive maintenance, and the method comprises the steps: training a multi-time scale sub-learner based on a multi-time scale degradation feature vector, executing the equipment residual life prediction operation at the same time, and generating a multi-time scale candidate prediction result packet; according to the multi-time-scale degradation feature vector, constructing a degradation track coordinate space, and performing clustering segmentation, distance operation and similarity operation to generate a degradation track space state record; and according to the degradation track space state record, carrying out dynamic weighted combination on the equipment residual life prediction values corresponding to different time scales in the multi-time scale candidate prediction result packet, and generating an equipment residual life prediction report. According to the invention, accurate identification of the equipment degradation state is realized, a basis is provided for selection of a machine learning sub-model, and the stage adaptability of prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance technology, and in particular to a method and system for predicting the remaining life of equipment based on AI ensemble learning. Background Technology

[0002] Against the backdrop of rapid development in industrial intelligence and predictive maintenance, machine learning-based equipment remaining life (RUL) prediction technology has become an important means to ensure the high reliability of critical equipment. In recent years, with the advancement of sensor technology, edge computing, and artificial intelligence algorithms, RUL prediction methods have gradually shifted from early approaches based on physical models or statistical regression to data-driven deep learning and ensemble learning paradigms. Especially in complex industrial scenarios, multi-source heterogeneous operating data (such as vibration, temperature, and current) provides a rich source of information for building high-precision degradation models.

[0003] Existing technologies have two limitations: First, most RUL prediction models rely solely on fixed-length time windows for feature extraction, making it difficult to account for the coupling effect between short-term sudden anomalies and long-term intermittent degradation during device degradation, thus limiting the model's generalization ability. Second, existing ensemble learning frameworks typically employ static weighting mechanisms to fuse the outputs of multiple sub-models, lacking the ability to dynamically perceive the similarity between the device's current degradation state and its historical trajectory. This prevents them from adaptively adjusting the contribution weights of each sub-learner according to the actual operating stage, thereby affecting the timeliness and accuracy of the prediction results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a device remaining lifetime prediction method based on AI ensemble learning to solve the problems of insufficient fusion of degradation features across multiple time scales and lack of state adaptability in the weights of the ensemble model.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting the remaining lifespan of a device based on AI ensemble learning, which includes, Collect multi-source operational data and preprocess it. Then, divide the preprocessed multi-source operational data into multiple sliding time windows to generate multi-timescale operational data packages. Different data operations are performed on the data packets running at multiple time scales, and the operation results are time-stamped and concatenated with feature dimensions to generate multi-time-scale degenerate feature vectors. A multi-timescale sub-learner is trained based on the multi-timescale degradation feature vector, and the remaining life of the device is predicted simultaneously to generate a multi-timescale candidate prediction result package. Based on the degradation feature vectors at multiple time scales, a degradation trajectory coordinate space is constructed, and clustering, distance calculation, and similarity calculation are performed to generate a degradation trajectory space state record. Based on the spatial state record of the degradation trajectory, the predicted remaining life of the equipment at different time scales in the multi-time scale candidate prediction result package is dynamically weighted and combined to generate a prediction report of the remaining life of the equipment.

[0007] As a preferred embodiment of the AI-based ensemble learning-based device remaining life prediction method of the present invention, the step of dividing the preprocessed multi-source operating data into multiple sliding time windows to generate multi-timescale operating data packages is as follows. Based on the preprocessed multi-source operational data, multiple sets of sliding time window parameters are configured at short, medium, and long time scales to generate a sliding time window configuration record set. Based on the sliding time window configuration record set, window data segments are extracted from the preprocessed multi-source running dataset and recombined into multi-time scale running data packets according to different time scales.

[0008] As a preferred embodiment of the AI-based ensemble learning-based device remaining lifetime prediction method of the present invention, the steps of performing different data operations in multi-timescale running data packets are as follows: Perform trend slope calculation, fluctuation amplitude calculation, and periodic change calculation on the vibration, current, and voltage data in the multi-timescale running data package to generate a short-timescale calculation record set; Perform trend slope calculation, fluctuation amplitude calculation, and periodic change calculation on the temperature, speed, torque, pressure, and flow rate in the multi-timescale running data package to generate a medium-timescale calculation record set; Perform cumulative and proportional calculations on the cumulative runtime, number of starts and stops, and maintenance intervals in the multi-timescale running data packets to generate a long-term scale calculation record set.

[0009] As a preferred embodiment of the AI-based ensemble learning-based device remaining life prediction method of the present invention, the steps of aligning the computation results with time stamps and concatenating feature dimensions to generate multi-timescale degradation feature vectors are as follows. For short-timescale operation record sets, medium-timescale operation record sets, and long-timescale operation record sets, time markers are searched and inserted, and aligned in chronological order to generate multi-timescale operation aligned record sets. Feature fields are extracted from the multi-timescale operation-aligned record set, and feature dimensions are concatenated to generate multi-timescale degradation feature vectors.

[0010] As a preferred embodiment of the AI-based ensemble learning-based device remaining life prediction method of the present invention, the steps of training a multi-timescale sub-learner based on multi-timescale degradation feature vectors are as follows: The multi-timescale degradation feature vectors are time-aligned with the historical equipment lifecycle records and the remaining equipment lifecycle label is attached to generate a multi-timescale training sample set. Training samples are extracted from a multi-timescale training sample set and input into multiple sets of machine learning sub-learners to perform parameter initialization, batch training, and parameter updates, thereby generating multi-timescale sub-learners.

[0011] As a preferred embodiment of the AI-based ensemble learning-based device remaining lifetime prediction method of the present invention, the steps for performing device remaining lifetime prediction calculations and generating a multi-timescale candidate prediction result package are as follows: The multi-timescale degradation feature vector is input into the multi-timescale sub-learner, forward operation is performed, the remaining life of the device is output as the predicted value of the device's remaining life is generated, and a multi-timescale candidate prediction intermediate record set is generated. The intermediate record set of multi-timescale candidate predictions is organized according to the machine learning sub-learner identifier, time scale identifier, and time stamp to generate a multi-timescale candidate prediction result package.

[0012] As a preferred embodiment of the AI-based ensemble learning-based device remaining life prediction method of the present invention, the step of constructing a degradation trajectory coordinate space based on multi-timescale degradation feature vectors is as follows: The numerical range of the degradation feature vectors at multiple time scales is unified and the features are grouped and organized. The degradation process features are generated by combining the equipment life cycle records. At the same time, the configuration parameters of the degradation trajectory coordinate space are set. The degradation process characteristics are converted into degradation trajectory curves according to the configuration parameters of the degradation trajectory coordinate space, and the equipment degradation stages are divided to form a set of degradation trajectory coordinate points.

[0013] As a preferred embodiment of the AI-based ensemble learning-based device remaining life prediction method of the present invention, the steps of performing clustering segmentation, distance calculation, and similarity calculation to generate a degradation trajectory spatial state record are as follows: Based on the set of coordinate points of the degradation trajectory, initial segmentation and clustering operations are performed during the equipment degradation stage to generate a degradation trajectory spatial cluster; The current time-scale degradation feature vector is mapped to the current degradation coordinate point, and distance and similarity calculations are performed with the degradation trajectory space cluster to generate a degradation trajectory space state record.

[0014] As a preferred embodiment of the AI-based ensemble learning-based equipment remaining lifetime prediction method of the present invention, the steps of dynamically weighting and combining the predicted equipment remaining lifetime values ​​corresponding to different time scales in the multi-time scale candidate prediction result package according to the degradation trajectory spatial state record to generate an equipment remaining lifetime prediction report are as follows. Based on the degradation trajectory space state record, basic weight values ​​are assigned to the predicted remaining life of the equipment corresponding to different time scales in the multi-time scale candidate prediction result package, and a dynamic weighted configuration record set of degradation trajectory space is generated. The predicted remaining lifespan of the equipment and its corresponding weights at different time scales are read from the dynamically weighted configuration record set of the degradation trajectory space, and then dynamically weighted and combined to generate a multi-time scale weight distribution record. The remaining life prediction of the equipment, the multi-timescale weight distribution record, the set of degradation trajectory coordinate points, and the multi-timescale degradation feature vector are integrated into the remaining life prediction report of the equipment.

[0015] Secondly, the present invention provides a device remaining life prediction system based on AI ensemble learning, including a data acquisition module for acquiring multi-source operating data and preprocessing it, and dividing the preprocessed multi-source operating data into multiple sliding time windows to generate multi-timescale operating data packets. The feature generation module is used to perform different data operations in the running data packets at multiple time scales, and to align the operation results with time stamps and concatenate the feature dimensions to generate multi-time scale degenerate feature vectors. The learner training module is used to train multi-timescale sub-learners based on multi-timescale degradation feature vectors, while performing device remaining lifetime prediction calculations and generating multi-timescale candidate prediction result packages. The degradation operation module is used to construct the degradation trajectory coordinate space based on the degradation feature vectors of multiple time scales, and perform clustering segmentation, distance operation and similarity operation to generate the degradation trajectory space state record; The report generation module is used to dynamically weight and combine the predicted remaining life of equipment at different time scales in the multi-time scale candidate prediction result package based on the degradation trajectory spatial state record, and generate an equipment remaining life prediction report.

[0016] The beneficial effects of this invention are as follows: by constructing a degradation trajectory coordinate space and performing clustering and similarity calculations, accurate identification of equipment degradation status is achieved, providing a basis for the selection of machine learning sub-models and improving the stage adaptability of prediction; by dynamically weighting and fusing multi-scale machine learning prediction results based on state records, adaptive optimization of the integrated strategy is achieved, improving the accuracy and robustness of RUL prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a device remaining life prediction method based on AI ensemble learning.

[0019] Figure 2 This is a schematic diagram of a device remaining life prediction system based on AI ensemble learning.

[0020] Figure 3 This is a flowchart for multi-timescale feature extraction and alignment.

[0021] Figure 4 This is a flowchart for degenerate trajectory mapping and state recognition. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a device remaining lifetime prediction method based on AI ensemble learning, including the following steps: S1: Collect multi-source operational data and preprocess it. Divide the preprocessed multi-source operational data into multiple sliding time windows to generate multi-timescale operational data packages.

[0026] S1.1: Multi-source operating data includes vibration, temperature, current, voltage, speed, torque, pressure, flow rate, as well as cumulative operating time, number of start-stop cycles, and maintenance intervals.

[0027] Furthermore, vibration data is recorded as a sequence of vibration amplitude changes within a continuous operating cycle using vibration acquisition equipment. Temperature data is recorded in chronological order using temperature sensing equipment to obtain the temperature of the equipment casing, bearing, or oil. Current and voltage data are recorded as current and voltage change sequences based on the instantaneous values ​​output from the electrical measurement channel. Speed ​​and torque data are recorded as speed and torque change curves on the same time axis using signal acquisition equipment. Pressure and flow sensing devices are used to collect pressure and flow value sequences at fixed sampling intervals (e.g., 1 second or 5 seconds in the example) during operation. Cumulative running time is continuously accumulated based on the cumulative time the equipment is powered on to form a cumulative running time record. Start-stop counts are recorded based on the time stamps of equipment start and stop events. Maintenance intervals are recorded based on the time difference between the maintenance event occurrence time and the current acquisition time.

[0028] S1.2: Preprocessing includes timestamp alignment, missing segment interpolation repair, outlier removal, noise suppression, and unit standardization.

[0029] Furthermore, by performing time stamp lookup, time difference comparison, and time interpolation on data channels with different sampling frequencies, the multi-source operational data are kept consistent on the same time series. For missing segments in the time series, local interpolation is performed to repair them based on the data change trends of adjacent moments (for example, using the change amplitude of two adjacent points to estimate the value of the intermediate moment to fill in the missing segment). Fluctuation anomalies generated during operation and maintenance are identified and removed through range comparison, rate of change detection, and time period continuity detection to maintain reasonable numerical change relationships. High-frequency noise in channels such as vibration, current, and pressure is suppressed through filtering strategies. The values ​​in all channels are converted to a uniform scale according to the unit standardization method, making the multi-source operational data comparable on a numerical scale.

[0030] S1.3: Based on the preprocessed multi-source running data, multiple sets of sliding time window parameters are configured through short time scale, medium time scale and long time scale to generate a sliding time window configuration record set.

[0031] Furthermore, based on the continuous distribution of preprocessed multi-source operational data on a unified time axis, time span analysis is performed. By reading the time stamp interval, data change density, and characteristic change cycle, multiple sets of sliding time window parameters are configured for short, medium, and long time scales. For example, a 10-second window with a 2-second step size can be used for the short time scale; a 5-minute window with a 30-second step size for the medium time scale; and a 1-hour window with a 5-minute step size for the long time scale. By recording the window length, step size, and time span for each set of sliding time window parameters, the sliding time windows at different time scales can match the change cycles of different data channels. All sliding time window parameters are organized according to the structure of short, medium, and long time scales to generate a sliding time window configuration record set.

[0032] S1.4: Based on the sliding time window configuration record set, extract window data segments from the preprocessed multi-source running dataset and recombine them into multi-time scale running data packets according to different time scales.

[0033] Furthermore, the window length and step size information corresponding to the short, medium, and long time scales are read. In the preprocessed multi-source running dataset, sliding truncation is performed sequentially using time markers as indices. Continuous data segments are truncated according to the window start time and window end time. The step size controls the window movement interval, so that each time scale forms a sequence of data segments covering a continuous time period. The data segments of the short time scale are grouped and recorded in chronological order to form a short time scale window data set. The data segments of the medium and long time scales are organized according to the window division method to form a medium time scale window data set and a long time scale window data set, respectively. The window data sets of the three time scales are recombined according to the structure of the short, medium, and long time scales, while maintaining consistency with the sliding time window configuration record set, to generate a multi-time scale running data package.

[0034] S2: Perform different data operations in the data package at multiple time scales, and align the operation results with time stamps and concatenate the feature dimensions to generate a multi-time scale degenerate feature vector.

[0035] S2.1: Perform trend slope calculation, fluctuation amplitude calculation, and periodic change calculation on the vibration, current, and voltage data in the multi-timescale running data package to generate a short-timescale calculation record set.

[0036] Furthermore, vibration, current, and voltage data from short-timescale data segments within the multi-timescale running data package are read segment by segment. The continuity of the time stamp of each data segment is checked, and the trend slope is calculated based on the numerical change trend between adjacent sampling points. For example, the rate of increase or decrease in the trend is determined by the magnitude of the numerical increase or decrease. The fluctuation amplitude is calculated by searching for the difference between the maximum and minimum values ​​within each data segment, thus quantifying the fluctuation range over a short period. The periodic change is calculated by detecting the repetitive change patterns of vibration, current, and voltage data within fixed time intervals. For example, the interval between recurring vibration peaks or the length of the current fluctuation cycle is identified. The trend slope, fluctuation amplitude, and periodic change results are integrated in chronological order and a short-timescale calculation record set is generated according to the short-timescale structure.

[0037] S2.2: Perform trend slope calculation, fluctuation amplitude calculation, and periodic change calculation on the temperature, speed, torque, pressure, and flow rate in the multi-timescale running data package to generate a medium-timescale calculation record set.

[0038] Furthermore, the data segments belonging to the medium time scale in the multi-time scale running data package are read segment by segment, including temperature data, speed data, torque data, pressure data, and flow data. A continuity check is performed on the time stamp of each data segment, and the numerical sequence is extracted. Within each segment, the trend slope is calculated based on the numerical difference relationship between adjacent sampling points, yielding the changing trends of temperature, speed, torque, pressure, and flow at the medium time scale. The fluctuation amplitude is calculated by retrieving the difference between the maximum and minimum values ​​within a segment, quantifying the fluctuation range at the medium time scale. By dividing the numerical sequence within the data segment into equally spaced time intervals, the similarity of the shape and the consistency of the change amplitude of the numerical change curves in adjacent time sub-intervals are compared. When the change shape and amplitude characteristics in multiple time sub-intervals repeat within the similarity range, the corresponding numerical change is determined to be a periodic change, and its change interval and amplitude characteristics are recorded, such as periodic temperature fluctuations or periodic pressure fluctuations. The trend slope calculation results, fluctuation amplitude calculation results, and periodic change calculation results are integrated into a continuous record in chronological order and written into the medium time scale calculation record set according to the medium time scale window division method.

[0039] It should be noted that the similarity range (example range: 0.75 to 0.90) is set based on the normal fluctuation level of the medium-time scale operating data after noise suppression and preprocessing, and the upper limit is set based on the maximum consistency of the change pattern of adjacent cycles under stable operating conditions.

[0040] S2.3: Perform cumulative and proportional calculations on the cumulative runtime, number of starts and stops, and maintenance intervals in the multi-timescale running data packets to generate a long-timescale calculation record set.

[0041] Furthermore, the data segments belonging to the long-term scale in the multi-timescale runtime data packet are read sequentially according to time, including cumulative runtime, number of starts and stops, and maintenance interval. The cumulative runtime is calculated by summing the cumulative runtime within each long-term scale segment. The number of starts and stops is summed within the same segment based on the number of start and stop events to form the cumulative number of starts and stops. The maintenance interval is calculated by the time difference between the maintenance event time and the segment end time. The cumulative runtime, number of starts and stops, and maintenance interval are calculated proportionally based on the total amount of the long-term scale segment and the segment duration. The cumulative and proportional calculation results are recorded as consecutive entries according to the long-term scale structure to generate a long-term scale calculation record set.

[0042] S2.4: Perform time stamp lookup and insertion on short-timescale operation record sets, medium-timescale operation record sets, and long-timescale operation record sets, and align them according to time order to generate multi-timescale operation aligned record sets.

[0043] Furthermore, timemark lookup is performed on the short-scale, medium-scale, and long-scale operation record sets according to their respective timemarks. By comparing the timemark differences of record entries in the three different time scales, it is determined whether there are any missing time points. For missing timemarks, time insertion is performed according to the time order of adjacent records, so that the record entries of the three scales form a corresponding set of time nodes on the time axis. For example, based on the time step of the short-scale time scale, the records of other scales are aligned or linearly interpolated at the most recent time point. By reading the timemarks of the data record entries of the three scales and performing time sorting processing, strict time order alignment is achieved on the same time axis. The trend slope operation records, fluctuation amplitude operation records, periodic change operation records, cumulative operation records, and proportion operation records of the aligned three scales are row-level mapped according to timemarks, so that the operation content of different scales corresponding to the same time node forms a one-to-one correspondence. All the aligned operation content is continuously combined in time order to generate a multi-time scale operation aligned record set.

[0044] S2.5: Extract feature fields from the multi-timescale operation aligned record set, and concatenate the feature dimensions to generate a multi-timescale degenerate feature vector.

[0045] Furthermore, each record is read from the multi-timescale operation-aligned record set in chronological order. By locating feature fields such as trend slope calculation results, fluctuation amplitude calculation results, periodic change calculation results, cumulative amount calculation results, and proportion calculation results in the record entries, short-timescale feature fields, medium-timescale feature fields, and long-timescale feature fields are extracted into concatenable data fragments. Feature dimension concatenation is performed on the extracted three types of feature fields. By using a fixed field concatenation order, short-timescale feature fields are placed at the beginning of the feature vector, medium-timescale feature fields are placed in the middle, and long-timescale feature fields are placed at the end, so that the feature content generated at different time scales maintains a stable arrangement in the feature dimension. The concatenated feature dimension content is mapped to the same multi-timescale degradation feature vector according to the time mark, so that each time node corresponds to a complete feature expression containing short-timescale, medium-timescale, and long-timescale feature content, generating a multi-timescale degradation feature vector.

[0046] It should be noted that the feature content refers to the numerical degradation characterization results obtained at different time scales under the same time mark. For example, the slope and fluctuation amplitude of the trend of vibration or current at a short time scale, the periodic changes and fluctuation range of temperature, rotation speed or pressure at a medium time scale, and the cumulative and proportional values ​​corresponding to the cumulative running time, number of start-stops or maintenance intervals at a long time scale. The above values ​​are spliced ​​together in a fixed order to form the multi-time scale degradation feature vector of that time node.

[0047] The purpose of generating multi-timescale degradation feature vectors is to encode information reflecting transient changes, phased fluctuations and long-term cumulative degradation of equipment at different time scales into a stable arrangement of feature expressions. This enables each time point to simultaneously characterize the short-term operating status, medium-term degradation trend and long-term life evolution characteristics of the equipment, thereby providing a complete and continuous degradation characterization basis for subsequent remaining life prediction.

[0048] S3: Train a multi-timescale sub-learner based on the multi-timescale degradation feature vector, and simultaneously perform device remaining lifetime prediction calculation to generate a multi-timescale candidate prediction result package.

[0049] S3.1: Align the multi-timescale degradation feature vectors with the historical device lifecycle records and attach the device remaining lifespan label to generate a multi-timescale training sample set.

[0050] Furthermore, the multi-timescale degradation feature vectors are arranged in time-stamped order, and the time stamps in the historical equipment lifecycle records are read. The correspondence between the two time series is found by comparing the time stamps, so that the multi-timescale degradation feature vectors are matched one-to-one with the equipment lifecycle records on the time axis. For each time node that has completed time alignment, the remaining lifespan label of the corresponding time node is obtained according to the total equipment operating time, failure time, or lifespan end time in the equipment lifecycle record. The remaining lifespan label is combined with the corresponding multi-timescale degradation feature vector, so that each multi-timescale degradation feature vector is appended with a remaining lifespan label field. All time nodes that have completed label appending are structured and organized so that the multi-timescale degradation feature vectors and the remaining lifespan labels together constitute the data entries for training, generating a multi-timescale training sample set.

[0051] S3.2: Extract training samples from the multi-timescale training sample set and input them into multiple sets of machine learning sub-learners to perform parameter initialization, batch training and parameter update, and generate multi-timescale sub-learners.

[0052] Furthermore, training samples containing multi-timescale degradation feature vectors and device remaining life labels are extracted from the multi-timescale training sample set in chronological order. Data entries from multiple time points are organized into training batches that can be input into multiple sets of machine learning sub-learners. Parameter initialization is performed on each machine learning sub-learner to ensure that the parameter set is updatable before training begins. The organized training batches are input into multiple sets of machine learning sub-learners. Within each training batch, forward computation is performed based on the feature dimensions of the multi-timescale degradation feature vectors, and error information between the device remaining life label and the prediction result is read to perform parameter update operations, such as adjusting the parameter values ​​in the parameter set by the error descent direction. Forward computation and parameter updates are repeated for all training batches to ensure that each machine learning sub-learner forms a stable parameter structure under the complete coverage of the multi-timescale training sample set. All trained machine learning sub-learners are classified and organized according to short-timescale, medium-timescale, and long-timescale, and combined to form multi-timescale sub-learners.

[0053] It should be noted that, based on the structural composition of multi-timescale degradation feature vectors, machine learning sub-learners are constructed. The input feature range of each sub-learner is determined according to short, medium, and long timescales, ensuring that each type of machine learning sub-learner only receives a set of feature fields matching its corresponding timescale. The corresponding parameter structure and operational hierarchy are established based on the number and order of the feature fields, thus forming multiple sets of machine learning sub-learners that are structurally independent and differ in their focus on different timescales. This construction method allows the short-timescale machine learning sub-learners to focus on depicting transient changes during equipment operation, the medium-timescale machine learning sub-learners to focus on depicting the phased evolution of equipment status, and the long-timescale machine learning sub-learners to focus on depicting the long-term cumulative degradation characteristics of the equipment. Each machine learning sub-learner forms a dedicated prediction unit for different degradation rhythms at the structural level. Its working principle lies in separating and modeling the degradation features at different timescales at the same time point, enabling the remaining life prediction of the equipment to simultaneously reflect short-term fluctuations, medium-term trend changes, and long-term life evolution patterns, providing a structural foundation for the generation of subsequent multi-timescale candidate prediction results.

[0054] S3.3: Input the multi-timescale degradation feature vector into the multi-timescale sub-learner, perform forward operation, output the device remaining lifetime prediction value, and generate a multi-timescale candidate prediction intermediate record set.

[0055] Furthermore, the multi-timescale degradation feature vectors are sequentially input into the machine learning sub-learners within the multi-timescale sub-learners according to the time stamp order. By reading the parameter sets of the machine learning sub-learners corresponding to different time scales, forward operations are performed, so that each multi-timescale degradation feature vector yields the corresponding device remaining lifetime prediction value in the short-timescale, medium-timescale, and long-timescale machine learning sub-learners, respectively. During the forward operation, the feature dimension mapping structure of the multi-timescale degradation feature vectors ensures that the input fields are consistent with the parameter sets of the machine learning sub-learners, enabling each timescale machine learning sub-learner to output prediction results within its respective feature interest region. The device remaining lifetime prediction values ​​output by all machine learning sub-learners are recorded according to the machine learning sub-learner identifier, time scale identifier, and time stamp, so that each time node forms a record entry containing prediction results from multiple time scales. All record entries are combined into a continuous data sequence in chronological order and organized into a structured data set to generate a multi-timescale candidate prediction intermediate record set.

[0056] The formula for calculating the predicted remaining life of equipment is: ; in, This represents the predicted remaining lifespan of the equipment. A time stamp indicating the predicted remaining lifespan of the equipment. Indicates that it is within the time window Historical time markers within, Indicates the time decay coefficient. Represents the set of timescale markers, where Indicates a short time scale. Indicates the medium time scale. Indicates a long time scale. Indicated on the time scale The index of machine learning sub-learners, Indicated on the time scale The number of machine learning sub-learners configured below. Indicated on the time scale Next Each machine learning sub-learner at time stamp The device's remaining lifetime prediction is output after performing a forward operation on the multi-timescale degradation feature vectors. Indicates the current time stamp Assigned to time scale The Middle The weight values ​​of each machine learning sub-learner This represents the normalization factor.

[0057] It should be noted that the time decay factor (example range 0.85 to 0.98) is defined based on the device sampling period and degradation rate, so that the weight of adjacent moments within the window gradually decreases at a decay rate of approximately 2% to 15%.

[0058] The weight values ​​of the machine learning sub-learners (example range: 0.40-0.55 for short-timescale sub-learners, 0.25-0.40 for medium-timescale sub-learners, and 0.10-0.25 for long-timescale sub-learners) are defined based on the sensitivity and contribution of different timescales in the degradation trajectory spatial state record to the current degradation stage.

[0059] S3.4: Organize the intermediate record set of multi-timescale candidate predictions according to the machine learning sub-learner identifier, time scale identifier, and time stamp to generate a multi-timescale candidate prediction result package.

[0060] Furthermore, the intermediate record set of multi-timescale candidate predictions is classified and organized according to the machine learning sub-learner identifier. By reading the machine learning sub-learner identifier in each record, the predicted values ​​belonging to the same machine learning sub-learner are grouped into the same category. The predicted values ​​in each category are grouped according to the time scale identifier, so that the short-timescale prediction records, medium-timescale prediction records, and long-timescale prediction records form a continuous data sequence within their respective time scales. In each time scale sequence, the time markers are sorted so that the prediction records are arranged from early to late according to the time markers, and the correspondence between machine learning sub-learner identifiers, time scale identifiers, and time markers is established through a time marker mapping method. The organized record entries are combined into structured data content according to the structure of machine learning sub-learner identifiers, time scale identifiers, and time markers, so that each time node contains predicted values ​​from different time scales and different machine learning sub-learners, generating a dataset containing complete multi-timescale prediction results.

[0061] S4: Based on the degradation feature vectors at multiple time scales, construct the degradation trajectory coordinate space, and perform clustering, distance calculation, and similarity calculation to generate the degradation trajectory space state record.

[0062] S4.1: Perform numerical range unification and feature grouping on the multi-timescale degradation feature vectors, and generate degradation process features in combination with equipment lifecycle records, while setting configuration parameters for the degradation trajectory coordinate space.

[0063] Furthermore, the values ​​of each degradation feature vector at multiple time scales are read one by one. The maximum and minimum values ​​of the feature dimensions are unified with the distribution range to make the short-term, medium-term, and long-term feature fields comparable. Based on the field structure of the multi-time scale degradation feature vectors, the feature content is grouped and organized. The trend slope feature group, fluctuation amplitude feature group, periodic change feature group, cumulative amount feature group, and proportion feature group are arranged in a fixed order to form a stable structure. The life cycle stage corresponding to each time node is read in combination with the equipment life cycle record. The life cycle stage is combined with the multi-time scale degradation feature vectors to generate degradation process features. Based on the distribution of degradation process features, the configuration parameters of the degradation trajectory coordinate space are set, including the number of coordinate dimensions, the range of coordinate values, and the mapping method, so that the degradation trajectory can form a continuous expression in the degradation trajectory coordinate space.

[0064] S4.2: Convert the degradation process characteristics into degradation trajectory curves according to the configuration parameters of the degradation trajectory coordinate space, divide the equipment degradation stages, and form a set of degradation trajectory coordinate points.

[0065] Furthermore, the degradation process features are read one by one according to the time stamp order, and coordinate mapping is performed on each degradation process feature according to the configuration parameters of the degradation trajectory coordinate space. By converting the feature grouping content in the degradation process features into corresponding coordinate values ​​according to the coordinate dimension mapping method specified by the configuration parameters, the degradation degree at different time nodes can form a continuous coordinate sequence in the degradation trajectory coordinate space. By performing curve connection processing on the mapped coordinate sequence, the coordinate values ​​of continuous time nodes are combined into a degradation trajectory curve in chronological order. According to the life cycle stage division method in the equipment life cycle record, the time nodes corresponding to the degradation trajectory curve are segmented according to the life cycle stage, so that different stages in the degradation process form distinguishable stage intervals in the degradation trajectory curve. The coordinate values ​​after mapping the degradation process features at each time node are structured and organized, and a correspondence is established through time stamp, coordinate dimension and life cycle stage, so that the coordinate content of all time nodes forms a set of degradation trajectory coordinate points.

[0066] S4.3: Based on the set of coordinate points of the degradation trajectory, perform initial segmentation and clustering operations during the equipment degradation stage to generate a degradation trajectory spatial cluster.

[0067] Furthermore, based on the set of degraded trajectory coordinate points, the time stamp, coordinate values, and lifecycle stages are read. Initial segmentation is performed on the set of degraded trajectory coordinate points using a lifecycle stage division method, forming continuous initial segment intervals on the time axis for coordinate points within the same lifecycle stage. Clustering operations are then performed on the corresponding coordinate point sequences for each initial segment interval based on the distance relationships, directional changes, and numerical trends between coordinate points, merging coordinate points with similar degradation feature change patterns within the same initial segment interval into the same cluster group. Cluster groups formed from different initial segment intervals are classified according to their lifecycle stages, establishing a mapping relationship between the lifecycle stages of the initial segment intervals and the cluster groups. All cluster groups are then organized according to the lifecycle stage order and coordinate point distribution, forming independent structured spatial blocks for each cluster group. These blocks are then aggregated into multiple cluster structures representing the feature distribution of different degradation stages, generating a degraded trajectory spatial cluster.

[0068] S4.4: Map the current time-scale degradation feature vector to the current degradation coordinate point, and perform distance and similarity calculations with the degradation trajectory space cluster to generate a degradation trajectory space state record.

[0069] Furthermore, the multi-timescale degradation feature vector at the current moment is subjected to feature grouping mapping according to the configuration parameters of the degradation trajectory coordinate space. By converting the trend slope feature group, fluctuation amplitude feature group, periodic change feature group, cumulative amount feature group, and proportion feature group into corresponding coordinate dimension values, the multi-timescale degradation feature vector at the current moment forms the current degradation coordinate point. The distribution of coordinate points within each cluster structure in the degradation trajectory space cluster is read, and distance calculation is performed between the current degradation coordinate point and the coordinate points in the cluster structure. The distance metric from the current degradation coordinate point to the cluster structure is obtained through multi-dimensional coordinate difference and change direction difference. Similarity calculation is performed on the distance metric. The similarity value between the current degradation coordinate point and the degradation trajectory space cluster is obtained by normalizing or reverse mapping the distance value, so that the degradation state at the current moment can be located based on the degradation trajectory space cluster. The current degradation coordinate point, distance calculation result, and similarity calculation result are recorded as the state content of the corresponding time node in a structured format to generate the degradation trajectory space state record.

[0070] It should be noted that the degradation trajectory space state record is used to reflect the overall degradation position and degradation stage at the current time point of the equipment. By simultaneously depicting the specific coordinate distribution of the current degradation coordinate point in the degradation trajectory coordinate space, the distance relationship with each degradation trajectory space cluster, and the corresponding similarity level, it comprehensively characterizes the current degree of degradation, the degradation evolution stage, and the degree of similarity with historical degradation patterns of the equipment. This provides a state basis for the subsequent dynamic weighted combination of the equipment remaining life prediction results at different time scales based on the differences in degradation status.

[0071] S5: Based on the spatial state record of the degradation trajectory, dynamically weight and combine the predicted remaining life of the equipment corresponding to different time scales in the multi-time scale candidate prediction result package to generate a predicted remaining life of the equipment.

[0072] S5.1: Based on the degradation trajectory space state record, assign basic weight values ​​to the predicted remaining life of the equipment corresponding to different time scales in the multi-time scale candidate prediction result package, and generate a dynamic weighted configuration record set for the degradation trajectory space.

[0073] Furthermore, based on the degradation trajectory spatial state record, the current degradation coordinate point, distance calculation result, and similarity calculation result are read. Interval mapping is performed on the similarity values ​​to determine the basic weight distribution pattern of the predicted remaining life of the equipment at different time scales. This ensures that the basic weight values ​​of the predicted remaining life of the equipment at short, medium, and long time scales correspond to the degree of degradation. Based on the similarity distribution results in the degradation trajectory spatial state record, basic weight values ​​are assigned to the predicted remaining life of the equipment at each time scale in the multi-time scale candidate prediction result package. Each basic weight value is then bound to its corresponding time scale identifier, ensuring that each predicted remaining life of the equipment at each time scale receives an independent basic weight configuration. All basic weight values ​​are then structured according to the time scale order and written into a unified dataset to generate a dynamically weighted configuration record set for the degradation trajectory space.

[0074] It should be noted that when the similarity is in the high similarity range (e.g., 0.7 to 1.0), the base weight values ​​of the corresponding multi-timescale sub-learners are allocated to the high weight range (e.g., 0.6 to 0.8), so that this time scale has a stronger contribution to the determination of the degradation state; when the similarity is in the medium similarity range (e.g., 0.4 to 0.7), the base weight values ​​of the corresponding multi-timescale sub-learners are allocated to the medium weight range (e.g., 0.3 to 0.6), so that it maintains a medium influence in dynamic weighting; when the similarity is in the low similarity range (e.g., 0.0 to 0.4), the base weight values ​​of the corresponding multi-timescale sub-learners are allocated to the low weight range (e.g., 0.1 to 0.3), so that its influence in the prediction of the degradation state is reduced. This segmented mapping method converts the similarity values ​​in the degradation trajectory spatial state record into the base weight values ​​of the multi-timescale sub-learners.

[0075] S5.2: Read the predicted remaining lifespan of the equipment and its corresponding weights at different time scales from the dynamically weighted configuration record set of the degradation trajectory space, and perform dynamic weighted combination to generate a multi-time scale weight distribution record.

[0076] Furthermore, the predicted remaining lifetime of equipment (RNB) and its corresponding basic weight values ​​for different time scales are read from the dynamic weighted configuration record set of the degradation trajectory space according to the time stamp order. The RSB predicted values ​​for short, medium, and long time scales under the same time stamp are then calculated with their corresponding basic weight values ​​to form weighted prediction quantities for each time scale. At the same time node, the weighted prediction quantities for each time scale are weighted and combined to merge the RSB predicted values ​​for multiple time scales into a unified weighted result. Based on the distribution of the basic weight values, the weight values ​​for each time scale are normalized to ensure comparability of the RSB predicted values ​​for different time scales within the overall combination. The normalized weight values, along with the weighted prediction quantities for each time scale, the corresponding time scale identifiers, and the time stamp, are recorded as structured fields. A continuous relationship is established through the time stamps to generate a multi-time scale weight distribution record containing the weighted result, normalized weights, and corresponding time scale information, which is used to generate the RSB prediction report.

[0077] S5.3: Integrate the equipment remaining life prediction value, multi-timescale weight distribution record, degradation trajectory coordinate point set, and multi-timescale degradation feature vector into an equipment remaining life prediction report.

[0078] Furthermore, the predicted remaining lifetime of the equipment at the corresponding time node in the multi-timescale candidate prediction result package is combined with the weighted results, normalized weights, and timescale structure in the multi-timescale weight distribution record according to the corresponding time markers. The degradation state expression is supplemented by reading the coordinate values ​​from the degradation trajectory coordinate point set and the lifecycle stage information. For example, when the degradation trajectory coordinate point is located within the trajectory space cluster corresponding to the accelerated degradation stage, and the lifecycle stage information indicates that the equipment has entered the accelerated degradation stage, the corresponding predicted remaining lifetime value is associated and labeled as the prediction result for the accelerated degradation stage in the equipment remaining lifetime prediction report, thus supplementing the expression of the equipment's current state. The system reads the short-term, medium-term, and long-term feature fields from the multi-timescale degradation feature vector at each time point, and combines the feature content with the predicted remaining life of the equipment, weight information, and the set of degradation trajectory coordinates into a unified structured record. The structured record is then uniformly arranged to ensure a complete representation of all information at each time point. Finally, the structured records from all time points are summarized in chronological order, ensuring a continuous correlation between changes in the predicted remaining life of the equipment, changes in multi-timescale weights, changes in the degradation trajectory, and feature evolution within the same report structure, generating a predicted remaining life of the equipment report.

[0079] This embodiment also provides a device remaining life prediction system based on AI ensemble learning, including: The data acquisition module is used to collect multi-source operational data and preprocess it. The preprocessed multi-source operational data is divided into multiple sliding time windows to generate multi-timescale operational data packages. The feature generation module is used to perform different data operations in the running data packets at multiple time scales, and to align the operation results with time stamps and concatenate the feature dimensions to generate multi-time scale degenerate feature vectors. The learner training module is used to train multi-timescale sub-learners based on multi-timescale degradation feature vectors, while performing device remaining lifetime prediction calculations and generating multi-timescale candidate prediction result packages. The degradation operation module is used to construct the degradation trajectory coordinate space based on the degradation feature vectors of multiple time scales, and perform clustering segmentation, distance operation and similarity operation to generate the degradation trajectory space state record; The report generation module is used to dynamically weight and combine the predicted remaining life of equipment at different time scales in the multi-time scale candidate prediction result package based on the degradation trajectory spatial state record, and generate an equipment remaining life prediction report.

[0080] In summary, this invention achieves accurate identification of equipment degradation status by constructing a degradation trajectory coordinate space and performing clustering and similarity calculations, providing a basis for the selection of machine learning sub-models and improving the stage adaptability of prediction; and improves the accuracy and robustness of RUL prediction by dynamically weighting and fusing multi-scale machine learning prediction results based on state records to achieve adaptive optimization of the integrated strategy.

[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A device residual life prediction method based on AI ensemble learning, characterized in that: The application relates to a device residual life prediction method and device. Collecting multi-source operation data and preprocessing, the preprocessed multi-source operation data is cut according to multiple groups of sliding time windows to generate multi-time scale operation data packets; Different data operations are performed in the multi-time scale operation data packets, and the operation results are time-labeled, aligned and feature-dimension spliced to generate multi-time scale degradation feature vectors; Based on the multi-time scale degradation feature vectors, multi-time scale sub-learners are trained, and device residual life prediction operations are simultaneously performed to generate multi-time scale candidate prediction result packets; According to the multi-time scale degradation feature vectors, a degradation trajectory coordinate space is constructed, and clustering segmentation, distance operation and similarity operation are performed to generate a degradation trajectory space state record; According to the degradation trajectory space state record, the device residual life prediction values corresponding to different time scales in the multi-time scale candidate prediction result packet are dynamically weighted and combined to generate a device residual life prediction report. 2.The AI ensemble learning-based device remaining life prediction method of claim 1, wherein: The preprocessed multi-source operation data is cut according to multiple groups of sliding time windows to generate multi-time scale operation data packets, and the steps are as follows, Based on the preprocessed multi-source operation data, multiple groups of sliding time window parameters are configured through short-time scale, medium-time scale and long-time scale to generate a sliding time window configuration record set; According to the sliding time window configuration record set, window data segments are intercepted in the preprocessed multi-source operation data set, and are recombined into multi-time scale operation data packets according to different time scales. 3.The AI ensemble learning-based device remaining life prediction method of claim 2, wherein: Different data operations are performed in the multi-time scale operation data packets, and the operation results are time-labeled, aligned and feature-dimension spliced to generate multi-time scale degradation feature vectors, and the steps are as follows, Trend slope operation, fluctuation amplitude operation and periodic change operation are performed on vibration, current and voltage data in the multi-time scale operation data packet to generate a short-time scale operation record set; Trend slope operation, fluctuation amplitude operation and periodic change operation are performed on temperature, speed, torque, pressure and flow in the multi-time scale operation data packet to generate a medium-time scale operation record set; Cumulative operation time, start-stop times and maintenance intervals in the multi-time scale operation data packet are subjected to cumulative operation and proportional operation to generate a long-time scale operation record set. 4.The AI ensemble learning-based device remaining life prediction method of claim 3, wherein: The operation results are time-labeled, aligned and feature-dimension spliced to generate multi-time scale degradation feature vectors, and the steps are as follows, Time label searching and inserting are performed on the short-time scale operation record set, the medium-time scale operation record set and the long-time scale operation record set, and the sets are aligned in time sequence to generate a multi-time scale operation alignment record set; Feature fields are extracted from the multi-time scale operation alignment record set, and feature dimension splicing is performed to generate a multi-time scale degradation feature vector. 5.The AI ensemble learning-based device remaining life prediction method of claim 4, wherein: The multi-time scale sub-learners are trained based on the multi-time scale degradation feature vectors, and the steps are as follows, The multi-time scale degradation feature vectors are time-aligned with the device life cycle records in the historical operation and are attached with device residual life labels to generate a multi-time scale training sample set; Training samples are extracted from the multi-time scale training sample set, and are input into multiple machine learning sub-learners to perform parameter initialization, batch training and parameter updating to generate multi-time scale sub-learners. 6.The AI ensemble learning-based device remaining life prediction method of claim 5, wherein: The remaining lifetime prediction calculation of the execution equipment generates a multi-timescale candidate prediction result package, and the steps are as follows. The multi-timescale degradation feature vector is input into the multi-timescale sub-learner, forward operation is performed, the remaining life of the device is output as the predicted value of the device's remaining life is generated, and a multi-timescale candidate prediction intermediate record set is generated. The intermediate record set of multi-timescale candidate predictions is organized according to the machine learning sub-learner identifier, time scale identifier, and time stamp to generate a multi-timescale candidate prediction result package. 7.The AI ensemble learning-based device remaining life prediction method of claim 6, wherein: The steps for constructing the degradation trajectory coordinate space based on multi-timescale degradation feature vectors are as follows. The numerical range of the degradation feature vectors at multiple time scales is unified and the features are grouped and organized. The degradation process features are generated by combining the equipment life cycle records. At the same time, the configuration parameters of the degradation trajectory coordinate space are set. The degradation process characteristics are converted into degradation trajectory curves according to the configuration parameters of the degradation trajectory coordinate space, and the equipment degradation stages are divided to form a set of degradation trajectory coordinate points. 8.The AI integration learning-based device remaining life prediction method of claim 7, wherein: The steps for performing clustering segmentation, distance calculation, and similarity calculation to generate degenerate trajectory spatial state records are as follows. Based on the set of coordinate points of the degradation trajectory, initial segmentation and clustering operations are performed during the equipment degradation stage to generate a degradation trajectory spatial cluster; The current time-scale degradation feature vector is mapped to the current degradation coordinate point, and distance and similarity calculations are performed with the degradation trajectory space cluster to generate a degradation trajectory space state record. 9.The AI integration learning-based device remaining life prediction method of claim 8, wherein: The step of dynamically weighting and combining the predicted remaining life of equipment at different time scales in the multi-time-scale candidate prediction result package based on the degradation trajectory spatial state record to generate an equipment remaining life prediction report is as follows. Based on the degradation trajectory space state record, basic weight values ​​are assigned to the predicted remaining life of the equipment corresponding to different time scales in the multi-time scale candidate prediction result package, and a dynamic weighted configuration record set of degradation trajectory space is generated. The predicted remaining lifespan of the equipment and its corresponding weights at different time scales are read from the dynamically weighted configuration record set of the degradation trajectory space, and then dynamically weighted and combined to generate a multi-time scale weight distribution record. The remaining life prediction of the equipment, the multi-timescale weight distribution record, the set of degradation trajectory coordinate points, and the multi-timescale degradation feature vector are integrated into the remaining life prediction report of the equipment.

10. An AI ensemble learning-based equipment residual life prediction system based on the AI ensemble learning-based equipment residual life prediction method of any one of claims 1 to 9, characterized by: include, The data acquisition module is used to collect multi-source operational data and preprocess it. The preprocessed multi-source operational data is divided into multiple sliding time windows to generate multi-timescale operational data packages. The feature generation module is used to perform different data operations in the running data packets at multiple time scales, and to align the operation results with time stamps and concatenate the feature dimensions to generate multi-time scale degenerate feature vectors. The learner training module is used to train multi-timescale sub-learners based on multi-timescale degradation feature vectors, while performing device remaining lifetime prediction calculations and generating multi-timescale candidate prediction result packages. The degradation operation module is used to construct the degradation trajectory coordinate space based on the degradation feature vectors of multiple time scales, and perform clustering segmentation, distance operation and similarity operation to generate the degradation trajectory space state record; The report generation module is configured to generate a device residual life prediction report by dynamically weighting and combining the device residual life prediction values corresponding to different time scales in the multi-time-scale candidate prediction result package according to the degradation trajectory space state records.

Images