A Smart Bus Operation and Maintenance Task Scheduling Method Based on Edge Intelligence Prediction Model

Abstract

The application discloses a smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model, belongs to the field of vehicle operation and maintenance scheduling, and comprises the following steps: performing segment division on a continuous sampling sequence according to a preset time window, extracting derived characteristic parameters used for operation and maintenance risk discrimination and prediction inference in each segment, and encapsulating the derived characteristic parameters and data quality information corresponding to the derived characteristic parameters to generate a characteristic certificate; performing comparable correction on the derived characteristic parameters based on the data quality information in the characteristic certificate; obtaining a certificate-driven operation and maintenance risk judgment result based on a derived characteristic drift index; generating a model input sequence and inputting the model input sequence into an edge intelligent prediction model to obtain a model-driven operation and maintenance risk judgment result; and the problems that different reporting strategy gears are prone to introducing derived characteristic deviation in the prior art, and then causing risk discrimination fluctuation and operation and maintenance scheduling jitter are solved.

Description

Technical Field

[0001] This invention relates to the field of vehicle operation and maintenance scheduling technology, and in particular to a smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model. Background Technology

[0002] With the informatization and intelligentization of public transportation operation and management, buses are typically equipped with driver display terminals, on-board gateways, and various sensors and on-board bus interfaces to collect vehicle operation data, station event data, driving behavior data, and environmental data. This data is then exchanged with edge computing nodes or a central platform via cellular networks. The driver display terminal can show the driver information such as departure intervals, arrival times, task instructions, alarm information, and passenger flow, and can receive dispatch instructions from the dispatch system to assist vehicle operation.

[0003] In terms of data processing and operation and maintenance management, existing systems typically upload data collected from the vehicle side to edge nodes or central platforms. On the platform side, the data is preprocessed based on timestamps, including sorting, deduplication, missing data handling, and resampling. Then, feature extraction, model reasoning, or rule discrimination are performed to obtain the vehicle health status, risk level, or operation and maintenance suggestions. Furthermore, operation and maintenance work orders, task instructions, or scheduling prompts are generated and sent to the vehicle-side terminals for execution.

[0004] In terms of intelligent methods, existing technologies include intelligent operation and maintenance methods based on knowledge organization and reasoning, as well as vehicle scheduling methods based on machine learning models.

[0005] For example, Chinese invention patent CN120671800A discloses an intelligent operation and maintenance method and system for the traction transmission system of rail trains. This method first collects multi-source heterogeneous knowledge, defines a fault knowledge graph ontology and its relationships to construct a knowledge graph, and stores the text data in a vector database after different types of parsing and segmentation to support subsequent accurate retrieval and reasoning. A multi-source knowledge retrieval mechanism is established to extract unstructured text and structured knowledge highly relevant to fault queries by retrieving the vector database and knowledge graph. For unstructured text knowledge, a reordering model is used to optimize the order of knowledge. A prompt word template is constructed to integrate text and structured knowledge, guiding a large language model to perform fault analysis and providing maintenance decision support.

[0006] For example, Chinese invention patent CN113344211A discloses a vehicle scheduling method using machine learning, comprising: a scheduling database storing vehicle data and constraint data for multiple vehicles available for scheduling. A processor is programmed to execute a scheduling server to perform operations including: receiving a scheduling request; using a machine learning model to identify one or more of the multiple vehicles in response to the scheduling request, the machine learning model using vehicle data and constraint data as input to determine one or more of the multiple vehicles; notifying one or more of the multiple vehicles of the scheduling request; receiving information indicating the result of the scheduling request; and using the vehicle data, constraint data, and result to update the training of the machine learning model to improve the machine learning model used for learning scheduling of mobility assistance and services.

[0007] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0008] In smart bus scenarios, the driver's display terminal needs to interact with edge nodes or platforms via cellular networks. During vehicle operation, it experiences 4G / 5G switching, weak coverage sections, and base station congestion, resulting in fluctuations in uplink bandwidth, latency, and packet loss. To adapt to these link fluctuations and ensure the transmission of critical business messages, the vehicle typically employs quantization, differential compression, and batch reporting methods for reporting high-frequency continuous signals. The quantization step size, differential anchor point period, reset method, and batch window length may be adjusted according to link conditions. These strategies introduce various structural errors and uncertainties: quantization can cause amplitude discretization errors; differential compression may result in reconstruction drift or error accumulation within the anchor point period, and statistical shifts under packet loss or reset conditions; batch reporting and supplementary reporting can lead to time delays, changes in the sample distribution structure within the window, and inconsistencies in time alignment. Since derived features are usually calculated from window statistics, peak values, and durations, the aforementioned errors and changes in time structure can cause systematic shifts and short-term fluctuations in derived features under different reporting strategy levels. This leads to unstable threshold comparison results, frequent switching of warning / dispatch status, and operational scheduling jitter and resource waste, making it difficult to achieve interpretable and convergent risk identification and task scheduling closed loops under low bandwidth conditions.

[0009] In existing technologies, since vehicles experience network quality fluctuations during operation, the system needs to adjust strategy parameters such as quantization, differential compression, and batch reporting to form different reporting strategy levels under bandwidth-limited and weak coverage conditions in order to ensure that critical messages can be transmitted. However, different reporting strategy levels will introduce systematic offsets and short-term jitters in derived features, which will lead to fluctuations in risk assessment and jitters in operation and maintenance scheduling. Summary of the Invention

[0010] To address the technical problems of risk assessment fluctuations and operation and maintenance scheduling jitter caused by derived feature offsets in existing technologies, this invention provides a smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model. The technical solution is as follows:

[0011] On the one hand, a smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model is provided, which includes:

[0012] S1. Acquire the operation signal of the smart bus and form a continuous sampling sequence with timestamps. Divide the continuous sampling sequence into segments according to a preset time window, extract the derived feature parameters used for operation and maintenance risk identification and prediction inference in each segment, and encapsulate the derived feature parameters and the corresponding data quality information to generate a feature certificate. The feature certificate is used to uniformly represent the derived feature parameters and their data quality information.

[0013] S2. Establish a certificate queue and a background data queue for dual-channel transmission. Perform comparability correction on the derived feature parameters based on the data quality information in the feature certificate to obtain comparable derived features. Obtain the derived feature drift index based on the difference between the comparable derived features and the historical baseline. Obtain the certificate-driven operation and maintenance risk judgment result based on the derived feature drift index. Generate a model input sequence and input it into the edge intelligent prediction model to obtain the model-driven operation and maintenance risk judgment result. The data quality information is used to indicate the quantization accuracy and missing degree of the derived feature parameters under the current reporting strategy. The model input sequence consists of the comparable derived features, the derived feature drift index, and the quality constraint information obtained from the data quality information.

[0014] S3. Based on the certificate-driven operation and maintenance risk assessment results and the model-driven operation and maintenance risk assessment results, generate an operation and maintenance task priority queue and work order assignment parameters as operation and maintenance scheduling results.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0016] 1. This invention explicitly incorporates the structural uncertainties introduced by the link and reporting strategy into the risk assessment process through the unified encapsulation of feature certificates and the coordination of three-state gating judgment of error interval boundaries. On the vehicle side, while generating derived feature parameters, the link quality level, strategy parameters related to quantization, differential compression, and batch reporting, as well as missing information, are written into the feature certificate, and error interval boundaries are constructed for quality-sensitive derived features. On the central side, instead of simply comparing a single feature value with a threshold, the system outputs a basic gating state based on the relative relationship between the error interval boundary and the judgment threshold, determining whether the system is safe, triggering a decision, or awaiting disambiguation. When awaiting disambiguation, enhanced backfilling is triggered to narrow the interval before further judgment. Unlike existing technologies that directly compare derived features such as "accumulated errors from quantization discretization, differential reconstruction, and batch alignment altering sample distribution" at different levels, leading to frequent threshold reversals, this invention transforms uncertainty into a calculable interval boundary and incorporates it into the decision logic. This makes the judgments on the safety and trigger sides conservatively verifiable, and the state to be disambiguated has a convergent processing path. This significantly reduces short-term fluctuations in the warning and dispatch states, reduces duplicate dispatches and resource waste caused by false triggers and false cancellations, and improves the interpretability and verifiability of the judgment conclusions.

[0017] 2. This invention achieves continuous scheduling under weak network conditions and traceable evidence at critical moments through the combination of dual-queue, dual-channel transmission and on-demand minimum replenishment convergence. A certificate queue and a background data queue are established on the vehicle side. The certificate queue periodically sends carrying feature certificates with high priority, enabling the central side to perform risk gating and scheduling based on certificates even when not continuously receiving large volumes of raw data. The background data queue carries raw sampling sequences or compressed sequences and sends or supplements them when bandwidth allows, for evidence replenishment and recalculation confirmation when disambiguation and verification are triggered. Unlike existing technologies that either continuously upload high-frequency raw sequences, leading to weak network congestion and the blocking of critical scheduling messages, or only upload compressed features, making verification impossible in disputed states, this invention maintains the continuity of discrimination and scheduling under network congestion, handover, and offline conditions. Simultaneously, when disambiguation is pending or fault modes need to be locked, it can perform minimum necessary replenishment for uncertain time ranges and quickly converge evidence boundaries, thus balancing bandwidth usage, decision continuity, and verifiability of verification.

[0018] 3. This invention improves the consistency of trend identification across different price levels and missing structures by combining comparable correction with baseline normalized drift measurement and introducing evidence credibility weights. This complements the predictive model's advance suggestion mechanism. The system uses data quality information from feature certificates to perform comparable correction on derived features, ensuring that features obtained under different quantization step sizes, missing proportions, and different differencing and batching strategies return to a unified statistical standard. Based on this, a normalized deviation measurement is performed with historical healthy baselines to obtain a derived feature drift index. The evidence uncertainty reflected by the error interval boundary width is used to generate feature credibility weights, thereby suppressing the amplification effect of low-quality evidence on the drift index and avoiding misjudging feature shifts caused by strategy switching as true deterioration trends. Furthermore, this invention limits the output of the edge intelligent prediction model to advance suggestions rather than as the basis for dispatch gating, so that the prediction results are used for resource pre-allocation, disambiguation priority and replenishment time limit adjustment, and its availability is controlled by prediction confidence and input quality gating, thereby improving the forward response capability to risk escalation without introducing the risk of mis-dispatch, realizing the collaborative work of certificate-driven stable dispatch and model-driven advance preparation, and improving the overall scheduling efficiency and timeliness of handling. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of a smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model provided in this application embodiment;

[0021] Figure 2 This is a block diagram of the edge intelligent prediction model structure provided in the embodiments of this application;

[0022] Figure 3 A flowchart for feature certificate generation and gating basic state determination provided in the embodiments of this application;

[0023] Figure 4 A flowchart for determining the disposal priority based on the drift index provided in this application embodiment;

[0024] Figure 5 A flowchart illustrating the joint scheduling of certificate-driven and model-driven methods provided in this application embodiment. Detailed Implementation

[0025] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0026] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0027] This embodiment provides a smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model, applicable to the operation and maintenance risk assessment and task assignment scheduling of smart buses in scenarios with network quality fluctuations, bandwidth limitations, or offline conditions. The method is implemented collaboratively by the vehicle-side edge and the central side. The vehicle-side edge is used to collect operating signals, generate feature certificates, perform comparability correction, drift measurement, and predictive inference, and trigger minimum necessary evidence supplementation when needed. The central side is used to aggregate risk results from multiple vehicles, generate an operation and maintenance task priority queue, determine work order assignment parameters, and issue scheduling instructions. For ease of understanding, this embodiment uses brake temperature rise, brake pressure, vibration, and key electrical quantities as typical operating signals, but the invention is not limited to these.

[0028] Operating signals refer to continuous time-series signals output by sensors or acquisition units during vehicle operation. A continuous sampling sequence refers to a sequence of sampling points formed by sampling the operating signal and recording timestamps. A segment refers to a data block obtained by dividing a continuous sampling sequence according to a preset segment time window. A micro-window refers to a sub-time window obtained by further dividing a segment according to shorter time windows, used to obtain more stable statistics. Data quality information refers to link and reporting strategy information related to the comparability and error boundaries of derived feature parameters, including at least link quality levels and strategy parameters related to quantization, differential compression, and batch reporting. Feature certificates are evidence carriers used to uniformly encapsulate derived feature parameters and their data quality information, and provide feature error interval boundaries.

[0029] Specific Implementation Example 1, such as Figure 1 The diagram shown is a flowchart of a smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model provided in this application embodiment. The method includes the following steps:

[0030] S1. Acquire the operation signals of the smart bus and form a continuous sampling sequence with timestamps. Divide the continuous sampling sequence into segments according to a preset time window, and divide each segment into multiple micro-windows according to a preset micro-window length. Use the micro-window as the smallest statistical and evidence collection unit to extract derived feature parameters for operation and maintenance risk identification and prediction. Encapsulate the derived feature parameters and their corresponding data quality information to generate a feature certificate. The feature certificate is used to uniformly represent the derived feature parameters and their data quality information.

[0031] Derived feature parameters refer to the set of parameters obtained from the original sampling points within a micro-window or segment through statistical, differential, energy calculation, or duration discrimination methods, which are used for risk identification and prediction.

[0032] The statistical calculation includes calculating the mean, median, or root mean square of the sampled values ​​within a micro-window; the difference calculation includes subtracting adjacent sampled points within a micro-window and taking the maximum value of the difference amplitude to characterize the intensity of the abrupt change; the energy calculation includes summing the squares of the sampled values ​​or taking the mean square within a micro-window to characterize the vibration or impact energy level; the duration discrimination includes comparing the sampled values ​​with a pre-configured discrimination threshold and accumulating the time corresponding to the sampled points that exceed the threshold to obtain the over-threshold duration, wherein the discrimination threshold is pre-configured according to vehicle type, signal type, or operating condition category.

[0033] The vehicle-side edge acquires the smart bus's operating signals and forms a continuous sampling sequence with timestamps. Operating signals can be one or more of the following: brake temperature rise, brake pressure, vibration, and critical electrical quantities. For each sampling point, a sampling timestamp and sample value are recorded and written to the vehicle's local cache. To ensure evidence retrieval and restoration even with poor network quality, this embodiment uses a circular cache to store the most recent original sampling point sequence. The length of the circular cache is configured and ensures it covers at least several time ranges that may be requested by subsequent steps. The reason for using a circular cache is that the certificate queue only carries low-bandwidth evidence representation, while when risks are in a state of disambiguation or conflict, it is necessary to retrieve original or higher-precision data locally to narrow uncertainty. The circular cache ensures that this restoration can be hit at the vehicle end and can be completed quickly.

[0034] The continuous sampling sequence is divided into segments according to a preset segment time window. The preset segment time window (e.g., 10 seconds) is determined by considering at least the following factors: the typical risk evolution time scale or fault symptom duration scale of the target operating signal, the fixed transmission cycle of the certificate queue and the expected risk decision delay upper limit, and the constraints of vehicle-side computing resources and cache capacity. Each segment should cover at least one complete risk discrimination statistical cycle to ensure that drift measurement and threshold discrimination have stable statistical significance. A segment identifier is generated for each segment. The segment identifier includes at least the vehicle identifier, signal type identifier, and segment number, with the segment number increasing sequentially according to the time of segment generation. To improve the stability and comparability of features, this embodiment further divides each segment into multiple micro-windows according to a preset micro-window length. The micro-window length is configurable, and the preset micro-window length (e.g., 2 seconds) must at least satisfy the following constraints: First, the number of sampling points within a micro-window is not less than the preset minimum number of samples; second, the micro-window length is not greater than the shortest duration scale of the signal under approximately stationary conditions to make the feature statistics within the micro-window more stable; third, the segment is continuously covered by an integer number of micro-windows so that the micro-window index can be used for accurate positioning and minimum necessary backfilling of subsequent enhancement requests. Micro-windows continuously cover within a segment, and the position index of the micro-window within the segment is recorded. The purpose of setting micro-windows is that when the link status or reporting strategy changes, micro-window-level statistics can limit unstable factors to a shorter time range, thereby making feature changes between adjacent micro-windows easier to interpret and improving the accuracy of drift measurement.

[0035] In this embodiment, a braking system-related risk occurs in vehicle BUS0173 during operation as an example. The vehicle collects operating signals including brake pressure, brake vibration, and brake temperature rise. The sampling frequency for brake pressure and brake vibration is fifty times per second, and for brake temperature rise, it is once per second. The vehicle-side divides the continuous sampling sequence into segments with ten-second time windows, and further subdivides each segment into micro-windows with a one-second micro-window length. The vehicle-side generates a feature certificate at the micro-window granularity. The certificate records the vehicle identifier, signal type identifier, segment number, micro-window index range, and timestamp range, as well as derived feature parameters, data quality information, and error interval boundaries.

[0036] Enhancement requests are control-type small messages, preferably sent via a certificate queue with the same priority or an independent control channel. Their purpose is to limit the scope and accuracy of the supplementation, ensuring that the vehicle only supplements the minimum necessary evidence without continuously uploading the original sequence, thus achieving evidence convergence and controllable bandwidth under weak network conditions. An enhancement request must include at least the following fields: vehicle identifier, signal type identifier, segment identifier and micro-window position index range, trigger reason identifier, supplementation data type, supplementation accuracy parameter, and supplementation time limit. The trigger reason identifier indicates that the trigger scenario for this enhancement request is one of the following: pending disambiguation processing, rapid review, or fault mode locking; the supplementation data type must include at least one of the following: requesting a compressed sequence or derived feature under a finer quantization step size, requesting the anchor point reset frame and differential increment corresponding to the micro-window, or requesting a short-term original subsequence of the micro-window; the supplementation accuracy parameter must include at least one or more of the following: target quantization step size, target supplementation length, or target number of sample points, used to limit the supplementation intensity; the supplementation time limit limits the latest completion time of the vehicle's supplementation to support the central-side scheduling closed loop.

[0037] Derived feature parameters are calculated based on the original sampling points covered by each micro-window. These derived feature parameters include at least two categories: quality-invariant derived features and quality-sensitive derived features. Quality-invariant derived features are used to maintain statistical consistency across different reporting strategy levels, typically including micro-window mean, micro-window median, and steady-state root mean square. Quality-invariant derived features are less sensitive to quantization and batch reporting under resampling or fixed-point statistical conditions. Quality-sensitive derived features are used to characterize risk information such as peak values, abrupt changes, energy levels, or durations, typically including peak value, peak-to-peak value, maximum first-order difference, maximum second-order difference, short-term energy, and over-threshold duration. Quality-invariant derived features ensure basic comparability across different levels, while quality-sensitive derived features provide risk evidence that is more sensitive to early signs of faults. The combination of both allows the system to continuously monitor under low bandwidth conditions and trigger the minimum necessary accuracy compensation at critical moments.

[0038] Data quality information corresponding to derived feature parameters is generated. This information includes at least the current link quality level and a set of strategy parameters related to quantization, differential compression, and batch reporting. The strategy parameter set includes at least one or more of the following: quantization step size, differential compression anchor period, and batch reporting window length. Optional parameters include missing proportion estimation and backfilling delay estimation. The link quality level is a predefined set of discrete levels, preferably multiple fixed levels, such as four levels, corresponding to different states of link quality from high to low. A mapping relationship between levels and strategy parameters is pre-established in the system configuration, meaning each link quality level uniquely corresponds to a set of strategy parameter values. During operation, the vehicle-mounted edge determines the current link quality level based on real-time link status or reported strategy control results. It then directly looks up the corresponding strategy parameters, such as the quantization step size, anchor period, and batch reporting window length, in a table. The current link quality level identifier and its corresponding strategy parameter values ​​are written into the feature certificate, ensuring consistent processing of subsequent comparability correction, error interval boundary construction, and risk assessment based on the same set of defined levels and parameter values.

[0039] In this embodiment, a braking system-related risk occurs in vehicle BUS0173 during operation as an example. Further, a mapping table of link quality levels and reporting strategy parameters is pre-established on the vehicle side, with four levels in total. Level one corresponds to high link quality, and level four corresponds to low link quality. Each level uniquely corresponds to three strategy parameters: quantization step size, differential compression anchor point period, and batch reporting window length. Certificate queues and background data queues are established separately, with the certificate queue carrying feature certificates and the background data queue carrying original sampling sequences or compressed sequences. The certificate queue has a higher priority than the background data queue. Between 10:15:20 and 10:15:40, the vehicle enters a base station congestion section, and the vehicle-side link quality assessment result switches from level one to level two, and then to level three. The corresponding reporting strategy parameters are updated by the vehicle-side table lookup as follows: Level 1: quantization step size 0.05, anchor point period 2 seconds, batch window 2 seconds; Level 2: quantization step size 0.2, anchor point period 4 seconds, batch window 4 seconds; Level 3: quantization step size 0.5, anchor point period 8 seconds, batch window 8 seconds. The vehicle-side writes the current link quality level and the above strategy parameters into the feature certificate of the current micro-window, ensuring that subsequent corrections, interval boundary construction, and gating decisions are all based on the same set of clearly defined levels for consistent processing.

[0040] After encapsulating derived feature parameters and data quality information to generate a feature certificate, this embodiment provides a corresponding error interval boundary for each quality-sensitive derived feature in the feature certificate to enable the feature certificate to support subsequent certificate-driven threshold discrimination and disambiguation control. The error interval boundary is used to provide a conservative uncertainty interval boundary for the quality-sensitive derived feature under the current reporting policy level, and is used to form a decision-making basis for determining safety, triggering, or pending disambiguation without continuously uploading the original sequence.

[0041] In this embodiment, the feature certificate is generated and sent at the micro-window granularity. The segment identifier and micro-window position index are used to achieve cross-end consistent positioning and minimum necessary backfilling. The feature certificate can be organized using the following fields. The certificate header fields include vehicle identifier, signal type identifier, segment identifier, segment sequence number, micro-window index range, and micro-window timestamp range. The certificate payload fields include a set of derived feature parameters, a set of data quality information, and a set of error interval boundaries. The data quality information set includes at least the link quality level L, quantization step size q, differential compression anchor period Pa, batch reporting window length B, missing proportion estimate ρ, and optional backfilling delay estimate τ. The missing proportion estimate is used to characterize the degree of missingness of the sampling points within the micro-window that can be restored to valid sample points on the central side under the current reporting strategy and link conditions. Specifically, let the expected number of sampling points of the micro-window be N. exp The number of effective sampling points that can be recovered from the continuity of received batch frames, differential anchor points, and sequence numbers on the central side is N. rec Then ρ is defined as ρ=1 N rec / N exp The edge device can obtain the Nrec based on the transmitted frame sequence number, anchor point reset record, and local expected point count, and write it to the certificate. Alternatively, the central device can recalculate it at the receiving end and write it back for consistency verification. The error interval bound set provides a lower bound f for each quality-sensitive derived feature f. min With upper bound f max .

[0042] In one feasible implementation, for a quality-sensitive feature recorded in the certificate, its feature value under the current reporting caliber is recorded as f1. To ensure that the error interval boundary simultaneously covers the structural uncertainties brought about by quantization discretization, missing links, and differential reconstruction and batch alignment, this embodiment constructs the error radius Δf into a form obtained by superimposing multiple monotonic uncertain terms.

[0043] When uniform quantization is used and mapped to the nearest quantization level using rounding, the upper bound of the absolute error of single-point quantization is q / 2. The missing proportion ρ is used to characterize the degree of missing recoverable valid sample points. To maintain conservatism and interpretability, a missing sensitivity coefficient k, configured according to signal type and feature type, is introduced. ρThis causes the uncertainty caused by the missing value to increase monotonically with ρ and the magnitude of the feature. When the sampled value exceeds the preset range and saturation truncation occurs, a saturation flag is recorded in the certificate, and the quantization error term is replaced by a conservative upper bound that includes the truncation effect.

[0044] When differential compression is used in the reporting and the anchor period Pa is used for resetting, packet loss or resetting will cause the differential reconstruction error to accumulate within the anchor period. To address this, a differential drift coefficient k, configured according to the feature type, is introduced. Pa A drift uncertainty term is constructed using the normalized ratio of Pa under a preset reference anchor point period Pa0, and it increases monotonically with Pa. Pa0 is the reference anchor point period constant in the system configuration, preferably taken as the anchor point period corresponding to the highest link quality level or a preset calibration period, used to normalize Pa under different quality levels.

[0045] When reporting uses a batch window length B and supplementary reporting may occur, window alignment errors can introduce uncertainty for time-sensitive features such as duration and count types. Therefore, a batch alignment coefficient k, configured according to feature type, is introduced. B Furthermore, an alignment uncertainty term is constructed using the normalized ratio of the backfill delay estimate τ to the batch window length B, which is then monotonically changed with τ and B.

[0046] Therefore, for amplitude-based quality-sensitive features, the error radius can be taken as... Sensitive to time structure characteristics, the error radius can be taken as... .

[0047] This leads to one way of constructing the error interval bound: , Where q, Pa, and B are derived from the current reporting strategy and written to the certificate, ρ and τ are obtained based on the current micro-window statistics or link-side estimation, and k ρ k Pa k B The signal type and feature type are preset and can be calibrated using historical data. The reason for adopting the above form is that each uncertainty is monotonically correlated with its source parameter and is configurable and calibrable, thereby obtaining a conservative interval boundary that is calculable and interpretable in engineering without introducing complex statistical assumptions, which can be used for subsequent threshold crossing discrimination and disambiguation control.

[0048] In this embodiment, a braking system-related risk occurs in vehicle BUS0173 during operation as an example. Due to the longer batch window in gear three and packet loss, the number of recoverable valid sample points at the receiving end is reduced on the central side. The vehicle-side or central side calculates the micro-window missing ratio accordingly. The missing ratio is defined as the proportion of the difference between the expected number of sample points and the number of recoverable valid sample points to the expected number of sample points, and writes the missing ratio into the data quality information of the certificate. Taking a micro-window from 10:15:26 to 10:15:27 as an example, this micro-window should theoretically contain 50 sampling points. Under gear two conditions, the central side can recover 45 valid sampling points, so the missing ratio is 0.1. The vehicle-side extracts derived feature parameters within this micro-window, including: peak braking pressure, maximum first-order difference of braking pressure, root mean square of braking vibration, short-time energy of braking vibration, and duration of braking pressure exceeding the threshold. The peak braking pressure of this micro-window is reported as 9.6 after quantization. The vehicle-side constructs the error interval boundary of the peak braking pressure of this micro-window based on the data quality information. When the quantization step size is 0.2, the contribution of the quantization uncertainty is set to 0.1. When the missing proportion is 0.1, to ensure conservatism and interpretability, a missing sensitivity coefficient is configured according to the peak-type feature, and the missing uncertainty is monotonically amplified according to the missing proportion and the feature magnitude to obtain the error radius. This yields a peak error interval bound of 9.26 to 9.94, which is written into the feature certificate. The same logic is used to construct corresponding error interval bounds for other quality-sensitive derived features, and the lower and upper bounds of the intervals are recorded.

[0049] like Figure 3 The flowchart for feature certificate generation and gating basic state determination provided in this application embodiment shows that the operation signal of the smart bus is acquired and a continuous sampling sequence with timestamps is formed. Then, segmentation is performed first, followed by micro-window segmentation, and derived feature parameters are extracted. The derived feature parameters and corresponding data quality information are encapsulated to generate a feature certificate. Based on this, a certificate queue and a background data queue are established, and an error interval boundary is generated. The gating basic state is determined according to the error interval boundary: when the upper boundary of the error interval boundary is less than the discrimination threshold, it is determined to be a confirmed safe state; when the lower boundary of the error interval boundary is greater than the discrimination threshold, it is determined to be a confirmed trigger state; when the discrimination threshold falls within the error interval boundary, it is determined to be a state awaiting disambiguation. This process realizes the entire process from data acquisition and feature extraction to evidence presentation, and establishes a stable and interpretable gating basic determination mechanism through error interval boundary comparison.

[0050] S2. Establish a certificate queue and a background data queue for dual-channel transmission. Perform comparability correction on the derived feature parameters based on the data quality information in the feature certificate to obtain comparable derived features. Obtain the derived feature drift index based on the difference between the comparable derived features and the historical baseline. Obtain the certificate-driven operation and maintenance risk judgment result based on the derived feature drift index. Generate a model input sequence and input it into the edge intelligent prediction model to obtain the model-driven operation and maintenance risk judgment result. The data quality information is used to indicate the quantization accuracy and missing degree of the derived feature parameters under the current reporting strategy. The model input sequence consists of the comparable derived features, the derived feature drift index, and the quality constraint information obtained from the data quality information.

[0051] like Figure 2 The edge intelligent prediction model structure diagram provided in this application embodiment is shown. The edge intelligent prediction model includes a preprocessing module, an input layer, an encoding layer, and an output layer. The preprocessing module is used to parse the feature certificate and perform comparability correction. It obtains three types of information for model input from the feature certificate: a comparable derived feature set to carry comparable derived features obtained by comparability correction of derived feature parameters; a drift information feature set to carry the derived feature drift index obtained by measuring the difference between comparable derived features and historical baselines; and a quality constraint information set to carry quality constraint information generated from the data quality information in the feature certificate. The quality constraint information is used to indicate the quantization accuracy and missing degree of derived feature parameters under the current reporting strategy and the resulting input reliability level. The input layer takes the three types of features mentioned above and arranges them into the most recent consecutive segment sequence as the temporal input; the encoding layer includes a lightweight temporal encoder and a quality information encoder, and fuses the temporal representation and quality representation through a fusion head; the output layer outputs a predicted risk score and a predicted confidence label, where the predicted risk score is used to characterize the probability and intensity of the increase in operation and maintenance risk within the preset prediction time domain, and the predicted confidence label is used to characterize whether the model output is reliable under the current input quality conditions.

[0052] On the vehicle-side edge, two types of logical queues are established in the communication module: a certificate queue and a background data queue. The certificate queue carries the feature certificates generated in step S1. Since the feature certificates only contain derived feature parameters of each micro-window, data quality information, and error interval boundaries, the data volume is small. Therefore, the certificate queue is set to high priority with a fixed sending period to ensure priority scheduling and continuous reporting even during network congestion. The background data queue carries the original sampling sequence or compressed sequence, employing quantization, differential compression, and batch reporting. Its priority is lower than the certificate queue, allowing for delayed sending or batch reporting when bandwidth is sufficient.

[0053] The process of establishing the certificate queue and background data queue is as follows: a first sending queue and a second sending queue are created in the sending buffer of the vehicle communication module, the first sending queue corresponds to the certificate queue, and the second sending queue corresponds to the background data queue; the feature certificate generated in step S1 is encapsulated into a certificate message and written into the first sending queue, and the original sampling sequence or its compressed sequence is encapsulated into a background data message and written into the second sending queue; the sending scheduler retrieves the message to be sent from the first sending queue and the second sending queue according to the preset queue scheduling rules, wherein the preset queue scheduling rules prioritize retrieving messages from the first sending queue for sending.

[0054] It should be noted that the reason for the dual-queue design is as follows: on the one hand, the certificate queue ensures that the central operation and maintenance platform can still complete risk assessment and scheduling based on certificates even when it does not receive a large amount of raw data; on the other hand, the background data queue provides original evidence support for subsequent enhancement requests and model calibration, but will not crowd out the scheduling message channel, thereby ensuring the continuity and availability of operation and maintenance scheduling instructions and risk certificates under weak network conditions.

[0055] To ensure that derived feature parameters obtained under different link tiers, reporting strategies, and quantization and missing structures are statistically comparable, this embodiment first performs comparability correction on the derived feature parameters in the feature certificate to obtain comparable derived features. Different correction formulas are selected for different types of derived features. The following explanation uses root mean square features and duration features as typical examples; other features can be configured using a similar approach.

[0056] Correction for root mean square (RMS) class features: Suppose that the observed value of the derived feature of braking vibration root mean square (RMS) of a certain microwindow is recorded as f in the certificate. （rms） , representing the root mean square value of the observations, including the effects of quantization and compression, under the current reporting strategy. Let the quantization step size given in the data quality information of this channel in step S1 be ΔQ. rms Its physical meaning is: the vibration signal is quantized at equal intervals during sampling, and the interval between each quantization level is ΔQ. rms The corresponding engineering units (e.g., g or m / s²) under the classic uniform quantization noise model can be approximated as the quantization error within the interval […]. ΔQ rms / 2,ΔQ rms The distribution is uniformly distributed on [ / 2], and its variance is... Therefore, the observed root mean square value can be approximated as the superposition of the true root mean square and the variance of the quantization noise, thus allowing us to deduce the estimated value of the true root mean square. This embodiment uses the following correction formula:

[0057] ;

[0058] Among them, fc,(rms) The root mean square feature after comparability correction is used for subsequent drift exponent calculation and model input; f (rms) For the observations recorded in the certificate, ΔQ rms This is the quantization step size field corresponding to the root mean square feature of this channel, recorded when the certificate is generated by S1; the max(·,0) operation is used to ensure that the expression within the parentheses will not have a negative value due to approximation error or special operating conditions, thereby avoiding abnormal square root calculations. The reason for using the above form is that... The theoretical variance, which comes directly from uniform quantization noise, has a clear physical background and does not depend on additional distribution assumptions. Subtracting this variance from the observed root mean square value and then taking the square root can be regarded as a debiased estimate of the true root mean square, thereby improving the comparability of features between different quantization levels.

[0059] Correction for duration-related features: Let f be the observed value of the derived feature "overthreshold duration" in the certificate. (dur) This represents the cumulative time during which the signal within the microwindow exceeds a certain threshold under the current missing structure. For example, if the sampling frequency is 50Hz, and 10 sampling points exceed the threshold within one second, then f... (dur) =10×(1 / 50)=0.2s. Because missing samples can cause segments exceeding the threshold to go unobserved, resulting in a systematic underestimation of the duration-type feature, compensation is needed based on the missing proportion. Let the missing proportion field recorded in the certificate for this micro-window be... This represents the missing number between the expected number of sample points and the actual number of sample points received. For example, if only 45 out of 50 expected sample points are received, then... =0.10. Under the simple assumption that the sampling loss is approximately uniform over time, the actual duration can be approximated as the inverse relationship between the observation duration and the effective sampling ratio. Therefore, this embodiment adopts the following correction formula:

[0060] ;

[0061] Where f c,(dur) For the duration characteristics after comparability correction; f (dur) The duration of observation recorded in the certificate; The missing percentage of certificate records is calculated by comparing the expected number of sampling points for this micro-window with the actual number of sampling points in S1; ε dur The lower limit constant of the denominator for duration correction is a positive number less than 1, used to prevent errors when the missing percentage is extremely high. The reason for adopting the above form to avoid numerical divergence caused by proximity to 0 is that, under the condition that the missing data is relatively mild and approximately uniform, the systematic underestimation of duration-type features can be compensated in the first order by dividing by the effective sampling ratio; at the same time, the lower limit of the denominator is used to control extreme missing conditions, ensuring numerical stability and engineering feasibility.

[0062] For other types of derived features, such as energy, peak value, and statistical amplitude, this embodiment can select the corresponding correction formula according to the physical mechanism by which they are affected by quantization and missing values. As long as sufficient quantization and missing value-related parameters are recorded in the certificate, similar comparability processing can be supported. In this embodiment, all derived features of a microwindow are corrected in the above manner and denoted as the corrected feature vector. , where n is the dimension of the microwindow-derived feature.

[0063] Establish historical baselines and calculate derived feature drift indices. Historical baselines can be established by vehicle, signal type, and operating condition type. Typically, the baseline mean μ for each correction feature is statistically obtained under healthy operating conditions. i Compared with baseline standard deviation σ i To ensure that the drift metric also reflects the quality of evidence, this embodiment introduces a feature confidence weight w. i The drift index is defined using a weighted normalized distance:

[0064] ;

[0065] in, For the i-th correction feature, ε1 is the lower limit constant of the denominator, the specific value of which shall be determined by those skilled in the art to avoid σ. i The small size leads to numerical instability. This form is used because the normalization term ( μ i ) / (σ i +ε1) enables unified comparison of features with different dimensions, weight w i By injecting certificate quality information into the drift metric, the contribution of low-quality features to drift is automatically suppressed, and the whole is used as a distance-type indicator to facilitate thresholding and state machine control.

[0066] Feature confidence weight W i,ref This is a reference width for the feature at an acceptable quality level, such as a reference value calculated when the network is good, the quantization is fine, and there are almost no missing values; W iLet be the width of the error interval for the i-th quality-sensitive feature in the current certificate. This width is determined by the error interval boundary derived in step S1 based on quantization and missing error, and is numerically equal to twice the error radius. It is used to quantify the uncertainty of the current feature value. `clip` is a truncation function that restricts the weights to the range of 0 to 1. This form is used because a wider interval indicates greater uncertainty in the evidence, thus requiring a lower weight, and linear truncation is simple and easy to tune in engineering.

[0067] After obtaining the derived feature drift index and the error interval boundaries of each key discriminant quantity in the feature certificate, this embodiment obtains the certificate-driven operation and maintenance risk assessment result as follows. First, it should be noted that key discriminant quantities refer to risk discrimination indicators that can be directly used for operation and maintenance decisions. They are obtained from derived feature parameters through a preset mapping and have a clear correspondence with specific fault risks. Key discriminant quantities can be one or more of the following: peak value, sudden change, energy, or duration-based discriminant quantities, such as peak braking temperature rise, sudden change in braking pressure intensity, vibration energy, and duration exceeding the threshold. For each type of key discriminant quantity, a corresponding discrimination threshold is preset. The discrimination threshold can be jointly determined based on vehicle manufacturer technical specifications, operation and maintenance experience rules, and historical fault statistics, and configured separately by vehicle model, signal type, or operating condition category to ensure that the threshold matches the actual vehicle operating conditions.

[0068] If the upper bound of the error interval is less than the discrimination threshold, the key discriminant is determined to be in a safe state, meaning that even considering the worst-case deviation caused by quantization error and missing error, the discriminant is still on the safe side of the threshold, and therefore a safe judgment can be made without immediately supplementing evidence. If the lower bound of the error interval is greater than the discrimination threshold, the key discriminant is determined to be in a triggered state, meaning that even considering the worst-case deviation, the discriminant is still on the triggered side of the threshold. If the discrimination threshold is in the error interval, the key discriminant is determined to be in a state of pending disambiguation, meaning that under the current data quality conditions, it is impossible to determine which side of the threshold the true value of the discriminant is on, and the interval needs to be narrowed by supplementing with the minimum necessary accuracy before a final judgment is made.

[0069] In this embodiment, a braking system-related risk occurs in vehicle BUS0173 during operation as an example. Further, the vehicle-side or central side performs comparability correction on the derived feature parameters in the certificate. For the root mean square (RMS) feature of braking vibration, the quantization noise is corrected based on the quantization step size recorded in the micro-window data quality information to obtain the corrected RMS value. For the over-threshold duration feature, the observation duration is compensated for in the first order based on the missing proportion, and a lower limit constant for the denominator is set to obtain the corrected duration value. A corrected feature vector is formed after correcting all derived features within the same micro-window. The central side establishes historical baselines for vehicle, signal type, and operating condition category. The historical baselines are obtained by statistically analyzing healthy operating condition samples to obtain the baseline mean and baseline dispersion of each corrected feature. To suppress the influence of low-quality evidence on drift measurement, the system calculates feature confidence weights based on the error interval boundary width; the wider the error interval boundary, the lower the feature weight. The system constructs a drift index using normalized difference and confidence weights. The drift index increases with the degree of deviation and is used for subsequent priority routing. In this embodiment, the peak braking pressure is used as one of the key discrimination parameters, and a corresponding discrimination threshold of 9.8 is configured. The system determines the basic gating state by comparing the error interval boundary with the discrimination threshold. Since the error interval boundary of the peak braking pressure of this micro-window is 9.26 to 9.94, and the discrimination threshold of 9.8 falls within the error interval boundary, this key discrimination parameter is determined to be in a state requiring disambiguation. At the same time, the system determines the processing priority by comparing the drift index with the drift threshold. The drift threshold includes a trend enhancement boundary and a significant anomaly boundary, both of which are boundary values ​​defined for the same drift index, used to distinguish between deviations into the trend enhancement interval or into the significant anomaly interval. In this embodiment, the current segment drift index exceeds the trend enhancement boundary but does not exceed the significant anomaly boundary, so the system marks this event requiring disambiguation as a high-priority disambiguation event and shortens the recovery time.

[0070] After obtaining the above three basic states, in order to further distinguish between significant overall deterioration trends and occasional fluctuations and determine operational priorities, this embodiment introduces a derived feature drift index and a drift threshold for joint discrimination. The drift threshold uses healthy operating condition data of the same vehicle model and the same signal type as the baseline sample. The healthy operating condition data can come from the new vehicle acceptance period, the operating interval with no faults and no work orders in a recent period, or the healthy labeled interval confirmed by operation and maintenance. The drift index is calculated on the baseline sample in the same way as online, forming a healthy baseline drift index distribution. After removing obvious outliers and abnormal operating condition segments, at least two drift thresholds are determined from this distribution, corresponding to the trend enhancement boundary and the significant anomaly boundary, respectively. The threshold is preferably taken as the high quantile of the healthy baseline distribution as the boundary value, and can be grouped according to vehicle type and seasonal operating conditions (for example, the 95th percentile can be taken as the trend enhancement boundary and the 99th percentile as the significant anomaly boundary). During operation, the drift threshold can be updated using a rolling window to the healthy baseline samples to adapt to component aging, seasonal changes, and line differences. However, the update must meet minimum sample size and consistency constraints to avoid threshold drift due to short-term noise. Based on the drift index and drift threshold, further operational routing is performed on the three basic states. The trend enhancement boundary and the significant anomaly boundary are both graded thresholds set for the drift index of derived features with the same calculation caliber. They correspond to the same granularity of the evaluation unit, and the significant anomaly boundary is larger than the trend enhancement boundary, used to grade the degree of drift.

[0071] For a confirmed safe state, when the drift index does not exceed the trend enhancement threshold, the output operation is normal monitoring, indicating that the regular sampling and certificate reporting cycle is maintained, no enhancement request is triggered, and no work order is entered. When the drift index exceeds the trend enhancement threshold, the output operation is trend attention, indicating that although the key discriminant is still considered safe in the interval sense, the overall deviation from the healthy baseline has been significantly enhanced. The vehicle or signal is included in the trend attention queue, and at least one enhancement measure is taken, such as increasing the certificate reporting frequency, increasing the observation weight of the key discriminant, or generating a weak early warning prompt in advance to reserve background resources, but no mandatory maintenance work order is directly generated.

[0072] For a confirmed trigger state, when the drift index does not exceed the significant anomaly threshold, the output operation is a stable trigger, indicating that the key discriminant has been determined to have crossed the threshold in an interval sense, but the overall drift has not yet shown as accelerated degradation or sudden escalation. The system enters the confirmation dispatch candidate and executes the anti-jitter state mechanism. The anti-jitter state mechanism confirms the persistence of the trigger by accumulating the number of triggers (for example, if the number of confirmed trigger states in the sliding window of the most recent 10 segments reaches 6 or more, the trigger is confirmed and a work order is generated or maintained; the trigger is only allowed to be canceled when the number of confirmed trigger states in the most recent 10 segments is less than 2 and there is no state to be disambiguated). After the conditions are met, an operation and maintenance task work order is generated or maintained. When the drift index exceeds the significant anomaly threshold, the output operation is a high-priority trigger, indicating that the key discriminant has been determined to have crossed the threshold and the overall drift is significantly abnormal. The system marks this event as a high-priority confirmation dispatch candidate. Under the premise that the anti-jitter conditions are met, a work order is generated with a more urgent handling time limit, and higher-priority evidence supplementation can be triggered at the same time to lock the fault mode and assist in localization, thereby improving the dispatch accuracy.

[0073] When dealing with the disambiguation state, if the drift index does not exceed the trend enhancement threshold, the output operation is low-priority disambiguation, indicating that the current evidence is insufficient to determine the cross-threshold but the overall drift has not significantly increased. The system only triggers a minimum enhancement request to narrow the interval. The minimum enhancement request is limited to the uncertain micro-window range and adopts a preset minimum complement accuracy method, such as requesting data with a smaller quantization step size within the range or requesting short-term original subsequences within the range. When the drift index exceeds the trend enhancement threshold, the output operation is high-priority disambiguation, indicating that the current evidence is insufficient and the overall deviation from the healthy baseline has significantly increased. The system triggers a high-priority minimum enhancement request, and accelerates evidence convergence by shortening the complement time, increasing the complement accuracy priority of key micro-windows, or increasing the certificate reporting frequency, so as to quickly transform the state to be disambiguated into a certain safe or certain triggered state and enter the corresponding subsequent process.

[0074] In this embodiment, a braking system-related risk occurs in vehicle BUS0173 during operation as an example. For this high-priority disambiguation event, the central side sends an enhancement request to the vehicle. The enhancement request is limited to an uncertain micro-window range, and the following objects are adjusted simultaneously to promote evidence convergence: First, the backfilling range is adjusted, limiting the backfilling micro-window index range to the micro-window itself and one micro-window immediately before and after it, avoiding bandwidth pressure caused by full-fragment backfilling. Second, the backfilling accuracy is adjusted, requesting a smaller quantization step size for the braking pressure signal backfilling, adjusting the quantization step size from 0.2 to 0.05, and requesting the corresponding differential anchor point reset within this range to reduce differential reconstruction drift. Third, the backfilling type is adjusted, requesting short-time original subsequence backfilling for the braking pressure signal, with a backfilling length of one second, so that the central side can recalculate peak values ​​and abrupt change characteristics and narrow the error interval. Fourth, the queue parameters are adjusted, temporarily reducing the maximum allowable accumulated cache size of the background data queue and increasing the queue priority attribute of the enhanced backfilling data, so that the backfilling data is enqueued first and dequeued as quickly as possible for uploading. After the data is replenished, the central system recalculates the peak braking pressure of the micro-window. If the recalculated peak value is 9.9, and the new error interval bound constructed under conditions of finer quantization and significantly reduced missing data is 9.86 to 9.94, then the lower bound of this error interval bound is greater than the discrimination threshold of 9.8, and the gating base state converges from pending disambiguation to a confirmed trigger state. The system enters the confirmation dispatch candidate and executes the anti-jitter state mechanism. The trigger persistence is confirmed by determining the number of times the trigger state occurs within the sliding window. After the conditions are met, an operation and maintenance task work order is generated or maintained, and work order assignment parameters are formed, including response time limit, skill tag, and vehicle entry suggestion. After receiving the replenished data, the central system recalculates the true peak value of the micro-window and reconstructs the error interval bound with a finer quantization step size and more complete sample points, significantly narrowing the error interval bound width.

[0075] like Figure 4 As shown in the flowchart of the handling priority determination based on the drift index provided in this application embodiment, after determining the gating base state, the handling priority is determined based on the derived feature drift index. First, it is determined whether the derived feature drift index does not exceed the trend enhancement boundary; if the derived feature drift index does not exceed the trend enhancement boundary, the handling level of the event corresponding to the gating base state is maintained; if the derived feature drift index exceeds the trend enhancement boundary but does not exceed the significant anomaly boundary, the handling priority of the event corresponding to the gating base state is increased; if the derived feature drift index exceeds the significant anomaly boundary, the event is marked as a high-priority handling event. Through the above hierarchical logic, the dynamic linkage between handling priority and system drift degree is realized, which can both ensure the stability of safety monitoring and prioritize the response to key risks during the deterioration acceleration phase.

[0076] The edge intelligent prediction model provides model-driven operational risk assessment results. This model is trained offline and deployed in the vehicle's edge computing unit. The model input consists of comparable derived features obtained from feature certificate parsing, along with their corresponding data quality information, ensuring that the model's inference does not depend on the continuous uploading of the original sequence.

[0077] The training data is derived from a joint sample of historical vehicle operation data and maintenance records. Historical operation data includes feature certificates generated according to the steps outlined in this application, derived features after comparability correction, and corresponding data quality information. Maintenance records include one or more of the following: fault repair work orders, key component replacement records, alarm confirmation results, or manual review conclusions. To ensure consistency between the training samples and the online inference input, the offline training phase employs the same feature extraction, comparability correction, and drift index calculation process as the online phase. Continuous segments are grouped into time series samples of a preset input length. Whether a confirmed maintenance-verified fault event occurs within a preset prediction time domain or whether a certificate-gated trigger state is entered is used as the supervision label, thus forming binary training samples of "high risk / non-high risk." Each training sample simultaneously contains an input sequence and a supervision target. The supervision target includes a high-risk or non-high-risk category label and a target prediction risk score corresponding to that category label. High-risk corresponds to a target prediction risk score of one, and non-high-risk corresponds to a target prediction risk score of zero.

[0078] The model structure employs a lightweight temporal modeling network to adapt to the vehicle's computing power. This network can be one of the following: a gated recurrent unit network, a long short-term memory network, a one-dimensional temporal convolutional network, or a lightweight attention network. The model input is a sequence formed by concatenating comparable derived features, drift information, and quality constraint information from multiple recent consecutive segments in chronological order. The model output is the predicted risk score. The training objective is to make the predicted risk score approach high values ​​on high-risk samples and low values ​​on non-high-risk samples, and to select thresholds through a validation set to meet preset false positive and false negative constraints. After training, the model parameters are fixed and distributed to the vehicle's edge computing unit. The vehicle only performs inference calculations and does not perform large-scale reverse training on the vehicle side.

[0079] To adapt to vehicle aging, seasonal changes, and route differences, the model allows for periodic retraining or small-step updates on the central side based on newly labeled samples. After meeting consistency constraints, the updated model parameters are distributed to the vehicle in batches. The consistency constraints include that the false alarm rate on healthy baseline samples does not increase and the recall rate on faulty samples does not decrease before and after the update, and there is no systematic conflict between the model output and the certificate gating conclusion.

[0080] The vehicle-mounted edge computing system reads derived feature parameters from the feature certificate and performs comparability correction to obtain comparable derived features. Simultaneously, it reads data quality information and error interval boundaries to generate quality constraint information characterizing input reliability. The model input consists of three types of information: the first type is comparable derived features, reflecting risk characteristics such as peak values, abrupt changes, energy, and duration within the current segment or micro-window; the second type is drift information, including at least the derived feature drift index, and optionally including the direction or intensity of change of the drift index in the most recent consecutive segments, used to characterize the deterioration trend relative to the historical baseline; the third type is quality constraint information, preferably derived from fields such as error interval boundary width, missing structure related indicators, anchor point period, and batch window, used to inform the model of the current input uncertainty level, enabling the model to adopt conservative outputs or reduce confidence for low-quality inputs.

[0081] To meet the temporal requirements of the prediction, this embodiment adopts a sliding time series organization method for the model input: the most recent K consecutive segments are used as an inference sample, and the above three types of information are spliced ​​into an input sequence in chronological order. The input length K is determined by configuration and matches the rate of change of vehicle operating conditions.

[0082] It should be noted that the output of the edge intelligent prediction model includes at least the predicted risk score R. pred With prediction confidence markers. Predicted risk score R pred To quantify the probability and intensity of risk escalation within the future prediction time domain, it is preferably normalized to the [0,1] interval. For ease of engineering implementation and scheduling rule configuration, this embodiment uses R... pred By a single threshold θ H Mapped to a two-level predicted risk level: when R pred ≥θ H When defined as high risk; when R pred <θ H The time frame is defined as non-high risk. Threshold θ H It can be configured according to vehicle model, signal type and operating condition category, and can be calibrated or updated on a rolling basis based on the statistical distribution of historical fault samples and healthy samples.

[0083] Prediction confidence markers are used to characterize whether the model output is stable and reliable under the current input quality conditions. This embodiment defines prediction confidence markers as either usable or unusable, where usable indicates that the following requirements are simultaneously met: intrinsic confidence requirement and input quality gating requirement. The intrinsic confidence requirement includes at least the maximum posterior probability p of the model output. max Not less than the threshold θ p And the highest probability and the second highest probability p 2nd The difference (p) max p 2nd Not less than the threshold θg Input quality gating requirements must include at least: the missing percentage ρ is not greater than ρ max Furthermore, the error interval width W corresponding to the key discriminant is not greater than W. max When all the above conditions are met, the prediction confidence label is considered available; otherwise, it is considered unavailable. Where θ H， ,θ p ,θ g ,ρ max W max All parameters are configurable, stored in groups by vehicle type / signal / operating condition and obtained by table lookup at runtime. Their initial values ​​are determined by offline calibration and are allowed to be updated on a rolling basis while meeting consistency constraints.

[0084] In this embodiment, a braking system-related risk occurs in vehicle BUS0173 during operation as an example. At the same time period, the vehicle-side system sequentially combines comparable derived features, drift information, and quality constraint information from multiple recent consecutive segments into a model input sequence, which is then input into the edge intelligent prediction model to output a predicted risk score. The predicted risk score is normalized to zero to one and mapped to high or low risk using a risk threshold. The predicted confidence flag is determined as available or unavailable based on the model's inherent confidence requirements and input quality gating requirements. In this embodiment, when the certificate gating has converged to a confirmed trigger and the model output is high risk with the predicted confidence flag being available, the central side increases the response time of the candidate work order by one level and issues resource pre-allocation parameters in advance for preparing spare parts and scheduling skilled personnel. It should be noted that the model output is not used to lower the certificate gating conclusion; the model is only used for advance suggestions and resource pre-allocation.

[0085] like Figure 5 As shown in the certificate-driven and model-driven joint scheduling flowchart provided in this application embodiment, the operation and maintenance risk assessment result of certificate-driven scheduling is used as the dispatch gating basis, and the operation and maintenance risk assessment result of model-driven scheduling is used as the advance input. When the model-driven scheduling determines that the event is high-risk and the prediction confidence mark is available, the event is marked as a high-priority event, and resource pre-allocation parameters are generated; when the model-driven scheduling determines that the event is not high-risk or the prediction confidence mark is unavailable, the default handling level is maintained. Through the collaborative judgment of certificate-driven and model-driven scheduling, the dynamic integration of evidence-based stable dispatch and prediction-based advance pre-allocation is achieved, which can maintain the continuity and foresight of scheduling decisions under weak network and incomplete information conditions.

[0086] S3. Based on the certificate-driven and model-driven operation and maintenance risk assessment results, generate operation and maintenance task priority queues and work order assignment parameters as operation and maintenance scheduling results.

[0087] The operation and maintenance task priority queue uses vehicle identifier, signal type identifier, and risk event identifier as queue element keys. Each queue element is associated with at least certificate gating status, drift level, disambiguation priority, replenishment time limit, and model prediction risk level and prediction confidence flag. The work order assignment parameters include at least response time limit, work confirmation level, suggested spare parts label, suggested skill label, review priority, and optional assignment site or team constraints, which are used by the central side to generate or update work orders and issue scheduling instructions.

[0088] It should be noted that in this embodiment, the certificate-driven operation and maintenance risk assessment results and the model-driven operation and maintenance risk assessment results are output in parallel but at different decision levels: the certificate-driven operation and maintenance risk assessment results serve as the basis for dispatch gating, used to determine whether to enter the dispatch candidate, whether to generate or maintain a work order, and whether to trigger disambiguation and replenishment; the model-driven operation and maintenance risk assessment results do not directly generate or cancel work orders, but their role is limited to advance suggestions and scheduling parameter corrections, used to adjust response time limits, resource pre-allocation and review priorities under the premise that the certificate gating conclusion remains unchanged, and used to determine the priority and replenishment time limit of the disambiguation to be handled, thereby realizing the collaborative work of stable certificate-side certification and dispatch and advance preparation on the model side.

[0089] When the certificate-driven determination is confirmed as triggered, the process enters the confirmation dispatch candidate and executes the anti-jitter state mechanism. At this time, the model output is used for advance support: when the predicted risk level is high risk and the predicted confidence mark is available, the response time limit requirement of the candidate work order is increased without changing the certificate gating conclusion, the suggested spare parts label and suggested skill label are generated or increased, and the review priority is increased; when the predicted risk level is not high risk or the predicted confidence mark is unavailable, the dispatch gating conclusion is not reduced, but the work order is marked for quick review and a minimum enhancement backfill limited to the critical micro-window is triggered once, so as to further lock the failure mode through evidence convergence and reduce the cost of mis-dispatch.

[0090] When the certificate-driven determination indicates safety, the system does not generate a mandatory maintenance work order and defaults to outputting normal monitoring or trend monitoring. In this case, the model output is used for advance warning suggestions: when the predicted risk level is high and the predicted confidence markers are all available in the most recent K consecutive assessment periods, an early warning is output with no work assignment and resource pre-allocation parameters, triggering a review enhancement to verify whether the trend can be substantiated on the certificate side, thus avoiding direct maintenance driven by the prediction conclusion but still allowing for advance resource preparation; when the predicted risk level is not high risk or the predicted confidence marker is unavailable, the system does not upgrade to a work assignment candidate based on the model output, but only maintains the original monitoring strategy or includes the vehicle or signal in the trend monitoring queue, adjusting the certificate reporting frequency to a preset higher frequency level or switching the assessment time granularity to a smaller micro-window configuration.

[0091] Among them, the resource pre-reservation parameters include at least a pre-reservation spare parts list, pre-reservation skill tags, and suggested review time limits, which are used to complete resource matching and scheduling preparation in advance without generating work orders; K is a configurable parameter, which is stored in groups by vehicle type / signal / operating condition and obtained by looking up a table at runtime. Its initial value is determined by offline calibration and is allowed to be updated on a rolling basis when consistency constraints are met.

[0092] When the certificate-driven decision indicates a need for disambiguation, the system does not generate a work order. Instead, it outputs disambiguation handling and issues a minimum enhancement request to converge the error interval boundary. In this scenario, the model output is used to determine the disambiguation priority and recovery time: when the predicted risk level is high and the predicted confidence marker is available, the system elevates disambiguation to high priority and shortens the recovery time; when the predicted risk level is not high or the predicted confidence marker is unavailable, the system maintains low-priority disambiguation or continues observation. After augmented data recovery, the feature certificate is regenerated, the error interval boundary and drift index are updated, and the certificate gating decision is re-executed to bring the state awaiting disambiguation to either confirmed safety or confirmed triggering, thus completing the minimum main process of evidence collection—decision—scheduling in a closed loop. Through the above scheduling arrangement, the final operation and maintenance scheduling result output in this embodiment includes at least two parts: the first part is the work order and assignment parameters, determined by the certificate-driven gating and anti-jitter state machine; the second part is the lead time suggestion and resource pre-allocation parameters, determined by the prediction model output and used to adjust monitoring density, review priority, and handling preparation. This enables certificate-driven stable certification and dispatching to work in tandem with model-driven advance warning and preparation, suppressing fluctuations in risk assessment and operational scheduling jitter under conditions of weak network and changes in reporting strategies.

[0093] In Specific Implementation Example 2, based on Specific Implementation Example 1, the augmentation process of the enhancement request under the disambiguation state is further constrained by a two-stage approach, and a clear stopping condition is introduced. This enables the system to achieve faster evidence convergence with a smaller amount of augmentation data under weak network conditions, avoiding unnecessary original data augmentation that could cause bandwidth occupation and scheduling jitter.

[0094] It should be noted that the two-stage disambiguation enhancement in this embodiment means that when the certificate gating determines that it needs to be disambiguated, the system does not directly request the original sequence, but instead prioritizes low-cost accuracy enhancement to narrow the error interval. Only when the low-cost enhancement still fails to make the gating converge, it is upgraded to short-time original subsequence backfilling, and the subsequent backfilling is terminated immediately after the convergence condition is met at any stage.

[0095] After receiving the feature certificate and completing the certificate-driven gating determination, if a key discriminant is in a state awaiting disambiguation, the central side generates an enhancement request and sends it to the vehicle edge side. The enhancement request adds an enhancement stage identifier E to the first embodiment, whereby the enhancement stage identifier indicates whether the current enhancement belongs to the first or second stage.

[0096] When the enhancement stage identifier E=1, the original sampling points are not re-added at the vehicle edge. Instead, the derived features or necessary statistical summaries related to the key discriminant are regenerated for the micro-window range specified in the enhancement request using a finer quantization step size, and the first-stage enhancement certificate is generated and sent back. The finer quantization step size is preferably obtained by reducing the quantization step size in the current certificate according to a preset ratio, for example, q. fine =α×q, where 0<α<1 is a configurable scaling factor. The first-stage enhancement certificate must include at least: enhanced key discriminant observations, and corresponding data quality information (including q). fine (and missing proportion estimation) and q-based fine The error interval bound is recalculated, thereby significantly reducing quantization uncertainty and accelerating interval convergence without backfilling the original sequence. After receiving the first-stage enhancement certificate, the central side re-executes the certificate gating judgment: if the updated error interval bound meets the convergence condition of "the upper bound of the error interval bound is less than the discrimination threshold" or "the upper bound of the error interval bound is greater than the discrimination threshold", thus converging the state to be disambiguated to a certain safety or a certain trigger, then the subsequent enhancement of the event is terminated and the corresponding process is entered; if the discrimination threshold is still within the error interval bound, then the first stage is determined to have not converged, and the central side sends a second-stage enhancement request E=2 to the vehicle end.

[0097] When the enhancement stage identifier E=2, the vehicle-side edge retrieves the short-time original subsequence within the micro-window range specified in the enhancement request from the circular buffer and sends it back to the central side. To ensure minimal back-up, the length of the short-time original subsequence is limited to no more than the preset maximum back-up duration (defined by those skilled in the art) and is limited to the range of the critical micro-window index that leads to the disambiguation. After receiving the backed-up original subsequence, the central side recalculates the critical discrimination quantity and reconstructs the error interval boundary or directly obtains the high-precision discrimination quantity value, which is used to complete the final gating decision and fault mode locking.

[0098] The central side regenerates or updates the feature certificate based on the second-stage backfilling results and re-executes the certificate gating judgment to converge the pending disambiguation state to either "definitely safe" or "definitely triggered," subsequently terminating the backfilling process for that event. If the convergence is to "definitely triggered," the system proceeds to the anti-jitter confirmation and dispatching process of Implementation Example 1; if the convergence is to "definitely safe," normal monitoring or trend monitoring resumes. The model-driven judgment results are still only used for advance suggestion and disambiguation priority selection and do not replace the certificate gating conclusion in generating work orders. When the backfilled data is still insufficient to form a definitive judgment, the system enters the fallback handling process.

[0099] As a fallback procedure, if the second-stage remediation still results in a pending disambiguation, the central side will not generate a work order, but will output a conservative disambiguation maintenance status and execute at least one of the following measures: First, expand the remediation window and increase the preset maximum remediation duration to the preset upper limit, but not exceeding the single remediation data volume constraint; Second, escalate the event to high-priority disambiguation and shorten the remediation time limit, and repeat the second-stage remediation without exceeding the original remediation retries limit. Through this fallback procedure, forced dispatching or frequent cancellations are avoided when there is insufficient evidence, thereby maintaining scheduling stability and interpretability.

[0100] Through the above two-stage disambiguation enhancement and early termination mechanism, this embodiment further achieves the following compared to Embodiment 1: in the case of disambiguation, priority is given to enhancing the convergence interval boundary with low cost and high accuracy, and the short-term original subsequence is only replenished when necessary, thereby significantly reducing the bandwidth occupation of replenishment, shortening the evidence convergence delay, and reducing the operation and maintenance scheduling jitter and resource waste caused by repeated replenishment under weak network conditions.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart bus operation and maintenance task scheduling method based on an edge intelligent prediction model, characterized in that, Includes the following steps: S1. Acquire the operation signal of the smart bus and form a continuous sampling sequence with timestamps. Divide the continuous sampling sequence into segments according to a preset time window, extract the derived feature parameters used for operation and maintenance risk identification and prediction in each segment, and encapsulate the derived feature parameters and the data quality information corresponding to the derived feature parameters to generate a feature certificate. S2. Based on the data quality information in the feature certificate, perform comparability correction on the derived feature parameters to obtain comparable derived features. Then, obtain the derived feature drift index based on the difference between the comparable derived features and the historical baseline. Based on the derived feature drift index, obtain the certificate-driven operation and maintenance risk assessment result. Generate a model input sequence and input it into the edge intelligent prediction model to obtain the model-driven operation and maintenance risk assessment result. The data quality information is used to indicate the quantization accuracy and missing value of the derived feature parameters under the current reporting strategy. The model input sequence consists of the comparable derived features, the derived feature drift index, and quality constraint information obtained from the data quality information. Establish a certificate queue and background. The data queue is carried out in a dual-channel manner. The specific process is as follows: a certificate queue and a background data queue are established separately, and the certificate queue is set to a higher sending priority than the background data queue. The certificate queue is used to cache and prioritize the sending queue of feature certificates, so that the derived feature parameters can be compared and corrected based on the data quality information in the feature certificates. The background data queue is used to cache and send the original sampling sequence or compressed sampling sequence of the running signal when the bandwidth allows. When the certificate driver determines that it is in a state to be disambiguated, it is used to enhance the request backfilling, update the feature certificates and converge the error interval boundary. During the comparison correction process, different correction methods are selected for different types of derived feature parameters. S3. Based on the certificate-driven operation and maintenance risk assessment results and the model-driven operation and maintenance risk assessment results, generate an operation and maintenance task priority queue and work order assignment parameters as operation and maintenance scheduling results.

2. The intelligent public transport operation and maintenance task scheduling method based on the edge intelligent prediction model as described in claim 1, characterized in that: The derived feature parameters extracted from each segment for operational risk identification and prediction include: Each segment is divided into micro-windows according to a preset micro-window length; Within each micro-window, one or more of the following are performed based on the original sampling points covered by the micro-window: statistical calculation, difference calculation, energy calculation, or duration discrimination, to obtain derived feature parameters for operation and maintenance risk discrimination and prediction inference. The derived feature parameters include quality-invariant derived features and quality-sensitive derived features.

3. The intelligent public transport operation and maintenance task scheduling method based on the edge intelligent prediction model as described in claim 1, characterized in that: The step of encapsulating the derived feature parameters and the corresponding data quality information to generate a feature certificate specifically involves: A feature certificate is generated for each micro-window with micro-window as the smallest granularity. The fragment identifier is written into the feature certificate and the position index of the micro-window within the fragment is recorded, thereby establishing an association between the feature certificate and the fragment identifier. The feature certificate consists of a certificate header and a certificate payload. The certificate header is used for cross-end consistent positioning, and the certificate payload is used to carry derived feature parameters and data quality information corresponding to the derived feature parameters. The data quality information includes the current link quality level and upload strategy parameters.

4. The intelligent public transport operation and maintenance task scheduling method based on the edge intelligent prediction model as described in claim 3, characterized in that, The process of generating the feature certificate also includes the following steps: Obtain the feature value of each quality-sensitive derived feature under the current reporting strategy. The error radius is determined based on the feature values ​​and the data quality information recorded in the feature certificate. The error radius is determined according to the feature type, which includes amplitude-type quality-sensitive features and time-structure-sensitive features. The lower and upper bounds of the error interval are determined based on the feature value and the error radius, and the lower and upper bounds are written into the feature certificate.

5. The intelligent public transport operation and maintenance task scheduling method based on the edge intelligent prediction model as described in claim 4, characterized in that, After obtaining the error interval boundary, the gating base state for determining operational risk assessment is determined based on the error interval boundary, including: The basic gating states include a determined security state, a determined trigger state, and a state to be disambiguated. For a pre-configured key discrimination quantity, obtain the error interval boundary and discrimination threshold corresponding to the key discrimination quantity, and compare the error interval boundary with the discrimination threshold; When the upper bound of the error interval is less than the discrimination threshold, the key discrimination quantity is determined to be in a safe state. When the lower bound of the error interval is greater than the discrimination threshold, the key discrimination quantity is determined to be in a confirmed trigger state. When the discrimination threshold is between the lower and upper bounds of the error interval, the key discrimination quantity is determined to be in a state to be disambiguated. The key discriminant quantity refers to the risk indicator obtained by derived feature parameters according to a preset mapping relationship, and the discrimination threshold refers to the pre-configured gate threshold for each key discriminant quantity.

6. The intelligent public transport operation and maintenance task scheduling method based on an edge intelligent prediction model as described in claim 5, characterized in that, After obtaining the basic gating state, the processing priority of the basic gating state is determined based on the derived feature drift index, specifically as follows: The derived feature drift index is compared with a pre-configured drift threshold, which includes a trend enhancement boundary and a significant anomaly boundary. The trend enhancement boundary is used to distinguish between normal fluctuations and continuous enhancement deviations under a healthy baseline, and the significant anomaly boundary is used to distinguish between continuous enhancement deviations and significant anomaly deviations. When the derived feature drift index does not exceed the trend enhancement boundary, the processing priority level corresponding to the gating base state is maintained. When the derived feature drift index exceeds the trend enhancement boundary but does not exceed the significant anomaly boundary, the processing priority of the event corresponding to the gated basic state is increased. When the derived feature drift index exceeds the significant anomaly boundary, the event corresponding to the determined trigger state or the state to be disambiguated will be marked as a high-priority event.

7. The intelligent public transport operation and maintenance task scheduling method based on the edge intelligent prediction model as described in claim 1, characterized in that: The specific process for obtaining the model-driven operation and maintenance risk assessment result based on the edge intelligent prediction model is as follows: The model input sequence is fed into the edge intelligent prediction model, and the edge intelligent prediction model outputs a predicted risk score and a predicted confidence label. The predicted risk level is determined based on the predicted risk score. The predicted risk level includes high risk and non-high risk, and the predicted confidence marker includes availability and unavailability; When the predicted risk score is higher than the preset risk threshold, it is judged as high risk; when the predicted risk score is not higher than the preset risk threshold, it is judged as not high risk. When the output of the edge intelligent prediction model meets the preset model intrinsic confidence requirement and the data quality information meets the preset input quality gating requirement, the prediction confidence flag is determined to be available; otherwise, it is determined to be unavailable. The predicted risk level and the predicted confidence level are output as the model-driven operation and maintenance risk assessment results.

8. The intelligent public transport operation and maintenance task scheduling method based on the edge intelligent prediction model as described in claim 1, characterized in that: The operation and maintenance scheduling result is jointly generated by the certificate-driven operation and maintenance risk assessment result and the model-driven operation and maintenance risk assessment result, including: When the gating base state is a confirmed trigger state, it enters the dispatch candidate and generates or maintains a work order. When the basic state of the gate control is a determined safe state, no mandatory maintenance work order is generated; When the basic state of the gating is pending disambiguation, no work order is generated and the enhancement request is triggered to compensate. Once the basic state of the gating is determined, the model-driven operation and maintenance risk assessment results are used as advance suggestions to adjust the handling priority and response time. When the model-driven assessment is high risk and the prediction confidence mark is available, the event priority is increased and resource pre-occupancy parameters are generated. When the model-driven assessment is not high risk or the prediction confidence mark is unavailable, the default handling level is maintained.

9. The intelligent public transport operation and maintenance task scheduling method based on the edge intelligent prediction model as described in claim 8, characterized in that: The enhancement request recovery process also includes the following steps: Two-phase disambiguation enhancement is performed during the enhancement request recovery process; The first stage of disambiguation enhancement involves generating an enhanced feature certificate only for the micro-window range corresponding to the state to be disambiguated after receiving an enhancement request feedback instruction for the event corresponding to the state to be disambiguated. The second-stage disambiguation enhancement involves extracting short-term original subsequences within the micro-window range from the local cache when the gating basic state determination is still pending after re-executing the enhanced feature certificate returned from the first-stage disambiguation enhancement.

Images