A yard operation autonomous decision-making scheduling method based on a multi-modal perception model

Abstract

The application belongs to the technical field of air logistics station operation control, and provides a station operation autonomous decision scheduling method based on a multi-modal perception model, specifically comprising the following steps: obtaining a plurality of regions of cargo sorting types in an air logistics station operation scene, obtaining the characteristic properties of each region, respectively determining the initial weights of the multi-modal data allocated to each region, constructing an initial multi-modal perception model, extracting regions similar in initial weight distribution as an initial similar region group; in the actual logistics sorting process, obtaining the real-time running characteristics of each region, extracting regions similar in real-time running characteristics in the initial similar region group as a running similar region group; extracting the region with the largest change in real-time running characteristics in the running similar region group to determine the precision evaluation target of the initial multi-modal perception model, and evaluating the precision of the initial multi-modal perception model.

Description

Technical Field

[0001] This invention belongs to the field of aviation logistics terminal operation technology, specifically a terminal operation autonomous decision-making and scheduling method based on a multimodal perception model. Background Technology

[0002] Air logistics terminal operations refer to a series of work activities related to cargo handling, transportation, storage, and information management carried out within logistics terminals such as airports, centered around the core link of air transport in the air logistics system.

[0003] Air logistics terminal operations involve multiple modalities of data, such as image data (from surveillance cameras, used to identify the appearance and placement of goods), text data (such as flight information and cargo manifests), and sensor data (such as temperature and humidity sensor data to ensure a suitable storage environment for goods). Air logistics terminals have different types of logistics areas, including cargo sorting areas, cargo storage areas, and aircraft docking areas.

[0004] When some air logistics terminals use multimodal perception models for autonomous decision-making and scheduling of each logistics area of ​​the same type, they mostly use the same model construction and the same weight allocation for multimodal data. However, in actual operation, for the same type of logistics area, such as cargo sorting area, there are differences in equipment configuration, spatial layout and operation stage. If a uniform allocation mechanism is used, it may reduce the accuracy of decision output and affect the efficiency of terminal operation. Therefore, it is necessary to set differentiated settings and dynamic adjustments for the weight allocation of multimodal data for several areas of the same type.

[0005] Therefore, this invention provides a method for autonomous decision-making and scheduling of field operations based on a multimodal perception model. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is: a method for autonomous decision-making and scheduling of field operations based on a multimodal perception model, specifically including the following steps:

[0008] In the operation scenario of air logistics terminal, several areas of cargo sorting type are obtained, the characteristics of each area are obtained, the initial weight of multimodal data is determined for each area, an initial multimodal perception model is constructed, and areas with similar initial weight allocation are extracted as the initial similar area group.

[0009] During the actual logistics sorting process, the real-time operating characteristics of each area are obtained, and areas with similar real-time operating characteristics are extracted from the initial similar area group as the operating similar area group.

[0010] The region with the greatest change in real-time operating characteristics in the similar operating region group is extracted to determine the accuracy evaluation target of the initial multimodal sensing model. The accuracy of the initial multimodal sensing model is evaluated. If the accuracy of the initial multimodal sensing model is low, the initial weights of the corresponding multimodal data in the similar operating region group are allocated and adjusted based on the real-time operating characteristics to obtain an adaptive multimodal sensing model.

[0011] Based on the adaptive multimodal perception model, the resource utilization rate of each region is output, the level of resource utilization is evaluated, and decision scheduling is performed based on the level of resource utilization.

[0012] As a further aspect of the present invention: the initial weights of each modal data in each region are determined by the analytic hierarchy process, and a multimodal perception model for each region is constructed using a neural network model based on the initial weights of each modal data in each region.

[0013] As a further aspect of the present invention: the process for obtaining the initial similar region group is as follows:

[0014] For any two regions, calculate the difference between the initial weights of the same modal data and take the absolute value as the absolute weight difference. Take the average of the absolute weight differences of all modal data to obtain the average absolute weight difference.

[0015] Extract the maximum and minimum values ​​of the absolute differences of weights for all modal data, calculate the difference, and sum the obtained difference with the mean of the absolute differences of weights to obtain the weight difference degree.

[0016] When the weight difference is less than the weight difference limit, the weight allocation of the two regions is similar, and the two regions are marked as similar regions. The regions marked as similar regions are extracted and integrated into an initial similar region group.

[0017] As a further aspect of the present invention: the process of obtaining the similar region group is as follows:

[0018] Extract the real-time operational characteristic data sequence of one region from the initial similar region group as the reference sequence, and use the real-time operational characteristic data of the remaining regions as the comparison sequence;

[0019] For each comparison sequence, calculate its correlation coefficient with the reference sequence at each time step, average the correlation coefficients at each time step to obtain the correlation degree between the comparison sequence and the reference sequence, extract regions with a correlation degree greater than a preset threshold, and integrate them into a running similar region group.

[0020] As a further aspect of the present invention: the process of determining the accuracy evaluation target of the initial multimodal sensing model is as follows:

[0021] For the real-time operational characteristic data of each region in the similar operational region group, calculate its comprehensive value of change, and extract the accuracy evaluation target of the initial multimodal sensing model of the region with the maximum value of the comprehensive value of change in each similar operational region group.

[0022] As a further aspect of the present invention: the process for obtaining the comprehensive value of the degree of change is as follows:

[0023] For the real-time operational characteristic data of each region in the similar operational region group, calculate its corresponding information entropy and variance.

[0024] The combined output of information entropy and variance yields a comprehensive value for the degree of change.

[0025] As a further aspect of the present invention: the process of evaluating the accuracy of the initial multimodal sensing model is as follows:

[0026] Obtain the resource utilization rate at multiple moments within the real-time monitoring period, and use it as the first resource utilization rate sequence;

[0027] Obtain the resource utilization rate output by the initial multimodal perception model as the second resource utilization rate sequence;

[0028] The coefficient of determination is calculated based on the first resource utilization rate sequence and the second resource utilization rate sequence;

[0029] If the coefficient of determination is less than the preset limit, it indicates that the initial multimodal sensing model has low accuracy.

[0030] As a further aspect of the present invention: the process of the adaptive multimodal sensing model is as follows:

[0031] The hierarchical structure model is reconstructed using the analytic hierarchy process (AHP), the judgment matrix is ​​updated, and the weights of each modality are redistributed. Based on the redistributed weights of each modality, an adaptive multimodal perception model is obtained.

[0032] As a further aspect of the present invention: the process of evaluating the level of resource utilization and making decision-making and scheduling is as follows:

[0033] Real-time multimodal data for each region is input into the adaptive multimodal sensing model, which outputs the resource utilization rate for each region.

[0034] If the resource utilization rate is less than or equal to the resource utilization level judgment value, the region is identified as having a low resource utilization level, and an optimization strategy is implemented to improve the resource utilization rate.

[0035] If the resource utilization rate is greater than the resource utilization level judgment value, then the resource utilization level of the area is identified as high, and the utilization rate of goods sorting is increased.

[0036] As a further aspect of the present invention: the execution decision scheduling process further includes:

[0037] Locate the operational similarity group corresponding to each region with low resource utilization, and obtain the corresponding comprehensive value of the degree of change.

[0038] For each region with low resource utilization, the corresponding comprehensive value of change is normalized to the resource utilization rate. The ratio of the normalized comprehensive value of change to the resource utilization rate is then calculated to obtain the ranking value.

[0039] The higher the ranking value, the higher the urgency of optimizing the corresponding low resource utilization area. Optimization strategies are then implemented for the low resource utilization area based on the urgency of optimization.

[0040] The beneficial effects of this invention are as follows:

[0041] This invention optimizes the limitations of the traditional unified weight allocation mechanism by setting and dynamically adjusting the weights of multimodal data for the same type of logistics area in a differentiated manner. It determines the initial weights based on the characteristics of the region and constructs a multimodal perception model, which can accurately fit the actual situation of different regions and improve the accuracy of decision output. At the same time, by dividing the initial similar region groups, it provides a scientific grouping basis for subsequent management and optimization.

[0042] In actual operation, this invention acquires regional operating characteristics in real time and extracts similar operating region groups, and extracts the initial multimodal sensing model corresponding to the region with the greatest degree of change to evaluate the model's accuracy. This not only reduces the computational effort and saves resources, but also provides representativeness.

[0043] Ultimately, the resource utilization rate output by the adaptive model enables the execution of decision-making and scheduling, which can accurately identify the resource utilization status, adopt optimization strategies and resource allocation for areas with high and low utilization rates respectively, and prioritize key areas by prioritizing them according to their urgency. This achieves balanced allocation and efficient utilization of resources within the terminal, improving the overall efficiency and intelligence level of air logistics terminal operations. Attached Figure Description

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart illustrating the steps of an autonomous decision-making and scheduling method for field operations based on a multimodal perception model, as described in this invention.

[0046] Figure 2This is a flowchart of the steps of a site operation autonomous decision-making and scheduling method based on a multimodal perception model according to the present invention;

[0047] Figure 3 This is a flowchart of a field operation autonomous decision-making and scheduling system based on a multimodal perception model, according to the present invention. Detailed Implementation

[0048] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0049] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a method for autonomous decision-making and scheduling of field operations based on a multimodal perception model specifically includes the following steps:

[0050] Step 1: In the operation scenario of the air logistics terminal, obtain several areas of cargo sorting type, obtain the characteristics of each area, determine the initial weight of multimodal data for each area, construct an initial multimodal perception model, and extract areas with similar initial weight allocation as the initial similar area group.

[0051] Specifically, in the scenario of air logistics terminal operations, for cargo sorting, the location of cargo sorting in the air logistics terminal is obtained as the cargo sorting area;

[0052] The cargo sorting area is divided into several zones, with the following criteria: based on the cargo's destination, the area is divided into domestic cargo sorting area and international cargo sorting area; based on cargo type, it is divided into general cargo sorting area and dangerous goods sorting area; and based on sorting stages, it is divided into initial sorting area and secondary sorting area.

[0053] Based on the zoning standards, physical and electronic tags are used to actually divide the goods sorting area; physical tags include fences and markings; electronic tags include RFID tags and QR codes.

[0054] Each divided area is numbered and marked to establish an area information database;

[0055] The characteristics of each region are obtained, including: spatial layout characteristics, equipment configuration characteristics, work process characteristics, and environmental characteristics.

[0056] Furthermore, the spatial layout features are measured using measuring tools such as laser rangefinders or total stations to measure the spatial layout of each area, and a spatial layout map of the area is drawn to intuitively show the physical structure of the area. The layout includes at least: the length, width, height, aisle width, and shelf arrangement of the area.

[0057] Equipment configuration characteristics are recorded by specifying the type, model, quantity, performance parameters, installation location, and operating status of the equipment; among which, the equipment type includes at least: sorting machine, conveyor belt, and barcode scanner; and the performance parameters include at least: processing speed, accuracy, and load capacity.

[0058] Environmental characteristics: Install environmental monitoring equipment to monitor environmental parameters in each area in real time. Environmental parameters include, but are not limited to: temperature, humidity, light intensity, and noise level.

[0059] The workflow characteristics are defined by recording the actual process of goods sorting and drawing workflow diagrams for each area, clarifying the flow path of goods, operational steps, and the connection between each step;

[0060] The collected multimodal data types include image data, text data, and sensor data;

[0061] For example, the image clearly shows the external structure of a certain type of sorting machine, including its conveyor track, sorting port, and other components; the type, model, quantity, performance parameters, installation location, and operating status information of the equipment are recorded, namely, "Equipment type: sorting machine, model: XYZ-100, quantity: 2 units, performance parameters: processing speed 1000 pieces / hour, accuracy 99.5%, load capacity 50kg, installation location: east side of the domestic goods sorting area, operating status: normal operation"; during a certain period of time, the temperature data recorded by the temperature sensor in the international goods sorting area is 22℃-25℃, and the humidity data recorded by the humidity sensor is 40%-50%;

[0062] The initial weights of each modal data in each region are determined by the analytic hierarchy process, and a multimodal sensing model is constructed based on the initial weights of each modal data in each region.

[0063] As will be understood by those skilled in the art, the Analytic Hierarchy Process (AHP) is used to determine the initial weights of each modality data in each region. First, a hierarchical model is constructed, comprising a target layer, a criterion layer, and a scheme layer. The target layer, criterion layer, and scheme layer determine the weight allocation, the factors influencing the weights, and the modality data for each region, respectively. Next, a judgment matrix is ​​constructed by comparing the scheme layer data pairwise for each criterion in the criterion layer. The hierarchical single-rank weight is obtained by calculating and normalizing the feature vectors, and a consistency check is performed. Then, the hierarchical total ranking weight is calculated and consistency is checked to obtain the initial weights. After obtaining the initial weights, the modality data is preprocessed, features are extracted, and a multimodal perception model is constructed by weighted fusion based on the weights. Finally, the model is trained and optimized using a labeled dataset.

[0064] Constructing multimodal perception models using machine learning or deep learning algorithms is a well-known method in the field. Taking a neural network model as an example, the model architecture includes an input layer, hidden layers, and an output layer. The input layer contains weighted and fused multimodal features, the hidden layers use fully connected layers or convolutional layers (for image data) to extract high-order features, and the output layer outputs the resource utilization of the output region. The model training process includes supervised learning using a dataset, optimizing network parameters using the backpropagation algorithm to minimize the loss function, adjusting hyperparameters using a validation set to prevent overfitting, accelerating convergence using the Adam optimizer, and using Dropout or L2 regularization to prevent overfitting.

[0065] Based on the initial weights of each modality data in each region, regions with similar initial weight assignments are extracted;

[0066] For any two regions, calculate the difference between the initial weights of the same modal data and take the absolute value as the absolute weight difference. Take the average of the absolute weight differences of all modal data to obtain the average absolute weight difference.

[0067] Extract the maximum and minimum values ​​of the absolute differences of weights for all modal data, calculate the difference, and sum the obtained difference with the mean of the absolute differences of weights to obtain the weight difference degree.

[0068] When the weight difference is greater than or equal to the weight difference limit, the weight allocation of the two regions is not similar. When the weight difference is less than the weight difference limit, the weight allocation of the two regions is similar. These two regions are marked as similar regions.

[0069] The difference limit is determined based on the experience of those skilled in the art or historical data, to determine the acceptable range of weight differences, in order to distinguish whether two regions are similar in their initial weight allocation;

[0070] The purpose of distinguishing whether two regions have similar initial weight allocations is twofold: firstly, to facilitate analysis and management, in air logistics terminals, similar regions can adopt the same resource allocation strategies, work process optimization schemes, or equipment maintenance plans to improve overall logistics efficiency and quality; secondly, by distinguishing regions with similar weight allocations, local optimization can be carried out for specific regions or groups of regions, reducing optimization complexity and improving optimization efficiency.

[0071] Regions marked as similar are extracted and integrated into an initial group of similar regions;

[0072] Step 2: During the actual logistics sorting process, obtain the real-time operating characteristics of each area, and extract areas with similar real-time operating characteristics from the initial similar area group as the similar operating area group;

[0073] Specifically, during the actual cargo sorting operation at the air logistics terminal, each area corresponds to its real-time operating characteristics. Therefore, the initial weight allocation may not be suitable for the real-time operating characteristics, and the weight allocation of the inapplicable areas needs to be adjusted. This step is to extract areas with similar real-time operating characteristics from the initial group of similar areas.

[0074] The reason for extracting regions with similar real-time operating characteristics is that different regions may exhibit different characteristics in actual operation. Therefore, it is necessary to formulate differentiated weight adjustment strategies. By extracting regions with similar real-time operating characteristics, a unified weight adjustment strategy can be applied to regions that are initially similar and have similar real-time operating characteristics, reducing management difficulty and improving the efficiency of weight adjustment.

[0075] Specifically, a real-time monitoring period is set to obtain real-time operational characteristic data for each area. This real-time operational characteristic data includes: cargo flow data, equipment operating efficiency data, and sorting efficiency data.

[0076] Cargo flow is the number of goods passing through the area per unit time. The number of goods entering and leaving the area is counted in real time by counting sensors at the cargo entrance and exit, and the cargo flow per unit time is calculated by combining the time information.

[0077] Equipment operating efficiency refers to various performance indicators of the equipment during operation, such as the processing speed of the sorting machine, the scanning success rate of the barcode scanner, and the running speed of the conveyor belt; the scanning success rate is calculated by recording the number of scans and the number of successful scans.

[0078] Sorting efficiency is the time required to complete a unit of goods sorting task and the accuracy of the operation. It is achieved by collecting the time to complete each goods sorting task, calculating the average operation time, establishing a quality inspection process, checking the operation results, and calculating the operation accuracy.

[0079] The real-time operating characteristics of each region are analyzed, and regions with similar real-time operating characteristics are extracted from the initial group of similar regions.

[0080] The grey relational analysis method has advantages in terms of data requirements, computational complexity, and sensitivity to trend changes.

[0081] Extract the real-time operational characteristic data sequence of one region from the initial similar region group as the reference sequence, and use the real-time operational characteristic data of the remaining regions as the comparison sequence;

[0082] Dimensionless processing is performed on real-time operational characteristic data, using methods such as initialization and meanization.

[0083] For each comparison sequence, calculate its correlation coefficient with the reference sequence at each time step, and take the average of the correlation coefficients at each time step to obtain the correlation degree between the comparison sequence and the reference sequence.

[0084] Regions with a correlation greater than a preset correlation threshold are extracted and integrated into a group of similar regions.

[0085] Step 3: Extract the region with the greatest change in real-time operating characteristics from the similar operating region group to determine the accuracy evaluation target of the initial multimodal sensing model. If the accuracy of the initial multimodal sensing model is low, the initial weights of the corresponding multimodal data in the similar operating region group are allocated and adjusted based on the real-time operating characteristics to obtain an adaptive multimodal sensing model.

[0086] Specifically, the process of determining the accuracy assessment target of the initial multimodal sensing model in each similar operating region is as follows:

[0087] In air logistics terminals, the changes in real-time operational characteristics data are complex and diverse. Information entropy can comprehensively consider all possible values ​​of the data and their probabilities of occurrence, fully reflecting the uncertainty of data changes.

[0088] For example, suppose there are two regions with cargo flow data. The cargo flow in region A is mostly concentrated between 200-300 pieces / hour, with only a small amount of data deviating from this range. The cargo flow in region B fluctuates more evenly between 100-400 pieces / hour. Intuitively, the data in region B has a larger fluctuation range. However, if variance is used to measure it, it may not accurately reflect the difference in uncertainty between the two. Information entropy can comprehensively consider the probability of each value interval and more accurately assess that the data in region B has higher uncertainty and greater variability.

[0089] For the real-time operational characteristic data of each region in the similar operational region group, calculate its corresponding information entropy;

[0090] For real-time runtime characteristic data, statistics Given its probability distribution, the formula for calculating information entropy is: Where n is the total amount of data, Let i be the probability distribution of the i-th data.

[0091] Variance reflects the degree of dispersion of data. When combined with information entropy, it can be used to analyze the degree of change in real-time operational characteristic data, which is complementary and increases the comprehensiveness of the analysis.

[0092] For the real-time operational characteristic data of each region in the similar operational region group, calculate its corresponding variance;

[0093] The weighted average of information entropy and variance is used to calculate the comprehensive value of the degree of change. The weight of information entropy can be assigned to 0.6 and the weight of variance to 0.4. This setting is because the core objective of air logistics is efficient scheduling and resource optimization, and it is necessary to focus on the complexity and diversity of data changes rather than the simple fluctuation range.

[0094] Before weighting information entropy and variance, these two metrics need to be normalized.

[0095] The accuracy evaluation target of the initial multimodal sensing model in each running similar region group is to extract the region with the maximum value of the comprehensive value of the degree of change.

[0096] The reason for selecting the region corresponding to the maximum value of the comprehensive value of the degree of change is that: extracting the region with the greatest degree of change as the "key sample" can reduce the amount of calculation from "full evaluation" to "single point breakthrough", reducing the computational resources required; the region with the greatest comprehensive value of the degree of change has the strongest fluctuation range and uncertainty in data such as cargo flow and equipment efficiency, and is the most complex in the entire group of similar operating regions, and is the "typical sample" that is most likely to cause model failure, and is representative.

[0097] Specifically, the process of evaluating the accuracy of the initial multimodal sensing model is as follows:

[0098] Obtain the resource utilization rate at multiple moments within the real-time monitoring period, and use it as the first resource utilization rate sequence;

[0099] Obtain the resource utilization rate output by the initial multimodal perception model as the second resource utilization rate sequence;

[0100] The coefficient of determination is calculated based on the first resource utilization rate sequence and the second resource utilization rate sequence;

[0101] If the coefficient of determination is greater than or equal to the preset limit, it indicates that the initial multimodal perception model meets the real-time operation characteristics; if the coefficient of determination is less than the preset limit, it indicates that the initial multimodal perception model does not meet the real-time operation characteristics; the initial weights of each modality data need to be redistributed.

[0102] Specifically, the process of redistributing the initial weights of each modality of data is as follows:

[0103] Since the initial weights were determined using the analytic hierarchy process, it is necessary to rebuild the hierarchical model, update the judgment matrix, and redistribute the weights of each modality.

[0104] S301: Reconstruct the hierarchical model;

[0105] Target layer: Define the decision-making objectives, namely, optimize the weight allocation of the multimodal perception model to improve the accuracy and efficiency of autonomous decision-making and scheduling of station operations;

[0106] Criteria Layer: Based on the actual needs of the site operations and real-time operational characteristic data, the key factors affecting the weight allocation are re-determined. These factors include, but are not limited to, cargo flow, equipment operating efficiency, sorting efficiency, and environmental parameters, which are summarized by those skilled in the art based on the actual needs of the site operations, real-time operational characteristic data, and experience.

[0107] Solution layer: Lists the data for each modality, serving as the specific objects for weight allocation;

[0108] S302: Update the judgment matrix, which mainly includes data collection and quantification and construction of the judgment matrix;

[0109] Collect the aforementioned numerical quantitative data, including the performance of each modality in real-time operational characteristics, such as cargo flow, equipment operating efficiency, and their impact on decision objectives; convert this quantitative data into a form that can be used to construct a judgment matrix;

[0110] For example, the importance of each modality of data at the criterion level can be quantitatively scored through statistical analysis or expert scoring.

[0111] Based on quantitative scoring, a judgment matrix is ​​constructed from the criterion layer to the scheme layer. The elements in the judgment matrix represent the relative importance of each modality data under the corresponding criterion.

[0112] For example, if cargo flow is one of the key factors affecting weight allocation, then the performance of each modality in terms of cargo flow can be compared and corresponding weights can be assigned to them.

[0113] S303: Recalculate the weights of each modal data;

[0114] Hierarchical single ranking: Calculate the eigenvectors of each element in the judgment matrix and normalize them to obtain the hierarchical single ranking weights, which represent the relative importance of each modality data under the corresponding criteria.

[0115] Overall hierarchical ranking: Based on the weight of the criterion layer and the hierarchical single ranking weight of each modality under the criterion layer, the overall hierarchical ranking weight of each modality is calculated, which represents the comprehensive importance of each modality under the decision objective;

[0116] Results Analysis and Adjustment: Analyze the calculated weight results to ensure that they meet the actual needs of the site operation and the real-time operating characteristics. If necessary, the weight results can be fine-tuned to further optimize the performance of the multimodal sensing model.

[0117] An adaptive multimodal perception model is obtained based on the weights of the reassigned modal data.

[0118] This embodiment has at least the following effects: by extracting regions with similar initial weight allocations to form an initial similar region group, it is convenient to apply the same resource allocation strategy, work process optimization scheme or equipment maintenance plan to similar regions, thereby improving the overall logistics efficiency and quality. At the same time, local optimization for specific regions or region groups can reduce optimization complexity and improve optimization efficiency.

[0119] In the actual logistics sorting process, obtaining the real-time operating characteristics of each area can promptly identify situations where the initial weight allocation is not applicable to the real-time operating characteristics, providing a basis for weight adjustment.

[0120] By extracting regions with similar real-time operating characteristics from the initial similar region group to form a similar operating region group, a unified weight adjustment strategy can be applied to regions that are initially similar and operate similarly in real time, reducing management difficulty and improving the efficiency of weight adjustment.

[0121] By calculating the weighted average change value of information entropy and variance, the region with the greatest change in real-time operating characteristics is determined as the accuracy evaluation target. If the initial model accuracy is low, the weights are reallocated so that the model can be dynamically optimized according to real-time operating characteristics, thereby improving the model's adaptability and accuracy.

[0122] Information entropy and variance fusion analysis of the degree of change in real-time operational characteristic data are complementary, and can comprehensively reflect the uncertainty and dispersion of data changes, making the assessment of the degree of regional change more accurate.

[0123] An adaptive multimodal perception model is obtained by adjusting the weights based on real-time operational characteristics. This model can more accurately reflect the actual situation in the region, thereby improving the accuracy and efficiency of autonomous decision-making and scheduling of field operations and optimizing resource utilization.

[0124] Example 2: Please refer to Figure 2 As shown in the embodiment of the present invention, a site operation autonomous decision-making and scheduling method based on a multimodal perception model further includes the following steps:

[0125] Step 4: Based on the adaptive multimodal perception model, output the resource utilization rate of each region, and execute decision scheduling based on the resource utilization rate;

[0126] Specifically, real-time multimodal data for each region is input into the adaptive multimodal perception model, which outputs the resource utilization rate for each region.

[0127] The resource utilization data of all regions are compiled and the resource utilization level of each region is evaluated to formulate autonomous decision-making and scheduling methods.

[0128] The resource utilization rate is compared with the resource utilization level judgment value. If the resource utilization rate is less than or equal to the resource utilization level judgment value, the resource utilization level of the area is identified as low. If the resource utilization rate is greater than the resource utilization level judgment value, the resource utilization level of the area is identified as high.

[0129] The resource utilization rate assessment value is set by those skilled in the art based on the characteristics of operations at air logistics terminals in this field;

[0130] For areas with low resource utilization, implement optimization strategies to improve resource utilization. Optimization strategies include, but are not limited to: increasing equipment investment, optimizing work processes, and adjusting personnel allocation.

[0131] For areas with high resource utilization, the utilization rate of cargo sorting in that area can be increased, including but not limited to: increasing the number of work shifts;

[0132] Optionally, extract all areas with low resource utilization and sort them by their urgency of optimization;

[0133] Optimizing the ranking of urgency levels can identify and prioritize areas that have the greatest impact on overall efficiency, reducing the blockage and delays in the scheduling process caused by these areas. This can gradually improve resource utilization, promote balanced use of resources within the terminal, reduce resource idleness and waste, and contribute to the overall optimization and efficiency improvement of air logistics terminal operations.

[0134] Specifically, locate the operational similarity group corresponding to each region with low resource utilization, and obtain the corresponding comprehensive value of the degree of change;

[0135] For each region with low resource utilization, the corresponding comprehensive value of change is normalized to the resource utilization rate. The ratio of the normalized comprehensive value of change to the resource utilization rate is then calculated to obtain the ranking value.

[0136] For ranking values, the larger the overall value of the degree of change, the greater the degree of change in the region during real-time operation, and the higher the urgency of optimization. The lower the resource utilization, the higher the urgency of optimization. Therefore, the larger the ranking value, the higher the urgency of optimization.

[0137] The higher the ranking value, the more urgent the optimization of the corresponding low resource utilization area.

[0138] The reason for using resource utilization rate and comprehensive change rate to prioritize optimization urgency is as follows: resource utilization rate reflects the current resource use efficiency of a region and is an important indicator for evaluating regional performance; the comprehensive change rate reflects the degree of change in the real-time operating characteristics of the region and can capture the dynamic changes in the region's status. Combining these two indicators allows for a more comprehensive assessment of the region's status. By calculating the ratio of resource utilization rate to comprehensive change rate, regions with low resource utilization and high degree of change can be identified. Prioritizing these regions can improve resource utilization more quickly, reduce resource idleness and waste, and thus improve overall efficiency.

[0139] This embodiment has at least the following effects: by outputting the resource utilization rate of each region through an adaptive multimodal perception model and comparing it with the judgment value, it can accurately identify the level of resource utilization and provide a quantitative basis for decision-making and scheduling.

[0140] By normalizing the comprehensive value of the degree of change and the resource utilization rate and calculating the ranking value, areas with low resource utilization rate can be ranked according to the degree of urgency, and areas with a large impact on overall efficiency can be prioritized to reduce the risk of scheduling congestion and gradually improve the resource utilization efficiency of the entire station.

[0141] The scheduling strategy is automatically executed based on the quantitative data output by the model, reducing manual intervention, improving the automation and intelligence of decision-making and scheduling, adapting to the high-efficiency operation requirements of air logistics terminals, and promoting the optimization of the overall operation process.

[0142] Example 3: Based on the same inventive concept as the multimodal perception model-based autonomous decision-making and scheduling method for station operations in the foregoing examples, such as... Figure 3 As shown, this application provides a field operation autonomous decision-making and scheduling system based on a multimodal perception model, wherein the system specifically includes:

[0143] Initial weight analysis module: In the operation scenario of air logistics terminal, several areas of cargo sorting type are obtained, the characteristics of each area are obtained, the initial weight of multimodal data is determined for each area, an initial multimodal perception model is constructed, and areas with similar initial weight allocation are extracted as the initial similar area group.

[0144] Real-time operation analysis module: During the actual logistics sorting process, the module acquires the real-time operation characteristics of each area and extracts areas with similar real-time operation characteristics from the initial similar area group as the operation similar area group.

[0145] Weight update module: Extract the region with the greatest change in real-time running characteristics in the similar running region group to determine the accuracy evaluation target of the initial multimodal perception model. If the accuracy of the initial multimodal perception model is low, the initial weights of the corresponding multimodal data in the similar running region group are allocated and adjusted based on the real-time running characteristics to obtain an adaptive multimodal perception model.

[0146] Decision execution module: Based on the adaptive multimodal perception model, outputs the resource utilization rate of each region and executes decision scheduling based on the resource utilization rate.

[0147] 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 illustrative of the principles of 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 yard operation autonomous decision-making scheduling method based on a multi-modal perception model, characterized in that: Specifically, the following steps are included: In the operation scenario of air logistics terminal, several areas of cargo sorting type are obtained, the characteristics of each area are obtained, the initial weight of multimodal data is determined for each area, an initial multimodal perception model is constructed, and areas with similar initial weight allocation are extracted as the initial similar area group. The process for obtaining the initial group of similar regions is as follows: For any two regions, calculate the difference between the initial weights of the same modal data and take the absolute value as the absolute weight difference. Take the average of the absolute weight differences of all modal data to obtain the average absolute weight difference. Extract the maximum and minimum values ​​of the absolute differences of weights for all modal data, calculate the difference, and sum the obtained difference with the mean of the absolute differences of weights to obtain the weight difference degree. When the weight difference is less than the weight difference limit, the weight allocation of the two regions is similar, and the two regions are marked as similar regions. Regions marked as similar are extracted and integrated into an initial group of similar regions; During the actual logistics sorting process, the real-time operating characteristics of each area are obtained, and areas with similar real-time operating characteristics are extracted from the initial similar area group as the operating similar area group. The region with the greatest change in real-time operating characteristics in the similar operating region group is extracted to determine the accuracy evaluation target of the initial multimodal sensing model. The accuracy of the initial multimodal sensing model is evaluated. If the accuracy of the initial multimodal sensing model is low, the initial weights of the corresponding multimodal data in the similar operating region group are allocated and adjusted based on the real-time operating characteristics to obtain an adaptive multimodal sensing model. The process for obtaining the similar region group is as follows: Extract the real-time operational characteristic data sequence of one region from the initial similar region group as the reference sequence, and use the real-time operational characteristic data of the remaining regions as the comparison sequence; For each comparison sequence, calculate its correlation coefficient with the reference sequence at each time step, average the correlation coefficients at each time step to obtain the correlation degree between the comparison sequence and the reference sequence, extract regions with a correlation degree greater than a preset threshold, and integrate them into a running similar region group. The process of determining the accuracy evaluation target of the initial multimodal sensing model is as follows: For the real-time operational characteristic data of each region in the similar operational region group, calculate its comprehensive value of change, and extract the accuracy evaluation target of the initial multimodal sensing model of the region with the maximum value of the comprehensive value of change in each similar operational region group; The process of obtaining the comprehensive value of the degree of change is as follows: calculate the information entropy corresponding to the real-time operating characteristic data of each region in the similar operating region group, and calculate the variance corresponding to the real-time operating characteristic data of each region in the similar operating region group. The combined output of information entropy and variance yields a comprehensive value for the degree of change. Based on the adaptive multimodal perception model, the resource utilization rate of each region is output, the level of resource utilization is evaluated, and decision scheduling is performed based on the level of resource utilization.

2. The method for autonomous decision-making and scheduling of station operations based on a multimodal perception model according to claim 1, characterized in that: The initial weights of each modality data in each region are determined by the analytic hierarchy process (AHP). Based on the initial weights of each modality data in each region, a neural network model is used to construct a multimodal perception model for each region.

3. The method for autonomous decision-making and scheduling of station operations based on a multimodal perception model according to claim 1, characterized in that: The process of evaluating the accuracy of the initial multimodal sensing model is as follows: Obtain the resource utilization rate at multiple moments within the real-time monitoring period, and use it as the first resource utilization rate sequence; Obtain the resource utilization rate output by the initial multimodal perception model as the second resource utilization rate sequence; The coefficient of determination is calculated based on the first resource utilization rate sequence and the second resource utilization rate sequence; If the coefficient of determination is less than the preset limit, it indicates that the initial multimodal sensing model has low accuracy.

4. The method for autonomous decision-making and scheduling of station operations based on a multimodal perception model according to claim 1, characterized in that: The process of the adaptive multimodal sensing model is as follows: The hierarchical structure model is reconstructed using the analytic hierarchy process (AHP), the judgment matrix is ​​updated, and the weights of each modality are redistributed. Based on the redistributed weights of each modality, an adaptive multimodal perception model is obtained.

5. The method for autonomous decision-making and scheduling of station operations based on a multimodal perception model according to claim 1, characterized in that: The process of assessing the level of resource utilization and making decision-making and scheduling is as follows: Real-time multimodal data for each region is input into the adaptive multimodal sensing model, which outputs the resource utilization rate for each region. If the resource utilization rate is less than or equal to the resource utilization level judgment value, the region is identified as having a low resource utilization level, and an optimization strategy is implemented to improve the resource utilization rate. If the resource utilization rate is greater than the resource utilization level judgment value, then the resource utilization level of the area is identified as high, and the utilization rate of goods sorting is increased.

6. The method for autonomous decision-making and scheduling of station operations based on a multimodal perception model according to claim 5, characterized in that: The execution decision scheduling process also includes: Locate the operational similarity group corresponding to each region with low resource utilization, and obtain the corresponding comprehensive value of the degree of change. For each region with low resource utilization, the corresponding comprehensive value of change is normalized to the resource utilization rate. The ratio of the normalized comprehensive value of change to the resource utilization rate is then calculated to obtain the ranking value. The higher the ranking value, the higher the urgency of optimizing the corresponding low resource utilization area. Optimization strategies are then implemented for the low resource utilization area based on the urgency of optimization.

Images