A multi-modal fusion-based train service work order priority dynamic scheduling method and system

Abstract

The present application relates to the technical field of train dispatching, in particular, the present application relates to a kind of train service work order priority dynamic scheduling method and system based on multi-modal fusion, real-time acquisition work order multimodal data in the present application, fusion structured field, fault text semantics and visual features by cross-modal semantic alignment model, generate work order feature vector, combine basic weight rule set to calculate initial priority;Monitoring work order time limit difference, emergency order superposition incremental emergency promotion coefficient with difference reduction;Based on the number of load work order, work hour saturation and mobile state calculation master resource utilization rate, high load area inhibits the priority of non-emergency work order, and is preferentially assigned to idle master;Finally, according to work order priority weight global ordering, combine master real-time state and position, generate optimal allocation instruction execution scheduling, realize multi-modal information and resource state coordination, balance work order time limit and regional load, improve scheduling efficiency and resource utilization rate.

Description

Technical Field

[0001] This invention relates to the field of vehicle dispatching technology, and more specifically, to a method and system for dynamic dispatching of vehicle work orders based on multimodal fusion. Background Technology

[0002] Vehicle dispatching technology is an important technology, specifically applied to the allocation of vehicle work orders and the scheduling of service resources. Its core function is to improve dispatching efficiency and user satisfaction by accurately matching work order priorities with service resources, thus meeting the core requirements of vehicle services for timeliness and resource utilization. Vehicle work orders contain multimodal data including structured information, fault text descriptions, fault images, and videos. Different work orders correspond to different business scenarios, and the resource utilization of service personnel is affected by current load, estimated working hours, and geographical location. Because these multi-dimensional information lacks effective coordination and integration, it is difficult to accurately quantify work order priorities, leading to scheduling imbalances such as excessive resource concentration in high-load areas, delayed response to urgent work orders, or backlog of ordinary work orders. To solve this technical problem, we provide a dynamic scheduling method and system for vehicle work order priorities based on multimodal fusion. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for dynamic scheduling of vehicle work orders based on multimodal fusion, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, one objective of this invention is to provide a dynamic scheduling method for vehicle work order priority based on multimodal fusion, comprising the following steps: S1. Real-time acquisition of multimodal data of vehicle service work orders to be dispatched, feature extraction and fusion of the multimodal data to generate work order feature vectors containing business scenario semantics; based on a preset set of basic weight rules corresponding to different vehicle service business scenario types, combined with the work order feature vectors, calculate the initial priority weight of each work order. S2. Monitor the difference between the estimated completion time and the current time of all work orders to be scheduled in real time. When the difference is found to be less than a preset first threshold, the work order is determined to be an emergency order. For the emergency order, calculate the emergency boosting coefficient based on the difference. The emergency boosting coefficient increases monotonically as the difference decreases. The emergency boosting coefficient is added to the initial priority weight of the work order to generate a priority weight. S3. Real-time acquisition of resource utilization rate for each service technician. The resource utilization rate is calculated based on the current number of work orders, expected work saturation, and geographical location movement status. A second threshold is set. When the average resource utilization rate of a specific group of technicians in the target area is continuously higher than the second threshold, the work order allocation strategy for the specific group of technicians in the target area is adjusted, including: Apply a suppression factor to the priority weight of non-urgent work orders, and prioritize assigning new work orders to idle workers whose resource utilization is below the second threshold; S4. Based on the priority weight of each work order and the adjusted work order allocation strategy, perform global priority sorting on all work orders to be scheduled. Based on the sorting results and the real-time status and location information of the service technicians, generate the optimal work order corresponding to the technician allocation instruction and execute the scheduling.

[0005] The second objective of this invention is to provide a system for implementing a multimodal fusion-based dynamic scheduling method for vehicle work orders, comprising: The multimodal data acquisition and fusion unit 1 extracts and fuses work order text, images and structured data through a cross-modal semantic alignment model to generate work order feature vectors containing business scenario semantics; The initial priority calculation unit 2 is connected to the multimodal data acquisition and fusion unit 1 to calculate the initial priority weight based on the preset business scenario rule set, and uses a piecewise nonlinear function to generate an emergency boosting coefficient. Combined with the emergency compensation coefficient generated inversely proportional to the resource utilization rate, the dynamic priority weight is output. Emergency order processing unit 3 and resource utilization calculation unit 4 calculate the worker resource utilization rate through the working time entropy value model, trigger the inhibition factor application mechanism according to the utilization rate threshold, and build a two-way matching queue to realize the real-time matching of work orders and idle workers; The global sorting and scheduling unit 5 integrates dynamic priority weights, suppression factor adjustment results, and regional delay coefficients, and generates a global sort through a multi-factor weighted aggregation engine to drive the execution of work orders and the optimal allocation instructions for workers.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention deeply integrates structured work order information, fault text, and visual data through a cross-modal semantic alignment model to generate feature vectors containing complete business scenario semantics. Combined with a basic weight rule set and fault severity confidence, it accurately calculates initial priorities, fully reflecting core business attributes such as fault risk and vehicle value. An emergency order is dynamically identified through a multi-level time window mechanism, and an emergency boosting coefficient is calculated using a piecewise nonlinear function. A compensation coefficient related to regional resource utilization is introduced during the superposition process, ensuring timely response to emergency orders while avoiding overload in high-load areas. Based on a work-hour entropy model, the service technician resource utilization rate is accurately calculated. In high-load areas, resource allocation is optimized by suppressing the priority of non-emergency work orders and bidirectional matching of idle technicians. Global sorting integrates multiple factors and dynamically adapts weight ratios, balancing work order priority, regional load, and timeliness requirements. This achieves deep collaboration between multi-modal information and resource status, effectively solving scheduling imbalance problems, improving work order response efficiency, service resource utilization, and user satisfaction, and adapting to the dynamic scheduling needs of vehicle services. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the overall workflow of the present invention; Figure 2 This is a schematic diagram of the overall structure of the present invention; The meanings of the labels in the diagram are as follows: 1. Multimodal data acquisition and fusion unit; 2. Initial priority calculation unit; 3. Emergency order processing unit; 4. Resource utilization calculation unit; 5. Global sorting and scheduling unit. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] Please see Figure 1 As shown, one of the objectives of this embodiment is to provide a dynamic scheduling method for vehicle work order priority based on multimodal fusion, including the following steps: S1. Real-time acquisition of multimodal data of vehicle service work orders to be dispatched, feature extraction and fusion of multimodal data to generate work order feature vectors containing business scenario semantics; based on a preset set of basic weight rules corresponding to different vehicle service business scenario types, combined with the work order feature vectors, calculate the initial priority weight of each work order. S2. Monitor the difference between the estimated completion time and the current time of all work orders to be scheduled in real time. When the difference is less than the preset first threshold, the work order is determined to be an emergency order. For emergency orders, calculate the emergency boosting coefficient based on the difference. The emergency boosting coefficient increases monotonically as the difference decreases. The emergency boosting coefficient is added to the initial priority weight of the work order to generate the priority weight. S3. Real-time acquisition of resource utilization rate for each service technician. Resource utilization rate is calculated based on the current number of work orders, expected work saturation, and geographical location movement status. A second threshold is set. When the average resource utilization rate of a specific group of technicians in the target area is continuously higher than the second threshold, the work order allocation strategy for that specific group of technicians in the target area is adjusted, including: Apply a suppression factor to the priority weight of non-urgent work orders, and prioritize assigning new work orders to idle workers whose resource utilization is below the second threshold; S4. Based on the priority weight of each work order and the adjusted work order allocation strategy, perform global priority sorting on all work orders to be scheduled. Based on the sorting results and the real-time status and location information of the service technicians, generate the optimal work order corresponding to the technician allocation instruction and execute the scheduling.

[0010] Feature extraction and fusion of multimodal data are achieved by establishing a cross-modal semantic alignment model. First, the structured fields in the basic information of the work order are encoded with key attributes to generate basic feature vectors. At the same time, the vehicle fault description text is input into a pre-trained language model to extract text semantic vectors. Lightweight convolutional networks are used to extract visual feature vectors from uploaded fault images or videos. The text semantic vectors and visual feature vectors are interactively aligned through a cross-modal attention mechanism to capture the semantic correlation between text description and image content. Finally, the aligned cross-modal semantic vectors are concatenated and fused with the basic feature vectors to generate a work order feature vector containing complete business scenario semantics.

[0011] When calculating the initial priority weight based on the basic weight rule set of the preset business scenario type, the work order feature vector is first input into the scenario classifier to identify the specific business scenario type. The basic calculation path corresponding to the scenario type is matched according to the preset rule set. The rule set contains the key factors affecting the weight under each scenario type and the logical relationship between the factors. By inputting the feature dimensions corresponding to the key factors in the work order feature vector into the dynamic calculation graph engine, multi-level weight calculation is performed according to the preset logical relationship. The key factors include the risk level of the faulty component, the vehicle value coefficient, and the user service level label. The calculation process introduces the fault severity confidence extracted from the feature vector as an adaptive adjustment parameter.

[0012] When monitoring urgent orders in real time, a multi-level time window mechanism is adopted. The preset first threshold includes an imminent timeout warning threshold and a severe timeout threshold. During the monitoring process, the work order state machine switching events are continuously tracked. When the work order enters the execution countdown stage and the difference between the estimated completion time and the current time is less than the imminent timeout warning threshold for the first time, a primary warning flag is triggered. When the difference is further less than the severe timeout threshold, it is upgraded to an urgent order flag. At the same time, the actual judgment value of the first threshold is dynamically adjusted according to the traffic congestion index of the geographical location of the work order.

[0013] The calculation of the emergency boost coefficient adopts a piecewise nonlinear function mapping strategy. When the difference is in the range of the near timeout warning threshold, the basic boost magnitude is generated by linear interpolation. When the difference enters the range of the severe timeout critical threshold, an exponential growth function is activated so that the boost magnitude increases as the remaining time decreases. The slope coefficient of the mapping function is dynamically adjusted by the confidence level of the fault severity.

[0014] When the emergency boost coefficient is superimposed on the initial priority weight, a weighted fusion mechanism is used to distinguish between normal weight and emergency increment. An initial priority base weight ratio coefficient is set so that the superimposed dynamic priority weight is expressed as the product of the initial priority weight and this coefficient plus the product of the emergency boost coefficient and the emergency compensation coefficient. The emergency compensation coefficient is generated inversely proportionally to the resource utilization rate of the target area. When the resource utilization rate is too high, the emergency compensation coefficient is reduced to balance the system load.

[0015] The calculation of service technician resource utilization rate is achieved through the time entropy model: the time prediction model is trained based on historical data, and the estimated time is output according to the current load work order type and fault feature vector. The sum of the estimated time of the current load work orders is divided by the technician's standard time capacity to obtain the time saturation. The time saturation is corrected by the mobility efficiency factor in combination with the driver's travel speed and destination distance in the geographical location movement status. The final resource utilization rate is the weighted geometric mean of the corrected time saturation and the current load work order number.

[0016] When applying a suppression factor to non-urgent work orders, the suppression intensity parameter is obtained by looking up the table based on the average resource utilization rate. The suppression factor is represented as the decay coefficient of the original priority of the non-urgent work order. This coefficient decreases monotonically from the benchmark value to the minimum value as the resource utilization rate increases. When prioritizing the allocation of idle workers, a two-way matching queue is established. Workers with resource utilization rates below the second threshold are arranged in descending order of idleness. New work orders are arranged in descending order of dynamic priority weight. Real-time matching of the work order with the highest weight and the worker with the highest idleness is achieved through a rolling time window.

[0017] The global priority ranking adopts a multi-factor weighted aggregation under dynamic strategy constraints. The dynamic priority weight of each work order is adjusted by introducing a suppression factor. At the same time, a heat map of resource utilization distribution of each master is loaded, and a regional delay coefficient is applied to work orders in high-load areas. The final ranking value is determined by the weighted sum of the adjusted priority weight, regional delay coefficient, and work order timeliness decay factor. The ranking engine recalculates the weight ratio of each factor at preset intervals, and the weight ratio is dynamically configured according to the system's average resource utilization.

[0018] It needs further explanation that after acquiring multimodal data of work orders to be dispatched in real time, the data covers different types such as structured information, text descriptions, and images / videos. The semantics of each modality are fragmented, and directly using them for priority calculation will lead to a bias in the understanding of the scenario. Therefore, it is necessary to achieve deep extraction and fusion of features through a cross-modal semantic alignment model to generate work order feature vectors that contain complete business scenario semantics. The specific implementation method is as follows: The feature extraction and fusion of multimodal data relies entirely on a cross-modal semantic alignment model. This model is specifically designed to integrate semantic information from different types of data. Its core capability is to break down barriers between structured, textual, and visual modalities, capturing the semantic relationships behind each modality. This ensures that the fused features fully reflect the business scenario of the vehicle service order. The entire process unfolds logically as follows: modal feature extraction—cross-modal semantic alignment—multi-feature concatenation and fusion. First, key attribute encoding is performed on the structured fields in the basic information of the service order to generate basic feature vectors. The basic information of the service order consists of data with a fixed format and clearly defined fields, including the service order number, the faulty vehicle, etc. Type, user ID, appointment time, and faulty component type are structured fields, which are fields with fixed formats and clear meanings, distinguishing them from unstructured data such as free text and images. Key attribute encoding converts the information in structured fields into numerical vectors that the model can recognize. For example, vehicle models are mapped to values ​​in the 0-1 range according to their value level, faulty component types are encoded into corresponding values ​​according to their risk level, and user service level labels are directly converted into quantified values. In this way, discrete structured information is transformed into continuous feature vectors, which are the basic feature vectors. Their dimension is determined by the number of key attributes, ensuring that the core business features in the structured information are fully covered. Simultaneously, features of unstructured data are extracted: on the one hand, the vehicle fault description text is input into a pre-trained language model to extract text semantic vectors. The vehicle fault description text is free text filled in by users or staff, such as abnormal noise when starting the engine or sluggish feedback from the brake pedal, containing key semantics such as fault phenomena and severity; the pre-trained language model uses a lightweight Chinese semantic understanding model. This type of model has been trained on massive amounts of text and has powerful semantic extraction capabilities. It can quickly output high-quality semantic vectors without retraining, and has high computational efficiency, adapting to the real-time processing requirements of work orders; during extraction, the fault description text is formatted according to the model requirements, and after inputting into the model, it outputs a fixed-dimensional text semantic vector. This vector can accurately capture the core fault information and semantic tendencies in the text. On the other hand, a lightweight convolutional network is used to extract visual feature vectors from uploaded fault images or videos. Fault images or videos are image data uploaded by users that reflect vehicle faults, such as photos of damaged tires or video clips of abnormal engine noises. The lightweight convolutional network uses models such as MobileNet and ShuffleNet. These networks significantly reduce the amount of computation while ensuring the accuracy of feature extraction by simplifying the structure and using depthwise separable convolutions, avoiding processing delays caused by large amounts of image / video data. When extracting image features, the image is directly input and the network outputs a visual feature vector. When extracting video features, keyframes are extracted at a frequency of 1 frame per second. The average value of the extracted features from each frame is taken as the visual feature vector of the video. This vector quantifies the visual features of the fault in the image, such as the degree of damage and the visual markings of the fault location.Subsequently, a cross-modal attention mechanism is used to interactively align the text semantic vector and the visual feature vector. This mechanism allows features from two different modalities—text and vision—to mutually focus on each other's key semantics. For example, the description of a brake pedal in the text guides the model to focus on the corresponding brake pedal area in the visual features, while the feature of tire damage in the vision strengthens the semantic weight of the related description in the text. This captures the semantic association between the text description and the image content, avoiding semantic fragmentation caused by the two modal features operating independently. In the specific alignment process, the model calculates the similarity matrix between the text semantic vector and the visual feature vector, assigns attention weights based on similarity, giving higher weights to feature dimensions with high relevance, and then generates an aligned cross-modal semantic vector through weighted fusion. This vector contains both the abstract semantics of the text and the concrete features of the image, achieving semantic unification of the two unstructured data. Finally, the aligned cross-modal semantic vector is concatenated and fused with the basic feature vector. This concatenation and fusion involves joining the two vectors end-to-end according to their feature dimensions to form a higher-dimensional comprehensive feature vector. For example, if the basic feature vector is 10-dimensional and the cross-modal semantic vector is 128-dimensional, the concatenation generates a 138-dimensional work order feature vector. This fusion method fully preserves the core attributes of structured data, the semantic information of text, and the visual features of images / videos, ensuring that the generated work order feature vector contains complete business scenario semantics. It reflects both the basic information of the work order and the specific circumstances and severity of the fault, providing comprehensive and accurate feature support for subsequent initial priority weight calculations. The entire feature extraction and fusion process improves efficiency through modal parallel processing, ensures semantic alignment through a cross-modal attention mechanism, and achieves complete information preservation through concatenation and fusion. This allows the work order feature vector to realistically and comprehensively depict the business scenario of the vehicle work order, providing reliable basic data for subsequent priority scheduling.

[0019] After generating the work order feature vector containing complete business scenario semantics, in order to ensure that the calculation of the initial priority weights aligns with the actual needs of different vehicle service businesses and avoids the situation where a single calculation standard leads to underweighting of high-risk scenarios or overweighting of ordinary scenarios, it is necessary to rely on a preset set of basic weight rules and combine it with the core semantic information in the feature vector to perform accurate calculations. The specific implementation method is as follows: The initial priority weights are calculated based on a pre-defined set of basic weight rules for different business scenario types. The process proceeds step-by-step, following a logic of scenario identification, path matching, factor extraction, and multi-layer computation. First, the work order feature vector is input into a scenario classifier to identify the specific business scenario type. This scenario classifier is a pre-trained classification model that learns from massive amounts of labeled work order data. It automatically categorizes the work order into the corresponding business scenario type based on semantic information such as faulty components, fault phenomena, and vehicle models within the work order feature vector. The classification accuracy must be verified to be no less than 98% to ensure bias-free scenario identification. During the classification process, the scenario classifier extracts dimensions from the feature vector that are strongly correlated with the scenario type and matches them with a pre-defined scenario type template to quickly output the specific business scenario type, thus locking in the appropriate rules for subsequent weight calculations. Next, the system matches the basic calculation path corresponding to the scenario type based on a preset set of basic weight rules. This set of rules is a collection of weight calculation specifications tailored to different vehicle service scenarios. Each scenario type corresponds to a unique calculation path. The rules clearly define the key factors influencing the weight in each scenario and the logical relationships between these factors. Examples include the superposition logic of faulty component risk level × weight percentage + vehicle value coefficient × weight percentage + user service level tag × weight percentage, or the priority progression logic of key factors in some scenarios. Specifically, the key factors include: the faulty component risk level, which is a level categorized according to the component's impact on vehicle driving safety (higher levels correspond to higher weight percentages); the vehicle value coefficient, a coefficient quantified based on factors such as market price and ownership (high-end models have higher coefficients than ordinary models), reflecting the service priority differences between vehicles of different values; and the user service level tag, a tag based on user historical consumption and membership level, corresponding to a fixed weight bonus. The logical relationships of these key factors' weight percentages are all verified through extensive business data statistics and expert experience to ensure they meet actual service priority requirements. Subsequently, a multi-level weight calculation is performed through a dynamic computation graph engine. This engine is an execution unit that supports flexible computational logic. Based on the matched computational path, it automatically parses the logical relationships between factors and constructs the corresponding computational flowchart, adapting to the computational needs of different scenarios without manual adjustment. Specifically, during the calculation, feature dimensions corresponding to key factors are first extracted from the work order feature vector. For example, the quantitative value corresponding to the risk level of the faulty component, the calibration value of the vehicle model value coefficient, and the weighted value corresponding to the user service level label are located from the feature vector, ensuring that each key factor can obtain accurate quantitative input from the feature vector. These extracted feature dimensions are then input into the dynamic computation graph engine according to a preset logical relationship. The engine performs multi-level computational operations according to a multi-level process: factor quantification calibration—single-factor weight calculation—multi-factor weighted fusion. For example, the input values ​​of each key factor are first standardized and calibrated (ensuring that the values ​​of different dimensions are on the same order of magnitude), then the basic weights corresponding to each factor are calculated separately, and finally, a weighted sum is performed according to the proportions set in the rule set to obtain the preliminary weight results.During the calculation process, the fault severity confidence score extracted from the feature vector needs to be introduced as an adaptive adjustment parameter. The fault severity confidence score is a quantitative indicator extracted from multimodal fusion features, reflecting the credibility of the actual severity of the fault, and its value ranges from 0 to 1. The role of this parameter is to dynamically correct the results of multi-level calculations: when the confidence score is high, it indicates that the fault description is true and credible, and the original calculation result is output; when the confidence score is low, the weight ratio of key factors is appropriately reduced to avoid misjudgment of weights due to ambiguity in fault information. For example, when the confidence score is below 0.5, the weight ratio of the risk level of the faulty component is reduced by 30%, ensuring that the initial priority weights are both in line with the characteristics of the scenario and avoid the bias caused by information uncertainty. Through the above process, the final value output by the dynamic calculation graph engine is the initial priority weight of the work order. This weight reflects the inherent priority of the business scenario, integrates personalized factors such as vehicle model and user level, and achieves adaptive adjustment through fault severity confidence score, ensuring that the initial weight of each work order is accurate and reasonable, laying a reliable foundation for the subsequent superposition of emergency boost coefficients.

[0020] After calculating the initial priority weight of work orders, in order to accurately control order timeliness and avoid default due to relying solely on static weights, a dynamic monitoring mechanism is needed to identify urgent orders. A multi-level time window mechanism can adapt to scenarios with different levels of timeliness, and combined with dynamic threshold adjustment based on road conditions, it ensures that the determination of urgent orders is both sensitive and accurate. The specific implementation method is as follows: The core of real-time monitoring of urgent orders lies in the implementation of a multi-level time window mechanism. This mechanism achieves end-to-end timeliness control through dual threshold division, status tracking, and dynamic correction. Firstly, the pre-defined first threshold comprises two levels: an imminent timeout warning threshold and a critical timeout threshold. The imminent timeout warning threshold is the alert line for orders that are about to enter a time-critical phase, used to mark orders requiring priority attention in advance. For example, based on the industry's standard service cycle, it is set at 2 hours (i.e., an alert is triggered when the difference between the estimated completion time and the current time is less than 2 hours). The critical timeout threshold is the emergency line for orders that are about to time out, used to define truly urgent orders that require priority processing, set at 30 minutes (i.e., orders are judged to be urgent when the difference is less than 30 minutes). The setting of these two thresholds not only allows for service response buffer time but also distinguishes between priority and urgent orders, avoiding misjudgments of urgency. During monitoring, it is necessary to continuously track work order state machine transition events. Work order state machine transition events refer to the transformation of different states in the lifecycle of a work order, including states such as pending assignment, dispatched, in progress, execution countdown, and completed. The execution countdown stage refers to the countdown phase after the work order is dispatched, based on the estimated service time and the time remaining until the completion deadline. Only after entering this stage will the time difference monitoring be initiated to avoid mistakenly triggering the emergency flag for pending orders that have not yet been dispatched. In specific tracking, the system captures work order status change signals in real time. When it detects that a work order has switched from dispatched to the execution countdown stage, it begins to calculate the difference between the estimated completion time and the current time, and formally initiates the emergency order identification process. The triggering of emergency orders follows a tiered logic: when the difference mentioned above first falls below the near-timeout warning threshold, the system automatically triggers a primary warning flag. At this point, only a highlight is displayed in the dispatching backend, without changing the original priority weight of the order. The purpose is to remind dispatchers to pay attention to the execution progress of the order. When the difference further decreases and falls below the severe timeout threshold, the primary warning flag is upgraded to an emergency order flag. At this time, the system will initiate the subsequent emergency escalation coefficient calculation process to ensure that the order priority is quickly increased. This tiered triggering mechanism avoids the gap problem caused by a single threshold, where either there is no warning or the order is already an emergency, and allows dispatching resources to be allocated as needed. At the same time, in order to adapt to the impact of road conditions in actual service scenarios and avoid the situation where drivers cannot arrive on time due to road congestion but are mistakenly judged as emergency orders, the system will dynamically adjust the actual judgment value of the first threshold based on the road congestion index of the work order's geographical location. The traffic congestion index is a quantitative indicator obtained by accessing real-time traffic data, ranging from 0 to 1. The higher the value, the more severe the congestion. The adjustment logic is as follows: when the congestion index is higher than 0.7 (severe congestion), the near-timeout warning threshold is increased by 30%, and the severe timeout threshold is increased by 50%. For example, the original 2-hour warning threshold is adjusted to 2 hours and 40 minutes, and the 30-minute critical threshold is adjusted to 45 minutes. When the congestion index is lower than 0.3 (smooth traffic), the threshold remains unchanged. When it is between the two, it is adjusted linearly according to the congestion index ratio.This dynamic adjustment allows thresholds to adapt to actual traffic conditions, ensuring that emergency order determination meets timeliness requirements without deviating from the actual execution capabilities of service personnel, thus preventing invalid emergency orders from consuming excessive scheduling resources. The entire monitoring process achieves urgency level classification through multi-level threshold division, ensures accurate monitoring timing through state machine tracking, and achieves dynamic threshold adaptation through traffic congestion index correction. This not only guarantees timely identification of emergency orders but also avoids misjudgments and omissions, providing a reliable basis for subsequent calculation of emergency escalation coefficients and priority adjustments.

[0021] After identifying urgent orders through a multi-level time window mechanism, to ensure that priority weights accurately match the urgency of the orders—avoiding slightly delayed orders from consuming too many resources while ensuring that extremely urgent orders receive sufficient priority support—the urgency enhancement coefficient is calculated using a piecewise nonlinear function mapping strategy. This strategy adapts to different time intervals through differentiated functions and dynamically adjusts the calculation based on fault characteristics. The specific implementation method is as follows: The calculation logic for the urgency boosting coefficient revolves around interval division, function adaptation, and dynamic coefficient adjustment. First, the segmentation basis and core strategy are clarified. The segmented nonlinear function mapping strategy refers to using different nonlinear functions to calculate the boosting coefficient based on the threshold interval where the difference between the estimated completion time and the current time falls. This allows the boosting magnitude to increase differentially as the difference decreases, ensuring a smooth transition within the interval while highlighting the urgency of severely overdue orders. Here, the difference refers to the difference between the estimated completion time and the current time calculated in real-time during the monitoring mentioned earlier. Its magnitude directly reflects the timeliness of the order; the smaller the difference, the higher the urgency, and the larger the corresponding boosting coefficient should be. Furthermore, the growth rate needs to change nonlinearly with interval switching. When the difference falls within the near-timeout warning threshold range (i.e., between the near-timeout warning threshold after dynamic road condition correction and the severe timeout threshold, for example, 1 hour to 30 minutes after correction), a basic increase is generated through linear interpolation. Linear interpolation means that within this range, the increase coefficient increases steadily at a fixed ratio as the difference decreases, forming a linear correlation. This avoids excessive increase leading to resource imbalance while assigning basic urgency weight to orders. In practice, the range of the increase coefficient for this range is first defined: when the difference equals the near-timeout warning threshold, the increase coefficient is the baseline value (e.g., 0.1); when the difference equals the severe timeout threshold, the increase coefficient is the upper limit of the range (e.g., 0.4). Then, the corresponding basic increase is calculated based on the current actual difference using the linear interpolation formula. For example, when the difference is 45 minutes (in the middle of the 1 hour to 30 minutes range), the increase coefficient is 0.25, ensuring that urgent orders within this range receive tiered priority increases, distinguishing them from ordinary orders without conflicting with severely urgent orders. When the time difference enters the critical threshold range for severe timeout (i.e., less than the dynamically corrected critical threshold for severe timeout, such as within 30 minutes after correction), an exponential growth function is activated. This causes the increase in priority to rise rapidly as the remaining time decreases. The exponential growth function means that the rate of increase of the increase coefficient accelerates exponentially as the time difference decreases. For example, when the time difference decreases from 30 minutes to 15 minutes, the increase coefficient increases from 0.4 to 0.8, and when it decreases from 15 minutes to 5 minutes, the increase coefficient jumps directly from 0.8 to 1.5. This growth pattern allows extremely urgent orders to quickly receive high priority, ensuring that service resources are prioritized and preventing order timeouts. In specific settings, the baseline parameters of the exponential function need to be calibrated using a large amount of business data to ensure that the increase is neither excessively exaggerated, leading to system imbalance, nor ineffective in highlighting the priority of severely urgent orders. For example, when the time difference approaches 0, the upper limit of the increase coefficient can be set to 2.0 to prevent a single order from consuming too many resources.Meanwhile, the slope coefficient of the mapping function is dynamically adjusted by the fault severity confidence level. The fault severity confidence level is a quantitative indicator extracted from the work order feature vector, reflecting the credibility of the actual severity of the fault (e.g., when the fault involves core safety components and there is clear image evidence, the confidence level is close to 1; when the fault description is vague and there is no visual evidence, the confidence level is low). Its core function is to ensure that the boost coefficient not only adapts to the urgency of timeliness but also matches the severity of the fault itself. The specific adjustment logic is as follows: when the fault severity confidence level is higher than 0.8, the slope coefficient of the linear interpolation and exponential growth function is increased by 20%, allowing urgent orders with severe faults to receive a higher boost; when the confidence level is between 0.5 and 0.8, the slope coefficient remains at its default value; when the confidence level is lower than 0.5, the slope coefficient is decreased by 30% to avoid inflated priority due to vague fault information, ensuring that the calculation of the boost coefficient is both timely and consistent with the risk level of the actual business scenario. The entire process of calculating the emergency escalation coefficient achieves precise quantification of the urgency level through piecewise functions, linear interpolation ensures smooth scheduling of medium-urgent orders, exponential growth highlights the priority of extremely urgent orders, and dynamic adjustment of the confidence level of fault severity makes the coefficient calculation more in line with the nature of the fault. The final output emergency escalation coefficient can comprehensively reflect the timeliness urgency and fault risk level of the order, laying a precise foundation for subsequent superposition with the initial priority weight.

[0022] After calculating the emergency boost coefficient using a piecewise nonlinear function mapping strategy, in order to both retain the inherent attributes of the business scenario reflected by the initial priority weight (such as fault risk, vehicle value, etc.) and reasonably incorporate the timeliness requirements of emergency increments, while avoiding system imbalance in high-load areas due to single superposition, a weighted fusion mechanism is needed to scientifically integrate the two to generate dynamic priority weights that fit actual scheduling needs. The specific implementation method is as follows: The core of adding the emergency boosting coefficient to the initial priority weight is to differentiate between normal weight and emergency increment through a weighted fusion mechanism. This mechanism involves configuring dedicated percentage parameters for both the initial priority weight and the emergency boosting coefficient, allowing them to contribute to the final result according to a preset ratio, preventing one from excessively dominating and distorting the priority. The normal weight, calculated earlier based on the business scenario rule set, reflects the inherent business value of the work order; the emergency increment, the emergency boosting coefficient, focuses on the timeliness and urgency of the work order. This differentiated weight configuration ensures that the dynamic priority aligns with both the business essence and timeliness requirements. First, an initial priority base weight percentage coefficient is set. This coefficient is the core parameter controlling the contribution of the initial priority weight to the final result. Its value range, verified through business testing, is set to 0.7 to 0.9. Specific calibration needs to be combined with the scenario type: for core safety scenarios such as engine failure and brake system repair, the coefficient is set to 0.9 to ensure a higher weight percentage for inherently high-risk attributes; for ordinary scenarios such as routine maintenance and interior repair, the coefficient is set to 0.7 to allow more adjustment space for emergency increments. Once calibrated, this coefficient remains fixed within the same scenario type to ensure the stability of the normal weight. Based on this, the superimposed dynamic priority weight is generated according to the logic of normal contribution + emergency contribution: the normal contribution is the product of the initial priority weight and the above-mentioned proportion coefficient. For example, if the initial weight is 0.8 and the proportion coefficient is 0.8, then the normal contribution is 0.64; the emergency contribution is the product of the emergency boost coefficient and the emergency compensation coefficient. The emergency compensation coefficient is a key adjustment parameter for balancing the system load. Its core characteristic is that it is generated inversely proportionally based on the resource utilization rate of the service providers in the target area. The target area refers to the geographical service area to which the work order belongs. The resource utilization rate of the service providers is the comprehensive load quantification value of the service providers in the target area calculated by the working hour entropy value model mentioned above (the value ranges from 0 to 1, and the higher the value, the busier the service provider is). The generation logic of the emergency compensation coefficient is as follows: The preset baseline compensation coefficient is 1.0. When the average resource utilization rate of the target area's workers is below the second threshold (e.g., 70%, indicating a relaxed system load), the compensation coefficient remains at 1.0, allowing the emergency boost coefficient to take full effect. When the resource utilization rate is between the second threshold and 85% (the high load threshold), the compensation coefficient decreases linearly with increasing utilization (e.g., at 80% utilization, the coefficient drops to 0.7). When the resource utilization rate is above 85% (extremely high load), the compensation coefficient is fixed at 0.5. This significantly reduces the proportion of emergency contributions, preventing excessive emergency orders from flooding into high-load areas and causing scheduling paralysis. This inverse proportional generation method allows the emergency compensation coefficient to accurately adapt to the regional load status, achieving a dynamic balance between peak-hour throttling and off-peak full utilization. Finally, the dynamic priority weight is obtained by summing the above two parts. For example, if the emergency boost coefficient is 0.5 and the emergency compensation coefficient is 0.7, the emergency contribution is 0.35. Combined with the previously mentioned normal contribution of 0.64, the final dynamic priority weight is 0.99.Throughout the entire process, the initial priority base weight ratio ensures that the inherent value of the business is not diluted, while the emergency compensation coefficient, through linkage with regional resource utilization, effectively balances the contradiction between emergency order response and system load, making the generated dynamic priority weight both timely and feasible, and providing a precise weight basis for subsequent adjustment of allocation strategies based on resource utilization.

[0023] After generating dynamic priority weights for work orders, in order to accurately determine the load status of service personnel in the target area, which is the core basis for whether to adjust the work order allocation strategy, the calculation of service personnel resource utilization needs to be achieved through the work hour entropy value model. This model can comprehensively consider the work order work hour load and the mobility efficiency of the personnel, avoiding load assessment bias caused by a single dimension judgment. The specific implementation method is as follows: The calculation of service technician resource utilization revolves around the logic of time prediction, saturation calculation, dynamic correction, and weighted fusion. The core relies on the time entropy model, a mathematical model that comprehensively quantifies technician workload and movement losses. By integrating time demand and movement status, it outputs a normalized index (ranging from 0 to 1) that accurately reflects the technician's workload. A value closer to 1 indicates higher resource utilization and a heavier technician workload. First, the time prediction model is trained based on historical data. This historical data covers information related to all completed work orders in the service area over the past six months, including work order type (e.g., engine repair, tire replacement), fault characteristics (faulty parts, severity), technician skill level, service hours, and weather conditions. This ensures comprehensive data coverage of key factors affecting work order processing time. The time prediction model uses a lightweight regression model (e.g., random forest regression, gradient boosting regression). This type of model can adapt to multiple feature inputs while ensuring real-time prediction efficiency, avoiding computational delays caused by model complexity. During training, work order types and fault characteristics from historical data are used as input features, and the corresponding actual processing time is used as the output label. The model parameters are optimized through iterative training to ensure that the model prediction error is controlled within a preset range (e.g., the mean absolute error does not exceed 10 minutes). After training, when it is necessary to calculate the resource utilization rate of a certain operator, each work order type currently being handled by the operator and its corresponding fault feature vector are input into the model. The model outputs the estimated processing time for each work order, which is the predicted total time required from accepting the work order to its completion, including the time spent on the entire process such as fault diagnosis, maintenance operations, and final confirmation. Next, the work saturation is calculated. First, the estimated working hours of the current work orders of the technician are summed (i.e., the estimated working hours of all unfinished work orders are added together; for example, if there are two work orders with estimated working hours of 1.5 hours and 2 hours respectively, the total is 3.5 hours). Then, the technician's standard working hour capacity is determined. The standard working hour capacity is the preset maximum effective working time of the technician per day. Combining the industry's standard work system and the characteristics of maintenance services, it is set at 8 hours / day (excluding lunch breaks, commuting, and other non-working time). Finally, the sum of the estimated working hours is divided by the standard working hour capacity to obtain the work saturation. For example, 3.5 hours divided by 8 hours results in a work saturation of 0.4375. This indicator quantifies the proportion of the technician's current work order workload, but does not consider the impact of travel time on the actual workload. The system then dynamically adjusts the time based on the driver's location and movement status. The location and movement status is obtained in real time through the driver's mobile device location tracking. The core information includes driving speed and destination distance: driving speed is the driver's current real-time speed (e.g., 30 km / h), reflecting the smoothness of the road conditions; destination distance is the straight-line distance from the driver's current location to the next service location of the pending work order (e.g., 5 km). Both factors together determine the travel time.The mobility efficiency factor is a correction coefficient calculated based on driving speed and destination distance. It is used to quantify the impact of travel time on work saturation. When the driving speed is high and the destination distance is short, the travel time is short, and the mobility efficiency factor is close to 1 (e.g., a distance of 5 kilometers, a speed of 30 kilometers per hour, and a travel time of about 10 minutes, with a factor of 0.95). When the driving speed is slow and the distance is long, the travel time is long, and the factor is less than 1 (e.g., a distance of 10 kilometers, a speed of 15 kilometers per hour, and a travel time of 40 minutes, with a factor of 0.8). The specific correction process is as follows: multiply the previously calculated work saturation by the mobility efficiency factor to obtain the corrected work saturation. For example, 0.4375 multiplied by 0.95 results in a corrected work saturation of 0.4156, which is closer to the actual effective work hours available for work order processing by the driver. Finally, the final resource utilization rate is calculated using the weighted geometric mean of the corrected work hour saturation and the current number of work orders: the current number of work orders is the total number of work orders that the technician is currently processing or that have been dispatched but not yet started (e.g., 2 in the example above); the weighted geometric mean is a fusion calculation that takes into account both dimensions, by assigning preset weights to the two indicators (according to experimental calibration, the weight of the corrected work hour saturation is 0.7, and the weight of the current number of work orders is 0.3), avoiding the one-sidedness of a single-dimensional evaluation (e.g., looking only at work hour saturation may ignore the scheduling pressure caused by too many work orders, and looking only at the number may ignore the load of a single work order with excessively long working hours). In the specific calculation, the two indicators are first geometrically averaged according to their corresponding weights, and the final output value is the resource utilization rate of the service technician (e.g., in the example above, the corrected work hour saturation is 0.4156, the number of work orders is 2, and the weighted geometric mean is approximately 0.57). This value can comprehensively and accurately reflect the overall load status of the technician, providing a core quantitative basis for the subsequent determination of the second threshold and adjustment of the allocation strategy. The entire calculation process ensures the accuracy of time estimation through a time prediction model, incorporates the actual mobility loss in the scenario through mobility efficiency factor correction, and balances the two dimensions of time and work order quantity through weighted geometric average. This makes the calculation of resource utilization both in line with business reality and operable, providing reliable data support for judging whether the average resource utilization rate is higher than the second threshold, and ensuring the rationality of subsequent allocation strategy adjustments.

[0024] After monitoring that the average resource utilization rate of a specific group of workers in the target area consistently exceeds the second threshold, in order to prevent further exacerbation of service pressure in high-load areas while ensuring the response efficiency of urgent orders, it is necessary to achieve system load balancing by suppressing the priority of non-urgent work orders and optimizing the worker allocation logic. The specific implementation method is as follows: The core of applying suppression factors to non-urgent work orders is to dynamically reduce their priority based on regional load. First, the average resource utilization rate, i.e., the arithmetic mean of the resource utilization rates of all service personnel within the target region, is defined. This value directly reflects the overall service pressure in the region and is the core basis for adjusting the suppression intensity. The system has a preset suppression intensity parameter reference table, with the average resource utilization rate on the horizontal axis and the suppression intensity parameter on the vertical axis. Adaptive parameters under different loads are calibrated using a large amount of scheduling data. For example, when the average resource utilization rate is 75% (slightly above the second threshold), the corresponding suppression intensity parameter is 0.8; when the utilization rate rises to 85%, the parameter drops to 0.5; and when the utilization rate reaches 90% or above, the parameter is fixed at 0.3 (i.e., the minimum suppression intensity parameter), ensuring that the suppression intensity is precisely matched to the load level. The suppression factor is essentially a decay coefficient for the original priority of non-urgent work orders. Its value is directly determined by the suppression strength parameter and follows a rule that it monotonically decreases from the baseline value to the minimum value as resource utilization increases: the baseline value is set to 1.0 (when the average resource utilization is below the second threshold, there is no suppression effect, and non-urgent work orders participate in scheduling according to their original priority); the minimum value is set to 0.3 (when the area is under extremely high load, the priority is decayed to the maximum extent to avoid occupying emergency order resources). For example, if the original dynamic priority weight of a non-urgent work order is 0.8, when the average resource utilization is 85%, the suppression factor is 0.5, and the decayed priority weight is 0.4, effectively reducing its competitiveness in scheduling. While prioritizing non-urgent work orders, new work orders (including non-urgent work orders and urgent work orders not yet handled by high-load areas) should be preferentially assigned to idle workers with resource utilization rates below the second threshold. This process is achieved through a two-way matching queue, which consists of two separate sorting queues for workers and work orders. The core principle is to ensure that the most idle workers receive the most urgent (or highest-weighted) work orders, thereby improving scheduling efficiency. First, a worker queue is constructed: workers with resource utilization rates below the second threshold within the target area are selected and sorted in descending order of idleness. Idleness is quantified by combining (1 - resource utilization rate) with the current waiting time without work orders. For example, workers with a resource utilization rate of 30% and no waiting work orders, and workers with an idleness rate higher than 50% utilization, are placed higher in the sorting, ensuring that truly idle workers get priority in matching opportunities. Next, a work order queue is constructed: new work orders to be assigned (including suppressed non-urgent work orders and unassigned urgent work orders) are sorted in descending order of dynamic priority weight. Work orders with higher weights are placed earlier in the queue, ensuring that high-priority work orders are matched first. Ultimately, real-time matching is achieved through a rolling time window, which is a fixed time period (verified to be 5 seconds per window based on scheduling efficiency). Within each window, the system will synchronously update the order of two queues: the worker queue will adjust its order according to real-time resource utilization changes (e.g., if a worker's resource utilization decreases after completing a work order, the order will automatically move forward), and the work order queue will add newly generated work orders and reorder them.Within each time window, the system matches the first item in the work order queue (highest weight) with the first item in the master queue (highest availability). If a match is found, the work order and master are removed from the queue, and the cycle continues to the next window. If the first master cannot accept the task due to geographical distance or skill mismatch, the system automatically matches the next master in the queue until a suitable master is found. This ensures optimal matching in each window, avoiding waste of idle master resources and guaranteeing that high-weight work orders are accepted promptly. The entire process dynamically balances regional load through a suppression factor and achieves precise matching of resources and demands through bidirectional matching queues and rolling time windows. This prevents overload in high-load areas and maximizes the utilization of idle master resources, ensuring the scheduling system operates efficiently and orderly under high pressure. This provides an optimized foundation for work order and master matching for subsequent global priority sorting.

[0025] After completing the suppression of non-urgent work orders and the two-way matching of idle workers, in order to achieve the optimal sorting of work orders to be scheduled across the entire system, taking into account both the differences in the priority of the work orders themselves and balancing the regional load pressure and the timeliness requirements of the work orders, the global priority sorting adopts a multi-factor weighted aggregation method under dynamic strategy constraints. By integrating multiple key influencing factors and dynamically adapting the weight ratio, it is ensured that the sorting results can accurately match the real-time scheduling scenario. The specific implementation method is as follows: The core logic of global priority ranking is factor integration—dynamic weighting—periodic update. First, it's important to clarify that multi-factor weighted aggregation under dynamic policy constraints refers to selecting core factors influencing the ranking and assigning dynamic weights to each factor, with the optimal system scheduling goal as the constraint. The final ranking value is generated through weighted summation, avoiding ranking imbalances caused by a single factor dominating and adapting to different system load states through weight adjustments. The first step in ranking is to introduce the result of a suppression factor adjustment to the dynamic priority weights of each work order. This suppression factor adjustment result is the priority weight after attenuation for non-urgent work orders in high-load areas, as mentioned earlier. It fully considers the scheduling limitations of regional load on non-urgent work orders and uses it as the basis for ranking, ensuring that the adjustment effect of the previous allocation strategy continues in the global ranking. For example, if the original dynamic priority weight of a non-urgent work order is 0.8, it becomes 0.4 after suppression factor adjustment. This adjusted value will directly serve as the core input for subsequent weighted calculations. Simultaneously, the system loads a heatmap showing the distribution of resource utilization rates for each service area (such as city zones or business districts). This heatmap visualizes the resource utilization rates of service areas, using color depth to quantify load levels. Darker colors indicate higher average resource utilization and heavier loads (e.g., red represents high-load areas with utilization rates above 85%, and green represents low-load areas with utilization rates below 60%). The data originates from the real-time calculations of resource utilization rates for each service area mentioned earlier, aggregated by region. Based on this heatmap, a regional delay coefficient is applied to all work orders awaiting scheduling in high-load areas. This regional delay coefficient is a quantitative indicator used to reduce the priority of work orders in high-load areas, ranging from 0 to 0.2. The heavier the regional load, the larger the coefficient. For example, the delay coefficient for the high-load red area is set to 0.2, the medium-load yellow area to 0.1, and the low-load green area to 0. Its core function is to prevent work orders in high-load areas from being excessively concentrated in the priority ranking, which would further exacerbate the regional load. The coefficient attenuation balances the scheduling pressure of each region. In addition, the sorting process also needs to introduce a work order timeliness decay factor. This factor quantifies the impact of work order backlog time, ranging from 0.9 to 1.0. The longer the time from work order creation to the present, the closer the factor is to 0.9, and vice versa. For example, a work order created 1 hour ago has a factor of 1.0, while a work order created 6 hours ago and not yet assigned has a factor of 0.92. The purpose is to avoid long-term work order backlogs, encourage the system to prioritize older work orders with tight deadlines, and ensure scheduling fairness. The final sorting value is determined by the weighted sum of the adjusted priority weight, the regional delay coefficient, and the work order timeliness decay factor. The weight percentages of these three factors are not fixed but dynamically configured by the sorting engine. The adjusted priority weight is the core factor with the highest basic weight, while the regional delay coefficient and the work order timeliness decay factor are auxiliary factors whose weights are dynamically adjusted according to the system status.The sorting engine recalculates the weighting of each factor at preset intervals (calibrated to 1 minute after verification of scheduling efficiency and resource consumption balance). The adjustment is based on the system's average resource utilization rate, which is the arithmetic mean of the resource utilization rates of all service technicians in the entire system, reflecting the overall load status. When the system's average resource utilization rate is below 60% (low overall load), the adjusted priority weighting is set to 0.8, the regional delay coefficient to 0.1, and the work order timeliness decay factor to 0.1, prioritizing the work order's own priority. When the utilization rate is between 60% and 80% (medium overall load), the three weightings are adjusted to 0.7, 0.15, and 0.15, balancing priority with regional and timeliness factors. When the utilization rate is above 80% (high overall load), the weightings are adjusted to 0.6, 0.2, and 0.2, focusing on alleviating high load pressure through the regional delay coefficient and preventing work order backlog through the timeliness decay factor. Within each cycle, the sorting engine synchronously updates the adjusted priority weights, regional latency coefficients, and timeliness decay factors of all work orders. Combined with the currently dynamically configured factor weight ratios, it performs weighted summation and calculation to obtain the final sorting value for each work order. Work orders across the entire system are then sorted from highest to lowest sorting value. This entire process integrates multiple factors covering three core dimensions: work order priority, regional load, and timeliness requirements. By dynamically adjusting weight ratios to adapt to changes in the overall system load, it ensures that the global sorting result is both optimal and feasible, providing a precise sorting basis for subsequently generating optimal work orders and assigning optimal dispatch instructions to workers.

[0026] The second objective of this invention is to provide a system for implementing a multimodal fusion-based dynamic scheduling method for vehicle work orders, comprising: The multimodal data acquisition and fusion unit 1 extracts and fuses work order text, images and structured data through a cross-modal semantic alignment model to generate work order feature vectors containing business scenario semantics; The initial priority calculation unit 2 is connected to the multimodal data acquisition and fusion unit 1 to calculate the initial priority weight based on the preset business scenario rule set, and uses a piecewise nonlinear function to generate an emergency boosting coefficient. Combined with the emergency compensation coefficient generated inversely proportional to the resource utilization rate, the dynamic priority weight is output. Emergency order processing unit 3 and resource utilization calculation unit 4 calculate the worker resource utilization rate through the working time entropy value model, trigger the inhibition factor application mechanism according to the utilization rate threshold, and build a two-way matching queue to realize the real-time matching of work orders and idle workers; The global sorting and scheduling unit 5 integrates dynamic priority weights, suppression factor adjustment results, and regional delay coefficients, and generates a global sort through a multi-factor weighted aggregation engine to drive the execution of work orders and the optimal allocation instructions for workers.

[0027] This invention acquires multimodal work order data in real time, integrates structured fields, fault text semantics, and visual features through a cross-modal semantic alignment model to generate work order feature vectors, and calculates initial priorities using a basic weight rule set. It monitors work order timeliness differences, and adds an emergency boost coefficient that increases as the difference decreases for urgent orders. Based on the number of work orders, work hour saturation, and mobility status, it calculates the worker resource utilization rate, suppressing the priority of non-urgent work orders in high-load areas and prioritizing their allocation to idle workers. Finally, it globally sorts work orders by priority weight, combines the worker's real-time status and location, and generates optimal allocation instructions for execution scheduling. This achieves multimodal information and resource status coordination, balances work order timeliness and regional load, and improves scheduling efficiency and resource utilization.

[0028] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic scheduling method for vehicle work order priority based on multimodal fusion, characterized in that: Includes the following steps: S1. Real-time acquisition of multimodal data of vehicle service work orders to be dispatched, feature extraction and fusion of the multimodal data to generate work order feature vectors containing business scenario semantics; based on a preset set of basic weight rules corresponding to different vehicle service business scenario types, combined with the work order feature vectors, calculate the initial priority weight of each work order. S2. Monitor the difference between the estimated completion time and the current time of all work orders to be scheduled in real time. When the difference is found to be less than a preset first threshold, the work order is determined to be an emergency order. For the emergency order, calculate the emergency boosting coefficient based on the difference. The emergency boosting coefficient increases monotonically as the difference decreases. The emergency boosting coefficient is added to the initial priority weight of the work order to generate a priority weight. S3. Real-time acquisition of resource utilization rate for each service technician. The resource utilization rate is calculated based on the current number of work orders, expected work saturation, and geographical location movement status. A second threshold is set. When the average resource utilization rate of a specific group of technicians in the target area is continuously higher than the second threshold, the work order allocation strategy for the specific group of technicians in the target area is adjusted, including: Apply a suppression factor to the priority weight of non-urgent work orders, and prioritize assigning new work orders to idle workers whose resource utilization is below the second threshold; S4. Based on the priority weight of each work order and the adjusted work order allocation strategy, perform global priority sorting on all work orders to be scheduled. Based on the sorting results and the real-time status and location information of the service technicians, generate the optimal work order corresponding to the technician allocation instruction and execute the scheduling.

2. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 1, characterized in that: Feature extraction and fusion of multimodal data are specifically achieved by establishing a cross-modal semantic alignment model. First, the structured fields in the basic information of the work order are encoded with key attributes to generate basic feature vectors. At the same time, the vehicle fault description text is input into a pre-trained language model to extract text semantic vectors. Lightweight convolutional networks are used to extract visual feature vectors from uploaded fault images or videos. The text semantic vectors and visual feature vectors are interactively aligned through a cross-modal attention mechanism to capture the semantic correlation between text description and image content. Finally, the aligned cross-modal semantic vectors are concatenated and fused with the basic feature vectors to generate a work order feature vector containing complete business scenario semantics.

3. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 2, characterized in that: When calculating the initial priority weight based on the basic weight rule set of the preset business scenario type, the work order feature vector is first input into the scenario classifier to identify the specific business scenario type. The basic calculation path corresponding to the scenario type is matched according to the preset rule set. The rule set contains the key factors affecting the weight under each scenario type and the logical relationship between the factors. By inputting the feature dimensions corresponding to the key factors in the work order feature vector into the dynamic calculation graph engine, multi-level weight calculation is performed according to the preset logical relationship. The key factors include the risk level of the faulty component, the vehicle value coefficient, and the user service level label. The calculation process introduces the fault severity confidence extracted from the feature vector as an adaptive adjustment parameter.

4. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 1, characterized in that: When monitoring urgent orders in real time, a multi-level time window mechanism is adopted. The preset first threshold includes an imminent timeout warning threshold and a severe timeout threshold. During the monitoring process, the work order state machine switching events are continuously tracked. When the work order enters the execution countdown stage and the difference between the estimated completion time and the current time is less than the imminent timeout warning threshold for the first time, a primary warning flag is triggered. When the difference is further less than the severe timeout threshold, it is upgraded to an urgent order flag. At the same time, the actual judgment value of the first threshold is dynamically adjusted according to the traffic congestion index of the geographical location of the work order.

5. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 1, characterized in that: The calculation of the emergency boost coefficient adopts a piecewise nonlinear function mapping strategy. When the difference is in the range of the near timeout warning threshold, the basic boost magnitude is generated by linear interpolation. When the difference enters the range of the severe timeout critical threshold, an exponential growth function is activated so that the boost magnitude increases as the remaining time decreases. The slope coefficient of the mapping function is dynamically adjusted by the confidence level of the fault severity.

6. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 5, characterized in that: When the emergency boost coefficient is superimposed on the initial priority weight, a weighted fusion mechanism is used to distinguish between normal weight and emergency increment. An initial priority base weight ratio coefficient is set so that the superimposed dynamic priority weight is expressed as the product of the initial priority weight and this coefficient plus the product of the emergency boost coefficient and the emergency compensation coefficient. The emergency compensation coefficient is generated inversely proportionally to the resource utilization rate of the target area. When the resource utilization rate is too high, the emergency compensation coefficient is reduced to balance the system load.

7. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 1, characterized in that: The calculation of service technician resource utilization rate is achieved through the time entropy model: the time prediction model is trained based on historical data, and the estimated time is output according to the current load work order type and fault feature vector. The sum of the estimated time of the current load work orders is divided by the technician's standard time capacity to obtain the time saturation. The time saturation is corrected by the mobility efficiency factor in combination with the driver's travel speed and destination distance in the geographical location movement status. The final resource utilization rate is the weighted geometric mean of the corrected time saturation and the current load work order number.

8. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 7, characterized in that: When applying a suppression factor to non-urgent work orders, the suppression intensity parameter is obtained by looking up the table based on the average resource utilization rate. The suppression factor is represented as the decay coefficient of the original priority of the non-urgent work order. This coefficient decreases monotonically from the benchmark value to the minimum value as the resource utilization rate increases. When prioritizing the allocation of idle workers, a two-way matching queue is established. Workers with resource utilization rates below the second threshold are arranged in descending order of idleness. New work orders are arranged in descending order of dynamic priority weight. Real-time matching of the work order with the highest weight and the worker with the highest idleness is achieved through a rolling time window.

9. The method for dynamic scheduling of vehicle work orders based on multimodal fusion according to claim 8, characterized in that: The global priority ranking adopts a multi-factor weighted aggregation under dynamic strategy constraints. The dynamic priority weight of each work order is adjusted by introducing a suppression factor. At the same time, a heat map of resource utilization distribution of each master is loaded, and a regional delay coefficient is applied to work orders in high-load areas. The final ranking value is determined by the weighted sum of the adjusted priority weight, regional delay coefficient, and work order timeliness decay factor. The ranking engine recalculates the weight ratio of each factor at preset intervals, and the weight ratio is dynamically configured according to the system's average resource utilization.

10. A system for implementing a multimodal fusion-based dynamic scheduling method for train work orders, comprising any one of claims 1-9, characterized in that, include: The multimodal data acquisition and fusion unit (1) extracts and fuses work order text, images and structured data through a cross-modal semantic alignment model to generate work order feature vectors containing business scenario semantics; The initial priority calculation unit (2) connects to the multimodal data acquisition and fusion unit (1) to calculate the initial priority weight based on the preset business scenario rule set, and uses a piecewise nonlinear function to generate an emergency boost coefficient. Combined with the emergency compensation coefficient generated inversely proportional to the resource utilization rate, the dynamic priority weight is output. The emergency order processing unit (3) and the resource utilization calculation unit (4) calculate the master's resource utilization rate through the working time entropy value model, trigger the inhibition factor application mechanism according to the utilization rate threshold, and build a two-way matching queue to realize the real-time matching of work orders and idle masters; The global sorting and scheduling unit (5) integrates dynamic priority weights, suppression factor adjustment results and regional delay coefficients, and generates a global sort through a multi-factor weighted aggregation engine to drive the execution of work orders and the optimal allocation instructions of the master.

Images