Large-scale vehicle sparse fault prediction method based on semantic self-adaptive label transfer

Abstract

This invention provides a large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer, belonging to the field of vehicle fault prediction. The method includes acquiring vehicle repair samples, fault label sequences, and corresponding textual descriptions of each fault label, and constructing historical vehicle repair sequence data. Based on this historical repair sequence data, an empirical transfer prior is constructed by statistically analyzing the temporal decay co-occurrence intensity of fault labels, and a semantic transfer prior is constructed using a semantic transfer projector. Based on the historical vehicle repair sequence data, empirical transfer prior, and semantic transfer prior, an enhanced temporal representation is obtained through sequence encoding and prior injection. Based on this enhanced temporal representation, features are extracted from both temporal and contextual perspectives, and residual fusion is performed to obtain multi-label prediction probabilities. This invention addresses the problem of insufficient empirical transfer coverage in existing technologies for large-scale vehicle fault label scenarios, which leads to inaccurate prediction of some vehicle faults.

Description

Technical Field

[0001] This invention belongs to the field of vehicle fault prediction, and particularly relates to a large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer. Background Technology

[0002] Vehicle maintenance failure prediction aims to predict future failure events based on historical maintenance records, playing a crucial role in predictive maintenance and vehicle safety management. Accurate failure prediction helps maintenance systems identify potential risks in advance, reduce unexpected failures and repair costs, and improve operational reliability. Unlike traditional fault diagnosis tasks that focus on specific components or single fault types, real-world vehicle maintenance records typically exhibit time dependence and fault co-occurrence; a single vehicle sample may be associated with multiple future fault labels.

[0003] Existing vehicle fault prediction methods mainly follow two technical routes. Sequence modeling-based methods use recurrent neural networks, attention mechanisms, or time encoders to capture the temporal dependencies of fault evolution from historical maintenance records; graph structure modeling-based methods construct fault label relationship graphs through statistical co-occurrence matrices or transition probabilities, and use graph neural networks for label dependency propagation and inference.

[0004] However, in large-scale vehicle repair scenarios, the fault label space can expand to thousands, and fault distribution generally exhibits long-tail characteristics, with a large number of low-frequency labels having extremely limited historical samples. Sequence-based modeling methods rely solely on single historical sequences for modeling, lacking explicit prior guidance on label relationships; graph-based methods heavily depend on co-occurrence or transition statistics already present in the training data. When transition relationships between a large number of low-frequency labels are rarely or never present in the data, the coverage of the empirical label graph is severely limited, failing to utilize the rich semantic information contained in the fault label text to supplement sparse observation relationships. Therefore, effective and accurate predictions are difficult to achieve when facing large-scale sparse fault prediction. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer. This method solves the problem that existing large-scale vehicle fault prediction technologies suffer from insufficient experience transfer coverage in scenarios with large-scale vehicle fault labels, leading to difficulties in accurately predicting some vehicle faults.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer, characterized by comprising the following steps: Obtain vehicle repair samples, fault label sequences, and text descriptions of each fault label, and construct historical vehicle repair sequence data; Based on historical vehicle maintenance sequence data, an experience transfer prior is constructed by statistically analyzing the time decay co-occurrence intensity of fault labels, and a semantic transfer prior is constructed by using a semantic transfer projector. Based on vehicle historical maintenance sequence data, experience transfer priors, and semantic transfer priors, an enhanced temporal representation is obtained through sequence encoding and prior injection processing. Based on the enhanced temporal representation, features are extracted from both temporal and contextual perspectives and residual fusion is performed to obtain the multi-label prediction probability.

[0007] Furthermore: the construction of the experience transfer prior specifically includes: Iterate through all time steps of the vehicle's historical maintenance sequence data; For each fault label in the vehicle's historical maintenance sequence data at the current time step, use that fault label as the source fault label and other fault labels other than the original fault label as the target fault label. Within the maximum future observation time window defined after the source fault label, count the number of occurrences of each target fault label; Based on the frequency of occurrence of each target fault label, a decay weight allocation is performed according to the time interval from the current time step to obtain the weight of each fault label; the decay weight allocation is specifically as follows: the farther away from the current time step, the smaller the weight. For each label pair consisting of a source fault label and a target fault label, the corresponding decay-weighted co-occurrence in the training set is accumulated to obtain the directional time decay co-occurrence intensity from the source fault label to the target fault label. Based on the frequency of occurrence of each source fault label in the training set, the corresponding directional time decay co-occurrence intensity is normalized to obtain the normalized transfer intensity. For each source fault label, retain the target fault labels whose normalized transfer strength is greater than the preset transfer strength threshold and which are ranked in the top preset number according to the normalized transfer strength. Based on the retained source fault labels, target fault labels and corresponding normalized transfer strengths, construct an empirical transfer prior.

[0008] Furthermore, the expression for the time-decay co-occurrence intensity is as follows:

[0009] in, To remove fault labels To the fault label The directional time decay co-occurrence intensity, fault label For source fault labels, fault labels For the target fault label, and These are all indexes of fault labels. This is a vehicle's historical maintenance sequence. For the historical maintenance sequence of all vehicles, For time steps, To step from the current time The offset step for counting backwards, ranging from 1 to... , For the maximum future observation time window, This is the time decay factor, and its value range is... , For sequence Mid-time step Are there any fault labels? The indicated value, For sequence Mid-time step Are there any fault labels? The indicated value; The expression for the prior knowledge transfer is as follows:

[0010]

[0011]

[0012] in, To remove fault labels To the fault label Experience transfer prior, To normalize the transfer intensity, Fault Label Total frequency of occurrence Fault Label The set of retained fault labels To preset the transfer intensity threshold, To select the top with the highest normalized transfer strength Fault label.

[0013] The further beneficial effects mentioned above are as follows: the present invention assigns higher weight to recent co-occurrences through time decay weighting, accurately capturing the timeliness characteristics of fault evolution; eliminates statistical bias of high-frequency labels through source frequency normalization; and filters out weak correlations and noise edges through thresholding and Top-K dual sparsification, so that the empirical transfer prior has good computational efficiency while maintaining high reliability.

[0014] Furthermore: the construction of semantic transfer priors specifically includes: Obtain text description information corresponding to each fault tag in the vehicle's historical maintenance sequence data. The text description information includes fault codes and component names. Based on the text description information of each fault label, construct a single-label text for each fault label; Based on the single-label text of each fault label, input them into the frozen large language model and output the hidden state of each fault label. Masked average pooling and norm normalization are performed on the effective words in each hidden state to obtain the fixed semantic vector of each fault label. Based on the fixed semantic vector of each fault label, combined with the empirical successor label of the empirical transfer prior, the nearest neighbor label in the semantic space, and the randomly sampled exploration label, a set of candidate fault labels is constructed for each fault label. For each fault label and each fault label in its candidate fault label set, the fixed semantic vectors of the two are combined with features and then fed into a lightweight projector to output a semantic transfer score. Using empirical transfer priors as supervision signals, a lightweight semantic transfer projector is trained. Based on the existence of pseudo-labels on edges and empirical transfer strength derived from empirical transfer priors, the lightweight semantic transfer projector is trained by minimizing the projector loss function to obtain a trained lightweight semantic transfer projector. The semantic transfer scores of all candidate label pairs output by the trained lightweight semantic transfer projector are sparsified. Candidate label pairs with scores greater than the preset semantic transfer score threshold are retained, and the top few faulty labels with the highest semantic transfer scores are retained to obtain the sparsified semantic transfer prior.

[0015] Furthermore, the expression for the semantic transfer score is as follows:

[0016]

[0017]

[0018]

[0019]

[0020] in, To remove fault labels To the fault label semantic transfer score, It is the sigmoid activation function. It is a multilayer perceptron network. and All labels are represented after bottleneck transformation. Fault Label Fixed semantic vectors, Fault Label Fixed semantic vectors, and These are all indexes of fault labels. This is a vector concatenation operation. For element-wise multiplication, The first in the candidate fault label set One fault label, For the set of candidate fault labels, All of these are fault labels from the candidate fault label set. The superscript represents the number of fault labels in the candidate fault label set. It is the transpose symbol. For the real number field, For fixed semantic vector dimensions, For fault labels The unnormalized semantic vector is obtained by masking and average pooling the corresponding effective word hidden states. for norm, This is the mask indicator value, which can be either 0 or 1. For the index of lexical elements, The first hidden state Each word element, It is in a hidden state. Indicates the first Each word element is a valid word element. Indicates the first Each word element is an invalid word element; The expression for the semantic transfer prior is as follows:

[0021]

[0022] in, To remove fault labels To the fault label semantic transfer prior, For the set of candidate fault labels, To preset the semantic transfer score threshold, Semantic transfer score in fault label The descending ranking of all semantic transfer scores. To select the top semantic transfer score One fault label, For the transfer of prior experience to subsequent labels, The nearest neighbor label in the semantic space. To explore labels for random sampling.

[0023] Furthermore, the expression for the loss function of the lightweight semantic transfer projector is as follows:

[0024] in, For the loss function of a lightweight semantic transfer projector, The cross-entropy loss function is used for binary classification. Let the mean squared error loss function be . The logit value output by the lightweight semantic transfer projector without sigmoid activation. To transfer prior knowledge from experience The exported edges contain pseudo-labels. The weighting coefficients are used to balance the cross-entropy loss and the mean squared error loss in binary classification. To remove fault labels To the fault label semantic transfer score, To remove fault labels To the fault label The prior knowledge transfer.

[0025] The further beneficial effects mentioned above are as follows: This invention utilizes a frozen large language model to extract the hidden states of each fault label. Compared with shallow text encoding, the hidden states of the large model can provide richer semantic representations for short and terminology-dense vehicle fault labels. The candidate set integrates three types of labels: empirical successor, semantic nearest neighbor, and random exploration, focusing semantic modeling on label pairs that are more likely to contain transition relationships, while also taking into account the discovery of potential new relationships. Empirical prior supervision and calibration align the semantic scores with historical maintenance patterns, rather than directly using LLM to generate numerical scores, avoiding the quadratic computational overhead and task mismatch problem of large model pairwise inference. Thresholding and Top-K dual sparsification ensure the efficiency and sparsity of semantic priors.

[0026] Further: the obtained enhanced timing representation specifically includes: Input the vehicle's historical maintenance sequence data into the sequence encoder and output the hidden state of each time step for each maintenance sample; The hidden states at each time step are projected onto the label space through a linear mapping to obtain the label space activation vector; Multiply the label space activation vector by the empirical transfer prior and the semantic transfer prior respectively to obtain the activation vector after empirical propagation and the activation vector after semantic propagation. Based on the activation vectors after empirical propagation and semantic propagation, the empirical enhancement amount and semantic enhancement amount are obtained by projecting them back into the latent space through linear mapping. The hidden state of the last valid time step of each maintenance sample in the vehicle history maintenance sequence data is input into a linear layer with sigmoid activation, and the sample-level gating coefficient of each maintenance sample is output. Based on the sample-level gating coefficients, the empirical enhancement and semantic enhancement are weighted and fused, and then multiplied by the learnable overall scaling coefficient and superimposed with the hidden states of each time step to obtain the sequence of hidden states for transition enhancement, which serves as the temporal representation of the enhancement.

[0027] Furthermore, the expression for the enhanced timing representation is as follows:

[0028]

[0029]

[0030]

[0031] in, For the first One repair sample at time step Enhanced timing representation, For the first One repair sample at time step The hidden state, The learnable overall scaling factor. For the first Sample-level gating coefficients for each maintenance sample. For the first One repair sample at time step The amount of experience enhancement, For the first One repair sample at time step The amount of semantic enhancement, It is the sigmoid activation function. This is the weight matrix for the gating coefficients. For the first The hidden state of the last valid time step of a maintenance sample This is the bias term for the gating coefficient. The weight matrix for projecting empirical enhancements back into the latent space. Let be the weight matrix mapping the hidden states to the label space. For the transfer of prior knowledge, The weight matrix for projecting semantic enhancements back into the latent space. Let be the weight matrix mapping the hidden states to the label space. This serves as a semantic transfer prior.

[0032] The further beneficial effects mentioned above are as follows: By mapping to the label space, propagating experience, and then back mapping to the latent space, this invention injects the experience transfer prior and semantic transfer prior into the latent state of each time step of the sequence encoder. Compared with existing methods that only introduce label relationships at the classifier level, this invention enables prior knowledge to deeply participate in the construction process of sequence representation, which can improve the prediction ability and prediction accuracy of sparse vehicle faults. The sample-level gating coefficient adaptively adjusts the fusion ratio of the two priors according to the sequence characteristics of each maintenance sample, so that frequent label samples can focus on reliable experience priors, and sparse label samples can obtain relationship completion with the help of semantic priors. This overcomes the problem that fixed ratio fusion is not adaptable to maintenance samples with different degrees of sparseness in vehicles, and improves the generalization ability for different fault frequency distributions. The residual superposition protects the main information of the original sequence encoding from being destroyed while introducing external prior knowledge, realizing the synergy of prior guidance and data-driven approach, so that the enhanced temporal representation has both sequence context information and label transfer structure knowledge.

[0033] Furthermore: obtaining the multi-label prediction probability specifically includes: Based on the enhanced temporal representation, average pooling is performed on multiple time windows through the temporal stream, and the pooled window representations of each window are weighted, summed, and then linearly mapped to obtain the temporal features; Based on the augmented temporal representation, the augmented temporal representation is averaged along the time dimension through the context stream to obtain the global sample representation. The global sample representation is used as the query and the augmented temporal representation is used as the key and value for cross-attention calculation to obtain the context features. Based on temporal and contextual features, a basic interaction representation is constructed, and the temporal and contextual features are adaptively fused through a routing mechanism to obtain a basic fused representation. Based on temporal and contextual features, residual interaction terms are constructed and superimposed with the basic fusion representation to obtain sample-level representations; Based on the sample-level representation, the data is mapped to the label space through a linear classification layer, and the independent prediction probability of each fault label is output through the sigmoid activation function to obtain the multi-label prediction probability.

[0034] Furthermore: the expression for obtaining the multi-label prediction probability is as follows:

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] in, For multi-label probability prediction, It is the sigmoid activation function. The weight matrix of the linear classification layer. For sample-level representation, For the bias term of the linear classification layer, Based on the fusion representation, The learnable residual scaling factor has a value range of [value range missing]. , For residual interaction terms, These are learnable interpolation coefficients, with a value range of [value range missing]. , Basic interactive representation, According to routing experts, Multilayer perceptron networks based on interactive representations, As a time series feature, For contextual features, This is a vector concatenation operation. For element-wise multiplication, The linear mapping matrix for time series features, , and All are learnable weights. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of a repair sample within a short time window. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of each maintenance sample within a mid-time window. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of each maintenance sample within the full time window. For cross-attention calculation, For global sample representation, To enhance the timing representation, For average pooling, For the first The repair sample corresponds to the first Routing weights of an expert network For the first An expert network with independent learnable parameters. For the first Temporal characteristics of a repair sample The shared routing space representation obtained after learnable projection. For the first Contextual features of a repair sample The shared routing space representation obtained after learnable projection. For the first The routing weight vector corresponding to each maintenance sample For indexing expert networks, For the number of expert networks, Here is the weight matrix of the router. This is the router's bias term. For the first Encoding of the effective sequence length of a repair sample The learnable projection matrix corresponding to the temporal features. The bias term corresponding to the time series features. The learnable projection matrix corresponding to the context features. The bias term corresponding to the context features. For sequence length encoding function, For the first The number of valid time steps for each repair sample.

[0044] The further beneficial effects mentioned above are as follows: This invention captures multi-scale evolution patterns from fine-grained recent changes to long-term historical patterns across multiple time windows through temporal flow; it uses context flow to extract historical information related to the overall state of the current sample through cross-attention adaptive extraction of global representation as a query, thus solving the problem that a single aggregation path cannot take into account both dynamic evolution and global context; it achieves adaptive feature fusion of samples through routing mechanism, enabling maintenance samples with different characteristics to obtain personalized fusion strategies; residual interaction further supplements fine-grained information such as differences and consistency between temporal and contextual information, improving the discriminative power of multi-label prediction representation and making the prediction of sparse faults of large-scale vehicles more accurate.

[0045] The beneficial effects of this invention are: This invention addresses the problem that existing methods, which rely solely on observed co-occurrence or statistics in training data, cannot utilize reasonable transfer relationships between low-frequency fault labels when the fault label space is large and long-tailed in large-scale vehicle maintenance scenarios, by constructing a dual-priority collaborative mechanism of empirical transfer prior and semantic transfer prior. This improves the overall vehicle fault prediction performance and the prediction ability of low-frequency fault labels. In the construction of the experience transfer prior, this invention uses time decay weighting to give higher weight to recent fault labels, accurately depicting the timeliness of the evolution of actual vehicle faults. It also eliminates the statistical dominance effect of high-frequency labels through frequency normalization, avoiding large errors caused by excessive frequency deviations of different fault labels. Furthermore, it improves the reliability and computational efficiency of the experience transfer prior by screening high-confidence experience transfer priors through thresholding and TopK sparsification.

[0046] This invention introduces a semantic transfer prior based on a frozen large language model, which encodes each fault label only once. It uses a lightweight semantic transfer projector with empirical transfer prior as supervision for calibration, avoiding the quadratic computational overhead of pairwise inference in large language models. This achieves low-cost, reproducible, and task-aligned semantic transfer prior construction, providing supplementary information for long-tailed fault labels that cannot be covered by empirical transfer prior.

[0047] This invention injects experience transfer priors and semantic transfer priors into the hidden state of each time step of the sequence encoder, enabling prior knowledge to deeply participate in the construction process of sequence representation; the sample-level gating coefficients adaptively adjust the dual prior fusion ratio according to the sequence characteristics of each maintenance sample, so that samples with different sparsity levels receive different prior guidance; the residual superposition design protects the main information of the original sequence encoding while introducing external priors, realizing the synergy between prior knowledge and data-driven approaches in large-scale sparse scenarios. This invention extracts global maintenance information from both temporal and contextual perspectives in the prediction head, overcoming the problem that a single path cannot take into account both dynamic evolution and global context. The routing mechanism enables adaptive feature fusion of samples, and residual interactive refinement further supplements the fine-grained information on the differences and consistency between temporal and contextual information, making the prediction representation both temporally sensitive and contextually global, thus improving the prediction accuracy in large-scale sparse fault label scenarios. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the process for a large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer. Detailed Implementation

[0049] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0050] Example 1 like Figure 1The diagram shows a flowchart of a large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer. This invention provides a large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer, comprising the following steps: Obtain vehicle repair samples, fault label sequences, and text descriptions of each fault label, and construct historical vehicle repair sequence data; Based on historical vehicle maintenance sequence data, an experience transfer prior is constructed by statistically analyzing the time decay co-occurrence intensity of fault labels, and a semantic transfer prior is constructed by using a semantic transfer projector. Based on vehicle historical maintenance sequence data, experience transfer priors, and semantic transfer priors, an enhanced temporal representation is obtained through sequence encoding and prior injection processing. Based on the enhanced temporal representation, features are extracted from both temporal and contextual perspectives and residual fusion is performed to obtain the multi-label prediction probability.

[0051] Vehicle maintenance fault prediction aims to jointly predict multiple future fault labels from historical maintenance sequences, which is crucial for improving vehicle operational safety and service reliability. Existing research typically focuses only on specific subsets of vehicle faults, such as component-level or single-fault-type scenarios like engine fault diagnosis, braking system fault prediction, or battery health status assessment. However, real-world vehicle maintenance records often exhibit time dependence and fault co-occurrence. A single vehicle sample may be associated with multiple future fault labels. In large-scale vehicle maintenance scenarios, the fault label space is extremely large and exhibits severe sparsity and a long-tailed distribution. Historical samples of many low-frequency labels are extremely limited, and many fault labels have only appeared a few times in historical records. Existing vehicle fault prediction methods that rely on training data struggle to reliably model all fault types. To address this, this invention proposes a large-scale vehicle sparse fault prediction method based on semantically adaptive label transfer. This method combines empirical transfer priors with semantic transfer priors derived from a large language model. The empirical transfer priors serve as reliable but sparse structural anchors, while fault label semantics act as supplementary signals for under-observed but semantically plausible relationships. Furthermore, instead of directly calling the large language model to assign numerical scores to label pairs, a frozen large language model is used as a label semantic encoder to extract the representation of each fault label offline. Under the supervision of empirical edges, a lightweight semantic transfer projector learns task-calibrated semantic transfer scores. This design avoids the high cost and potential instability of pairwise large language model scoring, while adapting semantic relationships to the target vehicle's fault prediction task.

[0052] After obtaining the empirical transfer prior and the semantic transfer prior, this invention introduces an adaptive prior injection mechanism. The empirical transfer prior and the semantic transfer prior are injected into the latent state of the historical maintenance sequence through label space propagation, sample-level gating, and residual fusion, generating a transfer-enhanced sequence representation. Subsequently, a prediction head aggregates the enhanced sequence representation from both temporal and contextual perspectives, and applies temporal-contextual routing and residual interaction fusion for adaptive fusion and interaction modeling. Through the adaptive prior injection mechanism, this invention improves the prediction capability and accuracy of low-frequency fault labels.

[0053] In a specific embodiment of the present invention, vehicle repair samples, fault tag sequences, and text description information corresponding to each fault tag are obtained, and historical vehicle repair sequence data is constructed. Specifically, original repair samples are obtained from the vehicle repair work order system. Each repair sample contains at least the following information: a unique vehicle identifier, a fault tag, and the date and time of occurrence of the fault tag. The fault tag includes a fault code and a component name. For each vehicle, the fault tags are sorted according to the order of their fault occurrence time to form the original repair event sequence of the vehicle. At the same time, the text information of the fault code and component name corresponding to each fault tag is extracted as the text description information of the fault tag. The original repair event sequence of the vehicle and the text description information of the fault tags are integrated to obtain the historical vehicle repair sequence data.

[0054] In a specific embodiment of the present invention, the present invention constructs an experience transfer prior based on vehicle historical maintenance sequence data by statistically analyzing the time decay co-occurrence intensity of fault labels, and constructs a semantic transfer prior through a semantic transfer projector, specifically including: Iterate through all time steps of the vehicle's historical maintenance sequence data; For each fault label in the vehicle's historical maintenance sequence data at the current time step Label the fault As the source fault label, other fault labels besides the original fault label will be used. As a target fault label; Within the maximum future observation time window defined after the source fault label, count the number of occurrences of each target fault label; Based on the frequency of occurrence of each target fault label, a decay weight allocation is performed according to the time interval from the current time step to obtain the weight of each fault label; the decay weight allocation is specifically as follows: the farther away from the current time step, the smaller the weight. For each label pair consisting of a source fault label and a target fault label, the corresponding decay-weighted co-occurrence in the training set is accumulated to obtain the directional time decay co-occurrence intensity from the source fault label to the target fault label. The time decay co-occurrence intensity of each fault tag in the vehicle's historical maintenance sequence data at the current time step is calculated; the expression for the time decay co-occurrence intensity is as follows:

[0055] in, To remove fault labels To the fault label The directional time decay co-occurrence intensity, fault label For source fault labels, fault labels For the target fault label, and These are all indexes of fault labels. This is a vehicle's historical maintenance sequence. For the historical maintenance sequence of all vehicles, For time steps, To step from the current time The offset step for counting backwards, ranging from 1 to... , For the maximum future observation time window, This is the time decay factor, and its value range is... , For sequence Mid-time step Are there any fault labels? The indicated value, For sequence Mid-time step Are there any fault labels? The indicated value; Based on the frequency of occurrence of each source fault label in the training set, the corresponding directional time decay co-occurrence intensity is normalized to obtain the normalized transfer intensity. For each source fault label, retain the target fault labels whose normalized transfer strength is greater than a preset transfer strength threshold and which rank among the top preset number according to their normalized transfer strength. Based on the retained source fault labels, target fault labels, and their corresponding normalized transfer strengths, construct an empirical transfer prior. The expression for the empirical transfer prior is as follows:

[0056]

[0057]

[0058] in, To remove fault labels To the fault label Experience transfer prior, To normalize the transfer intensity, Fault Label Total frequency of occurrence Fault Label The set of retained fault labels To preset the transfer intensity threshold, To select the top with the highest normalized transfer strength Fault label.

[0059] In one embodiment of the present invention, constructing a semantic transfer prior specifically includes: Obtain text description information corresponding to each fault tag in the vehicle's historical maintenance sequence data. The text description information includes fault codes and component names. Based on the text description information of each fault label, construct a single-label text for each fault label; Based on the single-label text of each fault label, input them into the frozen large language model and output the hidden state of each fault label. Masked average pooling and norm normalization are performed on the effective words in each hidden state to obtain the fixed semantic vector of each fault label. Based on the fixed semantic vector of each fault label, combined with the empirical successor label of the empirical transfer prior, the nearest neighbor label in the semantic space, and the randomly sampled exploration label, a set of candidate fault labels is constructed for each fault label. For each fault label and each fault label in its candidate fault label set, the fixed semantic vectors of the two labels are combined and fed into a lightweight projector to output a semantic transfer score. The expression for the semantic transfer score is as follows:

[0060]

[0061]

[0062]

[0063]

[0064] in, To remove fault labels To the fault label semantic transfer score, It is the sigmoid activation function. It is a multilayer perceptron network. and All labels are represented after bottleneck transformation. Fault Label Fixed semantic vectors, Fault Label Fixed semantic vectors, and These are all indexes of fault labels. This is a vector concatenation operation. For element-wise multiplication, The first in the candidate fault label set One fault label, For the set of candidate fault labels, All of these are fault labels from the candidate fault label set. The superscript represents the number of fault labels in the candidate fault label set. It is the transpose symbol. For the real number field, For fixed semantic vector dimensions, For fault labels The unnormalized semantic vector is obtained by masking and average pooling the corresponding effective word hidden states. for norm, This is the mask indicator value, which can be either 0 or 1. For the index of lexical elements, The first hidden state Each word element, It is in a hidden state. Indicates the first Each word element is a valid word element. Indicates the first Each word element is an invalid word element; Using empirical transfer priors as supervision signals, a lightweight semantic transfer projector is trained. Based on the existence of pseudo-labels for edges and empirical transfer strength derived from the empirical transfer priors, the lightweight semantic transfer projector is trained by minimizing the projector loss function, resulting in a trained lightweight semantic transfer projector. The expression for the loss function of the lightweight semantic transfer projector is as follows:

[0065] in, For the loss function of a lightweight semantic transfer projector, The cross-entropy loss function is used for binary classification. Let the mean squared error loss function be . The logit value output by the lightweight semantic transfer projector without sigmoid activation. To transfer prior knowledge from experience The exported edges contain pseudo-labels. The weighting coefficients are used to balance the cross-entropy loss and the mean squared error loss in binary classification. To remove fault labels To the fault label semantic transfer score, To remove fault labels To the fault label The prior knowledge transfer.

[0066] The semantic transfer scores of all candidate label pairs output by the trained lightweight semantic transfer projector are sparsified. Candidate label pairs with scores greater than a preset semantic transfer score threshold are retained, and the top few faulty labels with the highest semantic transfer scores are retained. This yields the sparsified semantic transfer prior, the expression of which is as follows:

[0067]

[0068] in, To remove fault labels To the fault label semantic transfer prior, For the set of candidate fault labels, To preset the semantic transfer score threshold, Semantic transfer score in fault label The descending ranking of all semantic transfer scores. To select the top semantic transfer score One fault label, For the transfer of prior experience to subsequent labels, The nearest neighbor label in the semantic space. To explore labels for random sampling.

[0069] In a specific embodiment of the present invention, the present invention obtains an enhanced temporal representation based on vehicle historical maintenance sequence data, experience transfer priors, and semantic transfer priors, through sequence encoding and prior injection processing, specifically including: Input the vehicle's historical maintenance sequence data into the sequence encoder and output the hidden state of each time step for each maintenance sample; The hidden states at each time step are projected onto the label space through a linear mapping to obtain the label space activation vector; Multiply the label space activation vector by the empirical transfer prior and the semantic transfer prior respectively to obtain the activation vector after empirical propagation and the activation vector after semantic propagation. Based on the activation vectors after empirical propagation and semantic propagation, the empirical enhancement amount and semantic enhancement amount are obtained by projecting them back into the latent space through linear mapping. The hidden state of the last valid time step of each maintenance sample in the vehicle history maintenance sequence data is input into a linear layer with sigmoid activation, and the sample-level gating coefficient of each maintenance sample is output. Based on the sample-level gating coefficients, the empirical enhancement and semantic enhancement are weighted and fused, and then multiplied by a learnable overall scaling factor before being superimposed with the hidden states at each time step to obtain the sequence of hidden states for transition enhancement, which serves as the temporal representation of the enhancement. The expression for the temporal representation of the enhancement is as follows:

[0070]

[0071]

[0072]

[0073] in, For the first One repair sample at time step Enhanced timing representation, For the first One repair sample at time step The hidden state, The learnable overall scaling factor. For the first Sample-level gating coefficients for each maintenance sample. For the first One repair sample at time step The amount of experience enhancement, For the first One repair sample at time step The amount of semantic enhancement, It is the sigmoid activation function. This is the weight matrix for the gating coefficients. For the first The hidden state of the last valid time step of a maintenance sample This is the bias term for the gating coefficient. The weight matrix for projecting empirical enhancements back into the latent space. Let be the weight matrix mapping the hidden states to the label space. For the transfer of prior knowledge, The weight matrix for projecting semantic enhancements back into the latent space. Let be the weight matrix mapping the hidden states to the label space. This serves as a semantic transfer prior.

[0074] In a specific embodiment of the present invention, the present invention extracts features from both temporal and contextual perspectives based on enhanced temporal representation and performs residual fusion to obtain multi-label prediction probabilities, specifically including: Based on the enhanced temporal representation, average pooling is performed on multiple time windows through the temporal stream, and the pooled window representations of each window are weighted, summed, and then linearly mapped to obtain the temporal features; Based on the augmented temporal representation, the augmented temporal representation is averaged along the time dimension through the context stream to obtain the global sample representation. The global sample representation is used as the query and the augmented temporal representation is used as the key and value for cross-attention calculation to obtain the context features. Based on temporal and contextual features, a basic interaction representation is constructed, and the temporal and contextual features are adaptively fused through a routing mechanism to obtain a basic fused representation. Based on temporal and contextual features, residual interaction terms are constructed and superimposed with the basic fusion representation to obtain sample-level representations; Based on the sample-level representation, the data is mapped to the label space through a linear classification layer. The independent predicted probabilities of each fault label are then output using the sigmoid activation function, yielding the multi-label predicted probabilities. The expression for the multi-label predicted probabilities is as follows:

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] in, For multi-label probability prediction, It is the sigmoid activation function. The weight matrix of the linear classification layer. For sample-level representation, For the bias term of the linear classification layer, Based on the fusion representation, The learnable residual scaling factor has a value range of [value range missing]. , For residual interaction terms, These are learnable interpolation coefficients, with a value range of [value range missing]. , Basic interactive representation, According to routing experts, Multilayer perceptron networks based on interactive representations, As a time series feature, For contextual features, This is a vector concatenation operation. For element-wise multiplication, The linear mapping matrix for time series features, , and All are learnable weights. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of a repair sample within a short time window. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of each maintenance sample within a mid-time window. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of each maintenance sample within the full time window. For cross-attention calculation, For global sample representation, To enhance the timing representation, For average pooling, For the first The repair sample corresponds to the first Routing weights of an expert network For the first An expert network with independent learnable parameters. For the first Temporal characteristics of a repair sample The shared routing space representation obtained after learnable projection. For the first Contextual features of a repair sample The shared routing space representation obtained after learnable projection. For the first The routing weight vector corresponding to each maintenance sample For indexing expert networks, For the number of expert networks, Here is the weight matrix of the router. This is the router's bias term. For the first Encoding of the effective sequence length of a repair sample The learnable projection matrix corresponding to the temporal features. The bias term corresponding to the time series features. The learnable projection matrix corresponding to the context features. The bias term corresponding to the context features. For sequence length encoding function, For the first The number of valid time steps for each repair sample.

[0084] In a specific embodiment of the present invention, the loss function of the lightweight semantic transfer projector includes edge existence constraints and transfer strength constraints, and its expression is as follows:

[0085] in, For the loss function of a lightweight semantic transfer projector, The cross-entropy loss function is used for binary classification. Let the mean squared error loss function be . This is the logit value without sigmoid activation output by the lightweight semantic transfer projector for the fault label pair (p,q). The semantic transfer score is the score from fault label p to fault label q. The edges derived from prior knowledge through empirical transfer may have pseudo-labels. For the empirical transfer prior from fault label p to fault label q, The weighting coefficients are used to balance the cross-entropy loss and the mean squared error loss in binary classification. Specifically, when the fault label pair (p,q) belongs to the high-confidence empirical transition edge in the empirical transition prior, When the fault label pair (p,q) is a non-empirical transition candidate edge obtained from sampling, The lightweight semantic transfer projector learns whether there is a valid transfer relationship between candidate fault label pairs through binary cross-entropy loss; through mean squared error loss, the lightweight semantic transfer projector further learns the transfer strength information in the empirical transfer prior, thereby obtaining a semantic transfer score calibrated by the empirical transfer relationship.

[0086] In a specific embodiment of the present invention, the large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer can form a large-scale vehicle sparse fault prediction model, namely the master prediction model. Based on the real multi-label fault data and multi-label prediction probabilities corresponding to vehicle repair samples, the master prediction model can be trained using a binary classification cross-entropy loss function. The expression of the prediction loss function of the master prediction model is as follows:

[0087] in, The prediction loss function of the main prediction model. The number of maintenance samples participating in the training. For fault tags total, No. The repair sample corresponds to the first The actual indication value of each fault tag is either 0 or 1. For the first The repair sample corresponds to the first The predicted probability of each fault label.

[0088] The beneficial effects of this invention are: This invention addresses the problem that existing methods, which rely solely on observed co-occurrence or statistics in training data, cannot utilize reasonable transfer relationships between low-frequency fault labels when the fault label space is large and long-tailed in large-scale vehicle maintenance scenarios, by constructing a dual-priority collaborative mechanism of empirical transfer prior and semantic transfer prior. This improves the overall vehicle fault prediction performance and the prediction ability of low-frequency fault labels. In the construction of the experience transfer prior, this invention uses time decay weighting to give higher weight to recent fault labels, accurately depicting the timeliness of the evolution of actual vehicle faults. It also eliminates the statistical dominance effect of high-frequency labels through frequency normalization, avoiding large errors caused by excessive frequency deviations of different fault labels. Furthermore, it improves the reliability and computational efficiency of the experience transfer prior by screening high-confidence experience transfer priors through thresholding and TopK sparsification.

[0089] This invention introduces a semantic transfer prior based on a frozen large language model, which encodes each fault label only once. It uses a lightweight semantic transfer projector with empirical transfer prior as supervision for calibration, avoiding the quadratic computational overhead of pairwise inference in large language models. This achieves low-cost, reproducible, and task-aligned semantic transfer prior construction, providing supplementary information for long-tailed fault labels that cannot be covered by empirical transfer prior.

[0090] This invention injects experience transfer priors and semantic transfer priors into the hidden state of each time step of the sequence encoder, enabling prior knowledge to deeply participate in the construction process of sequence representation; the sample-level gating coefficients adaptively adjust the dual prior fusion ratio according to the sequence characteristics of each maintenance sample, so that samples with different sparsity levels receive different prior guidance; the residual superposition design protects the main information of the original sequence encoding while introducing external priors, realizing the synergy between prior knowledge and data-driven approaches in large-scale sparse scenarios. This invention extracts global maintenance information from both temporal and contextual perspectives in the prediction head, overcoming the problem that a single path cannot take into account both dynamic evolution and global context. The routing mechanism enables adaptive feature fusion of samples, and residual interactive refinement further supplements the fine-grained information on the differences and consistency between temporal and contextual information, making the prediction representation both temporally sensitive and contextually global, thus improving the prediction accuracy in large-scale sparse fault label scenarios.

Claims

1. A method for large-scale vehicle sparse fault prediction based on semantic adaptive label transfer, characterized in that, Includes the following steps: Obtain vehicle repair samples, fault label sequences, and text descriptions of each fault label, and construct historical vehicle repair sequence data; Based on historical vehicle maintenance sequence data, an experience transfer prior is constructed by statistically analyzing the time decay co-occurrence intensity of fault labels, and a semantic transfer prior is constructed by using a semantic transfer projector. Based on vehicle historical maintenance sequence data, experience transfer priors, and semantic transfer priors, an enhanced temporal representation is obtained through sequence encoding and prior injection processing. Based on the enhanced temporal representation, features are extracted from both temporal and contextual perspectives and residual fusion is performed to obtain the multi-label prediction probability.

2. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 1, characterized in that, The aforementioned construction experience transfer prior specifically includes: Iterate through all time steps of the vehicle's historical maintenance sequence data; For each fault label in the vehicle's historical maintenance sequence data at the current time step, use that fault label as the source fault label and other fault labels other than the original fault label as the target fault label. Within the maximum future observation time window defined after the source fault label, count the number of occurrences of each target fault label; Based on the frequency of occurrence of each target fault label, a decay weight allocation is performed according to the time interval from the current time step to obtain the weight of each fault label; the decay weight allocation is specifically as follows: the farther away from the current time step, the smaller the weight. For each label pair consisting of a source fault label and a target fault label, the corresponding decay-weighted co-occurrence in the training set is accumulated to obtain the directional time decay co-occurrence intensity from the source fault label to the target fault label. Based on the frequency of occurrence of each source fault label in the training set, the corresponding directional time decay co-occurrence intensity is normalized to obtain the normalized transfer intensity. For each source fault label, retain the target fault labels whose normalized transfer strength is greater than the preset transfer strength threshold and which are ranked in the top preset number according to the normalized transfer strength. Based on the retained source fault labels, target fault labels and corresponding normalized transfer strengths, construct an empirical transfer prior.

3. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 2, characterized in that, The expression for the time-decay co-occurrence intensity is as follows: in, To remove fault labels To the fault label The directional time decay co-occurrence intensity, fault label For source fault labels, fault labels For the target fault label, and These are all indexes of fault labels. This is a vehicle's historical maintenance sequence. For the historical maintenance sequence of all vehicles, For time steps, To step from the current time The offset step for counting backwards, ranging from 1 to... , For the maximum future observation time window, This is the time decay factor, and its value range is... , For sequence Mid-time step Are there any fault labels? The indicated value, For sequence Mid-time step Are there any fault labels? The indicated value; The expression for the prior knowledge transfer is as follows: in, To remove fault labels To the fault label The transfer of prior knowledge, To normalize the transfer intensity, Fault Label Total frequency of occurrence Fault Label The set of retained fault labels To preset the transfer intensity threshold, To select the top with the highest normalized transfer strength Fault label.

4. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 1, characterized in that, The construction of semantic transfer priors specifically includes: Obtain text description information corresponding to each fault tag in the vehicle's historical maintenance sequence data. The text description information includes fault codes and component names. Based on the text description information of each fault label, construct a single-label text for each fault label; Based on the single-label text of each fault label, input them into the frozen large language model and output the hidden state of each fault label. Masked average pooling and norm normalization are performed on the effective words in each hidden state to obtain the fixed semantic vector of each fault label. Based on the fixed semantic vector of each fault label, combined with the empirical successor label of the empirical transfer prior, the nearest neighbor label in the semantic space, and the randomly sampled exploration label, a set of candidate fault labels is constructed for each fault label. For each fault label and each fault label in its candidate fault label set, the fixed semantic vectors of the two are combined with features and then fed into a lightweight projector to output a semantic transfer score. Using empirical transfer priors as supervision signals, a lightweight semantic transfer projector is trained. Based on the existence of pseudo-labels on edges and empirical transfer strength derived from empirical transfer priors, the lightweight semantic transfer projector is trained by minimizing the projector loss function, resulting in a well-trained lightweight semantic transfer projector. The semantic transfer scores of all candidate label pairs output by the trained lightweight semantic transfer projector are sparsified. Candidate label pairs with scores greater than the preset semantic transfer score threshold are retained, and the top few faulty labels with the highest semantic transfer scores are retained to obtain the sparsified semantic transfer prior.

5. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 4, characterized in that, The expression for the semantic transfer score is as follows: in, To remove fault labels To the fault label semantic transfer score, It is the sigmoid activation function. It is a multilayer perceptron network. and All labels are represented after bottleneck transformation. Fault Label Fixed semantic vectors, Fault Label Fixed semantic vectors, and These are all indexes of fault labels. This is a vector concatenation operation. For element-wise multiplication, The first in the candidate fault label set One fault label, For the set of candidate fault labels, All of these are fault labels from the candidate fault label set. The superscript represents the number of fault labels in the candidate fault label set. It is the transpose symbol. For the real number field, For fixed semantic vector dimensions, For fault labels The unnormalized semantic vector is obtained by masking and average pooling the corresponding effective lexical hidden states. for norm, This is the mask indicator value, which can be either 0 or 1. For the index of lexical elements, The first hidden state Each word element, It is in a hidden state. Indicates the first Each word element is a valid word element. Indicates the first Each word element is an invalid word element; The expression for the semantic transfer prior is as follows: in, To remove fault labels To the fault label semantic transfer prior, For the set of candidate fault labels, To preset the semantic transfer score threshold, Semantic transfer score in fault label The descending ranking of all semantic transfer scores. To select the top semantic transfer score One fault label, For the transfer of prior experience to subsequent labels, The nearest neighbor label in the semantic space. To explore labels for random sampling.

6. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 5, characterized in that, The loss function of the lightweight semantic transfer projector is expressed as follows: in, For the loss function of a lightweight semantic transfer projector, The cross-entropy loss function is used for binary classification. Let the mean squared error loss function be . The logit value output by the lightweight semantic transfer projector without sigmoid activation. To transfer prior knowledge from experience The exported edges contain pseudo-labels. The weighting coefficients are used to balance the cross-entropy loss and the mean squared error loss in binary classification. To remove fault labels To the fault label semantic transfer score, To remove fault labels To the fault label The prior knowledge transfer.

7. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 1, characterized in that, The obtained enhanced temporal representation specifically includes: Input the vehicle's historical maintenance sequence data into the sequence encoder and output the hidden state of each time step for each maintenance sample; The hidden states at each time step are projected onto the label space through a linear mapping to obtain the label space activation vector; Multiply the label space activation vector by the empirical transfer prior and the semantic transfer prior respectively to obtain the activation vector after empirical propagation and the activation vector after semantic propagation. Based on the activation vectors after empirical propagation and semantic propagation, the empirical enhancement amount and semantic enhancement amount are obtained by projecting them back into the latent space through linear mapping. Input the hidden state of the last valid time step of each maintenance sample in the vehicle history maintenance sequence data into a linear layer with sigmoid activation, and output the sample-level gating coefficient of each maintenance sample. Based on the sample-level gating coefficients, the empirical enhancement and semantic enhancement are weighted and fused, and then multiplied by the learnable overall scaling coefficient and superimposed with the hidden states of each time step to obtain the sequence of hidden states for transition enhancement, which serves as the temporal representation of the enhancement.

8. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 7, characterized in that, The expression for the enhanced timing representation is as follows: in, For the first One repair sample at time step Enhanced timing representation, For the first One repair sample at time step The hidden state, The learnable overall scaling factor. For the first Sample-level gating coefficients for each maintenance sample. For the first One repair sample at time step The amount of experience enhancement For the first One repair sample at time step The amount of semantic enhancement, It is the sigmoid activation function. This is the weight matrix for the gating coefficients. For the first The hidden state of the last valid time step of a maintenance sample This is the bias term for the gating coefficient. The weight matrix for projecting empirical enhancements back into the latent space. This is the weight matrix mapping the hidden states to the label space. For the transfer of prior knowledge, The weight matrix for projecting semantic enhancements back into the latent space. This is the weight matrix mapping the hidden states to the label space. This serves as a semantic transfer prior.

9. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 1, characterized in that, The process of obtaining the multi-label prediction probability specifically includes: Based on the enhanced temporal representation, average pooling is performed on multiple time windows through the temporal stream, and the pooled window representations of each window are weighted, summed, and then linearly mapped to obtain the temporal features; Based on the augmented temporal representation, the augmented temporal representation is averaged along the time dimension through the context stream to obtain the global sample representation. The global sample representation is used as the query and the augmented temporal representation is used as the key and value for cross-attention calculation to obtain the context features. Based on temporal and contextual features, a basic interaction representation is constructed, and the temporal and contextual features are adaptively fused through a routing mechanism to obtain a basic fused representation. Based on temporal and contextual features, residual interaction terms are constructed and superimposed with the basic fusion representation to obtain sample-level representations; Based on the sample-level representation, the data is mapped to the label space through a linear classification layer, and the independent prediction probability of each fault label is output through the sigmoid activation function to obtain the multi-label prediction probability.

10. The large-scale vehicle sparse fault prediction method based on semantic adaptive label transfer according to claim 9, characterized in that, The expression for obtaining the multi-label prediction probability is as follows: in, For multi-label probability prediction, It is the sigmoid activation function. The weight matrix of the linear classification layer. For sample-level representation, For the bias term of the linear classification layer, Based on the fusion representation, The learnable residual scaling factor has a value range of [value range missing]. , For residual interaction terms, These are learnable interpolation coefficients, with a value range of [value range missing]. , Basic interactive representation, According to routing experts, Multilayer perceptron networks based on interactive representations, As a time series feature, For contextual features, This is a vector concatenation operation. For element-wise multiplication, The linear mapping matrix for time series features, , and All are learnable weights. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of a maintenance sample within a short time window. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of each maintenance sample within a mid-time window. For the first The windowed aggregate representation is obtained by average pooling the augmented temporal representation of each maintenance sample within the full time window. For cross-attention calculation, For global sample representation, To enhance the timing representation, For average pooling, For the first The repair sample corresponds to the first Routing weights of an expert network For the first An expert network with independent learnable parameters. For the first Temporal characteristics of a repair sample The shared routing space representation obtained after learnable projection. For the first Contextual features of a repair sample The shared routing space representation obtained after learnable projection. For the first The routing weight vector corresponding to each maintenance sample For the index of expert networks, For the number of expert networks, Here is the weight matrix of the router. This is the router's bias term. For the first Encoding of the effective sequence length of a repair sample The learnable projection matrix corresponding to the temporal features. The bias term corresponding to the time series features. The learnable projection matrix corresponding to the context features. The bias term corresponding to the context features. For sequence length encoding function, For the first The number of valid time steps for each repair sample.

Images