A deep learning-based automobile part life prediction method

Abstract

The application discloses a kind of based on deep learning's automobile parts life prediction method, comprising: collecting the multi-source heterogeneous data of vehicle end, pre-processing forms multivariate input sequence;Cut multivariate input sequence, construct life learning sample set;Improved PatchTST network is constructed, and output condition representation vector;Diffusion training target is constructed, and constitutes training sample;In prediction stage, initialize future degradation trajectory label;Construct reverse diffusion sampling module, output candidate future degradation trajectory;Life inference is carried out, and maintenance decision result is output.The application realizes automobile parts life prediction result and maintenance decision by combining improved PatchTST network and reverse diffusion sampling module.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a deep learning-based method for predicting the lifespan of automotive parts. Background Technology

[0002] Existing automotive component life prediction technologies mainly include life distribution fitting methods based on reliability statistics, degradation modeling methods based on physical mechanisms, and feature regression methods based on traditional machine learning. Current methods typically rely on artificially constructed degradation features, fixed operating condition assumptions, or specific failure mechanism parameters, making it difficult to maintain stable predictive performance in real-world vehicle operating environments. In actual vehicle scenarios, components operate under variable conditions for extended periods, with frequent changes in load, temperature, speed, and braking behavior, resulting in non-stationary and phased degradation processes. Vehicle data exhibits significant multi-source heterogeneity, including sensor time-series data with different sampling frequencies, low-frequency or event-triggered vehicle operating condition data, and discrete maintenance and replacement event data. Data acquisition often suffers from timestamp asynchrony, inconsistent sampling frequencies, missing observations, noise interference, and outliers, making data alignment, cleaning, and fusion difficult. This hinders the direct formation of structured, reusable life learning samples, leading to unstable degradation state representations, insufficient sample consistency, and training susceptibility to data quality fluctuations.

[0003] In fleet-based applications and cross-vehicle model deployments, individual vehicle differences, component batch differences, and driving behavior variations significantly alter degradation rates and failure modes, making life prediction models difficult to generalize across vehicles and operating conditions. Lifespan labels typically rely on replacement or failure events; however, real-world maintenance systems often involve a mix of maintenance-related and failure-related replacements, delayed or incomplete event records, and varying fault code triggering conditions. This results in sparse and incomplete lifespan labels, with a large number of right-truncated samples, further increasing the difficulty of model learning. Existing deep learning methods often output remaining lifespan through single-point regression or predictive values ​​in a fixed format, making it difficult to provide a reliable expression of uncertainty under conditions of incomplete labels and distribution drift. Furthermore, they struggle to generate lifespan ranges, failure risk curves, and early warning thresholds that can be directly used for maintenance decisions.

[0004] Therefore, how to provide a deep learning-based method for predicting the lifespan of automotive parts 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 deep learning-based method for predicting the lifespan of automotive components. This invention comprehensively utilizes artificial intelligence and data analysis methods, including multi-source heterogeneous vehicle data fusion, event-driven sample construction, improved PatchTST network temporal representation, and diffusion-generative uncertainty prediction. It details the entire process from collecting sensor time-series data, vehicle operating condition data, and maintenance / replacement event data from the vehicle, preprocessing to form a multivariate input sequence, constructing a lifespan learning sample set according to event time, generating conditional representation vectors, constructing and training diffusion training objectives, backsampling to generate candidate future degradation trajectories in the prediction phase, and outputting lifespan inference and maintenance decision results. Structurally, this invention innovatively constructs an improved PatchTST network, introducing non-uniform pulse grouping, sector splitting and recoding, and site-driven bursts to obtain reliable degradation state representations. It also constructs a backdiffusion sampling module, generating reliable uncertainty lifespan prediction results through asymmetric closed-loop sampling and dynamic step-size segmentation of the sampling stream. Compared with existing technologies, this invention has advantages such as strong robustness to noise-induced missing and heterogeneous sampling data, high adaptability across operating conditions and vehicles, ability to output lifespan ranges and failure risks, and ease of engineering deployment and maintenance decision-making.

[0006] A deep learning-based method for predicting the lifespan of automotive components according to an embodiment of the present invention includes: Collect multi-source heterogeneous data from the vehicle end, preprocess the multi-source heterogeneous data, and form a multivariate input sequence; Extract multivariate input sequences from the event time backward according to historical length to construct a lifetime learning sample set; An improved PatchTST network is constructed. The lifetime learning sample set is divided into variable-length grouping units by a non-uniform pulse grouping layer. A sector splitting recoding layer is executed to generate grouping representations. A site-driven burst layer is introduced to insert site markers and participate in sequence encoding, and a conditional representation vector is output. Construct a diffusion training objective, train to obtain future degradation trajectory labels, and combine the future degradation trajectory labels with the conditional representation vector to form training samples; During the prediction phase, the training samples are input into the improved PatchTST network to obtain the current conditional representation vector, and the initial state of the future degenerate trajectory label is initialized with random noise. A backdiffusion sampling module is constructed to sample the current conditional representation vector and the future degradation trajectory label, perform asymmetric closed-loop sampling to update the sampling path, adjust the back iteration step size by dynamically cutting the sampling stream, and output candidate future degradation trajectories. The lifetime of candidate future degradation trajectories is inferred, and the remaining lifetime sample corresponding to each candidate future degradation trajectory is calculated based on the failure criterion. The maintenance decision results are then output.

[0007] Optionally, the multi-source heterogeneous data includes sensor time-series data of the component to be predicted, vehicle operating condition data, and maintenance and replacement event data.

[0008] Optionally, forming a multivariate input sequence includes: Collect multi-source heterogeneous data from the vehicle. The sensor time-series data in the multi-source heterogeneous data includes temperature, current, voltage and speed; the vehicle operating condition data includes vehicle speed, load and braking frequency; and the maintenance and replacement event data includes replacement records and fault code records. Time alignment and resampling fusion are performed on multi-source heterogeneous data. A unified time base is selected, and various types of data are mapped to the same time series according to timestamps. Maintenance and replacement event data are converted into event channel sequences and marked as valid at the time of event occurrence, and invalid at the remaining time. The aligned multi-source heterogeneous data is subjected to outlier removal and missing marker generation. Abnormal observations are identified and set as missing. A missing marker sequence of the same length as the time series is generated for each variable. When an observation exists, the missing marker is set to the non-missing state. When an observation is missing, the missing marker is set to the missing state. The observation sequence, missing marker sequence and event channel sequence are combined in chronological order to form a multivariate input sequence.

[0009] Optionally, constructing the lifetime learning sample set includes: From the maintenance and replacement event data, determine the set of event moments corresponding to the automotive parts to be predicted. The set of event moments includes replacement event moments and failure event moments. Associate each event moment with a multivariate input sequence of the same vehicle and the same part. For each event moment, a continuous time period is extracted backward from the event moment according to the historical length to form a sample input sequence. The sample input sequence is then filtered for validity. The validity filtering includes removing sample input sequences in which the proportion of missing markers in the extracted time period exceeds a certain threshold, and removing sample input sequences in which the event channel markers in the extracted time period are valid. A lifetime label is generated for each sample input sequence. The lifetime label is the time difference between the event time and the end time of the sample input sequence. It is expressed in words as the time interval obtained by subtracting the end time of the sample from the event time. The sample input sequence and the lifetime label are combined to form a lifetime learning sample. All lifetime learning samples are summarized to form a lifetime learning sample set.

[0010] Optionally, the output conditional representation vector includes: An improved PatchTST network was constructed, including a non-uniform pulse grouping layer, a sector splitting recoding layer, and a site-driven burst layer. Pulse sequences were generated based on the lifetime learning sample set. The pulse sequences were generated by the change amplitude of the observations at adjacent time points and the state changes of the missing labels. The non-uniform pulse grouping layer determines the grouping boundary based on the pulse peak position and pulse density change position in the pulse sequence. The sample input sequence is divided into variable-length grouping units according to the grouping boundary, and each variable-length grouping unit is mapped to a grouping input vector. The sector splitting and recoding layer splits each variable-length grouping unit into sector subsequences in chronological order, encodes each sector subsequence to obtain a sector representation, and concatenates and maps the sector representations in sequence to obtain a grouping representation. Construct a grouped representation sequence and perform patching encoding. Arrange all grouped representations in chronological order to form a grouped representation sequence. Perform patching mapping on the grouped representation sequence to obtain a patched sequence. After assigning position index values ​​to the patched sequence, input it into the encoder for sequence encoding to obtain the encoded sequence. The site-driven burst layer determines the set of site time points from maintenance and replacement event data and vehicle operating condition data, maps them to the corresponding positions in the patch sequence, inserts site markers at the corresponding positions, and inputs them into the encoder along with the patch sequence to participate in sequence encoding. The conditional representation vector is extracted and output from the encoded sequence.

[0011] Optionally, the step of composing training samples from future degradation trajectory labels and conditional representation vectors includes: For each sample input sequence in the lifetime learning sample set, the future observation sequence corresponding to the automotive parts to be predicted is extracted within the prediction period, starting from the end of the sample input sequence. The future observation sequences are arranged in chronological order to form future degradation trajectory labels. Set up a diffusion step sequence and a noise intensity sequence, arrange the diffusion step sequence and the noise intensity sequence in a corresponding relationship, and perform stepwise noise addition processing on the future degraded trajectory labels according to the diffusion step sequence to form a noise-added label sequence corresponding to the diffusion step sequence; Construct and train training samples by combining conditional representation vectors with noisy label sequences to form training samples. The input of the training samples includes noisy label states, diffusion step identifiers, and conditional representation vectors. The output of the training samples includes noise prediction values ​​or denoised label prediction values. Iterative training based on the training samples yields future degradation trajectory labels.

[0012] Optionally, obtaining the current conditional representation vector includes: Obtain the multi-source heterogeneous data of the vehicle-side corresponding to the automotive parts to be predicted at the current moment, and obtain the current multivariate input sequence through preprocessing; The current multivariate input sequence is input into the improved PatchTST network, and the non-uniform pulse grouping layer, sector splitting recoding layer and site-driven burst layer are processed sequentially and sequence encoding is performed to output the current conditional representation vector. Determine the predicted length and output dimension of the future degraded trajectory label, generate a random noise sequence consistent with the predicted length and output dimension, and use the random noise sequence as the initial state of the future degraded trajectory label.

[0013] Optionally, the output candidate future degradation trajectory includes: A reverse diffusion sampling module is constructed, which includes asymmetric closed-loop sampling and dynamic step-size cutting sampling stream. It receives the current condition representation vector, the initial state of the future degradation trajectory, the diffusion step sequence and the noise intensity sequence as input to generate a reverse iteration sequence. The intermediate generated trajectory is obtained by performing a denoising update. The initial state of the future degenerate trajectory is then gradually denoised and updated according to the reverse iteration sequence to obtain the intermediate generated trajectory composed of the reverse iteration states. Perform asymmetric closed-loop sampling to update the sampling path, extract the terminal trend summary from the intermediate generated trajectory, inject the terminal trend summary back into the input of the previous reverse iteration state, and perform noise reduction and update on the previous reverse iteration state based on the back-injected input to form the sampling path after closed-loop update, and output the intermediate generated trajectory after closed-loop update. The dynamic step-size cutting sampling stream is executed to adjust the reverse iteration step size. The state change amplitude sequence is calculated for the intermediate generated trajectory after the closed loop update. The step-size cutting point is determined based on the generated state change amplitude sequence, and a variable step-size reverse iteration sequence is generated. The denoising update is completed according to the variable step-size reverse iteration sequence to obtain the final generated trajectory. The final generated trajectory is output as a candidate future degradation trajectory.

[0014] Optionally, the output maintenance decision result includes: The system receives candidate future degradation trajectories and generates a degradation index sequence. It then expands the candidate future degradation trajectories into a future degradation state sequence in chronological order. For each moment in the future degradation state sequence, it generates a degradation index sequence based on the observed values, missing markers, and event channels. The failure time is determined based on the failure criteria and the remaining lifetime sample is obtained. The time when the failure criteria are first met is determined as the failure time, and the remaining lifetime sample is defined as the time interval between the failure time and the start time of the candidate future degradation trajectory. The remaining lifetime samples are statistically analyzed and maintenance decision results are output. The remaining lifetime samples corresponding to the candidate future degradation trajectories are statistically analyzed to obtain the remaining lifetime point estimate, remaining lifetime interval and failure risk curve. Based on the remaining lifetime point estimate, remaining lifetime interval and failure risk curve and the preset maintenance threshold, the maintenance decision results are output.

[0015] The beneficial effects of this invention are: This invention improves the stability of automotive component degradation state characterization and lifetime learning sample construction by fusing multi-source heterogeneous data from the vehicle end and forming a multivariate input sequence containing observations and missing labels. Compared with existing methods that rely on single signals or artificial features and struggle to handle heterogeneous sampling and missing noise data, this invention constructs a lifetime learning sample set based on event moments, significantly improving the consistency and usability of training samples. The improved PatchTST network employs a structured coding approach with non-uniform pulse grouping, sector splitting and recoding, and site-driven bursts, which can effectively extract degradation trends and key site information under conditions of changing operating conditions and fluctuating data quality, thereby reducing prediction bias caused by unreliable degradation characterization.

[0016] This invention achieves distributed output of remaining lifetime and maintenance decision support capabilities by constructing a diffusion training objective and generating candidate future degradation trajectories during the prediction phase. Compared with existing single-point lifetime regression or fixed-form output technologies that struggle to provide reliable uncertainty results, the reverse diffusion sampling module of this invention updates the sampling path through asymmetric closed-loop sampling and adjusts the reverse iteration step size by dynamically cutting the sampling stream. This enables it to output a set of candidate future degradation trajectories even under conditions of significant differences across vehicles and incomplete lifetime labels. Furthermore, it obtains remaining lifetime point estimates, lifetime intervals, and failure risk curves, supporting risk classification, maintenance window determination, and alarm triggering, thereby improving the availability of engineering deployment and the closed-loop effect of operation and maintenance. Attached Figure Description

[0017] 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: Figure 1 This is a flowchart of a deep learning-based method for predicting the lifespan of automotive parts proposed in this invention. Figure 2 This is a block diagram of the improved PatchTST network for a deep learning-based method for predicting the lifespan of automotive parts proposed in this invention. Figure 3 This is a functional diagram of the backdiffusion sampling module of a deep learning-based method for predicting the lifespan of automotive parts proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figure 1 , Figure 2 and Figure 3A deep learning-based method for predicting the lifespan of automotive parts includes: Collect multi-source heterogeneous data from the vehicle end, preprocess the multi-source heterogeneous data, and form a multivariate input sequence; Extract multivariate input sequences from the event time backward according to historical length to construct a lifetime learning sample set; An improved PatchTST network is constructed. The lifetime learning sample set is divided into variable-length grouping units by a non-uniform pulse grouping layer. A sector splitting recoding layer is executed to generate grouping representations. A site-driven burst layer is introduced to insert site markers and participate in sequence encoding, and a conditional representation vector is output. Construct a diffusion training objective, train to obtain future degradation trajectory labels, and combine the future degradation trajectory labels with the conditional representation vector to form training samples; During the prediction phase, the training samples are input into the improved PatchTST network to obtain the current conditional representation vector, and the initial state of the future degenerate trajectory label is initialized with random noise. A backdiffusion sampling module is constructed to sample the current conditional representation vector and the future degradation trajectory label, perform asymmetric closed-loop sampling to update the sampling path, adjust the back iteration step size by dynamically cutting the sampling stream, and output candidate future degradation trajectories. The lifetime of candidate future degradation trajectories is inferred, and the remaining lifetime sample corresponding to each candidate future degradation trajectory is calculated based on the failure criterion. The maintenance decision results are then output.

[0020] In this embodiment, the multi-source heterogeneous data includes sensor time-series data of the component to be predicted, vehicle operating condition data, and maintenance and replacement event data.

[0021] In this embodiment, forming a multivariate input sequence includes: Collect multi-source heterogeneous data from the vehicle. The sensor time-series data in the multi-source heterogeneous data includes temperature, current, voltage and speed; the vehicle operating condition data includes vehicle speed, load and braking frequency; and the maintenance and replacement event data includes replacement records and fault code records. Time alignment and resampling fusion are performed on multi-source heterogeneous data. A unified time base is selected, and various types of data are mapped to the same time series according to timestamps. Maintenance and replacement event data are converted into event channel sequences, with the event occurrence time marked as valid and the remaining time time marked as invalid. Specifically: Repair and replacement event data are aggregated by vehicle number and part number. The event occurrence time of each event record is extracted, and the event occurrence time is first mapped to the discrete time corresponding to a unified time base. The time step of the unified time base is used as the time interval. The time difference is obtained by subtracting the start time of the time series from the event occurrence time. The time difference is then divided by the time interval and rounded down to obtain the discrete time index. An event channel sequence of the same length as the unified time series is established for each vehicle and part and initialized to all zeros. When there is a repair or replacement event corresponding to the discrete time index, the event channel value of the discrete time is set to one, and the remaining discrete times are kept to zero. When there are multiple event records within the same discrete time, the event channel value is still set to one. To avoid alignment errors between event timestamps and the unified time reference boundary, the event occurrence time within one time interval before and after the unified time reference discrete time is considered as the same discrete time. If the time difference between the event occurrence time and the target discrete time does not exceed half a time interval, it is mapped to the discrete time; otherwise, it is mapped to the adjacent discrete time. The aligned multi-source heterogeneous data is subjected to outlier removal and missing marker generation. Abnormal observations are identified and set as missing. A missing marker sequence of the same length as the time series is generated for each variable. When an observation exists, the missing marker is set to the non-missing state. When an observation is missing, the missing marker is set to the missing state. The observation sequence, missing marker sequence and event channel sequence are combined in chronological order to form a multivariate input sequence.

[0022] In this embodiment, constructing the lifetime learning sample set includes: From the maintenance and replacement event data, determine the set of event moments corresponding to the automotive parts to be predicted. The set of event moments includes replacement event moments and failure event moments. Associate each event moment with a multivariate input sequence of the same vehicle and the same part. For each event moment, a continuous time period is extracted backward from the event moment according to historical length to form a sample input sequence. The sample input sequence is then subjected to validity screening. Validity screening includes removing sample input sequences where the proportion of missing labels within the extracted time period exceeds a certain threshold, and removing sample input sequences where the event channel labels within the extracted time period are valid. Specifically, the validity screening of the sample input sequence is as follows: The missing label percentage is calculated by counting the number of elements with missing labels across all time points and variables covered by the sample input sequence. This number is then divided by the product of the number of time points and the number of variables in the sample input sequence to obtain the missing label percentage. If the missing label percentage is greater than 30%, the sample input sequence is deemed invalid and removed. Event channel labels are then checked by counting the number of time points with valid event channel labels within the time period covered by the sample input sequence. If the number of valid labels is greater than zero, the sample input sequence is deemed invalid and removed. This ensures that the time period of the sample input sequence does not contain state reset segments caused by maintenance or replacement events. When the same event time corresponds to multiple capture windows, each window is filtered and the valid sample input sequences are retained to enter the lifetime learning sample set. A lifetime label is generated for each sample input sequence. The lifetime label is the time difference between the event time and the end time of the sample input sequence. It is expressed in words as the time interval obtained by subtracting the end time of the sample from the event time. The sample input sequence and the lifetime label are combined to form a lifetime learning sample. All lifetime learning samples are summarized to form a lifetime learning sample set.

[0023] In this embodiment, the output conditional representation vector includes: An improved PatchTST network is constructed, comprising a non-uniform pulse grouping layer, a sector splitting recoding layer, and a site-driven burst layer. Pulse sequences are generated based on a lifetime learning sample set. These pulse sequences are jointly generated by the amplitude changes of observations at adjacent time points and the state changes of missing markers. The specific details of constructing the improved PatchTST network are as follows: The non-uniform pulse grouping layer calculates the pulse sequence for each multivariate input sequence in the lifetime learning sample set. At the same time, it determines whether the missing label changes at adjacent time points. The amplitude of each variable change is summed at the same time and the missing label change term is added to obtain the pulse value at the current time. The grouping boundary is obtained and the original sequence is divided into variable-length grouping units. The sector splitting and recoding layer divides each variable-length grouping unit into sector subsequences in chronological order. Each sector subsequence is encoded and sequentially spliced ​​and mapped to generate grouping representations, so that each grouping unit is replaced by an intra-segment structured representation. The site-driven burst layer inserts site markers into the grouped representation sequence according to the maintenance and replacement event time and the time of sudden change in operating conditions. The sequence containing the site markers is sent to the PatchTST sequence encoder for encoding, thereby outputting a conditional representation vector containing degradation trend, intra-segment morphology and key site information. The non-uniform pulse grouping layer determines the grouping boundaries based on the pulse peak positions and pulse density change positions in the pulse sequence. The sample input sequence is divided into variable-length grouping units according to these boundaries, and each variable-length grouping unit is mapped to a grouping input vector, where: The non-uniform pulse grouping layer determines the grouping boundaries based on the pulse peak positions and pulse density change positions in the pulse sequence. Specifically, the pulse sequence is standardized, and the mean and standard deviation of the pulse sequence within the sample time period are calculated. The moment when the pulse value is greater than the mean plus twice the standard deviation is defined as the pulse peak position. The pulse density change position is calculated by dividing the number of pulse peaks in the statistical window by the window length by a fixed window sliding on the time axis. The window length is taken as sixty consecutive moments under a unified time base, and the sliding step size is one moment. When the absolute value of the pulse density difference between two adjacent windows is greater than 0.1, the moment when the density change occurs is defined as the pulse density change position. All pulse peak positions and pulse density change positions are merged and sorted by time. If the interval between two adjacent boundaries is less than ten moments under a unified time base, only the earlier boundary is retained, thus obtaining the final set of grouping boundaries. The sample input sequence is divided into variable-length grouping units according to the grouping boundaries. Each variable-length grouping unit is mapped to a grouping input vector. Specifically, the start time of the sample input sequence is used as the first boundary, and the end time of the sample input sequence is used as the last boundary. The continuous time segment between two adjacent boundaries is truncated into a variable-length grouping unit. When mapping each variable-length grouping unit to a grouping input vector, the mean, standard deviation, maximum value, minimum value, and first-to-last difference of the observed values ​​in the group are calculated for each variable in the group. The proportion of the current variable marked as missing in the group is also calculated, and the duration of the grouping unit is calculated. All statistics are concatenated in the order of variables, and the duration of the grouping unit is appended to the end to obtain a fixed-dimensional grouping input vector. The sector splitting and recoding layer splits each variable-length grouping unit into sector subsequences in chronological order. Each sector subsequence is then encoded to obtain a sector representation. These sector representations are then concatenated and mapped sequentially to obtain a grouping representation. Specifically, the encoding of each sector subsequence to obtain the sector representation involves: The observed values ​​of each variable in the sector subsequence are normalized according to the mean and standard deviation within the sector. The missing markers and the normalized observed values ​​are concatenated along the variable dimension to form the sector input. The sector input is padded or truncated to a uniform length. The uniform length is set to twenty consecutive time points under a uniform time base. The padded part of the observed values ​​is set to zero and the missing marker is set to the missing state. The sector input of uniform length is sent to the sector encoder for encoding. The sector encoder is composed of a one-dimensional convolutional layer and a gated recurrent unit connected in series. The length of the one-dimensional convolutional kernel is set to three time points and the stride is set to one time point. The hidden state at the last time point output by the gated recurrent unit is used as the sector representation. A grouped representation sequence is constructed and patched encoding is performed. All grouped representations are arranged in chronological order to form a grouped representation sequence. Patching mapping is performed on the grouped representation sequence to obtain a patched sequence. After assigning position indices to the patched sequence, it is input into the encoder for sequence encoding to obtain the encoded sequence. Specifically, the patched sequence is input into the encoder after assigning position indices to obtain the encoded sequence. Each patch in the patch sequence is assigned a position index in chronological order. The position index is determined by the sequence number of the patch in the sequence. The position index of the first patch is the starting sequence number, and the remaining patches are incremented sequentially. A position vector is generated based on the position index. The dimension of the position vector is the same as that of the patch vector. The position vector is calculated in sine and cosine form. The position index and the vector dimension together determine the sine and cosine values ​​of each dimension. The position vector is added element by element to the corresponding patch vector to obtain a patch vector sequence with position information. The patch vector sequence with position information is input into the encoder for sequence encoding. The encoder is composed of a multi-head self-attention layer and a feedforward network layer stacked together. The hidden states of each patch output by the encoder are arranged in chronological order to form an encoding sequence. When the length of the patch sequence exceeds the maximum length, only the most recent maximum length patch is retained in chronological order. The maximum length is set to 128 patches. The site-driven burst layer determines the set of site time points from maintenance and replacement event data and vehicle operating condition data, maps them to the corresponding positions in the patch sequence, inserts site markers at the corresponding positions, and inputs them into the encoder along with the patch sequence to participate in sequence encoding. The conditional representation vector is extracted and output from the encoded sequence.

[0024] In this embodiment, the step of composing training samples from future degradation trajectory labels and conditional representation vectors includes: For each sample input sequence in the lifetime learning sample set, the future observation sequence corresponding to the automotive parts to be predicted is extracted within the prediction period, starting from the end of the sample input sequence. The future observation sequences are arranged in chronological order to form future degradation trajectory labels. A diffusion step sequence and a noise intensity sequence are defined, arranged in a corresponding relationship. Future degradation trajectory labels are then subjected to progressive noise addition according to the diffusion step sequence, forming a noisy label sequence corresponding to the diffusion step sequence, where: The diffusion step sequence and noise intensity sequence are set as follows: the total number of diffusion steps is set to one hundred steps, the diffusion step sequence is set as an ordered sequence decreasing from the maximum diffusion step to the minimum diffusion step, and the noise intensity sequence is set as an increasing sequence corresponding to the diffusion step sequence. The noise intensity is generated using a linear scheduling method, the initial noise intensity is set to one ten-thousandth, and the final noise intensity is set to two percent. The difference between the noise intensities of two adjacent steps is set as the final noise intensity minus the initial noise intensity and then divided by the total number of diffusion steps minus one, to obtain the noise intensity corresponding to each diffusion step. The cumulative noise intensity is calculated for each diffusion step. The cumulative noise intensity is the product obtained by multiplying the noise intensities from the initial step to the current step step by step, which is used to control the degree of accumulation of gradually added noise. The degraded trajectory labels are subjected to progressive noise addition processing according to the diffusion step sequence. Specifically, the degraded trajectory labels are used as the initial label states. An independent and identically distributed Gaussian random noise sequence with the same dimension as the initial label state is generated. The mean of the Gaussian random noise sequence is zero and the standard deviation is one. The noise-added label states are generated progressively according to the diffusion step sequence. In each diffusion step, the current label state and the Gaussian random noise are linearly combined according to the noise intensity corresponding to the current step to obtain the next label state. The linear combination is calculated by multiplying the current label state by the attenuation coefficient corresponding to the current step and adding the Gaussian random noise by multiplying the noise addition coefficient corresponding to the current step. The attenuation coefficient and the noise addition coefficient are determined by the cumulative noise intensity of the current step. The label states obtained from each diffusion step are saved sequentially according to the diffusion step order to form a noise-added label sequence corresponding to the diffusion step sequence. Construct and train training samples by combining conditional representation vectors with noisy label sequences to form training samples. The input of the training samples includes noisy label states, diffusion step identifiers, and conditional representation vectors. The output of the training samples includes noise prediction values ​​or denoised label prediction values. Iterative training based on the training samples yields future degradation trajectory labels.

[0025] In this embodiment, obtaining the current conditional representation vector includes: Obtain the multi-source heterogeneous data of the vehicle-side corresponding to the automotive parts to be predicted at the current moment, and obtain the current multivariate input sequence through preprocessing; The current multivariate input sequence is input into the improved PatchTST network, and the non-uniform pulse grouping layer, sector splitting recoding layer and site-driven burst layer are processed sequentially and sequence encoding is performed to output the current conditional representation vector. Determine the predicted length and output dimension of the future degraded trajectory label, generate a random noise sequence with the same predicted length and output dimension, and use the random noise sequence as the initial state of the future degraded trajectory label. Specifically, determining the predicted length and output dimension of the future degraded trajectory label involves: The prediction length is defined based on the prediction period used when constructing the life learning sample set. The prediction length is the number of consecutive moments of a unified time reference covered by the last moment of the sample input sequence, which is fixed at three hundred consecutive moments under the unified time reference. The output dimension is determined based on the variable composition of the future degradation trajectory labels. The output dimension is the sum of the number of labeled variables used to characterize the degradation of the parts and the number of missing labeled variables. The labeled variables contain the set of future observation variables used to characterize the degradation, and the missing labeled variables are the missing labeled channels corresponding to each future observation variable. When the number of labeled variables exceeds the upper limit, the top ten labeled variables are retained according to their relevance to the parts, and the missing labeled variables are retained simultaneously, so that the output dimension is twenty.

[0026] In this embodiment, the output candidate future degradation trajectory includes: A back-diffusion sampling module is constructed, comprising asymmetric closed-loop sampling and dynamic step-size segmented sampling streams. It receives the current condition representation vector, the initial state of the future degradation trajectory, the diffusion step sequence, and the noise intensity sequence as input to generate a back-iteration sequence. Specifically, the generation of the back-iteration sequence is as follows: Arrange the diffusion step sequence in order from the largest diffusion step to the smallest diffusion step as the basic step sequence. Match the noise intensity corresponding to the noise intensity in the noise intensity sequence for each step in the basic step sequence. Calculate the absolute value of the noise intensity difference between two adjacent steps as the inter-step variation value. When the inter-step variation value is greater than 0.05%, insert an intermediate step between the two steps. The noise intensity of the intermediate step is the arithmetic mean of the noise intensities of the two steps, and the step identifier of the intermediate step is the midpoint of the step identifiers of the two steps. When the inter-step variation value is not greater than 0.05%, do not insert an intermediate step. Merge the basic step sequence and the inserted intermediate step in order and renumber them to obtain an ordered set of steps containing step identifiers and noise intensities. The ordered set of steps constitutes the reverse iteration sequence. The intermediate generated trajectory is obtained by performing a denoising update. The initial state of the future degenerate trajectory is then gradually denoised and updated according to the reverse iteration sequence to obtain the intermediate generated trajectory composed of the reverse iteration states. Perform asymmetric closed-loop sampling to update the sampling path, extract the terminal trend summary from the intermediate generated trajectory, inject the terminal trend summary back into the input of the previous reverse iteration state, and perform noise reduction and update on the previous reverse iteration state based on the back-injected input to form the sampling path after closed-loop update, and output the intermediate generated trajectory after closed-loop update. The sampling stream is dynamically segmented and the reverse iteration step size is adjusted. A sequence of state change amplitudes is calculated for the intermediate generated trajectory after the closed-loop update. The step size segmentation point is determined based on this sequence, and a variable step size reverse iteration sequence is generated. Denoising and updating are performed according to this sequence to obtain the final generated trajectory. This final generated trajectory is then output as a candidate future degenerate trajectory. The intermediate generated trajectory after the closed-loop update is used to calculate the sequence of generated state change amplitudes. Specifically, the intermediate generated trajectory is represented as a continuous generated state sequence in reverse iteration order. The state increment is obtained by subtracting each element of the adjacent generated state. The absolute value of each element of the state increment is taken and summed over all time positions and all output dimensions. The summation is divided by the number of output dimensions to obtain the normalized change amplitude of two adjacent steps. All normalized change amplitudes of two adjacent steps are arranged in order to form the sequence of generated state change amplitudes. In order to weaken the influence of occasional spikes, the sequence of generated state change amplitudes is smoothed by sliding. The sliding window takes five consecutive steps, and the window output is the median of the change amplitude within the window. Based on the sequence of changes in generated states, the step size cut-off point is determined, and a variable step size reverse iteration sequence is generated. Specifically, the mean and standard deviation of the smoothed sequence are calculated, and the mean plus the standard deviation is used as the cut-off threshold. The step position where the change amplitude is greater than the cut-off threshold is defined as the cut-off point. The reverse iteration sequence is divided into a refined segment and a non-refined segment by the cut-off point. The refined segment retains the original step size, and the non-refined segment performs step size merging. Two adjacent steps are merged into one step, and the noise intensity of the merged step is taken as the arithmetic mean of the noise intensity of the two steps, and the step identifier is taken as the midpoint of the step identifiers of the two steps. The steps of the refined segment and the non-refined segment are combined in chronological order to form a variable step size reverse iteration sequence.

[0027] In this embodiment, the output maintenance decision result includes: The system receives candidate future degradation trajectories and generates a degradation index sequence. It then expands the candidate future degradation trajectories into a future degradation state sequence in chronological order. For each moment in the future degradation state sequence, it generates a degradation index sequence based on the observed values, missing markers, and event channels. The failure time is determined based on the failure criteria, and a remaining lifetime sample is obtained. The moment when the failure criteria are first met is defined as the failure time. The remaining lifetime sample is defined as the time interval between the failure time and the start time of the candidate future degradation trajectory. Specifically, determining the failure time and obtaining the remaining lifetime sample based on the failure criteria involves: Degradation indices are calculated time-by-time for candidate future degradation trajectories. The degradation index is the average value obtained by summing the observed values ​​of each degradation variable in the candidate future degradation trajectory at the same time and then dividing by the number of degradation variables. Failure threshold and mutation threshold are statistically determined in the lifetime learning sample set. The failure threshold is the 95th percentile of the end value of the degradation index of the sample corresponding to the failure event, and the mutation threshold is the 95th percentile of the absolute value of the difference between adjacent time points of the degradation index. The duration is set to ten consecutive moments under a unified time base. Starting from the start time of the candidate future degradation trajectory, the current moment is determined as the failure moment when the degradation index is not less than the failure threshold, or the absolute value of the difference between adjacent moments of the degradation index is not less than the mutation threshold, or the number of consecutive moments of the degradation index not less than the failure threshold reaches the duration. The remaining lifetime sample is calculated as the time interval between the failure moment and the start time of the candidate future degradation trajectory. The time interval is the index difference between the two moments on the unified time base multiplied by the time step of the unified time base. The remaining lifetime samples are statistically analyzed, and maintenance decision results are output. The remaining lifetime samples corresponding to candidate future degradation trajectories are statistically analyzed to obtain remaining lifetime point estimates, remaining lifetime intervals, and failure risk curves. Maintenance decision results are output based on the remaining lifetime point estimates, remaining lifetime intervals, failure risk curves, and preset maintenance thresholds. Specifically, the output of maintenance decision results based on the remaining lifetime point estimates, remaining lifetime intervals, failure risk curves, and preset maintenance thresholds is as follows: The remaining lifetime samples are sorted and the median is taken as the remaining lifetime point estimate. The fifth percentile and the ninety-fifth percentile are taken as the lower and upper bounds of the remaining lifetime interval. The cumulative failure probability is calculated time by time during the future assessment period. The cumulative failure probability is defined as the number of remaining lifetime samples that are not greater than the time interval corresponding to the current time divided by the total number of remaining lifetime samples. The cumulative failure probability is arranged over time to form a failure risk curve. The preset maintenance thresholds include a risk threshold and a time threshold. The risk threshold is set to 20% and the time threshold is set to the time interval corresponding to seven days. An emergency maintenance decision is output when the cumulative failure probability of the failure risk curve at the time threshold is not less than the risk threshold; a planned maintenance decision is output when the cumulative failure probability of the failure risk curve at the time threshold is less than the risk threshold and the lower bound of the remaining lifespan interval is not greater than the time threshold; and a continue operation decision is output when the cumulative failure probability of the failure risk curve at the time threshold is less than the risk threshold and the lower bound of the remaining lifespan interval is greater than the time threshold.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a large-scale fleet maintenance scenario. The components to be predicted operate under variable conditions for extended periods. Sensor data comes from different controllers and buses, resulting in inconsistent sampling frequencies, short-term data interruptions, noise impacts, and abnormal drift. Furthermore, maintenance and replacement records and fault code records are incomplete and delayed, leading to sparse lifespan labels and the inclusion of right-truncated samples. Traditional methods typically rely on single signal thresholds or human experience features. When faced with differences in driving behavior, load variations, and environmental changes, the degradation representation is easily distorted, and training samples are difficult to construct stably. Ultimately, this manifests as large prediction deviations for the same component on different vehicles, inconsistent alarm timing, and difficulty in outputting lifespan ranges and risk results that can be used for maintenance decisions. This fails to meet the fleet's requirements for fewer false alarms, interpretability, and decision-making support.

[0029] The system integrates time-series data from vehicle sensors, vehicle operating condition data, and maintenance / replacement event data. First, it maps all data types to a unified time base and generates missing label channels. Simultaneously, it converts maintenance / replacement event data into event channel sequences to mark state reset points. A fixed historical length is truncated backward from the event time to construct a lifetime learning sample set. This data is then fed into an improved PatchTST network. Pulse sequences are generated based on the magnitude of changes in observed values ​​and changes in missing states. These sequences are then non-uniformly grouped, and each variable-length group unit is split into sectors and re-encoded to obtain intra-segment structural representations. Maintenance / replacement event times and abrupt changes in operating conditions are mapped to site-specific labels inserted into the encoding sequence, outputting a conditional representation vector. During the training phase, future degradation trajectory labels are generated based on future observations and noise is gradually added to learn denoising prediction. In the prediction phase, the initial state of the future degradation trajectory is initialized with random noise. A closed-loop sampling path update is performed in the backdiffusion sampling module. The backdiffusion step size is cut and adjusted based on the magnitude of generated state changes, outputting multiple candidate future degradation trajectories. This yields the remaining lifetime sample set, lifetime interval, and failure risk curve for maintenance decision-making.

[0030] During deployment and validation, after applying this method, event channels and missing markers are explicitly incorporated into the model input. The sample construction process removes high-missing windows and windows containing event reset fragments, significantly improving the consistency of training samples. The degradation representation of the improved PatchTST output is more stable under cross-vehicle conditions, and the candidate future degradation trajectory can provide a lifetime distribution rather than a single point result. The maintenance side can set alarm thresholds based on the risk curve and output maintenance suggestions of different levels.

[0031] Table 1. Comprehensive Comparison Index of Vehicle Life Prediction Methods

[0032] As can be seen from Table 1, in terms of lifetime prediction accuracy, the present invention has the lowest mean absolute error and root mean square error, which are 8.7h and 12.5h respectively. This is significantly better than the threshold method, the two types of tree models and the time series deep model. This indicates that under the conditions of vehicle noise, missing data and heterosampling, the present invention has a more stable characterization of the degradation state and the lifetime point estimation is closer to the true remaining lifetime.

[0033] In terms of uncertainty and risk discrimination capabilities, the coverage of this invention reaches 91%, while the interval width is 18.9h, which shows that the interval does not expand excessively while maintaining high coverage; the risk AUC reaches 92%, which is better than the comparison method, indicating that the failure risk curve based on the candidate future degradation trajectory has a stronger ability to distinguish high-risk samples and can more reliably support maintenance decisions.

[0034] From the perspective of operational availability, the lead time of this invention is 9.1 hours, which is higher than that of the comparison method, meaning that effective early warnings can be given earlier while meeting accuracy requirements. The false alarm rate and false negative rate are 7% and 9% respectively, both of which are the lowest or close to the lowest. Compared with the high false alarm and false negative performance of the threshold method, this invention is more conducive to reducing unnecessary maintenance and reducing the risk of missed detection in actual fleet applications, thus achieving more robust closed-loop operation and maintenance.

[0035] 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 deep learning-based method for predicting the lifespan of automotive parts, characterized in that, include: Collect multi-source heterogeneous data from the vehicle end, preprocess the multi-source heterogeneous data, and form a multivariate input sequence; Extract multivariate input sequences from the event time backward according to historical length to construct a lifetime learning sample set; An improved PatchTST network is constructed. The lifetime learning sample set is divided into variable-length grouping units by a non-uniform pulse grouping layer. A sector splitting recoding layer is executed to generate grouping representations. A site-driven burst layer is introduced to insert site markers and participate in sequence encoding, and a conditional representation vector is output. Construct a diffusion training objective, train to obtain future degradation trajectory labels, and combine the future degradation trajectory labels with the conditional representation vector to form training samples; During the prediction phase, the training samples are input into the improved PatchTST network to obtain the current conditional representation vector, and the initial state of the future degenerate trajectory label is initialized with random noise. A backdiffusion sampling module is constructed to sample the current conditional representation vector and the future degradation trajectory label, perform asymmetric closed-loop sampling to update the sampling path, adjust the back iteration step size by dynamically cutting the sampling stream, and output candidate future degradation trajectories. The lifetime of candidate future degradation trajectories is inferred, and the remaining lifetime sample corresponding to each candidate future degradation trajectory is calculated based on the failure criterion. The maintenance decision results are then output.

2. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The multi-source heterogeneous data includes sensor time-series data of the components to be predicted, vehicle operating condition data, and maintenance and replacement event data.

3. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The formation of the multivariate input sequence includes: Collect multi-source heterogeneous data from the vehicle. The sensor time-series data in the multi-source heterogeneous data includes temperature, current, voltage and speed; the vehicle operating condition data includes vehicle speed, load and braking frequency; and the maintenance and replacement event data includes replacement records and fault code records. Time alignment and resampling fusion are performed on multi-source heterogeneous data. A unified time base is selected, and various types of data are mapped to the same time series according to timestamps. Maintenance and replacement event data are converted into event channel sequences and marked as valid at the time of event occurrence, and invalid at the remaining time. The aligned multi-source heterogeneous data is subjected to outlier removal and missing marker generation. Abnormal observations are identified and set as missing. A missing marker sequence of the same length as the time series is generated for each variable. When an observation exists, the missing marker is set to the non-missing state. When an observation is missing, the missing marker is set to the missing state. The observation sequence, missing marker sequence and event channel sequence are combined in chronological order to form a multivariate input sequence.

4. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The construction of the lifetime learning sample set includes: From the maintenance and replacement event data, determine the set of event moments corresponding to the automotive parts to be predicted. The set of event moments includes replacement event moments and failure event moments. Associate each event moment with a multivariate input sequence of the same vehicle and the same part. For each event moment, a continuous time period is extracted backward from the event moment according to the historical length to form a sample input sequence. The sample input sequence is then filtered for validity. The validity filtering includes removing sample input sequences in which the proportion of missing markers in the extracted time period exceeds a certain threshold, and removing sample input sequences in which the event channel markers in the extracted time period are valid. A lifetime label is generated for each sample input sequence. The lifetime label is the time difference between the event time and the end time of the sample input sequence. It is expressed in words as the time interval obtained by subtracting the end time of the sample from the event time. The sample input sequence and the lifetime label are combined to form a lifetime learning sample. All lifetime learning samples are summarized to form a lifetime learning sample set.

5. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The output conditional representation vector includes: An improved PatchTST network was constructed, including a non-uniform pulse grouping layer, a sector splitting recoding layer, and a site-driven burst layer. Pulse sequences were generated based on the lifetime learning sample set. The pulse sequences were generated by the change amplitude of the observations at adjacent time points and the state changes of the missing labels. The non-uniform pulse grouping layer determines the grouping boundary based on the pulse peak position and pulse density change position in the pulse sequence. The sample input sequence is divided into variable-length grouping units according to the grouping boundary, and each variable-length grouping unit is mapped to a grouping input vector. The sector splitting and recoding layer splits each variable-length grouping unit into sector subsequences in chronological order, encodes each sector subsequence to obtain a sector representation, and concatenates and maps the sector representations in sequence to obtain a grouping representation. Construct a grouped representation sequence and perform patching encoding. Arrange all grouped representations in chronological order to form a grouped representation sequence. Perform patching mapping on the grouped representation sequence to obtain a patched sequence. After assigning position index values ​​to the patched sequence, input it into the encoder for sequence encoding to obtain the encoded sequence. The site-driven burst layer determines the set of site time points from maintenance and replacement event data and vehicle operating condition data, maps them to the corresponding positions in the patch sequence, inserts site markers at the corresponding positions, and inputs them into the encoder along with the patch sequence to participate in sequence encoding. The conditional representation vector is extracted and output from the encoded sequence.

6. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The step of combining future degradation trajectory labels and conditional representation vectors to form training samples includes: For each sample input sequence in the lifetime learning sample set, the future observation sequence corresponding to the automotive parts to be predicted is extracted within the prediction period, starting from the end time of the sample input sequence. The future observation sequences are arranged in chronological order to form future degradation trajectory labels. Set up a diffusion step sequence and a noise intensity sequence, arrange the diffusion step sequence and the noise intensity sequence in a corresponding relationship, and perform stepwise noise addition processing on the future degraded trajectory labels according to the diffusion step sequence to form a noise-added label sequence corresponding to the diffusion step sequence; Construct and train training samples by combining conditional representation vectors with noisy label sequences to form training samples. The input of the training samples includes noisy label states, diffusion step identifiers, and conditional representation vectors. The output of the training samples includes noise prediction values ​​or denoised label prediction values. Iterative training based on the training samples yields future degradation trajectory labels.

7. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The process of obtaining the current conditional representation vector includes: Obtain the multi-source heterogeneous data of the vehicle-side corresponding to the automotive parts to be predicted at the current moment, and obtain the current multivariate input sequence through preprocessing; The current multivariate input sequence is input into the improved PatchTST network, and the non-uniform pulse grouping layer, sector splitting recoding layer and site-driven burst layer are processed sequentially and sequence encoding is performed to output the current conditional representation vector. Determine the predicted length and output dimension of the future degraded trajectory label, generate a random noise sequence consistent with the predicted length and output dimension, and use the random noise sequence as the initial state of the future degraded trajectory label.

8. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The output candidate future degradation trajectory includes: A reverse diffusion sampling module is constructed, which includes asymmetric closed-loop sampling and dynamic step-size cutting sampling stream. It receives the current condition representation vector, the initial state of the future degradation trajectory, the diffusion step sequence and the noise intensity sequence as input to generate a reverse iteration sequence. Perform denoising update to obtain intermediate generated trajectory, and perform stepwise denoising update on the initial state of future degenerate trajectory according to the reverse iteration sequence to obtain intermediate generated trajectory composed of reverse iteration state; Perform asymmetric closed-loop sampling to update the sampling path, extract the terminal trend summary from the intermediate generated trajectory, inject the terminal trend summary back into the input of the previous reverse iteration state, and perform noise reduction and update on the previous reverse iteration state based on the back-injected input to form the sampling path after closed-loop update, and output the intermediate generated trajectory after closed-loop update. The dynamic step-size cutting sampling stream is executed to adjust the reverse iteration step size. The state change amplitude sequence is calculated for the intermediate generated trajectory after the closed loop update. The step-size cutting point is determined based on the generated state change amplitude sequence, and a variable step-size reverse iteration sequence is generated. The denoising update is completed according to the variable step-size reverse iteration sequence to obtain the final generated trajectory. The final generated trajectory is output as a candidate future degradation trajectory.

9. The method for predicting the lifespan of automotive parts based on deep learning according to claim 1, characterized in that, The output maintenance decision results include: The system receives candidate future degradation trajectories and generates a degradation index sequence. It then expands the candidate future degradation trajectories into a future degradation state sequence in chronological order. For each moment in the future degradation state sequence, it generates a degradation index sequence based on the observed values, missing markers, and event channels. The failure time is determined based on the failure criteria and the remaining lifetime sample is obtained. The time when the failure criteria are first met is determined as the failure time, and the remaining lifetime sample is defined as the time interval between the failure time and the start time of the candidate future degradation trajectory. The remaining lifetime samples are statistically analyzed and maintenance decision results are output. The remaining lifetime samples corresponding to the candidate future degradation trajectories are statistically analyzed to obtain the remaining lifetime point estimate, remaining lifetime interval and failure risk curve. Based on the remaining lifetime point estimate, remaining lifetime interval and failure risk curve and the preset maintenance threshold, the maintenance decision results are output.

Images