Bulk bulk cargo intelligent logistics scheduling method and system based on artificial intelligence

Abstract

The invention discloses a bulk cargo intelligent logistics scheduling method and system based on artificial intelligence. The method comprises the steps that firstly, bulk cargo transportation tasks are acquired, and to-be-scheduled task features are extracted through a task feature extraction network; obtaining transport capacity attribute data of to-be-selected transport capacity, and generating to-be-selected transport capacity features through the transport capacity feature extraction network; screening the optimal transport capacity of which the scheduling priority coefficient is higher than a reference value, and extracting the characteristics of the optimal transport capacity; optimizing the optimal transport capacity feature based on the adaptation coefficient of the task feature and the optimal transport capacity feature; and determining a final scheduling transport capacity and distributing tasks through matching of the optimization features and the to-be-selected transport capacity features. According to the design, multi-dimensional feature extraction of tasks and transport capacity is realized by utilizing an artificial intelligence network, the adaptation precision is improved in combination with a dynamic optimization mechanism, the problems of low matching efficiency and poor real-time performance of traditional scheduling are solved, and the intelligent scheduling level of bulk cargo logistics is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation, and in particular to an artificial intelligence-based intelligent logistics scheduling method and system for bulk cargo. Background Art

[0002] The logistics scheduling of bulk cargo is a core link in supply chain management. Its transportation needs are usually characterized by large volume, strong timeliness, and many special requirements. Traditional scheduling methods rely on manual experience to match transportation tasks with transportation resources, which has significant flaws. It is difficult for humans to fully analyze multi-dimensional information such as the transportation volume, time window, cargo characteristics, and special requirements of the task. At the same time, it is impossible to efficiently evaluate the dynamic attributes of the transportation capacity, such as the load capacity, real-time location, historical punctuality, and equipment configuration. This leads to low accuracy in matching transportation capacity with tasks, and is prone to idle capacity or task delays. In addition, traditional methods lack in-depth mining of task and capacity characteristics and dynamic optimization mechanisms, making it difficult to adapt to the real-time needs of bulk cargo logistics, which restricts the efficient use of logistics resources. Although existing technologies have attempted to record task and capacity data through information systems, they have not achieved feature extraction and intelligent matching based on artificial intelligence, and scheduling efficiency still needs to be improved. Summary of the Invention

[0003] The purpose of the present invention is to provide an artificial intelligence-based intelligent logistics scheduling method and system for bulk cargo.

[0004] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based intelligent logistics scheduling method for bulk cargo, comprising:

[0005] Obtain bulk cargo transportation tasks;

[0006] Extracting task features of the bulk cargo transportation task to obtain features of the task to be scheduled;

[0007] Acquire multiple transport capacity characteristics to be selected; each transport capacity characteristic to be selected is obtained by extracting transport characteristics from transport attribute data of each transport capacity to be selected;

[0008] Obtaining a preferred transport capacity corresponding to the bulk cargo transportation task, and determining a preferred transport capacity characteristic corresponding to the preferred transport capacity; the preferred transport capacity is a transport capacity having a scheduling priority coefficient greater than a transport capacity matching benchmark value; the scheduling priority coefficient represents the degree of adaptability of the preferred transport capacity to the bulk cargo transportation task;

[0009] Optimizing the preferred capacity characteristics according to the adaptation coefficient between the characteristics of the task to be scheduled and the preferred capacity characteristics to obtain an optimized preferred capacity characteristics;

[0010] Matching a dispatching capacity characteristic from the plurality of candidate capacity characteristics according to the adaptation coefficient between the optimized preferred capacity characteristic and each candidate capacity characteristic;

[0011] The bulk cargo transportation task is dispatched to the dispatching capacity corresponding to the dispatching capacity characteristics.

[0012] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0013] Compared with the existing technology, the beneficial effects provided by the present invention include: using the artificial intelligence-based intelligent logistics scheduling method and system for bulk cargo disclosed by the present invention, by obtaining bulk cargo transportation tasks and extracting the characteristics of the tasks to be scheduled through the task feature extraction network; obtaining the capacity attribute data of the selected capacity and generating the characteristics of the selected capacity through the capacity feature extraction network; screening the preferred capacity with a scheduling priority coefficient higher than the benchmark value and extracting its characteristics; optimizing the preferred capacity characteristics based on the adaptation coefficient of the task characteristics and the preferred capacity characteristics; determining the final scheduling capacity and dispatching the tasks by matching the optimized characteristics with the characteristics of the selected capacity. This design utilizes the artificial intelligence network to realize the multi-dimensional feature extraction of tasks and capacity, and combines the dynamic optimization mechanism to improve the adaptation accuracy, thus solving the problems of low efficiency and poor real-time performance of traditional scheduling matching, and improving the level of intelligent scheduling of bulk cargo logistics. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0015] Figure 1 A schematic diagram of the steps of the intelligent logistics scheduling method for bulk cargo based on artificial intelligence provided by an embodiment of the present invention;

[0016] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0018] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0019] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of the intelligent logistics scheduling method for bulk cargo based on artificial intelligence provided in an embodiment of the present disclosure. The following is a detailed introduction to the intelligent logistics scheduling method for bulk cargo based on artificial intelligence.

[0020] Step S201, obtaining bulk cargo transportation tasks;

[0021] Step S202: extracting task features of the bulk cargo transportation task to obtain features of the task to be scheduled;

[0022] Step S203, obtaining multiple transport capacity characteristics to be selected; each transport capacity characteristic to be selected is obtained by extracting the transport attribute data of each transport capacity to be selected;

[0023] Step S204: Obtain the preferred transport capacity corresponding to the bulk cargo transportation task, and determine the preferred transport capacity characteristics corresponding to the preferred transport capacity; the preferred transport capacity is the transport capacity whose scheduling priority coefficient is greater than the transport capacity matching benchmark value; the scheduling priority coefficient represents the adaptability of the preferred transport capacity to the bulk cargo transportation task;

[0024] Step S205: optimizing the preferred capacity characteristics according to the adaptation coefficient between the characteristics of the task to be scheduled and the preferred capacity characteristics to obtain an optimized preferred capacity characteristics;

[0025] Step S206, matching a dispatching capacity characteristic from the plurality of candidate capacity characteristics according to the adaptation coefficient between the optimized preferred capacity characteristic and each candidate capacity characteristic;

[0026] Step S207: dispatching the bulk cargo transportation task to the dispatching transport capacity corresponding to the dispatching transport capacity characteristics.

[0027] In an embodiment of the present invention, for example, in the field of bulk cargo logistics, a common transportation task is that a steel company needs to transport 5,000 tons of iron ore from Port A to Steel Plant B. The transportation time requirement is September 1 to September 5, 2024, the cargo type is lump iron ore (density 1.8 tons / cubic meter), the transportation method is required to be road transportation (due to tight railway transportation capacity), and a dump truck with a load capacity of more than 40 tons is required. At this time, the cargo owner submits the bulk cargo transportation task to the server through the logistics platform, and the server first performs the first step: obtaining the bulk cargo transportation task. The server receives the structured data of the task through the API interface of the logistics platform, including but not limited to the transportation volume (5,000 tons), the starting point (Port A), the destination (Steel Plant B), the time window (September 1-5), the cargo type (lump iron ore), the transportation method (road dump truck), special requirements (load capacity ≥ 40 tons), etc., to form complete transportation task data.

[0028] Next, the server needs to extract features from the transport task to obtain the features of the task to be scheduled. This process relies on a pre-trained task feature extraction network. This network is trained based on historical data: the server has collected a large number of transport task instances (e.g., 100,000 bulk cargo transport orders for ore, coal, and other bulk cargoes over the past year) and corresponding transport capacity instances (e.g., historical dispatch records for 3,000 dump trucks). Some of these instances are annotated with dispatch results (e.g., "efficient transport task instances" with a first score are assigned, while "inefficient transport task instances" with a second score are assigned). For example, if a historical task is "2,000 tons of coal from C to D, within a 3-day time window, requiring trucks with a payload of at least 30 tons," a successful dispatch of a fleet (90% of the fleet's vehicles completed on time with no cargo damage) is assigned a first score (e.g., 0.9 points). If the dispatch fails due to insufficient vehicle payload (only 25 tons), the second score is assigned (e.g., 0.1 points).

[0029] When training the task feature extraction network, the server first processes a transport task instance. For example, the initial task model comprises an initial transport indicator feature extraction layer, an initial waybill description feature extraction layer, an initial feature alignment layer, and an initial attention unit. For the aforementioned 2,000-ton coal transport example, the initial transport indicator feature extraction layer extracts numerical features such as transport volume (2,000 tons), distance (approximately 300 kilometers), time window (3 days), and cargo density (1.5 tons / cubic meter). The initial waybill description feature extraction layer uses natural language processing (e.g., BERT) to extract key descriptions from the waybill text (e.g., "urgent," "no special loading and unloading requirements," and "requires tarpaulin cover for dust protection") and converts them into text feature vectors. The initial feature alignment layer then linearly transforms the numerical transport indicator features (128 dimensions) into a pre-set 256-dimensional feature space. Global average pooling is then performed on the text feature vectors (originally 768 dimensions) to reduce their dimensionality to 256, ensuring consistency between the two dimensions. Next, the initial attention unit uses the aligned transport indicator features as the query vector (Q), the waybill description features as the key vector (K), and the value vector (V). The dot product of Q and K is calculated to obtain a matching score (e.g., [0.8, 0.2, ...]), which is then scaled by √256 (i.e., 16) to stabilize training. A softmax function is then used to obtain an attention weight distribution (e.g., [0.9, 0.1, ...]). Finally, this weight is used to perform a weighted sum on V to obtain a context vector (e.g., focusing on "time window" and "tarpaulin requirements"). This is then superimposed with the transport indicator features to generate a task feature instance (dimension 256) that integrates task requirements and capacity characteristics.

[0030] At the same time, the server processes capacity instances (such as a truck's capacity attribute data instance) and extracts capacity features through an initial capacity model (i.e., the initial capacity feature extraction network). Capacity attribute data includes static attributes (vehicle load capacity of 50 tons, vehicle length of 12 meters, dump function, and tarpaulin equipped) and dynamic attributes (current location at D, estimated arrival time at Port A at 6:00 PM on August 31st, 95% on-time performance over the past three months, and 0% cargo damage rate). The initial capacity model processes static and dynamic attributes through fully connected layers and LSTM layers, respectively, and fuses them into a 256-dimensional capacity feature instance.

[0031] The server calculates the adaptation coefficient (e.g., cosine similarity) between the task feature instance and the transport feature instance, and obtains the sample adaptation coefficient evaluation value. For the efficient instance of successful scheduling (e.g., the 2,000-ton coal task successfully matched with a 50-ton truck), the cosine similarity between the task feature instance and the transport feature instance is 0.85, corresponding to the first score of 0.9, and the first error is calculated (0.9-0.85=0.05); for the failed instance (e.g., the task requires a load of more than 30 tons, and a 25-ton truck is matched), the cosine similarity is 0.3, corresponding to the second score of 0.1, and the second error is calculated (0.1-0.3=-0.2). The server takes the weighted sum of the first error and the second error (such as weights 0.7 and 0.3) to obtain the target scheduling error value (0.05×0.7+(-0.2)×0.3=-0.025), and updates the network parameters of the initial task model and the initial capacity model (such as the weight matrix of the attention unit and the bias term of the fully connected layer) through backpropagation until the error is less than the threshold (such as 0.01) or the maximum number of iterations (such as 100 rounds) is reached, and finally obtains the trained task feature extraction network and capacity feature extraction network.

[0032] For the current 5,000-ton iron ore transportation task, the server uses a trained task feature extraction network for feature extraction: the transportation index feature extraction layer extracts numerical features such as transportation volume (5,000 tons), distance (approximately 400 kilometers), time window (5 days), and cargo density (1.8 tons / cubic meter); the waybill description feature extraction layer processes text such as "dump truck required" and "anti-dust tarpaulin" and converts them into text features; the feature alignment layer maps numerical features (originally 128 dimensions) to 256 dimensions and reduces text features (originally 768 dimensions) to 256 dimensions; the attention unit uses transportation index features as Q and text features as K and V, calculates attention weights (for example, giving high weights to "dump truck" and "5-day time window"), generates a context vector, and superimposes it with the transportation index features to obtain 256-dimensional features of the task to be scheduled (focusing on reflecting key requirements such as "large transportation volume", "long distance", "dumping requirements", and "relaxed time but tarpaulin coverage").

[0033] Next, the server obtains multiple candidate capacity features. These features come from a pre-built capacity database: The server collects capacity attribute information (vehicle load, vehicle model, equipment configuration) and capacity interaction information (historical order volume, on-time performance, cargo damage rate, current location, and available time) from 2,000 initial capacity sources (such as cooperating logistics enterprise fleets and individual drivers) as capacity attribute data. A capacity feature extraction network extracts features from each initial capacity attribute data (for example, fleet A's capacity features are [0.7, 0.8, ...], indicating a 50-ton load, current location, 98% on-time performance, and a tarpaulin; fleet B's features are [0.5, 0.6, ...], indicating a 45-ton load, current location, 90% on-time performance, and no tarpaulin). These 256-dimensional features are stored in the capacity database. In daily operations, the server scans the capacity database every 24 hours to obtain the latest attribute data of each capacity (for example, the location of a truck is updated to Tianjin after completing the previous order). If data changes are found (such as the location changes from A to Tianjin), it will be marked as capacity to be optimized, and the optimized capacity features (such as the updated location features) will be extracted again through the capacity feature extraction network, and the old features in the database will be replaced to ensure the real-time nature of the capacity features.

[0034] Then, the server needs to determine the preferred capacity corresponding to the transport task. The preferred capacity is the capacity whose scheduling priority coefficient is greater than the capacity matching benchmark value (such as 0.7). The scheduling priority coefficient indicates the adaptability of the capacity to the task. The server first calculates the adaptation coefficient (such as cosine similarity) between the characteristics of the task to be scheduled and the characteristics of each candidate capacity in the capacity database to obtain the initial adaptation coefficient of each candidate capacity. For example, the adaptation coefficient of fleet A is 0.85 (the load capacity is 50 tons, which matches the task requirements. It can arrive on time at A and has a tarpaulin to prevent dust), the adaptation coefficient of fleet B is 0.6 (the load capacity is 45 tons, which is close to the requirements but slightly lower. It currently needs to travel 1 hour more in Baoding and is not satisfied without a tarpaulin), and the adaptation coefficient of fleet C is 0.9 (the load capacity is 60 tons, it is currently in A, the punctuality rate is 100%, and there is a tarpaulin). The server selects fleets A and C with adaptation coefficients greater than 0.7 as preferred transport capacities and extracts their preferred transport capacity features (256-dimensional feature vector of fleet A and 256-dimensional feature vector of fleet C).

[0035] Next, the server needs to optimize the preferred capacity characteristics based on the fit coefficient between the characteristics of the task to be scheduled and the preferred capacity characteristics to obtain the optimized preferred capacity characteristics. For example, the characteristics of the task to be scheduled emphasize "5,000 tons of large transport volume" (corresponding to the value of feature dimension 10 of 0.9) and "dumping requirement" (the value of dimension 20 is 0.8). However, in the preferred capacity characteristics of fleet C, the load characteristic (dimension 10) is 0.95 (corresponding to 60 tons) and the dump function characteristic (dimension 20) is 0.9 (with dumping), and the cosine similarity between the two is 0.9. The server adjusts the feature vector of fleet C through an optimization algorithm (such as gradient ascent), increasing the weight of dimensions highly correlated with the task characteristics (for example, increasing the value of dimension 10 from 0.95 to 0.98 and dimension 20 from 0.9 to 0.95), while reducing the weight of less relevant dimensions (such as the maintenance record dimension, which has no special requirements for the task), thereby generating an optimized preferred capacity characteristic (more in line with the core requirements of "large load" and "dumping" of the task).

[0036] Subsequently, the server matches the dispatch capacity characteristics based on the adaptation coefficient between the optimized preferred capacity characteristics and the characteristics of each candidate capacity. For example, the cosine similarity is calculated between the optimized fleet C characteristics and the candidate capacity characteristics (including fleets A, B, D, etc.), and it is found that the adaptation coefficient of fleet C itself is still the highest (0.98), while the adaptation coefficient of fleet A is 0.88, and the adaptation coefficient of fleet D (55 tons of load, currently in A, with tarpaulin) is 0.92. However, the server may further consider the real-time status of the capacity: although fleet C currently has 5 vehicles available, 2 of them are under maintenance, and 3 are actually available; fleet D has 8 vehicles available, and all are equipped with tarpaulins. At this time, the server combines the "number of available vehicles" dimension in the optimized preferred capacity characteristics and adjusts the adaptation coefficient calculation (such as adding a weight of 0.2 for the number of available vehicles). Finally, the comprehensive adaptation coefficient of fleet D (0.92×0.8+number of available vehicles 8 / 10×0.2=0.736+0.16=0.896) exceeds that of fleet C (0.98×0.8+number of available vehicles 3 / 5×0.2=0.784+0.12=0.904), so fleet D is selected as the scheduling capacity.

[0037] Finally, the server dispatches a transport task to Fleet D corresponding to the dispatched capacity characteristics. The server sends the task details (5,000 tons of iron ore, from A to B, September 1st to 5th, requiring a dump truck and tarpaulin) to Fleet D's management system via the logistics platform, along with a dispatch confirmation link. After Fleet D confirms the order, the server records the dispatch result (success) and uses the task characteristics, capacity characteristics, and dispatch results for this task as new training data. Subsequently, the server periodically performs transfer learning training. When a certain amount of new data has accumulated (e.g., 1,000 new dispatch records), it selects a subset of tasks (e.g., 100 recent ore transport tasks). Based on the fit coefficient between the characteristics of the task to be dispatched and the capacity characteristics trained on the baseline model (characteristics of the dispatched capacity), the server calculates the error based on the actual dispatch results (0.9 for success and 0.1 for failure), and then updates the parameters of the task and capacity feature extraction networks. For example, the server adjusts the weight of the "number of available vehicles" dimension in the attention unit to make the model more attentive to the real-time availability of transport resources in future dispatches, improving the accuracy of the fit coefficient calculation.

[0038] In summary, this method uses artificial intelligence technology to achieve intelligent matching of bulk cargo transportation tasks and transportation capacity. From task feature extraction, transportation capacity feature management, optimal transportation capacity screening to final scheduling, the entire process combines historical data training and real-time data optimization, ensuring the efficiency and accuracy of scheduling, and significantly improving the operational efficiency of bulk cargo logistics.

[0039] In the embodiment of the present invention, extracting the task features of the bulk cargo transportation task to obtain the features of the task to be scheduled can be implemented through the following examples.

[0040] The task feature extraction network is used to extract the task features of the bulk cargo transportation task to obtain the task features to be scheduled; wherein the task feature extraction network is generated by training the initial task model based on the error between the sample adaptation coefficient evaluation value and the sample scheduling result annotation; the sample adaptation coefficient evaluation value is the adaptation coefficient between the task feature instance and the capacity feature instance, the task feature instance is obtained by extracting the task features of the transportation task instance according to the initial task model, and the capacity feature instance is obtained by extracting the capacity features of the capacity attribute data instance according to the initial capacity model; the sample scheduling result annotation indicates that the scheduling result of the transportation task instance is dispatched to the capacity instance;

[0041] The acquisition of multiple transport capacity characteristics to be selected may be implemented through the following examples.

[0042] The capacity feature extraction network is used to extract the capacity attribute data of each candidate capacity to obtain a plurality of candidate capacity features; the capacity feature extraction network is used to obtain, train and generate the initial capacity model based on the error between the sample adaptation coefficient evaluation value and the sample scheduling result label.

[0043] In an embodiment of the present invention, illustratively, taking a 5,000-ton iron ore transportation task commissioned by a steel company as an example, the specific process of the server performing task feature extraction and obtaining the characteristics of the transport capacity to be selected is as follows: First, the server receives the complete data of the transportation task, including structured information (transportation volume 5,000 tons, starting point A port, destination B steel plant, time window September 1-5, cargo density 1.8 tons / cubic meter, transportation method requires road dump trucks, load ≥ 40 tons) and unstructured descriptions ("needs to be covered with dust-proof tarpaulins" and "priority is given to vehicles with no recent cargo damage records"). At this time, the server calls the trained task feature extraction network to extract features for the task. The training of the task feature extraction network is based on historical sample data. For example, the server collected 100,000 bulk cargo transportation task instances (such as coal and ore transportation orders) from 2022 to 2023 and corresponding 30,000 transportation capacity instances (such as truck data of cooperative fleets). Some of the instances were marked with scheduling results: "efficient instances" of successful scheduling (such as a 3,000-ton coal task matched with a dump truck fleet with a load capacity of 50 tons and a punctuality rate of 98%, marked as 0.9 points) and "inefficient instances" of failed scheduling (such as a 2,000-ton ore task that was timed out due to matching with a non-dump truck with a load capacity of 35 tons, marked as 0.1 points). During training, the server first uses the initial task model to process transport task instances. For high-efficiency instances (such as a 3,000-ton coal task), the initial transport indicator feature extraction layer extracts numerical features (128-dimensional vectors) such as transport volume (3,000 tons), distance (280 kilometers), and time window (4 days). The initial waybill description feature extraction layer uses the BERT model to process text such as "self-unloading required" and "dust prevention" to generate a 768-dimensional text feature vector. Subsequently, the initial feature alignment layer linearly transforms the numerical features to 256 dimensions and performs global average pooling on the text features to reduce the dimensionality to 256 to ensure dimensionality consistency. The initial attention unit uses numerical features as query vectors (Q), text features as key vectors (K) and value vectors (V), calculates the dot product of Q and K (e.g., [0.7, 0.3, ...]), divides by √256 (16), and scales it to obtain attention weights through Softmax (e.g., assigning 0.8 weights to "dumping" and "time window"). The weighted fusion V generates a context vector, which is then superimposed with the numerical features to obtain a 256-dimensional task feature instance (focusing on "large transportation volume" and "dumping demand"). At the same time, the initial capacity model processes capacity instances (e.g., attribute data of a dump truck fleet): static attributes (load capacity 50 tons, vehicle length 12 meters, tarpaulin) are generated into 128-dimensional vectors through the fully connected layer, and dynamic attributes (current location A, 98% on-time rate in the past three months, 0% cargo damage rate) are generated into 128-dimensional vectors through the LSTM layer. After fusion, a 256-dimensional capacity feature instance is obtained.The server calculates the cosine similarity between task feature instances and capacity feature instances (e.g., the similarity for efficient instances is 0.85), uses this as the sample fit coefficient, and calculates an error (0.05) compared to the annotation score (0.9). For inefficient instances (e.g., a task requiring a payload ≥ 40 tons, matching a fleet with a payload of 35 tons, with a similarity of 0.3 and an annotation of 0.1), the error is -0.2. The network parameters of the initial task model (e.g., the weight matrix of the attention unit) and the initial capacity model (e.g., the hidden layer parameters of the LSTM) are adjusted through backpropagation until the error is less than 0.01, ultimately yielding the trained task feature extraction network and capacity feature extraction network. For the current 5,000-ton iron ore task, the server calls the task feature extraction network: the transportation indicator layer extracts numerical features (128 dimensions) such as 5,000 tons, 400 kilometers, and 5 days, and the waybill description layer processes text such as "anti-dust tarpaulin" to generate a 768-dimensional vector; the feature alignment layer maps the numerical features to 256 dimensions, and the text features are reduced to 256 dimensions; the attention unit calculates the weight (giving a weight of 0.9 to "5,000 tons" and "dumping"), generates a context vector, and then superimposes the numerical features to obtain a 256-dimensional feature of the task to be scheduled (with a focus on "super-large transportation volume" and "dumping + tarpaulin urgent need"). To obtain the characteristics of the candidate transport capacity, the server retrieves the attribute data of 2,000 cooperating transport capacities from the transport capacity database (e.g., Fleet A: 50-ton load, current location A, 98% punctuality, with tarpaulin; Fleet B: 45-ton load, current location Baoding, 90% punctuality, no tarpaulin). This is then processed through a transport feature extraction network: static attributes (load, vehicle type) are passed through a fully connected layer, and dynamic attributes (location, punctuality) are passed through an LSTM layer. These attributes are then integrated to generate a 256-dimensional feature vector for each transport capacity (e.g., Fleet A's features are [0.8, 0.9, ...], while Fleet B's are [0.6, 0.7, ...]). These features serve as the candidate transport capacity characteristics for subsequent scheduling and matching. In summary, through the trained dual networks (task feature extraction network and capacity feature extraction network), the server transforms the complex information of transport tasks and capacity into high-dimensional feature vectors, providing a standardized and computable matching foundation for subsequent intelligent scheduling.

[0044] In the embodiments of the present invention, the following implementation modes are also provided.

[0045] Obtaining the transport task instance and the transport attribute data instance corresponding to the transport capacity instance, wherein the transport task instance is configured with the sample scheduling result annotation;

[0046] Extracting task features from the transportation task instance according to the initial task model to obtain the task feature instance;

[0047] Extracting capacity features from the capacity attribute data instance according to the initial capacity model to obtain the capacity feature instance;

[0048] Determining a fit coefficient between the task feature instance and the capacity feature instance, and obtaining an evaluation value of the sample fit coefficient;

[0049] The initial task model and the initial capacity model are trained according to the error between the sample adaptation coefficient evaluation value and the sample scheduling result label to obtain the task feature extraction network and the capacity feature extraction network.

[0050] In this embodiment of the present invention, for example, a logistics platform server trains an initial task model and an initial transport capacity model. The specific process is as follows: The server first extracts the raw data required for training from the historical scheduling database. For example, from January to December 2023, the platform recorded a total of 80,000 bulk cargo transport task instances and corresponding transport capacity instances, of which 50,000 were marked as "successful scheduling" (with a first score, such as 0.9 points) and 30,000 were marked as "failed scheduling" (with a second score, such as 0.1 points). Take one of the efficient transportation task instances as an example: the transportation task instance is "On May 10, 2023, 3,000 tons of coke will be transported from E to F, with a time window of 3 days (May 12-14), requiring a box dump truck with a load capacity of more than 40 tons, and must be equipped with a dust-proof tarpaulin"; the corresponding transportation capacity instance is a logistics fleet G, and its transportation capacity attribute data instance includes static attributes (vehicle load capacity of 50 tons, vehicle type is a 12-meter box dump truck, equipped with a tarpaulin) and dynamic attributes (vehicle position at E at 18:00 on May 11, 2023, punctuality rate of 97% in the past three months, cargo damage rate of 0%). The instance was marked as 0.9 points for successfully completing the scheduling. Another inefficient transport task example is "1,500 tons of iron powder transported from H to I on July 20, 2023, within a two-day window (July 22-23), requiring a dump truck with a capacity of 35 tons or more." The corresponding capacity example is fleet H, whose capacity attribute data shows a vehicle with a capacity of only 30 tons, an on-time rate of 80% over the past three months, and no dump truck capability. This example is labeled 0.1 due to the vehicle's insufficient capacity and inability to arrive on time. After obtaining these capacity examples, transport task examples, and annotations, the server begins processing the data using the initial task model and initial capacity model. For the efficient transportation task instance, the initial task model's processing steps are as follows: The initial transportation indicator feature extraction layer first extracts numerical indicators from the transportation task instance, including transportation volume (3,000 tons), transportation distance (approximately 350 kilometers), time window (3 days), and cargo density (1.4 tons / cubic meter), generating a 128-dimensional sample transportation indicator feature vector. The initial waybill description feature extraction layer processes key descriptions such as "van dump truck" and "anti-dust tarpaulin" in the task text through a pre-trained BERT model to generate a 768-dimensional sample waybill description feature vector. Subsequently, the initial feature alignment layer performs a linear transformation on the sample transportation indicator features (such as matrix multiplication W1×vector+b1), mapping them to a preset 256-dimensional feature space. At the same time, the sample waybill description features are globally average pooled (averaging the 768-dimensional vector by dimension) to reduce the dimensionality to 256 dimensions to ensure consistency between the two dimensions.The initial attention unit uses the aligned transport indicator features as the query vector (Q), the waybill description features as the key vector (K) and the value vector (V), calculates the dot product of Q and K to obtain a matching score (e.g., [0.8, 0.2, ..., 0.5]), divides it by √256 (i.e., 16) to scale it to stabilize the gradient, and then uses the Softmax function to convert the score into an attention weight distribution (e.g., assigning weights of 0.7 and 0.2 to the dimensions corresponding to "van dump" and "time window"). Finally, the attention unit performs a weighted summation of V according to the weights to generate a context vector that reflects the key requirements of the task (e.g., focusing on "dumping function" and "3-day time limit"), and superimposes it with the aligned transport indicator features to obtain a 256-dimensional task feature instance. For a capacity instance (e.g., fleet G), the initial capacity model processes its capacity attribute data instance: static attributes (50-ton load capacity, vehicle model, and tarpaulin configuration) are processed through a fully connected layer (parameter matrix W2 and bias b2) to generate a 128-dimensional static feature vector. Dynamic attributes (location, on-time performance, and cargo damage rate) are processed through an LSTM layer (hidden layer size 128) to process time series data (e.g., on-time performance changes over the past three months) to generate a 128-dimensional dynamic feature vector. The static and dynamic feature vectors are fused by element-wise addition to obtain a 256-dimensional capacity feature instance. The server calculates the cosine similarity between the task feature instance and the capacity feature instance as the sample adaptation coefficient evaluation value. For example, in the efficient instance, the cosine similarity between the task feature instance (focusing on "3,000 tons," "dumping," and "3 days") and the capacity feature instance of fleet G (50-ton load, dumping, and high punctuality) is 0.88, corresponding to a sample adaptation coefficient evaluation value of 0.88; in the inefficient instance, the cosine similarity between the task feature instance ("1,500 tons," "over 35 tons," and "2 days") and the capacity feature instance of fleet H (30-ton load, no dumping, and low punctuality) is 0.25, and the evaluation value is 0.25. Subsequently, the server calculates the error between the sample adaptation coefficient evaluation value and the sample scheduling result annotation: the error of the efficient instance is 0.9 (annotation) - 0.88 (evaluation value) = 0.02; the error of the inefficient instance is 0.1 (annotation) - 0.25 (evaluation value) = -0.15. The server takes a weighted average of the errors across all instances (e.g., a weight of 0.6 for efficient instances and 0.4 for inefficient instances) to obtain the target scheduling error (0.02 × 0.6 + (-0.15) × 0.4 = -0.048). It then uses the backpropagation algorithm to update the network parameters of the initial task model (e.g., the weight matrix W1 of the attention unit) and the initial capacity model (e.g., the gating parameters of the LSTM layer). After multiple rounds of iterative training (e.g., 100 rounds), training terminates when the target scheduling error drops below 0.01. The initial task model and initial capacity model are then designated as the task feature extraction network and the capacity feature extraction network, respectively, ready for feature extraction and matching in actual tasks.

[0051] In an embodiment of the present invention, the transport task instance includes an efficient transport task instance that is successfully dispatched to an efficient transport capacity instance, and an inefficient transport task instance that fails to be dispatched to an inefficient transport capacity instance. The sample scheduling result annotation includes a first score annotation and a second score annotation, and the first score is greater than the second score; the efficient transport task instance is configured with the first score annotation, and the inefficient transport task instance is configured with the second score annotation;

[0052] The determination of the adaptation coefficient between the task feature instance and the capacity feature instance and the acquisition of the sample adaptation coefficient evaluation value may be implemented through the following example.

[0053] Determine an adaptation coefficient between an efficient task feature instance and an efficient transport capacity feature instance corresponding to the efficient transport capacity instance, and obtain a first sample adaptation coefficient evaluation value; the efficient task feature instance is a task feature corresponding to the efficient transport task instance;

[0054] Determining a fit coefficient between an inefficient task feature instance and an inefficient transport capacity feature instance corresponding to the inefficient transport capacity instance, to obtain a second sample fit coefficient evaluation value; the inefficient task feature instance is a task feature corresponding to the inefficient transport task instance;

[0055] The initial task model and the initial capacity model are trained according to the error between the sample adaptation coefficient evaluation value and the sample scheduling result label to obtain the task feature extraction network and the capacity feature extraction network, which can be implemented through the following examples.

[0056] Training the initial task model and the initial capacity model until a training termination state is met according to a first error between the first sample fitness coefficient evaluation value and the first score annotation, and a second error between the second sample fitness coefficient evaluation value and the second score annotation;

[0057] The initial task model upon completion of training is determined as the task feature extraction network, and the initial capacity model upon completion of training is determined as the capacity feature extraction network.

[0058] In an embodiment of the present invention, illustratively, taking the actual scenario of a logistics platform server training a task feature extraction network and a transportation capacity feature extraction network as an example, the specific process is as follows: the server selects two types of transportation task instances from the historical scheduling records: one type is an efficient transportation task instance that is successfully dispatched to an efficient transportation capacity instance (such as a 5,000-ton coal transportation task that is successfully matched with a dump truck fleet with a load capacity of 60 tons and a punctuality rate of 98%), and the other type is an inefficient transportation task instance that fails to be dispatched to an inefficient transportation capacity instance (such as a 2,000-ton ore transportation task that is timed out due to matching a fleet with a load capacity of 30 tons and no dump truck function). The efficient instance is configured with a first score label (such as 0.9 points), and the inefficient instance is configured with a second score label (such as 0.1 points). First, the server processes the efficient transportation task instance. For example, consider the task of transporting 5,000 tons of coal from J to K in October 2023, with a four-day time window (October 15-18), requiring dump trucks with a capacity of 50 tons or more and equipped with dust-proof tarpaulins. The corresponding efficient transport instance is fleet X (12-meter dump trucks with a capacity of 60 tons, located at J on October 14, 2023, with a 98% on-time rate and a 0% cargo damage rate over the past three months). The server uses the initial task model to extract features for this efficient transport task instance: the initial transport indicator feature extraction layer extracts numerical features (128-dimensional vectors) such as transport volume (5,000 tons), distance (approximately 400 kilometers), time window (4 days), and cargo density (1.3 tons / cubic meter). The initial waybill description feature extraction layer processes text such as "dump truck" and "dust-proof" using the BERT model to generate a 768-dimensional text feature vector. The initial feature alignment layer linearly transforms the numerical features to 256 dimensions, and the text features are reduced to 256 dimensions through global average pooling to ensure dimensional consistency. The initial attention unit uses the numerical features as the query vector (Q), the text features as the key vector (K) and the value vector (V), calculates the dot product of Q and K (e.g., [0.85, 0.1, ...]), divides by √256 (16), and scales it to obtain the attention weight through Softmax (giving a weight of 0.7 to "dumping" and "4-day time window"). The context vector V is weighted and fused, and then superimposed with the numerical features to obtain a 256-dimensional efficient task feature instance (focusing on reflecting "large transportation volume" and "dumping demand"). At the same time, the server uses the initial capacity model to process fleet X's capacity attribute data instance: static attributes (60-ton load capacity, vehicle model, and tarpaulin configuration) are processed through a fully connected layer to generate 128-dimensional static features; dynamic attributes (location, punctuality, and cargo damage rate) are processed through an LSTM layer to generate 128-dimensional dynamic features. After fusion, these fusions produce a 256-dimensional high-efficiency capacity feature instance (focusing on "high load capacity" and "high punctuality"). The server calculates the cosine similarity (e.g., 0.88) between the high-efficiency task feature instance and the high-efficiency capacity feature instance as the first sample adaptation coefficient evaluation value.Then, we process inefficient transportation task instances, such as "In August 2023, 2,000 tons of ore will be transported from L to M, with a time window of 2 days (August 20-21), requiring a dump truck with a load capacity of more than 35 tons". The corresponding inefficient transportation capacity instance is fleet Y (vehicle load capacity of 30 tons, no dump function, location in Lianyungang on August 19, 2023, 80% punctuality rate in the past 3 months, and 1% cargo damage rate). The server uses the initial task model to extract inefficient task feature instances (focusing on "2,000 tons", "more than 35 tons", and "2 days"), and the initial transportation capacity model to extract inefficient transportation capacity feature instances (focusing on "low load", "no dump", and "low punctuality"), and calculates the cosine similarity of the two (such as 0.25) as the second sample adaptation coefficient evaluation value. The server calculates the error: the first error is the first score annotation (0.9) - the first sample adaptation coefficient evaluation value (0.88) = 0.02; the second error is the second score annotation (0.1) - the second sample adaptation coefficient evaluation value (0.25) = -0.15. The server takes a weighted average of the first errors of all efficient instances and the second errors of all inefficient instances (e.g., 60% efficient instances and 40% inefficient instances) to obtain the target scheduling error value (0.02 × 0.6 + (-0.15) × 0.4 = -0.048). Using the backpropagation algorithm, the server adjusts the network parameters of the initial task model (e.g., the weight matrix of the attention unit) and the initial capacity model (e.g., the gating parameters of the LSTM layer) to gradually reduce the target error. After multiple rounds of iteration (e.g., 100 rounds), training terminates when the target error drops below 0.01. At this time, the initial task model is determined as the task feature extraction network, and the initial capacity model is determined as the capacity feature extraction network, which can accurately extract the core features of tasks and capacity, providing a basis for subsequent intelligent scheduling.

[0059] In an embodiment of the present invention, the initial task model and the initial capacity model are trained according to the first error between the first sample adaptation coefficient evaluation value and the first score annotation, and the second error between the second sample adaptation coefficient evaluation value and the second score annotation until the training termination state is met, which can be implemented through the following examples.

[0060] Determining a first scheduling error value based on a first error between the first sample adaptation coefficient evaluation value and the first scoring label;

[0061] determining a second scheduling error value according to a second error between the second sample adaptation coefficient evaluation value and the second scoring label;

[0062] determining a target scheduling error value according to the first scheduling error value and the second scheduling error value;

[0063] The network parameters corresponding to the initial task model and the initial capacity model are updated according to the target scheduling error value until the training termination state is met.

[0064] In the embodiment of the present invention, for example, taking the actual scenario of a logistics platform server training the initial task model and the initial transportation capacity model as an example, the server's error calculation and model parameter update process based on efficient and inefficient transportation task instances is as follows: the server selects two typical instances from historical data: the efficient transportation task instance is "4,000 tons of cement will be transported from N to O in November 2023, with a time window of 3 days (November 20-22), requiring a dump truck with a load capacity of more than 45 tons and equipped with an automatic tarpaulin", and the corresponding efficient transportation capacity instance is fleet M ( For example, a 50-ton vehicle, located in N on November 19, 2023, with a 99% on-time rate and a 0% cargo damage rate over the past three months, is assigned a first score of 0.9. An example of an inefficient transport task is "1,500 tons of sand and gravel transported from O to Q in June 2023, with a one-day time window (June 10), requiring a dump truck with a capacity of 30 tons or more." This corresponds to an inefficient transport capacity example, fleet N (25-ton vehicle, no dump truck, located in Hangzhou on June 9, 2023, with a 75% on-time rate and a 2% cargo damage rate over the past three months), and is assigned a second score of 0.1. The server first extracts features using the initial task model and the initial transport capacity model and calculates the adaptation coefficient. For high-efficiency examples, the initial task model extracts high-efficiency task feature examples (focusing on "4,000 tons," "3-day time window," and "automatic tarpaulin"). The initial capacity model extracts high-efficiency capacity feature examples (focusing on "50-ton load," "high punctuality," and "automatic tarpaulin"). The cosine similarity between the two is 0.89, indicating that the first sample fitness coefficient is 0.89. The first error is the first score annotation (0.9) minus the first sample fitness coefficient (0.89), i.e., 0.9 - 0.89 = 0.01. For low-efficiency examples, the initial task model extracts low-efficiency task feature examples (focusing on "1,500 tons," "1-day time window," and "over 30 tons"). The initial capacity model extracts low-efficiency capacity feature examples (focusing on "25-ton load," "low punctuality," and "no dump"). The cosine similarity between the two is 0.22, indicating that the second sample fitness coefficient is 0.22. The second error is the second score label (0.1) minus the second sample fitness coefficient evaluation value (0.22), that is, 0.1-0.22=-0.12. The server needs to integrate the errors of all training instances to guide model optimization. Assuming that the current batch contains 100 efficient instances and 50 inefficient instances, the server calculates the target error weighted by the number of instances: the average first error of efficient instances is 0.01 (the errors of all 100 instances are close to 0.01); the average second error of inefficient instances is -0.12 (the errors of all 50 instances are close to -0.12); the target scheduling error value = (100×0.01+50×(-0.12)) / (100+50) = (1-6) / 150 = -5 / 150≈-0.033. The server uses the target scheduling error value as the optimization target and updates the network parameters of the initial task model and initial capacity model through the backpropagation algorithm.For example: Initial task model: The original parameter of the weight matrix of the "time window" dimension in its attention unit is W. q (query vector weight), after calculating the gradient, it is found that the “time window” does not contribute enough to the adaptation coefficient of the efficient instance (error 0.01), so W q The column parameter corresponding to the "time window" in the initial capacity model is increased by 0.005 to enhance the importance of this dimension in feature extraction; the hidden layer bias b of its LSTM layer processing dynamic attributes (such as punctuality) h The original parameter is 0.3. After calculating the gradient, it is found that the inefficient instance has a high adaptation coefficient (error -0.12) due to the "low punctuality". Therefore, b h The parameter W1 in the linear transformation matrix of the feature alignment layer of the initial task model was adjusted to 0.25 to reduce the feature score of low punctuality capacity. At the same time, the parameter of a row of the linear transformation matrix W1 in the feature alignment layer of the initial task model was originally 0.7. Because the "automatic tarpaulin" feature of the efficient instance matched the "automatic tarpaulin" feature of the capacity well but the error still existed, the parameter of this row was fine-tuned to 0.72 to enhance the alignment effect of text features and numerical features. The server sets the training termination condition as the absolute value of the target scheduling error value is less than 0.01 or the number of iterations reaches 200 rounds. After multiple rounds of iterations (such as the 150th round), the target error value gradually narrowed to -0.008 (absolute value 0.008<0.01), meeting the termination condition and completing the training. At this point, the parameters of the initial task model and the initial capacity model have been optimized to a stable state and are respectively determined as the task feature extraction network and the capacity feature extraction network, which can accurately capture the core matching features of tasks and capacity, providing a reliable feature representation basis for subsequent intelligent scheduling.

[0065] In an embodiment of the present invention, the initial task model includes an initial transport index feature extraction layer, an initial waybill description feature extraction layer, an initial feature alignment layer, and an initial attention unit. The task feature extraction of the transport task instance according to the initial task model to obtain the task feature instance can be implemented through the following example.

[0066] Extracting transport index features from the transport task instance according to the initial transport index feature extraction layer to obtain sample transport index features;

[0067] Extracting waybill description features from the transport task instance according to the initial waybill description feature extraction layer to obtain sample waybill description features;

[0068] According to the initial feature alignment layer, the sample transport index features are mapped to a preset feature dimension through linear transformation to generate aligned transport index features; global feature extraction is performed on the sample waybill description features to generate aligned waybill description features consistent with the transport index feature dimension;

[0069] According to the initial attention unit, the aligned transport indicator features are used as query vectors, and the aligned waybill description features are used as key vectors and value vectors; the matching score between the query vector and the key vector is calculated, and the score is scaled in accordance with the feature dimension; the scaled matching score is converted into an attention weight distribution through a normalization function; the value vector is dynamically weighted fused according to the attention weight distribution to generate a context vector reflecting the key features of the task; the context vector and the aligned transport indicator features are cross-modally superimposed to obtain the task feature instance that integrates the task requirements and capacity characteristics.

[0070] In an embodiment of the present invention, illustratively, taking a logistics platform server using an initial task model to process the "5,000-ton coal transportation task instance in October 2023", the generation process of the task feature instance is described in detail as follows: The specific information of the transportation task instance is: transportation volume 5,000 tons, starting point J, destination K, transportation distance of approximately 400 kilometers, time window 4 days (October 15-18), cargo density 1.3 tons / cubic meter, and the waybill description text is "A van dump truck with a load capacity of more than 50 tons is required, equipped with a dust-proof tarpaulin". The server gradually processes the instance through the four core modules of the initial task model (initial transportation indicator feature extraction layer, initial waybill description feature extraction layer, initial feature alignment layer, and initial attention unit). The initial transportation indicator feature extraction layer is responsible for extracting structured numerical indicators in the task. The server first parses out the numerical parameters from the transport task instance: transport volume (5000 tons), transport distance (400 kilometers), time window (4 days), cargo density (1.3 tons / cubic meter), minimum load requirement (50 tons), etc. These parameters are standardized (such as transport volume divided by 1000 tons, distance divided by 100 kilometers, time window divided by 1 day), and then converted into a 128-dimensional vector through a fully connected layer. For example, the transport volume of 5000 tons is standardized to 5.0, and the corresponding vector dimension 10 is 5.0; the time window of 4 days is standardized to 4.0, and the corresponding vector dimension 20 is 4.0. Finally, a 128-dimensional sample transport index feature vector (denoted as F) is generated. metric ). The initial waybill description feature extraction layer processes the unstructured text description in the task. The server inputs the waybill text "a dump truck with a load capacity of more than 50 tons is required, equipped with a dust-proof tarpaulin" into the pre-trained BERT model, and generates a 768-dimensional word vector sequence through word embedding, position embedding, and segment embedding (each word corresponds to a 768-dimensional vector). Subsequently, the output corresponding to the [CLS] tag is taken as the global text feature (representing the semantics of the entire sentence), and a 768-dimensional sample waybill description feature vector (denoted as F textFor example, the dimension values ​​corresponding to “dumping” and “tarpaulin” in this vector are relatively high (such as 0.8 and 0.7), reflecting the task’s requirements for dumping and dust prevention. The function of the initial feature alignment layer is to transform F metric and F text Mapped to the same preset feature dimension (such as 256 dimensions). metric (128 dimensions), the server performs dimensionality increase through linear transformation (matrix multiplication plus bias): using the 128×256 weight matrix W1 and the 256-dimensional bias vector b1, calculate F metric,aligned =F metric ×W1+b1, and obtain the 256-dimensional aligned transport index feature (denoted as F metric,aligned ). For F text (768 dimensions), the server reduces the dimension by global average pooling: the average value of each dimension of the 768-dimensional vector is calculated (for example, the average value of the first to the 768th dimension is calculated separately), and the 256-dimensional aligned waybill description feature is obtained (denoted as F text,aligned ), ensure that the metric,aligned The initial attention unit fuses the aligned features through the self-attention mechanism. The server will F metric,aligned As the query vector (Q), F text,aligned It serves as both a key vector (K) and a value vector (V). First, the dot product of Q and K is calculated to obtain a matching score matrix (such as a 256×256 matrix, where each element represents the degree of matching between the i-th dimension of Q and the j-th dimension of K). To stabilize the training, the score matrix is ​​scaled by dividing it by √256 (i.e., 16). Subsequently, the scaled score is converted into an attention weight distribution through the Softmax function (the sum of each row in the weight matrix is ​​1, and a high weight indicates that the dimension is more critical to the task). For example, the K dimension corresponding to "dumping" has a high matching degree with the "load requirement" dimension of Q, with a weight of 0.9; the K dimension corresponding to "time window" has the second highest matching degree with the "time limit" dimension of Q, with a weight of 0.8. The server performs a weighted fusion of V according to the attention weight to generate a context vector (C), where each dimension is the weighted sum of the dimensions of V (such as the i-th dimension of C = Σ(the i-th row and j-th column of the weight matrix × the j-th dimension of V)). The context vector C focuses on reflecting the key requirements of the task (such as "dumping function" and "anti-dust tarpaulin"). Finally, C is combined with F metric,aligned Perform cross-modal superposition (add corresponding dimensions, such as the 10th dimension of C + F metric,aligned 10th dimension), and obtain a 256-dimensional task feature instance (F task), which combines numerical metrics (e.g., 5,000 tons, 4 days) with textual requirements (e.g., "dumping," "tarpaulin") to fully reflect the core matching characteristics of the task. Through these steps, the initial task model transforms complex transport task instances into high-dimensional feature vectors, providing a standardized input foundation for subsequent adaptive calculations based on transport capacity characteristics.

[0071] In the embodiments of the present invention, the following implementation modes are also provided.

[0072] Obtaining capacity attribute data corresponding to multiple initial capacities;

[0073] Extracting capacity characteristics from the capacity attribute data of each initial capacity according to the capacity characteristic extraction network to obtain initial capacity characteristics corresponding to each initial capacity;

[0074] Building a capacity database according to the initial capacity characteristics corresponding to each initial capacity;

[0075] The acquisition of multiple transport capacity characteristics to be selected may be implemented through the following examples.

[0076] According to the transport capacity database, the plurality of transport capacity characteristics to be selected are obtained.

[0077] In an embodiment of the present invention, for example, taking the actual scenario of a logistics platform server building a bulk cargo capacity database as an example, the process of collecting attribute data of initial capacity, extracting features and storing them is as follows: the server first obtains capacity attribute data of multiple initial capacities from the logistics companies and individual drivers cooperating with the platform. These initial capacities include 2,000 cooperative fleets and 5,000 individual drivers, and their capacity attribute data covers static attributes and dynamic attributes. For example: Fleet A (large logistics company): static attributes are "vehicle type: 12-meter box dump truck, single vehicle load capacity of 50 tons, equipped with automatic tarpaulin, and GPS positioning function"; dynamic attributes are "current location: Port A (10:00 on August 20, 2024), 80 orders received in the past three months, 98% on-time rate, 0% cargo damage rate, next available time: 0:00 on August 21". Driver B (individual transport operator): Static attributes include "Vehicle type: 8-meter ordinary truck (without dump function), single vehicle load capacity of 35 tons, no tarpaulin"; dynamic attributes include "Current location: Terminal H (August 20, 2024, 12:00 PM), 20 orders received in the past three months, 85% on-time rate, 1% cargo damage rate, next available time: August 22, 6:00 PM." The server organizes these structured (such as load capacity, on-time rate) and unstructured (such as "equipped with automatic tarpaulin") attribute data and calls a trained capacity feature extraction network for feature extraction. This network, trained based on historical data, can convert complex capacity information into 256-dimensional feature vectors. Taking fleet A as an example: static attributes are processed through the fully connected layer: keywords such as "50-ton load capacity", "dumping function", and "automatic tarpaulin" are encoded into 128-dimensional static feature vectors (for example, the 10th dimension is 1.0, indicating "dumping function", and the 20th dimension is 0.9, indicating "automatic tarpaulin"); dynamic attributes are processed through the LSTM layer: time series data such as "current location", "98% punctuality", and "available time" are encoded into 128-dimensional dynamic feature vectors (for example, the 50th dimension is 0.8, indicating "high punctuality", and the 60th dimension is 0.7, indicating "nearly available"); static and dynamic feature vectors are fused by element-by-element addition to obtain a 256-dimensional initial capacity feature (for example, [0.9, 0.8, 0.7, ...]), which comprehensively reflects the core advantages of fleet A: "high load capacity, high reliability, and near-term dispatchability". For driver B, the capacity feature extraction network processes his attribute data and generates initial capacity features of [0.5, 0.3, 0.4, ...]. Low numerical dimensions correspond to disadvantages such as "no dump function," "low punctuality," and "no tarpaulin." The server associates all 256-dimensional feature vectors of the initial capacity with basic information (such as capacity ID, fleet affiliation, and contact information) to construct a capacity database.The database uses a distributed storage architecture. Each capacity record contains: a basic identifier: the capacity ID (e.g., F001 for fleet A, D001 for driver B); a feature vector: a 256-dimensional floating-point array (e.g., [0.9, 0.8, ...] for fleet A); raw attribute data: the original values ​​of static and dynamic attributes (e.g., "load capacity 50 tons," "punctuality 98%"); and a timestamp: the time the feature was updated (e.g., 10:30 AM, August 20, 2024). When the server needs to obtain the characteristics of a candidate capacity, it selects the capacity from the capacity database that meets the basic requirements (e.g., load capacity ≥ 40 tons, with dump truck capability) based on the current transport task requirements (e.g., "5,000 tons of iron ore from A to B requires a dump truck with a load capacity of 40 tons or more") and retrieves the stored initial capacity characteristics. For example, if the server recognizes that a task requires a "dump truck," it will filter out from the database the transport capacity with a "dump function" dimension value ≥ 0.8 in the feature vector (such as Fleet A, Fleet C, etc.), obtain its 256-dimensional feature vector as the candidate transport capacity feature, and use it for subsequent adaptive calculations with the task features. Through this process, the server converts dispersed transport capacity information into standardized, computable feature vectors, constructing a dynamically updated transport capacity database, and providing an efficient feature retrieval and matching foundation for the intelligent scheduling of bulk cargo logistics.

[0078] In the embodiments of the present invention, the following implementation modes are also provided.

[0079] Acquire the transport attribute data corresponding to each initial transport capacity in the transport capacity database according to a preset time period;

[0080] Determine the initial transport capacity with changed transport capacity attribute data as the transport capacity to be optimized;

[0081] Extracting capacity characteristics from the optimized capacity attribute data of the capacity to be optimized according to the capacity characteristic extraction network to obtain optimized capacity characteristics of the capacity to be optimized;

[0082] The transport capacity database is optimized according to the optimized transport capacity characteristics of the transport capacity to be optimized.

[0083] In an embodiment of the present invention, for example, taking the actual scenario in which a logistics platform server updates the bulk freight capacity database daily as an example, the process of optimizing the capacity characteristics according to the preset period is as follows: the server sets the preset time period to 24 hours (3 am every day) to trigger the capacity database optimization process. First, the server obtains the latest capacity attribute data of all initial capacities in the capacity database through the Internet of Things interface of the logistics platform, the driver APP positioning system and the historical order system. These data include static attributes (such as vehicle type and load capacity, which usually do not change) and dynamic attributes (such as current location, punctuality rate in the past 3 months, and next available time, which change in real time). The server compares the newly acquired capacity attribute data with the old data stored in the database field by field, and filters out the initial capacity whose attribute data has changed as the capacity to be optimized. For example, Fleet X (original data: location A at 10:00 on August 20, 2024, 98% on-time performance over the past three months, available until 0:00 on August 21): New data shows that it completed the A-B transport task at 0:00 on August 21, 2024, and its current location has been updated to B. Its on-time performance over the past three months has increased to 99% due to the addition of one new order, and its available time has been adjusted to 0:00 on August 22. Due to changes in location, on-time performance, and available time, it is marked as a capacity for optimization. Driver Y (original data: location H at 12:00 on August 20, 2024, 85% on-time performance over the past three months, available until 18:00 on August 22): New data shows that it has not received any orders, its location remains H, its on-time performance remains unchanged, its available time remains unchanged, and it is not marked as a capacity for optimization. The server invokes the trained capacity feature extraction network to extract features from the latest attribute data for each capacity for optimization. Taking fleet X as an example: static attributes (vehicle type: 12-meter van dump truck, load capacity: 50 tons, equipped with automatic tarpaulin) are generated through the fully connected layer to generate a 128-dimensional static feature vector (for example, the 10th dimension = 1.0 represents "dumping function", and the 20th dimension = 0.9 represents "automatic tarpaulin"); dynamic attributes (current location: B, on-time rate of 99% in the past three months, available time: 00:00 on August 22) are processed through the LSTM layer to process time series data (such as location change trajectory, on-time rate improvement trend) to generate a 128-dimensional Dynamic feature vectors (e.g., dimension 50 = 0.95 represents "extremely high punctuality," dimension 60 = 0.8 represents "available tomorrow"); static and dynamic feature vectors are combined element-wise to create a 256-dimensional optimized capacity feature (e.g., [0.92, 0.85, 0.95, ...]). Compared to the old feature (original [0.9, 0.8, 0.9, ...]), the values ​​in the "location," "punctuality," and "available time" dimensions are higher, more accurately reflecting fleet X's current reliability and dispatch availability. The server replaces the optimized capacity feature with the old feature in the database.For example, the database record for fleet X originally had a feature vector of [0.9, 0.8, 0.9, ...], which is updated to [0.92, 0.85, 0.95, ...]. The original attribute data (location, on-time performance, and availability) and the timestamp (marked as 3:30 AM, August 21, 2024) are also updated simultaneously. For unchanged capacity (such as driver Y), the database retains the original record. Through this process, the capacity database always stores the latest capacity characteristics, ensuring that the capacity information used in subsequent dispatch truly reflects the current status. For example, when a steel company submits a new task of "transporting 5,000 tons of iron ore from B to I", the server retrieves the optimized features of fleet X from the database (showing that it is currently at B, has a 99% punctuality rate, and is available tomorrow). When calculating its adaptation coefficient with the task characteristics ("originating at B", "large transportation volume", and "relaxed time window"), it will obtain a higher matching score due to the high values ​​of the "location matching" and "high punctuality rate" dimensions, and will be scheduled first, thereby improving the real-time matching accuracy of tasks and transportation capacity.

[0084] In the embodiment of the present invention, the acquisition of the transport capacity attribute data corresponding to the multiple initial transport capacities may be implemented through the following examples.

[0085] Obtaining capacity attribute information and capacity interaction information corresponding to multiple initial capacities;

[0086] The transport capacity attribute information and transport capacity interaction information corresponding to each initial transport capacity are determined as the transport capacity attribute data corresponding to each initial transport capacity.

[0087] In an embodiment of the present invention, illustratively, taking the case of a logistics platform server collecting capacity attribute data of initial bulk cargo capacity, the process of obtaining capacity attribute information and capacity interaction information and integrating them is as follows: the server obtains two types of core information through a multi-source data interface for all initial capacities registered on the platform (including logistics enterprise fleets, individual transport households, etc.): capacity attribute information (static data reflecting the inherent characteristics of capacity) and capacity interaction information (dynamic data reflecting the operational performance of capacity), and merges the two into complete capacity attribute data. Capacity attribute information is the basic attribute of the initial capacity, which is mainly obtained through platform registration information and hardware equipment data. For example: Logistics enterprise fleet (initial capacity A): the server extracts static attribute information from the enterprise's registration information on the platform, including vehicle type (such as 12-meter box dump truck), single vehicle load (50 tons), number of vehicles (20 vehicles), equipment configuration (whether equipped with automatic tarpaulin, GPS positioning device, weighing sensor), vehicle certificate validity (whether the driving license and road transport license are within the validity period), etc. This information is entered into the platform by uploading documents such as business licenses and vehicle driving licenses when the fleet is registered, and the server stores it after OCR recognition and manual review and verification. Individual transport operators (initial capacity B): The server extracts attribute information from the driver's personal registration information, including vehicle type (8-meter ordinary truck), single vehicle load (35 tons), vehicle color (red), whether there is a dump function (no), whether a cargo monitoring camera is installed (no), etc. This information is filled in by the driver APP and the vehicle photo and electronic driving license are uploaded, and then automatically verified and stored by the server. Capacity interaction information is the operational behavior data of the initial capacity on the platform, which is collected in real time through IoT devices, order systems and user evaluation systems. For example, for logistics company fleet A, the server obtains its vehicles' real-time locations (e.g., at Port A at 10:00 AM on September 1, 2024) and driving trajectories (route from Port B back to Port A over the past 24 hours) through onboard GPS devices. It also obtains historical interaction data from the order management system, including the number of orders received (60), completed order types (40 for ore transport, 20 for coal transport), and average transport distance (380 kilometers) over the past three months. It also obtains cargo damage data from the cargo monitoring system (0% damage rate over the past three months). It also obtains shipper ratings from the user review system (average 4.9 out of 5). For individual transporter B, the server obtains its current location (at Terminal H at 12:00 PM on September 1, 2024) and online status (idle / accepting orders) through the driver app. It also obtains historical interaction data from the order system, including the number of orders received (15), number of timeouts (2), and number of order cancellations (1) over the past three months. It also obtains specific issues reported by shippers (e.g., "On August 10, 2024, sand and gravel were transported without a tarp, resulting in some spillage"). The server associates the attribute information of the same initial capacity with the interaction information to form complete capacity attribute data of the capacity.For example, fleet A's capacity attribute data includes the following: "Vehicle type: 12-meter van dump truck, 50-ton load capacity, 20 vehicles, equipped with automatic tarpaulin, valid GPS positioning, and valid documents"; interactive information: "Location at 10:00 on September 1, 2024: Port A, 60 orders received in the past three months, 98% on-time rate, 0% cargo damage rate, and 4.9 shipper rating." Individual transport operator B's capacity attribute data includes the following: "Vehicle type: 8-meter ordinary truck, 35-ton load capacity, no dump truck, no tarpaulin installed, and no cargo monitoring cameras"; interactive information: "Location at 12:00 on September 1, 2024: Terminal H, 15 orders received in the past three months, 80% on-time rate, 1% cargo damage rate, and shipper feedback: 'No tarpaulin covering caused spillage.'" Through this process, the server integrates scattered static attributes and dynamic behavior data into structured capacity attribute data, providing a comprehensive input basis for the subsequent generation of initial capacity characteristics through the capacity feature extraction network and the construction of a capacity database, ensuring that the capacity portrait not only contains inherent capabilities (such as load capacity, vehicle type), but also reflects actual performance (such as punctuality, cargo damage), thereby improving the accuracy of subsequent intelligent scheduling.

[0088] In an embodiment of the present invention, after extracting the task features of the bulk cargo transportation task based on the task feature extraction network and obtaining the features of the task to be scheduled, the embodiment of the present invention further provides the following implementation methods.

[0089] Obtaining multiple basic model training capacities corresponding to the bulk cargo transportation task and basic model training scheduling result labels corresponding to each basic model training capacity;

[0090] Extracting capacity features from each basic model training capacity according to the capacity feature extraction network to obtain basic model training capacity features of each basic model training capacity;

[0091] Determine the basic model training adaptation coefficient evaluation value corresponding to each basic model training capacity according to the adaptation coefficient between the characteristics of the task to be scheduled and the characteristics of each basic model training capacity;

[0092] According to the error between the basic model training adaptation coefficient evaluation value corresponding to each basic model training capacity and the basic model training scheduling result labeling corresponding to each basic model training capacity, the task feature extraction network and the capacity feature extraction network are subjected to transfer learning training to obtain an optimized task feature extraction network and an optimized capacity feature extraction network.

[0093] In an embodiment of the present invention, for example, taking the actual scenario of a logistics platform server performing transfer learning training for the "5,000 tons of iron ore from A to B" transportation task as an example, the process of optimizing the task feature extraction network and the capacity feature extraction network is as follows: After completing the extraction of the task features to be scheduled for the transportation task, the server selects 5 cooperative capacities as the basic model training capacity (such as fleets P, Q, R, S, and T), assigns tasks to them, and collects actual scheduling results as annotations. For example, fleet P (50-ton dump trucks, currently stationed at A, with a 98% on-time performance) successfully completed a delivery (no cargo damage, on-time delivery), and is labeled 0.9. Fleet Q (45-ton dump trucks, currently stationed at Baoding, with a 90% on-time performance) was delayed by one day due to a vehicle failure, and is labeled 0.3. Fleet R (35-ton trucks, non-dump trucks, with a 85% on-time performance) declined an order due to insufficient capacity, and is labeled 0.1. Fleet S (60-ton dump trucks, currently stationed at B, with a 100% on-time performance) declined an order due to the high cost of having to travel 400 kilometers empty to A to load cargo, and is labeled 0.2. Fleet T (55-ton dump trucks, currently stationed at A, with a 95% on-time performance) successfully completed a delivery and received high ratings from the shipper, and is labeled 0.95. The server invokes the existing capacity feature extraction network to extract features from the latest attribute data of the capacity trained on the five basic models. For example, fleet P's capacity attribute data includes "50-ton load, dump truck function, current location A, and 98% punctuality." After network processing, it generates 256-dimensional basic model training capacity features (e.g., [0.85, 0.9, 0.88, ...]). Fleet R's capacity attribute data includes "35-ton load, no dump truck function, and 85% punctuality," generating features of [0.3, 0.2, 0.4, ...] (low values ​​reflect low matching ability). The server calculates the cosine similarity between the characteristics of the scheduled task (the extracted 256-dimensional vector, focusing on "5,000 tons," "dump truck," and "origin location A") and the training capacity features of each basic model as the adaptation coefficient evaluation value. For example, the similarity between fleet P's features and task features is 0.88 (evaluation value); the similarity between fleet R's features and task features is 0.22 (evaluation value); and the similarity between fleet T's features and task features is 0.92 (evaluation value). The server calculated the error between each evaluation value and the annotation: Fleet P: 0.9 (annotation) - 0.88 (evaluation) = 0.02; Fleet R: 0.1 (annotation) - 0.22 (evaluation) = -0.12; Fleet T: 0.95 (annotation) - 0.92 (evaluation) = 0.03. Based on these errors, the server fine-tuned the parameters of the task feature extraction network and the capacity feature extraction network through transfer learning.For example, in the task feature extraction network, the original attention unit had a low weight for the "origin location match" dimension (e.g., 0.6). However, fleet T had a high match with "currently in A" but a slightly lower evaluation value (error 0.03). Therefore, the weight of this dimension was increased to 0.65, thereby strengthening the task feature's focus on the "origin location." In the capacity feature extraction network, the original LSTM layer processing "punctuality" had a hidden layer bias of 0.3. Fleet P's high punctuality (98%) corresponded to an evaluation value of 0.88 (error 0.02), indicating that this dimension was underrepresented. Therefore, the bias was adjusted to 0.32, thereby strengthening the feature score for high punctuality capacity. In response to fleet R's low evaluation error (-0.12), the parameters of the fully connected layer of the capacity feature extraction network were adjusted, reducing the values ​​of the "low load" and "no dump" dimensions (e.g., from 0.3 to 0.25) to make its features more accurately reflect low matching ability. After multiple rounds of transfer learning training (e.g., 10 rounds), training was terminated when the average error dropped below 0.01. The resulting optimized task feature extraction network and capacity feature extraction network can more accurately capture key dimensions such as "origin location matching" and "high punctuality," improving the accuracy of the calculation of the adaptation coefficient between tasks and capacity in subsequent scheduling. For example, in subsequent similar tasks, the capacity with a high punctuality rate at the task origin will receive a higher evaluation value and be prioritized in scheduling, further improving logistics efficiency.

[0094] In an embodiment of the present invention, there are multiple transport capacities to be selected, and the acquisition of multiple basic model training transport capacities corresponding to the bulk cargo transportation task and the basic model training scheduling result annotations corresponding to each basic model training transport capacity can be implemented through the following examples.

[0095] Obtaining a candidate adaptation coefficient corresponding to each candidate transport capacity characteristic according to the adaptation coefficient between the characteristics of the task to be scheduled and the characteristics of each candidate transport capacity;

[0096] Matching the multiple basic model training capacities from the multiple candidate capacities according to the candidate adaptation coefficients corresponding to the characteristics of the candidate capacities;

[0097] Dispatching the bulk cargo transportation task to the plurality of basic model training transport capabilities;

[0098] According to the scheduling results corresponding to the multiple basic model training capacities, the basic model training scheduling result labels corresponding to the respective basic model training capacities are determined.

[0099] In an embodiment of the present invention, for example, taking a logistics platform server processing the transportation task of "5,000 tons of iron ore from port A to steel plant B", the specific process of selecting the basic model training capacity and generating annotations is as follows: the server has extracted the features of the task to be scheduled (256-dimensional vector, focusing on "5,000 tons", "starting point A", "dump truck", and "anti-dust tarpaulin") through the task feature extraction network, and obtained 20 candidate capacity features from the capacity database (such as the 256-dimensional vector of fleet A to fleet T). The server calculates the cosine similarity between each candidate capacity feature and the feature of the task to be scheduled as the candidate adaptation coefficient. For example: the adaptation coefficient of fleet A (60-ton dump truck, currently in A, 99% punctuality, with tarpaulin) is 0.92; the adaptation coefficient of fleet B (50-ton dump truck, currently in Baoding, 95% punctuality, with tarpaulin) is 0.85 (due to the location, it takes an extra hour to reach A); the adaptation coefficient of fleet C (45-ton dump truck, currently in A, 90% punctuality, without tarpaulin) is 0.78 (due to the lack of tarpaulin, it does not meet the dust prevention requirements); the adaptation coefficient of fleet D (35-ton ordinary truck, currently in A, 85% punctuality, without dump truck) is 0.3 (insufficient load and no dump truck function); the adaptation coefficient of fleet E (55-ton dump truck, currently in B, 100% punctuality, with tarpaulin) is 0.6 (it needs to drive 400 kilometers empty to A to load the cargo, which is costly). Based on the candidate fitness coefficients, the server selects transport capacities covering high, medium, and low fitness levels as the basic model training capacities to comprehensively evaluate model performance. The specific selection rules are: the top three transport capacities with the best fitness coefficients (high match), the middle two (medium match), and the last two (low match), for a total of seven capacities. In this example, the following were selected: High match: Fleet A (0.92), Fleet B (0.85); Medium match: Fleet C (0.78), Fleet E (0.6); Low match: Fleet D (0.3), and another low-fit fleet F (fitness coefficient 0.25). The server sends task details (5,000 tons of iron ore, A to B, time window September 1st-5th, dump truck and tarpaulin required) to these seven basic model training capacities and obtains order acceptance feedback and actual execution results for each capacity through the logistics platform API. For example: Fleet A confirmed the order, departed from A at 8:00 AM on September 1st, and arrived on time at 6:00 PM on September 3rd with no cargo damage. Fleet B confirmed the order, but due to road congestion from Baoding to A, departed at 12:00 PM on September 1st and arrived at 10:00 AM on September 4th (a 16-hour delay). Fleet C was rejected by the cargo owner due to lack of tarpaulin and did not accept the order. Fleet E actively declined the order due to high empty-run costs. Fleet D declined the order due to insufficient load capacity (35 tons) and inability to transport 5,000 tons (requiring 143 trips, exceeding its capacity of 10 vehicles). Fleet F declined the order due to vehicle failures (two vehicles under repair, three vehicles in transit) and no available vehicles. The server assigns a labeling score (0-1, with higher scores indicating better scheduling) to each base model training capacity based on the actual scheduling results.The specific rules are: successful completion without problems (such as fleet A): 0.95 points; successful but with delays (such as fleet B): 0.5 points; shipper or transport capacity actively refuses (such as fleets C and E): 0.2 points; unable to undertake due to insufficient capacity (such as fleets D and F): 0.1 points. In the end, the annotations of the training capacity of each basic model are as follows: fleet A: 0.95; fleet B: 0.5; fleet C: 0.2; fleet E: 0.2; fleet D: 0.1; fleet F: 0.1. Through this process, the server generates labeled data based on the actual scheduling results, providing real feedback for subsequent transfer learning, ensuring that the optimized task feature extraction network and capacity feature extraction network can more accurately capture key factors such as "location matching", "equipment compliance", and "capacity sufficiency", thereby improving the accuracy of future scheduling. For example, if fleet C is rejected due to the lack of a tarpaulin (marked as 0.2), the model will adjust the weight of the "tarpaulin equipment" dimension so that the "anti-dust" requirement is more emphasized when extracting subsequent task features, and the adaptation coefficient of the capacity without a tarpaulin is reduced when extracting capacity features, thereby reducing similar ineffective scheduling.

[0100] In the embodiments of the present invention, the following implementation modes are also provided.

[0101] Extracting task features of the bulk cargo transportation task according to the optimized task feature extraction network to obtain optimized task features to be scheduled;

[0102] The method further comprises:

[0103] Extracting capacity characteristics from the capacity attribute data corresponding to the preferred capacity according to the optimized capacity characteristic extraction network to obtain target preferred capacity characteristics corresponding to the preferred capacity;

[0104] Optimizing the preferred transport capacity characteristics according to the adaptation coefficient between the characteristics of the task to be scheduled and the preferred transport capacity characteristics to obtain the optimized preferred transport capacity characteristics includes:

[0105] According to the adaptation coefficient between the optimized task characteristics to be scheduled and the target preferred capacity characteristics, the preferred capacity characteristics are optimized to obtain the optimized preferred capacity characteristics.

[0106] In an embodiment of the present invention, for example, taking a logistics platform server processing the transportation task of "5,000 tons of iron ore from port A to steel plant B", the specific process of applying the optimized network and optimizing the preferred transportation capacity features is as follows: After completing the transfer learning training of the task feature extraction network and the transportation capacity feature extraction network, the server first calls the optimized task feature extraction network to reprocess the original transportation task. The original information of the task includes: transportation volume of 5,000 tons, starting point port A, destination steel plant B, time window from September 1st to 5th, and a dump truck with a load of more than 40 tons and anti-dust tarpaulin. The optimized task feature extraction network adjusts the weights of the attention units in transfer learning (such as enhancing attention to the "starting point matching" and "tarpaulin equipment" dimensions). When the server processes tasks through this network: the initial transport indicator feature extraction layer extracts numerical features (128 dimensions) such as transport volume (5,000 tons), distance (400 kilometers), and time window (5 days); the initial waybill description feature extraction layer processes text such as "anti-dust tarpaulin" and "dump truck" to generate 768-dimensional text features; the feature alignment layer maps numerical features to 256 dimensions, and reduces text features to 256 dimensions; the attention unit uses numerical features as query vectors (Q) and text features as key vectors (K) and value vectors (V). When calculating the matching score, the K dimension weights corresponding to "origin A" and "tarpaulin equipment" are increased from 0.6 to 0.8 (because transfer learning found that these two dimensions have a significant impact on scheduling results). In the final optimized features (256 dimensions) for the tasks to be scheduled, the value of the "origin matching" dimension is increased from 0.7 to 0.85, and the value of the "tarpaulin requirement" dimension is increased from 0.8 to 0.92, more accurately reflecting the core requirements of the task. The server selects the original preferred capacity from the capacity database (e.g., Fleet A, 60-ton dump truck, currently located at Port A, 99% on-time performance, with tarpaulin), and calls the optimized capacity feature extraction network to process its latest capacity attribute data (location: Port A at 10:00 AM on September 1, 2024, 99% on-time performance over the past three months, 15 available vehicles). The optimized capacity feature extraction network adjusts the parameters of the LSTM layer during transfer learning (for example, enhancing the representation of the "on-time performance" and "current location" dimensions). During processing: static attributes (60-ton load capacity, dump function, and tarpaulin) are converted into 128-dimensional static features through the fully connected layer (the value of the "dump" dimension is increased from the original 0.9 to 0.95); dynamic attributes (current location A, 99% punctuality, and 15 available vehicles) are converted into 128-dimensional dynamic features through the LSTM layer (the value of the "current location matching" dimension is increased from the original 0.8 to 0.9, and the value of the "high punctuality" dimension is increased from the original 0.9 to 0.98). After the static and dynamic features are fused, a 256-dimensional target optimization capacity feature is obtained, among which the values ​​of key dimensions such as "origin matching", "high punctuality", and "tarpaulin configuration" are significantly improved (for example, [0.92, 0.98, 0.95,...]).The server calculates the cosine similarity (such as 0.96) between the optimized features of the task to be scheduled and the target preferred capacity features, and uses this as the basis for optimizing the preferred capacity features. The optimization logic is: increase the weight of the dimension with high correlation with the task features and reduce the weight of the dimension with low correlation. For example: the value of the "starting place matching" dimension of the optimized feature of the task to be scheduled is 0.85, and the value of the "current location matching" dimension of the target preferred capacity feature is 0.9 (corresponding to Port A). The two are highly matched, so the value of this dimension is increased to 0.95; the value of the "tarpaulin demand" dimension of the optimized feature of the task to be scheduled is 0.92, and the value of the "tarpaulin equipment" dimension of the target preferred capacity feature is 0.95 (already met). The value of this dimension is maintained and fine-tuned to 0.96 to strengthen the match; the value of the "maintenance record" dimension in the target preferred capacity feature (no special requirements for the task) is 0.7. Because it has a low correlation with the task requirements, it is reduced to 0.6 to reduce interference with the overall adaptation coefficient. Ultimately, the optimized preferred capacity characteristics were adjusted to [0.95, 0.98, 0.96, 0.6, ...], focusing more on the core requirements of the mission (origin, tarpaulin, and high punctuality). When the server subsequently schedules based on the compatibility coefficients of these optimized characteristics with the other candidate capacity characteristics, Fleet A will receive a higher matching score, ensuring that missions are prioritized for the most suitable capacity, thereby improving logistics efficiency.

[0107] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent logistics scheduling method for bulk cargo based on artificial intelligence. Figure 2 As shown, Figure 2 This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0108] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. An artificial intelligence-based intelligent logistics dispatching method for bulk cargo, characterized by: include: Obtain bulk cargo transportation tasks; Extracting task features of the bulk cargo transportation task to obtain features of the task to be scheduled; Obtaining multiple characteristics of transport capacity to be selected; The characteristics of each transport capacity to be selected are obtained by extracting the transport characteristics from the transport attribute data of each transport capacity to be selected; Obtaining a preferred transport capacity corresponding to the bulk cargo transportation task, and determining a preferred transport capacity characteristic corresponding to the preferred transport capacity; the preferred transport capacity is a transport capacity having a scheduling priority coefficient greater than a transport capacity matching benchmark value; the scheduling priority coefficient represents the degree of adaptability of the preferred transport capacity to the bulk cargo transportation task; Optimizing the preferred capacity characteristics according to the adaptation coefficient between the characteristics of the task to be scheduled and the preferred capacity characteristics to obtain an optimized preferred capacity characteristics; Matching a dispatching capacity characteristic from the plurality of candidate capacity characteristics according to the adaptation coefficient between the optimized preferred capacity characteristic and each candidate capacity characteristic; The bulk cargo transportation task is dispatched to the dispatching capacity corresponding to the dispatching capacity characteristics.

2. The method according to claim 1, characterized in that Extracting the task features of the bulk cargo transportation task to obtain the features of the task to be scheduled includes: The task feature extraction network is used to extract the task features of the bulk cargo transportation task to obtain the task features to be scheduled; wherein the task feature extraction network is generated by training the initial task model based on the error between the sample adaptation coefficient evaluation value and the sample scheduling result annotation; the sample adaptation coefficient evaluation value is the adaptation coefficient between the task feature instance and the capacity feature instance, the task feature instance is obtained by extracting the task features of the transportation task instance according to the initial task model, and the capacity feature instance is obtained by extracting the capacity features of the capacity attribute data instance according to the initial capacity model; the sample scheduling result annotation indicates that the scheduling result of the transportation task instance is dispatched to the capacity instance; The acquiring of multiple characteristics of transport capacity to be selected includes: The capacity feature extraction network is used to extract the capacity attribute data of each candidate capacity to obtain a plurality of candidate capacity features; the capacity feature extraction network is used to obtain, train and generate the initial capacity model based on the error between the sample adaptation coefficient evaluation value and the sample scheduling result label.

3. The method according to claim 2, characterized in that The method further comprises: Obtaining the transport task instance and the transport attribute data instance corresponding to the transport capacity instance, wherein the transport task instance is configured with the sample scheduling result annotation; Extracting task features from the transportation task instance according to the initial task model to obtain the task feature instance; Extracting capacity features from the capacity attribute data instance according to the initial capacity model to obtain the capacity feature instance; Determining a fit coefficient between the task feature instance and the capacity feature instance, and obtaining an evaluation value of the sample fit coefficient; The initial task model and the initial capacity model are trained according to the error between the sample adaptation coefficient evaluation value and the sample scheduling result label to obtain the task feature extraction network and the capacity feature extraction network.

4. The method according to claim 3, characterized in that The transport task instances include efficient transport task instances that are successfully dispatched to efficient transport capacity instances, and inefficient transport task instances that fail to be dispatched to inefficient transport capacity instances. The sample scheduling result annotations include a first score annotation and a second score annotation, where the first score is greater than the second score. The efficient transport task instances are configured with the first score annotation, and the inefficient transport task instances are configured with the second score annotation. The determining the adaptation coefficient between the task feature instance and the capacity feature instance to obtain the sample adaptation coefficient evaluation value includes: Determine an adaptation coefficient between an efficient task feature instance and an efficient transport capacity feature instance corresponding to the efficient transport capacity instance, and obtain a first sample adaptation coefficient evaluation value; the efficient task feature instance is a task feature corresponding to the efficient transport task instance; Determining a fit coefficient between an inefficient task feature instance and an inefficient transport capacity feature instance corresponding to the inefficient transport capacity instance, to obtain a second sample fit coefficient evaluation value; the inefficient task feature instance is a task feature corresponding to the inefficient transport task instance; The training of the initial task model and the initial capacity model according to the error between the sample adaptation coefficient evaluation value and the sample scheduling result label to obtain the task feature extraction network and the capacity feature extraction network includes: Determining a first scheduling error value based on a first error between the first sample adaptation coefficient evaluation value and the first scoring label; determining a second scheduling error value according to a second error between the second sample adaptation coefficient evaluation value and the second scoring label; determining a target scheduling error value according to the first scheduling error value and the second scheduling error value; Update the network parameters corresponding to the initial task model and the initial capacity model according to the target scheduling error value until the training termination state is met; The initial task model upon completion of training is determined as the task feature extraction network, and the initial capacity model upon completion of training is determined as the capacity feature extraction network.

5. The method according to claim 3, characterized in that The initial task model includes an initial transport index feature extraction layer, an initial waybill description feature extraction layer, an initial feature alignment layer, and an initial attention unit. The task feature extraction of the transport task instance according to the initial task model to obtain the task feature instance includes: Extracting transport index features from the transport task instance according to the initial transport index feature extraction layer to obtain sample transport index features; Extracting waybill description features from the transport task instance according to the initial waybill description feature extraction layer to obtain sample waybill description features; According to the initial feature alignment layer, the sample transport index features are mapped to a preset feature dimension through linear transformation to generate aligned transport index features; global feature extraction is performed on the sample waybill description features to generate aligned waybill description features consistent with the transport index feature dimension; According to the initial attention unit, the aligned transport indicator features are used as query vectors, and the aligned waybill description features are used as key vectors and value vectors; the matching score between the query vector and the key vector is calculated, and the score is scaled in accordance with the feature dimension; the scaled matching score is converted into an attention weight distribution through a normalization function; the value vector is dynamically weighted fused according to the attention weight distribution to generate a context vector reflecting the key features of the task; the context vector and the aligned transport indicator features are cross-modally superimposed to obtain the task feature instance that integrates the task requirements and capacity characteristics.

6. The method according to claim 2, characterized in that The method further comprises: Obtaining capacity attribute information and capacity interaction information corresponding to multiple initial capacities; Determining the transport capacity attribute information and transport capacity interaction information corresponding to each initial transport capacity as the transport capacity attribute data corresponding to each initial transport capacity; Extracting capacity characteristics from the capacity attribute data of each initial capacity according to the capacity characteristic extraction network to obtain initial capacity characteristics corresponding to each initial capacity; Building a capacity database according to the initial capacity characteristics corresponding to each initial capacity; The acquiring of multiple characteristics of transport capacity to be selected includes: According to the transport capacity database, the plurality of transport capacity characteristics to be selected are obtained.

7. The method according to claim 6, characterized in that The method further comprises: Acquire the transport attribute data corresponding to each initial transport capacity in the transport capacity database according to a preset time period; Determine the initial transport capacity with changed transport capacity attribute data as the transport capacity to be optimized; Extracting capacity characteristics from the optimized capacity attribute data of the capacity to be optimized according to the capacity characteristic extraction network to obtain optimized capacity characteristics of the capacity to be optimized; The transport capacity database is optimized according to the optimized transport capacity characteristics of the transport capacity to be optimized.

8. The method according to claim 6, characterized in that There are multiple transport capacities to be selected, and after extracting the task features of the bulk cargo transportation task based on the task feature extraction network and obtaining the features of the task to be scheduled, the method further includes: Obtaining a candidate adaptation coefficient corresponding to each candidate transport capacity characteristic according to the adaptation coefficient between the characteristics of the task to be scheduled and the characteristics of each candidate transport capacity; Matching a plurality of basic model training capacities from the plurality of candidate capacities according to the candidate adaptation coefficients corresponding to the characteristics of the candidate capacities; Dispatching the bulk cargo transportation task to the plurality of basic model training transport capabilities; Determine, according to the scheduling results corresponding to the multiple basic model training capacities, the basic model training scheduling result label corresponding to each basic model training capacity; Extracting capacity features from each basic model training capacity according to the capacity feature extraction network to obtain basic model training capacity features of each basic model training capacity; Determine the basic model training adaptation coefficient evaluation value corresponding to each basic model training capacity according to the adaptation coefficient between the characteristics of the task to be scheduled and the characteristics of each basic model training capacity; According to the error between the basic model training adaptation coefficient evaluation value corresponding to each basic model training capacity and the basic model training scheduling result labeling corresponding to each basic model training capacity, the task feature extraction network and the capacity feature extraction network are subjected to transfer learning training to obtain an optimized task feature extraction network and an optimized capacity feature extraction network.

9. The method according to claim 8, characterized in that The method further comprises: Extracting task features of the bulk cargo transportation task according to the optimized task feature extraction network to obtain optimized task features to be scheduled; The method further comprises: Extracting capacity characteristics from the capacity attribute data corresponding to the preferred capacity according to the optimized capacity characteristic extraction network to obtain target preferred capacity characteristics corresponding to the preferred capacity; Optimizing the preferred transport capacity characteristics according to the adaptation coefficient between the characteristics of the task to be scheduled and the preferred transport capacity characteristics to obtain the optimized preferred transport capacity characteristics includes: According to the adaptation coefficient between the optimized task characteristics to be scheduled and the target preferred capacity characteristics, the preferred capacity characteristics are optimized to obtain the optimized preferred capacity characteristics.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images