Continuous learning waybill matching method based on rules and experience

Abstract

The invention provides a continuous learning waybill matching method based on rules and experience, and aims to solve the problem that manual waybill scheduling is complex, time-consuming and labor-consuming in the marine transportation process. According to the method, the three aspects of saving the empty container running mileage proportion, the remaining time interval and the route included angle are considered from the rule level, comprehensive sorting is carried out according to the three aspects, a large language model is established in the aspect of experience, the waybill is vectorized, experience similar to historical orders is obtained, and then sorting is carried out according to the difference value. And comprehensive sorting is carried out by integrating the two aspects of rules and experience, and waybill matching is carried out.

Description

Technical Field

[0001] The present invention relates to the fields of logistics and e-commerce, and in particular to a waybill matching method based on sustainable learning from two perspectives: rules and experience. Background Art

[0002] With the development of international trade, the transportation volume of import and export goods has increased dramatically. As the most cost-effective international transportation method, sea transportation has a large number of containers flowing in and out of the terminal every day. Imported containers will be transported from the terminal to the corresponding customer's door by truck for unpacking and unloading, and then the empty containers will be sent back to the terminal; while for exports, an empty container will be taken from the terminal to the corresponding customer's door, and the goods will be packed and sent to the terminal. Whether it is the import process or the export process, trucks are used as the main container transportation tool, and there will always be empty containers in the process. Studies have pointed out that trucks, as high-energy-consuming transportation tools, account for nearly 65% ​​of the country's average annual gasoline consumption. Therefore, through a reasonable algorithm, reducing the mileage of empty containers during container transportation can save fuel consumption in terms of economic benefits, and promote carbon emission reduction in the transportation industry in terms of social benefits.

[0003] The transport company uses manual dispatching to select import and export waybills from the currently unprocessed transport waybills and match them. In this way, after the import waybills are unpacked and unloaded at the customer's door, there is no need to transport the empty containers back to the terminal. Instead, they are directly transported to the customer's door of the export waybill to load the goods, and then return to the terminal with the goods.

[0004] This manual dispatching method requires a lot of work to select the appropriate waybills from a large number of waybills. If each waybill is compared, there will be m*n combinations for m import waybills and n export waybills. In fact, dispatchers do not perform such a large amount of calculation comparison, but perform matching operations based on past experience.

[0005] However, if the system is used for matching, the rules are too rigid. In actual applications, users may be dissatisfied with the recommendation results if they rely solely on preset rules for matching. This is because there are often complex conditions in actual scenarios that are difficult to clearly define and extract in advance. For example, in scenarios involving export goods, if the goods are of food type, the containers used for loading should avoid being containers that have previously loaded toxic chemicals (especially during the import process). Such implicit and specific applicability criteria are difficult to be fully covered by simple rules. Summary of the invention

[0006] The purpose of the present invention is to solve the shortcomings in the prior art.

[0007] To achieve the above object, the present invention adopts the following technical solutions: S1. Analyze the import and export waybill groups in the historical data and formulate rules. The steps are as follows: S11. Calculate the data distribution of three dimensions: the proportion of empty box running mileage saved, the remaining time interval, and the route angle; S12. Calculate the thresholds of three dimensions: the ratio of empty box running mileage saved, the remaining time interval, and the route angle; S13. Filter out the import and export orders that do not meet the threshold, and retain the import and export orders that meet the threshold; S2. Use RAG technology to create a large language model experience knowledge base based on the Chinese large language model. The steps are as follows: S21. Use the embedding capability of the large speech model to vectorize the text information in the waybill data; S22. Combining the numerical dimension and the vectorized information of the text, a vector of the waybill is generated; S23. Convert the waybills in the import and export waybill group in the historical data into vectors, and save the mapping relationship between the import vector and the export vector as a piece of knowledge in the large language model experience knowledge base; S3. Use contrastive learning technology to create an experience knowledge base based on industry information to make up for the deficiency that the large language model only has generalized understanding ability and is not sensitive to industry information knowledge. The steps are as follows: S31. Use historical waybill group data to train the model; S32. Using the trained model, convert the import and export waybill group information into a comparison vector; S33. Calculate the difference between the import and export waybill and the experience; S34. Sort the difference values, S4. Matching of waybills based on rules and experience knowledge base, the steps are as follows: S41. For waybills that meet the rules, which refers to waybills that can be matched with the waybills to be matched, the matching order is sorted by route angle in ascending order, empty container mileage ratio in descending order, and remaining time interval in ascending order; S42. Find the waybill that matches the waybill to be matched from the large language model experience knowledge base as the matching result; S43. For each matching result in step S42, calculate the similarity with the knowledge in the experience knowledge base of the industry information; and use the similarity with the experience knowledge of the industry information and the similarity with the large language model knowledge to sort them; S44. Based on the results of step S41 and step S43, the waybills that meet both the matching rules and experience are prioritized.

[0008] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention recommends matching solutions based on rules and experience, which complements and improves accuracy. It is more in line with the working habits of dispatchers and greatly improves the adoption rate of solutions.

[0009] 2. The rules and experience used in the present invention are automatically adjusted according to the actual waybill matching plan and have continuous learning capabilities.

[0010] 3. The empirical scheme of the present invention uses both a large language model and a model trained by contrastive learning. The large language model can achieve a generalized understanding of Chinese semantics; contrastive learning can achieve an accurate match with business waybill data.

[0011] The large language model is used to compare the similarity of the waybill itself; contrastive learning is used to judge the similarity of the entire import and export waybill pair. Through two rounds of comparison and filtering, the query efficiency can be improved.

[0012] 4. In the experience matching process of the present invention, the vector fusion solution of the numerical dimension and the text dimension solves the problem of inaccurate judgment of numerical similarity of large language models.

[0013] 5. In the experience matching process of the present invention, the inverse geographic transformation and time processing in the comparative learning process solve the problem of insufficient information when using comparative learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a system module diagram of the present invention; Figure 2 is a learning process diagram of the rule module of the present invention; Figure 3 This is a diagram of the matching process of the rule module of the present invention; Figure 4 A learning process diagram of the experience module of the present invention; Figure 5 It is a diagram of the matching process of the experience module of the present invention; DETAILED DESCRIPTION

[0015] The present invention provides a waybill matching method based on sustainable learning from two perspectives: rules and experience. The specific steps are as follows: S100. Study of historical waybill group: S101. Interface input parameters: - Import waybill number; - Import container model; - description of the imported goods; - Latitude and longitude of the import port; - Import customer number; - Import customer door point longitude and latitude; - Arrival time required by import customers; -Export waybill number; - Export container model; - Description of the exported goods; - Latitude and longitude of export port; - Export customer number; - Latitude and longitude of export customer gate; - Export customers require arrival time.

[0016] S102.Interface return value: None.

[0017] S103. Process description: The facade module passes the request to the historical waybill group learning module, and then passes the waybill group information to the rule module and the experience module respectively.

[0018] The rule module and experience module will process and calculate according to their respective requirements, and the results will be used as the basis for subsequent matching processing. The results will be saved in the database.

[0019] S200. Synchronization of waybills to be loaded: S201. Interface input parameters: - Waybill number; - Waybill type (import waybill / export waybill); -Container model; - description of the goods; -Port latitude and longitude; - Customer number; - Customer door point longitude and latitude; -Customer requested arrival time.

[0020] S202. Interface return value: None.

[0021] S203. Process description: The facade module saves the waybill data to be loaded into the database for use by the rule module and experience module in waybill matching.

[0022] S204. Waybill Matching: a. Interface input parameters: - Import waybill number.

[0023] b. Interface return value: The matching interface returns information (matching results). A list sorted by matching priority. The structure of each matching result is as follows: Table 1. List of matching priorities No Field Name illustrate 1 seq The matching sequence number starts from 1 2 order_no Waybill number of the matched export waybill 3 type 1. Matching based on rules only; 2. Matching based on experience only; 3. Matching based on rules + experience 4 ref_orders When matching is based on historical experience, the referenced historical waybills will be returned in the structure of "import waybill number, export waybill number". c. Process description: The facade module passes the request to the waybill matching module, and then passes the import waybill number to the rule module and the experience module respectively.

[0024] The rule module and the experience module will match and process according to their respective logics. The waybill matching module integrates the export waybill results matched by the two logics and sorts them according to the comprehensive score.

[0025] The specific steps of the rule module are as follows: 1. The calculation steps for the waybill group information are as follows: a. According to the earth plane distance calculation formula, use the "import port longitude and latitude" and "import customer door longitude and latitude" to calculate the distance between the import waybill's port and the customer door. Recorded as distance d1, the unit is kilometers.

[0026] The formula is as follows: d1=R*arccos(cos(latitude A)cos(latitude B)cos(longitude A-longitude B)+sin(latitude A)sin(latitude B)

[0027] A represents the longitude and latitude of the import port, and B represents the longitude and latitude of the import customer's door point. There is no order of precedence between the two. R is the radius of the earth, which is 6371, and the unit is kilometers.

[0028] b. According to the earth plane distance calculation formula, use the "export port longitude and latitude" and "export customer door longitude and latitude" to calculate the distance between the export waybill's port and the customer door. Record it as distance d2, in kilometers.

[0029] d2=R*arccos(cos(latitude A)cos(latitude B)cos(longitude A-longitude B)+sin(latitude A)sin(latitude B) A represents the longitude and latitude of the export port, and B represents the longitude and latitude of the export customer's door point. There is no order of precedence between the two.

[0030] c. According to the earth plane distance calculation formula, use the "import customer door point longitude and latitude" and "export customer door point longitude and latitude" to calculate the distance between the import and export customer door points. Recorded as distance d3, the unit is kilometers.

[0031] d3=R*arccos(cos(latitude A)cos(latitude B)cos(longitude A-longitude B)+sin(latitude A)sin(latitude B) A represents the longitude and latitude of the import customer's door point, and B represents the longitude and latitude of the export customer's door point.

[0032] d. The ratio of empty-tank mileage saved = (d1+d2-d3) / (d1+d2).

[0033] e. Use "import customer door latitude and longitude" - "import port latitude and longitude" as the import waybill vector, recorded as: , use "export customer door point longitude and latitude" - "export port longitude and latitude" as the export waybill vector, recorded as: , use the following formula to calculate the angle:

[0034] f. The remaining time interval refers to the effective interval between the "arrival time required by the export customer" and the "arrival time required by the import customer". The effective time interval here means that after unloading the goods at the import customer's door for 1 hour, the goods can arrive at the export customer's door at a speed of 50 kilometers per hour. For situations that do not fall within the effective time interval, a penalty value needs to be added. The calculation formula for the remaining time interval is as follows, where timestamp1 and timestamp2 are the arrival times required by the import and export customers respectively:

[0035]

[0036]

[0037] Among them, 96 is the penalty value and rst is the remaining time interval.

[0038] g. Save the calculation results to the database. The table structure is as follows: Table 2. order_pair_for_rule No Field Name type illustrate 1 id bigint(20) Primary Key 2 import_order_no varchar(32) Import waybill number 3 export_order_no varchar(32) Export waybill number 4 saving_empty_ran_rate double Saving the proportion of empty box mileage 5 angle_degree_of_routes double Route Angle 6 time_gap int Remaining time interval 2. The threshold calculation process is as follows: Since the current data is a set of waybills that have been successfully matched in the past, they are all reasonable matching methods. Therefore, it can be guaranteed that: The mileage saved by running with an empty tank must be greater than 0.

[0039] The route angle must be an acute angle.

[0040] The remaining time interval must be greater than 0, and no penalty value is added.

[0041] The steps for calculating the route angle threshold are as follows: a. Get the minimum and maximum values ​​of the route angles in the order_pair_for_rule table.

[0042] Calculation: Group interval = (maximum value - minimum value) / 20.

[0043] b. Traverse each record in order_pair_for_rule and obtain the route angle value.

[0044] Calculation: Group number = floor((Current included angle - Minimum value) / Group interval).

[0045] Statistical record quantity of each group: Increment the count of this group number by 1.

[0046] c. Calculate the threshold of the record quantity: threshold = floor(Record quantity in the order_pair_for_rule table * 0.95).

[0047] d. Traverse in ascending order of the group number to find the target group number n that satisfies the following conditions:

[0048]

[0049] where is the record quantity of the m-th group.

[0050] e. When amount < threshold: Arrange the elements of the (n + 1)-th group in ascending order and take the element at the (threshold - amount) position.

[0051] The value of this element is the threshold.

[0052] f. When amount = threshold, take the maximum value of the elements in the n-th group as the threshold.

[0053] Steps for calculating the threshold of the proportion of empty container running mileage saved are as follows: a. Obtain the minimum value and the maximum value of the proportion of empty container running mileage saved in the order_pair_for_rule table.

[0054] Calculation: Group interval = (Maximum value - Minimum value) / 20.

[0055] b. Traverse each record in order_pair_for_rule to obtain the interval value of the proportion of empty container running mileage saved.

[0056] Calculation: Group number = floor((Current proportion of empty container running mileage saved - Minimum value) / Group interval).

[0057] Statistical record quantity of each group: Increment the count of this group number by 1.

[0058] c. Calculate the threshold of the record quantity: threshold = floor(Record quantity in the order_pair_for_rule table * 0.95).

[0059] d. Traverse in ascending order of the group number to find the target group number n such that n satisfies the following conditions:

[0060] where is the number of records in the m-th group.

[0061] e. When amount < threshold: Arrange the elements of the (n + 1)-th group in ascending order and take the element at the (threshold - amount)-th position.

[0062] The value of this element is the threshold.

[0063] f. When amount = threshold, take the maximum value of the elements in the n-th group as the threshold.

[0064] The steps for calculating the threshold for the remaining time intervals are as follows: a. Obtain the minimum and maximum values of the remaining time intervals in the order_pair_for_rule table.

[0065] Calculate: group interval = (maximum value - minimum value) / 20.

[0066] b. Traverse each record in order_pair_for_rule to obtain the remaining time interval value.

[0067] Calculate: the group number to which it belongs = floor((current remaining time interval - minimum value) / group interval).

[0068] Count the number of records in each group: increment the count for this group number by 1.

[0069] c. Calculate the threshold for the number of records: threshold = floor(0.95 * the number of records in the order_pair_for_rule table).

[0070] d. Traverse in ascending order of the group number to find the target group number n such that n satisfies the following conditions:

[0071]

[0072]

[0073] where is the number of records in the m-th group.

[0074] e. When amount < threshold: Arrange the elements of the (n + 1)-th group in ascending order and take the element at the (threshold - amount)-th position.

[0075] The value of this element is the threshold.

[0076] f. When amount = threshold, take the maximum value of the elements in the nth group as the threshold.

[0077] After calculating the remaining time interval, route angle, and the threshold value of the empty box running mileage ratio, they will be saved in the database. The table structure is as follows: Table 3. rule_threshold No Field Name type illustrate 1 id bigint(20) Primary Key 2 min_saving_empty_ran_rate double Threshold value (minimum value) for saving empty box mileage ratio 3 max_angle_degree_of_routes double Threshold of route angle (maximum value) 4 max_time_gap int The remaining time interval threshold (maximum value) 5 update_datetime timestamp Update time of rule threshold 3. The specific steps of the rule-based waybill matching function are as follows:

[0078] (1) According to the import waybill number, obtain the import container model and the arrival time required by the import customer. Then obtain the export waybill from the waiting loading pool in the database. The obtained export waybill must meet the following conditions: Export waybill container model = import waybill container model; The arrival time required by the customer for export waybills is greater than the arrival time required by the customer for import waybills.

[0079] (2) Traverse each export waybill and calculate its "saving empty container mileage ratio", "remaining time interval" and "route angle" information with the import waybill to be matched. The calculation process is the same as the historical waybill group learning function.

[0080] (3) Obtain the thresholds of the three dimensions of "saving empty box mileage ratio", "remaining time interval" and "route angle" from the rule_threshold table in the database, and filter out the waybills that do not meet the threshold requirements.

[0081] (4) For export waybills that meet the threshold conditions, they are sorted in ascending order based on the route angle, descending order based on the percentage of empty container mileage saved, and ascending order based on the remaining time interval.

[0082] The specific steps of the experience module are as follows:

[0083] 1. The vectorization of large language models is achieved through embedding. Embedding is the process of converting unstructured information (images, texts, etc.) into vectors. The role of embedding is to project real-world discrete data into a high-dimensional data space through deep learning training, and reflect the similarity of the real world through the distance between data in the space.

[0084] The present invention selects the open source Chinese text embedding model m3e-base to perform the Embedding operation.

[0085] 2. The steps for vectorizing import waybills using a large language model are as follows: a. Perform an Embedding operation on the “Description of Imported Goods” to generate a 508-dimensional vector.

[0086] b. Multiply each longitude and latitude of the “Import Port Latitude and Longitude” and “Import Customer Gate Latitude and Longitude” by 2000 to generate 4 float type numbers.

[0087] c. Increase the 508-dimensional vector in step a by 4 dimensions, using the 4 numbers obtained in step b as the dimension values. Generate a new 512-dimensional vector. This 512-dimensional vector is the vector corresponding to the import order.

[0088] 3. The steps for vectorizing export waybills using a large language model are as follows: a. Perform an Embedding operation on the “Export Goods Description” to generate a 508-dimensional vector.

[0089] b. Multiply each longitude and latitude of the "Export Port Latitude and Longitude" and "Export Customer Gate Latitude and Longitude" by 2000 to generate 4 float type numbers.

[0090] c. Increase the 508-dimensional vector in step a by 4 dimensions, using the 4 numbers obtained in step b as the values ​​of the dimensions. Generate a new 512-dimensional vector. This 512-dimensional vector is the vector corresponding to the export order.

[0091] 4. Contrastive learning is a machine learning method that aims to learn useful representations or features by comparing the similarities and differences between samples. Simply put, in multidimensional representation, similar samples are pulled closer and different samples are pulled farther apart, so that the multidimensional representation of the samples is used to represent the feature differences between the sample data.

[0092] In contrastive learning, self-supervised learning is generally the mainstream method, which is trained without counterexamples and is suitable for the scenario of waybill matching.

[0093] Take the import and export waybills as a whole and use the open source algorithm ESimCSE to train a model based on business features. The specific steps are as follows: a. Use the reverse geographic transformation function provided by Baidu Maps to convert the "import port longitude and latitude", "import customer door longitude and latitude", "export port longitude and latitude", and "export customer door longitude and latitude" into specific address text information, and save them in the "province, city, district, street, and house number" format, respectively called "import port address", "import customer door address", "export port address", and "export customer door address".

[0094] b. Classify the "import customer required arrival time" and "export customer required arrival time" into three categories: "morning", "afternoon" and "night" according to the hour information and save them, which are called "import time period" and "export time period" respectively.

[0095] c. Subtract the "import customer required arrival time" from the "export customer required arrival time" to get the time difference in hours.

[0096] d. Use "import port address", "import customer door address", "export port address", "export customer door address", "import time period", "export time period", "import goods description", and "export goods description" as the content, splice them into a sentence as input to train the model.

[0097] 5. The process of comparative learning vectorization is consistent with the model training process in terms of data processing. The calculated "import port address", "import customer door address", "export port address", "export customer door address", "import time period", "export time period" are combined with "import goods description" and "export goods description" to form a sentence. Then this sentence is passed into the trained model to obtain a 1536-dimensional vector.

[0098] 6. Save the import and export waybill group and the generated "import waybill vector", "export waybill vector" and "comparative learning vector" into the database as an experience knowledge base. The table structure is as follows: Table 4. experience_lib No Field Name type illustrate 1 id bigint(20) Primary Key 2 import_order_no varchar(32) Import waybill number 3 export_order_no varchar(32) Export waybill number 4 import_vector vector(512) Import waybill vector 5 export_vector vector(512) Export waybill vector 6 simcse_vector vector(1536) Contrastive learning of waybill group vectors 7 create_datetime timestamp Knowledge creation time 7. The steps of the waybill matching function based on experience are as follows:

[0099] (1) Obtain detailed information of the import waybill according to the import waybill number. The steps to quantize the import waybill are as follows: a. Perform an Embedding operation on the “Description of Imported Goods” to generate a 508-dimensional vector.

[0100] b. Multiply each longitude and latitude of the “Import Port Latitude and Longitude” and “Import Customer Gate Latitude and Longitude” by 2000 to generate 4 float type numbers.

[0101] c. Increase the 508-dimensional vector in step a by 4 dimensions, using the 4 numbers obtained in step b as the dimension values. Generate a new 512-dimensional vector. This 512-dimensional vector is the vector corresponding to the import order.

[0102] (2) Obtain experience similar to that of import waybills. The specific steps are as follows: Using this import order vector, find the experience whose Euclidean distance to the import_vector field is less than 1 from the experience_lib table of the knowledge base. And record the distance as vector_d1 Euclidean distance formula:

[0103] (3) Obtain export waybills similar to those in experience from the database's pool of waiting items. The obtained export waybills must meet the following conditions: Export waybill container model = import waybill container model; The arrival time required by the customer for export waybills is greater than the arrival time required by the customer for import waybills; The Euclidean distance between the vector of the export waybill and the export vector export_vector field in the experience information is <= 1.

[0104] This distance is recorded as vector_d2.

[0105] (4) Traverse each export waybill in step (3) and calculate the difference between the import and export waybill and the experience. The specific steps are as follows: a. Use the reverse geographic transformation function provided by Baidu Maps to convert the "import port longitude and latitude", "import customer door longitude and latitude", "export port longitude and latitude", and "export customer door longitude and latitude" into specific address text information, and save them in the "province, city, district, street, and house number" format, respectively called "import port address", "import customer door address", "export port address", and "export customer door address".

[0106] b. Classify the "import customer required arrival time" and "export customer required arrival time" into three categories: "morning", "afternoon" and "night" according to the hour information and save them, which are called "import time period" and "export time period" respectively.

[0107] c. Subtract the "import customer required arrival time" from the "export customer required arrival time" to get the time difference in hours.

[0108] d. Use "import port address", "import customer door address", "export port address", "export customer door address", "import time period", "export time period", "import goods description", "export goods description" as the content and splice them into a sentence.

[0109] e. Then pass this sentence into the trained model to obtain a 1536-dimensional vector.

[0110] f. Calculate the Euclidean distance between the vector obtained in step e and simcse_vector in the empirical table. This distance is the difference between the import and export waybill and the empirical value. Recorded as vector_d3 (5) Use the vector_d1, vector_d2, and vector_d3 values ​​calculated in steps (2) to (4) and use the geometric mean as the comprehensive difference value.

[0111]

[0112] (6) For all export orders, they are sorted in ascending order according to the comprehensive difference value total_diff. For export orders with the same comprehensive difference value, the experience with more times is given priority, and the latest experience is used as the recommended result for matching.

[0113] The detailed steps of the order matching module are as follows: The order matching functions of the rule module and the experience module are called respectively to perform the matching calculation of the waybill. Then the export waybill results matched by the two logics are integrated and sorted according to the comprehensive evaluation score. The sorting rules are as follows: First, the matching results that have both rules and experience are recommended first. Among these results, the order of export waybills in the rule module is used for recommendation.

[0114] Secondly, recommendations only exist in the matching results of the experience module. Recommendations are made in the order of experience matching.

[0115] Finally, recommendations only exist in the matching results of the rule module. Recommendations are made in the order in which the rules are matched.

[0116] The top 10 of the sorted export waybills are returned as the final result.

[0117] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A waybill matching method based on sustainable learning from both rules and experience perspectives. Features: S1. Analyze the import and export waybill groups in the historical data and formulate rules. The steps are as follows: S11. Calculate the data distribution of three dimensions: the proportion of empty box running mileage saved, the remaining time interval, and the route angle; S12. Calculate the thresholds of three dimensions: the ratio of empty box running mileage saved, the remaining time interval, and the route angle; S13. Filter out the import and export orders that do not meet the threshold, and retain the import and export orders that meet the threshold; S2. Use RAG technology to create a large language model experience knowledge base based on the Chinese large language model. The steps are as follows: S21. Use the embedding capability of the large speech model to vectorize the text information in the waybill data; S22. Combining the numerical dimension and the vectorized information of the text, a vector of the waybill is generated; S23. Convert the waybills in the import and export waybill group in the historical data into vectors, and save the mapping relationship between the import vector and the export vector as a piece of knowledge in the large language model experience knowledge base; S3. Use contrastive learning technology to create an experience knowledge base based on industry information to make up for the deficiency that the large language model only has generalized understanding ability and is not sensitive to industry information knowledge. The steps are as follows: S31. Use historical waybill group data to train the model; S32. Using the trained model, convert the import and export waybill group information into a comparison vector; S33. Calculate the difference between the import and export waybill and the experience; S34. Sort the difference values, S4. Matching of waybills based on rules and experience knowledge base, the steps are as follows: S41. For waybills that meet the rules, which refers to waybills that are paired with the waybills to be matched, they are sorted by the rules according to the ascending order of route angle, the descending order of the proportion of empty container mileage saved, and the ascending order of the remaining time interval; S42. Find the waybill that matches the waybill to be matched from the large language model experience knowledge base as the matching result; S43. For each matching result in step S42, calculate the similarity with the knowledge in the experience knowledge base of the industry information; and use the similarity with the experience knowledge of the industry information and the similarity with the large language model knowledge to sort them; S44. Based on the results of step S41 and step S43, the waybills that meet both the matching rules and experience are prioritized.

2. A waybill matching method based on continuous learning from both rule and experience perspectives as claimed in claim 1, Features: The steps for calculating the ratio of empty box mileage saved in S1 are as follows: A1. According to the earth plane distance calculation formula, use the "Import port longitude and latitude" and "Import customer door longitude and latitude" to calculate the distance between the port of the import waybill and the customer door; record it as distance d1, in kilometers; The formula is as follows: , A represents the longitude and latitude of the import port, B represents the longitude and latitude of the import customer's door point, and there is no order between the two. R is the radius of the earth, which is 6371, in kilometers. B1. According to the Earth's plane distance calculation formula, use the "Export port longitude and latitude" and "Export customer door longitude and latitude" to calculate the distance between the export waybill's port and the customer door; record it as distance d2, in kilometers; ， A represents the latitude and longitude of the export port, and B represents the latitude and longitude of the export customer's door point. There is no order between the two. C1. According to the Earth's plane distance calculation formula, use the "Import Customer Gate Longitude and Latitude" and "Export Customer Gate Longitude and Latitude" to calculate the distance between the import and export customer gates; record it as distance d3, in kilometers; ， A represents the longitude and latitude of the import customer's door point, and B represents the longitude and latitude of the export customer's door point. There is no order between the two. D1. The proportion of empty-tank mileage saved = (d1+d2-d3) / (d1+d2).

3. A waybill matching method based on continuous learning from both rule and experience perspectives as claimed in claim 1, Features: The remaining time interval in S1 is calculated as follows: It refers to the effective interval between the "arrival time required by export customers" and the "arrival time required by import customers". The effective time here means that the goods can arrive at the export customer's door in time at a speed of 50 kilometers per hour one hour after unloading at the import customer's door. For situations that do not belong to the effective time interval, a penalty value needs to be added. The calculation formula for the remaining time interval is as follows, where timestamp1 and timestamp2 are the arrival times required by import and export customers respectively. ， Among them, 96 is the penalty value and rst is the remaining time interval.

4. A waybill matching method based on continuous learning from both rule and experience perspectives as claimed in claim 1, Features: The steps of calculating the route angle in S1 are as follows: Use "Import customer door point longitude and latitude" - "Import port longitude and latitude" as the import waybill vector, recorded as: , use "export customer door point longitude and latitude" - "export port longitude and latitude" as the export waybill vector, recorded as: , the angle is calculated as follows: 。 5. A waybill matching method based on continuous learning from both rule and experience perspectives as claimed in claim 1, Features: In S2, the embedding capability of the large speech model is used to vectorize the text information in the waybill data. The steps for vectorizing the import waybill are as follows: A2. Perform an Embedding operation on "Imported Goods Description" to generate a 508-dimensional vector; B2. Multiply each longitude and latitude of "Import Port Latitude and Longitude" and "Import Customer Gate Latitude and Longitude" by 2000 to generate 4 float type numbers; C2. Increase the 508-dimensional vector in step A2 by 4 dimensions, using the 4 numbers obtained in step B2 as the values ​​of the dimensions respectively; generate a new 512-dimensional vector; this 512-dimensional vector is the vector corresponding to the import order; The steps to quantize export waybills are as follows: A3. Perform an Embedding operation on "Export Goods Description" to generate a 508-dimensional vector; B3. Multiply each longitude and latitude of "Export port longitude and latitude" and "Export customer gate longitude and latitude" by 2000 to generate 4 float type numbers; C3. Increase 4 dimensions of the 508-dimensional vector in step A3, using the 4 numbers obtained in step B3 as the values ​​of the dimensions respectively; generate a new 512-dimensional vector; this 512-dimensional vector is the vector corresponding to the export order.

6. A waybill matching method based on continuous learning from both rule and experience perspectives as claimed in claim 1, Features: The calculation steps for establishing the comparison vector in S3 are as follows: A4. Use the reverse geographic transformation function provided by Baidu Maps to convert "Import port longitude and latitude", "Import customer door longitude and latitude", "Export port longitude and latitude", "Export customer door longitude and latitude" into specific address text information, and save them in the format of "Province, City, District, Street, House Number", and they are called "Import port address", "Import customer door address", "Export port address", "Export customer door address"; B4. Classify the "import customer required arrival time" and "export customer required arrival time" into "morning", "afternoon" and "night" according to the hour information and save them in three categories, namely "import time" and "export time"; C4. Subtract "Import Customer Requested Arrival Time" from "Export Customer Requested Arrival Time" to get the time difference in hours; D4. Take "import port address", "import customer door address", "export port address", "export customer door address", "import time", "export time", "import goods description", "export goods description" as the content, splice them into a sentence as input to form a contrast vector for training the model.

7. A waybill matching method based on continuous learning from both rule and experience perspectives as claimed in claim 1, Features: The steps for calculating the difference value in S3 are as follows: A5. Use the reverse geographic transformation function provided by Baidu Maps to convert "Import port longitude and latitude", "Import customer door longitude and latitude", "Export port longitude and latitude", "Export customer door longitude and latitude" into specific address text information, and save them in the format of "Province, City, District, Street, House Number", and they are called "Import port address", "Import customer door address", "Export port address", "Export customer door address"; B5. Classify the "import customer required arrival time" and "export customer required arrival time" into "morning", "afternoon" and "night" according to the hour information and save them in three categories, namely "import time" and "export time"; C5. Subtract "Import Customer Requested Arrival Time" from "Export Customer Requested Arrival Time" to get the time difference in hours; D5. Use "import port address", "import customer door address", "export port address", "export customer door address", "import time", "export time", "import goods description", "export goods description" as the content, and stitch them into a sentence; E5. Then pass this sentence into the trained model to obtain a 1536-dimensional vector; F5. Calculate the Euclidean distance between the vector obtained in step E5 and simcse_vector in the experience table; this distance is the difference between the import and export waybill and the experience, recorded as vector_d3.

Images