Automobile filter abnormity monitoring method based on multi-sensor fusion

Abstract

The invention discloses an automobile filter abnormity monitoring method based on multi-sensor fusion, which comprises the following steps: acquiring and preprocessing multi-modal sensing data to generate standardized multi-modal sensing data; carrying out working condition identification and standardization on multi-modal sensing data to obtain working condition state variables and a monitoring sequence; an improved Holt-Winters model is constructed, and an automobile filter prediction baseline is obtained; calculating a residual error sequence and performing hierarchical processing to obtain a low-frequency residual error and a high-frequency residual error; executing a CUSUM algorithm to obtain a bidirectional CUSUM statistical magnitude; statistical analysis is carried out, a self-adaptive threshold value is constructed, and the category and severity of the abnormal event are judged. According to the method, the improved Holt-Winters model and the CUSUM algorithm are combined, so that accurate abnormal event identification and classification of the automobile filter in a complex multi-working-condition environment are realized.

Description

Technical Field

[0001] This invention relates to the fields of vehicle engineering and intelligent condition monitoring technology, and in particular to a method for monitoring abnormalities in automotive filters based on multi-sensor fusion. Background Technology

[0002] With the continuous increase in vehicle ownership and the gradual tightening of emission regulations, the scale of operational data, engine condition data, intake air flow data, and differential pressure sensing data generated by vehicle intake systems continues to grow. Engine operating conditions and usage environments are becoming increasingly complex, increasing the demand for online monitoring of intake filter clogging, leakage, and performance degradation. Traditional filter condition monitoring often relies on single differential pressure thresholds, empirical rules, or diagnostic models based on a limited number of statistical characteristics. When faced with frequent changes in vehicle operating conditions, variable environmental conditions, and the superposition of long-term degradation and short-term disturbances, it is often difficult to distinguish between normal fluctuations and genuine anomalies in a timely and accurate manner.

[0003] In existing technologies, to address the need for filter monitoring under complex operating conditions, some solutions have begun to introduce multi-sensor data fusion and data-driven models. By fusing multi-dimensional information such as differential pressure, intake airflow, engine speed, load, and vehicle speed, a comprehensive assessment of the filter's condition can be conducted. Other technologies attempt to use time-series models or simple trend analysis methods to characterize the evolution of filter clogging, thereby improving sensitivity to abnormal changes. However, these methods often rely on static models calibrated under fixed operating conditions or single-scale residual analysis, lacking the ability to dynamically adjust the model structure and thresholds for different operating conditions. They also struggle to effectively separate different types of components, such as long-term degradation, periodic fluctuations, and instantaneous impacts. Furthermore, their monitoring stability and fine-grained classification of abnormal behavior under multi-condition switching scenarios remain insufficient, hindering reliable online identification and early warning of automotive filter anomalies in complex real-world operating environments.

[0004] Therefore, how to provide a method for detecting abnormalities in automotive filters based on multi-sensor fusion 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 multi-sensor fusion-based method for monitoring anomalies in automotive filters. This invention utilizes multi-source sensor data, including differential pressure, intake airflow, engine operating conditions, environmental parameters, and vibration acoustics. It introduces operating condition identification and weighted fusion mechanisms to construct a monitoring sequence and employs an improved Holt-Winters model to generate a filter prediction baseline that can be dynamically adjusted according to operating conditions. The invention details key processing steps such as operating condition-driven data fusion modeling, separation of trend and periodic changes, long-term degradation accumulation modeling, and model confidence gating fusion. Simultaneously, a residual hierarchical mechanism is constructed to extract low-frequency and high-frequency residuals, and an improved CUSUM algorithm and adaptive threshold mechanism are used to achieve anomaly triggering and anomaly event classification, enabling reliable determination of anomaly types such as blockage, leakage, and performance degradation. By combining multi-sensor fusion, the improved Holt-Winters model, and the CUSUM algorithm, this invention achieves higher robustness and detection accuracy than single-sensor threshold methods, possessing advantages such as strong resistance to operating condition disturbances, low false alarm rate, high identification sensitivity, and good long-term online stability.

[0006] A method for detecting abnormalities in automotive filters based on multi-sensor fusion according to an embodiment of the present invention includes:

[0007] Collect multimodal sensing data from automotive filters and engines, preprocess the multimodal sensor data, and generate standardized multimodal sensing data;

[0008] Based on standardized multimodal sensor data, the vehicle operating status is identified to obtain operating status variables, and the standardized multimodal sensor data is fused to form a monitoring sequence.

[0009] An improved Holt–Winters model was constructed, and an adaptive smoothing mechanism and a trend freezing mechanism were introduced. Degenerate input correction, operating parameter selection, trend constraint processing and multi-level recursive update were performed on the monitoring sequence and operating state variables to obtain the prediction baseline of the automotive filter.

[0010] The residual sequence is calculated based on the prediction baseline and monitoring sequence of the car filter, and the residual sequence is stratified to obtain low-frequency residual and high-frequency residual;

[0011] The low-frequency and high-frequency residuals are standardized to obtain standardized residuals. The CUSUM algorithm, which introduces a variable shrinkage factor and a joint suppression mechanism of operating conditions, is executed. By accumulating the positive and negative offsets of the standardized residuals, a bidirectional CUSUM statistic containing positive and negative CUSUM statistics is obtained.

[0012] Statistical analysis is performed on the bidirectional CUSUM statistic, an adaptive threshold is constructed, and anomalies are determined based on whether the CUSUM statistic exceeds the threshold in continuous sampling points. The anomaly category and severity are then output.

[0013] Optionally, the multimodal sensing data specifically includes filter-related data, intake fluid data, engine operating condition data, environmental data, and vibration and acoustic data.

[0014] Optionally, the preprocessing of multimodal sensor data specifically includes time synchronization, missing value completion, noise filtering, outlier correction, and normalization.

[0015] Optionally, forming the monitoring sequence includes:

[0016] Based on standardized multimodal sensor data, the data is classified and organized according to filter-related data, intake fluid data, engine operating condition data, environmental data, and vibration and acoustic data to form a multi-channel standardized dataset arranged in chronological order.

[0017] Engine operating condition data is extracted based on a multi-channel standardized dataset to obtain operating condition features including engine speed, engine load, throttle opening, vehicle speed, and intake manifold pressure.

[0018] The operating condition characteristics at each time point are classified into several discrete vehicle operating condition states, and the operating condition states at each time point are recorded as operating condition state variables.

[0019] Based on the operating condition variables, the fusion weights of various types of sensor data under different operating conditions are determined. The standardized multimodal sensor data at each time point are weighted and combined according to the fusion weights to obtain the monitoring sequence that changes over time.

[0020] Optionally, obtaining the vehicle filter prediction baseline includes:

[0021] An improved Holt-Winters model is constructed, which consists of a health degradation layer, an adaptive working condition level layer, a constrained trend layer, and a seasonal working condition layer.

[0022] The health degradation layer processes filter-related data and historical low-frequency residuals in the monitoring sequence to obtain the clogging drive and long-term offset. It also determines the degradation contribution factor based on the operating condition variables and low-frequency residual volatility, adjusts the contribution of the clogging drive and long-term offset, and outputs the long-term degradation component representing the degree of filter element degradation accumulation at the current moment based on the monotonic accumulation rule.

[0023] Based on the long-term degradation component, the monitoring sequence is subjected to degradation correction processing to remove the cumulative effects caused by long-term degradation from the monitoring sequence and generate a corrected monitoring sequence.

[0024] The adaptive working condition level layer processes the calibration monitoring sequence and introduces an adaptive working condition smoothing mechanism. It selects the corresponding smoothing coefficient according to the working condition state variables, takes the difference between the calibration monitoring sequence and the benchmark level component of the previous time as the input difference, performs input difference limiting processing, determines the input confidence factor according to the residual volatility within the sliding window, and performs weighted calculation on the input after limiting processing to obtain the benchmark level component determined by the current working condition state.

[0025] The restricted trend layer uses the difference between the baseline level components at the current time and the previous time as the trend processing input, and restricts the trend change amplitude based on the difference between the long-term degradation components at adjacent times. Before executing the trend update, the consistency of the forward and backward directions of the trend processing input is checked, and a trend freezing mechanism is introduced. When the trend direction is frequently reversed, the trend update is frozen according to the operating condition variables. When the freezing condition is not triggered, the trend processing input is processed to obtain the trend components that characterize the long-term change trend in the monitoring sequence.

[0026] Based on the operating condition state variables, the operating condition segment is determined. The seasonal layer selects seasonal items within the operating condition segment and introduces a periodic structure matching mechanism. For each candidate seasonal item, the periodic fitting residual of the correction monitoring sequence is calculated. The seasonal item with the smallest fitting error is selected according to the matching degree of the periodic fitting residual. Before performing periodic residual processing, a self-check of the seasonal fitting residual is performed on the selected seasonal item. When the seasonal fitting residual exceeds the threshold, the update weight of the seasonal component is reduced. Based on the finally determined seasonal item, periodic residual processing is performed on the correction monitoring sequence and the baseline level component to obtain the periodic component that characterizes the periodic change of the monitoring sequence under the current operating condition.

[0027] A confidence gating mechanism is introduced to calculate confidence scores for long-term degradation components, baseline components, trend components, and periodic components. The confidence scores are dynamically determined based on the current operating condition variables and historical residual fluctuation levels. Each component is multiplied by its corresponding confidence score as a weighting coefficient, and weighted fusion is performed to obtain the current vehicle filter prediction baseline.

[0028] Optionally, the step of performing layered processing on the residual sequence to obtain low-frequency residuals and high-frequency residuals includes:

[0029] Aligning the predicted baseline and monitoring sequence based on the time index of the vehicle filter prediction baseline, the predicted baseline value at each time moment is paired with the monitoring sequence value at the corresponding time moment to form a paired dataset arranged in chronological order;

[0030] The difference between the monitored sequence values ​​and the predicted baseline values ​​at each time point in the paired dataset is calculated to obtain the residual sequence arranged in chronological order;

[0031] The residual sequence is processed in layers. Low-frequency components are extracted from the residual sequence using a preset filtering rule to obtain low-frequency residuals. The low-frequency residuals are then subtracted from the residual sequence to obtain high-frequency residuals.

[0032] Optionally, obtaining the bidirectional CUSUM statistic, which includes both positive and negative CUSUM statistics, includes:

[0033] The low-frequency and high-frequency residuals at each time point are paired according to a unified time index to form a residual dataset arranged in chronological order. Statistical features are calculated and standardized for the low-frequency and high-frequency residuals at each time point in the residual dataset to obtain low-frequency standardized residuals and high-frequency standardized residuals.

[0034] A standardized residual sequence is constructed based on the low-frequency standardized residual and the high-frequency standardized residual. The low-frequency standardized residual and the high-frequency standardized residual at each time step are weighted and summed to form the standardized residual input.

[0035] The CUSUM algorithm is executed. At the initial moment, the positive and negative CUSUM statistics are initialized to zero. At non-initial moments, the positive and negative CUSUM statistics of the previous moment are used as the accumulation basis for the current moment. Based on the magnitude of the operating condition variables and the standardized residual input, the corresponding variable shrinkage factor is determined for the current moment. The positive and negative CUSUM statistics are shrunk to obtain the intermediate positive and intermediate negative CUSUM statistics adjusted by the variable shrinkage factor.

[0036] A joint suppression mechanism for operating conditions is introduced. When the operating condition variable indicates a range of drastic changes in operating conditions, the intermediate positive CUSUM statistic and the intermediate negative CUSUM statistic are suppressed. When the operating condition variable indicates a range of stable operating conditions, the CUSUM statistic is updated based on the standardized residual sequence.

[0037] Under the combined suppression mechanism of variable shrinkage factor and operating condition, the standardized residual input at each time point is updated by positive accumulation and negative accumulation respectively, and the positive CUSUM statistic and negative CUSUM statistic at each time point are obtained.

[0038] The positive and negative CUSUM statistics at each time point are combined to form a bidirectional CUSUM statistics sequence arranged in chronological order.

[0039] Optionally, the output anomaly event category and severity include:

[0040] Statistical analysis was performed on the positive and negative CUSUM statistics at each time point, and statistical characteristics including the maximum value, average value, and continuous overthreshold length were calculated.

[0041] Based on the baseline features of statistical characteristics and historical bidirectional CUSUM statistical sequences, adaptive positive thresholds and adaptive negative thresholds are constructed respectively.

[0042] The positive CUSUM statistic is compared with the adaptive positive threshold, and the negative CUSUM statistic is compared with the adaptive negative threshold. Positive and negative anomalous events are marked according to the threshold exceedance of the positive and negative CUSUM statistic in continuous sampling points.

[0043] The abnormal events are classified based on the statistical characteristics, threshold amplitude, and duration corresponding to positive and negative abnormal events. The abnormality category and severity of each abnormal event are determined and output.

[0044] The beneficial effects of this invention are:

[0045] This invention proposes a multi-sensor fusion-based method for monitoring anomalies in automotive filters. By utilizing multi-source sensor information such as differential pressure, intake airflow, engine operating parameters, environmental data, and vibration acoustic data, and combining it with operating condition state variables, a multi-modal fusion monitoring sequence is constructed to achieve a unified representation of the filter's state under different operating conditions. An improved Holt-Winters model is introduced, decomposing the monitoring sequence into long-term degradation components, baseline components, trend components, and periodic components. Through mechanisms such as health degradation modeling, adaptive operating condition level adjustment, constrained trend updates, and operating condition seasonal term matching, a predictive baseline that dynamically changes with operating conditions is generated. Furthermore, a confidence gating mechanism is incorporated to dynamically adjust the contribution of different components to the predictive baseline based on the operating condition and residual fluctuation level, enhancing the model's robustness in complex operating environments.

[0046] The method of this invention separates the long-term slowly varying low-frequency residuals from the high-frequency residuals representing instantaneous shocks through a residual hierarchical mechanism and merges them into a standardized residual sequence. It introduces an improved CUSUM algorithm, a variable shrinkage factor, and a joint suppression mechanism of operating conditions to ensure responsiveness to highly sensitive anomalies while suppressing false accumulations caused by operating condition fluctuations. Finally, it achieves anomaly event identification and severity classification through adaptive positive and negative thresholds.

[0047] The method of this invention realizes the effective fusion and dynamic modeling of multi-sensor information under complex working conditions. It can simultaneously characterize the long-term degradation trend and short-term shock fluctuations of filters. It has higher accuracy and stability in the identification of multiple types of anomalies such as blockage, leakage and performance degradation. Compared with the traditional single sensor threshold method and static model, it can improve sensitivity and online detection reliability while reducing false alarms and missed alarms. It has good engineering application value for automotive filter health monitoring. Attached Figure Description

[0048] 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:

[0049] Figure 1 This is a flowchart of a method for detecting abnormalities in automotive filters based on multi-sensor fusion proposed in this invention;

[0050] Figure 2 This is a schematic diagram of the improved Holt-Winters model structure of an automotive filter anomaly monitoring method based on multi-sensor fusion proposed in this invention.

[0051] Figure 3 This is a flowchart of the CUSUM algorithm processing for an anomaly monitoring method for automotive filters based on multi-sensor fusion proposed in this invention. Detailed Implementation

[0052] 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.

[0053] refer to Figure 1 , Figure 2 and Figure 3 A method for detecting abnormalities in automotive filters based on multi-sensor fusion, comprising:

[0054] Collect multimodal sensing data from automotive filters and engines, preprocess the multimodal sensor data, and generate standardized multimodal sensing data;

[0055] Based on standardized multimodal sensor data, the vehicle operating status is identified to obtain operating status variables, and the standardized multimodal sensor data is fused to form a monitoring sequence.

[0056] An improved Holt–Winters model was constructed, and an adaptive smoothing mechanism and a trend freezing mechanism were introduced. Degenerate input correction, operating parameter selection, trend constraint processing and multi-level recursive update were performed on the monitoring sequence and operating state variables to obtain the prediction baseline of the automotive filter.

[0057] The residual sequence is calculated based on the prediction baseline and monitoring sequence of the car filter, and the residual sequence is stratified to obtain low-frequency residual and high-frequency residual;

[0058] The low-frequency and high-frequency residuals are standardized to obtain standardized residuals. The CUSUM algorithm, which introduces a variable shrinkage factor and a joint suppression mechanism of operating conditions, is executed. By accumulating the positive and negative offsets of the standardized residuals, a bidirectional CUSUM statistic containing positive and negative CUSUM statistics is obtained.

[0059] Statistical analysis is performed on the bidirectional CUSUM statistic, an adaptive threshold is constructed, and anomalies are determined based on whether the CUSUM statistic exceeds the threshold in continuous sampling points. The anomaly category and severity are then output.

[0060] In this embodiment, the multimodal sensing data specifically includes filter-related data, intake fluid data, engine operating data, environmental data, and vibration and acoustic data.

[0061] In this embodiment, the preprocessing of multimodal sensor data specifically includes time synchronization, missing value completion, noise filtering, outlier correction, and normalization.

[0062] In this embodiment, forming the monitoring sequence includes:

[0063] Based on standardized multimodal sensor data, the data is classified and organized according to filter-related data, intake fluid data, engine operating condition data, environmental data, and vibration and acoustic data to form a multi-channel standardized dataset arranged in chronological order.

[0064] Engine operating condition data is extracted based on a multi-channel standardized dataset to obtain operating condition features including engine speed, engine load, throttle opening, vehicle speed, and intake manifold pressure.

[0065] The operating condition characteristics at each time point are classified into several discrete vehicle operating condition states, and the operating condition states at each time point are recorded as operating condition state variables.

[0066] The fusion weights of various types of sensor data under different operating conditions are determined based on the operating condition state variables. Standardized multimodal sensor data at each time point are then weighted and combined according to the fusion weights to obtain a monitoring sequence that changes over time. Specifically, determining the fusion weights of various types of sensor data under different operating conditions based on the operating condition state variables involves:

[0067] The operating status is classified according to the operating condition variables including engine speed, load, throttle opening and vehicle speed. The entire operating condition is divided into several typical operating conditions. During the calibration phase, multiple types of sensor data, including filter-related data, intake fluid data, engine operating condition data and environmental data, are collected under each operating condition. Combined with the filter health calibration results, the correlation and noise level between various types of sensor data and filter status are analyzed. Sensor data with strong correlation and low noise are given a larger fusion weight, while sensor data with weak correlation and susceptible to operating condition disturbances are given a smaller fusion weight.

[0068] In this embodiment, obtaining the predicted baseline for the automotive filter includes:

[0069] An improved Holt-Winters model is constructed, which consists of a health degradation layer, a working condition adaptive level layer, a constrained trend layer, and a working condition seasonal layer. Specifically, the construction of the improved Holt-Winters model involves:

[0070] The newly added health degradation layer is the first step in processing the input of the improved Holt-Winters model. Based on the horizontal layer of the Holt-Winters model, an adaptive smoothing mechanism for operating conditions is introduced to improve it into an adaptive horizontal layer for operating conditions. An operating condition-driven trend freezing mechanism is added to the trend layer of the Holt-Winters model to form a restricted trend layer. In the seasonal layer of the Holt-Winters model, an operating condition seasonal layer is formed by introducing a segmented operating condition and a multi-period structure matching mechanism and seasonal residual self-checking.

[0071] The health degradation layer processes filter-related data and historical low-frequency residuals in the monitoring sequence to obtain the clogging drive and long-term offset. It then determines the degradation contribution factor based on the operating condition variables and low-frequency residual volatility, adjusting the contribution of the clogging drive and long-term offset. Based on a monotonic accumulation rule, it outputs the long-term degradation component representing the degree of filter element degradation accumulation at the current moment, where:

[0072] The specific details of obtaining the blockage drive amount and long-term offset are as follows:

[0073] Based on filter-related data related to filter differential pressure and intake flow in the monitoring sequence, the amplitude of differential pressure data at each time point is extracted. The original differential pressure sequence is denoised and RMS value is obtained to get the effective differential pressure amplitude at the corresponding time point. The intake flow rate at the corresponding time point is normalized. The change in effective differential pressure amplitude is combined with the normalized intake flow rate to obtain the blockage driving amount. Using historical low-frequency residuals as input, the low-frequency residuals are segmented into time windows. The average residuals within each time window are extracted, and the averages of adjacent time windows are accumulated to obtain the long-term offset.

[0074] The determination of the degradation contribution factor is as follows:

[0075] Based on the operating condition variables, it is determined whether the current vehicle is in a stable operating condition range or a rapidly changing range. Based on the observability of differential pressure and the stability of intake flow characteristics, various operating conditions are classified. Based on the classification results, a basic contribution coefficient is set for different operating condition types. The basic contribution coefficient for the stable operating condition range is higher than that for the rapidly changing range. The basic contribution coefficient is adjusted according to the fluctuation of low-frequency residuals within the time window. The contribution coefficient is increased for periods with gentle changes and small fluctuations in low-frequency residuals, and decreased for periods with drastic fluctuations and large fluctuations in low-frequency residuals. The adjusted contribution coefficient is used as the degradation contribution factor at the current moment.

[0076] The adjustment of the contribution of the blocking drive amount and the long-term offset is specifically as follows:

[0077] Using the degradation contribution factor at the current moment as the adjustment coefficient, the blockage driving amount and long-term offset are weighted according to the adjustment coefficient to obtain the weighted blockage driving amount and long-term offset.

[0078] When updating the long-term degradation component, the monotonic accumulation rule updates the long-term degradation component to the sum of the weighted congestion drive and long-term offset when the sum of the weighted congestion drive and long-term offset is greater than the long-term degradation component at the previous time step; otherwise, it keeps the long-term degradation component at the previous time step unchanged, forming a degradation accumulation that monotonically does not decrease over time.

[0079] Based on long-term degradation components, degradation correction processing is performed on the monitoring sequence to remove the cumulative effects of long-term degradation from the monitoring sequence, generating a corrected monitoring sequence. Specifically, the degradation correction processing of the monitoring sequence involves:

[0080] Using the long-term degradation component at the current moment as the correction benchmark, the original monitoring value at the corresponding moment of the monitoring sequence is subjected to difference processing. The long-term degradation component at the current moment is subtracted from the original monitoring value to eliminate the cumulative offset caused by long-term degradation from the monitoring sequence. The correction results at each moment after difference processing are arranged in chronological order to form the corrected monitoring sequence.

[0081] The adaptive operating condition level layer processes the calibration monitoring sequence by introducing an adaptive smoothing mechanism. It selects a smoothing coefficient based on the operating condition state variables, uses the difference between the calibration monitoring sequence and the previous time-bound baseline level component as the input difference, performs input difference limiting, and determines the input confidence factor based on the residual volatility within the sliding window. The weighted input difference is then calculated to obtain the baseline level component determined by the current operating condition state, where:

[0082] When processing the calibration monitoring sequence, the working condition adaptive smoothing mechanism uses the working condition state variable as the basis for selecting the smoothing coefficient. Different smoothing coefficients are called under different working condition types. The response speed of the smoothing processing to the calibration monitoring sequence is adjusted with the change of working condition. When the working condition state variable indicates that the vehicle is in a stable working condition, a relatively large smoothing coefficient is selected. When the working condition state variable indicates that the vehicle is in a rapidly changing working condition, a relatively small smoothing coefficient is selected. The input difference is weighted according to the smoothing coefficient and then subjected to amplitude limiting processing.

[0083] The weighted calculation of the input difference after amplitude limiting is specifically as follows:

[0084] The input confidence factor is determined based on the volatility of the low-frequency residuals within the time window. Periods with lower low-frequency residual volatility correspond to higher input confidence factors, while periods with higher residual volatility correspond to lower input confidence factors. The input difference after amplitude limiting is weighted according to the input confidence factor. At the same time, the benchmark level component of the previous moment is weighted according to the weights complementary to the input confidence factor. The weighted result of the weighted input difference and the benchmark level component of the previous moment is synthesized at the current moment as the benchmark level component of the current moment.

[0085] The constrained trend layer uses the difference between the baseline level components at the current and previous times as the trend processing input, and restricts the trend change amplitude based on the difference between long-term degradation components at adjacent times. Before executing the trend update, the consistency of the forward and backward directions of the trend processing input is checked, and a trend freezing mechanism is introduced. When the trend direction frequently reverses, the trend update is frozen according to the operating condition variables. When the freezing condition is not triggered, the trend processing input is processed to obtain the trend components representing the long-term change trend in the monitoring sequence, where:

[0086] The verification of the consistency of the trend processing input in the forward and backward directions specifically includes:

[0087] The current trend processing input sign is compared with the previous trend processing input sign to determine whether the trend change direction is consistent. When the current trend processing input direction is consistent with the previous trend processing input direction, it is determined that the direction is consistent. When the sign changes, it is determined that the trend direction is reversed.

[0088] The trend freezing mechanism is activated when the trend direction frequently reverses. It determines whether the vehicle is in a drastically fluctuating condition or a stable condition based on the operating condition state variable. When the operating condition state variable indicates that the vehicle is in a drastically changing condition, the current trend processing is temporarily suspended. Instead, the trend component of the previous moment is used as the current trend output. The trend update remains unchanged during the freezing period. When the operating condition state variable indicates that the vehicle is in a stable condition, the trend processing input is processed. The current trend processing input is combined with the trend component of the previous moment according to the weight to obtain the current trend component.

[0089] Based on the operating condition state variables, the operating condition segment is determined. Within each operating condition segment, a seasonal term is selected, and a periodic structure matching mechanism is introduced. For each candidate seasonal term, the periodic fitting residual of the corrected monitoring sequence is calculated. The seasonal term with the smallest fitting error is selected based on the matching degree of the periodic fitting residual. Before processing the periodic residual, a self-check of the seasonal fitting residual is performed on the selected seasonal term. When the seasonal fitting residual exceeds a threshold, the update weight of the seasonal component is reduced. Based on the finally determined seasonal term, periodic residual processing is performed on the corrected monitoring sequence and the baseline level component to obtain the periodic component characterizing the periodic change of the monitoring sequence under the current operating condition, where:

[0090] The periodic structure matching mechanism, within a given operating condition segment, utilizes historical monitoring data to perform periodic alignment. Data from the same periodic phase are statistically analyzed and averaged to obtain the periodic variation profile under the operating condition segment. This profile is denoted as a set of candidate seasonal terms corresponding to the operating condition segment. Each candidate seasonal term is periodically fitted to the calibration monitoring sequence, and the corresponding periodic fitting residual is calculated. The seasonal term with the smallest periodic fitting residual is selected as the seasonal term used under the current operating condition segment. Periodic residual processing is then applied to the calibration monitoring sequence and the baseline level components to obtain the periodic components, where:

[0091] The calculation of the corresponding periodic fitting residual is specifically as follows:

[0092] For each candidate seasonal item, the value of the candidate seasonal item in each phase is used as the seasonal component. Periodic fitting is performed on the calibration monitoring sequence. For each sampling time in the working condition segment, the seasonal value of the candidate seasonal item in the corresponding phase is read. The seasonal value is superimposed with the reference level component at the same time to obtain the fitted value under the candidate seasonal item. The periodic fitting residual is obtained by subtracting the fitted value from the actual value of the calibration monitoring sequence.

[0093] The periodic residual processing of the calibration monitoring sequence and the baseline level components specifically involves:

[0094] Read the seasonal value of the corresponding phase from the seasonal item, take the seasonal value as the periodic component of the corresponding time, subtract the corresponding periodic component from the actual value of the correction monitoring sequence at the current time to obtain the correction monitoring sequence after removing the periodic component, and subtract the corresponding periodic component from the value of the reference level component at the current time to obtain the reference level component after removing the periodic component.

[0095] A confidence gating mechanism is introduced to calculate confidence scores for long-term degradation components, baseline components, trend components, and periodic components. These confidence scores are dynamically determined based on the current operating condition variables and historical residual fluctuation levels. Each component is multiplied by its corresponding confidence score as a weighting coefficient, and weighted fusion is performed to obtain the current vehicle filter prediction baseline.

[0096] The confidence gating mechanism determines whether the vehicle is in a stable operating condition, or an accelerating, decelerating, or rapidly changing operating condition based on the current operating condition state variables. It sets corresponding operating condition factor values ​​for different operating conditions, assigning a operating condition factor close to 1 to a stable operating condition, a medium value to a slowly changing operating condition, and a small value to a rapidly changing operating condition.

[0097] Using the sampling time as the endpoint, historical residual sequences corresponding to the long-term degradation component, the baseline component, the trend component, and the periodic component are extracted within the time window. The average absolute value of the historical residual sequence of each component is calculated to obtain the residual volatility index of each component. The index is then normalized and converted into a volatility confidence factor in the range of 0 to 1. The smaller the residual volatility index, the closer the corresponding volatility confidence factor is to 1; the larger the residual volatility index, the closer it is to 0.

[0098] The confidence score of the corresponding component under the current operating condition is obtained by weighted summation of the operating condition factor and the fluctuation confidence factor.

[0099] In this embodiment, the step of performing layered processing on the residual sequence to obtain low-frequency residuals and high-frequency residuals includes:

[0100] Aligning the predicted baseline and monitoring sequence based on the time index of the vehicle filter prediction baseline, the predicted baseline value at each time moment is paired with the monitoring sequence value at the corresponding time moment to form a paired dataset arranged in chronological order;

[0101] The difference between the monitored sequence values ​​and the predicted baseline values ​​at each time point in the paired dataset is calculated to obtain the residual sequence arranged in chronological order;

[0102] The residual sequence is layered, and low-frequency components are extracted from the residual sequence using a preset filtering rule to obtain low-frequency residuals. The low-frequency residuals are then subtracted from the residual sequence to obtain high-frequency residuals. Where:

[0103] The preset filtering rules refer to the selection of filter type and parameter configuration based on the sampling period of the monitoring sequence and the frequency range to be separated. By combining the spectral distribution of the residual sequence under typical operating conditions during the calibration stage, a set of fixed filtering parameters is determined that can effectively retain the low-frequency components reflecting the long-term degradation and slow changes in operating conditions of the filter, and suppress the high-frequency components of transient impact and random noise. During online operation, the residual sequence is processed according to the filtering parameters. The residual value at each moment is input into the filter in chronological order. The residual sequence is weighted and summed point by point to obtain the low-frequency residual.

[0104] In this embodiment, obtaining the bidirectional CUSUM statistic, which includes both positive and negative CUSUM statistics, includes:

[0105] The low-frequency and high-frequency residuals at each time step are paired according to a unified time index to form a residual dataset arranged in chronological order. Statistical features are then calculated and standardized for the low-frequency and high-frequency residuals at each time step in the residual dataset to obtain low-frequency standardized residuals and high-frequency standardized residuals, where:

[0106] The calculation of the statistical features is performed as follows:

[0107] The mean and standard deviation of the low-frequency and high-frequency residuals within the time window were calculated respectively to obtain the corresponding statistical characteristics;

[0108] The standardization process is carried out as follows:

[0109] Subtract the mean of the low-frequency residuals within the time window at each time step from the low-frequency residual at each time step, and then divide by the corresponding standard deviation to obtain the low-frequency standardized residual at the current time step. Subtract the mean of the high-frequency residuals within the time window at each time step from the high-frequency residual at each time step, and then divide by the corresponding standard deviation to obtain the high-frequency standardized residual at the current time step.

[0110] A standardized residual sequence is constructed based on the low-frequency standardized residual and the high-frequency standardized residual. The low-frequency standardized residual and the high-frequency standardized residual at each time step are weighted and summed to form the standardized residual input.

[0111] The CUSUM algorithm is executed. Initially, both the positive and negative CUSUM statistics are initialized to zero. At non-initial times, the positive and negative CUSUM statistics from the previous time step are used as the cumulative basis for the current time step. Based on the magnitudes of the operating condition variables and the standardized residual input, a corresponding variable shrinkage factor is determined for the current time step. The positive and negative CUSUM statistics are then shrunk to obtain intermediate positive and negative CUSUM statistics adjusted by the variable shrinkage factor, where:

[0112] The determination of the corresponding variable contraction factor at the current moment is specifically as follows:

[0113] At each sampling moment, the current operating condition is determined as a stable, slowly changing, or rapidly changing condition based on the operating condition state variables. The basic shrinkage factor corresponding to the stable operating condition is close to 1, while the basic shrinkage factor corresponding to the rapidly changing operating condition is smaller. Then, the amplitude of the standardized residual input at the current moment is read, and the amplitude of the standardized residual input is compared with the amplitude threshold and divided into small, medium, and large deviation levels. Different amplitude correction coefficients are set accordingly. The larger the residual amplitude, the smaller the correction coefficient. The basic shrinkage factor determined by the operating condition state is multiplied by the correction coefficient determined by the residual amplitude to obtain the variable shrinkage factor at the current moment.

[0114] The shrinking process for the positive and negative CUSUM statistics is as follows:

[0115] After determining the variable shrinkage factor corresponding to the current time, read the positive CUSUM statistic and the negative CUSUM statistic before the current time update respectively. Multiply the positive CUSUM statistic by the variable shrinkage factor to obtain the intermediate positive CUSUM statistic after shrinkage. Multiply the negative CUSUM statistic by the variable shrinkage factor to obtain the intermediate negative CUSUM statistic after shrinkage.

[0116] A joint suppression mechanism for operating conditions is introduced. When the operating condition variable indicates a period of drastic change, the intermediate positive and negative CUSUM statistics are suppressed. When the operating condition variable indicates a period of stable operation, the CUSUM statistics are updated based on the standardized residual sequence.

[0117] At each sampling time, the joint suppression mechanism for operating conditions reads the operating condition variables to determine whether the current operating condition is in a period of drastic change. When the operating condition variables indicate that the operating condition is in a period of drastic change, a unified suppression process is applied to the intermediate positive CUSUM statistic and the intermediate negative CUSUM statistic, which have been adjusted by the variable shrinkage factor. An suppression coefficient less than 1 is assigned, and the intermediate positive CUSUM statistic and the intermediate negative CUSUM statistic are multiplied by the suppression coefficient respectively to obtain the compressed positive and negative CUSUM statistics at the current time.

[0118] When the operating condition state variable indicates a stable operating range, the positive CUSUM statistic of the previous time step is used as the basis. The positive offset control is subtracted from the standardized residual sequence of the current time step and then added to the positive CUSUM statistic to obtain the positive cumulative result of the current time step. If the result is less than zero, it is truncated to zero. Similarly, the negative CUSUM statistic of the previous time step is used as the basis. The negative offset control is subtracted from the standardized residual sequence of the current time step and then added to the negative CUSUM statistic to obtain the negative cumulative result of the current time step. If the result is less than zero, it is also truncated to zero.

[0119] Under the combined suppression mechanism of variable shrinkage factor and operating condition, the standardized residual input at each time point is updated by positive accumulation and negative accumulation respectively, and the positive CUSUM statistic and negative CUSUM statistic at each time point are obtained.

[0120] The positive and negative CUSUM statistics at each time point are combined to form a bidirectional CUSUM statistics sequence arranged in chronological order. Specifically, the combination of the positive and negative CUSUM statistics at each time point involves:

[0121] At each sampling time, the positive and negative CUSUM statistics corresponding to the current time are read and paired according to a unified time index. The positive and negative CUSUM statistics at the same time are combined into a set of bidirectional statistics data. Then, the bidirectional statistics data of each time are arranged in chronological order to form a bidirectional CUSUM statistics sequence.

[0122] In this embodiment, the output anomaly event category and severity include:

[0123] Statistical analysis was performed on the positive and negative CUSUM statistics at each time point, and statistical characteristics including the maximum value, average value, and continuous overthreshold length were calculated.

[0124] Based on the baseline features of statistical characteristics and historical bidirectional CUSUM statistical sequences, adaptive positive thresholds and adaptive negative thresholds are constructed respectively. The construction of adaptive positive thresholds and adaptive negative thresholds specifically involves:

[0125] Baseline features of positive and negative CUSUM statistics are extracted from historical bidirectional CUSUM statistic sequences, including their respective means and standard deviations. Then, within the current detection window, based on the statistical features corresponding to the current bidirectional CUSUM statistics, the current statistical features are combined with historical baseline features. For the positive direction, the mean of the historical positive CUSUM statistic baseline features plus a certain multiple of the baseline standard deviation is used as the base threshold. This base threshold is then adjusted according to changes in the statistical features within the current window to obtain an adaptive positive threshold. For the negative direction, the mean of the historical negative CUSUM statistic baseline features plus a certain multiple of the baseline standard deviation is used as the base threshold. This base threshold is then adjusted according to changes in the statistical features within the current window to obtain an adaptive negative threshold. Specifically, adjusting the base threshold according to changes in the statistical features within the current window involves:

[0126] Calculate the ratio of the current window statistical feature to the historical baseline statistical feature. When the ratio is greater than 1, multiply the base threshold by a coefficient that monotonically increases with the ratio and is greater than or equal to 1 to amplify the base threshold. When the ratio is less than 1, multiply the base threshold by a coefficient that monotonically changes with the ratio and is less than or equal to 1 to shrink the base threshold.

[0127] The positive CUSUM statistic is compared with the adaptive positive threshold, and the negative CUSUM statistic is compared with the adaptive negative threshold. Positive and negative anomalous events are marked according to the threshold exceedance of the positive and negative CUSUM statistic in continuous sampling points.

[0128] Anomalies are classified based on statistical characteristics, threshold amplitude, and duration corresponding to positive and negative anomalies. The anomaly category and severity of each anomaly are determined, and the anomaly category and severity are output. Specifically, determining the anomaly category and severity of each anomaly involves:

[0129] Based on whether the abnormal event corresponds to a positive or negative abnormal event, and combined with the offset pattern of the bidirectional CUSUM statistic in the corresponding direction, it is mapped to different types of physical meanings to determine the basic category of the abnormal event. The maximum over-threshold amplitude, average over-threshold amplitude, and continuous over-threshold duration of the abnormal event are statistically analyzed. The over-threshold amplitude and duration are compared with multi-level classification thresholds. When the over-threshold amplitude and duration fall into different intervals, the abnormal event is classified into different severity levels such as mild, moderate, or severe.

[0130] Example 1:

[0131] To verify the feasibility of this invention in practice, it was applied to a pilot project for intelligent monitoring of vehicle filters on a comprehensive mountain road. The test route covered highways, gravel roads, long mountain slopes, and congested urban roads, comprehensively encompassing typical operating conditions of vehicles in real-world use. From March to July 2025, multiple multimodal monitoring terminals were deployed on three 2.0T turbocharged prototype vehicles to collect real-time data on differential pressure, intake airflow, engine speed, load, vehicle speed, throttle opening, vibration and acoustic information, as well as environmental parameters such as temperature, humidity, and particulate matter concentration. Vehicle operating rules, component attribute information, and historical operating records were also collected. All data was uploaded to a cloud monitoring platform via an onboard industrial communication unit. After standardization verification, anomaly removal, and time-series alignment, the data entered the operating condition identification, monitoring sequence construction, and predictive baseline modeling process proposed in this invention.

[0132] During the pilot phase, vehicle operation exhibited various typical fluctuation characteristics due to seasonal changes and road conditions. In late March, temperatures plummeted on mountain roads, and frequent switching between high engine speeds and sloping sections caused short-term pressure spikes. In mid-April, increased dust concentration on gravel roads led to fluctuations in intake airflow and short-term resistance changes. In late May, the duration of high-speed operation prolonged, increasing intake load and adding high-frequency disturbances to the monitoring data. In early June, frequent start-stop cycles during urban congestion significantly altered the transient changes in various sensor data. Traditional single-threshold monitoring methods struggled to distinguish between operating condition fluctuations and filter degradation, resulting in erroneous alarms under multiple typical conditions. The method of this invention adaptively adjusts the fusion weights of multi-source data using operating condition state variables, generates a predictive baseline that dynamically changes with operating conditions using an improved Holt-Winters model, and combines low-frequency and high-frequency residual hierarchical processing with an improved CUSUM algorithm to achieve a structured characterization of filter health changes.

[0133] During the pilot test, test data from different dates were automatically recorded into a results table, including operating condition identification stability, monitoring sequence noise suppression rate, prediction baseline fitting error, low-frequency residuals, RMS values ​​of high-frequency residuals, maximum values ​​of positive and negative CUSUM, adaptive threshold level, and the number and classification results of abnormal events provided by the system. Complete data records were generated for different operating phases, including March 26, April 18, May 12, June 3, June 9, and June 18, 2025. The testing process covered operating condition disturbance environments, dust interference environments, early degradation stages, and significant clogging stages. All monitoring data were processed in the cloud platform according to the method of this invention. To verify the accuracy of the monitoring results, three test vehicles were physically disassembled and inspected on June 20, 2025. The actual filter element clogging status was recorded and archived for data analysis and comparison.

[0134] Table 1. Statistical data of the method of the present invention in the pilot project of comprehensive mountain roads from March to July 2025.

[0135] date Multimodal data acquisition volume (MB) Stability of operating condition identification (%) Monitoring sequence noise suppression rate (%) Predicted baseline fit error (MSE) Low-frequency residual mean Root mean square (RMS) of high-frequency residuals Bidirectional CUSUM maximum value (positive / negative) Adaptive threshold (positive / negative) Determine the number of abnormal events Anomaly classification results 2025-03-26 285 93.4 41.2 0.018 0.05 0.27 0.12 / 0.10 0.35 / 0.33 0 No abnormalities (operating condition disturbances) 2025-04-18 310 94.1 38.6 0.020 0.06 0.31 0.15 / 0.11 0.37 / 0.34 0 No abnormalities (dust disturbance) 2025-05-12 295 95.3 46.8 0.017 0.21 0.24 0.46 / 0.14 0.39 / 0.35 2 Mild congestion trend 2025-06-03 305 96.0 49.7 0.015 0.33 0.22 0.58 / 0.18 0.42 / 0.37 3 Mild blockage 2025-06-09 330 96.4 52.1 0.014 0.41 0.20 0.63 / 0.21 0.44 / 0.39 3 Mild congestion persists 2025-06-18 345 97.2 54.8 0.012 0.74 0.18 0.92 / 0.27 0.51 / 0.42 2 Moderate congestion 2025-06-20 — — — — — — — — — Disassembly and inspection revealed blockages of 36%–47% (verified). total 1870 Average 95.4% Average 47.2% Average 0.016 — — — — 10 —

[0136] As shown in Table 1, during the pilot project on integrated mountain roads from March to July 2025, a total of 1870MB of multi-source monitoring data was collected, with an average data volume of approximately 300MB per monitoring session. This data fully covered typical operating conditions such as sudden temperature changes, dust interference, high-speed heavy loads, and urban congestion. Throughout the pilot project, the system completed the entire process of processing monitoring data from different dates, from operating condition identification, data fusion, and prediction baseline calculation to CUSUM statistic generation. Each indicator exhibited clear phased characteristics. On March 26th and April 18th, when operating condition disturbances were dominant, the noise suppression rate of the monitoring sequence reached 41.2% and 38.6%, respectively. The prediction baseline fitting error remained within the range of 0.018–0.020, and both positive and negative CUSUM values ​​were significantly lower than the threshold, enabling stable identification as non-abnormal states. This demonstrates that the present invention possesses good anti-interference capabilities under highly fluctuating environments.

[0137] As time progressed, the true degradation of the vehicle's filter gradually became apparent. Starting May 12th, the low-frequency residual and positive CUSUM values ​​steadily increased, indicating that the system had detected a slight clogging trend. By June 3rd and June 9th, the low-frequency residual had risen to 0.33 and 0.41 respectively, and the maximum CUSUM value had increased to 0.58 and 0.63. The adaptive threshold also dynamically adjusted with the fluctuation level, making the anomaly identification process more robust and reliable. On June 18th, the low-frequency residual reached 0.74, and the maximum CUSUM value was 0.92, which the system determined to be moderate clogging. Subsequent vehicle disassembly and inspection revealed that the filter element clogging area was measured to be 36%–47%, highly consistent with the anomaly level given by the system based on the residual structure and CUSUM statistics.

[0138] During the entire pilot phase, the system identified 10 abnormal events with no missed detections and only one minor false alarm, with an average positioning time of approximately 2.2 minutes. This demonstrates that the invention maintains high fitting accuracy, stable residual signal layering capability, and reliable anomaly classification performance under various operating conditions and high disturbance environments, fully verifying the effectiveness and engineering application value of the invention in actual road environments.

[0139] 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 method for detecting abnormalities in automotive filters based on multi-sensor fusion, characterized in that, include: Collect multimodal sensing data from automotive filters and engines, preprocess the multimodal sensor data, and generate standardized multimodal sensing data; Based on standardized multimodal sensor data, the vehicle operating status is identified to obtain operating status variables, and the standardized multimodal sensor data is fused to form a monitoring sequence. An improved Holt–Winters model was constructed, and an adaptive smoothing mechanism and a trend freezing mechanism were introduced. Degenerate input correction, operating parameter selection, trend constraint processing and multi-level recursive update were performed on the monitoring sequence and operating state variables to obtain the prediction baseline of the automotive filter. The residual sequence is calculated based on the prediction baseline and monitoring sequence of the car filter, and the residual sequence is stratified to obtain low-frequency residual and high-frequency residual; The low-frequency and high-frequency residuals are standardized to obtain standardized residuals. The CUSUM algorithm, which introduces a variable shrinkage factor and a joint suppression mechanism of operating conditions, is executed. By accumulating the positive and negative offsets of the standardized residuals, a bidirectional CUSUM statistic containing positive and negative CUSUM statistics is obtained. Statistical analysis is performed on the bidirectional CUSUM statistic, an adaptive threshold is constructed, and anomalies are determined based on whether the CUSUM statistic exceeds the threshold in continuous sampling points. The anomaly category and severity are then output.

2. The method for detecting abnormalities in automotive filters based on multi-sensor fusion according to claim 1, characterized in that, The multimodal sensing data specifically includes filter-related data, intake fluid data, engine operating data, environmental data, and vibration and acoustic data.

3. The method for detecting abnormalities in automotive filters based on multi-sensor fusion according to claim 1, characterized in that, The preprocessing of multimodal sensor data specifically includes time synchronization, missing value completion, noise filtering, outlier correction, and normalization.

4. The method for detecting abnormalities in automotive filters based on multi-sensor fusion according to claim 1, characterized in that, The formation of the monitoring sequence includes: Based on standardized multimodal sensor data, the data is classified and organized according to filter-related data, intake fluid data, engine operating condition data, environmental data, and vibration and acoustic data to form a multi-channel standardized dataset arranged in chronological order. Engine operating condition data is extracted based on a multi-channel standardized dataset to obtain operating condition features including engine speed, engine load, throttle opening, vehicle speed, and intake manifold pressure. The operating condition characteristics at each time point are classified into several discrete vehicle operating condition states, and the operating condition states at each time point are recorded as operating condition state variables. Based on the operating condition variables, the fusion weights of various types of sensor data under different operating conditions are determined. The standardized multimodal sensor data at each time point are weighted and combined according to the fusion weights to obtain the monitoring sequence that changes over time.

5. The method for detecting abnormalities in automotive filters based on multi-sensor fusion according to claim 1, characterized in that, The method for obtaining the automotive filter prediction baseline includes: An improved Holt-Winters model is constructed, which consists of a health degradation layer, an adaptive working condition level layer, a constrained trend layer, and a seasonal working condition layer. The health degradation layer processes filter-related data and historical low-frequency residuals in the monitoring sequence to obtain the clogging drive and long-term offset. It also determines the degradation contribution factor based on the operating condition variables and low-frequency residual volatility, adjusts the contribution of the clogging drive and long-term offset, and outputs the long-term degradation component representing the degree of filter element degradation accumulation at the current moment based on the monotonic accumulation rule. Based on the long-term degradation component, the monitoring sequence is subjected to degradation correction processing to remove the cumulative effects caused by long-term degradation from the monitoring sequence and generate a corrected monitoring sequence. The adaptive working condition level layer processes the calibration monitoring sequence and introduces an adaptive working condition smoothing mechanism. It selects the corresponding smoothing coefficient according to the working condition state variables, takes the difference between the calibration monitoring sequence and the benchmark level component of the previous time as the input difference, performs input difference limiting processing, determines the input confidence factor according to the residual volatility within the sliding window, and performs weighted calculation on the input after limiting processing to obtain the benchmark level component determined by the current working condition state. The restricted trend layer uses the difference between the baseline level components at the current time and the previous time as the trend processing input, and restricts the trend change amplitude based on the difference between the long-term degradation components at adjacent times. Before executing the trend update, the consistency of the forward and backward directions of the trend processing input is checked, and a trend freezing mechanism is introduced. When the trend direction is frequently reversed, the trend update is frozen according to the operating condition variables. When the freezing condition is not triggered, the trend processing input is processed to obtain the trend components that characterize the long-term change trend in the monitoring sequence. Based on the operating condition state variables, the operating condition segment is determined. The seasonal layer selects seasonal items within the operating condition segment and introduces a periodic structure matching mechanism. For each candidate seasonal item, the periodic fitting residual of the correction monitoring sequence is calculated. The seasonal item with the smallest fitting error is selected according to the matching degree of the periodic fitting residual. Before performing periodic residual processing, a self-check of the seasonal fitting residual is performed on the selected seasonal item. When the seasonal fitting residual exceeds the threshold, the update weight of the seasonal component is reduced. Based on the finally determined seasonal item, periodic residual processing is performed on the correction monitoring sequence and the baseline level component to obtain the periodic component that characterizes the periodic change of the monitoring sequence under the current operating condition. A confidence gating mechanism is introduced to calculate confidence scores for long-term degradation components, baseline components, trend components, and periodic components. The confidence scores are dynamically determined based on the current operating condition variables and historical residual fluctuation levels. Each component is multiplied by its corresponding confidence score as a weighting coefficient, and weighted fusion is performed to obtain the current vehicle filter prediction baseline.

6. The method for detecting abnormalities in automotive filters based on multi-sensor fusion according to claim 1, characterized in that, The step of performing hierarchical processing on the residual sequence to obtain low-frequency residuals and high-frequency residuals includes: Aligning the predicted baseline and monitoring sequence based on the time index of the vehicle filter prediction baseline, the predicted baseline value at each time moment is paired with the monitoring sequence value at the corresponding time moment to form a paired dataset arranged in chronological order; The difference between the monitored sequence values ​​and the predicted baseline values ​​at each time point in the paired dataset is calculated to obtain the residual sequence arranged in chronological order; The residual sequence is processed in layers. Low-frequency components are extracted from the residual sequence using a preset filtering rule to obtain low-frequency residuals. The low-frequency residuals are then subtracted from the residual sequence to obtain high-frequency residuals.

7. The method for detecting abnormalities in automotive filters based on multi-sensor fusion according to claim 1, characterized in that, The method for obtaining a bidirectional CUSUM statistic that includes both positive and negative CUSUM statistics includes: The low-frequency and high-frequency residuals at each time point are paired according to a unified time index to form a residual dataset arranged in chronological order. Statistical features are calculated and standardized for the low-frequency and high-frequency residuals at each time point in the residual dataset to obtain low-frequency standardized residuals and high-frequency standardized residuals. A standardized residual sequence is constructed based on the low-frequency standardized residual and the high-frequency standardized residual. The low-frequency standardized residual and the high-frequency standardized residual at each time step are weighted and summed to form the standardized residual input. The CUSUM algorithm is executed. At the initial moment, the positive and negative CUSUM statistics are initialized to zero. At non-initial moments, the positive and negative CUSUM statistics of the previous moment are used as the accumulation basis for the current moment. Based on the magnitude of the operating condition variables and the standardized residual input, the corresponding variable shrinkage factor is determined for the current moment. The positive and negative CUSUM statistics are shrunk to obtain the intermediate positive and intermediate negative CUSUM statistics adjusted by the variable shrinkage factor. A joint suppression mechanism for operating conditions is introduced. When the operating condition variable indicates a range of drastic changes in operating conditions, the intermediate positive CUSUM statistic and the intermediate negative CUSUM statistic are suppressed. When the operating condition variable indicates a range of stable operating conditions, the CUSUM statistic is updated based on the standardized residual sequence. Under the combined suppression mechanism of variable shrinkage factor and operating condition, the standardized residual input at each time point is updated by positive accumulation and negative accumulation respectively, and the positive CUSUM statistic and negative CUSUM statistic at each time point are obtained. The positive and negative CUSUM statistics at each time point are combined to form a bidirectional CUSUM statistics sequence arranged in chronological order.

8. The method for detecting abnormalities in automotive filters based on multi-sensor fusion according to claim 1, characterized in that, The output anomaly event categories and severity include: Statistical analysis was performed on the positive and negative CUSUM statistics at each time point, and statistical characteristics including the maximum value, average value, and continuous overthreshold length were calculated. Based on the baseline features of statistical characteristics and historical bidirectional CUSUM statistical sequences, adaptive positive thresholds and adaptive negative thresholds are constructed respectively. The positive CUSUM statistic is compared with the adaptive positive threshold, and the negative CUSUM statistic is compared with the adaptive negative threshold. Positive and negative anomalous events are marked according to the threshold exceedance of the positive and negative CUSUM statistic in continuous sampling points. The abnormal events are classified based on the statistical characteristics, threshold amplitude, and duration corresponding to positive and negative abnormal events. The abnormality category and severity of each abnormal event are determined and output.

Images