Business data mining method and system applied to intelligent port scheduling

Abstract

The invention provides a business data mining method and system applied to intelligent port scheduling, and the method comprises the steps: firstly obtaining real-time operation data containing an equipment operation state, a transportation task queue and environment monitoring information during the loading and unloading operation of a target port, then carrying out the dynamic preprocessing of the real-time operation data, screening and marking conflict tasks according to a preset rule, and obtaining effective operation data; and then performing multi-dimensional feature coding on the effective data to generate an operation feature set containing equipment operation features, task distribution features and environment constraint features, and based on the operation feature set, generating a port equipment scheduling strategy meeting balance of operation efficiency and resource consumption through a pre-training model, covering an equipment start-stop sequence and the like. And finally, dynamically adjusting port equipment to execute tasks according to a scheduling strategy, and updating real-time operation data, thereby realizing efficient and reasonable scheduling of the intelligent port.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology. Specifically, it relates to a business data mining method and system applied to intelligent port scheduling. Background Art

[0002] With the continuous growth of global trade and the rapid development of the logistics industry, intelligent ports, as an important part of the modern logistics system, their efficient and intelligent scheduling management has become the key to improving port operation efficiency. Intelligent ports achieve comprehensive monitoring and intelligent scheduling of the port operation process by integrating advanced information technology, Internet of Things technology, big data analysis, and artificial intelligence algorithms. However, in a complex port environment, how to effectively mine and utilize business data to optimize the port equipment scheduling strategy is still a challenging topic.

[0003] Traditional port scheduling methods mainly rely on manual experience and simple rule systems. These methods can work when dealing with small-scale and low-complexity port operations, but as the port scale expands and the operation complexity increases, their limitations become increasingly prominent. Manual scheduling is not only time-consuming and laborious but also difficult to respond to the dynamic changes in the port operation process in real time. Simple rule systems lack flexibility and intelligence and cannot be dynamically adjusted and optimized according to the actual situation of port operations.

[0004] To overcome the deficiencies of traditional scheduling methods, in related technologies, some data-driven port scheduling methods collect and analyze historical data during the port operation process, try to discover operation rules, and formulate scheduling strategies accordingly. However, most of these methods focus on the static analysis of historical data and ignore the importance of real-time operation data. During the port operation process, information such as equipment operation status, transportation task queue, and environmental monitoring changes in real time, and these changes have an important impact on the formulation and execution of scheduling strategies. Therefore, how to effectively utilize real-time operation data to achieve dynamic optimization of port scheduling has become an urgent problem to be solved. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of this application, embodiments of this application provide a business data mining method applied to intelligent port scheduling. The method includes: Obtain real-time operation data generated during the loading and unloading operations of the target port. The real-time operation data includes equipment operation status information, transportation task queue information, and environmental monitoring information; Perform dynamic preprocessing on the real-time operation data to obtain effective operation data. The dynamic preprocessing includes screening and priority marking of conflicting tasks in the real-time operation data according to preset port operation rules; Perform multi-dimensional feature encoding on the effective operation data to generate an operation feature set related to port scheduling, where the operation feature set includes equipment operation features, task distribution features, and environmental constraint features; Based on the operation feature set, generate a port equipment scheduling strategy for the current period through a pre-trained scheduling decision model. The port equipment scheduling strategy includes equipment start-stop sequences, task allocation paths, and resource occupancy ratios, and the scheduling strategy satisfies the balance condition of port operation efficiency and resource consumption; Dynamically adjust port equipment according to the scheduling strategy to perform loading and unloading tasks, and update the real-time operation data based on the adjusted equipment operation status.

[0006] In another aspect, an embodiment of the present application further provides a smart port scheduling system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or codes, and the processor is used to run the programs, instructions, or codes in the machine-readable storage medium to implement the above method.

[0007] Based on the above aspects, the embodiment of the present application obtains real-time operation data covering equipment operation status information, transportation task queue information, and environmental monitoring information, and then screens and marks the priority of conflicting tasks according to preset port operation rules, improving the data quality. Then, it generates an operation feature set including equipment operation features, task distribution features, and environmental constraint features, deeply excavates the data value from multiple key dimensions, comprehensively depicts various factors related to port scheduling, and improves the scientificity and rationality of scheduling decisions. Based on the operation feature set, a port equipment scheduling strategy that satisfies the balance condition of port operation efficiency and resource consumption is generated by using a pre-trained scheduling decision model, realizing intelligent and precise scheduling decisions. The generated scheduling strategies such as equipment start-stop sequences, task allocation paths, and resource occupancy ratios can effectively coordinate the operation of port equipment, improve port operation efficiency while reasonably controlling resource consumption. Finally, the port equipment is dynamically adjusted according to the scheduling strategy and the real-time operation data is updated, enabling the port equipment to optimize its operation in real time according to the actual operation situation, continuously maintaining a high-efficiency and stable working state. At the same time, the update of the real-time operation data provides the latest and accurate data support for the next round of scheduling decisions, further enhancing the adaptive ability and overall effectiveness of port scheduling, and enabling the smart port scheduling to always maintain an efficient, intelligent, and sustainable operation mode in a complex and changeable actual operation environment. Description of the Drawings

[0008] Figure 1 It is a schematic execution flowchart of a business data mining method applied to smart port scheduling provided by an embodiment of the present application.

[0009] Figure 2 It is a schematic diagram of the hardware architecture of the intelligent port scheduling system provided by the embodiments of the present application. Specific embodiments

[0010] The present application will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of a business data mining method applied to intelligent port scheduling provided by an embodiment of the present application. The business data mining method applied to intelligent port scheduling will be introduced in detail below.

[0011] Step S110: Obtain real-time operation data generated during the loading and unloading operations of the target port. The real-time operation data includes equipment operation status information, transportation task queue information, and environmental monitoring information.

[0012] Taking a large coastal port as an example, during the loading and unloading operations, there are numerous devices operating in the port. The equipment operation status information covers the status of devices such as cranes, forklifts, and conveyor belts. For example, the operation status information of a crane includes the lifting weight, lifting height, extension angle of the boom, operating speed, and operating power of the motor of the crane. The equipment operation status information of a forklift includes its traveling speed, the weight of the goods being lifted, battery power (if it is an electric forklift), and tire pressure. The conveyor belt has information such as operating speed, the flow rate of goods on the conveyor belt, and the temperature of the conveyor belt motor.

[0013] The transportation task queue information reflects the task arrangements for cargo loading, unloading, handling, and transfer within the port. For example, there is a batch of containers from a large cargo ship that need to be unloaded. These containers are arranged with corresponding subsequent transportation tasks according to different destinations and cargo types. Some containers are to be transported to nearby warehouses by trucks, and some are to be transported to inland cities by rail. Each task has a clear starting point (such as a certain cabin of the cargo ship), destination (such as a specific warehouse number or railway freight yard), cargo type (such as electronic products, food, etc.), and estimated start time and completion time, etc.

[0014] Regarding the environmental monitoring information, environmental monitoring devices continuously monitor environmental factors such as wind speed, wind direction, visibility, humidity, and temperature. For example, when the wind speed is too high, the lifting operation of the crane may be restricted because excessive wind speed may cause the lifted goods to sway, increasing safety risks. Low visibility may affect the traveling speed and safety of forklifts within the port, and humidity also requires special attention for the loading, unloading, and storage of certain humidity-sensitive goods (such as electronic products). By obtaining the above real-time operation data, the port's scheduling center can comprehensively understand the current status of port operations.

[0015] Step S120: Dynamically preprocess the real-time job data to obtain valid job data. The dynamic preprocessing includes screening and priority marking of conflicting tasks in the real-time job data according to preset port operation rules.

[0016] Continuing with the above coastal port as an example, in the actual loading and unloading operation process, task conflicts often occur. Separate the subset of equipment status data corresponding to the equipment operation status information and the subset of task data corresponding to the transportation task queue information from the real-time job data. For example, in the subset of equipment status data, it is found that the motor operating power of crane A is close to its rated power upper limit, which is an operating status that needs attention. At the same time, in the subset of task data, there are multiple transportation tasks that all need to use crane A for loading and unloading operations within the same time period.

[0017] Perform running noise filtering on the subset of equipment status data. Assume that the instantaneous power of crane A occasionally exceeds the preset threshold, which may be caused by momentary load fluctuations or measurement errors. Thus, perform piecewise interpolation processing on the equipment status data containing such abnormal operation parameters to fill in the data missing. For example, within the time period when the power suddenly rises and then quickly returns to normal, calculate a reasonable power value through interpolation, and then merge the interpolated equipment status data with the equipment status data that does not contain abnormal operation parameters into the initial denoised equipment status data. Then, use the sliding window mean algorithm to perform local average calculation on the equipment operation parameters of the initial denoised equipment status data. According to the type of crane, set the length of the sliding window to an appropriate value, such as 5 minutes, calculate the average of the power data within these 5 minutes, and finally perform normalization processing on the equipment operation parameters after the local average calculation to make the equipment operation parameters fall within the preset numerical interval.

[0018] Perform task conflict detection on the subset of task data and analyze the task attributes in the subset of task data. Suppose task 1 is to unload a batch of valuable electronic products from a cargo ship to a temporary storage area, and task 2 is to unload a batch of large mechanical equipment from the same area of the cargo ship to another location, and both tasks need to use the same large crane, and task 1 requires completion within a specific time to avoid the electronic products being exposed to high temperature for a long time. Construct a task dependency graph according to the task attributes and find that there is a conflict between these two tasks in terms of crane resources. Traverse all paths in the task dependency graph, detect this resource overrun conflict, and sort the detected conflicting tasks by priority.

[0019] Re - prioritize the transportation task queue information according to the denoising device status data and the conflict task list, in combination with the preset port operation rules. Extract the current load rate of each port device from the denoising device status data. For example, crane A has been operating continuously for a long time and has a high load rate. Extract the attribute parameters of each conflict task from the conflict task list. Since the goods of task 1 are valuable electronic products and there is a time limit, and the difference between its task deadline and the current system time is small, the task urgency score is relatively high. According to the environmental safety threshold defined in the port operation rules, match the environmental monitoring information corresponding to the task execution period in the conflict task list. Assume that the current humidity is high, which has a certain impact on the storage of electronic products in task 1, and generate environmental safety constraint parameters. Based on the device load rate index, task urgency score, and environmental safety constraint parameters, construct the comprehensive priority score for each conflict task. For example, the inverse proportional value of the device load rate index of task 1 is relatively high (because high - load tasks should not be arranged when the crane load rate is high), the direct proportional value of the task urgency score is relatively high, and the direct proportional value of the environmental safety constraint parameter is also relatively high, so the comprehensive priority score is relatively high. Sort the tasks in the conflict task list in descending order according to the comprehensive priority score to generate a conflict task sub - queue with priority sorting. Traverse the conflict task sub - queue with priority sorting and insert each conflict task into the non - conflict task gap in the transportation task queue information in turn. Assume that task 1 is inserted into the idle period after a crane has completed other non - critical tasks. Based on the inserted transportation task queue information, detect whether there is an over - limit of device resources or a violation of environmental safety constraints. If so, perform a secondary screening of the conflict tasks according to the comprehensive priority score, and eliminate the conflict tasks with over - limit of device resources or violation of environmental safety constraints. Finally, align the time stamps of the transportation task queue information after insertion and screening with the denoising device status data to generate valid operation data containing the complete task sequence and device status mapping relationship.

[0020] Step S130: Perform multi - dimensional feature encoding on the valid operation data to generate an operation feature set related to port scheduling. The operation feature set includes device operation features, task distribution features, and environmental constraint features.

[0021] In the above coastal port scenario, time series features associated with equipment operating status information and spatial distribution features associated with transportation task queue information can be extracted from the effective operation data. Taking the crane as an example, the time series features of its equipment operating status information may include the changes in parameters such as the lifting weight, lifting height, and operating speed over time in the past few hours. Then, this time series feature is input into a temporal convolutional network for multi-scale feature extraction. The temporal convolutional network contains multiple layers of causal convolutional modules. The time series feature is intercepted by a sliding window through multiple layers of causal convolutional modules. Assuming the window length is 10 minutes, multiple local time segments are obtained. Different convolutional kernel weights are assigned to each local time segment. For example, different convolutional kernel weights are assigned according to different operating modes of the crane (such as light-load lifting and heavy-load lifting). Then, based on the convolutional kernel weight assignment results, the features of adjacent local time segments are superimposed to obtain the global time dependence relationship. According to the global time dependence relationship, a device dynamic feature vector is generated. This device dynamic feature vector can reflect the operating modes of the crane at different time periods, such as different operating characteristics during peak operation periods and off-peak operation periods.

[0022] The spatial distribution features associated with transportation task queue information are input into a graph attention network for node relationship modeling. For example, in a transportation task, multiple goods are transported from different cargo ship compartments to different warehouses within the port. These tasks are regarded as nodes in the graph attention network, and the edges represent the spatial proximity relationships between tasks (such as the transportation tasks of goods in adjacent cargo ship compartments may have a cooperative relationship). Attention coefficients are calculated for each task node, comprehensively considering the geographical location of the task node (such as the task of the cargo ship compartment near the dock entrance may have a higher priority), the overlap degree of equipment requirements (such as tasks that all require forklift handling), and the overlap degree of time windows (such as tasks for loading and unloading within the same time period). According to the attention coefficients, the features of adjacent task nodes are weighted and aggregated to generate a task topology feature vector. This task topology feature vector contains the cooperative operation information of the task group, such as which tasks can be carried out simultaneously to improve efficiency.

[0023] The environmental constraint features are independently encoded, considering the environmental monitoring information of the port, such as wind speed, humidity, and visibility. Assuming that too high a wind speed will have a greater impact on the lifting operation of the crane, high humidity may affect the storage of goods, and low visibility will affect the vehicle driving speed. According to the degree of influence of these factors on operation safety, environmental constraint features are generated. For example, a higher influence weight is assigned to the wind speed because it is directly related to the safety of the lifting operation.

[0024] Finally, the device dynamic feature vector, task topology feature vector, and environmental constraint features can be cross-modally fused, and the weights of different features are dynamically allocated through an attention mechanism. For example, when deciding whether to start a certain crane for operation, the operating characteristics of the crane (device dynamic feature vector), the coordination relationship of related tasks (task topology feature vector), and the current environmental conditions (environmental constraint features) will be comprehensively considered. If the current wind speed is relatively high, even if the crane is idle and there is a task demand, the start may be delayed according to the weight allocation to ensure the safety of the operation.

[0025] Step S140, based on the job feature set, generate the port equipment scheduling strategy for the current period through a pre-trained scheduling decision model. The port equipment scheduling strategy includes the equipment start-stop sequence, task allocation path, and resource occupancy ratio, and the scheduling strategy satisfies the balance condition of port operation efficiency and resource consumption.

[0026] Still taking this coastal port as an example, based on the job feature set generated previously, the pre-trained scheduling decision model starts to generate the scheduling strategy. Suppose there are 5 cranes, 10 forklifts, 3 conveyor belts and other equipment in the port, as well as multiple different types of transportation tasks. The scheduling decision model first determines the equipment start-stop sequence. According to the equipment operation characteristics and task distribution characteristics, if there is a batch of large goods to be unloaded currently, and one of the large cranes C has a low load rate and is idle in the previous operation, the scheduling decision model will rank crane C at the front of the equipment start-stop sequence and give priority to starting it to perform this unloading task. At the same time, considering the environmental constraint features, if the current wind speed is relatively high but within the safe operation range of crane C, a start decision will also be made.

[0027] For the task allocation path, according to the task topology feature vector, the task of unloading goods from the cargo ship is allocated to the most suitable equipment and subsequent transportation path. For example, for small pieces of goods that need to be quickly transported to a nearby warehouse, the scheduling decision model can allocate them to a forklift and specify the shortest path from the dock directly to the warehouse. For goods that need to be transported to inland cities and are in large quantities, it will be arranged to be first transported to the railway freight yard through the conveyor belt and then transported by rail.

[0028] In determining the resource occupancy ratio, resources are reasonably allocated according to the load capacity of the equipment and the task requirements. For example, for the lifting task of large goods, sufficient time and energy resources are allocated to the crane to ensure that the task can be completed safely and efficiently, while avoiding excessive resource occupancy that causes other tasks to wait. This scheduling strategy takes into account the balance condition of resource consumption (such as equipment energy consumption, equipment wear, etc.) while meeting the port operation efficiency (such as completing cargo handling and transfer tasks as soon as possible). For example, not all equipment will be run at full load simultaneously in the pursuit of speed. Instead, the working intensity of the equipment is reasonably arranged according to the urgency of the task and the actual situation of the equipment to extend the service life of the equipment and reduce energy consumption.

[0029] Step S150, dynamically adjust the port equipment according to the scheduling strategy to perform the loading and unloading tasks, and update the real-time operation data based on the adjusted equipment operation status.

[0030] According to the above scheduling strategy, in this coastal port, analyze the equipment start-stop sequence in the scheduling strategy to generate an equipment control instruction set. For example, for the crane C to be started, the equipment control instruction set includes the equipment number of the crane C, the operation type (start), and the execution timestamp (execute immediately). Send this equipment control instruction set to the corresponding port equipment controller and monitor the equipment execution status in real time. Suppose during the execution process, it is found that the instruction response time of the crane C is longer than expected, the task execution progress is also slower than planned, and the equipment does not send a fault signal. At this time, collect the difference parameters between the current equipment execution status and the expected status. The uncompleted task list is the information of the goods that have not been unloaded yet. The equipment delay time is obtained by comparing the actual execution time and the planned execution time. The remaining resources (such as the remaining lifting capacity of the crane C, the remaining energy consumption, etc.) are also counted.

[0031] Then, perform feature transformation on the difference parameters to generate difference feature vectors, concatenate the difference feature vectors with the original job feature set to obtain an enhanced job feature set. For example, the device delay time information contained in the difference feature vectors will be combined with the time series features in the original job feature set. Re-encode the enhanced job feature set through the multi-layer neural network of the scheduling decision model to generate an updated feature set adapted to real-time deviations. Input the enhanced job feature set into the embedding layer of the multi-layer neural network to align the dimensions of the device dynamic feature vectors, task topology feature vectors, and environmental constraint features therein, and generate intermediate feature vectors of unified dimensions. In the attention calculation layer, based on the interaction relationship between the device dynamic feature vectors and task topology feature vectors in the intermediate feature vectors, calculate the attention weight matrix between features. This attention weight matrix can reflect the dynamic association strength between device operation features and task distribution features. For example, if device delays cause task distribution adjustments, this attention weight matrix will reflect the new association relationship between the device and the task. Aggregate the intermediate feature vectors by weighting with the attention weight matrix to generate a weighted global feature vector, which contains the cross-modal dependence relationship between device operation states and task distributions. Then, extract the temporal change patterns of the device operation parameters in the weighted global feature vector in the time series analysis layer to generate a temporally enhanced feature vector, which reflects the dynamic trend of device states with task adjustments. In the spatial relationship encoding layer, combine the environmental constraint features to correct and constrain the spatial dependence paths of task distributions, and generate spatial constraint feature vectors, which contain the task collaborative execution paths under environmental safety condition restrictions. Concatenate the temporally enhanced feature vector and the spatial constraint feature vector in the channel dimension to generate a fused feature vector, and input it into the normalization layer for feature scaling to generate a normalized fused feature vector. Then input it into the fully connected layer, perform high-order feature mapping through a non-linear activation function to generate high-order feature vectors. Input the high-order feature vectors into the residual connection layer, perform skip connection with the intermediate feature vectors and execute feature superposition to generate residual enhanced feature vectors. Finally, input the residual enhanced feature vectors into the output layer, and map them to the feature space with the same dimension as the original job feature set through linear transformation to generate an updated feature set adapted to real-time deviations.

[0032] Generate a corrected device control instruction set based on the updated feature set, such as adjusting the lifting speed of crane C or allocating more time for it to complete tasks, and send the corrected device control instruction set to the port device controller to overwrite the original device control instruction set.

[0033] Meanwhile, collect the adjusted device operation status data, such as the actual lifting time, lifting weight, energy consumption, etc. of crane C. Compare the device operation status data with the original real-time operation data to generate a data difference report. For example, it is found that the lifting weight is a little less than the original plan and the lifting time is extended. Incrementally update the device operation status information and transportation task queue information in the real-time operation data according to the data difference report, update the new lifting weight and lifting time to the device operation status information, and adjust the subsequent task time in the transportation task queue information according to the extension of the lifting time, so that the subsequent scheduling decisions are generated based on the latest operation status.

[0034] Based on the above steps, in the embodiment of the present application, by obtaining real-time operation data covering device operation status information, transportation task queue information, and environmental monitoring information, then screening and prioritizing conflicting tasks according to the preset port operation rules, the data quality is improved. Then, a job feature set including device operation characteristics, task distribution characteristics, and environmental constraint characteristics is generated, deeply mining the data value from multiple key dimensions, comprehensively depicting various factors related to port scheduling, and enhancing the scientificity and rationality of scheduling decisions. Based on the job feature set, a port device scheduling strategy that meets the balance condition of port operation efficiency and resource consumption is generated by using a pre-trained scheduling decision model, realizing intelligent and precise scheduling decisions. The generated scheduling strategies such as device start-stop sequences, task allocation paths, and resource occupancy ratios can effectively coordinate the operation of port devices, while improving port operation efficiency, reasonably controlling resource consumption. Finally, dynamically adjust the port devices according to the scheduling strategy and update the real-time operation data, enabling the port devices to optimize their operation in real time according to the actual operation situation, continuously maintaining a high-efficiency and stable working state. At the same time, the update of the real-time operation data provides the latest and accurate data support for the next round of scheduling decisions, further enhancing the adaptive ability and overall effectiveness of port scheduling, and enabling the intelligent port scheduling to always maintain an efficient, intelligent, and sustainable operation mode in the complex and changeable actual operation environment.

[0035] In a possible implementation manner, step S120 includes: Step S121, separate the device status data subset corresponding to the device operation status information and the task data subset corresponding to the transportation task queue information from the real-time operation data.

[0036] Among the above-mentioned coastal ports, the equipment operation status information contains relevant data of numerous devices, such as various operation parameters of devices like cranes, forklifts, conveyor belts, etc., which together constitute a subset of equipment status data. For example, in the subset of equipment status data of cranes, there are data such as lifting weight, lifting height, boom extension angle, operating speed, motor power, etc.; in the subset of equipment status data of forklifts, there are data such as traveling speed, weight of lifted goods, battery power, tire pressure, etc.; in the subset of equipment status data of conveyor belts, there are data such as operating speed, conveying cargo flow, motor temperature, etc. And the task data subset corresponding to the transportation task queue information contains detailed arrangements for cargo loading, handling, and transfer tasks within the port, such as the source of the cargo (which cabin of which cargo ship), destination (which warehouse or railway freight yard), cargo type (electronic products, mechanical products, etc.), estimated start time, and estimated completion time, etc.

[0037] Step S122, perform running noise filtering on the equipment status data subset to obtain denoised equipment status data, and perform task conflict detection on the task data subset to obtain a conflict task list.

[0038] Step S123, according to the denoised equipment status data and the conflict task list, combined with the preset port operation rules, re - sort the priority of the transportation task queue information to generate the effective operation data, and the priority re - sorting is comprehensively determined based on the equipment load rate, task urgency, and environmental safety conditions.

[0039] In a possible implementation manner, step S122 includes: Step S1221, identify abnormal operation parameters in the equipment status data subset, where the abnormal operation parameters include operation parameters whose equipment instantaneous power exceeds a preset threshold or whose equipment continuous idle duration exceeds a preset range.

[0040] Step S1222, perform piecewise interpolation processing on the equipment status data containing the abnormal operation parameters to fill in the missing data, and merge the interpolated equipment status data with the equipment status data not containing abnormal operation parameters into initial denoised equipment status data.

[0041] Step S1223, use the sliding window mean algorithm to perform local average calculation on the equipment operation parameters of the initial denoised equipment status data, and the length of the sliding window is dynamically adjusted according to the equipment type.

[0042] Step S1224, perform normalization processing on the equipment operation parameters after local average calculation, so that the equipment operation parameters fall within a preset numerical interval for joint encoding with the characteristics of the task data subset to obtain the denoised equipment status data.

[0043] Taking the crane at the port as an example, during the operation of the crane, the instantaneous power of its motor may exceed the preset threshold, which may be caused by suddenly lifting overweight goods or a short-term fluctuation of the motor itself. In addition, the continuous idle time of the equipment exceeding the preset range also belongs to abnormal operation parameters. For example, during the normal operation of a forklift, if the continuous idle time exceeds the range set according to the previous operation rules, it may indicate that there is a fault with the forklift or the task arrangement is unreasonable. For the equipment status data containing abnormal operation parameters, piecewise interpolation processing is performed to fill in the missing data. For example, when the motor power of the crane suddenly increases and then quickly returns to normal at a certain moment, within the time period of this abnormal power fluctuation, a reasonable power value is calculated through piecewise interpolation processing, and then the interpolated equipment status data is merged with the equipment status data that does not contain abnormal operation parameters to obtain the initial denoised equipment status data. Then, the sliding window mean algorithm is used to perform local average calculation on the equipment operation parameters of the initial denoised equipment status data, and the length of the sliding window is dynamically adjusted according to the equipment type. For large equipment such as cranes, since the change of its operation parameters is relatively slow, the length of the sliding window may be set to a relatively long time, such as 10 minutes; while for relatively flexible equipment such as forklifts with faster-changing operation parameters, the length of the sliding window may be set to 3 minutes. Through this local average calculation, some small fluctuations can be smoothed out. Finally, the equipment operation parameters after the local average calculation are normalized so that the equipment operation parameters fall within the preset numerical range for joint encoding with the features of the task data subset, thereby obtaining the denoised equipment status data. In this way, parameters such as the lifting weight and lifting height of the crane, and parameters such as the driving speed and the weight of the goods picked up by the forklift are processed into appropriate numerical ranges, which is convenient for subsequent comprehensive analysis.

[0044] Step S1225, parse the task attributes in the task data subset, where the task attributes include the task start time, task end time, required equipment type, and task dependencies.

[0045] Step S1226, construct a task dependency graph according to the task attributes, where the nodes in the task dependency graph represent independent tasks, and the edges represent the dependencies between tasks.

[0046] Step S1227, traverse all paths in the task dependency graph, detect whether there are path loops or resource overrun conflicts, and after prioritizing the detected conflicting tasks, obtain the list of conflicting tasks.

[0047] Specifically, in the cargo handling and transportation tasks at coastal ports, each task has clear attributes such as the task start time, task end time, required equipment type, and task dependencies. For example, there is a task to unload a batch of electronic products from a certain cargo ship to a temporary storage area. The task start time is set within 1 hour after the cargo ship docks, the task end time is within 3 hours after docking, the required equipment types are cranes and forklifts, and this task depends on the docking status of the cargo ship and the preparation work of the dock workers. Based on these task attributes, a task dependency graph is constructed. The nodes in the task dependency graph represent independent tasks, and the edges represent the dependencies between tasks. For example, the task of unloading electronic products from the cargo ship to the temporary storage area is a node, and the task of transferring them from the temporary storage area to the warehouse is another node. If the previous task is not completed, the latter task cannot start, which constitutes the dependency between tasks and is represented by an edge in the task dependency graph. Traverse all paths in the task dependency graph to detect whether there are path loops or resource overlimit conflicts. For example, assume there are three tasks. Task A needs to use a certain crane from 9 am to 10 am, Task B also needs to use this crane from 9:30 am to 10:30 am, and Task C needs to use the same crane from 10 am to 11 am. This constitutes a resource overlimit conflict because the same crane is assigned multiple tasks within the same time period. After prioritizing the detected conflicting tasks, a list of conflicting tasks is obtained. For the above conflicting tasks, prioritize them according to factors such as the urgency of the tasks and the types of goods. If Task A is transporting time-sensitive electronic products, Task B is transporting ordinary goods, and Task C is transporting large mechanical equipment, since electronic products have high time requirements and are easily affected by the environment, the priority of Task A may be the highest, followed by Task C, and finally Task B, thus obtaining a list of conflicting tasks.

[0048] Further, step S123 includes: Step S1231, extract the current load rate of each port device from the denoised device status data to generate a device load rate index, where the device load rate index is the ratio of the average operation duration to the idle duration of the port device within a preset time window.

[0049] In this embodiment, in the scenario of the above-mentioned coastal port, considering a crane device, a preset time window is set, for example, the past 3 hours. Within these 3 hours, if the operating duration of the crane is 2 hours and the idle duration is 1 hour, then according to the calculation method of the equipment load rate index, that is, the ratio of the average operating duration to the idle duration, the equipment load rate index of this crane is 2 / 1 = 2. For a forklift device, similarly within the past 3 hours, if the operating duration is 1.5 hours and the idle duration is 1.5 hours, its equipment load rate index is 1.5 / 1.5 = 1. These equipment load rate indexes reflect the current busy degree of the equipment and are of great significance for subsequent task arrangements.

[0050] Step S1232, extract the attribute parameters of each conflict task from the conflict task list, where the attribute parameters include the task deadline, the type of goods, and the priority mark, and generate a task urgency score based on the difference between the task deadline and the current system time.

[0051] For example, there are three tasks in the conflict task list. Task A is to unload a batch of electronic products from a cargo ship to a temporary storage area, and its task deadline is 1 hour after the current time; Task B is to transfer a batch of mechanical equipment from another cabin of the cargo ship to a warehouse, and its task deadline is 3 hours after the current time; Task C is to transport a batch of food from the dock to the warehouse, and its task deadline is 2 hours after the current time. The current system time is 10 am. Then the difference between the task deadline of Task A and the current system time is 1 hour, the difference of Task B is 3 hours, and the difference of Task C is 2 hours. Since the task urgency score is generated based on this difference, the smaller the difference, the higher the task urgency score. Therefore, the task urgency score of Task A is the highest, Task C is the second, and Task B is the lowest. At the same time, the type of goods will also affect the task urgency score. Electronic products may need to be processed as soon as possible because they are more sensitive to the environment, such as humidity, temperature, etc. Therefore, the task urgency score of Task A will be further improved after considering the type of goods.

[0052] Step S1233, according to the environmental safety threshold defined in the port operation rules, match the environmental monitoring information corresponding to the task execution period in the conflict task list, and generate an environmental safety constraint parameter, where the environmental safety constraint parameter includes a combined quantization value of the wind speed level, the visibility level, and the humidity level.

[0053] In this coastal port, environmental monitoring information includes wind speed level, visibility level, humidity level, etc. Suppose the port operation rules stipulate that when the wind speed exceeds a certain level, there is a safety risk for using a crane for lifting operations; when the visibility is lower than a certain level, the driving speed of a forklift in the port needs to be reduced; when the humidity is higher than a certain value, special handling is required for the loading, unloading, and storage of certain goods (such as electronic products). For task A, the environmental monitoring information corresponding to its execution period shows that the wind speed level is close to the upper limit of safe crane operation, the humidity has a certain impact on the storage of electronic products, and the visibility is normal. Then, according to the port operation rules, these environmental factors are combined in a certain quantitative manner to generate the environmental safety constraint parameters for task A. For example, it is set that the quantitative weight of the wind speed level is relatively high, the humidity level is the second, and the visibility level is the lowest, and the environmental safety constraint parameters for task A are obtained through corresponding calculations.

[0054] Step S1234, based on the equipment load rate index, the task urgency score, and the environmental safety constraint parameters, construct the comprehensive priority score for each conflicting task. The comprehensive priority score superimposes the inverse proportional value of the equipment load rate index, the proportional value of the task urgency score, and the proportional value of the environmental safety constraint parameters through linear weighting.

[0055] For task A, the equipment load rate index is 2 (assuming a high crane load rate), and its inverse proportional value is relatively high (because high-load tasks should not be arranged when the equipment load rate is high), the task urgency score is relatively high (because the deadline is tight and the goods are electronic products), and the environmental safety constraint parameters are relatively high (because the wind speed and humidity affect the task). Through linear weighting, the inverse proportional value of the equipment load rate index, the proportional value of the task urgency score, and the proportional value of the environmental safety constraint parameters are superimposed to obtain the comprehensive priority score for task A. For example, let the weight of the inverse proportional value of the equipment load rate index be 0.3, the weight of the task urgency score be 0.4, and the weight of the environmental safety constraint parameters be 0.3, and calculate that the comprehensive priority score for task A is a specific value. The comprehensive priority scores for task B and task C are calculated in the same way.

[0056] Step S1235, according to the comprehensive priority scores, sort the tasks in the conflicting task list in descending order to generate a sub-queue of conflicting tasks with sorted priorities.

[0057] For example, according to the calculated comprehensive priority scores, the score of task A is the highest, the score of task C is the second, and the score of task B is the lowest. Therefore, the tasks are sorted in descending order in this order to form a sub-queue of conflicting tasks with sorted priorities.

[0058] Step S1236: Traverse the conflict task sub-queue after priority sorting, and insert each conflict task into the non-conflict task gap in the transportation task queue information in sequence. The non-conflict task gap is the device idle period between adjacent tasks in the transportation task queue and satisfies the condition that there is no cycle in the task dependency relationship.

[0059] The transportation task queue information contains multiple tasks, such as tasks of unloading goods from different holds of a cargo ship and transporting them to different warehouses. There are device idle periods between adjacent tasks, and these periods meet the requirement of no cycle in the task dependency relationship. Take task A from the conflict task sub-queue after priority sorting, search for the non-conflict task gap in the transportation task queue information. If a gap is found where the devices (such as cranes and forklifts) are idle during this gap and inserting task A will not cause a cycle in the task dependency relationship, then insert task A into this gap. Process conflict tasks such as task C and task B in the same way in sequence.

[0060] Step S1237: Based on the transportation task queue information after insertion, detect whether there is any overrun of device resources or violation of environmental safety constraints. If there is an overrun of device resources or violation of environmental safety constraints, perform a secondary screening of the conflict tasks according to the comprehensive priority score, and eliminate the conflict tasks with overrun of device resources or violation of environmental safety constraints.

[0061] For example, after inserting task A, check whether the task volume of the involved device (such as a crane) during this period exceeds its load capacity. If it does, there is an overrun of device resources. At the same time, check whether the environmental conditions during the execution period of task A violate the environmental safety constraints. For example, if the electronic products in task A do not take corresponding protective measures in a high-humidity environment, there is a violation of environmental safety constraints. If there is an overrun of device resources or violation of environmental safety constraints, perform a secondary screening of the conflict tasks according to the comprehensive priority score, and eliminate the conflict tasks with overrun of device resources or violation of environmental safety constraints. Suppose task A has an overrun of device resources. Since the comprehensive priority score of task C is second only to that of task A, and task C has no overrun of device resources and violation of environmental safety constraints, then task A will be eliminated and the insertion position of task C in the transportation task queue will be retained.

[0062] Step S1238: Align the timestamp of the transportation task queue information after insertion and screening with the denoised device status data to generate the effective operation data containing the complete task sequence and the device status mapping relationship.

[0063] In the transportation task queue information, each task has corresponding time information such as start time and end time. The denoised device status data also contains the status information of the device at different times. The two are accurately aligned through timestamps. For example, if the start time of task A is 11 am and the end time is 12 pm, then in the denoised device status data, search for the status information of devices such as cranes and forklifts between 11 am and 12 pm, and establish a mapping relationship between task A and the status of these devices. For other tasks in the transportation task queue, the same timestamp alignment operation is performed, thereby generating valid operation data containing the complete task sequence and the device status mapping relationship, providing an accurate and comprehensive data basis for subsequent port operation scheduling.

[0064] In a possible implementation manner, step S130 includes: Step S131, extract the time series features associated with the device operation status information and the spatial distribution features associated with the transportation task queue information from the valid operation data.

[0065] Still taking the crane device as an example, the time series features associated with the device operation status information include the changes in parameters such as the lifting weight, lifting height, running speed, and motor power of the crane over a period of time. For example, the lifting weight of the crane is recorded every 10 minutes in the past 5 hours, and the lifting weight data at these different time points constitute a time series feature. Similarly, for the forklift device, the changes in parameters such as its driving speed and the weight of the goods picked up over time also belong to the time series features. The spatial distribution features associated with the transportation task queue information reflect the distribution of tasks within the port space. For example, there are multiple transportation tasks, including unloading goods from different cargo ship compartments to different areas of the dock and then transferring them to different warehouses within the port. The spatial information such as the starting points (cargo ship compartments and dock areas) and ending points (warehouse locations) of these tasks, as well as the spatial relationships between the tasks, constitute the spatial distribution features.

[0066] Step S132, input the time series features into a temporal convolutional network for multi-scale feature extraction to obtain a device dynamic feature vector, and the device dynamic feature vector serves as the device operation feature, reflecting the operation mode of the device at different time periods.

[0067] In this embodiment, for the lifting weight time series characteristics of the crane, assuming that the length of the sliding window is 30 minutes, then the entire 5-hour time series is intercepted with a window of 30 minutes to obtain multiple local time segments. Convolution kernel weights are assigned to each local time segment respectively, and the convolution kernel weights are assigned according to the different operating states and task requirements of the crane. For example, when the crane is lifting heavier goods, the corresponding local time segment may be assigned a larger convolution kernel weight, because in this case the change in lifting weight has a greater impact on the equipment operation mode. Based on the convolution kernel weight assignment result, the features of adjacent local time segments are superimposed to obtain a global time dependency. For example, the lifting weight features in several adjacent 30-minute windows are superimposed according to the convolution kernel weight assignment result, so that the global time dependency of the crane's lifting weight within the entire 5 hours can be obtained, which reflects the overall change trend of the lifting weight over time and the correlation between different time periods. According to this global time dependency, a device dynamic feature vector is generated, and the dimension of the device dynamic feature vector is proportional to the number of port equipment and the number of task types. Since there are various types of equipment (such as cranes, forklifts, conveyor belts, etc.) and various types of tasks (such as cargo unloading, transshipment, storage, etc.) in the port, the dynamic feature vector of the equipment can comprehensively reflect the operating modes of different equipment under different tasks, such as the busyness of different equipment in different time periods, the impact of different tasks on the equipment operating parameters, etc.

[0068] Step S133, input the spatial distribution feature into the graph attention network to perform node relationship modeling to obtain a task topology feature vector, and the task topology feature vector serves as the task distribution feature to reflect the spatial dependency relationship between tasks.

[0069] For example, the task of unloading a batch of electronic products from the No. 1 cargo hold of cargo ship A to Area 1 of the dock and then transferring them to Warehouse 1 is a node. The edges represent the spatial proximity relationship between tasks. For example, if the starting or ending points of two tasks are close in space, there is a spatial proximity relationship. Calculate the attention coefficient for each task node, which is comprehensively determined based on the geographical location of the task node, the overlap degree of equipment requirements, and the overlap degree of time windows. Taking the geographical location as an example, if the starting point of a task is close to the dock entrance, then this task node may be assigned a higher attention coefficient because it has an advantage in geographical location and may be more likely to be prioritized. For the overlap degree of equipment requirements, if multiple tasks all need to use the same crane, then the overlap degree of equipment requirements between these task nodes is high, and this factor will be considered when calculating the attention coefficient. In terms of the overlap degree of time windows, if two tasks are carried out for loading, unloading, or transferring within the same time period, their overlap degree of time windows is high. Aggregate the features of adjacent task nodes according to the attention coefficient to generate a task topology feature vector, which contains the collaborative operation information of the task group. For example, if several tasks are spatially adjacent, have a high overlap degree of equipment requirements, and also have a high overlap degree of time windows, by aggregating the features of these task nodes through weighting, the task topology feature vector can reflect that these tasks can be collaboratively operated. For example, the same equipment can be arranged to complete these tasks in sequence to improve the operation efficiency.

[0070] Step S134, independently encode the environmental constraint features to generate the environmental constraint features, where the environmental constraint features include the influence weights of wind speed, humidity, and visibility on operation safety.

[0071] For example, in coastal ports, the environmental constraint features include the influence weights of wind speed, humidity, and visibility on operation safety. Wind speed has an important impact on the hoisting operation safety of cranes. When the wind speed is relatively high, the hoisting weight of the crane needs to be restricted, and the hoisting height may also need to be reduced, so a relatively high influence weight is assigned to wind speed. Humidity affects the loading, unloading, and storage of humidity-sensitive goods such as electronic products. For example, moisture-proof measures may need to be taken in a high-humidity environment, so a certain influence weight is also assigned to humidity. When the visibility is low, the driving speed of forklifts in the port needs to be reduced, which affects the cargo transportation tasks in the port, so visibility is also considered in the environmental constraint features and an appropriate weight is assigned. These influence weights are determined according to the actual situation and safety requirements of port operations. For example, the specific influence weights of wind speed, humidity, and visibility are determined through the analysis of historical data, expert experience, and the safety standards of port operations, so as to generate the environmental constraint features.

[0072] Step S135: Perform cross-modal fusion on the device dynamic feature vector, the task topology feature vector, and the environmental constraint features to generate the job feature set. The cross-modal fusion dynamically assigns weights to different features through an attention mechanism.

[0073] For example, when deciding whether to start a certain crane for operation, the attention mechanism comprehensively considers the current operating state of the crane in the device dynamic feature vector (such as load rate, operating speed, etc.), the task collaboration information related to this crane in the task topology feature vector (such as whether there are other tasks that can collaborate with the current task), and environmental factors such as wind speed in the environmental constraint features. If the current wind speed is relatively high, although the load rate of the crane is low and there is a task requirement, according to the weights dynamically assigned by the attention mechanism, the possibility of starting this crane may be reduced to ensure operation safety. Through this cross-modal fusion, the job feature set can comprehensively consider the device operation features, task distribution features, and environmental constraint features, providing a comprehensive decision-making basis for port equipment scheduling.

[0074] In a possible implementation manner, the temporal convolutional network includes multiple layers of causal convolutional modules. Step S132 includes: Step S1321: Perform sliding window interception on the time series features through the multiple layers of causal convolutional modules to obtain multiple local time segments.

[0075] Step S1322: Perform convolutional kernel weight assignment on each local time segment respectively, and superimpose the features of adjacent local time segments based on the convolutional kernel weight assignment result to obtain the global time dependence.

[0076] Step S1323: Generate the device dynamic feature vector according to the global time dependence. The dimension of the device dynamic feature vector is proportional to the number of port devices and the number of task types.

[0077] In the above coastal port scenario, the process of using a multi-layer causal convolution module to intercept time series features through a sliding window is to capture the local information of time series features at different time scales. Taking the time series feature of the lifting weight of a crane as an example, the first-layer causal convolution in the multi-layer causal convolution module may use a relatively small sliding window length, such as 10 minutes, so as to capture the rapid changes in the lifting weight in a short period of time and obtain some local time segments. For each such local time segment, the convolution kernel weight is assigned according to factors such as its position in the entire time series and the change trend of the lifting weight. For example, a relatively large convolution kernel weight may be assigned to a local time segment where the lifting weight suddenly increases, because this change is crucial for reflecting the operating state of the crane. Then, based on the convolution kernel weight assignment result, the features of adjacent local time segments are superimposed, and the local time segments obtained by the first-layer causal convolution are superimposed according to the weights to obtain a preliminary global time dependence relationship.

[0078] Next, the second-layer causal convolution may use a relatively large sliding window length, such as 30 minutes, and intercept the preliminary global time dependence relationship obtained by the first-layer causal convolution through a sliding window again to obtain new local time segments. Similarly, the convolution kernel weights are assigned to these new local time segments. This time, the weight assignment may take into account the features obtained by the first-layer causal convolution and the change trend of the lifting weight on a larger time scale and other factors. Based on the new convolution kernel weight assignment result, the features of adjacent local time segments are superimposed to obtain a more comprehensive global time dependence relationship. This process may continue in the multi-layer causal convolution module, and each layer captures information at different time scales and superimposes features on the basis of the previous layer.

[0079] Finally, a device dynamic feature vector is generated according to the global time dependence relationship finally obtained by this multi-layer causal convolution module. Since there are a large number of port devices (such as multiple cranes, forklifts, etc.) and various task types (such as loading, unloading, and transferring of different goods), the dimension of the device dynamic feature vector needs to be large enough to reflect the operating modes of different devices under different tasks. For example, a certain dimension of the device dynamic feature vector may correspond to the operating state features of a specific crane under a specific task. In this way, the device dynamic feature vector can comprehensively reflect the operating modes of port devices at different time periods and provide detailed device operating feature information for subsequent port device scheduling.

[0080] Step S133 includes: Step S1331, inputting the spatial distribution feature into a graph attention network to construct a task node graph, where each node in the task node graph represents a transportation task, and the edge represents the spatial proximity relationship between tasks.

[0081] Step S1332: Calculate the attention coefficient for each task node, where the attention coefficient is comprehensively determined based on the geographical location of the task node, the overlap degree of equipment requirements, and the overlap degree of time windows.

[0082] In the coastal port scenario, after inputting the spatial distribution characteristics into the graph attention network to construct the task node graph, calculating the attention coefficient for each task node is a process that comprehensively considers multiple factors. Taking the transportation task of unloading goods from the cargo hold of a freighter to the dock area and then transferring them to the warehouse as an example, the influence of the geographical location factor of the task node on the attention coefficient is reflected in multiple aspects. If the starting cargo hold of a task is close to the loading and unloading equipment at the dock, then this task node has an advantage in terms of geographical location and may be assigned a relatively high attention coefficient because this can reduce the distance and time of goods transportation. For the overlap degree of equipment requirements, if multiple tasks all need to use the same special equipment (such as a large crane) to complete the loading and unloading tasks, then the overlap degree of equipment requirements between these task nodes is relatively high. For example, if three tasks all need to use a certain large crane, the overlap degree of equipment requirements between these three task nodes will be considered key when calculating the attention coefficient. In terms of the overlap degree of time windows, if the estimated start times and estimated completion times of several tasks overlap, then their overlap degree of time windows is relatively high. For example, if two tasks are both scheduled to perform loading and unloading operations between 9 am and 11 am, then the overlap degree of time windows between these two task nodes is relatively high.

[0083] Step S1333: Weightedly aggregate the features of adjacent task nodes according to the attention coefficient to generate the task topology feature vector, where the task topology feature vector contains the collaborative operation information of the task group.

[0084] Specifically, for adjacent task nodes, their features are weighted according to the magnitude of their attention coefficients. For example, the features of a task node with a relatively high attention coefficient will be given a larger weight during weighted aggregation, while the features of a task node with a relatively low attention coefficient will be given a smaller weight. Through this weighted aggregation method, a task topology feature vector is generated, and this task topology feature vector contains the collaborative operation information of the task group. For example, if there is a group of task nodes that are geographically close, have a high overlap degree of equipment requirements, and a high overlap degree of time windows, then the task topology feature vector obtained by weighted aggregating the features of these task nodes can reflect the information that these tasks can be collaboratively operated. For instance, the same piece of equipment can be arranged to complete these tasks in a certain order, or the start times and resource allocations of these tasks can be coordinated to improve the efficiency of the entire port operation.

[0085] In a possible implementation manner, the training process of the pre-trained scheduling decision model includes: Step S210, obtain a historical port operation dataset, where the historical port operation dataset includes operation data samples of multiple historical periods and corresponding scheduling policy labels.

[0086] In this coastal port, the historical operation data samples cover the port operation information over a long past period. For example, these operation data samples record the equipment operation status information at different dates and different time periods, such as the lifting weight, lifting height, running speed, motor power, etc. of the crane at each moment, the traveling speed, the weight of the goods lifted, the battery power, etc. of the forklift, and the running speed, the flow of the transported goods, the motor temperature, etc. of the conveyor belt. At the same time, the transportation task queue information is also included, such as the source (which cabin of which cargo ship) and destination (which warehouse or railway freight yard) of each cargo, the type of goods (electronic products, mechanical products, etc.), the estimated start time and the estimated completion time, etc. The corresponding scheduling policy labels clearly record the scheduling policies such as the equipment start-stop sequence, the task allocation path, and the resource occupancy ratio adopted by the port under the operation data at that time. For example, during a certain historical period, for the unloading task of a certain cargo ship, the scheduling policy label records which cranes to start, the working path of the forklift, and the resource allocation of each equipment, etc.

[0087] Step S220, perform sliding window segmentation on the historical port operation dataset to generate multiple training sample sequences, and a single training sample sequence covers the port operation data of a continuous preset time period.

[0088] Step S230, input the training sample sequence into the initial scheduling decision model for policy prediction to obtain a predicted scheduling policy.

[0089] The initial scheduling decision model is a model constructed based on technologies such as neural networks, and receives data such as the equipment operation status information and the transportation task queue information in the training sample sequence as input. In this coastal port scenario, the initial scheduling decision model can predict the predicted scheduling policies such as the equipment start-stop sequence, the task allocation path, and the resource occupancy ratio according to the input information. For example, for a batch of cargo handling tasks in a certain training sample sequence, the initial scheduling decision model may predict which cranes to start, allocate the goods to different transportation equipment in a specific order, and the resource occupancy ratio of each equipment during the task execution, etc.

[0090] Step S240, construct a loss function based on the difference between the predicted scheduling policy and the scheduling policy label, where the loss function is constructed by combining the equipment utilization rate, the task completion rate, and the energy consumption index.

[0091] Step S250: Update the parameters of the initial scheduling decision model through backpropagation based on the constructed loss function until convergence, obtaining a pre-trained scheduling decision model that can be generalized to real-time job data.

[0092] During the training process, each time a training sample sequence is input into the initial scheduling decision model to obtain a predicted scheduling strategy, according to the calculated loss function, the parameters of the model are adjusted through the backpropagation algorithm. For example, if the value of the loss function is large, it indicates that the difference between the predicted scheduling strategy and the actual scheduling strategy label is large. Then, through the backpropagation algorithm, the initial scheduling decision model will adjust its internal weights and other parameters in the direction of reducing this difference. This process will be repeated continuously. As the training progresses, the value of the loss function will gradually decrease. When the value of the loss function converges to a small value, it means that the model has learned enough knowledge and can better predict a reasonable scheduling strategy based on the input job data. At this time, the obtained model is a pre-trained scheduling decision model that can be generalized to real-time job data. This pre-trained scheduling decision model can be applied to actual port operation scheduling to predict equipment scheduling strategies based on real-time job data, so as to improve the efficiency of port operations and the rationality of resource utilization.

[0093] In a possible implementation manner, step S220 includes: Step S221: Set the length and step size of the sliding window according to the port operation cycle. The length is used to define the time span of a single training sample sequence, and the step size is used to define the overlapping ratio between adjacent training sample sequences.

[0094] Step S222: Continuously intercept the historical port operation data set according to the length and step size, and retain the scheduling strategy label corresponding to each training sample sequence during the interception process.

[0095] Step S223: Perform data augmentation on the intercepted training sample sequences to generate multiple training sample sequences. The data augmentation includes randomly masking some task attributes or simulating equipment failure events.

[0096] The port operation cycle has a certain regularity. For example, considering the daily peak and trough hours of port operations, as well as factors such as the general process time for cargo loading, unloading, and transshipment. Suppose the length of the sliding window is set to 8 hours, which defines the time span of a single training sample sequence, meaning each training sample sequence will contain 8 consecutive hours of port operation data. The step size is set to 4 hours, which is used to define the overlapping ratio between adjacent training sample sequences. Continuously intercept the historical port operation dataset according to this length and step size. For example, starting from the starting moment of the historical data, intercept the first 8 hours of operation data as the first training sample sequence, and then start from the 4th hour and intercept another 8 hours of operation data as the second training sample sequence, and so on. During the interception process, the scheduling policy label corresponding to each training sample sequence should be retained. Perform data augmentation on the intercepted training sample sequences to generate multiple training sample sequences. Data augmentation includes randomly masking some task attributes or simulating equipment failure events. For randomly masking some task attributes, for example, in a certain training sample sequence, randomly select some transportation tasks and mask their cargo type attributes, which can increase the model's adaptability to incomplete task attributes. Regarding simulating equipment failure events, for example, in a certain training sample sequence, simulate that a certain crane fails during a certain period of time, making it unable to work properly during this period, which can enable the model to learn how to make reasonable scheduling in case of sudden equipment failures.

[0097] Moreover, step S240 includes: Step S241, calculate the equipment utilization rate loss term, where the equipment utilization rate loss term is the mean square error of the predicted scheduling policy and the equipment idle duration in the scheduling policy label.

[0098] Step S242, calculate the task completion rate loss term, where the task completion rate loss term is the cross entropy of the predicted scheduling policy and the number of task delays in the scheduling policy label.

[0099] Step S243, calculate the energy consumption loss term, where the energy consumption loss term is the relative error of the predicted scheduling policy and the total energy consumption of the equipment in the scheduling policy label.

[0100] Step S244, perform weighted summation on the equipment utilization rate loss term, the task completion rate loss term, and the energy consumption loss term to obtain the loss function.

[0101] In coastal ports, the idle time of equipment is an important indicator to measure equipment utilization. For example, in the actual scheduling policy label corresponding to a certain training sample sequence, the idle time of a certain crane within 8 hours is 2 hours, while the idle time of this crane in the predicted scheduling policy is 3 hours. The mean square error between the two is calculated to obtain the equipment utilization loss term. Calculate the task completion rate loss term. The task completion rate loss term is the cross-entropy between the predicted scheduling policy and the number of task delays in the scheduling policy label. If in the scheduling policy label, 3 tasks are delayed, while in the predicted scheduling policy, 5 tasks are predicted to be delayed, the cross-entropy between the two is calculated to measure the difference in task completion rate, thereby obtaining the task completion rate loss term. Calculate the energy consumption loss term. The energy consumption loss term is the relative error between the predicted scheduling policy and the total energy consumption of the equipment in the scheduling policy label. Assume that in the scheduling policy label, the total energy consumption of all equipment within 8 hours is 1000 kilowatts, while the total energy consumption calculated in the predicted scheduling policy is 1200 kilowatts. The relative error between the two is calculated to obtain the energy consumption loss term. Finally, the equipment utilization loss term, the task completion rate loss term, and the energy consumption loss term are weighted and summed to obtain the loss function. For example, let the weight of the equipment utilization loss term be 0.3, the weight of the task completion rate loss term be 0.4, and the weight of the energy consumption loss term be 0.3. According to this weight, the three loss terms are weighted and summed to obtain the final loss function.

[0102] In one possible implementation, step S150 includes: Step S151, parsing the equipment start-stop sequence in the scheduling policy to generate an equipment control instruction set, where the equipment control instruction set includes equipment number, operation type, and execution timestamp.

[0103] In this embodiment, in this coastal port, the scheduling policy determines the equipment start-stop sequence. For example, for the cargo handling task, it may be determined to first start several cranes to unload goods from the cargo ship, then start forklifts to transport the goods to the designated area, and then start conveyor belts to transport the goods to the warehouse, etc. An equipment control instruction set is generated according to this equipment start-stop sequence. The equipment control instruction set includes equipment number, operation type, and execution timestamp. Taking the crane as an example, the equipment number may be a specific number assigned according to the numbering rules of the port equipment management system, such as "CR - 001"; the operation type may be "start", indicating that this crane is to be started; the execution timestamp is accurate to the second, such as "2023 - 09 - 15 10:00:00", indicating that this crane is to be started at this time point. For equipment such as forklifts and conveyor belts, the equipment control instruction set is generated in the same way to clarify the operation tasks and execution times of each piece of equipment.

[0104] Step S152: Send the device control instruction set to the corresponding port device controller and monitor the device execution status in real time. The device execution status includes instruction response time, task execution progress, and device fault signal.

[0105] Send the device control instruction set generated for the crane "CR - 001" to the corresponding crane controller. After receiving the instruction, the controller will perform corresponding operations. At the same time, monitor the device execution status in real time. The device execution status includes instruction response time, task execution progress, and device fault signal. The instruction response time refers to the time interval from when the device controller receives the instruction to when the device starts to respond. For example, if the crane "CR - 001" receives a start instruction at 10:00:00 on September 15, 2023, and it starts to have an action response at 10:00:05, then the instruction response time is 5 seconds. The task execution progress refers to the degree of completion of the device during the task execution. For a crane, it can be measured by comparing parameters such as the lifting weight and lifting height with the total task volume. For example, to unload 100 tons of goods from a certain cargo ship, and 30 tons have been lifted at a certain moment, then the task execution progress is 30%. The device fault signal is the signal feedback from the device's own fault detection system. If there is a fault in the crane's motor, a fault signal will be sent.

[0106] Step S153: If it is detected that there is a deviation between the device execution status and the device control instruction set, collect the difference parameters between the current device execution status and the expected status. The difference parameters include the list of unfinished tasks, device delay time, and remaining resources.

[0107] For example, it is found that the task execution progress of the crane "CR - 001" is slower than expected. Originally, it was planned to unload 50 tons of goods within 1 hour, but actually only 30 tons were unloaded within 1 hour. At this time, collect the difference parameters. The list of unfinished tasks is the relevant information of the remaining 20 tons of goods, including the location of the goods (which cabin of the cargo ship, etc.); the device delay time is calculated by comparing the actual time used to unload 30 tons of goods with the expected time. Assuming that it is expected to take 40 minutes to unload 30 tons of goods, and actually it took 50 minutes, then the device delay time is 10 minutes; in terms of the remaining resources, for a crane, it may include the remaining lifting capacity, remaining energy consumption, etc. For example, if the maximum lifting capacity of the crane is 80 tons and 30 tons have been lifted, then the remaining lifting capacity is 50 tons. If the total energy consumption is 1000 kilowatts and 400 kilowatts have been consumed, then the remaining energy consumption is 600 kilowatts.

[0108] Step S154: Perform feature transformation on the difference parameters to generate a difference feature vector, splice the difference feature vector with the original job feature set to obtain an enhanced job feature set, and re-encode the enhanced job feature set through the multi-layer neural network of the scheduling decision model to generate an updated feature set adapted to real-time deviation.

[0109] For example, convert information such as the weight and position of unfinished tasks into specific numerical representations, convert the equipment delay time into a proportional value related to the entire job cycle, and also convert the remaining amount of resources into a unified numerical format. These converted numerical values form the difference feature vector. Then splice the difference feature vector with the original job feature set, which includes the equipment dynamic feature vector, task topology feature vector, and environmental constraint features, etc. The spliced set is the enhanced job feature set.

[0110] Step S155: Generate a corrected equipment control instruction set based on the updated feature set, and send the corrected equipment control instruction set to the port equipment controller to overwrite the original equipment control instruction set.

[0111] In this embodiment, according to the information in the updated feature set, such as the new operating state of the equipment, the adjustment of tasks, and the changes in environmental constraints, etc., re-generate the equipment control instruction set. For the crane "CR - 001", its operating parameters such as the lifting weight, lifting speed, or execution time may be adjusted according to the new task requirements. Send the corrected equipment control instruction set to the corresponding port equipment controller again to overwrite the original equipment control instruction set, so that the equipment operates according to the new instructions to adapt to the real-time deviation situation and ensure the smooth progress of port loading and unloading tasks.

[0112] For example, in a possible implementation manner, step S154 includes: Step S1541: Input the enhanced job feature set into the embedding layer of the multi-layer neural network, align the dimensions of the equipment dynamic feature vector, task topology feature vector, and environmental constraint features in the enhanced job feature set to generate an intermediate feature vector with a unified dimension.

[0113] In this coastal port scenario, the equipment dynamic feature vector may include the operating mode information of multiple devices (such as cranes, forklifts, conveyor belts, etc.) at different time periods, the task topology feature vector includes information such as the spatial dependence relationship between tasks, and the environmental constraint features include the influence weight information of wind speed, humidity, visibility, etc. on job safety. These different types of feature vectors may have different dimensions, and in the embedding layer, they are converted into a unified dimension through a specific mapping method to generate an intermediate feature vector.

[0114] Step S1542: Input the intermediate feature vector into the attention calculation layer of the multi-layer neural network. Based on the interaction relationship between the device dynamic feature vector and the task topology feature vector in the intermediate feature vector, calculate the attention weight matrix between features. The attention weight matrix reflects the dynamic association strength between the device operation features and the task distribution features.

[0115] In port operations, the device operation mode in the device dynamic feature vector is closely related to the task coordination relationship in the task topology feature vector. For example, the lifting speed and efficiency of a crane will affect the task allocation and execution order, and the task distribution will in turn affect the crane's work arrangement. Calculate the attention weight matrix to reflect the dynamic association strength between the device operation features and the task distribution features. If at a certain moment the load rate of the crane is high and there are multiple tasks waiting to be executed, and there is a spatial coordination relationship between some tasks (such as the goods in adjacent cargo ship compartments can be loaded and unloaded together), then in the attention weight matrix, the weights between these relevant device operation features and task distribution features will be adjusted accordingly.

[0116] Step S1543: Weightedly aggregate the intermediate feature vector through the attention weight matrix to generate a weighted global feature vector. The weighted global feature vector contains the cross-modal dependence relationship between the device operation state and the task distribution.

[0117] For example, in the weighted aggregation process, for the device operation features and task distribution features that are closely related to the current task execution, assign larger weights according to the attention weight matrix, and assign smaller weights to the features with less association. The resulting weighted global feature vector can comprehensively reflect the cross-modal dependence relationship between the device operation state and the task distribution, such as how the current operation state of the device affects the task allocation and execution order, and how the task adjustment reacts on the device operation arrangement.

[0118] Step S1544: Input the weighted global feature vector into the time series analysis layer of the multi-layer neural network, extract the time series change pattern of the device operation parameters in the weighted global feature vector, and generate a time series enhanced feature vector. The time series enhanced feature vector reflects the dynamic trend of the device state with the task adjustment.

[0119] In coastal port operations, the variation of equipment operation parameters (such as the lifting weight, lifting height, running speed of a crane, etc.) over time follows certain patterns. Through the time series analysis layer, the temporal variation patterns of these parameters during the equipment's task execution can be analyzed. For example, for the lifting weight of a crane, it may gradually increase at the beginning of the task, remain stable after reaching a peak, and then gradually decrease as the task approaches completion. This temporal variation pattern reflects the dynamic trend of the equipment state adjusting with the task, and the generated temporal enhanced feature vector can capture this dynamic trend, providing a more accurate basis for subsequent decision-making.

[0120] Step S1545: Input the temporal enhanced feature vector into the spatial relationship encoding layer of the multi-layer neural network, and combine the environmental constraint features to perform constraint correction on the spatial dependence path of the task distribution, generating a spatial constraint feature vector, where the spatial constraint feature vector includes the task collaborative execution path under the limitation of environmental safety conditions.

[0121] In a port, the distribution of tasks has spatial dependence. For example, after goods are unloaded from a cargo ship and need to be transported to a warehouse, different warehouse locations and cargo stacking requirements will affect the task execution path. At the same time, environmental constraint features (such as wind speed, humidity, visibility, etc.) will also have an impact on the spatial dependence path of the task distribution. If the wind speed is relatively high, for open-air operation areas, the task path may need to be adjusted to avoid safety risks. Through the spatial relationship encoding layer, combining environmental constraint features to perform constraint correction on the spatial dependence path of the task distribution, the generated spatial constraint feature vector includes the task collaborative execution path under the limitation of environmental safety conditions. For example, in a high-humidity environment, for the transportation task of electronic products, a path passing through a moisture-proof treatment area may be planned to ensure the safety of the goods.

[0122] Step S1546: Perform channel splicing on the temporal enhanced feature vector and the spatial constraint feature vector to generate a fused feature vector, and input the fused feature vector into the normalization layer of the multi-layer neural network for feature scaling to generate a normalized fused feature vector.

[0123] In this process, the temporal enhanced feature vector and the spatial constraint feature vector are combined together through channel splicing to form a fused feature vector. Then the fused feature vector is input into the normalization layer, and the normalization layer performs feature scaling on the fused feature vector according to the set rules, adjusting its numerical range to a suitable interval to generate a normalized fused feature vector. This helps to improve the stability and training efficiency of the model, and is also convenient for the subsequent fully connected layer to operate.

[0124] Step S1547: Input the normalized fused feature vector into the fully connected layer of the multi-layer neural network, and perform high-order feature mapping on the normalized fused feature vector through a non-linear activation function to generate a high-order feature vector.

[0125] The neurons in the fully connected layer are connected to all neurons in the previous layer. By processing the normalized fused feature vector through a non-linear activation function (such as the ReLU function), high-order feature relationships in the data can be mined. In the port operation scenario, these high-order feature relationships may involve more complex interaction relationships between equipment operation and task scheduling, such as the optimal combination method of different equipment under different tasks and environmental conditions. The generated high-order feature vector can better represent this complex relationship and provide richer information for the final decision-making.

[0126] Step S1548: Input the high-order feature vector into the residual connection layer of the multi-layer neural network, perform skip connection with the intermediate feature vector and execute feature superposition to generate a residual enhanced feature vector.

[0127] The role of the residual connection layer is to solve the problem of gradient disappearance in deep neural networks. By performing skip connection and superposition between the high-order feature vector and the intermediate feature vector, more original information can be retained, while enhancing the feature expression ability. In port operations, this residual enhanced feature vector can more comprehensively reflect information in multiple aspects such as equipment operation, task distribution, and environmental constraints, enabling the model to better adapt to real-time deviation situations.

[0128] Step S1549: Input the residual enhanced feature vector into the output layer of the multi-layer neural network, and map the residual enhanced feature vector to a feature space with the same dimension as the original job feature set through linear transformation to generate the updated feature set that adapts to real-time deviation.

[0129] The role of the output layer is to convert the feature vector processed by the multi-layer neural network into a feature space with the same dimension as the original job feature set, so as to generate a corrected equipment control instruction set based on this updated feature set in the future. This updated feature set contains re-evaluation information on aspects such as equipment operation status, task distribution, and environmental constraints, and can adapt to the deviation between the equipment execution status and the equipment control instruction set.

[0130] And, updating the real-time job data based on the adjusted equipment operation status includes: Step S156: Collect the adjusted equipment operation status data, where the equipment operation status data includes the actual start and stop times of the equipment, the task execution time consumption, and the resource consumption.

[0131] In coastal ports, for the crane "CR - 001", the actual start and stop times may be 10:00:00 on September 15, 2023 for start and 12:00:00 on September 15, 2023 for stop; the task execution time is 2 hours, which is different from the original plan of 1.5 hours; in terms of resource consumption, the total energy consumption may be 800 kilowatts, while the original plan was 700 kilowatts. For other equipment such as forklifts and conveyor belts, data such as their actual start and stop times, task execution time, and resource consumption are also collected.

[0132] Step S157: Compare the equipment operation status data with the original real - time operation data to generate a data difference report.

[0133] For example, compare the data such as the actual start and stop times, task execution time, and resource consumption of the crane "CR - 001" collected with the corresponding data in the original real - time operation data. For example, in the original real - time operation data, the planned start and stop times are 10:00:00 on September 15, 2023 for start and 11:30:00 on September 15, 2023 for stop, the task execution time is 1.5 hours, and the planned energy consumption is 700 kilowatts. Through comparison, it is found that the start and stop times are delayed by 30 minutes, the task execution time increases by 30 minutes, and the resource consumption increases by 100 kilowatts. Organize these comparison results into a data difference report, which details the differences of each piece of equipment in each data item.

[0134] Step S158: Incrementally update the equipment operation status information and transportation task queue information in the real - time operation data according to the data difference report, so that subsequent scheduling decisions are generated based on the latest operation status.

[0135] For example, update the equipment operation status information in the real - time operation data according to the content in the data difference report. For the crane "CR - 001", update its actual start and stop times, task execution time, and resource consumption data into the equipment operation status information. At the same time, since the change in the equipment operation status may affect the transportation task queue information, for example, due to the delay of the crane, the task times of subsequent forklifts and conveyor belts need to be adjusted. Make corresponding adjustments to the transportation task queue information, such as postponing the start time of the forklift by 30 minutes and also postponing the starting time of the conveyor belt for transporting goods accordingly, to ensure the correct logical order of the entire transportation task queue. In this way, subsequent scheduling decisions can be made based on these latest operation status data, improving the accuracy and efficiency of port operation scheduling.

[0136] Figure 2The figure shows a hardware structure diagram of a smart port scheduling system 100 provided by an embodiment of the present application for implementing the above-mentioned business data mining method applied to smart port scheduling, as Figure 2 shown, the smart port scheduling system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0137] In a possible design, the smart port scheduling system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the smart port scheduling system 100 may be a distributed system). In some embodiments, the smart port scheduling system 100 may be local or remote. For example, the smart port scheduling system 100 may access information and / or data stored in the machine-readable storage medium 120 via a network. Alternatively, the smart port scheduling system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the smart port scheduling system 100 may be implemented on a smart port scheduling system. By way of example only, the smart port scheduling system may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc., or any aggregation thereof.

[0138] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the smart port scheduling system 100 uses to execute or use to complete the exemplary methods described in the present application.

[0139] In a specific implementation process, one or more processors 110 execute computer-executable instructions stored in the machine-readable storage medium 120, so that the processors 110 can execute the business data mining method applied to smart port scheduling in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected via the bus 130, and the processors 110 may be used to control the transceiver actions of the communication unit 140.

[0140] For the specific implementation process of the processor 110, reference may be made to the various method embodiments executed by the above-mentioned smart port scheduling system 100. The implementation principles and technical effects are similar, and will not be elaborated herein in this embodiment.

[0141] In addition, an embodiment of the present application further provides a readable storage medium, in which computer-executable instructions are set. When a processor runs the computer-executable instructions, the above-mentioned business data mining method applied to smart port scheduling is implemented.

[0142] It should be noted that, in order to simplify the description of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are incorporated into one embodiment, drawing or description thereof. Similarly, it should be noted that, in order to simplify the description of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are incorporated into one embodiment, drawing or description thereof.

Claims

1. A business data mining method applied to smart port scheduling, characterized in that: The method comprises: Acquire real-time operation data generated during the loading and unloading operation of the target port, the real-time operation data including equipment operation status information, transportation task queue information and environmental monitoring information; Dynamically preprocessing the real-time operation data to obtain effective operation data, wherein the dynamic preprocessing includes screening and priority marking conflicting tasks in the real-time operation data according to preset port operation rules; Performing multi-dimensional feature coding on the effective operation data to generate an operation feature set related to port scheduling, wherein the operation feature set includes equipment operation features, task distribution features, and environmental constraint features; Based on the operation feature set, a port equipment scheduling strategy for the current period is generated through a pre-trained scheduling decision model, wherein the port equipment scheduling strategy includes equipment start and stop sequence, task allocation path and resource occupancy ratio, and the scheduling strategy satisfies the balance condition between port operation efficiency and resource consumption; The port equipment is dynamically adjusted according to the scheduling strategy to perform loading and unloading tasks, and the real-time operation data is updated based on the adjusted equipment operation status.

2. The business data mining method applied to smart port scheduling according to claim 1 is characterized in that: The dynamically preprocessing the real-time operation data to obtain effective operation data includes: Separating a device status data subset corresponding to the device operation status information and a task data subset corresponding to the transport task queue information from the real-time operation data; Performing operation noise filtering on the device status data subset to obtain denoised device status data, and performing task conflict detection on the task data subset to obtain a conflicting task list; According to the denoising equipment status data and the conflicting task list, the transport task queue information is prioritized and re-ordered in combination with preset port operation rules to generate the effective operation data. The priority re-ordering is based on a comprehensive judgment of equipment load rate, task urgency and environmental safety conditions.

3. The business data mining method applied to smart port scheduling according to claim 2 is characterized in that: The performing operation noise filtering on the device status data subset to obtain denoised device status data includes: Identifying abnormal operating parameters in the device status data subset, wherein the abnormal operating parameters include operating parameters in which the instantaneous power of the device exceeds a preset threshold or the continuous idle time of the device exceeds a preset range; Performing segmented interpolation processing on the device state data containing the abnormal operation parameters to fill in the data missing, and merging the interpolated device state data with the device state data not containing the abnormal operation parameters into initial denoised device state data; A sliding window mean algorithm is used to perform local average calculation on the equipment operating parameters of the initial denoised equipment state data, and the length of the sliding window is dynamically adjusted according to the equipment type; The equipment operating parameters after local average mean calculation are normalized so that the equipment operating parameters fall within a preset numerical range so as to be jointly encoded with the features of the task data subset to obtain the denoised equipment status data.

4. The business data mining method applied to smart port scheduling according to claim 2 is characterized in that: The performing task conflict detection on the task data subset to obtain a conflicting task list includes: Parsing the task attributes in the task data subset, wherein the task attributes include task start time, task deadline, required equipment type and task dependency; Constructing a task dependency graph according to the task attributes, wherein nodes in the task dependency graph represent independent tasks, and edges represent dependencies between tasks; Traversing all paths in the task dependency graph, detecting whether there is a path cycle or resource overrun conflict, and prioritizing the detected conflicting tasks to obtain the conflicting task list; The step of re-prioritizing the transport task queue information according to the denoising equipment status data and the conflicting task list in combination with preset port operation rules to generate the effective operation data includes: Extracting the current load rate of each port equipment from the denoising equipment status data, and generating an equipment load rate index, wherein the equipment load rate index is the ratio of the average operating time to the idle time of the port equipment within a preset time window; Extracting attribute parameters of each conflicting task from the conflicting task list, the attribute parameters including task deadline, cargo type and priority mark, and generating a task urgency score based on the difference between the task deadline and the current system time; According to the environmental safety threshold defined in the port operation rules, the environmental monitoring information corresponding to the task execution period in the conflicting task list is matched to generate environmental safety constraint parameters, wherein the environmental safety constraint parameters include a combined quantized value of wind speed level, visibility level and humidity level; Based on the equipment load rate index, the task urgency score and the environmental safety constraint parameter, a comprehensive priority score for each conflicting task is constructed, wherein the comprehensive priority score is superimposed by linearly weighting the inverse proportional value of the equipment load rate index, the positive proportional value of the task urgency score and the positive proportional value of the environmental safety constraint parameter; Sort the tasks in the conflicting task list in descending order according to the comprehensive priority scores to generate a priority-sorted conflicting task subqueue; Traversing the priority-sorted conflicting task subqueues, inserting each conflicting task into the non-conflicting task gap in the transport task queue information in turn, wherein the non-conflicting task gap is a device idle period between adjacent tasks in the transport task queue and a period that satisfies the task dependency relationship and has no loop; Based on the inserted transport task queue information, detect whether there is equipment resource over-limit or environmental safety constraint violation. If there is equipment resource over-limit or environmental safety constraint violation, perform secondary screening on the conflicting tasks according to the comprehensive priority score to eliminate the conflicting tasks with equipment resource over-limit or environmental safety constraint violation. The inserted and filtered transport task queue information is time-stamp aligned with the denoised equipment status data to generate the valid operation data including the complete task sequence and equipment status mapping relationship.

5. The business data mining method applied to smart port scheduling according to claim 1 is characterized in that: The multi-dimensional feature encoding of the effective operation data to generate an operation feature set related to port scheduling includes: Extracting time series features associated with equipment operation status information and spatial distribution features associated with transportation task queue information from the effective operation data; Inputting the time series features into a time convolutional network for multi-scale feature extraction to obtain a device dynamic feature vector, wherein the device dynamic feature vector is used as the device operation feature to reflect the operation mode of the device in different time periods; Input the spatial distribution feature into the graph attention network to perform node relationship modeling to obtain a task topology feature vector, wherein the task topology feature vector serves as the task distribution feature and reflects the spatial dependency relationship between tasks; Independently encoding the environmental constraint characteristics to generate the environmental constraint characteristics, wherein the environmental constraint characteristics include the weights of the impact of wind speed, humidity and visibility on operation safety; The device dynamic feature vector, the task topology feature vector and the environmental constraint feature are cross-modally fused to generate the job feature set. The cross-modal fusion dynamically allocates weights of different features through an attention mechanism.

6. The business data mining method applied to smart port scheduling according to claim 5 is characterized in that: The temporal convolutional network comprises a multi-layer causal convolutional module, and the time series features are input into the temporal convolutional network for multi-scale feature extraction, including: Performing sliding window interception on the time series features through the multi-layer causal convolution module to obtain multiple local time segments; Convolution kernel weights are assigned to each local time segment, and the features of adjacent local time segments are superimposed based on the convolution kernel weight assignment results to obtain the global time dependency relationship; Generating the equipment dynamic feature vector according to the global time dependency, wherein the dimension of the equipment dynamic feature vector is proportional to the number of port equipment and the number of task types; And, the spatial distribution features are input into the graph attention network to perform node relationship modeling to obtain the task topology feature vector, including: Inputting the spatial distribution features into a graph attention network to construct a task node graph, wherein each node in the task node graph represents a transportation task, and the edge represents the spatial proximity relationship between tasks; Calculate an attention coefficient for each task node, where the attention coefficient is determined based on the geographical location of the task node, the overlap of device requirements, and the overlap of time windows; The features of adjacent task nodes are weighted and aggregated according to the attention coefficient to generate the task topology feature vector, which includes the collaborative work information of the task group.

7. The business data mining method applied to smart port scheduling according to claim 1 is characterized in that: The training process of the pre-trained scheduling decision model includes: Acquire a historical port operation data set, wherein the historical port operation data set includes operation data samples of multiple historical time periods and corresponding scheduling strategy labels; Performing sliding window segmentation on the historical port operation data set to generate multiple training sample sequences, wherein a single training sample sequence covers the port operation data of a continuous preset time period; Inputting the training sample sequence into the initial scheduling decision model for strategy prediction to obtain a predicted scheduling strategy; Constructing a loss function based on the difference between the predicted scheduling strategy and the scheduling strategy label, wherein the loss function is constructed by combining equipment utilization, task completion rate and energy consumption index; Based on the constructed loss function, the parameters of the initial scheduling decision model are updated by back propagation until convergence, thereby obtaining a pre-trained scheduling decision model that can be generalized to real-time job data.

8. The business data mining method applied to smart port scheduling according to claim 7 is characterized in that: The performing sliding window segmentation on the historical port operation data set to generate a plurality of training sample sequences comprises: The length and step size of the sliding window are set according to the port operation cycle, wherein the length is used to define the time span of a single training sample sequence, and the step size is used to define the overlap ratio between adjacent training sample sequences; Continuously intercepting the historical port operation data set according to the length and step size, and retaining the scheduling strategy label corresponding to each training sample sequence during the interception process; Performing data enhancement on the intercepted training sample sequence to generate multiple training sample sequences, wherein the data enhancement includes randomly masking part of the task attributes or simulating equipment failure events; And, the loss function is constructed by combining equipment utilization, task completion rate and energy consumption index, including: Calculate the equipment utilization loss term, where the equipment utilization loss term is the mean square error between the predicted scheduling strategy and the equipment idle time in the scheduling strategy label; Calculate the task completion rate loss term, where the task completion rate loss term is the cross entropy of the predicted scheduling strategy and the number of task delays in the scheduling strategy label; Calculate the energy consumption loss term, where the energy consumption loss term is the relative error between the predicted scheduling strategy and the total energy consumption of the equipment in the scheduling strategy label; The loss function is obtained by weighted summing the equipment utilization loss term, the task completion rate loss term and the energy consumption loss term.

9. The business data mining method applied to smart port scheduling according to claim 1 is characterized in that: The dynamically adjusting the port equipment according to the scheduling strategy to perform the loading and unloading tasks includes: Parsing the equipment start and stop sequence in the scheduling strategy to generate an equipment control instruction set, wherein the equipment control instruction set includes an equipment number, an operation type, and an execution timestamp; Send the equipment control instruction set to the corresponding port equipment controller, and monitor the equipment execution status in real time, wherein the equipment execution status includes instruction response time, task execution progress and equipment fault signal; If it is detected that the execution state of the device deviates from the device control instruction set, the difference parameters between the current execution state of the device and the expected state are collected, and the difference parameters include the unfinished task list, the device delay time and the remaining amount of resources; Performing feature conversion on the difference parameter to generate a difference feature vector, concatenating the difference feature vector with the original job feature set to obtain an enhanced job feature set, and re-encoding the enhanced job feature set through the multi-layer neural network of the scheduling decision model to generate an updated feature set that adapts to the real-time deviation; generating a revised device control instruction set based on the updated feature set, and issuing the revised device control instruction set to the port device controller to overwrite the original device control instruction set; And, updating the real-time operation data based on the adjusted equipment operation status includes: Collecting the adjusted equipment operation status data, wherein the equipment operation status data includes the actual start and stop time of the equipment, the task execution time and the resource consumption; Compare the equipment operation status data with the original real-time operation data to generate a data difference report; The equipment operation status information and transportation task queue information in the real-time operation data are incrementally updated according to the data difference report, so that subsequent scheduling decisions are generated based on the latest operation status.

10. A smart port dispatching system, characterized in that: The smart port scheduling system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to run the programs, instructions or codes in the memory to implement the business data mining method applied to smart port scheduling as described in any one of claims 1 to 9 above.

Images