A diesel engine fault prediction method based on big data analysis

Abstract

The application discloses a diesel engine fault prediction method based on big data analysis and relates to the technical field of big data analysis, which comprises the following steps: collecting diesel engine multi-source state data and preprocessing to generate a standardized data set; performing working condition segmentation, phase marking and transition arrangement to generate a working condition chain sample set; constructing a working condition parameter baseline, extracting a deviation to generate a multi-dimensional running residual set; decomposing the residual by using MSTL and correcting the misclassified fragments to generate a degradation residual feature package; constructing an improved MLP-Mixer network to generate a diesel engine degradation prediction package; performing degradation trajectory segmentation and risk mapping to generate a fault risk prediction package; marking the early warning level and matching the guarantee task window to generate a maintenance scheduling. By introducing the MSTL algorithm and the improved MLP-Mixer network, the application realizes accurate extraction of early degradation features of the diesel engine under complex task working conditions, joint prediction of fault risks and generation of predictive maintenance scheduling.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a method for predicting diesel engine faults based on big data analytics. Background Technology

[0002] With the development of diesel engine condition monitoring technology, sensor acquisition technology, and big data analysis technology, diesel engine fault prediction is gradually shifting from periodic maintenance, manual experience judgment, and single-parameter threshold alarms to a health management approach based on multi-source condition data. Existing methods typically collect data such as engine speed and load, combustion pressure, lubricating oil temperature and pressure, coolant temperature and flow rate, intake and exhaust pressure and temperature, engine vibration, electronic control response, operating conditions, historical faults, and maintenance records. Through data cleaning, time-series alignment, feature extraction, and model prediction, abnormal operating conditions of the diesel engine are identified. Compared to fixed-cycle maintenance, diesel engine fault prediction based on big data analysis can continuously track the operating status of the diesel engine, identifying early signs of faults such as combustion abnormalities, lubrication degradation, cooling abnormalities, intake and exhaust abnormalities, vibration transmission abnormalities, and electronic control response deviations, providing a data foundation for diesel engine operation assurance and predictive maintenance.

[0003] Existing diesel engine fault prediction methods still have shortcomings under complex task conditions. Diesel engine operating conditions are affected by speed cycles, load cycles, task phase switching, start-stop processes, and maintenance recovery processes. Normal operating condition fluctuations and early degradation characteristics are easily confused. Traditional fixed threshold, overall trend, or ordinary time-series characteristic analysis methods are insufficient to distinguish between normal fluctuations and abnormal residuals. Early fault characteristics are easily absorbed by periodic or trend components. Some methods lack sufficient segmentation of task conditions, task phase marking, and arrangement of condition transition relationships, making it difficult to establish baselines for sub-condition parameters and extract comparable multi-dimensional operating residuals. Prediction models fail to adequately express the sequential relationships of cross-component responses among combustion, intake and exhaust, vibration, cooling, lubrication, and electronic control responses, easily misjudging component propagation processes as independent anomalies. Prediction results often remain at the level of anomaly scoring, health index, or fault category judgment, failing to link fault probability, fault component attribution, remaining service life, maintenance deadline, and support task window, making it difficult to directly translate into dynamic risk warning results and predictive maintenance scheduling.

[0004] Therefore, how to provide a diesel engine fault prediction method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a diesel engine fault prediction method based on big data analysis. This invention utilizes multi-source state data fusion of diesel engines, task condition chain arrangement, baseline construction of sub-condition parameters, MSTL degradation residual development, and improved MLP-Mixer degradation prediction technology. It details the implementation process of diesel engine state data standardization, condition phase division, operational residual extraction, degradation feature decomposition, cross-component response prediction, and maintenance schedule generation under complex task conditions. This achieves the joint output of diesel engine fault probability, fault component attribution, remaining service life, dynamic risk warning, and predictive maintenance schedule. Compared to traditional single-parameter threshold alarms, fixed-cycle maintenance, and ordinary time-series prediction methods, this invention has the advantages of strong adaptability to complex operating conditions, low interference from periodic fluctuations, accurate identification of early degradation features, sufficient expression of cross-component response relationships, and fault prediction results that can directly support maintenance decisions within the task window.

[0006] A diesel engine fault prediction method based on big data analysis according to an embodiment of the present invention includes:

[0007] Collect multi-source state data of diesel engines, perform preprocessing on the multi-source state data of diesel engines, and generate a standardized multi-source state dataset of diesel engines;

[0008] Based on the standardized diesel engine multi-source state dataset, task condition segmentation, task phase labeling, and condition transition relationship arrangement are performed to generate a diesel engine condition chain sample set.

[0009] Based on the diesel engine operating condition chain sample set, construct the sub-operating condition parameter baseline, extract the deviation between the actual operating parameters and the sub-operating condition parameter baseline, and generate a multi-dimensional operating residual set;

[0010] The MSTL algorithm is used to perform multi-period trend decomposition on the multidimensional running residual set. By replacing abnormal residual segments and correcting misclassified degenerate segments, a diesel engine degradation residual feature package is generated.

[0011] An improved MLP-Mixer network is constructed to convert diesel engine degradation residual feature packets into operating window tokens and difference tokens. By embedding the baseline difference under the same operating condition and the phase difference under the same task, separating the convergence recovery residual and the continuous deviation residual, and aligning the cross-component misalignment response segments, a diesel engine degradation prediction packet is generated.

[0012] Based on the diesel engine degradation prediction package, degradation trajectory segmentation and risk state mapping are performed to calculate the failure probability, fault component attribution and remaining service life, and a diesel engine failure risk prediction package is generated.

[0013] Based on the diesel engine failure risk prediction package, the warning level is marked, the maintenance deadline is deduced from the remaining service life, the support task window is matched, and dynamic risk warning results and predictive maintenance schedules are generated.

[0014] Optionally, the multi-source status data of the diesel engine specifically includes diesel engine speed and load data, combustion pressure data, lubricating oil temperature and pressure data, cooling water temperature and flow data, intake and exhaust pressure and temperature data, engine vibration data, electronic control response data, task operating condition data, historical fault data, business maintenance records, and task support data.

[0015] Optionally, the preprocessing of the multi-source state data of the diesel engine includes:

[0016] According to the diesel engine number, component number, sensor number and acquisition time, the multi-source status data of the diesel engine are merged to form diesel engine status merged data;

[0017] The diesel engine status data is merged and the unit is unified, the sampling interval is reorganized, and the timestamp is corrected to form a sequence of diesel engine operating parameters arranged according to the time of collection.

[0018] Based on the task conditions, start-stop status, and business maintenance nodes, the diesel engine operating parameter sequence is marked with task stages, start-stop boundaries, load switching points, and maintenance recovery sections.

[0019] Remove abnormal sampling values ​​that exceed the sensor's range, deviate from the adjacent sampling distribution under the same operating conditions, or do not meet the continuity of component operation. Complete the missing sampling segments according to the adjacent cycles under the same operating conditions and the response changes of adjacent components.

[0020] The diesel engine operating parameter sequence is time-aligned with task phases, start-stop boundaries, load switching points, maintenance and recovery phases, historical fault nodes, and business maintenance nodes to generate a standardized diesel engine multi-source state dataset.

[0021] Optionally, generating the diesel engine operating condition chain sample set includes:

[0022] The standardized diesel engine multi-source state dataset is arranged according to the collection time, and the continuous operation segments of the diesel engine are divided according to the start-stop boundary, load switching point and maintenance recovery section.

[0023] Based on the changes in engine speed, load, combustion pressure, coolant temperature, and engine vibration during continuous operation segments of the diesel engine, mark the starting, idling, acceleration, high load, load reduction, cooling recovery, shutdown, and maintenance recovery conditions;

[0024] Based on the task condition data, the corresponding running segments of each condition are marked with task phases, and task phase segments containing phase number, start and end time, duration and condition category are generated.

[0025] Compare the operating condition categories, load change directions, and maintenance recovery status of adjacent task phase segments, and arrange the transition directions and transition sequences between adjacent operating conditions;

[0026] The task phase segments, transition directions, transition sequences, operating parameter sequences, historical fault nodes, and operational maintenance nodes are combined according to the diesel engine number to generate a diesel engine operating condition chain sample set.

[0027] Optionally, generating the multidimensional running residual set includes:

[0028] The diesel engine operating condition chain sample set was grouped according to diesel engine number, task phase, operating condition category and transition order, and the corresponding diesel engine operating parameters of each group were extracted.

[0029] Steady-state segments are screened for various operating parameters under the same working condition category and the same task phase. Non-steady-state sampled values ​​within the start-stop boundary, load switching point and maintenance recovery section are removed to form a baseline sample for each working condition.

[0030] Based on the time of data collection, the median value and fluctuation range of parameters of the baseline samples under different working conditions are calculated to generate the parameter baselines for different working conditions.

[0031] The actual operating parameters are matched with the baseline parameters of the sub-operating conditions under the corresponding operating condition category, task phase and acquisition time. The operating parameter deviation is obtained by subtracting the parameter baseline value from the actual operating parameter value.

[0032] A multidimensional operational residual set is generated by combining the operating parameter deviations based on the diesel engine number, component number, task phase, and data acquisition time.

[0033] Optionally, the generation of the diesel engine degradation residual feature package includes:

[0034] The multidimensional operational residual set is organized into a residual decomposition sequence according to the diesel engine number, component number, task phase and residual category. The MSTL decomposition cycle order is determined according to the speed cycle, load cycle, task phase cycle and maintenance recovery cycle.

[0035] Before fitting the seasonal term in each decomposition cycle, a periodic segment replacement process is performed. Segments that continuously deviate from the fluctuation range of the same working condition and have the same residual direction are marked as abnormal residual segments. Normal residual segments at the same position are selected in adjacent phase cycles of the same task to replace abnormal residual segments, thus forming a seasonal term fitting sequence.

[0036] MSTL seasonal term fitting is performed using the seasonal term fitting sequence to generate the periodic components of each decomposition period, and the residual decomposition sequence is successively subtracted from the corresponding periodic components to generate the remaining residual sequence.

[0037] The remaining residual sequence is fitted with the MSTL trend term to generate the trend component, and the initial degraded residual sequence is generated by subtracting the trend component from the remaining residual sequence.

[0038] Component back-injection and decomposition are performed between the periodic component, trend component, and initial degraded residual sequence. Fragments in the periodic component and trend component that overlap with abnormal residual fragments and are not retained in the initial degraded residual sequence are marked as misclassified degraded fragments.

[0039] The misclassified degenerate segments are subtracted from the periodic and trend components, and the misclassified degenerate segments are reinjected into the initial degenerate residual sequence. Then, MSTL decomposition is performed again to generate degenerate residual components.

[0040] By combining the diesel engine number, component number, task phase, residual category, and acquisition time, degraded residual components, periodic components, trend components, abnormal residual segment markers, and misclassified degraded segment markers, a diesel engine degradation residual feature package is generated.

[0041] Optionally, the generation of the diesel engine degradation prediction package includes:

[0042] An improved MLP-Mixer network is constructed, which includes a differential token embedding layer, a recovery degradation splitting layer, and a component misalignment mixing layer.

[0043] Diesel engine degradation residual feature package is divided into operating windows. Degradation residual components, periodic components, trend components and segment markers in each operating window are arranged to generate operating window tokens.

[0044] The difference token embedding layer calculates the degradation residual difference between the current running window and the baseline window of the same operating condition, as well as the degradation residual difference between the current running window and the historical window of the same task phase. It embeds the two types of degradation residual differences into the running window token to generate the difference token.

[0045] The degradation shunt layer divides the difference tokens into recovery tokens and degradation tokens based on the difference magnitude change, residual direction, and convergence status of the maintenance recovery section during the continuous operation window.

[0046] The component misalignment mixing layer performs front-to-back misalignment alignment and channel mixing on the combustion residuals, intake and exhaust residuals, vibration residuals, cooling residuals, lubrication residuals and electronic control response residuals in the recovery state tokens and degradation state tokens to generate cross-component mixed tokens;

[0047] Perform time-dimensional token mixing and residual channel mixing on cross-component mixed tokens to generate degradation trend values, degradation state markers and prediction time steps for each component, and combine them to form a diesel engine degradation prediction package;

[0048] The improved MLP-Mixer network is trained by labeling samples before historical fault nodes, samples after maintenance and recovery stages, and baseline samples in the same task phase as degraded samples, recovered samples, and baseline samples, respectively. Degradation trend error, degradation state labeling error, prediction time step error, recovery degradation shunting error, and component misalignment error are calculated. The various errors are combined as a comprehensive error and the network parameters of each layer are updated. The training of the improved MLP-Mixer network is completed when the comprehensive error of the verification samples in the same task phase decreases by less than 0.002 for four consecutive rounds.

[0049] Optionally, the generation of the diesel engine failure risk prediction package includes:

[0050] According to the diesel engine number, component number, task phase, and prediction time step, sort the degradation trend value and degradation status marker in the diesel engine degradation prediction package to generate the degradation trajectory of each component;

[0051] The degradation trajectory of each component is segmented based on the continuous rise, sudden increase, plateau, and fall of the degradation trend value, and mapped to normal, attention, alarm, and fault imminent risk states.

[0052] The failure probability is calculated based on the risk status, degradation trend value, predicted time step and historical failure node matching results. The component with the highest failure probability and consistent with the cross-component mixed token response order is marked as the failure component.

[0053] The time interval between the time of the near-fault risk state and the current data collection time is determined as the remaining service life;

[0054] A diesel engine failure risk prediction package is generated by combining the diesel engine number, task phase, failure probability, fault component attribution, remaining service life, and risk status.

[0055] Optionally, the generation of dynamic risk warning results and predictive maintenance schedules includes:

[0056] The diesel engine failure risk prediction package is analyzed to extract failure probability, failure component attribution, remaining service life and risk status, and warning levels are marked according to failure probability and risk status.

[0057] Add the remaining service life to the current data collection time to obtain the maintenance deadline, and combine the faulty component attribution, warning level and maintenance deadline to generate a fault warning item;

[0058] Match fault warning items to the support task window no later than the maintenance deadline, and determine the maintenance sequence according to the fault probability from high to low and the remaining service life from short to long;

[0059] Based on the attribution of the faulty component, the warning level, the repair deadline, the support task window, and the repair sequence, dynamic risk warning results and predictive maintenance schedules are generated.

[0060] Optionally, the support task window specifically includes the task execution time period, the downtime for maintenance, the spare parts availability time period, the availability time period for maintenance personnel, the availability time period for special tools, and the allowable time period for component replacement. Maintenance periods that overlap with the task execution time period are excluded, and the overlapping time periods of the downtime for maintenance, the availability time period for spare parts, the availability time period for maintenance personnel, the availability time period for special tools, and the allowable time period for component replacement are used as the support task window.

[0061] The beneficial effects of this invention are:

[0062] This invention proposes a diesel engine fault prediction method based on big data analysis. It collects multi-source state data of the diesel engine and performs standardized preprocessing. Combining task condition segmentation, task phase marking, and arrangement of condition transition relationships, it generates a diesel engine condition chain sample set. Compared to traditional methods relying on single sensor parameters, fixed thresholds, or overall trend judgments, this invention can maintain the consistency of operating state samples and phase comparability under complex task conditions, reducing the interference of start-stop, load switching, and maintenance recovery processes on fault prediction results.

[0063] This invention constructs baseline parameters for different operating conditions based on a diesel engine operating condition chain sample set, extracts a multi-dimensional operating residual set, and uses the MSTL algorithm for multi-period trend decomposition. Through periodic segment replacement processing and component back-injection and re-decomposition processing, a diesel engine degradation residual feature package is generated. This process can distinguish between normal periodic fluctuations, trend changes, and early degradation residuals, reducing the risk of early fault features being absorbed by periodic or trend components, and improving the accuracy of identifying fault precursors such as combustion anomalies, lubrication degradation, cooling anomalies, intake and exhaust anomalies, vibration transmission anomalies, and electronic control response deviations.

[0064] This invention generates diesel engine degradation prediction packets by improving the differential token embedding layer, the recovery degradation shunt layer, and the component misalignment mixing layer in the MLP-Mixer network. It then calculates the failure probability, faulty component attribution, and remaining service life, generating dynamic risk warning results and predictive maintenance schedules. This method can express the response sequence across components, reducing the possibility of components being misjudged as independent anomalies during propagation. It also enables the failure prediction results to be linked to maintenance deadlines and support task windows, improving the reliability of diesel engine operation support and predictive maintenance decisions. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a flowchart of a diesel engine fault prediction method based on big data analysis proposed in this invention;

[0067] Figure 2 This is a schematic diagram of the structure of the diesel engine degradation residual feature package generated by the MSTL algorithm in the diesel engine fault prediction method based on big data analysis proposed in this invention.

[0068] Figure 3 This is a schematic diagram of the structure of an improved MLP-Mixer network for a diesel engine fault prediction method based on big data analysis proposed in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figure 1 , Figure 2 and Figure 3 A diesel engine fault prediction method based on big data analysis includes:

[0071] Collect multi-source state data of diesel engines, perform preprocessing on the multi-source state data of diesel engines, and generate a standardized multi-source state dataset of diesel engines;

[0072] Based on the standardized diesel engine multi-source state dataset, task condition segmentation, task phase labeling, and condition transition relationship arrangement are performed to generate a diesel engine condition chain sample set.

[0073] Based on the diesel engine operating condition chain sample set, construct the sub-operating condition parameter baseline, extract the deviation between the actual operating parameters and the sub-operating condition parameter baseline, and generate a multi-dimensional operating residual set;

[0074] The MSTL algorithm is used to perform multi-period trend decomposition on the multidimensional running residual set. By replacing abnormal residual segments and correcting misclassified degenerate segments, a diesel engine degradation residual feature package is generated.

[0075] An improved MLP-Mixer network is constructed to convert diesel engine degradation residual feature packets into operating window tokens and difference tokens. By embedding the baseline difference under the same operating condition and the phase difference under the same task, separating the convergence recovery residual and the continuous deviation residual, and aligning the cross-component misalignment response segments, a diesel engine degradation prediction packet is generated.

[0076] Based on the diesel engine degradation prediction package, degradation trajectory segmentation and risk state mapping are performed to calculate the failure probability, fault component attribution and remaining service life, and a diesel engine failure risk prediction package is generated.

[0077] Based on the diesel engine failure risk prediction package, the warning level is marked, the maintenance deadline is deduced from the remaining service life, the support task window is matched, and dynamic risk warning results and predictive maintenance schedules are generated.

[0078] In this embodiment, the multi-source status data of the diesel engine specifically includes diesel engine speed and load data, combustion pressure data, lubricating oil temperature and pressure data, cooling water temperature and flow data, intake and exhaust pressure and temperature data, engine vibration data, electronic control response data, task operating condition data, historical fault data, business maintenance records, and task support data.

[0079] In this embodiment, the preprocessing of the multi-source state data of the diesel engine includes:

[0080] According to the diesel engine number, component number, sensor number and acquisition time, the multi-source status data of the diesel engine are merged to form diesel engine status merged data;

[0081] The diesel engine status data is processed by unifying the unit of analysis, resetting the sampling interval, and correcting the timestamp, resulting in a sequence of diesel engine operating parameters arranged according to the time of data collection.

[0082] The diesel engine operating parameter sequence is formed as follows:

[0083] The units of parameters in the diesel engine status data are unified: the speed is unified to revolutions per minute, the load is unified to a percentage, the combustion pressure, lubricating oil pressure and intake pressure are unified to megapascals, the lubricating oil temperature, coolant temperature and exhaust temperature are unified to degrees Celsius, the engine vibration is unified to multiples of gravitational acceleration, and the electronic control response delay is unified to milliseconds.

[0084] A sequence of acquisition time points is established with a 1-second sampling interval. For high-frequency acquisition of combustion pressure data and body vibration data, the peak value of combustion pressure, the average value of combustion pressure, and the root mean square value of vibration are calculated within each 1-second window. For low-frequency acquisition of temperature, pressure, and flow data, the parameter values ​​at the corresponding acquisition time points are supplemented according to the slope of change between adjacent sampling points.

[0085] The time of data collection by each sensor is compared with the diesel engine control clock corresponding to the electronic control response data. When the time offset exceeds 0.5 seconds, timestamp correction is performed based on adjacent stable speed points, load rapid change inflection points, and electronic control response trigger points.

[0086] After calibration, the various operating parameters are arranged according to the time of acquisition to form a diesel engine operating parameter sequence;

[0087] Based on the task conditions, start-stop status, and operational maintenance nodes, the diesel engine operating parameter sequence is marked with task stages, start-stop boundaries, load switching points, and maintenance recovery segments, where:

[0088] The task phase, start / stop boundary, load switching point, and maintenance recovery section are marked as follows:

[0089] Based on the task number, task intensity, and task duration range in the task condition data, the diesel engine operating parameter sequence is marked as the task stage, the continuous power output range is marked as the task execution stage, the low load waiting range is marked as the standby stage, and the temperature drop range after the task ends is marked as the cooling recovery stage.

[0090] The start-stop boundary is marked according to the start-stop status and speed change. When the speed rises from 0 to the idle stable speed and lasts for more than 30 seconds, the starting point of the speed rise is marked as the start boundary. When the speed drops from the idle stable speed to 0 and lasts for more than 30 seconds, the ending point of the speed drop is marked as the stop boundary.

[0091] The load switching point is marked based on the magnitude and duration of the load change. When the load changes by more than 15% within 20 consecutive seconds, and at least one of the speed, combustion pressure or exhaust temperature changes synchronously, the starting point of the change is marked as the load switching point.

[0092] The maintenance recovery segment is marked according to the business maintenance node and the parameter fallback status after maintenance. The corresponding time of the business maintenance record is taken as the maintenance start point. When the lubricating oil pressure, cooling water temperature, machine vibration or exhaust temperature gradually return to the stable operating range near the same task stage before maintenance within the continuous operation window after maintenance, the corresponding continuous fallback interval is marked as the maintenance recovery segment.

[0093] Abnormal sampling values ​​that exceed the sensor's range, deviate from the adjacent sampling distribution under the same operating condition, or do not meet the continuity of component operation are removed. Missing sampling segments are then supplemented according to adjacent cycles under the same operating condition and changes in the response of adjacent components. Among them:

[0094] The distribution of adjacent sampling under the same operating condition is determined by the task stage, the similar speed and load range, and the collection time. The adjacent cycles under the same operating condition are determined by the adjacent operating cycles within the similar speed and load range under the same task stage.

[0095] The process of removing outliers and completing missing sampling segments is as follows:

[0096] The sampled values ​​are screened according to the sensor's calibrated range. Sampled values ​​that exceed the upper limit of the range, are below the lower limit of the range, or do not conform to the physical operation relationship are marked as range abnormal values ​​and removed.

[0097] Select adjacent sampling windows according to the task stage, similar speed load range and collection time, calculate the median value of the same type of parameter, and use the difference between the upper quartile value and the lower quartile value as the fluctuation range. When the sampled value deviates from the median value of the adjacent window by more than 3 times the fluctuation range, and there is no continuous change in the same direction between the sampling points before and after, the sampled value is marked as an isolated jump anomaly and removed.

[0098] The speed, load and related component parameters are continuously verified. When the speed and load are stable, a single component parameter changes instantaneously and there is no corresponding response from adjacent components, the change value is marked as a continuous outlier and removed.

[0099] When multiple related component parameters deviate continuously in the same direction, the deviation data is retained as the actual operational deviation data;

[0100] For sampling segments with a missing time of no more than 10 seconds, reference segments are selected according to the corresponding positions in the same task stage, similar speed load range and adjacent operating cycles. The missing sampling values ​​are supplemented by combining the slope of the change in the sampling values ​​before and after the missing segment, and the supplementary values ​​with component linkage relationship are corrected according to the response changes of adjacent components.

[0101] After completing the removal of outlier samples and the completion of missing sample segments, record the outlier removal flag, the completion flag, and the completion source flag;

[0102] The diesel engine operating parameter sequence is time-aligned with task phases, start-stop boundaries, load switching points, maintenance and recovery phases, historical fault nodes, and business maintenance nodes to generate a standardized diesel engine multi-source state dataset.

[0103] In this embodiment, generating the diesel engine operating condition chain sample set includes:

[0104] The standardized diesel engine multi-source state dataset is arranged according to the collection time, and the continuous operation segments of the diesel engine are divided according to the start-stop boundary, load switching point and maintenance recovery section.

[0105] Based on the changes in engine speed, load, combustion pressure, coolant temperature, and engine vibration during continuous operation segments of the diesel engine, the following operating conditions are marked: start-up, idling, acceleration, high load, load reduction, cooling recovery, shutdown, and maintenance recovery.

[0106] The marking of the working condition category is as follows:

[0107] The following parameters were calculated for continuous operation segments of the diesel engine: rate of change of speed, rate of change of load, range of change of combustion pressure, range of change of cooling water temperature, and root mean square range of change of vibration.

[0108] The criteria for determining continuous rise, continuous fall, and stable state are that the corresponding parameter changes in the same direction or the change amplitude is smaller than the fluctuation range of the same type of parameter in adjacent windows within three or more consecutive sampling windows.

[0109] When the engine speed increases from 0 to a non-zero operating state, and the combustion pressure changes from an unpressurized state to a periodic fluctuation, it is marked as a start-up condition.

[0110] When the engine speed remains non-zero, the speed change rate is less than 2%, the load is less than 20%, and the combustion pressure and engine vibration fluctuations are stable, it is marked as idling condition.

[0111] When the speed or load increases continuously, and the combustion pressure and engine vibration increase in the same or adjacent sampling windows, it is marked as an acceleration condition.

[0112] When the load is consistently above 70%, the speed change rate is less than 3%, and the changes in combustion pressure and cooling water temperature are less than the fluctuation range of the adjacent window, it is marked as a high load condition.

[0113] When the load decreases continuously and the combustion pressure drops before the cooling water temperature and engine vibration, it is marked as a load drop condition.

[0114] When the load is below 20% after the task is completed, the cooling water temperature and machine vibration continue to decrease, and the speed does not drop to 0, it is marked as a cooling recovery condition.

[0115] When the speed drops from a non-zero operating state to 0 and the periodic fluctuation of combustion pressure disappears, it is marked as a shutdown condition;

[0116] When the operating segment is located after the business maintenance node or within the maintenance recovery segment, and the cooling water temperature, engine vibration or combustion pressure continuously drop back to the stable operating range adjacent to the same task stage before maintenance, it is marked as maintenance recovery condition.

[0117] When the same operating segment meets multiple operating conditions, the operating condition category is determined in the order of start-stop status, maintenance and recovery status, load change direction, and stable load status.

[0118] Based on the task condition data, the corresponding runtime segments for each condition are labeled with task phases, generating task phase segments that include phase number, start and end times, duration, and condition category, where:

[0119] Mark the task phase, specifically:

[0120] The continuous operation segments of the diesel engine with marked operating condition categories are time-matched with the task number, task intensity and task duration interval in the task operating condition data to determine the task process interval to which each operation segment belongs.

[0121] When a single running segment spans two task process intervals, the running segment is divided according to the boundary of the task process interval;

[0122] Phase numbers are generated according to task number, task process interval sequence and working condition category, and the start and end times of the corresponding running segment are recorded. The time difference between the start and end times is used as the duration.

[0123] Write the phase number, start and end times, duration and operating condition type into the corresponding running segment to generate the task phase segment;

[0124] Compare the operating condition categories, load change directions, and maintenance recovery status of adjacent task phase segments, and arrange the transition directions and sequence between adjacent operating conditions, where:

[0125] The jump direction and jump order are arranged as follows:

[0126] The task phase segments under the same diesel engine number are sorted according to the acquisition time, and two adjacent task phase segments are selected as the previous segment and the next segment in sequence.

[0127] Compare the operating condition categories of the preceding and following segments, and mark the change relationship between the operating condition category of the preceding segment and the operating condition category of the following segment as the transition direction;

[0128] Compare the average load values ​​of the preceding and following segments. When the average load value of the following segment is higher than that of the preceding segment, it is marked as a load increase transition. When the average load value of the following segment is lower than that of the preceding segment, it is marked as a load decrease transition. When the difference between the two does not exceed the load fluctuation range of the adjacent window, it is marked as a load smooth transition.

[0129] Combining the maintenance and recovery status of the preceding and following segments, the change relationship from maintenance and recovery condition to non-maintenance and recovery condition is marked as a post-maintenance recovery transition, and the change relationship from non-maintenance and recovery condition to maintenance and recovery condition is marked as a maintenance entry transition.

[0130] The transition sequence is generated according to the order of adjacent task phase segments, and the transition direction, load change direction, maintenance and recovery status and transition sequence are written between adjacent task phase segments.

[0131] The task phase segments, transition directions, transition sequences, operating parameter sequences, historical fault nodes, and operational maintenance nodes are combined according to the diesel engine number to generate a diesel engine operating condition chain sample set.

[0132] In this embodiment, generating the multidimensional operational residual set includes:

[0133] The diesel engine operating condition chain sample set was grouped according to diesel engine number, task phase, operating condition category and transition order, and the corresponding diesel engine operating parameters of each group were extracted.

[0134] Steady-state segments are screened for various operating parameters under the same operating condition category and the same task phase. Non-steady-state sampled values ​​within start-stop boundaries, load switching points, and maintenance recovery sections are removed to form baseline samples for different operating conditions, including:

[0135] Steady-state fragment screening is performed, specifically as follows:

[0136] Under the same operating condition category and the same task phase, sampling windows are divided according to the acquisition time for various operating parameters. The sampling window length is 30 seconds and the sliding step size is 10 seconds. The speed change rate, load change rate, combustion pressure change amplitude, cooling water temperature change amplitude and vibration root mean square change amplitude are calculated in each sampling window.

[0137] The sampled values ​​located 30 seconds before and after the start-up boundary, 30 seconds before and after the shutdown boundary, 20 seconds before and after the load switching point, and within the maintenance and recovery section are marked as non-steady-state sampled values ​​and removed.

[0138] The remaining sampling window is judged for stability. When the speed change rate is less than 2%, the load change rate is less than 3%, and the root mean square change amplitude of combustion pressure, cooling water temperature and vibration is less than the fluctuation range of adjacent windows of the same type of parameter, the corresponding sampling window is marked as a steady state segment.

[0139] The various operating parameters within the steady-state segment are combined according to the diesel engine number, task phase, operating condition category, acquisition time, and component number to form a baseline sample for each operating condition.

[0140] Based on the data collection time, the median value and fluctuation range of parameters are calculated for the baseline samples under different working conditions to generate parameter baselines for each working condition, where:

[0141] Generate baseline parameters for each operating condition, specifically as follows:

[0142] The baseline samples for different operating conditions are grouped according to diesel engine number, task phase, operating condition category, component number, and parameter category, and the acquisition time within the task phase is converted into relative acquisition time.

[0143] Within the same group, the same type of parameter values ​​in the steady-state segment are collected according to the relative acquisition time, the median value of the parameter is calculated, and the median value of the parameter is used as the parameter baseline value at the corresponding relative acquisition time.

[0144] Calculate the upper quartile and lower quartile values ​​of the same type of parameter at the corresponding relative acquisition time, and use the difference between the upper quartile and lower quartile values ​​as the parameter fluctuation range;

[0145] When the number of steady-state samples at the corresponding relative acquisition time is less than 5, steady-state samples with the same working condition category and the same task phase at adjacent acquisition times will be combined for calculation.

[0146] The sub-condition parameter baselines are generated by combining the task phase, working condition category, component number, parameter category, relative acquisition time, parameter baseline value and parameter fluctuation range. The acquisition time in the sub-condition parameter baseline is the relative acquisition time within the task phase. The relative acquisition time is obtained by subtracting the start time of the corresponding task phase from the acquisition time.

[0147] The actual operating parameters are matched with the baseline parameters of the sub-operating conditions under the corresponding operating condition category, task phase and acquisition time. The operating parameter deviation is obtained by subtracting the parameter baseline value from the actual operating parameter value.

[0148] A multidimensional operational residual set is generated by combining the operating parameter deviations based on the diesel engine number, component number, task phase, and data acquisition time.

[0149] In this embodiment, the generation of the diesel engine degradation residual feature package includes:

[0150] The multidimensional operational residual set is organized into a residual decomposition sequence according to the diesel engine number, component number, task phase, and residual category. The MSTL decomposition cycle order is determined based on the speed cycle, load cycle, task phase cycle, and maintenance recovery cycle, where:

[0151] The residual decomposition sequence and MSTL decomposition periodicity are generated as follows:

[0152] The multidimensional operational residual set is classified according to diesel engine number, component number, task phase and residual category. The residuals are combustion pressure residual, lubricating oil temperature and pressure residual, cooling water temperature and flow residual, intake and exhaust pressure and temperature residual, engine vibration residual and electronic control response residual.

[0153] Under the same diesel engine number, component number, task phase and residual category, the operating parameter deviations are arranged according to the acquisition time, and the operating condition category, transition order, anomaly removal mark and completion mark are retained to form a residual decomposition sequence;

[0154] Speed ​​cycles are determined based on the repeated fluctuation intervals of speed; load cycles are determined based on the repeated intervals of load rise, stabilization, and fall; task phase cycles are determined based on the repeated occurrence intervals of task phase numbers; and maintenance recovery cycles are determined based on the parameter fall-off intervals after business maintenance nodes.

[0155] Arranged from shortest to longest cycle time, the speed cycle, load cycle, task phase cycle, and maintenance recovery cycle are determined as the MSTL decomposition cycle sequence;

[0156] When the cycle time is the same or similar, the cycles should be arranged in the order of speed cycle, load cycle, task phase cycle, and maintenance and recovery cycle.

[0157] Before fitting the seasonal term in each decomposition cycle, a periodic segment replacement process is performed. Segments that continuously deviate from the fluctuation range of the same operating condition and have the same residual direction are marked as abnormal residual segments. Normal residual segments at the same position in adjacent phase cycles of the same task are selected to replace the abnormal residual segments, forming the seasonal term fitting sequence, where:

[0158] The execution cycle fragment placeholder replacement process is as follows:

[0159] Before fitting the seasonal term in each MSTL decomposition cycle, the residual decomposition sequence is divided into several cycle segments according to the decomposition cycle length, and the residual values ​​in each cycle segment are aligned according to the relative acquisition time.

[0160] Segments in which the residual values ​​in three or more consecutive sampling windows all exceed the fluctuation range of the same operating condition parameters, and the residual values ​​are all positively biased or all negatively biased, are marked as abnormal residual segments.

[0161] In the adjacent phase cycles of the same task in the abnormal residual segment, find the candidate residual segments with the same relative acquisition time position, and remove the candidate residual segments that exceed the fluctuation range of the same working condition parameters, have abnormal removal marks, or have continuous deviation in residual direction.

[0162] From the remaining candidate residual segments, select the normal residual segment with the smallest average residual amplitude and no continuous deviation in the same direction. Replace the abnormal residual segment with the same relative acquisition time to form a placeholder segment.

[0163] The unreplaced residual segments and placeholder segments are rearranged according to the acquisition time to generate a seasonal term fitting sequence, while retaining the abnormal residual segment markers and replacement source markers;

[0164] MSTL seasonal term fitting is performed using the seasonal term fitting sequence to generate the periodic components of each decomposition period, and the residual decomposition sequence is successively subtracted from the corresponding periodic components to generate the remaining residual sequence.

[0165] The remaining residual sequence is fitted with the MSTL trend term to generate the trend component, and the initial degraded residual sequence is generated by subtracting the trend component from the remaining residual sequence.

[0166] Component back-injection and re-decomposition are performed between the periodic component, trend component, and initial degraded residual sequence. Fragments in the periodic and trend components that overlap with anomalous residual segments but are not retained in the initial degraded residual sequence are marked as misclassified degraded segments.

[0167] Perform component reinjection and re-decomposition processing, specifically as follows:

[0168] The start and end acquisition times of abnormal residual segments are mapped to the periodic component, trend component, and initial degraded residual sequence, respectively, and periodic segments, trend segments, and initial degraded residual segments within the same acquisition time range are extracted.

[0169] Calculate the average residual amplitude and residual direction for the periodic segment, trend segment, and initial degraded residual segment respectively. The residual direction is determined according to the number of positively biased sampling points and the number of negatively biased sampling points within the segment.

[0170] When the residual direction of a periodic segment or trend segment is consistent with that of an abnormal residual segment, and the average residual amplitude exceeds the fluctuation range of the same operating condition parameters, the corresponding periodic segment or trend segment is taken as a candidate misclassified segment.

[0171] When the average residual amplitude of the initial degraded residual corresponding to the candidate misclassified segment is less than 50% of the average residual amplitude of the abnormal residual segment, or the residual direction of the abnormal residual segment is not continuously maintained, the candidate misclassified segment is marked as a misclassified degraded segment.

[0172] After subtracting misclassified degenerate segments from the periodic and trend components, the misclassified degenerate segments are reinjected into the initial degenerate residual sequence, and MSTL decomposition is performed again to generate degenerate residual components, where:

[0173] The degraded residual components are generated as follows:

[0174] Based on the start and end acquisition times of the misclassified degenerate segments, locate the corresponding segments in the periodic and trend components, and extract the directional misclassification residual values ​​of the corresponding sampling points;

[0175] For misclassified degenerate segments originating from periodic components, the directional misclassification residual value is subtracted from the sampling point corresponding to the periodic component;

[0176] For misclassified and degraded segments originating from trend components, the directional misclassification residual value is subtracted from the sampling point corresponding to the trend component;

[0177] The directional misclassified residual values ​​obtained after deduction are superimposed onto the initial degraded residual sequence at the same acquisition time to generate the reinjection residual sequence;

[0178] Following the MSTL decomposition cycle order, seasonal and trend term fitting are performed again on the back-injected residual sequence to generate the re-decomposed cycle component, re-decomposed trend component, and re-decomposed residual sequence.

[0179] The remaining residual sequence after further decomposition is identified as the degenerate residual component;

[0180] By combining the diesel engine number, component number, task phase, residual category, and acquisition time, degraded residual components, periodic components, trend components, abnormal residual segment markers, and misclassified degraded segment markers, a diesel engine degradation residual feature package is generated.

[0181] In this embodiment, generating the diesel engine degradation prediction package includes:

[0182] An improved MLP-Mixer network is constructed, comprising a differential token embedding layer, a recovery degradation splitting layer, and a component misalignment mixing layer, wherein:

[0183] The improved MLP-Mixer network is constructed as follows:

[0184] In the traditional MLP-Mixer network, the input token is linearly embedded and then enters the Mixer block, which consists of a temporal token mixing layer and a residual channel mixing layer, and the prediction result is generated through the output mapping layer.

[0185] A differential token embedding layer is added after the traditional linear embedding structure to enhance the expression of working condition differences in the running window token;

[0186] A recovery degradation offloading layer is added between the differential token embedding layer and the Mixer block to distinguish between maintenance recovery fluctuations and continuous degradation deviations.

[0187] The traditional residual channel mixing layer improves the direct channel mixing method at the same acquisition time into a component misalignment mixing layer, which expresses the sequential transmission relationship between the residual responses of different components.

[0188] The differential token embedding layer, the recovery degradation offloading layer, and the component misalignment mixing layer are connected sequentially to the traditional MLP-Mixer network to obtain an improved MLP-Mixer network.

[0189] The diesel engine degradation residual feature package is divided into operating windows. The degradation residual components, periodic components, trend components, and segment markers within each operating window are arranged to generate operating window tokens, where:

[0190] Generate a run window token, specifically as follows:

[0191] According to the diesel engine number, component number, task phase and residual category, the diesel engine degradation residual feature package is divided into operating windows according to the acquisition time. The operating window length is 180 seconds and the sliding step size is 60 seconds.

[0192] Within each running window, degraded residual components, periodic components, trend components, abnormal residual segment markers, and misclassified degraded segment markers are arranged according to the acquisition time.

[0193] The degraded residual components, periodic components, trend components, and fragment labels at the same acquisition time are concatenated to generate a feature vector of the sampling point;

[0194] Arrange the feature vectors of each sampling point within the running window in the order of acquisition time, and add the diesel engine number, component number, task phase, residual category and window start and end time to generate a running window token;

[0195] The difference token embedding layer calculates the degradation residual difference between the current running window and the baseline window under the same operating condition, as well as the degradation residual difference between the current running window and the historical window of the same task phase. It embeds these two types of degradation residual differences into the running window token to generate a difference token, where:

[0196] The differential token embedding layer includes:

[0197] Window index queue: caches the tokens of the currently running window;

[0198] Same-condition baseline window lookup table: stores the same-condition baseline window index and baseline residual values;

[0199] Same task phase history window queue: caches tokens for running windows of the same task phase history;

[0200] Parallel Calculator for Differences: Calculates the difference between two types of degraded residuals;

[0201] Difference Direction Register: Stores the difference direction and difference magnitude;

[0202] Difference Embedding Mapper: Generates difference embedding vectors;

[0203] Token concatenation buffer: concatenates the difference embedding vector to the running window token to generate the difference token;

[0204] The processing procedure of the differential token embedding layer is as follows:

[0205] The window index queue receives running window tokens. Each running window contains L sampling points, each sampling point has a feature dimension of C, and the running window token size is L×C. When the sampling interval is 1 second and the running window length is 180 seconds, L takes the value of 180.

[0206] The window index queue extracts the corresponding operating condition category based on the task phase and window start and end times in the running window token;

[0207] The same-condition baseline window lookup table matches the same-condition baseline window tokens according to the diesel engine number, component number, operating condition category, task phase, and residual category, and extracts the baseline degradation residual components in the same-condition baseline window tokens to form a baseline residual vector of size L×1.

[0208] The historical window queue of the same task phase selects K historical running window tokens before the current running window according to diesel engine number, component number, task phase, residual category and acquisition time, and extracts the historical degradation residual components from the K historical running window tokens to form a historical residual tensor of size K×L×1, where K is 3;

[0209] The differential parallel calculator extracts the current degraded residual component from the current running window token to form a current residual vector of size L×1. The current residual vector is subtracted from the baseline residual vector to generate a baseline differential vector of size L×1 for the same working condition. The current residual vector is subtracted from the mean of K historical residual vectors in the window dimension to generate a phase differential vector of size L×1 for the same task.

[0210] The differential direction register receives the baseline differential vector under the same operating condition and the phase differential vector under the same task. It marks the sampling points with a differential value greater than 0 as positive bias and assigns a value of 1, marks the sampling points with a differential value less than 0 as negative bias and assigns a value of -1, and marks the sampling points with a differential value equal to 0 as unbiased and assigns a value of 0. It also calculates the absolute amplitude of the two types of differentials respectively.

[0211] The differential embedding mapper concatenates the baseline difference, phase difference, direction of baseline difference, direction of phase difference, magnitude of baseline difference, and magnitude of phase difference corresponding to each sampling point into a 6-dimensional differential feature vector. It then maps this feature vector into a D-dimensional differential embedding vector through an embedding weight matrix of size 6×D and an embedding bias vector of length D, where D is 16, generating a differential embedding matrix of size L×D.

[0212] The token concatenation buffer concatenates a difference embedding matrix of size L×D to a run window token of size L×C, generating a difference token of size L×(C+D);

[0213] The degradation recovery layer divides the difference tokens into recovery-state tokens and degradation-state tokens based on the difference magnitude change, residual direction, and convergence status of the maintenance recovery segment during continuous operation.

[0214] Restoring a degraded shunt layer includes:

[0215] Continuous window cache queue: caches consecutive difference tokens;

[0216] Difference Amplitude Change Register: Stores the difference amplitude change between adjacent running windows;

[0217] Residual direction comparator: compares the residual directions of consecutive running windows;

[0218] Maintenance and recovery status lookup table: stores maintenance and recovery segment markers and recovery start and end times;

[0219] Convergence state determiner: determines whether the difference magnitude is continuously decreasing;

[0220] Degradation deviation detector: Determines whether the difference amplitude is continuously increasing;

[0221] Shunt gate: Generates recoverable shunt markers and degenerate shunt markers;

[0222] Token distribution buffer: outputs recoverable and degenerate tokens;

[0223] The process for restoring a degraded shunt layer is as follows:

[0224] The continuous window buffer queue receives difference tokens. Let the size of a single difference token be L×(C+D). The difference tokens of M running windows are continuously buffered to form a continuous difference token tensor of size M×L×(C+D), where M is 5.

[0225] The differential amplitude change register extracts the baseline differential amplitude and phase differential amplitude of the same working condition from the differential feature field corresponding to the continuous differential tokens, calculates the differential amplitude change according to adjacent operating windows, and generates a differential amplitude change tensor of size (M-1)×L×2.

[0226] The residual direction comparator extracts the baseline difference direction and the phase difference direction of the same working condition from the difference feature field corresponding to the continuous difference tokens, performs consistency comparison on the difference directions of adjacent running windows, and generates a direction consistency label tensor with a size of (M-1)×L×2.

[0227] The maintenance and recovery status lookup table matches the maintenance and recovery segment marker and recovery start and end time according to the diesel engine number, component number, task phase and window start and end time, and generates a maintenance and recovery status vector of size M×1.

[0228] The convergence state determiner receives the difference magnitude change tensor, the direction consistency label tensor, and the maintenance recovery state vector. When the continuous operation window is located within the maintenance recovery segment, and the difference magnitudes of the two types decrease continuously within no less than three adjacent windows, and the residual direction continues to fall back to zero, a recovery state shunting label is generated.

[0229] The degradation deviation determiner receives the difference magnitude change tensor and the direction consistency label tensor. When the difference magnitudes of the two types increase continuously within no less than 3 adjacent windows and the residual direction remains positive or negative, a degradation state shunting label is generated.

[0230] The shunt gate assigns a value of 1 to the recovery state shunt marker, a value of 1 to the degradation state shunt marker, and a value of 0 to the remaining sampling points that do not meet the shunt conditions, and generates a shunt gate matrix of size L×2;

[0231] The token distribution buffer divides the current difference tokens according to the split gating matrix, writes the sampling point token marked as 1 in the recovery state split to the recovery state token, writes the sampling point token marked as 1 in the degenerate state split to the degenerate state token, and outputs the recovery state token and the degenerate state token.

[0232] The component misalignment mixing layer performs front-to-back misalignment alignment and channel mixing on the combustion residuals, intake and exhaust residuals, vibration residuals, cooling residuals, lubrication residuals, and electronic control response residuals in the recovery state token and the degradation state token, generating a cross-component mixed token, wherein:

[0233] The component misalignment hybrid layer includes:

[0234] Token input queue: cached recovery-state tokens and degradation-state tokens;

[0235] Component Residual Path Index Table: Marks residual paths for combustion, intake and exhaust, vibration, cooling, lubrication, and electronic control response;

[0236] Misalignment step size register: stores the forward and backward misalignment step sizes between the residual channels of each component;

[0237] Component response sequence lookup table: stores the order of residual responses across components;

[0238] Misalignment buffer: rearranges the residual channels of each component according to the misalignment step size;

[0239] Channel blending weight register: stores cross-component channel blending weights;

[0240] Cross-component blending calculator: Performs weighted blending on the residual channels of misaligned components;

[0241] Mixed token output buffer: Outputs mixed tokens across components;

[0242] The process of handling the component misalignment mixing layer is as follows:

[0243] The diversion token input queue receives recovery tokens and degradation tokens, and establishes recovery token groups and degradation token groups according to diesel engine number, task phase, window start and end time and diversion mark respectively;

[0244] The component residual channel index table maps combustion residual, intake and exhaust residual, vibration residual, cooling residual, lubrication residual and electronic control response residual to 6 component residual channels. Let the number of component residual channels be P, and P takes the value of 6. The shunt tokens in the same running window are organized into a component residual token tensor of size L×P×(C+D).

[0245] The misalignment step register stores the misalignment step between the residual channels of each component according to the component response order lookup table. The misalignment step value ranges from -2 to 2 sampling points. A positive value indicates rear alignment, a negative value indicates front alignment, and 0 indicates alignment at the same acquisition time.

[0246] The misalignment buffer shifts each component residual channel in the component residual token tensor back and forth according to the misalignment step size. Sampling points that exceed the boundary of the running window are filled with zero values, generating a misalignment token tensor of size L×P×(C+D).

[0247] The channel mixing weight register is configured with a channel mixing weight matrix of size P×P, where each weight represents the response propagation strength of one component residual channel to another component residual channel;

[0248] The cross-component mixing calculator performs a weighted summation of the six component residual channels at the same acquisition time in the misalignment token tensor according to the channel mixing weight matrix. The mixing result of the target component residual channel is obtained by multiplying the misalignment tokens of each source component residual channel by the corresponding channel mixing weight and then summing them to generate a cross-component mixing tensor of size L×P×(C+D).

[0249] The hybrid token output buffer rearranges the cross-component hybrid tensors corresponding to the recoverable token group and the degenerate token group according to the acquisition time, and adds a shunt flag and a component residual channel flag to generate a cross-component hybrid token;

[0250] Time-dimensional token mixing and residual channel mixing are performed on the cross-component mixed tokens to generate degradation trend values, degradation state markers, and prediction time steps for each component. These are then combined to form a diesel engine degradation prediction package, where:

[0251] A diesel engine degradation prediction package is generated, specifically as follows:

[0252] Receive cross-component mixed tokens and categorize them according to diesel engine number, component number, task phase, window start and end time, and component residual channel marker;

[0253] Within the residual channel of the same component, cross-component hybrid tokens are arranged according to the acquisition time, and the token features at consecutive acquisition times are mixed in the time dimension to obtain the component time hybrid features.

[0254] Within the same operating window, residual channel mixing is performed on the component time-mixing characteristics corresponding to combustion, intake and exhaust, vibration, cooling, lubrication and electronic control response to obtain component degradation characterization;

[0255] The mean and maximum values ​​of degradation residuals are calculated from the degradation characterization of components. The number of recovery-state shunt markers and degradation-state shunt markers are also counted. The mean and maximum values ​​of degradation residuals are normalized to the range of 0 to 1. The proportion of degradation-state shunt markers to the number of sampling points within the operating window is taken as the degradation-state shunt marker percentage. The normalized mean degradation residual, the normalized maximum value of degradation residual, and the degradation-state shunt marker percentage are weighted and summed to obtain the degradation trend value corresponding to each component. The weight of the mean degradation residual is 0.4, the weight of the maximum value of degradation residual is 0.3, and the weight of the degradation-state shunt marker percentage is 0.3.

[0256] When the degradation trend value decreases continuously and the number of recovery state shunt markers is higher than the number of degradation state shunt markers, the corresponding component is marked as recovery state;

[0257] When the degradation trend value increases continuously and the number of degradation state shunt markers is higher than the number of recovery state shunt markers, the corresponding component will be marked as degradation state;

[0258] The remaining cases are marked as baseline states, generating degradation state labels;

[0259] The prediction span is set to 6 future operating windows. The slope of change is calculated based on the average difference of the degradation trend values ​​in the most recent 3 consecutive operating windows, and the degradation trend values ​​of the future operating windows are extrapolated based on the slope of change.

[0260] When the degradation trend value of the future running window continues to rise relative to the current running window, the sequence number of the future running window that first satisfies the continuous upward relationship is taken as the prediction time step;

[0261] When the degradation trend values ​​do not form a continuous upward relationship within 6 future operating windows, the prediction time step is marked as 7;

[0262] The diesel engine degradation prediction package is formed by combining the diesel engine number, component number, task phase, window start and end time, degradation trend value, degradation status mark and prediction time step.

[0263] The improved MLP-Mixer network is trained by labeling samples before historical fault nodes, samples after maintenance and recovery phases, and baseline samples in the same task phase as degraded samples, recovered samples, and baseline samples, respectively. Degradation trend error, degradation state labeling error, prediction time step error, recovery degradation shunting error, and component misalignment error are calculated. These errors are combined as a comprehensive error to update the network parameters of each layer. The improved MLP-Mixer network training is complete when the comprehensive error of the verification samples in the same task phase decreases by less than 0.002 for four consecutive rounds.

[0264] The improved MLP-Mixer network is trained as follows:

[0265] The samples within the first 6 operating windows of the historical fault node are marked as degraded samples, the samples within the last 6 operating windows of the maintenance and recovery phase are marked as recovery samples, and the samples that do not have abnormal residual segments and misclassified degraded segments under the same task phase are marked as baseline samples. The operating window tokens corresponding to each type of sample are input into the improved MLP-Mixer network, and the degradation trend value, degradation state label, prediction time step, recovery state split label, degradation state split label and component misalignment alignment result are output for each component.

[0266] The degradation trend error is calculated based on the difference between the degradation trend label and the degradation trend value corresponding to the degradation sample, recovery sample, and baseline sample. The degradation state label error is calculated based on the difference between the recovery state, degradation state, and baseline state labels and the degradation state label. The prediction time step error is calculated based on the difference between the window sequence number corresponding to the historical fault node and the prediction time step. The recovery degradation shunting error is calculated based on the difference between the recovery state shunting label, the degradation state shunting label, and the corresponding shunting label. The component misalignment alignment error is calculated based on the difference between the component residual response sequence, the misalignment step size label, and the component misalignment alignment result.

[0267] The degradation trend error, degradation state labeling error, prediction time step error, restoration degradation splitting error, and component misalignment alignment error are multiplied by 0.30, 0.20, 0.20, 0.15, and 0.15 respectively, and then summed to obtain the comprehensive error. The gradient value corresponding to the comprehensive error is calculated according to 64 running window samples per batch. The network parameters of the differential token embedding layer, restoration degradation splitting layer, component misalignment mixing layer, time dimension token mixing layer, residual channel mixing layer, and output mapping layer are updated with a learning rate of 0.001. When the comprehensive error of the phase verification sample of the same task decreases by less than 0.002 for 4 consecutive rounds, the training of the improved MLP-Mixer network is completed.

[0268] In this embodiment, the generation of the diesel engine failure risk prediction package includes:

[0269] According to the diesel engine number, component number, task phase, and prediction time step, sort the degradation trend value and degradation status marker in the diesel engine degradation prediction package to generate the degradation trajectory of each component;

[0270] The degradation trajectory of each component is segmented based on the continuous rise, sudden increase, plateau, and decline of the degradation trend value, and mapped to normal, watchful, alarm, and near-fault risk states, where:

[0271] The degradation trajectory of each component is segmented, specifically as follows:

[0272] According to the predicted time step, arrange the degradation trend value and degradation state mark in the degradation trajectory of each component, and calculate the degradation trend difference between adjacent predicted time steps;

[0273] When the degradation trend difference of more than three consecutive predicted time steps in the degradation trajectory of the same component is greater than 0, the corresponding trajectory interval is marked as a continuous rising segment.

[0274] When the degradation trend difference between adjacent prediction time steps is greater than twice the average absolute value of the degradation trend difference of the most recent 3 prediction time steps, and the degradation trend difference is greater than 0, the corresponding trajectory interval is marked as a sudden increase segment.

[0275] When the absolute value of the degradation trend difference over three or more consecutive prediction time steps is less than 5% of the average degradation trend value over the most recent three prediction time steps, and the degradation trend value has not decreased, the corresponding trajectory interval is marked as a platform holding segment.

[0276] When the degradation trend difference is less than 0 for more than 3 consecutive prediction time steps, and the degradation state is marked as recovery state or baseline state, the corresponding trajectory interval is marked as fallback segment.

[0277] The trajectory intervals of the pullback segment and the segment without continuous rising segment, sudden increase segment and platform holding segment are mapped as normal risk state; the continuous rising segment is mapped as risk state of concern; the platform holding segment under sudden increase segment or degenerate state is mapped as alarm risk state; the trajectory interval of continuous rising segment followed by sudden increase segment and the predicted time step does not exceed 2 future operating windows is mapped as fault imminent risk state.

[0278] The failure probability is calculated based on risk status, degradation trend value, predicted time step, and historical failure node matching results. The component with the highest failure probability and consistent with the cross-component mixed token response order is marked as the faulty component.

[0279] The failure probability is calculated as follows:

[0280] Based on the diesel engine number, component number, task phase, and prediction time step, extract the risk status, degradation trend value, and prediction time step corresponding to the degradation trajectory of each component;

[0281] The risk states of normal, attention, alarm, and impending fault are converted into risk state coefficients of 0.10, 0.40, 0.70, and 0.90, respectively.

[0282] Normalize the degradation trend values ​​of each component under the same task phase to the range of 0 to 1 to obtain the degradation trend coefficient;

[0283] The time proximity coefficient is calculated based on the prediction time step. When the prediction time step is 1 to 6, the time proximity coefficient is 7 minus the prediction time step and then divided by 6. When the prediction time step is 7, the time proximity coefficient is 0.

[0284] Match the current component and the current task phase with historical fault nodes. If there are historical fault nodes with the same component and the same task phase, mark the historical fault matching coefficient as 1. If there are no matching historical fault nodes, mark the historical fault matching coefficient as 0.

[0285] The risk status coefficient, degradation trend coefficient, time proximity coefficient, and historical failure matching coefficient are multiplied by 0.35, 0.30, 0.20, and 0.15 respectively, and then summed to obtain the failure probability corresponding to each component.

[0286] The components are arranged from highest to lowest failure probability, and the component with the highest failure probability and at the forefront of the cross-component response is selected as the faulty component based on the component response order in the cross-component mixed token.

[0287] The time interval between the time of the near-fault risk state and the current data collection time is determined as the remaining service life, where:

[0288] Determine the remaining useful life, specifically as follows:

[0289] Find the predicted time step that is first mapped to the near-fault risk state in the component degradation trajectory corresponding to the faulty component;

[0290] Add the current acquisition time to the running window step size corresponding to the prediction time step to obtain the time corresponding to the fault near risk state, where the running window step size is the time interval between adjacent prediction time steps;

[0291] The time interval between the time of the near-fault risk state and the current data collection time is determined as the remaining service life;

[0292] When no component degradation trajectory shows a near-failure risk state within the current prediction span, the remaining service life is marked as exceeding the current prediction span;

[0293] A diesel engine failure risk prediction package is generated by combining the diesel engine number, task phase, failure probability, fault component attribution, remaining service life, and risk status.

[0294] In this embodiment, generating dynamic risk warning results and predictive maintenance schedules includes:

[0295] The diesel engine failure risk prediction package is analyzed to extract failure probability, failure component attribution, remaining service life, and risk status. Warning levels are then assigned based on failure probability and risk status.

[0296] The warning level is marked as follows:

[0297] When the risk status is that a fault is imminent, or the probability of a fault is greater than or equal to 0.85, the corresponding faulty component will be marked as a Level 1 warning.

[0298] When the risk status is alarm, or the failure probability is greater than or equal to 0.70 and less than 0.85, the corresponding faulty component will be marked as a level 2 warning.

[0299] When the risk status is "attention" or the failure probability is greater than or equal to 0.40 and less than 0.70, the corresponding faulty component will be marked as a Level 3 warning.

[0300] When the risk status is normal and the failure probability is less than 0.40, the corresponding faulty component will be marked as being in a tracking status.

[0301] The warning level labeling results are generated according to the diesel engine number, mission phase, faulty component attribution, fault probability, remaining service life, risk status, and warning level.

[0302] Adding the current data collection time to the remaining service life yields the maintenance deadline. Combining the faulty component attribution, warning level, and maintenance deadline generates a fault warning item, where:

[0303] The following fault warning items are generated:

[0304] When the remaining service life is equal to the effective time interval, add the remaining service life to the current data collection time to obtain the maintenance deadline.

[0305] When the remaining service life is marked as exceeding the current forecast span, no maintenance deadline is generated, and the corresponding faulty component is retained as a tracking and warning item.

[0306] The faulty component is associated with the warning level, repair deadline, failure probability, remaining service life and risk status to generate a fault warning item;

[0307] Match fault warning items to the maintenance task window no later than the maintenance deadline, and determine the maintenance sequence according to the fault probability from high to low and the remaining service life from short to long, wherein:

[0308] The repair sequence is determined as follows:

[0309] For fault warning items with generated maintenance deadlines, filter out support task windows whose end time is no later than the maintenance deadline according to the maintenance deadline in the fault warning item, and exclude maintenance periods that overlap with the task execution time period.

[0310] Match the faulty component attribution, warning level, and repair deadline with the spare parts arrival time, maintenance personnel availability time, special tool availability time, and component replacement allowable time in the support task window, and select the support task window that meets both time and resource requirements.

[0311] When a single fault warning item matches multiple support task windows, select the support task window that is closest to the maintenance deadline and no later than the maintenance deadline.

[0312] When multiple fault warning items match the same support task window, they are sorted from high to low fault probability. If the fault probabilities are the same, they are sorted from short to long remaining service life.

[0313] When multiple fault warning items exist for the same faulty component or adjacent related components, and the repair deadline falls within the same support task window, the multiple fault warning items will be merged into the same support task window.

[0314] The maintenance sequence is determined based on the task window, the failure probability ranking, and the remaining service life ranking.

[0315] Based on the attribution of the faulty component, the warning level, the repair deadline, the support task window, and the repair sequence, dynamic risk warning results and predictive maintenance schedules are generated.

[0316] In this embodiment, the guarantee task window specifically includes the task execution time period, the downtime for maintenance period, the spare parts availability time period, the availability of maintenance personnel, the availability of special tools, and the allowable replacement time period. Maintenance periods that overlap with the task execution time period are excluded, and the overlapping time periods of the downtime for maintenance period, the availability of spare parts, the availability of maintenance personnel, the availability of special tools, and the allowable replacement time period are used as the guarantee task window.

[0317] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous power support task cycle. A certain type of diesel engine completed bench loading, task simulation operation, and intermittent maintenance preparation. The diesel engine, designated D-06, had a sampling period of 1 second, continuously recording 96 operating hours, generating 345,600 sets of raw sampling data. The data included engine speed, load, peak combustion pressure, lubricating oil pressure, lubricating oil temperature, coolant temperature, coolant flow rate, intake pressure, exhaust temperature, engine vibration, electronic fuel injection response delay, task phase markers, and maintenance records. During this task cycle, the load switched between 28% and 82%, the exhaust temperature fluctuated between 286℃ and 461℃, and the root mean square value of vibration varied between 0.21g and 0.51g, none of which continuously exceeded the traditional alarm threshold. However, in the middle of the operation, a combination of a slight decrease in combustion pressure, a slow increase in exhaust temperature, and an increase in fuel injection response delay occurred.

[0318] After data processing begins, this method merges data according to diesel engine number, component number, sensor number, and acquisition time. The raw data revealed 438 sets of instantaneous out-of-bounds sensor values ​​and 226 short-term missing segments, with the longest missing time being 7 seconds. After unit unification, sampling interval reshaping, and timestamp correction, the 438 sets of abnormal samples were removed, and the 226 missing segments were supplemented based on changes in adjacent cycles under the same operating conditions, resulting in 345,388 sets of valid sampled data. Before correction, the maximum time offset between different sensors was 2.8 seconds; after correction, the maximum time offset was reduced to 0.2 seconds, ensuring that combustion pressure, exhaust temperature, vibration response, and injection response delay correspond within the same operating window.

[0319] Subsequently, this method divides continuous operation segments based on start-stop boundaries, load switching points, and maintenance recovery phases. The entire task cycle is divided into 126 task phase segments, including 12 start-up segments, 18 idling segments, 21 acceleration segments, 31 high-load segments, 20 load reduction segments, 15 cooling recovery segments, and 9 maintenance recovery segments, forming 382 operating condition transition relationships. Taking the segment from hour 41.6 to hour 42.3 as an example, this segment transitions from idling to acceleration, and then to high-load operation, with the load increasing from 31% to 74% and the speed increasing from 820 rpm to 1504 rpm. Traditional methods can only observe that temperature and vibration increase with the load, making it difficult to determine whether it belongs to normal load response or a precursor to degradation; this method classifies this segment into a high-load operating condition under the same task phase, allowing it to be compared with similar historical segments.

[0320] In the process of constructing baseline parameters for different operating conditions, this method selects 18,240 historical operating windows of the same type of diesel engine as training samples. Each operating window is 180 seconds long with a slip step of 60 seconds, including 11,160 baseline samples, 3,180 maintenance and recovery samples, and 3,900 degradation samples. For the high-load phase, the calculated baseline ranges are: engine speed 1491–1513 rpm, peak combustion pressure 8.16–8.43 MPa, exhaust temperature 417–436 °C, root mean square vibration 0.30g–0.38g, lubricating oil pressure 0.43–0.47 MPa, and injection response delay 12.4–15.1 ms. During the 42.1-hour operating window, the measured engine speed was 1502 rpm, the load was 74.6%, the peak combustion pressure was 8.02 MPa, the exhaust temperature was 451℃, the root mean square value of vibration was 0.44g, and the injection response delay was 21.6 ms. Although the exhaust temperature did not exceed the 465℃ threshold and the vibration did not exceed the 0.52g threshold, compared with the baseline under the same operating conditions, the residual combustion pressure was -0.25 MPa, the residual exhaust temperature was +24℃, the residual vibration was +0.08g, and the residual injection response delay was +6.5 ms, indicating a multidimensional consistent deviation.

[0321] During MSTL decomposition, this method organizes the window and its preceding and following windows into a residual decomposition sequence, and determines the decomposition cycle according to speed cycle, load cycle, task phase cycle, and maintenance recovery cycle. Between hours 40.8 and 42.4, a total of 17 abnormal residual segments were identified, including 13 consecutive positively biased exhaust temperature residuals, 11 consecutive negatively biased combustion pressure residuals, and 15 consecutive positively biased injection response delay residuals. If ordinary trend decomposition were performed directly, approximately 38% of these abnormal segments would be absorbed into the load cycle component. This method uses adjacent normal segments in the same task phase for placeholder replacement before fitting the seasonal term. After the first decomposition, the mean value of the cycle component in the exhaust temperature residual was +11.2℃, the mean value of the trend component was +4.6℃, and the mean value of the initial degradation residual was +8.1℃. Further inspection revealed that +3.4℃ in the trend component still overlapped with abnormal segments and exhibited continuous deviation characteristics; therefore, this misclassified degradation segment was reinjected into the initial degradation residual sequence. After further decomposition, the average residual value of exhaust temperature degradation increased to +11.5℃, and the average residual value of fuel injection response delay degradation increased to +5.9ms, effectively preserving early degradation characteristics.

[0322] In the improved MLP-Mixer network processing, this method divides the diesel engine degradation residual feature package into operating window tokens. Taking the 42.1-hour operating window as an example, the difference token embedding layer calculates the difference between the current window and the baseline window under the same operating condition, as well as the historical window under the same task phase. Among them, the injection response delay difference is +6.5ms, the exhaust temperature difference is +24℃, and the combustion pressure difference is -0.25MPa. The degradation recovery layer determines that this window does not belong to the normal convergence fluctuation after maintenance recovery because the injection response delay difference has been positively increasing for five consecutive windows, with an average increase of 0.72ms. The component misalignment mixing layer further found that the combustion pressure decrease precedes the exhaust temperature increase by about two windows, and the exhaust temperature increase precedes the vibration enhancement by about one window, which is consistent with the propagation relationship of injector atomization degradation leading to decreased combustion efficiency, increased exhaust temperature, and induced vibration enhancement. Finally, the faulty component corresponding to this window is marked as an injector-related component with a failure probability of 0.86 and a predicted remaining service life of 5.6 operating hours.

[0323] The diesel engine entered a shutdown maintenance window after 47.3 hours of operation. Disassembly and inspection records showed carbon deposits in the injector nozzles, a spray cone angle below the normal range, and decreased combustion uniformity. Calculations indicated that the actual stable operating time was approximately 5.2 hours, 0.4 hours less than the 5.6 hours predicted by this method. Using the traditional fixed threshold method, this event wouldn't trigger an alarm until the 44.8th hour due to a brief exhaust temperature reaching 466℃, an advance of approximately 2.5 hours. Using the ordinary time-series prediction method, an anomaly score of 0.63 was given at the 43.2th hour, but due to a temporary decrease in residuals during the load reduction phase, the anomaly score dropped to 0.49, causing the warning status to be interrupted. This method provides continuous warnings at the 42.1st hour, an advance of 5.2 hours, and schedules the maintenance of injector-related components to a maintenance window that does not occupy the task execution section.

[0324] In the same batch of verification samples, 3200 operating windows were selected for comparison, including 780 actual degradation windows, 1900 normal baseline windows, and 520 maintenance and recovery windows.

[0325] The traditional fixed threshold method has an early degradation identification accuracy of 71.8%, a false alarm rate of 12.6%, a false negative rate of 19.4%, a fault component attribution accuracy of 63.5%, and an average error of 3.8 operating hours for remaining service life.

[0326] The accuracy rate of early degradation identification using the conventional time-series prediction method is 79.6%, the false alarm rate is 10.9%, the false negative rate is 13.7%, the accuracy rate of fault component attribution is 72.4%, and the average error of remaining service life is 2.6 operating hours.

[0327] The method achieves an early degradation identification accuracy of 91.3%, reduces the false alarm rate to 5.8%, the false negative rate to 6.1%, the fault component attribution accuracy to 87.9%, and reduces the average error of remaining service life to 1.1 operating hours.

[0328] As can be seen from Example 1, the present invention can distinguish between normal periodic fluctuations and early degradation residuals of components under complex task conditions, and determine the fault propagation process based on the sequential relationship of responses across components. This method not only improves the accuracy of early fault precursor identification in diesel engines, but also outputs fault probability, fault component attribution, remaining service life, and predictive maintenance schedule, enabling diesel engine fault prediction results to directly serve operational support and maintenance arrangements.

[0329] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A diesel engine fault prediction method based on big data analysis, characterized in that, include: Collect multi-source state data of diesel engines, perform preprocessing on the multi-source state data of diesel engines, and generate a standardized multi-source state dataset of diesel engines; Based on the standardized diesel engine multi-source state dataset, task condition segmentation, task phase labeling, and condition transition relationship arrangement are performed to generate a diesel engine condition chain sample set. Based on the diesel engine operating condition chain sample set, construct the sub-operating condition parameter baseline, extract the deviation between the actual operating parameters and the sub-operating condition parameter baseline, and generate a multi-dimensional operating residual set; The MSTL algorithm is used to perform multi-period trend decomposition on the multidimensional running residual set. By replacing abnormal residual segments and correcting misclassified degenerate segments, a diesel engine degradation residual feature package is generated. An improved MLP-Mixer network is constructed to convert diesel engine degradation residual feature packets into operating window tokens and difference tokens. By embedding the baseline difference under the same operating condition and the phase difference under the same task, separating the convergence recovery residual and the continuous deviation residual, and aligning the cross-component misalignment response segments, a diesel engine degradation prediction packet is generated. Based on the diesel engine degradation prediction package, degradation trajectory segmentation and risk state mapping are performed to calculate the failure probability, fault component attribution and remaining service life, and a diesel engine failure risk prediction package is generated. Based on the diesel engine failure risk prediction package, the warning level is marked, the maintenance deadline is deduced from the remaining service life, the support task window is matched, and dynamic risk warning results and predictive maintenance schedules are generated.

2. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The diesel engine multi-source status data specifically includes diesel engine speed and load data, combustion pressure data, lubricating oil temperature and pressure data, cooling water temperature and flow data, intake and exhaust pressure and temperature data, engine vibration data, electronic control response data, task operating condition data, historical fault data, business maintenance records, and task support data.

3. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The preprocessing of the multi-source state data of the diesel engine includes: According to the diesel engine number, component number, sensor number and acquisition time, the multi-source status data of the diesel engine are merged to form diesel engine status merged data; The diesel engine status data is merged and the unit is unified, the sampling interval is reorganized, and the timestamp is corrected to form a sequence of diesel engine operating parameters arranged according to the time of collection. Based on the task conditions, start-stop status, and business maintenance nodes, the diesel engine operating parameter sequence is marked with task stages, start-stop boundaries, load switching points, and maintenance recovery sections. Remove abnormal sampling values ​​that exceed the sensor's range, deviate from the adjacent sampling distribution under the same operating conditions, or do not meet the continuity of component operation. Complete the missing sampling segments according to the adjacent cycles under the same operating conditions and the response changes of adjacent components. The diesel engine operating parameter sequence is time-aligned with task phases, start-stop boundaries, load switching points, maintenance and recovery phases, historical fault nodes, and business maintenance nodes to generate a standardized diesel engine multi-source state dataset.

4. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The generated diesel engine operating condition chain sample set includes: The standardized diesel engine multi-source state dataset is arranged according to the collection time, and the continuous operation segments of the diesel engine are divided according to the start-stop boundary, load switching point and maintenance recovery section. Based on the changes in engine speed, load, combustion pressure, coolant temperature, and engine vibration during continuous operation segments of the diesel engine, mark the starting, idling, acceleration, high load, load reduction, cooling recovery, shutdown, and maintenance recovery conditions; Based on the task condition data, the corresponding running segments of each condition are marked with task phases, and task phase segments containing phase number, start and end time, duration and condition category are generated. Compare the operating condition categories, load change directions, and maintenance recovery status of adjacent task phase segments, and arrange the transition directions and transition sequences between adjacent operating conditions; The task phase segments, transition directions, transition sequences, operating parameter sequences, historical fault nodes, and operational maintenance nodes are combined according to the diesel engine number to generate a diesel engine operating condition chain sample set.

5. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The generation of the multidimensional operational residual set includes: The diesel engine operating condition chain sample set was grouped according to diesel engine number, task phase, operating condition category and transition order, and the corresponding diesel engine operating parameters of each group were extracted. Steady-state segments are screened for various operating parameters under the same working condition category and the same task phase. Non-steady-state sampled values ​​within the start-stop boundary, load switching point and maintenance recovery section are removed to form a baseline sample for each working condition. Based on the time of data collection, the median value and fluctuation range of parameters of the baseline samples under different working conditions are calculated to generate the parameter baselines for different working conditions. The actual operating parameters are matched with the baseline parameters of the sub-operating conditions under the corresponding operating condition category, task phase and acquisition time. The operating parameter deviation is obtained by subtracting the parameter baseline value from the actual operating parameter value. A multidimensional operational residual set is generated by combining the operating parameter deviations based on the diesel engine number, component number, task phase, and data acquisition time.

6. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The generated diesel engine degradation residual feature package includes: The multidimensional operational residual set is organized into a residual decomposition sequence according to the diesel engine number, component number, task phase and residual category. The MSTL decomposition cycle order is determined according to the speed cycle, load cycle, task phase cycle and maintenance recovery cycle. Before fitting the seasonal term in each decomposition cycle, a periodic segment replacement process is performed. Segments that continuously deviate from the fluctuation range of the same working condition and have the same residual direction are marked as abnormal residual segments. Normal residual segments at the same position are selected in adjacent phase cycles of the same task to replace abnormal residual segments, thus forming a seasonal term fitting sequence. MSTL seasonal term fitting is performed using the seasonal term fitting sequence to generate the periodic components of each decomposition period, and the residual decomposition sequence is successively subtracted from the corresponding periodic components to generate the remaining residual sequence. The remaining residual sequence is fitted with the MSTL trend term to generate the trend component, and the initial degraded residual sequence is generated by subtracting the trend component from the remaining residual sequence. Component back-injection and decomposition are performed between the periodic component, trend component, and initial degraded residual sequence. Fragments in the periodic component and trend component that overlap with abnormal residual fragments and are not retained in the initial degraded residual sequence are marked as misclassified degraded fragments. The misclassified degenerate segments are subtracted from the periodic and trend components, and the misclassified degenerate segments are reinjected into the initial degenerate residual sequence. Then, MSTL decomposition is performed again to generate degenerate residual components. By combining the diesel engine number, component number, task phase, residual category, and acquisition time, degraded residual components, periodic components, trend components, abnormal residual segment markers, and misclassified degraded segment markers, a diesel engine degradation residual feature package is generated.

7. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The generated diesel engine degradation prediction package includes: An improved MLP-Mixer network is constructed, which includes a differential token embedding layer, a recovery degradation splitting layer, and a component misalignment mixing layer. Diesel engine degradation residual feature package is divided into operating windows. Degradation residual components, periodic components, trend components and segment markers in each operating window are arranged to generate operating window tokens. The difference token embedding layer calculates the degradation residual difference between the current running window and the baseline window of the same operating condition, as well as the degradation residual difference between the current running window and the historical window of the same task phase. It embeds the two types of degradation residual differences into the running window token to generate the difference token. The degradation shunt layer divides the difference tokens into recovery tokens and degradation tokens based on the difference magnitude change, residual direction, and convergence status of the maintenance recovery section during the continuous operation window. The component misalignment mixing layer performs front-to-back misalignment alignment and channel mixing on the combustion residuals, intake and exhaust residuals, vibration residuals, cooling residuals, lubrication residuals and electronic control response residuals in the recovery state tokens and degradation state tokens to generate cross-component mixed tokens; Perform time-dimensional token mixing and residual channel mixing on cross-component mixed tokens to generate degradation trend values, degradation state markers and prediction time steps for each component, and combine them to form a diesel engine degradation prediction package; The improved MLP-Mixer network is trained by labeling samples before historical fault nodes, samples after maintenance and recovery stages, and baseline samples in the same task phase as degraded samples, recovered samples, and baseline samples, respectively. Degradation trend error, degradation state labeling error, prediction time step error, recovery degradation shunting error, and component misalignment error are calculated. The various errors are combined as a comprehensive error and the network parameters of each layer are updated. The training of the improved MLP-Mixer network is completed when the comprehensive error of the verification samples in the same task phase decreases by less than 0.002 for four consecutive rounds.

8. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The generated diesel engine failure risk prediction package includes: According to the diesel engine number, component number, task phase, and prediction time step, sort the degradation trend value and degradation status marker in the diesel engine degradation prediction package to generate the degradation trajectory of each component; The degradation trajectory of each component is segmented based on the continuous rise, sudden increase, plateau, and fall of the degradation trend value, and mapped to normal, attention, alarm, and fault imminent risk states. The failure probability is calculated based on the risk status, degradation trend value, predicted time step and historical failure node matching results. The component with the highest failure probability and consistent with the cross-component mixed token response order is marked as the failure component. The time interval between the time of the near-fault risk state and the current data collection time is determined as the remaining service life; A diesel engine failure risk prediction package is generated by combining the diesel engine number, task phase, failure probability, fault component attribution, remaining service life, and risk status.

9. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The generation of dynamic risk warning results and predictive maintenance schedules includes: The diesel engine failure risk prediction package is analyzed to extract failure probability, failure component attribution, remaining service life and risk status, and warning levels are marked according to failure probability and risk status. Add the remaining service life to the current data collection time to obtain the maintenance deadline, and combine the faulty component attribution, warning level and maintenance deadline to generate a fault warning item; Match fault warning items to the support task window no later than the maintenance deadline, and determine the maintenance sequence according to the fault probability from high to low and the remaining service life from short to long; Based on the attribution of the faulty component, the warning level, the repair deadline, the support task window, and the repair sequence, dynamic risk warning results and predictive maintenance schedules are generated.

10. The diesel engine fault prediction method based on big data analysis according to claim 1, characterized in that, The support task window specifically includes the task execution time period, the downtime for maintenance, the spare parts availability time period, the availability of maintenance personnel, the availability of special tools, and the permitted replacement time period. Maintenance periods that overlap with the task execution time period are excluded, and the overlapping time periods of the downtime for maintenance, the availability of spare parts, the availability of maintenance personnel, the availability of special tools, and the permitted replacement time period are used as the support task window.

Images