Dangerous article inspection item data prediction method and system, electronic equipment and program product

Abstract

The invention relates to the technical field of dangerous goods transportation, and provides a dangerous goods inspection item data prediction method and system, electronic equipment and a program product. The method comprises the following steps: extracting a plurality of business fields from multi-source business data; based on the plurality of service fields, obtaining embedded vectors corresponding to the plurality of service fields respectively; the multi-source business data comprises at least one of a dangerous goods waybill, a dangerous goods declaration bill and a dangerous goods transportation rule; according to the embedded vectors corresponding to the plurality of business fields respectively, obtaining sample data of the state; training the deep reinforcement learning network at least according to the sample data of the state to obtain a trained deep reinforcement learning network; and in response to a received inspection request for the to-be-inspected dangerous goods, through the trained deep reinforcement learning network, predicting an inspection strategy corresponding to the inspection request, the inspection strategy being used for recommending an inspection sequence of the plurality of inspection items to the user. According to the scheme, the checking efficiency of the dangerous goods waybill can be improved.

Description

Technical Field

[0001] This invention relates to the field of dangerous goods transportation technology, and in particular to a method and system for predicting dangerous goods inspection data, electronic equipment, and program products. Background Technology

[0002] Dangerous goods (or hazardous materials, or hazmat for short) are substances or articles whose chemical, physical, or biological properties, during production, transportation, storage, and use, may pose a significant hazard to health, safety, property, or the environment. The transportation of dangerous goods involves various modes of transport, including air, sea, and land, and its safety management relies on rigorous inspection procedures and compliant declarations. Traditional methods for inspecting dangerous goods often depend on manual inspection experience, making it difficult to cover complex risk combinations, or the inspection standards are static, relying on fixed checklists and unable to dynamically adapt to different transportation scenarios. Summary of the Invention

[0003] This application provides a method, system, electronic device, and program product for predicting data on dangerous goods inspection items, in order to solve the problem of relying on manual inspection experience or static inspection standards.

[0004] In a first aspect, embodiments of this application provide a method for predicting dangerous goods inspection item data, comprising: extracting multiple business fields from multi-source business data; obtaining embedding vectors corresponding to the multiple business fields respectively; the multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules; obtaining sample data of states according to the embedding vectors corresponding to the multiple business fields respectively; training a deep reinforcement learning network based on at least the sample data of states to obtain a trained deep reinforcement learning network; and, in response to receiving an inspection request for dangerous goods to be inspected, predicting an inspection strategy corresponding to the inspection request through the trained deep reinforcement learning network, wherein the inspection strategy is used to recommend the inspection order of multiple inspection items to the user.

[0005] In some embodiments, extracting multiple business fields from multi-source business data includes: obtaining multiple business data tables from multi-source business data stored in multiple business systems corresponding to dangerous goods transportation; each of the multiple business data tables includes multiple business fields.

[0006] In some embodiments, based on multiple business fields, the embedding vectors corresponding to each business field are obtained, including: performing word segmentation on the data of multiple business fields to obtain a word sequence; and inputting the word sequence into the BERT model to obtain the embedding vector corresponding to each word in the word sequence.

[0007] In some embodiments, the deep reinforcement learning network includes a main network and a target network; training the deep reinforcement learning network, at least based on sample data of the state, includes: initializing trainable parameters in the main network and the target network, and obtaining an initial state based on the sample data of the state; initializing a set of inspection items; initializing an experience replay buffer; determining a reward function; and continuously executing a training process of interactive loops: the agent selects an inspection item from the set of inspection items according to an ε-greedy policy as the inspection object of the action in the t-th time step, and performs the action in the environment; obtaining the inspection result corresponding to the executed action; the environment calculates the reward score corresponding to the action according to the reward function; and so on. The status information of the check item corresponding to the action is updated to "checked," and the updated status information and check result are written into the status to obtain the next status. Experience data from multiple time steps is stored in the experience replay buffer. The experience data includes status, action, reward score, and the next status. The training process of the learning loop is triggered periodically: a batch is randomly sampled from the multiple experience data stored in the experience replay buffer; the batch of experience data is input into the main network, and the main network predicts the first Q-value of the action; the target network predicts the second Q-value of the action; the loss function is calculated based on the first and second Q-values; and the trainable parameters in the main network are updated based on the loss function.

[0008] In some embodiments, determining the reward function includes: determining reward scores corresponding to different inspection results, and determining additional reward parameters; wherein, the inspection results include one or more of triggering major risk, high-risk violation, medium-risk violation, low-risk violation, inspection passed, and repeated invalid inspection; the reward scores corresponding to triggering major risk, high-risk violation, medium-risk violation, and low-risk violation decrease sequentially; the reward parameters include the reward scores corresponding to the discount factor, learning rate, time cost, and completion reward; wherein, the reward score for completion reward is a positive value, and the reward score corresponding to the time cost of each training step is a negative value; the discount factor is a decimal between 0 and 1; the discount factor and learning rate are both hyperparameters.

[0009] In some embodiments, after updating the trainable parameters in the main network, the method further includes: periodically synchronizing the values ​​of the trainable parameters in the main network to the target network.

[0010] In some embodiments, selecting the next check item according to an ε-greedy policy includes: randomly selecting an unchecked check item from the check item set with probability ε; or selecting the unchecked check item with the largest Q value from the check item set with probability 1-ε; where the Q value is predicted by the agent.

[0011] Secondly, embodiments of this application also provide a dangerous goods inspection item data prediction system, which includes: a feature extraction module, used to extract multiple business fields from multi-source business data; and to obtain embedding vectors corresponding to the multiple business fields respectively; the multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules; a state acquisition module, used to obtain state sample data according to the embedding vectors corresponding to the multiple business fields respectively; a training module, used to train a deep reinforcement learning network according to the state sample data and the sample data of the inspection item set, to obtain a trained deep reinforcement learning network; and a prediction module, used to predict the inspection strategy corresponding to the inspection request in response to receiving an inspection request, through the trained deep reinforcement learning network, wherein the inspection strategy is used to recommend the order of multiple inspection items to the user.

[0012] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the hazardous materials inspection item data prediction method described in the first aspect.

[0013] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the hazardous materials inspection item data prediction method described in the first aspect.

[0014] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the hazardous materials inspection item data prediction method described in the first aspect.

[0015] The dangerous goods inspection item data prediction method, system, electronic device, and program product provided in this application extract multiple business fields from multi-source business data such as dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules. These business fields are then vectorized to obtain embedding vectors, thereby acquiring sample data of the inspection item set and sample data of the state. Based on the sample data of the state and the sample data of the inspection item set, a deep reinforcement learning network is trained. When a new inspection request is received, the trained deep reinforcement learning network can predict the inspection order (i.e., inspection strategy) of the inspection items in the current dangerous goods inspection task. In other words, the deep reinforcement learning network is used for prediction. The priority of inspection items is determined by checking higher-priority items (those ranked higher in the inspection strategy). This is the optimal inspection strategy learned by the model from a large number of samples of multi-source business data. Users can improve inspection efficiency and achieve cost reduction and efficiency improvement by checking items in the order predicted by the model. In addition, the model (i.e., deep reinforcement learning network) can intelligently and automatically predict outputs without relying on human experience. On the other hand, it can make predictions in real time according to different inspection tasks. Different tasks can output different inspection strategies, so it can be applied to different dangerous goods inspection tasks in different transportation scenarios. The inspection standards are no longer static and are freed from the constraints of fixed checklists. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the learning architecture proposed in this application embodiment for applying reinforcement learning to dangerous goods inspection scenarios; Figure 2 This is one of the flowcharts illustrating the method for predicting dangerous goods inspection items provided in the embodiments of this application; Figure 3 This is a second schematic flowchart of the method for predicting dangerous goods inspection items provided in the embodiments of this application; Figure 4 This is a flowchart illustrating the extraction and embedding process of business fields from multi-source business data in some embodiments of the dangerous goods inspection item data prediction method provided in this application. Figure 5 This is a schematic diagram illustrating the use of the BERT model to obtain embedding vectors in some embodiments of the dangerous goods inspection item data prediction method provided in this application. Figure 6This is a schematic diagram of the process of training DQN in some embodiments of the dangerous goods inspection item data prediction method provided in this application; Figure 7 This is a schematic diagram of the software system architecture of the hazardous materials inspection item data prediction system provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] Currently, with the rapid expansion of dangerous goods transportation and the demand for dangerous goods transportation from airports and transportation hubs, the method currently used for cargo acceptance and verification at most airport cargo terminals or other transportation hubs is the experience-based decision-making method. This method is highly dependent on the experience level and professional knowledge of airport cargo terminal personnel.

[0020] For example, in hazardous materials collection and inspection scenarios, most rely on human experience. By consulting specified rules in paper / electronic books and making human decisions on the order and priority of each inspection item, the following drawbacks may exist: The inspection standards are static: they rely on fixed checklists and cannot be dynamically adapted to different transportation scenarios (such as different flights, different shippers, sales agents, and different types of dangerous goods).

[0021] High risk of missed detection: Manual inspection relies on experience and is difficult to cover complex risk combinations.

[0022] Low compliance efficiency: Collection and transportation personnel often fill in the same inspection items repeatedly or fill in unrelated inspection items on the collection and transportation verification form, which reduces the efficiency of collection and transportation inspection and causes the risk of transportation delays.

[0023] While related technologies have proposed improved solutions, some technical problems still exist, such as the following solutions: Employs a rule engine: an inspection system based on IF-THEN rules, for example, if a substance has a flash point below 60°C, then it must be classified as a Class 3 - flammable liquid, etc.

[0024] In practical applications, this detection system lacks adaptability. For example, there are nine categories of dangerous goods, totaling over 3,000 types, each with different inspection items. The IF-THEN-based approach cannot fully define the acceptance and inspection rules for each category of goods, limiting its application scope and making it unable to solve more complex dangerous goods inspection problems. Furthermore, the IF-THEN-based reasoning method is slow and inefficient, and it cannot incorporate the analysis results of large amounts of historical data, making it difficult to achieve rapid matching of dangerous goods rules.

[0025] Optical Character Recognition (OCR) technology is used for recognition: Currently, character recognition technology is only used for document recognition, which only makes it convenient for users to extract text and other character information, but does not further analyze and process the data, nor is it deeply integrated with the receiving and transportation inspection form and the user's operation and decision-making experience.

[0026] Currently, deep reinforcement learning methods have not been applied in the field of civil aviation cargo transportation, nor have they been used for the dynamic adjustment of dangerous goods inspection strategies.

[0027] In view of this, this application proposes a method and system for predicting dangerous goods inspection items, as well as electronic devices and program products. It proposes a dangerous goods intelligent acceptance and verification form generation model based on deep reinforcement learning. Based on multi-source business data such as paper waybills, electronic waybills, shipper's dangerous goods declaration forms, and domestic and international dangerous goods transportation rules, it systematically applies a deep reinforcement learning recommendation algorithm. By inputting data such as electronic waybills and dangerous goods declarations, it can intelligently recommend the order of inspection items to the user, providing auxiliary decision-making for intelligent dangerous goods acceptance and dynamically generating the items to be verified and their priorities on the acceptance and verification form. Deep reinforcement learning is a method that combines deep learning and reinforcement learning. It aims to use deep neural networks to process complex environments with high-dimensional state and action spaces, fit value functions, and directly control the behavior of the intelligent agent through learning from high-dimensional perceptual inputs.

[0028] To facilitate understanding of the solutions proposed in this application, a brief explanation of reinforcement learning is provided first. As a type of machine learning, reinforcement learning obtains feedback and updates its strategy through continuous interaction with the environment. In this application embodiment, the optimal inspection strategy can be inferred by analyzing the experience gained during the dangerous goods collection and inspection process. During the learning period, the action strategy (i.e., the inspection strategy) is continuously updated based on the reward signal, which evaluates the merits of the action strategy. The environment covered by reinforcement learning is mainly provided by a Markov Decision Process (MDP). The overall architecture of reinforcement learning applied to dangerous goods inspection scenarios is as follows: Figure 1 As shown.

[0029] Reinforcement learning generally consists of four elements: environment, policy, reward, and value function. The policy determines the action an agent should take in a given state; the reward is the feedback signal given by the environment based on the agent's actions; the value function evaluates the value of the action or state; and the environment model describes the transition patterns of the environment's states. These elements together constitute the reinforcement learning framework. Applying reinforcement learning to hazardous materials inspection scenarios, the user (e.g., freight station receiving personnel) is considered part of the external environment. The inspection process may include multiple inspection items. Selecting one from multiple candidate items is considered action 'a'. Executing action 'a' updates the current state 's' based on action 'a', resulting in the next state 's'. The environment calculates a reward score 'r' based on the inspection result and a pre-defined reward function, and then inputs the reward score 'r' and the next state 's' into the system. Figure 1 In the dangerous goods collection and inspection item generation and recommendation agent (hereinafter referred to as agent) shown, the agent predicts the next action based on the reward score (hereinafter referred to as reward) r and the next state s'.

[0030] In related technologies, rewards in reinforcement learning are assessed by evaluating the benefits of each possible action in a fixed state. For example, in the process of hazardous materials acceptance and inspection, the reward in traditional reinforcement learning methods is the time and value gain brought by the user choosing which hazardous materials inspection item to view. However, the long-term benefits and value of an action need to be obtained through multiple iterations of the Bellman equation and converged. The number of iterations is positively correlated with the complexity of the system. For the hazardous materials acceptance and inspection process, the acceptance and inspection forms contain many categories and involve complex hazardous materials knowledge. The embodiments of this application use deep neural networks to fit the value function. Thus, replacing the lengthy Bellman equation iteration with the deep neural network training process becomes a method that can effectively reduce optimization time and improve optimization efficiency.

[0031] Specific learning framework such as Figure 2 As shown. This learning framework may include a dangerous goods collection and inspection item generation agent (hereinafter referred to as agent), a DQN network, and an environment. The DQN network includes an input layer, multiple hidden layers, and an output layer. In the output layer, the Q-values ​​corresponding to each candidate action are obtained, and one action is selected as the recommended action 'a'. The inspection object of action 'a' is the next inspection item, that is, the next inspection item is selected based on the Q-value. The dangerous goods collection and inspection item generation agent applies the recommended action 'a' and / or the generated inspection item to the environment, receives the reward 'r' from the environment, and obtains the updated state 's'. In the embodiments of this application, in the deep reinforcement learning modeling, the dangerous goods collection and inspection item recommendation system is regarded as an agent, and the user (e.g., Figure 2The hazardous materials collection user shown is considered the environment. The interaction between the agent and the environment is the interaction between the recommendation algorithm and the user. The agent can be given a recommendation list (inspection strategy), which includes multiple inspection items and their order that need to be performed for the current inspection task. That is, it satisfies the user's needs by recommending a list of items. This recommendation process follows a sequential decision process and can be modeled as a Markov decision process. By recommending inspection items sequentially over a series of time steps, the cumulative reward and value are maximized.

[0032] Among them, the Markov Decision Process (MDP) can be defined as a quintuple.<S,A,R,P, γ> ,in: S represents the set of states. (States) , This represents the state at the t-th time step. Representing state It includes features corresponding to n inspection items. Each inspection item's features include the embedding vector of the corresponding business field, status information, and inspection result. This can be obtained from the user's collection and inspection operation history, that is, the user has performed n dangerous goods inspection items before time t. The inspection items can be arranged in chronological order.

[0033] A represents the set of actions that an agent can perform. , This represents the action taken by the agent at time t (the t-th time step or the t-th iteration).

[0034] R represents the reward function. Indicates the agent's state Execute action The immediate reward value (i.e., reward points) is earned. After the agent recommends inspection strategies to the user (e.g., a list of items), the user browses these inspection items and provides feedback. Based on the user's feedback, the agent will receive a reward. .

[0035] P represents the transition probability. Define the agent to take action Time state from Transferred to The probability. If the user skips or ignores all hazardous materials inspection items, the status... ; This indicates the initial state. If the user clicks on a specific item to inspect, the next state will be... It will be updated.

[0036] γ represents the discount factor. γ∈[0,1] represents the trade-off between future rewards and current rewards. When γ=0, the agent only considers the current reward, and future rewards have no impact on the current action; when γ=1, all future rewards are fully included in the reward of the current action, and future rewards have the greatest impact on the current action.

[0037] Specifically, in this embodiment, the Q-learning algorithm is used to train the deep reinforcement learning network. First, the agent obtains the current state of the system in the environment and selects an action based on the current state, the selection policy, and the Q-value table corresponding to each candidate check item. Second, the action is applied to the environment, changing it and resulting in a new state and corresponding reward score. Third, the agent updates the internal data of the Q-value table and the selection policy based on the state-action-reward information, facilitating subsequent selections aimed at pursuing higher returns. These three steps are continuously iterated until the selection policy converges iteratively, at which point the agent generates the optimal action for the current environment.

[0038] The Q-value, also known as the value of the Q-function (Q(s, a)), can be understood as the cumulative expected value of all future reward points that can be obtained after choosing and executing action a in state s.

[0039] The goal of training is to learn a policy that maximizes the cumulative expected reward, i.e., maximizes the value Q. The optimal Q function satisfies the Bellman optimality equation:

[0040] in Let represent the optimal value function, r represent the immediate reward (i.e., reward score) that the agent obtains by performing action a in state s, and γ represent the discount factor. This represents the maximum Q-value among all possible actions to take in the next state s'. a' represents the next action, for example, the action at time step t is a'. t The next action is a t+1 s' represents the next state; for example, the action at time step t is s. t The next action is s t+1 .

[0041] For an agent based on the Q-learning algorithm, at each time t, it obtains state information s from the environment. t And analyze the choice of decision action a t The reward score r for an action performed on the environment. t And new collection and verification status information t+1 Then, select the new state s. t+1The Q-value matrix Q is updated, and the corresponding probability matrix is ​​also updated. Finally, a new action a is selected based on the probability matrix. t+1 Output the results.

[0042] The Q-learning algorithm updates matrix Q using the Bellman optimal equation as follows:

[0043] in, This is the learning rate.

[0044] Specifically, such as Figure 3 As shown in the embodiments of this application, the method for predicting dangerous goods inspection items can first collect a large amount of historical business data from multiple dangerous goods freight-related business systems, obtain sample data based on the historical business data, and use the sample data to train a deep reinforcement learning model (Deep Q-Network, DQN). Specifically, it can include the following process: Step 301: Extract multiple business fields from multi-source business data.

[0045] Multi-source business data includes at least one of the following: dangerous goods waybills, dangerous goods declaration forms, dangerous goods transportation rules, etc.

[0046] For example, multiple business data tables can be obtained from the multi-source business data stored in multiple business systems corresponding to dangerous goods transportation. Each of the multiple business data tables includes multiple business fields, thus obtaining multiple business fields.

[0047] Step 302: Based on multiple business fields, obtain the embedding vectors corresponding to each business field.

[0048] For example, data from multiple business fields can be segmented to obtain a token ID sequence composed of multiple token IDs. Then, the token ID sequence can be input into a BERT (Bidirectional Encoder Representations from Transformers) model or other models with natural language processing capabilities to obtain the embedding vector corresponding to each token in the token sequence.

[0049] Step 303: Obtain sample data of the state based on the embedding vectors corresponding to the multiple business fields.

[0050] Different inspection tasks can correspond to different business fields, and different inspection tasks can correspond to different states. Therefore, the business fields corresponding to the states can be obtained based on the business fields corresponding to different inspection tasks, and the embedding vectors of the corresponding business fields can be used as features of the elements in the states.

[0051] For example, data samples such as waybill declaration data, dangerous goods declaration data, and dangerous goods rule data can be integrated into a state s. Multiple business fields related to waybill declarations are extracted from the waybill declaration data. Multiple business fields are also extracted from the dangerous goods declaration data and the dangerous goods rule data. These extracted business fields are then converted into embedding vectors, and the embedding vectors are integrated into a state s.

[0052] Each sample in the status sample data includes at least several check items, each corresponding to a check status and check result. Status information includes two states: checked and unchecked. The check result indicates whether a check item passed the check, and can generally be categorized into several types, such as passed, failed, invalid duplicate check, etc. Failed checks can be further classified into one or more of the following: triggering significant risk, high-risk violation, medium-risk violation, and low-risk violation.

[0053] Step 304: Train the deep reinforcement learning network based on at least the sample data of the state to obtain the trained deep reinforcement learning network.

[0054] It should be noted that training a deep reinforcement learning network can be done offline, online, or a combination of offline and online training.

[0055] Step 305: In response to receiving an inspection request for hazardous materials to be inspected, predict the inspection strategy corresponding to the inspection request through a trained deep reinforcement learning network.

[0056] The inspection strategy is used to recommend the order in which multiple inspection items are inspected to the user.

[0057] For example, in practical application scenarios, based on the business data such as waybills and declaration forms for the currently declared dangerous goods, the embedding vectors of multiple business fields are first extracted, and then the initial state is obtained. The initial state is then input into the trained deep reinforcement learning network, which can automatically output a detection strategy, usually in the form of a list of inspection items. This list includes which inspection items need to be inspected for the dangerous goods to be inspected, as predicted by the model, and the model will also recommend the inspection order of these items.

[0058] In practical applications, users can perform checks based on the list of checks recommended by the model and feed the actual check results and other data back to the deep reinforcement learning network. The environment in the deep reinforcement learning network will combine the detection results data fed back by the user to further optimize the trainable parameters in the current agent. This process is the online training process in the application stage of the model.

[0059] The following are specific examples based on practical application scenarios.

[0060] In one specific embodiment, such as Figure 4 As shown, the method proposed in this application embodiment may include the following process: Step 401: Obtain multiple business data tables from the multi-source business data stored in multiple business systems corresponding to dangerous goods transportation.

[0061] Each of the multiple business data tables includes multiple business fields.

[0062] Step 401 can be understood as the process of collecting and cleaning the original historical dangerous goods freight inspection data.

[0063] Freight-related business systems (such as freight information systems, security inspection information systems, and dangerous goods management systems) possess fine-grained data on waybill declarations and user interaction history, as well as detailed dangerous goods business data dictionaries and related dangerous goods rules (such as the air transport dangerous goods catalog, dangerous chemicals catalog, dangerous goods packaging instructions, dangerous goods classification, and labeling). This data is persistently and securely stored in a relational database. This database is not only the core support for platform operation but also the key data source for generating intelligent dangerous goods collection and transportation inspection reports.

[0064] In order to train an efficient and accurate collection and inspection checklist recommendation model, this application embodiment collects, cleans and transforms key business data from a relational database to ensure that the data not only retains its originality and integrity during the transformation process, but also meets the format requirements of the model input data.

[0065] Specifically, key business fields can be extracted from the raw data stored in multiple two-dimensional database tables. This process involves data collection and transformation, ensuring that the data retains its originality and integrity while meeting the format requirements of the recommendation system's input data. Specifically, this application embodiment requires extracting the following key data tables and corresponding business fields from the freight-related business system. For example, Table 1 is an example of an extracted business data table: Table 1

[0066] The master shipment declaration information form can be data selected from the freight declaration information forms recorded in the above-mentioned multiple business systems, with the freight nature (e.g., air freight nature) being "dangerous goods".

[0067] For example, "Master Order Number" and "Air Cargo Type" in Table 1 are business fields.

[0068] In analyzing business data, for image data such as electronic waybill attachments, unpacked photos, packaging instructions, and labels, automated recognition tools can be used to convert images or scanned paper documents into structured data to further extract useful information. Through data analysis, most image information is printed text, requiring the extraction of local features of the printed text to capture stroke and structural information. This application uses an OCR recognition engine, which can effectively extract and recognize features of printed English and numbers on the master waybill. After manual verification, the recognition rate for printed text is over 95%, identifying erroneous text and data, which can be further corrected and revised manually.

[0069] The collected data is standardized by unifying data formats and encoding rules to ensure data consistency. Simultaneously, duplicate, erroneous, and missing data are removed to improve data quality.

[0070] Step 402: Perform word segmentation on the data of multiple business fields to obtain a word sequence. Input the word sequence into the BERT model to obtain the embedding vector corresponding to each word in the word sequence.

[0071] Step 402 is the process of text vectorization, state and action representation.

[0072] Since the input waybill and dangerous goods declaration data have many fields, and most of them are text data, it is necessary to represent the key business data in numerical and vector form in order to further analyze the key business data.

[0073] Specifically, such as Figure 5 As shown, for text data that needs to be converted into vectors, the BERT model is selected for text vectorization. BERT is a pre-trained language model based on the Transformer architecture, capable of capturing the contextual information of text and generating high-quality word vector representations. The cleaned waybill information from the waybill declaration request and the text from the declaration information table in the dangerous goods declaration request are input into the BERT model to obtain the embedding vector corresponding to each text. These vectors contain the semantic information of the text, providing a data foundation for subsequent deep reinforcement learning. The specific process includes the following steps: Input text standardization: For example, Jieba or NLTK word segmentation tools can be used to segment the text to be processed, standardize the input text, and obtain multiple tokens.

[0074] Tokenization: The tokens after word segmentation are further segmented into sub-words. For example, the tokenizer method of the model can be used for automatic processing to output the token ID sequence. For example, the output token ID sequence is [101,6207,1523,4672,3422,5303,102].

[0075] Wherein, [CLS] (101) represents the start marker of the sequence; [SEP] (102) represents the end marker of the sequence; the others are the sub-word IDs after the actual words are segmented.

[0076] Next, the token ID is input into the BERT model. After processing by multiple encoding layers within the BERT model, an embedding vector of a specified dimension is output. For example, the BERT model may include 12 encoding layers, namely... Figure 5 The diagram shows encoding layers 1, 2, ..., 12. The BERT model uses text embedding to implement 12 layers of Transformer encoding, outputting a feature vector (embedding vector) of a specified dimension for each token, for example, a 768-dimensional vector. This converts the token ID into a 768-dimensional word vector and adds corresponding positional encoding. The model's output vector is then saved as a vector data file.

[0077] in, Figure 5 The numbers E0, E1, E2...E6 shown represent 6 tokens. A Represents a business field (Tablet PC), E B This indicates another business field (smartphone). E [CLS] This indicates the token corresponding to the start of the sequence. E [SEP] This indicates the token corresponding to the end marker of the sequence or the end marker of the business field. E 平板 E represents the token (embedding vector) corresponding to the word "flat". 电脑 E represents the token (embedding vector) corresponding to the word "computer". 智能 E represents the token (embedding vector) corresponding to the word "intelligent". 手机This represents the token (embedding vector) corresponding to the word "mobile phone". R[CLS] represents the feature vector (or output vector) output by the encoder after processing the token corresponding to the start of sequence marker. Rtablet, Rcomputer, R[SEP], RSmart, RMobile, and R[SEP] represent the feature vectors (or output vectors) output by the encoder after processing the tokens corresponding to the words "tablet", "computer", "smart", "mobile phone", and the end of sequence marker, respectively.

[0078] For example, after the above transformation, the key business fields and the target formats that need to be transformed are shown in Table 2 below: Table 2

[0079] After obtaining the embedding vectors of key business fields, business fields relevant to deep reinforcement learning training can be extracted and used as sample data to train the deep reinforcement learning network. For example, at least structured data of the states can be obtained.

[0080] In this embodiment, a deep reinforcement learning network (DQN) is used to implement intelligent recommendation of inspection strategies.

[0081] Deep Q-Network (DQN) is an algorithm that combines deep learning with Q-learning. This application's embodiments approximate the Q-function through a neural network to solve complex high-dimensional state problems. Specifically, the Q-Network algorithm is used to construct a dangerous goods acceptance and inspection item recommendation system. This system can dynamically update the priority and order of inspection items based on waybill declaration data, dangerous goods rule data, and historical inspection results. The system has a feedback mechanism, enabling adjustments based on operator feedback (such as missed or incorrect inspections). The main process is as follows: In high-dimensional state spaces, representing Q-values ​​using matrices is no longer feasible. DQN uses a neural network to approximate Q-values. The neural network takes state s as input and outputs the Q-values ​​corresponding to each action a, with trainable parameters. Optimization is achieved through gradient descent. The Q-function expression in the current deep Q-neural network to be trained is defined as follows: The expression for the Q-function in the target neural network is defined as follows: The training process is essentially the process of gradually optimizing the trainable parameters in the Q-function, that is, optimizing the trainable parameters... Updated after several steps .

[0082] Specifically, sample data is first collected by interacting with the environment. At each time step, action a is selected based on the current state s and executed. The next state s' and the immediate reward score r are observed, and (s, a, s', r) are stored as a sample.

[0083] A batch of samples is randomly drawn from the experience pool for network updates. For each sample (s, a, s', r), the target Q-value is updated according to the following formula:

[0084] The goal of training a Q-network is to minimize the Temporal Difference Error (TD), that is, to make the network's prediction of (s, a) Q(s, a; Get as close as possible to the calculated target value. The network parameters are updated by minimizing the squared error between the predicted Q-value and the target Q-value. Here, s' represents the next state, and a' represents the next action. This represents the updated trainable parameters (or network parameters).

[0085] The formula for calculating the mean squared error loss function is as follows:

[0086] Based on the value of the loss function, the trainable parameters of the DQN network (online or offline) are updated using gradient descent. Specifically, the loss function... Regarding trainable parameters in DQN networks Gradient descent is performed as follows:

[0087] in It is the learning rate.

[0088] For example, a deep reinforcement learning network (DQN) includes a main network and a target network. (e.g.) Figure 6 As shown, the specific steps for training DQN are as follows: Step 601: Model initialization.

[0089] Model initialization specifically includes performing the following initialization operations: Initialize the trainable parameters in the main network and the target network, obtain the initial state based on the sample data of the state, initialize the check item set, initialize the experience replay buffer, and determine the reward function.

[0090] This includes initializing the trainable parameters in the main network and the target network, specifically including initializing the weight parameters of each layer in the main network and the weight parameters of each layer in the target network, with the initial weight parameters of the main network and the target network being the same.

[0091] Initialize the experience replay buffer D by creating a fixed-size, first-in-first-out queue, initially empty. For example, the capacity can be set to N, and the update frequency C=10.

[0092] Initialization State: This refers to the initial state of the current waybill. For each new waybill, the initial state s0 includes the waybill declaration data, dangerous goods rule data, and the initial inspection state (all inspection items have not been inspected). It should be noted that the initial state may not include an inspection result field, but after at least one inspection item has been executed, state s1 must include an inspection result field to record the results corresponding to the inspected items.

[0093] Initializing the check item set: In one approach, the check item set can include a set of checked check items and a set of unchecked check items. Each time a check item is checked, it is removed from the unchecked set and added to the checked set. In this approach, during initialization, the checked item set can be set to empty, and the check items extracted from the standard transport check dictionary can be used as the initial set of items to be checked (i.e., the unchecked check item set). Alternatively, in other approaches, the check item set can include all check items, each with a flag indicating whether the check item has been checked.

[0094] In addition, action initialization is required, which is the process of determining the set of inspection items. The set of inspection items includes multiple candidate inspection items corresponding to the current inspection task. The action space (set of inspection items) is discrete, and each inspection item in the receiving and transporting inspection form corresponds to one action.

[0095] For example, some of the actions are shown in Table 3 below: Table 3

[0096] Among them, "Check aircraft type" is an action, "aircraft type" is the object of the action to be checked, and "if an inapplicable aircraft type is deleted or not displayed, it is unqualified" is a rule.

[0097] Determining the reward function, for example, may specifically involve determining the reward score corresponding to different inspection results, as well as determining the specific values ​​of the parameters that may be involved in the reward function.

[0098] For example, determining the reward function includes determining the values ​​corresponding to different inspection results: Triggering a major risk: r = +20, for example, the detection of absolutely prohibited items, such as Class 1 explosives; Risk violation inspection item: r = +10, for example, undeclared dangerous goods were found; Medium-risk violations: r = +5, such as packaging discrepancies, unclear labels, discrepancies between the net weight of the packaging and the declared weight, etc. Low-risk violation: r = +2, for example, inaccurate information provided by the sales agent; Check passed: r = +0.1. This bonus score is set to encourage efficiency, that is, for each item checked and the result is passed, the bonus score is increased by 0.1 points. Repeated invalid checks: r = -3, this design is to penalize redundant checks and improve efficiency; In addition, the reward function can include some additional reward parameters, such as time cost, completion reward and discount factor, learning rate and other parameters.

[0099] For example: time cost r_step = -0.1; completion reward r_final = 20; initial discount factor γ = 0.8; learning rate... =0.001.

[0100] The above reward function setting is only an example. In actual applications, the reward function can be set according to the following rules: determine the reward score corresponding to different inspection results, and determine the additional reward parameters: where the inspection result includes one or more of the following: triggering major risk, high-risk violation, medium-risk violation, low-risk violation, inspection passed, and repeated invalid inspection; the reward score corresponding to triggering major risk, high-risk violation, medium-risk violation, and low-risk violation decreases in that order.

[0101] The reward parameters include the discount factor, learning rate, time cost, and reward scores corresponding to completion rewards. The completion reward score is positive, while the reward score corresponding to the time cost of each training step is negative. The discount factor is a decimal between 0 and 1. Both the discount factor and the learning rate are hyperparameters.

[0102] After completing the initialization process described above, the DQN training process can begin. This training process can be online or offline, or it can involve first training offline using historical sample data, then applying the trained model to actual inspections, and finally training the model online in real time based on user feedback during actual inspections. The training data during the online training process is real-time.

[0103] The training process of DQN consists of two parts: an interactive loop and a learning loop. The two loops run in parallel until the preset number of training rounds is reached or the model performance converges.

[0104] The training process involves continuously executing interactive loops: Step 602: The agent selects an item from the set of items according to the ε-greedy policy, and uses it as action a in the t-th time step. t The object being inspected performs an action a in the environment. t。

[0105] For example, based on the current state s, an ε-greedy strategy is used to select action a.t That is, select the most recent next check item: In the initial training step, an unchecked item can be randomly selected with probability ε.

[0106] In subsequent training steps, the agent calculates the Q-values ​​corresponding to all items to be checked, and selects the item with the largest current Q-value with a probability of 1-ε.

[0107] In the early stages of training, the ε value is set relatively high to encourage exploration. As training progresses, ε gradually decays, and the agent increasingly relies on the learned policy (exploitation). Ideally, with iterations, the Q-value predicted by the main network will get closer and closer to the true target Q-value, and the loss will gradually decrease. The agent's policy will also tend towards optimality, i.e., to obtain the highest cumulative reward in the environment.

[0108] Step 603: Obtain the action to be executed, a t The corresponding inspection results.

[0109] Perform action a t This means checking the randomly selected item or the item with the largest Q value to obtain the check result (e.g., check result is pass or fail).

[0110] Step 604: The environment calculates action a based on the reward function. t The corresponding reward score r t .

[0111] The reward score is calculated based on the reward function determined in the initialization process described above.

[0112] Different reward scores are awarded based on different inspection results, and corresponding additional reward parameters are set according to different inspection types. For example, each step of the execution action has a small negative reward (time cost), such as r_step = -0.1. Therefore, the total reward is the above reward plus r_step. If all inspection items are completed and no problems are found, a completion reward r_final = 20 is given. If the inspection is completed but a problem is found, an additional reward can be added on the basis of the completion reward based on the inspection result. For example, if the inspection result is that a major risk is triggered, an additional 20 points will be added to the current reward score.

[0113] Step 605, Perform action a t The status information of the corresponding inspection item is updated to "inspected", and the updated status information and inspection result are written to status S. t In the middle, we obtain state S. t+1 .

[0114] Step 606: Store the experience data from multiple time steps into the experience playback buffer. The experience data includes state S. t Action a t Rewards r t and state S t+1 .

[0115] For example, the sample (s, a, r, s') is stored in the experience playback buffer D.

[0116] The training process that triggers the learning loop at regular intervals: Step 607: Randomly sample a batch from the multiple empirical data stored in the empirical playback buffer.

[0117] Step 608: Input a batch of empirical data into the main network, and predict action a through the main network. t The first Q value.

[0118] Randomly sample a batch of experience data from the experience playback buffer D. For example, a batch may include 32 experience data points.

[0119] Step 609: Predict action a using the target network. t The second Q value.

[0120] Step 610: Calculate the loss function based on the first Q value and the second Q value; update the trainable parameters in the main network based on the loss function.

[0121] For each sampled empirical data point (s) j , a j , r j , s j' ), calculate the Q value through the main network and the target network respectively, if s j' If the state is terminated (i.e., check completed), then the target value y j = r j , where j represents the training step in the most recent interactive loop.

[0122] otherwise,

[0123] Calculate Q(s) using the main network and the target network respectively. j , a j ; The Q-value calculated by the main network is the first Q-value, and the Q-value calculated by the target network is the second Q-value.

[0124] By minimizing the loss function This can be used to update the parameters of the current main network. For example, backpropagation can be performed on the main network, and gradient descent can be used to minimize the loss calculated in the previous step, thereby updating the parameters of the main network.

[0125] The loss function can be either mean squared error or Huber loss.

[0126] Step 611: Periodically synchronize the values ​​of trainable parameters in the main network to the target network.

[0127] Periodic updates can be set at predetermined intervals of a number of periods, where one period can be a training step. For example, the target network can be updated every C training steps (time steps). ' = .

[0128] For example, with C=1000, every 1000 steps, the parameters of the main network are completely copied to the target network. This provides a stable target for the learning process and is crucial for ensuring training convergence.

[0129] Following step 611, a dynamic update and feedback mechanism can be further employed: In practical applications, the inspection process for each waybill is a decision-making process, so new waybill data can be continuously collected and used for continuous training.

[0130] Feedback mechanism: The result of each check, i.e., the reward score r t This feedback will be sent to the system to update the Q network and improve strategies. Over time, the system will adjust the priority and order of inspection items based on new waybill data.

[0131] In practical use, for a new waybill, based on the current initial state, a pre-trained DQN can be used to select the unchecked item with the largest Q-value at each step until all checks are completed. The resulting order is the recommended check order, and the Q-value of each check item reflects its priority; items with higher Q-values ​​should be recommended for priority checking.

[0132] In summary, the solution proposed in this application uses the DQN (Dual Q Network) algorithm to solve the problem of recommending inspection items for dangerous goods collection and transportation. This solution combines waybill declaration data, dangerous goods rule data, and historical inspection feedback to dynamically generate the priority and order of inspection items and has the ability to update online.

[0133] The dangerous goods inspection item data prediction system provided in the embodiments of this application is described below. The dangerous goods inspection item data prediction system described below can be referred to in correspondence with the dangerous goods inspection item data prediction method described above.

[0134] This application provides a data prediction system for hazardous materials inspection items. The corresponding product for this system can be an electronic device deployed on the server side to implement the corresponding functions, such as a server or server cluster. Specifically, as... Figure 7 As shown, the system may include: The feature extraction module 701 is used to extract multiple business fields from multi-source business data; based on the multiple business fields, it obtains the embedding vectors corresponding to each business field. The multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules.

[0135] The status acquisition module 702 is used to obtain sample data of the status based on the embedding vectors corresponding to multiple business fields.

[0136] The training module 703 is used to train the deep reinforcement learning network based on the sample data of the state and the sample data of the check item set, so as to obtain the trained deep reinforcement learning network.

[0137] The prediction module 704 is used to predict the inspection policy corresponding to the inspection request in response to receiving an inspection request, through a trained deep reinforcement learning network. The inspection policy is used to recommend the order of multiple inspection items to the user.

[0138] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call a computer program in the memory 830 to execute the steps of a hazardous materials inspection item data prediction method, such as including: Multiple business fields are extracted from multi-source business data; based on these multiple business fields, embedding vectors corresponding to each business field are obtained; the multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules; sample data of states are obtained based on the embedding vectors corresponding to each business field; a deep reinforcement learning network is trained based on at least the sample data of states to obtain a trained deep reinforcement learning network; in response to receiving an inspection request for dangerous goods to be inspected, the trained deep reinforcement learning network is used to predict the inspection strategy corresponding to the inspection request, and the inspection strategy is used to recommend the inspection order of multiple inspection items to the user.

[0139] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0140] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the hazardous materials inspection item data prediction method provided in the above embodiments, such as including: Multiple business fields are extracted from multi-source business data; based on these multiple business fields, embedding vectors corresponding to each business field are obtained; the multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules; sample data of states are obtained based on the embedding vectors corresponding to each business field; a deep reinforcement learning network is trained based on at least the sample data of states to obtain a trained deep reinforcement learning network; in response to receiving an inspection request for dangerous goods to be inspected, the trained deep reinforcement learning network is used to predict the inspection strategy corresponding to the inspection request, and the inspection strategy is used to recommend the inspection order of multiple inspection items to the user.

[0141] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program for causing a processor to perform the steps of the methods provided in the above embodiments, such as including: Multiple business fields are extracted from multi-source business data; based on these multiple business fields, embedding vectors corresponding to each business field are obtained; the multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules; sample data of states are obtained based on the embedding vectors corresponding to each business field; a deep reinforcement learning network is trained based on at least the sample data of states to obtain a trained deep reinforcement learning network; in response to receiving an inspection request for dangerous goods to be inspected, the trained deep reinforcement learning network is used to predict the inspection strategy corresponding to the inspection request, and the inspection strategy is used to recommend the inspection order of multiple inspection items to the user.

[0142] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0143] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting data on hazardous materials inspection items, characterized in that, The method includes: Multiple business fields are extracted from multi-source business data; based on the multiple business fields, embedding vectors corresponding to the multiple business fields are obtained respectively; the multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules; Based on the embedding vectors corresponding to the multiple business fields, sample data of the state is obtained; At least based on the sample data of the stated state, the deep reinforcement learning network is trained to obtain the trained deep reinforcement learning network; In response to receiving an inspection request for hazardous materials to be inspected, the trained deep reinforcement learning network predicts the inspection strategy corresponding to the inspection request. The inspection strategy is used to recommend the order of inspection of multiple inspection items to the user.

2. The method according to claim 1, characterized in that, Extract multiple business fields from multi-source business data, including: Multiple business data tables are obtained from the multi-source business data stored in multiple business systems corresponding to dangerous goods transportation; each of the multiple business data tables includes multiple business fields.

3. The method according to claim 1, characterized in that, Based on the multiple business fields, the embedding vectors corresponding to each of the multiple business fields are obtained, including: The data from the multiple business fields are segmented into words to obtain a word sequence; The word sequence is input into the BERT model to obtain the embedding vector corresponding to each word in the word sequence.

4. The method according to any one of claims 1-3, characterized in that, The deep reinforcement learning network includes a main network and a target network; Training a deep reinforcement learning network based at least on sample data from the stated state includes: Initialize the trainable parameters in the main network and the target network, and obtain the initial state based on the sample data of the state; Initialize the check item set; Initialize the experience replay buffer; Determine the reward function; The training process involves continuously executing interactive loops: The agent selects an item from the set of items according to an ε-greedy policy, and uses it as the object of inspection for the action in the t-th time step, and then performs the action in the environment. Obtain the inspection result corresponding to the execution of the action; The environment calculates the reward score corresponding to the action based on the reward function; Update the status information of the check item corresponding to the action to "checked", and write the updated status information and the check result into the status to obtain the next status; The experience data from multiple time steps is stored in the experience replay buffer, and the experience data includes the state, the action, the reward score, and the next state; The training process that triggers the learning loop at regular intervals: A batch is randomly sampled from multiple empirical data stored in the empirical playback buffer; The empirical data of the batch is input into the main network, and the first Q value of the action is predicted by the main network. Predict the second Q value of the action using the target network; Calculate the loss function based on the first Q value and the second Q value; The trainable parameters in the main network are updated according to the loss function.

5. The method according to claim 4, characterized in that, Determine the reward function, including: Determine the corresponding reward scores for different inspection results, and determine the additional reward parameters; The inspection results include one or more of the following: triggering major risk, high-risk violation, medium-risk violation, low-risk violation, inspection passed, and repeated invalid inspection; the reward points corresponding to triggering major risk, high-risk violation, medium-risk violation, and low-risk violation decrease in that order. The reward parameters include a discount factor, a learning rate, a time cost, and a reward score corresponding to the completion reward; wherein, the reward score for the completion reward is a positive value, and the reward score corresponding to the time cost of each training step is a negative value; the discount factor is a decimal between 0 and 1; both the discount factor and the learning rate are hyperparameters.

6. The method according to claim 4, characterized in that, After updating the trainable parameters in the main network, the method further includes: The values ​​of trainable parameters in the main network are periodically synchronized to the target network.

7. The method according to claim 4, characterized in that, The next item to be checked is selected according to the ε-greedy strategy, including: Randomly select an unchecked item from the set of check items with probability ε; or, The unchecked item with the largest Q value is selected from the set of check items with a probability of 1-ε; the Q value is predicted by the agent.

8. A data prediction system for dangerous goods inspection items, characterized in that, The system includes: The feature extraction module is used to extract multiple business fields from multi-source business data; and based on the multiple business fields, to obtain the embedding vectors corresponding to the multiple business fields respectively; the multi-source business data includes at least one of dangerous goods waybills, dangerous goods declaration forms, and dangerous goods transportation rules; The status acquisition module is used to obtain sample data of the status based on the embedding vectors corresponding to the multiple business fields respectively; The training module is used to train the deep reinforcement learning network based on the sample data of the state and the sample data of the check item set, so as to obtain the trained deep reinforcement learning network. The prediction module is used to predict the inspection strategy corresponding to the inspection request in response to receiving an inspection request, through the trained deep reinforcement learning network. The inspection strategy is used to recommend the order of multiple inspection items to the user.

9. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the dangerous goods inspection item data prediction method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the dangerous goods inspection item data prediction method according to any one of claims 1 to 7.

Images